Final Program and Abstracts

Sponsored by the SIAM Activity Group on Optimization

The SIAM Activity Group on Optimization fosters the development of optimization theory, methods, and software -- and in particular the development and analysis of efficient and effective methods, as well as their implementation in high-quality software. It provides and encourages an environment for interaction among applied mathematicians, computer scientists, engineers, scientists, and others active in optimization. In addition to endorsing various optimization meetings, this activity group organizes the triennial SIAM Conference on Optimization, awards the SIAG/Optimization Prize, and sponsors minisymposia at the annual meeting and other SIAM conferences. The activity group maintains a wiki, a membership directory, and an electronic discussion group in which members can participate, and also publishes a newsletter - SIAG/OPT News and Views.

Society for Industrial and Applied Mathematics 3600 Market Street, 6th Floor Philadelphia, PA 19104-2688 USA Telephone: +1-215-382-9800 Fax: +1-215-386-7999 Conference E-mail: [email protected] SIAM 2014 Events Mobile App Conference Web: www.siam.org/meetings/ Scan the QR code with any QR reader and Membership and Customer Service: download the TripBuilder EventMobile™ (800) 447-7426 (US & Canada) or app to your iPhone, iPad, iTouch or +1-215-382-9800 (worldwide) Android mobile device.You can also visit www.siam.org/meetings/op14 www.tripbuilder.com/siam2014events 2 2014 SIAM Conference on Optimization

Table of Contents SIAM Registration Desk Child Care The SIAM registration desk is located in For local childcare information, please Program-at-a-Glance the Golden Foyer. It is open during the contact the concierge at the Town and ...... Separate handout following hours: Country Resort & Convention Center at General Information...... 2 Sunday, May 18 +1-619-291-7131. Get-togethers...... 4 5:00 PM - 8:00 PM Invited Plenary Presentations...... 6 Prize Lecture...... 8 Corporate Members Monday, May 19 Minitutorials...... 9 and Affiliates 7:00 AM - 5:00 PM Program Schedule...... 11 SIAM corporate members provide Poster Session...... 25 their employees with knowledge about, access to, and contacts in the applied Abstracts...... 67 Tuesday, May 20 mathematics and computational sciences 7:45 AM - 5:30 PM Speaker and Organizer Index...... 187 community through their membership Conference Budget..... Inside Back Cover benefits. Corporate membership is more Property Map...... Back Cover Wednesday, May 21 than just a bundle of tangible products

7:45 AM - 5:00 PM and services; it is an expression of Organizing Committee Co-Chairs support for SIAM and its programs. SIAM is pleased to acknowledge its Miguel F. Anjos Thursday, May 22 corporate members and sponsors. Polytechnique Montreal, Canada 7:45 AM - 5:00 PM In recognition of their support, non- member attendees who are employed by Michael Jeremy Todd the following organizations are entitled Cornell University, USA Hotel Address to the SIAM member registration rate. Town and Country Resort Organizing Committee & Convention Center Aharon Ben-Tal 500 Hotel Circle North Corporate Institutional Technion - Israel Institute of Technology, Israel Members San Diego, California 92108 USA Andrew Conn The Aerospace Corporation IBM Research, USA Phone Number: +1-619-291-7131 Air Force Office of Scientific Research Mirjam Dür Toll Free Reservations: (800)-772-8527 University of Trier, (USA and Canada) AT&T Laboratories - Research Michael Hintermüller Fax: +1-619-294-4681 Bechtel Marine Propulsion Laboratory Humboldt-Universität zu Berlin, Germany Hotel Website: www.towncountry.com/ The Boeing Company Etienne de Klerk CEA/DAM Tilburg University, The Netherlands Department of National Defence (DND/ Jon Lee Hotel Telephone Number CSEC) University of Michigan, USA To reach an attendee or leave a message, DSTO- Defence Science and Todd Munson call +1-619-291-7131. If the attendee Technology Organisation Argonne National Laboratory, USA is a hotel guest, the hotel operator can Hewlett-Packard Warren Powell connect you with the attendee’s room. Princeton University, USA IBM Corporation Daniel Ralph IDA Center for Communications University of Cambridge, United Kingdom Hotel Check-in Research, La Jolla Ariela Sofer and Check-out Times IDA Center for Communications George Mason University, USA Check-in time is 3:00 PM and check-out Research, Princeton Akiko Yoshise time is 11:00 PM. Institute for Computational and University of Tsukuba, Japan Experimental Research in Mathematics (ICERM) 2014 SIAM Conference on Optimization 3

Institute for Defense Analyses, Center Leading the applied Join onsite at the registration desk, go to for Computing Sciences mathematics community… www.siam.org/joinsiam to join online or Lawrence Berkeley National Laboratory download an application form, or contact Join SIAM and save! SIAM Customer Service: Lockheed Martin SIAM members save up to $130 on Los Alamos National Laboratory full registration for the 2014 SIAM Mathematical Sciences Research Conference on Optimization! Join Telephone: +1-215-382-9800 Institute your peers in supporting the premier (worldwide); or 800-447-7426 (U.S. and professional society for applied Canada only) Max-Planck-Institute for Dynamics of mathematicians and computational Fax: +1-215-386-7999 Complex Technical Systems scientists. SIAM members receive E-mail: [email protected] Mentor Graphics subscriptions to SIAM Review, SIAM Postal mail: Society for Industrial and National Institute of Standards and News, and Unwrapped, and enjoy Applied Mathematics, 3600 Market Technology (NIST) substantial discounts on SIAM books, Street, 6th floor, Philadelphia, PA 19104- journal subscriptions, and conference 2688 USA National Security Agency (DIRNSA) registrations. Oak Ridge National Laboratory, managed by UT-Battelle for the Department of Energy If you are not a SIAM member and paid Standard Audio/Visual the Non-Member or Non-Member Set-Up in Meeting Rooms Sandia National Laboratories Mini Speaker/Organizer rate to attend the SIAM does not provide computers for Schlumberger-Doll Research conference, you can apply the difference any speaker. When giving an electronic between what you paid and what a Tech X Corporation presentation, speakers must provide their member would have paid ($130 for a U.S. Army Corps of Engineers, Engineer own computers. SIAM is not responsible Non-Member and $65 for a Non-Member Research and Development Center for the safety and security of speakers’ Mini Speaker/Organizer) towards a computers. United States Department of Energy SIAM membership. Contact SIAM Customer Service for details or join at List current April 2014. the conference registration desk. The Plenary Session Room will have two (2) screens, one (1) data projector and one (1) overhead projector. Cables If you are a SIAM member, it only costs or adaptors for Apple computers are not $10 to join the SIAM Activity Group Funding Agency supplied, as they vary for each model. on Optimization (SIAG/OPT). As a SIAM and the Conference Organizing Please bring your own cable/adaptor if SIAG/OPT member, you are eligible Committee wish to extend their thanks using an Apple computer. for an additional $10 discount on this and appreciation to the U.S. National conference, so if you paid the SIAM All other concurrent/breakout rooms Science Foundation and the U.S. member rate to attend the conference, will have one (1) screen and one (1) data Department of Energy (DOE), Office you might be eligible for a free SIAG/ projector. Cables or adaptors for Apple of Science, for their support of this OPT membership. Check at the computers are not supplied, as they vary conference. registration desk. for each model. Please bring your own cable/adaptor if using an Apple computer. Overhead projectors will be provided Free Student Memberships are available only if requested. to students who attend an institution that is an Academic Member of SIAM, are members of Student Chapters of SIAM, If you have questions regarding or are nominated by a Regular Member availability of equipment in the meeting of SIAM. room of your presentation, or to request an overhead projector for your session, please see a SIAM staff member at the registration desk. 4 2014 SIAM Conference on Optimization

E-mail Access SIAM Books and Journals Get-togethers SIAM attendees will have complimentary Display copies of books and • Welcome Reception wireless Internet access in the sleeping complimentary copies of journals and Poster Session rooms, public areas and meeting space are available on site. SIAM books Monday, May 19 of The Town and Country Resort & are available at a discounted price 8:00 PM – 10:00 PM Convention Center. during the conference. If a SIAM In addition, a limited number of books representative is not available, • Business Meeting computers with Internet access will be completed order forms and payment (open to all SIAG/OPT members) available during registration hours. (credit cards are preferred) may be taken Wednesday, May 21 to the SIAM registration desk. The 8:00 PM - 8:45 PM books table will close at 4:30 PM on Registration Fee Includes Thursday, May 22. Complimentary beer and wine will be served. • Admission to all technical sessions

• Business Meeting (open to SIAG/ Table Top Displays OPT members) AMPL • Coffee breaks daily Now Publishers Please Note SIAM is not responsible for the safety • Room set-ups and audio/visual SIAM equipment and security of attendees’ computers. Do not leave your laptop computers • Welcome Reception and Poster Conference Sponsor unattended. Please remember to turn Session off your cell phones, pagers, etc. during sessions. Job Postings Please check with the SIAM registration Recording of Presentations desk regarding the availability of job Audio and video recording of postings or visit http://jobs.siam.org. presentations at SIAM meetings is prohibited without the written Name Badges permission of the presenter and SIAM. Important Notice to Poster A space for emergency contact Presenters information is provided on the back of The poster session is scheduled for your name badge. Help us help you in Social Media Monday, May 19 at 8:00 PM. Poster the event of an emergency! SIAM is promoting the use of social presenters are requested to set up their media, such as Facebook and Twitter, poster material on the provided 4’ x 8’ in order to enhance scientific discussion poster boards in the Town and Country Comments? at its meetings and enable attendees Room between the hours of 2:00 PM Comments about SIAM meetings are to connect with each other prior to, and 8:00 PM. All materials must be encouraged! Please send to: during and after conferences. If you posted by 8:00 PM on Monday, May are tweeting about a conference, please Cynthia Phillips, SIAM Vice President 19, the official start time of the session. use the designated hashtag to enable for Programs ([email protected]). Posters will remain on display through other attendees to keep up with the Thursday, May 22 at 10:00 AM, at which Twitter conversation and to allow better time poster displays must be removed. archiving of our conference discussions. Posters remaining after this time will be The hashtag for this meeting is discarded. SIAM is not responsible for #SIAMOP14. discarded posters. 2014 SIAM Conference on Optimization 5

The SIAM 2014 Events Mobile App Powered by

TripBuilder® To enhance your conference experience, we’re providing a state-of-the-art mobile app to give you important conference information right at your fingertips. With this TripBuilder EventMobile™ app, you can: • Create your own custom schedule • View Sessions, Speakers, Exhibitors and more • Take notes and export them to your email • View Award-Winning TripBuilder Recommendations for the meeting location • Get instant Alerts about important conference info

SIAM 2014 Events Mobile App Scan the QR code with any QR reader and download the TripBuilder EventMobile™ app to your iPhone, iPad, iTouch or Android mobile device. To access the app or the HTMLS version, you can also visit www. tripbuilder.com/siam2014events 6 2014 SIAM Conference on Optimization

Invited Plenary Speakers ** All Invited Plenary Presentations will take place in the Golden Ballroom **

Monday, May 19

8:15 AM - 9:00 AM IP1 Universal Gradient Methods , Université Catholique de Louvain, Belgium

1:00 PM - 1:45 PM IP2 Optimizing and Coordinating Healthcare Networks and Markets Retsef Levi, Massachusetts Institute of Technology, USA

Tuesday, May 20

8:15 AM - 9:00 AM IP3 Recent Progress on the Diameter of Polyhedra and Simplicial Complexes Francisco Santos, University of Cantabria,

9:00 AM - 9:45 AM IP4 Combinatorial Optimization for National Security Applications Cynthia Phillips, Sandia National Laboratories, USA 2014 SIAM Conference on Optimization 7

Invited Plenary Speakers ** All Invited Plenary Presentations will take place in the Golden Ballroom **

Wednesday, May 21

8:15 AM - 9:00 AM IP5 The Euclidean Distance Degree of an Algebraic Set Rekha Thomas, University of Washington, USA

1:00 PM - 1:45 PM IP6 Modeling Wholesale Electricity Markets with Hydro Storage Andy Philpott, University of Auckland, New Zealand

Thursday, May 22

8:15 AM - 9:00 AM IP7 Large-Scale Optimization of Multidisciplinary Engineering Systems Joaquim Martins, University of Michigan, USA

1:00 PM - 1:45 PM IP8 A Projection Hierarchy for Some NP-hard Optimization Problems Franz Rendl, Universitat Klagenfurt, 8 2014 SIAM Conference on Optimization

Prize Lecture ** The SIAG/OPT Prize Lecture will take place in the Golden Ballroom **

Tuesday, May 20 1:30 PM - 2:15 PM

SP1 SIAG/OPT Prize Lecture: Efficiency of the Simplex and Policy Iteration Methods for Markov Decision Processes , Stanford University, USA 2014 SIAM Conference on Optimization 9

Minitutorials ** The Minitutorials will take place in the Golden Ballroom **

Monday, May 19 9:30 AM - 11:30 AM

MT1: Polynomial Optimization

Polynomial optimization, which asks to minimize a polynomial function over a domain defined by polynomial inequalities, models broad classes of hard problems from optimization, control, or geometry. New algorithmic approaches have emerged recently for computing the global minimum, by combining tools from algebra (sums of squares of polynomials) and functional analysis (moments of measures) with semidefinite optimization. In this minitutorial we will present the general methodology, illustrate it on several selected examples and discuss existing software. Organizers and Speakers: Didier Henrion, University of Toulouse, France Monique Laurent, CWI, Amsterdam, Netherlands

Wednesday, May 21 9:30 AM - 11:30 AM

MT2: Mixed-integer Nonlinear Optimization

Mixed-integer nonlinear programming problems (MINLPs) combine the combinatorial complexity of discrete decisions with the numerical challenges of nonlinear functions. In this minitutorial, we will describe applications of MINLP in science and engineering, demonstrate the importance of building “good” MINLP models, discuss numerical techniques for solving MINLPs, and survey the forefront of ongoing research topics in this important and emerging area. Organizers and Speakers: Sven Leyffer, Argonne National Laboratory, USA Jeff Linderoth, University of Wisconsin, Madison, USA James Luedtke, University of Wisconsin, Madison, USA 10 2014 SIAM Conference on Optimization

SIAM Activity Group on Optimization (SIAG/OPT) www.siam.org/activity/optimization

A GREAT WAY TO GET INVOLVED! Collaborate and interact with mathematicians and applied scientists whose work involves optimization.

ACTIVITIES INCLUDE: • Special sessions at SIAM Annual Meetings • Triennial conference • SIAG/Optimization Prize • SIAG/OPTnewsletter News and Views • Wiki

BENEFITS OF SIAG/OPT MEMBERSHIP: • Listing in the SIAG’s online membership directory • Additional $10 discount on registration for the SIAM Conference on Optimization (excludes students) • Electronic communications about recent developments in your specialty • Eligibility for candidacy for SIAG/OPT office • Participation in the selection of SIAG/OPT officers ELIGIBILITY: • Be a current SIAM member. COST: • $10 per year • Student SIAM members can join 2 activity groups for free! 2014-16 SIAG/OPT OFFICERS • Chair: Juan Meza, University of California Merced • Vice Chair: Martine Labbe, Université Libre de Bruxelles • Program Director: Michael Friedlander, University of British Columbia • Secretary/Treasurer: Kim-Chuan Toh, National University of Singapore • Newsletters Editor: Sven Leyffer, Argonne National Laboratory and Franz Rendl, Universität Klagenfurt

TO JOIN:

SIAG/OPT: my.siam.org/forms/join_siag.htm SIAM: www.siam.org/joinsiam 2014 SIAM Conference on Optimization 11

OP14 Program 12 2014 SIAM Conference on Optimization

Sunday, May 18 Monday, May 19 Monday, May 19 MT1 Polynomial Optimization Registration Registration 9:30 AM-11:30 AM 5:00 PM-8:00 PM 7:00 AM-5:00 PM Room:Golden Ballroom Room:Golden Foyer Room:Golden Foyer Chair: Monique Laurent, CWI, Amsterdam, Netherlands Chair: Didier Henrion, University of Toulouse, France Welcoming Remarks Polynomial optimization, which asks 8:00 AM-8:15 AM to minimize a polynomial function Room:Golden Ballroom over a domain defined by polynomial inequalities, models broad classes of hard problems from optimization, control, or geometry. New algorithmic approaches have emerged recently for IP1 computing the global minimum, by combining tools from algebra (sums of Universal Gradient Methods squares of polynomials) and functional 8:15 AM-9:00 AM analysis (moments of measures) with semidefinite optimization. In this Room:Golden Ballroom minitutorial we will present the general Chair: Michael J. Todd, Cornell University, methodology, illustrate it on several USA selected examples and discuss existing In Convex Optimization, numerical software. schemes are always developed for Speakers: some specific problem classes. One of Didier Henrion, University of Toulouse, the most important characteristics of France such classes is the level of smoothness Monique Laurent, CWI, Amsterdam, of the objective function. Methods Netherlands for nonsmooth functions are different from the methods for smooth ones. However, very often the level of smoothness of the objective is difficult to estimate in advance. In this talk we present algorithms which adjust their behavior in accordance to the actual level of smoothness observed during the minimization process. Their only input parameter is the required accuracy of the solution. We discuss also the abilities of these schemes in reconstructing the dual solutions. Yurii Nesterov Université Catholique de Louvain, Belgium

Coffee Break 9:00 AM-9:30 AM Room:Town & Country 2014 SIAM Conference on Optimization 13

Monday, May 19 Monday, May 19 Monday, May 19 MS1 MS3 MS4 Applications in Healthcare: Advances in Modeling Advances in Mixed-Integer Patient Scheduling and Algorithms for Linear and Nonlinear 9:30 AM-11:30 AM Distributionally Robust Optimization Room:Pacific Salon 1 Optimization 9:30 AM-11:30 AM There are multiples challenges in the 9:30 AM-11:30 AM Room:Pacific Salon 4 patient scheduling problem. In this Room:Pacific Salon 3 Mixed-Integer Linear and Nonlinear session we present different application Robust optimization is a rapidly Optimization is an exciting research settings such as the operating room and advancing field. In this minisymposium area that is fundamental in all kinds of cancer treatment centers. Although the we will bring together researchers applied domains. This session provides patient is central, many related resources working in the organizer’s group to a glimpse of such an excitement with have to be aligned to optimize the flow present both modeling and algorithmic talks ranging from theory to applications and reduce delays. Techniques such advances on this topic. More specifically, with a strong methodological flavor in as stochastic programming and online the speakers will present the work on common. algorithms are developed and prove to methods for solving distributionally Organizer: Andrea Lodi be very effective. robust optimization problems where University of Bologna, Italy Organizer: Nadia Lahrichi more than just the first two moments are 9:30-9:55 How Good Are Sparse École Polytechnique de Montréal, Canada specified. They will present results for Cutting-Planes? 9:30-9:55 Online Stochastic models where robustness is specified in Santanu S. Dey, Marco Molinaro, and Qianyi Optimization of Radiotherapy Patient both the first and the second stage, and Wang, Georgia Institute of Technology, Scheduling problems where the uncertainty is in the USA Antoine Legrain, École Polytechnique function elicitation. Specific applications 10:00-10:25 The Two Facility Splittable de Montréal, Canada; Marie-Andrée will be discussed. Flow Arc Set Fortin, Centre Intégré de Cancérologie, Sanjeeb Dash, and Oktay Gunluk, IBM T.J. Organizer: Sanjay Mehrotra Canada; Nadia Lahrichi and Louis- Watson Research Center, USA; Laurence , USA Martin Rousseau, École Polytechnique de Wolsey, Université Catholique de Louvain, Montréal, Canada 9:30-9:55 Distributionally Robust Belgium Modeling Frameworks and Results for 10:00-10:25 Dynamic Extensible Bin 10:30-10:55 Finitely Convergent Least-Squares Packing with Applications to Patient Decomposition Algorithms for Sanjay Mehrotra, Northwestern University, Scheduling Two-Stage Stochastic Pure Integer USA Bjorn Berg, George Mason University, USA; Programs Brian Denton, University of Michigan, 10:00-10:25 A Cutting Surface Simge Kucukyavuz, Ohio State University, USA Algorithm for Semi-infinite Convex USA; Minjiao Zhang, University of 10:30-10:55 Scheduling Programming and Moment Robust Alabama, USA Optimization Chemotherapy Patient Appointments 11:00-11:25 Water Networks David Papp, Massachusetts General Hospital Martin L. Puterman, University of British Optimization: MILP vs MINLP Insights and Harvard Medical School, USA; Sanjay Columbia, Canada Cristiana Bragalli, University of Bologna, Mehrotra, Northwestern University, USA 11:00-11:25 Chance-Constrained Italy; Claudia D’Ambrosio, CNRS, Surgery Planning under Uncertain or 10:30-10:55 A Robust Additive France; Andrea Lodi and Sven Wiese, Ambiguous Surgery Time Multiattribute Preference Model using University of Bologna, Italy Yan Deng, Siqian Shen, and Brian Denton, a Nonparametric Shape-Preserving University of Michigan, USA Perturbation Jian Hu and Yung-wen Liu, University of Michigan, Dearborn, USA; Sanjay Mehrotra, Northwestern University, USA 11:00-11:25 Distributionally-Robust Support Vector Machines for Machine Learning Changhyeok Lee and Sanjay Mehrotra, Northwestern University, USA 14 2014 SIAM Conference on Optimization

Monday, May 19 Monday, May 19 Monday, May 19 MS6 MS7 MS5 Linear and Nonlinear First Order Methods and Multidisciplinary Design Optimization and Applications - Part I of II Optimization Applications 9:30 AM-11:30 AM 9:30 AM-11:30 AM 9:30 AM-11:30 AM Room:Sunrise Room:Pacific Salon 5 Room:Pacific Salon 6 For Part 2 see MS19 In this minisymposium we address The presenters will talk about recent The renewed interest for gradient computational challenges arising in advances in linear and nonlinear methods dates back to the seminal work the design optimization of engineering optimization algorithms, developed for of Barzilai and Borwein in 1988, and systems that involve multiple applications in machine learning, support since then, several new gradient methods disciplines and thus multiple physics. vector machines, and compressed have been proposed. The research Multidisciplinary design optimization sensing. about first-order methods has been also (MDO) problems involve the solution stimulated by their successful use in Organizer: Hande Y. Benson practice, for instance in the application of coupled simulations, some of which Drexel University, USA might require the solution of PDEs with to certain ill-posed inverse problems, 9:30-9:55 A Stochastic Optimization millions of degrees of freedom. Both where they exhibit a significant Algorithm to Solve Sparse Additive the objective and constraint functions regularizing effect. The goal of this Model minisymposium is to present both theory involved are typically nonlinear and Ethan Fang, Han Liu, and Mengdi Wang, and applications of state-of-the-art are expensive to compute. The talks Princeton University, USA in this minisymposium will discuss gradient methods, to better understand 10:00-10:25 An Exterior-Point Method their features and potentialities. methods for handling discipline coupling, for Support Vector Machines decompositions techniques, and efficient Igor Griva, George Mason University, USA Organizer: Gerardo Toraldo methods for computing the gradients of University of Naples “Frederico II”, Naples, 10:30-10:55 Cubic Regularization in coupled systems. Italy Interior-Point Methods Organizer: Joaquim Martins Hande Y. Benson, Drexel University, USA; Organizer: Luca Zanni University of Michigan, USA David Shanno, Rutgers University, USA University of Modena and Reggio Emilia, Italy 9:30-9:55 A General Theory for 11:00-11:25 L1 Regularization Via the Coupling Computational Models and Parametric Simplex Method 9:30-9:55 Faster Gradient Descent their Derivatives Robert J. Vanderbei, Princeton University, Kees van den Doel and Uri M. Ascher, John T. Hwang and Joaquim Martins, USA University of British Columbia, Canada University of Michigan, USA 10:00-10:25 An Efficient Gradient 10:00-10:25 A Matrix-free Trust-region Method Using the Yuan Steplength SQP Method for Reduced-space PDE- Roberta D. De Asmundis, Sapienza – constrained Optimization: Enabling Università di Roma, Italy; Daniela di Modular Multidisciplinary Design Serafino, Second University of Naples, Optimization Italy; Gerardo Toraldo, University of Jason E. Hicken and Alp Dener, Rensselaer Naples “Frederico II”, Naples, Italy Polytechnic Institute, USA 10:30-10:55 Iteration-Complexity with 10:30-10:55 Using Graph Theory Averaged Nonexpansive Operators: to Define Problem Formulation in Application to Convex Programming Openmdao Jingwei Liang and Jalal Fadili, Université de Justin Gray, NASA, USA Caen, France; Gabriel Peyre, CEREMADE Universite Paris 9 Dauphine, France 11:00-11:25 Rethinking MDO for Dynamic Engineering System Design 11:00-11:25 On Subspace James T. Allison, University of Illinois at Minimization Conjugate Gradient Urbana-Champaign, USA Method Yuhong Dai, Chinese Academy of Sciences, China 2014 SIAM Conference on Optimization 15

Monday, May 19 Monday, May 19 Monday, May 19 MS8 MS9 MS10 Large-Scale, Distributed, Optimization Under Optimization Methods and and Multilevel Optimization Stochastic Order Constraints Applications Algorithms 9:30 AM-11:30 AM 9:30 AM-11:30 AM 9:30 AM-11:30 AM Room:Towne Room:Royal Palm 1 Room:Sunset Optimization problems with stochastic This session contains four talks about This minisymposium highlights order relations as constraints are newly numerical methods for nonlinear recent developments in the design of proposed structure in optimization. It optimization and applications, including algorithms for solving constrained is motivated by reflecting risk aversion. limited memory techniques, trust region nonlinear optimization problems. In These type of constraints impose a algorithms and algorithms for wirless particular, the focus is on solving very requirement on distribution functions communication problems and stocastic large problems that are intractable for or their transforms thus controlling the nonlinear programming. current methods, such as those arising probability of undesirable events. The Organizer: Ya-Xiang Yuan in PDE-constrained optimization, approach exhibits interesting relations to Chinese Academy of Sciences, P.R. of China expected utility theory, rank-dependent and those that involve incorporating 9:30-9:55 On a Reduction of utility theory, and to the theory of data from huge datasets. The methods Cardinality to Complementarity in discussed in this minisymposium coherent measures of risk. The talk Sparse Optimization are designed to be computationally of this minisymposium will address Oleg Burdakov, Linköping University, efficient, but are also designed with theoretical and numerical questions Sweden; Christian Kanzow and Alexandra careful consideration of convergence associated with these problems, as well Schwartz, University of Würzburg, guarantees. as some applications. Germany Organizer: Frank E. Curtis Organizer: Darinka Dentcheva 10:00-10:25 Sum Rate Maximization Lehigh University, USA Stevens Institute of Technology, USA Algorithms for Mimo Relay Networks in Wireless Communications 9:30-9:55 Stochastic Dominance in Organizer: Daniel Robinson Cong Sun, University of Posts and Shape Optimization under Uncertain Johns Hopkins University, USA Telecommunications, China; Eduard Loading Jorswieck, Dresden University of 9:30-9:55 Adaptive Observations Ruediger Schultz, University of Duisburg- Technology, Germany; Yaxiang Yuan, And Multilevel Optimization In Data Essen, Germany Assimilation Chinese Academy of Sciences, China 10:00-10:25 Stability of Optimization Philippe L. Toint, Facultés Universitaires 10:30-10:55 Quasi-Newton methods for Problems with Stochastic Dominance Notre-Dame de la Paix, Belgium; Serge the Trust-Region Subproblem Constraints Gratton, ENSEEIHT, Toulouse, France; Jennifer Erway, Wake Forest University, Werner Roemisch, Humboldt University at Monserrat Rincon-Camacho, CERFACS, USA; Roummel F. Marcia, University of Berlin, Germany France California, Merced, USA 10:30-10:55 Regularization Methods 10:00-10:25 A Class of Distributed 11:00-11:25 Penalty Methods with for Stochastic Order Constrained Optimization Methods with Event- Stochastic Approximation for Problems Triggered Communication Stochastic Nonlinear Programming Gabriela Martinez, Cornell University, USA; Michael Ulbrich and Martin Meinel, Xiao Wang, University of Chinese Academy Darinka Dentcheva, Stevens Institute of Technische Universität München, of Sciences, China; Shiqian Ma, Chinese Technology, USA Germany University of Hong Kong, Hong Kong; 10:30-10:55 Multilevel Methods for 11:00-11:25 Numerical Methods for Ya-Xiang Yuan, Chinese Academy of Very Large Scale NLPs Resulting from Problems with Multivariate Stochastic Sciences, P.R. of China PDE Constrained Optimization Dominance Constraints Stefan Ulbrich, Technische Universität Darinka Dentcheva and Eli Wolfhagen, Darmstadt, Germany Stevens Institute of Technology, USA 11:00-11:25 A Two-Phase Augmented Lagrangian Filter Method Sven Leyffer, Argonne National Laboratory, USA 16 2014 SIAM Conference on Optimization

Monday, May 19 Monday, May 19 Monday, May 19 MS11 MS88 Lunch Break Evaluating Nonsmooth Advanced Methods 11:30 AM-1:00 PM Sensitivity Information for Exploiting Structures of Conic Attendees on their own Optimization Progamming 9:30 AM-11:30 AM 9:30 AM-11:30 AM Room:Royal Palm 2 Room:Pacific Salon 2 Methods for local optimization of Efficient methods for conic programming nonsmooth functions require local problems provide computational sensitivity information in the form framework for various applications of generalized derivatives, such as including statistics, quantum chemistry subgradients, elements of Clarke’s and combinatorial optimization. We generalized gradient, piecewise discuss several advanced methods which linearizations, or lexicographic exploit the properties of the cones, derivatives. Since generalized derivatives in particular, positive semidefinite are required at each iteration of many matrices, the second-order cone and methods, numerical methods for their the doubly-nonnegative cone. Software evaluation must be computationally implementation for the methods will be cheap and automatable. However, the also properties discussed in this session. failure of conventional calculus rules Organizer: Makoto Yamashita for generalized derivatives complicates Tokyo Institute of Technology, Japan this evaluation significantly. This Organizer: Mituhiro Fukuda minisymposium presents recent advances Tokyo Institute of Technology, Japan in generalized derivative evaluation, including extensions of established 9:30-9:55 Dual Approach Based on Spectral Projected Gradient Method methods to nonsmooth dynamic systems, for Log-Det Sdp with L1 Norm and improved methods for computing Makoto Yamashita and Mituhiro Fukuda, descent directions. Tokyo Institute of Technology, Japan; Organizer: Paul I. Barton Takashi Nakagaki, Yahoo Japan Massachusetts Institute of Technology, USA Corporation, Japan 9:30-9:55 Plan-C, A C-Package for 10:00-10:25 Improved Implementation Piecewise Linear Models in Abs- of Positive Matrix Completion Normal Form Interior-Point Method for Semidefinite Andreas Griewank, Humboldt University Programs Berlin, Germany Kazuhide Nakata and Makoto Yamashita, Tokyo Institute of Technology, Japan 10:00-10:25 Evaluating Generalized Derivatives for Nonsmooth Dynamic 10:30-10:55 A Lagrangian-Dnn Systems Relaxation: a Fast Method for Kamil Khan and Paul I. Barton, Computing Tight Lower Bounds for Massachusetts Institute of Technology, a Class of Quadratic Optimization USA Problems Sunyoung Kim, Ewha W. University, Korea; 10:30-10:55 Evaluating Generalized Masakazu Kojima, Chuo University, Japan; Derivatives of Dynamic Systems with a Kim-Chuan Toh, National University of Linear Program Embedded Singapore, Republic of Singapore Kai Hoeffner, Massachusetts Institute of Technology, USA 11:00-11:25 Mixed Integer Second- Order Cone Programming Formulations 11:00-11:25 Adjoint Mode for Variable Selection Computation of Subgradients for Ryuhei Miyashiro, Tokyo University of McCormick Relaxations Agriculture and Technology, Japan; Yuichi Markus Beckers, German Research School Takano, Tokyo Institute of Technology, for Simulation Sciences, Germany; Viktor Japan Mosenkis, RWTH Aachen University, Germany; Uwe Naumann, RWTH Aachen, Germany 2014 SIAM Conference on Optimization 17

Monday, May 19 Monday, May 19 Monday, May 19 IP2 MS2 MS13 Optimizing and Coordinating (Quasi)Variational Advances in Stochastic Healthcare Networks and Inequalities and Hierarchical Dynamic Programming Markets Equilibrium Problems 2:00 PM-4:00 PM 1:00 PM-1:45 PM 2:00 PM-4:00 PM Room:Pacific Salon 1 Room:Golden Ballroom Room:Golden Ballroom The goal of this session will be to Chair: Ariela Sofer, George Mason Variational inequalities provide a very present some recent developments in University, USA powerful framework for modeling approximate dynamic programming, with a particular emphasis on duality Healthcare systems pose a range of processes that involve switching, results and methods for calculating major access, quality and cost challenges nonsmooth behavior, and related performance bounds. Applications in in the U.S. and globally. We survey phenomena. Whenever they occur as OR will also be discussed. several healthcare network optimization constraints in an optimization problem, problems that are typically challenging the derivation of stationarity conditions Organizer: David Brown in practice and theory. The challenge and the associated numerical treatment Duke University, USA comes from network affects, including are significantly challenged by constraint 2:00-2:25 Information Relaxations, the fact that in many cases the network degeneracy. This minisymposium is Duality, and Convex Stochastic consists of selfish and competing players. devoted to recent advances in the field Dynamic Programs Incorporating the complex dynamics of variational inequalities, ranging David Brown and James Smith, Duke University, USA into the optimization is rather essential, from quasi-variational inequalities to but hard to models and often leads to hierarchical equilibrium problems and 2:30-2:55 An Approximate Dynamic computationally intractable models. multi-objective optimization problems Programming Approach to Stochastic We focus attention to computationally with equilibrium constraints. Matching Nikhil Bhat and Ciamac C. Moallemi, tractable and practical policies, and Organizer: Michael Hintermueller Columbia University, USA analyze their worst-case performance Humboldt University Berlin, Germany compared to optimal policies that 3:00-3:25 High Dimensional Revenue Organizer: Michael Ulbrich Management are computationally intractable and Technische Universität München, Germany conceptually impractical. Vivek Farias and Dragos Ciocan, 2:00-2:25 Multiple Optimization Massachusetts Institute of Technology, Retsef Levi Problems with Equilibrium Constraints USA Massachusetts Institute of Technology, USA (MOPEC) 3:30-3:55 Dynamic Robust Michael C. Ferris, University of Wisconsin, Optimization and Applications USA Dan A. Iancu, Stanford University, USA; Intermission 2:30-2:55 On the Optimal Control of a Mayank Sharma and Maxim Sviridenko, Class of Variational Inequalities of the IBM T.J. Watson Research Center, USA 1:45 PM-2:00 PM Second Kind Thomas M. Surowiec, Humboldt University at Berlin, Germany 3:00-3:25 Semismooth Newton Method for Differential Variational Inequalities Sonja Steffensen, RWTH Aachen University, Germany 3:30-3:55 Methods of Computing Individual Confidence Intervals for Solutions to Stochastic Variational Inequalities Michael Lamm, University of North Carolina, USA; Amarjit Budhiraja, University of North Carolina at Chapel Hill, USA; Shu Lu, University of North Carolina at Chapel Hill, USA 18 2014 SIAM Conference on Optimization

Monday, May 19 Monday, May 19 Monday, May 19 MS15 MS16 MS17 Stochastic Optimal Control Convegrence Rate Analysis Optimization Models and Applications of Inexact Second-order of Healthcare Delivery 2:00 PM-4:00 PM Methods in Nonlinear Networks Room:Pacific Salon 3 Optimization 2:00 PM-4:00 PM This minisymposium is focussed on 2:00 PM-4:00 PM Room:Pacific Salon 5 continuous time stochastic optimal Room:Pacific Salon 4 In this session we will discuss various control theory and its applications to While first-order methods have seen optimization models that arise in the Mathematical Finance. The main topics great renewal of interest recently, use context of healthcare delivery networks are: of second order information remains Organizer: Retsef Levi 1. Optimal control problems with an important goal for construction Massachusetts Institute of Technology, USA singular terminal values. of efficient optimization methods. 2:00-2:25 The Effectiveness of Uniform 2. Sensitivity analysis of stochastic However, exact second order models of Subsidies in Increasing Market optimal control problems. the objective function are often either Consumption to expensive to compute or too hard Retsef Levi, Georgia Perakis, and Gonzalo 3. Optimal stopping problems. to optimize accurately. Recently there Romero Yanez, Massachusetts Institute of 4. Investment models and Applications. has been many interesting methods Technology, USA Organizer: Francisco J. Silva developed that in one way or another 2:30-2:55 Efficient Resource Université de Limoges, France relax the second-order approximation, Allocation in Hospital Networks 2:00-2:25 Probabilistic and Smooth either by constructing inexact second- Marcus Braun, Fernanda Bravo, Vivek Solutions to Stochastic Optimal order models, or by inexactly minimizing Farias, and Retsef Levi, Massachusetts Control Problems with Singular exact models, producing approximate Institute of Technology, USA Terminal Values with Application to Newton steps. This minisymposium will 3:00-3:25 Discounted Referral Pricing Portfolio Liquidation Problems contain talks covering convergence and Between Healthcare Networks Ulrich Horst, Humboldt University Berlin, convergence rates analysis for different Fernanda Bravo, Angela King, Retsef Germany type of inexact second-order methods, Levi, Georgia Perakis, and Gonzalo 2:30-2:55 On Some Sensitivity Results for large scale convex optimization, Romero Yanez, Massachusetts Institute of Technology, USA in Stochastic Optimal Control Theory: nonconvex structured optimization and A Lagrange Multiplier Point of View stochastic optimization. 3:30-3:55 On Efficiency of Revenue- Julio Backhoff, Humboldt University Berlin, sharing Contracts in Joint Ventures in Germany Organizer: Katya Scheinberg Operations Management Lehigh University, USA 3:00-3:25 Some Irreversible Cong Shi, University of Michigan, USA; Investment Models with Finite Fuel 2:00-2:25 Convergence Rates of Line- Retsef Levi and Georgia Perakis, and Random Initial Time for Real Search and Trust Region Methods Massachusetts Institute of Technology, Options Analysis on Local Electricity Based on Probabilistic Models USA; Wei Sun, IBM T.J. Watson Markets Katya Scheinberg, Lehigh University, USA Research Center, USA Tiziano De Angelis, University of 2:30-2:55 Alternating Linearization Manchester, United Kingdom; Giorgio Methods for Quadratic Least-Square Ferrari, Bielefeld University, Germany; Problem John Moriarty, University of Manchester, Xi Bai and Katya Scheinberg, Lehigh United Kingdom University, USA 3:30-3:55 A Stochastic Reversible 3:00-3:25 Sampling Within Algorithmic Investment Problem on a Finite-Time Recursions Horizon: Free-Boundary Analysis Raghu Pasupathy, Virginia Tech, USA Giorgio Ferrari, Bielefeld University, 3:30-3:55 On the Use of Second Order Germany; Tiziano De Angelis, University Methods in Stochastic Optimization of Manchester, United Kingdom Stefan Solntsev, Northwestern University, USA 2014 SIAM Conference on Optimization 19

Monday, May 19 Monday, May 19 Monday, May 19 MS18 MS19 MS20 Decomposition Techniques in First Order Methods and Advances and Applications Conic Programming Applications - Part II of II in PDE Constrained 2:00 PM-4:00 PM 2:00 PM-4:00 PM Optimization Room:Pacific Salon 6 Room:Sunrise 2:00 PM-4:00 PM Computational methods for conic, in For Part 1 see MS7 Room:Sunset particular semidefinite optimization are The renewed interest for gradient Optimization problems with partial largely based on interior-point idea, methods dates back to the seminal work differential equations arise in a multitude and as such are limited to medium size of Barzilai and Borwein in 1988, and of applications. Their infinite-dimensional problems. The last decade has seen many since then, several new gradient methods origin leads to very large scale problems attepmts to break this size limitation. have been proposed. The research after discretization and poses significant Some of the most promising techniques about first-order methods has been also challenges: preserving and exploiting to solve (very) large scale conic problems stimulated by their successful use in structure, mastering high complexity, are based on decomposition of the practice, for instance in the application to achieving mesh-independence, matrix-free original problem into several small(er) certain ill-posed inverse problems, where implementation, error control etc. The scale problems. In this minisymposium they exhibit a significant regularizing minsymposium presents several recent we colect talks presenting various aspect effect. The goal of this minisymposium developments in this dynamic field. of this technique. is to present both theory and applications Organizer: Michael Ulbrich of state-of-the-art gradient methods, Organizer: Michal Kocvara Technische Universität München, Germany University of Birmingham, United Kingdom to better understand their features and potentialities. Organizer: Stefan Ulbrich Organizer: Michael Stingl Technische Universität Darmstadt, Germany Universität Erlangen-Nürnberg, Germany Organizer: Gerardo Toraldo University of Naples “Frederico II”, Naples, 2:00-2:25 Adaptive Reduced Basis 2:00-2:25 Reduced-Complexity Italy Stochastic Collocation for PDE Semidefinite Relaxations Via Chordal Optimization Under Uncertainty Decomposition Organizer: Luca Zanni Matthias Heinkenschloss, , Martin S. Andersen, Technical University University of Modena and Reggio Emilia, USA of Denmark, Denmark; Anders Hansson, Italy 2:30-2:55 A-optimal Sensor Placement Linköping University, Sweden; Lieven 2:00-2:25 First Order Methods for for Large-scale Nonlinear Bayesian Vandenberghe, University of California, Los Image Deconvolution in Microscopy Inverse Problems Angeles, USA Giuseppe Vicidomini, Istituto Italiano di Alen Alexanderian, Noemi Petra, Georg 2:30-2:55 Decomposition and Partial Tecnologia, Italy; Riccardo Zanella and Stadler, and Omar Ghattas, University of Separability in Sparse Semidefinite Gaetano Zanghirati, University of Ferrara, Texas at Austin, USA Optimization Italy; Luca Zanni, University of Modena 3:00-3:25 Risk-Averse PDE-Constrained Yifan Sun and Lieven Vandenberghe, and Reggio Emilia, Italy Optimization using the Conditional University of California, Los Angeles, USA 2:30-2:55 On the Regularizing Effect Value-At-Risk 3:00-3:25 Domain Decomposition of a New Gradient Method with Drew P. Kouri, Sandia National Laboratories, in Semidefinite Programming with Applications to Image Processing USA Application to Topology Optimization Roberta D. De Asmundis, Sapienza – 3:30-3:55 Generalized Nash Equilibrium Michal Kocvara, University of Birmingham, Università di Roma, Italy; Daniela di Problems in Banach Spaces United Kingdom Serafino, Second University of Naples, Italy; Germana Landi, University of Michael Hintermueller, Humboldt University 3:30-3:55 Preconditioned Newton- Bologna, Italy Berlin, Germany Krylov Methods for Topology Optimization 3:00-3:25 Gradient Descent James Turner, Michal Kocvara, and Daniel Approaches to Image Registration Loghin, University of Birmingham, United Roger Eastman, Loyola College, Maryland, Kingdom USA; Arlene Cole-Rhodes, Morgan State University, USA 3:30-3:55 Adaptive Hard Thresholding Algorithm For Constrained l0 Problem Fengmin Xu, Xi’an Jiaotong University, P.R. China 20 2014 SIAM Conference on Optimization

Monday, May 19 Monday, May 19 Monday, May 19 MS21 MS22 MS23 Advances in Deterministic Optimization of Natural Gas Global Efficiency of Global Optimization Networks Algorithms for Nonconvex 2:00 PM-4:00 PM 2:00 PM-4:00 PM Optimization Room:Towne Room:Royal Palm 1 2:00 PM-4:00 PM This session will focus on recent Optimization method can play an Room:Royal Palm 2 advances in theory, algorithms, and important role in the planning and Establishing the global rate of convergence applications of deterministic global operation of natural gas networks. The or the global evaluation complexity of optimization. Promising approaches will minisymposium describes approaches algorithms for nonconvex optimization be presented that include reduced space to solve the difficult mixed-integer problems is a natural but challenging global optimization methods, feasibility nonlinear programming problems that aspect of algorithm analysis that has based bounds tightening, and new class arise in the planning and operation of been little explored in the nonconvex of convex underestimators for NLP and such networks. We show how these setting until recently (while being very MINLP problems. problems can be approached from a well researched in the convex case). Organizer: Christodoulos A. discrete and a nonlinear optimization This minisymposium brings together Floudas perspective. A special focus lies on some of the latest important results Princeton University, USA problems that arise through legal on this topic, namely, recent work on requirements (e.g. non-discrimination 2:00-2:25 Globie: An Algorithm for the cubic regularization methods for smooth Deterministic Global Optimization of policies) and the incorporation of problems, complexity of stochastic Box-Constrained Nlps unconventional gas sources. gradient methods and other methods for Nikolaos Kazazakis and Claire Adjiman, Organizer: Lars Schewe nonsmooth nonconvex problems and that Imperial College London, United Kingdom Friedrich-Alexander-Universität Erlangen- of derivative-free optimization. 2:30-2:55 Prospects for Reduced- Nürnberg, Germany Organizer: Coralia Cartis Space Global Optimization 2:00-2:25 Mixed-Integer Nonlinear University of Oxford, United Kingdom Paul I. Barton, Harry Watson, and Achim Programming Problems in the 2:00-2:25 On the Use of Iterative Wechsung, Massachusetts Institute of Optimization of Gas Networks Methods in Adaptive Cubic Technology, USA -- Mixed- Regularization Algorithms for 3:00-3:25 On Feasibility-based Bounds Approaches Unconstrained Optimization Tightening Lars Schewe, Friedrich-Alexander- Tommaso Bianconcini, Universita degli Studi Leo Liberti, IBM T.J. Watson Research Universität Erlangen-Nürnberg, Germany di Firenze, Italy; Giampaolo Liuzzi, CNR, Center, USA; Pietro Belotti, FICO, United 2:30-2:55 Nonlinear Programming Italy; Benedetta Morini, Universita’ di Kingdom; Sonia Cafieri, ENAC, Toulouse, Techniques for Gas Transport Firenze, Italy; Marco Sciandrone, Università France; Jon Lee, University of Michigan, Optimization Problems degli Studi di Firenze, Italy USA Martin Schmidt, Leibniz University 2:30-2:55 Accelerated Gradient Hannover, Germany 3:30-3:55 New Classes of Convex Methods for Nonconvex Nonlinear and Underestimators for Global 3:00-3:25 Nonlinear Optimization in Stochastic Programming Optimization Gas Networks for Storage of Electric Saeed Ghadimi and Guanghui Lan, University Christodoulos A. Floudas, Princeton Energy of Florida, USA University, USA Jan Thiedau and Marc C. Steinbach, Leibniz 3:00-3:25 Subspace Decomposition University Hannover, Germany 4:30-4:45 Globie: An Algorithm for the Methods Deterministic Global Optimization of 3:30-3:55 Hard Pareto Optimization Serge Gratton, ENSEEIHT, Toulouse, France Box-Constrained Nlps Problems for Determinig Gas Network 3:30-3:55 Global Convergence and Nikolaos Kazazakis and Claire Adjiman, Capacities Complexity of a Derivative-Free Imperial College London, United Kingdom Marc C. Steinbach, Martin Schmidt, and Method for Composite Nonsmooth Bernhard Willert, Leibniz University Optimization Hannover, Germany Geovani Grapiglia and Jinyun Yuan, Federal University of Paraná, Brazil; Ya-Xiang Yuan, Chinese Academy of Sciences, P.R. of China

Coffee Break 4:00 PM-4:30 PM Room:Town & Country 2014 SIAM Conference on Optimization 21

Monday, May 19 Monday, May 19 Monday, May 19 CP1 CP2 CP3 Complementarity and Interior-Point Methods and Algorithms for Conic Equilibria Related Topics Optimization 4:30 PM-6:30 PM 4:30 PM-6:10 PM 4:30 PM-6:30 PM Room:Golden Ballroom Room:Pacific Salon 1 Room:Pacific Salon 2 Chair: Joaquim Judice, Instituto de Chair: Alexander Engau, University of Chair: Sergio Assuncao Monteiro, Federal Telecomunicações, Portugal Colorado, Denver, USA University of Rio de Janerio, Brazil 4:30-4:45 Computing a Cournot 4:30-4:45 An Efficient Algorithm 4:30-4:45 Gradient Methods for Least- Equilibrium in Integers for the Projection of a Point on the Squares Over Cones Michael J. Todd, Cornell University, USA Intersection of M Hyperplanes with Yu Xia, Lakehead University, Canada Nonnegative Coefficients and The 4:50-5:05 Complementarity 4:50-5:05 The Simplex Method and 0-1 Nonnegative Orthant Polytopes Formulations of l0-Norm Optimization Claudio Santiago, Lawrence Livermore Tomonari Kitahara and Shinji Mizuno, Tokyo Problems National Laboratory, USA; Nelson Institute of Technology, Japan John E. Mitchell, Rensselaer Polytechnic Maculan, Universidade Federal de Rio de Institute, USA; Mingbin Feng, Janeiro, Brazil; Helder Inacio, Georgia 5:10-5:25 Augmented Lagrangian- Northwestern University, USA; Jong-Shi Institute of Technology, USA; Sergio Type Solution Methods for Second Pang, University of Southern California, Assuncao Monteiro, Federal University of Order Conic Optimization USA; Xin Shen, Rensselaer Polytechnic Rio de Janerio, Brazil Alain B. Zemkoho and Michal Kocvara, Institute, USA; Andreas Waechter, University of Birmingham, United 4:50-5:05 A Quadratic Cone Northwestern University, USA Kingdom Relaxation-Based Algorithm for Linear 5:10-5:25 Superlinearly Convergent Programming 5:30-5:45 On a Polynomial Choice of Smoothing Continuation Algorithms for Mutiara Sondjaja, Cornell University, USA Parameters for Interior Point Methods Nonlinear Complementarity Problems Luiz-Rafael Santos and Fernando Villas- 5:10-5:25 FAIPA\SDP: A Feasible Arc over Definable Convex Cones Bôas, IMECC-UNICAMP, Brazil; Aurelio Interior Point Algorithm for Nonlinear Chek Beng Chua, Nanyang Technological Oliveira, Universidade Estadual de Semidefinite Programming University, Singapore Campinas, Brazil; Clovis Perin, IMECC- Jose Herskovits, COPPE/Universidade UNICAMP, Brazil 5:30-5:45 Global Convergence of Federal do Rio e Janeiro, Brazil; Smoothing and Scaling Conjugate Jean-Rodolphe Roche, University 5:50-6:05 A Low-Rank Algorithm Gradient Methods for Systems of of Lorraine, France; Elmer Bazan, to Solve SDP’s with Block Diagonal Nonsmooth Equations COPPE/Universidade Federal do Rio e Constraints via Riemannian Yasushi Narushima, Yokohama National Janeiro, Brazil; Jose Miguel Aroztegui, Optimization University, Japan; Hiroshi Yabe and Universidade Federal de Paraiba, Brazil Nicolas Boumal and P.-A. Absil, Université Ryousuke Ootani, Tokyo University of Catholique de Louvain, Belgium Science, Japan 5:30-5:45 Newton’s Method for Solving Generalized Equations Using Set- 6:10-6:25 An Analytic Center Algorithm 5:50-6:05 Competitive Equilibrium Valued Approximations for Semidefinite Programming Relaxations in General Auctions Samir Adly, Université de Limoges, France; Sergio Assuncao Monteiro, Federal University Johannes C. Müller, University of Erlangen- Radek Cibulka, University of West of Rio de Janerio, Brazil; Claudio Prata Nürnberg, Germany Bohemia, Pilsen, Czech Republic; Huynh Santiago, Lawrence Livermore National 6:10-6:25 On the Quadratic Van Ngai, Quy Nhon University, Vietnam Laboratory, USA; Paulo Roberto Oliveira, COPPE/Universidade Federal do Rio Eigenvalue Complementarity Problem 5:50-6:05 Polynomiality of Inequality- e Janeiro, Brazil; Nelson Maculan, Joaquim J. Júdice, Instituto de Selecting Interior-Point Methods for Universidade Federal de Rio de Janeiro, Telecomunicações, Portugal; Carmo Brás, Linear Programming Brazil Universidade Nova de Lisboa, Portugal; Alexander Engau, University of Colorado, Alfredo N. Iusem, IMPA, Rio de Janeiro, Denver, USA; Miguel F. Anjos, Brazil Mathematics and Industrial Engineering & GERAD, Ecole Polytechnique de Montreal, Canada 22 2014 SIAM Conference on Optimization

Monday, May 19 Monday, May 19 Monday, May 19 CP4 CP5 CP6 Integer and Combinatorial Stochastic Optimization I Stochastic Optimization II Optimization 4:30 PM-6:30 PM 4:30 PM-6:30 PM 4:30 PM-6:30 PM Room:Pacific Salon 4 Room:Pacific Salon 5 Room:Pacific Salon 3 Chair: Fu Lin, Argonne National Laboratory, Chair: Dzung Phan, IBM T.J. Watson USA Chair: Justo Puerto, Universidad de Sevilla, Research Center, USA Spain 4:30-4:45 An Integer L-Shaped 4:30-4:45 Two-Stage Portfolio Algorithm for the Single-Vehicle 4:30-4:45 Binary Steiner Trees and Their Optimization with Higher-Order Routing Problem with Simultaneous Application in Biological Sequence Conditional Measures of Risk Delivery and Stochastic Pickup Analysis Sitki Gulten and Andrzej Ruszczynski, Rutgers Nadine Wollenberg, University of Duisburg- Susanne Pape, Frauke Liers, and Alexander University, USA Essen, Germany; Michel Gendreau, Martin, Universität Erlangen-Nürnberg, 4:50-5:05 Risk Averse Computational Université de Montréal, Canada; Walter Germany Stochastic Programming Rei, Universite du Quebec a Montreal, Somayeh Moazeni and Warren Powell, 4:50-5:05 A New Exact Appoach for Canada; Rüdiger Schultz, University of Princeton University, USA a Generalized Quadratic Assignment Duisburg-Essen, Germany Problem 5:10-5:25 Optimal Low-Rank Inverse 4:50-5:05 Parallelized Branch-and- Lena Hupp, Friedrich-Alexander-Universität Matrices for Inverse Problems Fix Coordination Algorithm (P-Bfc) Erlangen-Nürnberg, Germany; Laura Matthias Chung and Julianne Chung, Virginia for Solving Large-Scale Multistage Klein, Technische Universität Dortmund, Tech, USA; Dianne P. O’Leary, University Stochastic Mixed 0-1 Problems Germany; Frauke Liers, Universität of Maryland, College Park, USA Laureano F. Escudero, Universidad Rey Erlangen-Nürnberg, Germany Juan Carlos, Spain; Unai Aldasoro, Maria 5:30-5:45 Support Vector Machines Via 5:10-5:25 Nonmonotone Grasp Merino, and Gloria Perez, Universidad del Stochastic Optimization Paola Festa, University of Napoli, Italy; País Vasco, Spain Konstantin Patev, Rutgers University, USA; Marianna De Santis, University of Rome La Darinka Dentcheva, Stevens Institute of 5:10-5:25 On Solving Multistage Mixed Sapienza, Italy; Giampaolo Liuzzi, CNR, Technology, USA 0-1 Optimization Problems with Some Italy; Stefano Lucidi, University of Rome Risk Averse Strategies 5:50-6:05 Semi-Stochastic Gradient La Sapienza, Italy; Francesco Rinaldi, Araceli Garin, University of the Basque Descent Methods University of Rome, Italy Country, Spain; Laureano F. Escudero, Jakub Konecny and Peter Richtarik, University 5:30-5:45 The Graph Partitioning into Universidad Rey Juan Carlos, Spain; Maria of Edinburgh, United Kingdom Constrained Spanning Sub-Trees Merino and Gloria Perez, Universidad del 6:10-6:25 Decomposition Algorithms Sangho Shim, Sunil Chopra, and Diego País Vasco, Spain for the Optimal Power Flow Problem Klabjan, Northwestern University, USA; 5:30-5:45 Linear Programming Twin under Uncertainty Dabeen Lee, POSTECH, Korea Support Vector Regression Using Dzung Phan, IBM T.J. Watson Research 5:50-6:05 Effects of the Lll Reduction on Newton Method Center, USA Integer Least Squares Estimation Mohammad Tanveer, The LNM Institute of Jinming Wen, Xiao-Wen Chang, and Xiaohu Information Technology, Jaipur, India Xie, McGill University, Canada 5:50-6:05 Robust Decision Analysis in 6:10-6:25 The Discrete Ordered Median Discrete Scenario Spaces Problem Revisited Saman Eskandarzadeh, Sharif University of Justo Puerto, Universidad de Sevilla, Spain Technology, Iran 6:10-6:25 Distributed Generation for Energy-Efficient Buildings: A Mixed- Integer Multi-Period Optimization Approach Fu Lin, Sven Leyffer, and Todd Munson, Argonne National Laboratory, USA 2014 SIAM Conference on Optimization 23

Monday, May 19 Monday, May 19 Monday, May 19 CP7 CP8 CP9 Optimal Control I Optimal Control II Nonlinear Optimization I 4:30 PM-6:30 PM 4:30 PM-6:30 PM 4:30 PM-6:30 PM Room:Pacific Salon 6 Room:Pacific Salon 7 Room:Sunrise Chair: Hilke Stibbe, University of Marburg, Chair: Francisco J. Silva, Université de Chair: Panos Parpas, Imperial College Germany Limoges, France London, United Kingdom 4:30-4:45 Optimal Control of 4:30-4:45 Branch and Bound 4:30-4:45 On the Full Convergence Nonlinear Hyperbolic Conservation Approach to the Optimal Control of of Descent Methods for Convex Laws with Switching Non-Smooth Dynamic Systems Optimization Sebastian Pfaff and Stefan Ulbrich, Ingmar Vierhaus, Zuse Institute Berlin, Clovis C. Gonzaga, Universidade Federal de Technische Universität Darmstadt, Germany; Armin Fügenschuh, Helmut- Santa Catarina, Brazil Germany Schmidt-University, Germany 4:50-5:05 A Flexible Inexact 4:50-5:05 Adaptive Multilevel Sqp 4:50-5:05 A Semismooth Newton-Cg Restoration Method and Application Method for State Constrained Method for Full Waveform Seismic to Multiobjective Constrained Optimization with Navier-Stokes Tomography Optimization Equations Christian Boehm, Technical University of Luis Felipe Bueno and Gabriel Haeser, Stefanie Bott and Stefan Ulbrich, Technische Munich, Germany; Michael Ulbrich, Federal University of Sao Paulo, Brazil; Universität Darmstadt, Germany Technische Universität München, José Mario Martínez, State University of Germany Campinas-Unicamp, Brazil 5:10-5:25 Optimal Control of Index 1 {PDAE}s 5:10-5:25 A New Semi-Smooth 5:10-5:25 The Alpha-BB Cutting Hannes Meinlschmidt and Stefan Ulbrich, Newton Multigrid Method for Control- Plane Algorithm for Semi-Infinite Technische Universität Darmstadt, Constrained Semi-Linear Elliptic PDE Programming Problem with Multi- Germany Problems Dimensional Index Set Jun Liu, Southern Illinois University, Shunsuke Hayashi, Tohoku University, 5:30-5:45 Recent Advances in USA; Mingqing Xiao, Southern Illinois Japan; Soon-Yi Wu, National Cheng Kung Optimum Experimental Design for University, Carbondale, USA University, PDE Models Gregor Kriwet, Philipps-Universität 5:30-5:45 Stochastic Optimal Control 5:30-5:45 Global Convergence Marburg, Germany; Ekaterina Kostina, Problem for Delayed Switching Analysis of Several Riemannian Fachbereich Mathematik und Informatik, Systems Conjugate Gradient Methods Philipps-Universität Marburg, Germany Charkaz A. Aghayeva, Anadolu University, Hiroyuki Sato, Kyoto University, Japan Turkey 5:50-6:05 Optimal Decay Rate for the 5:50-6:05 Optimality Conditions Indirect Stabilization of Hyperbolic 5:50-6:05 Iterative Solvers for for Global Minima of Nonconvex Systems Time-Dependent Pde-Constrained Functions on Riemannian Manifolds Roberto Guglielmi, University of Bayreuth, Optimization Problems Seyedehsomayeh Hosseini, ETH Zürich, Germany John W. Pearson, University of Edinburgh, Switzerland United Kingdom 6:10-6:25 A Numerical Method for 6:10-6:25 A Multiresolution Proximal Design of Optimal Experiments for 6:10-6:25 Second Order Conditions Method for Large Scale Nonlinear Model Discrimination under Model for Strong Minima in the Optimal Optimisation and Data Uncertainties Control of Partial Differential Panos Parpas, Imperial College London, Hilke Stibbe, University of Marburg, Equations United Kingdom Germany; Ekaterina Kostina, Fachbereich Francisco J. Silva, Université de Limoges, Mathematik und Informatik, Philipps- France Universität Marburg, Germany 24 2014 SIAM Conference on Optimization

Monday, May 19 Monday, May 19 Monday, May 19 CP10 CP11 CP12 Global Optimization Applications in Networks Algorithms 4:30 PM-6:10 PM 4:30 PM-6:30 PM 4:30 PM-6:30 PM Room:Sunset Room:Towne Room:Royal Palm 1 Chair: Jeffrey M. Larson, KTH Royal Chair: Bissan Ghaddar, IBM Research, Chair: Dimitri Papadimitriou, Bell Institute of Technology, Sweden Ireland Laboratories, Lucent Technologies, USA 4:30-4:45 ParDYCORS for Global 4:30-4:45 Stochastic On-time Arrival 4:30-4:45 Nonlinear Material Optimization Problem for Public Transportation Architecture Using Topology Tipaluck Krityakierne and Christine A. Networks Optimization Shoemaker, Cornell University, USA Sebastien Blandin, IBM Research, Reza Lotfi, Johns Hopkins University, USA Singapore; Samitha Samaranayake, 4:50-5:05 Approaching Forex 4:50-5:05 Topology Optimization of University of California, Berkeley, USA Trading Through Global Optimization Geometrically Nonlinear Structures Techniques 4:50-5:05 Generation of Relevant Francisco M. Gomes, University of Umberto Dellepiane, Alberto De Santis, Elementary Flux Modes in Metabolic Campinas, Brazil; Thadeu Senne, Stefano Lucidi, and Stefania Renzi, Networks Universidade Federal do Triangulo University of Rome La Sapienza, Italy Hildur Æsa Oddsdóttir, Erika Hagrot, Mineiro, Brazil Véronique Chotteau, and Anders Forsgren, 5:10-5:25 Complexity and Optimality 5:10-5:25 A Parallel Optimization KTH Royal Institute of Technology, of Subclasses of Zero-Order Global Algorithm for Nonnegative Tensor Sweden Optimization Factorization Serge L. Shishkin and Alan Finn, United 5:10-5:25 Discrete Optimization Ilker S. Birbil, Sabanci University, Turkey; Technologies Research Center, USA Methods for Pipe Routing Figen Oztoprak, Istanbul Technical Jakob Schelbert and Lars Schewe, Friedrich- University, Turkey; Ali Taylan Cemgil, 5:30-5:45 An Interval Algorithm for Alexander-Universität Erlangen-Nürnberg, Bogazici University, Turkey Global Optimization under Linear and Germany Box Constraints 5:30-5:45 Hipad - A Hybrid Interior- Tiago M. Montanher, University of Sao 5:30-5:45 Optimal Unit Selection in Point Alternating Direction Algorithm Paulo, Brazil Batched Liquid Pipelines with DRA for Knowledge-Based Svm and Richard Carter, GL Group, USA Feature Selection 5:50-6:05 Derivative-Free Multi-Agent Ioannis Akrotirianakis, Siemens Optimization 5:50-6:05 Pump Scheduling in Water Corporation, USA; Zhiwei Qin, Columbia Jeffrey M. Larson, Euhanna Ghadimi, and Networks University, USA; Xiaocheng Tang, Lehigh Mikael Johansson, KTH Royal Institute of Joe Naoum-Sawaya and Bissan Ghaddar, University, USA; Amit Chakraborty, Technology, Sweden IBM Research, Ireland Siemens Corporation, USA 6:10-6:25 Polynomial Optimization for 5:50-6:05 A Barycenter Method for Water Distribution Networks Direct Optimization Bissan Ghaddar, IBM Research, Ireland; Felipe M. Pait and Diego Colón, Mathieu Claeys, University of Cambridge, Universidade de Sao Paulo, Brazil United Kingdom; Martin Mevissen, IBM Research, USA 6:10-6:25 Distributed Benders Decomposition Method Dimitri Papadimitriou, Bell Laboratories, Lucent Technologies, USA; Bernard Fortz, Université Libre de Bruxelles, Belgium 2014 SIAM Conference on Optimization 25

Monday, May 19 Monday, May 19 Damage Detection Using a Hybrid Optimization Algorithm CP13 PP1 Rafael H. Lopez and Leandro Miguel, Universidade Federal de Santa Catarina, Applications of Optimization Welcome Reception Brazil; André Torii, Universidade Federal 4:30 PM-6:30 PM and Poster Session da Paraiba, Brazil Social Welfare Comparisons with Room:Royal Palm 2 8:00 PM-10:00 PM Fourth-Order Utility Derivatives Chair: Matthew J. Zahr, University Room:Town & Country Christophe Muller, Aix-Marseille Université, of California, Berkeley and Stanford Design of a Continuous Bioreactor France University, USA Via Multiobjective Optimization Nonlinear Model Reduction in 4:30-4:45 Riemannian Optimization in Coupled with Cfd Modeling Optimization with Implicit Constraints Multidisciplinary Design Optimization Jonas L. Ansoni and Paulo Seleghim Jr., Marielba Rojas, Delft University of Craig K. Bakker and Geoffrey Parks, University of Sao Paulo, Brazil Technology, Netherlands University of Cambridge, United Kingdom Interior Point Algorithm As Applied to An Optimization Model for Truck 4:50-5:05 A Unified Approach for the Transmission Network Expansion Tyres’ Selection Solving Continuous Multifacility Planning Problem Zuzana Sabartova, Chalmers University of Ordered Median Location Problems Carlos Becerril, Ricardo Mota, and Technology, Sweden Victor Blanco, Universidad de Granada, Mohamed Badaoui, National Polytechnic Spain; Justo Puerto and Safae ElHaj Institute, Mexico Analysis of Gmres Convergence Via BenAli, Universidad de Sevilla, Spain Convex Optimization Meshless Approximation Minimizing Hassane Sadok, Université du Littoral Calais 5:10-5:25 Linear Regression Estimation Divergence-free Energy Cedex, France under Different Censoring Models of Abderrahman Bouhamidi, University Littoral Survival Data Cote d’Opale, Calais, France In-Flight Trajectory Redesign Using Sequential Convex Programming Ke Wu, California State University, Fresno, Robust Security-Constrained Unit USA Eric Trumbauer and Benjamin F. Villac, Commitment with Dynamic Ratings University of California, Irvine, USA 5:30-5:45 A Near-Optimal Dynamic Anna Danandeh and Bo Zeng, University Learning Algorithm for Online of South Florida, USA; Brent Caldwell, Riemannian Pursuit for Big Matrix Matching Problems with Concave Tampa Electric Company, USA Recovery Returns under Random Permutations Li Wang, University of California, San Preventive Maintenance Scheduling Diego, USA; Mingkui Tan and Ivor Model of Multi-Component Systems with Xiao Chen and Zizhuo Wang, University of Tsang, Nanyang Technological University, Interval Costs Singapore Minnesota, Twin Cities, USA Emil Gustavsson, Chalmers University of 5:50-6:05 A Scenario Based Technology, Sweden Stochastic Optimization of the Topology of a Combustion-based Optimization of Resilient Supply A Minimization Based Krylov Method Chain with Dynamic Fortification Cost Power Plant for Maximum Energy for Ill-Posed Problems and Image Efficiency Amirhossein Meisami, Texas A&M Restoration University, USA; Nima Salehi, West Rebecca Zarin Pass and Chris Edwards, Khalide Jbilou, University of the Littoral Stanford University, USA Virginia University, USA Opal Coast, France 6:10-6:25 PDE-Constrained The Mesh Adaptive Direct Search Introducing Penlab, a Matlab Code Algorithm for Blackbox Optimization Optimization Using Hyper-Reduced for Nonlinear (and) Semidefinite Models with Linear Equalities Optimization Mathilde Peyrega, Charles Audet, Matthew J. Zahr, University of California, Michal Kocvara, University of Birmingham, Berkeley and Stanford University, USA; and Sebastien Le Digabel, École United Kingdom; Michael Stingl, Polytechnique de Montréal, Canada Charbel Farhat, Stanford University, USA Universität Erlangen-Nürnberg, Germany; Jan Fiala, Numerical Algorithms Group Ltd, United Kingdom Dinner Break Aircraft Structure Design Optimization Via a Matrix-Free Approach 6:30 PM-8:00 PM Andrew Lambe, University of Toronto, Attendees on their own Canada; Joaquim Martins, University of Michigan, USA; Sylvain Arreckx, and Dominique Orban, École Polytechnique de Montréal, Canada 26 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 IP3 IP4 Recent Progress on the Combinatorial Optimization Registration Diameter of Polyhedra and for National Security 7:45 AM-5:30 PM Simplicial Complexes Applications Room:Golden Foyer 8:15 AM-9:00 AM 9:00 AM-9:45 AM Room:Golden Ballroom Room:Golden Ballroom Chair: Mirjam Dür, University of Trier, Chair: Jon Lee, University of Michigan, Remarks Germany USA 8:10 AM-8:15 AM The Hirsch conjecture, posed in 1957, National-security optimization problems Room:Golden Ballroom stated that the graph of a d-dimensional emphasize safety, cost, and compliance, polytope or polyhedron with n facets frequently with some twist. They may cannot have diameter greater than n-d. merit parallelization, have to run on The conjecture itself has been disproved small, weak platforms, or have unusual (Klee-Walkup (1967) for unbounded constraints or objectives. We describe polyhedra, Santos (2010) for bounded several recent such applications. polytopes), but what we know about We summarize a massively-parallel the underlying question is quite scarce. branch-and-bound implementation for Most notably, no polynomial upper a core task in machine classification. A bound is known for the diameters challenging spam-classification problem that were conjectured to be linear. In scales perfectly to over 6000 processors. contrast, no polyhedron violating the We discuss math programming for Hirsch bound by more than 25% is benchmarking practical wireless-sensor- known. In this talk we review several management heuristics. We sketch recent attempts and progress on the theoretically justified algorithms for question. Some of these work in the a scheduling problem motivated by world of polyhedra or (more often) nuclear weapons inspections. Time bounded polytopes, but some try to shed permitting, we will present open light on the question by generalizing it problems. For example, can modern to simplicial complexes. In particular, nonlinear solvers benchmark a simple we show that the maximum diameter linearization heuristic for placing of arbitrary simplicial complexes is in imperfect sensors in a municipal water nΘ(d), we sketch the proof of Hirsch’s network? bound for ‘flag’ polyhedra (and more Cynthia Phillips general objects) by Adiprasito and Sandia National Laboratories, USA Benedetti, and we summarize the main ideas in the polymath 3 project, a web- based collective effort trying to prove an upper bound of type nd for the Coffee Break diameters of polyhedra. 9:45 AM-10:15 AM Francisco Santos Room:Town & Country University of Cantabria, Spain 2014 SIAM Conference on Optimization 27

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS24 MS25 MS26 Stochastic Optimization Derivative Free Optimization Optimization and of Complex Systems - 10:15 AM-12:15 PM Equilibrium Part I of II Room:Pacific Salon 1 10:15 AM-12:15 PM 10:15 AM-12:15 PM Due to the growing sophistication and Room:Pacific Salon 2 Room:Golden Ballroom efficiency of computer simulations This session will be dedicated to For Part 2 see MS36 (generally understood) as well as legacy new research into algorithms and This minisymposium seeks to showcase codes and other applications, there applications of equilibrium constraints. is an increasing demand to perform and motivate new mathematical Organizer: Todd Munson optimization of complex systems where developments in the field of stochastic Argonne National Laboratory, USA optimization with an eye on the unique (at least some) derivative information is not available. This minisymposium will 10:15-10:40 A SLPEC/EQP Method challenges posed by complex systems for Mathematical Programs with present some recent research in the area, such as buildings and electrical/ Variational Inequality Constraints gas/water/transportation networks: including comstrained, global, stochastic Todd Munson and Sven Leyffer, Argonne high dimensionality, spatio-temporal and application considerations. National Laboratory, USA phenomena, and behavior-driven Organizer: Andrew R. Conn 10:45-11:10 Polyhedral Variational uncertainty. IBM T.J. Watson Research Center, USA Inequalities Organizer: Victor Zavala 10:15-10:40 DIRECT-Type Algorithms Michael C. Ferris, University of Wisconsin, Argonne National Laboratory, USA for Derivative-Free Constrained USA; Youngdae Kim, University of Global Optimization Wisconsin, Madison, USA Organizer: Mihai Anitescu Gianni Di Pillo, Università di Roma “La Argonne National Laboratory, USA 11:15-11:40 Stochastic Variational Sapienza”, Italy; Giampaolo Liuzzi, Inequalities: Analysis and Algorithms 10:15-10:40 Modeling Energy CNR, Italy; Stefano Lucidi, University Aswin Kannan and Uday Shanbhag, Markets: An Illustration of the Four of Rome La Sapienza, Italy; Veronica Pennsylvania State University, USA Classes of Policies Piccialli, Università degli Studi di Roma Warren Powell and Hugo Simao, Princeton Tor Vergata, Italy; Francesco Rinaldi, 11:45-12:10 Vulnerability Analysis of University, USA University of Padova, Italy Power Systems: A Bilevel Optimization Approach 10:45-11:10 Stochastic Day-Ahead 10:45-11:10 Using Probabilistic Taedong Kim, University of Wisconsin, Clearing of Energy Markets Regression Models in Stochastic Madison, USA; Stephen Wright, Mihai Anitescu, Argonne National Derivative Free Optimization University of Wisconsin, USA Laboratory, USA Ruobing Chen and Katya Scheinberg, Lehigh University, USA 11:15-11:40 Cutting Planes for the Multi-Stage Stochastic Unit 11:15-11:40 Formulations for Commitment Problem Constrained Blackbox Optimization Yongpei Guan, University of Florida, USA Using Statistical Surrogates Sebastien Le Digabel, École Polytechnique 11:45-12:10 New Lower Bounding de Montréal, Canada; Bastien Talgorn, Results for Scenario-Based GERAD, Canada; Michael Kokkolaras, Decomposition Algorithms for McGill University, Canada Stochastic Mixed-Integer Programs Jean-Paul Watson, Sandia National 11:45-12:10 A World with Oil and Laboratories, USA; Sarah Ryan, Iowa Without Derivatives State University, USA; David Woodruff, David Echeverria Ciaurri, IBM Research, University of California, Davis, USA USA 28 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS27 MS28 MS29 Stochastic, Noisy, and Cuts and Computation for Recent Results in Conic Mixed-Integer Nonlinear Mixed Integer Nonlinear Programming Optimization Programming 10:15 AM-12:15 PM 10:15 AM-12:15 PM 10:15 AM-12:15 PM Room:Pacific Salon 5 Room:Pacific Salon 3 Room:Pacific Salon 4 Conic programming is a striving This minisymposium highlights While Mixed Integer Linear new area in nonlinear optimization, recent developments in the design Programming (MILP) problems are which comprises optimization over of algorithms for solving nonlinear extremely versatile, it is often useful matrix cones such as the semidefinite optimization problems. In particular, the to augment them with nonlinear cone, second order cone, as well as focus is on challenges that arise when constraints to better represent complex the copositive cone and its dual, the problems must be solved in real-time scientific and engineering problems. completely positive cone. This session and/or when they involve stochastic Unfortunately, the theory and will collect some recent results in this or noisy functions. In all such cases, technology for solving such Mixed area. optimization algorithms must take care Integer Nonlinear Programming Organizer: Mirjam Duer to balance the computational expense of (MINLP) problems is significantly University of Trier, Germany the “core” algorithm with the expense less developed than that for MILP. For 10:15-10:40 Combinatorial Proofs of of calculations imposed by the inherent this reason there has recently been Infeasibility, and of Positive Gaps in uncertainty in the problem statement significant interest on extending to Semidefinite Programming itself. These types of challenges MINLP one of the most crucial tools in Minghui Liu and Gabor Pataki, University of represent are some of the most effective black-box solvers for MILP: North Carolina at Chapel Hill, USA important that must be faced in the new cutting planes. This minisymposium 10:45-11:10 The Order of Solutions in frontiers of optimization research. considers MINLP cutting planes Conic Programming Bolor Jargalsaikhan, University of Organizer: Frank E. Curtis from their theoretical properties, their Groningen, The Netherlands; Mirjam Lehigh University, USA computational effectiveness and their integration into emerging black-box Duer, University of Trier, Germany; Organizer: Daniel Robinson solvers for MINLP. Georg Still, University of Twente, The Johns Hopkins University, USA Netherlands Organizer: Juan Pablo Vielma 10:15-10:40 An Optimization 11:15-11:40 A Copositive Approach Massachusetts Institute of Technology, USA Algorithm for Nonlinear Problems with to the Graph Isomorphism Problem Numerical Noise 10:15-10:40 Understanding Structure Luuk Gijben, University of Groningen, The Andreas Waechter, Alvaro Maggiar, in Conic Mixed Integer Programs: Netherlands; Mirjam Dür, University of and Irina Dolinskaya, Northwestern From Minimal Linear Inequalities to Trier, Germany University, USA Conic Disjunctive Cuts Fatma Kilinc-Karzan, Carnegie Mellon 11:45-12:10 Cutting Planes for 10:45-11:10 Stochastic Nonlinear University, USA Completely Positive Conic Problems Programming for Natural Gas Mirjam Duer, University of Trier, Germany Networks 10:45-11:10 Elementary Closures in Naiyuan Chiang and Victor Zavala, Argonne Nonlinear Integer Programming National Laboratory, USA Daniel Dadush, , USA 11:15-11:40 A Sequential Linear- 11:15-11:40 Valid Inequalities and Quadratic Programming Method for Computations for a Nonlinear Flow the Online Solution of Mixed-Integer Set Optimal Control Problems Jeff Linderoth, James Luedtke, and Hyemin Christian Kirches, University of Heidelberg, Jeon, University of Wisconsin, Madison, Germany USA 11:45-12:10 A Taxonomy of 11:45-12:10 MINOTAUR Framework: Constraints in Grey-Box Optimization New Developments and an Stefan Wild, Argonne National Laboratory, Application USA; Sebastien Le Digabel, École Ashutosh Mahajan, IIT Bombay, India Polytechnique de Montréal, Canada 2014 SIAM Conference on Optimization 29

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS30 MS31 MS32 Conic Programming, Coordinate Descent Geometry of Linear Symmetry, and Graphs Methods: Parallelism, Programming Part I 10:15 AM-12:15 PM Duality and Non- of III: Alternatives to Room:Pacific Salon 6 smoothness- Part I of III traditional Algorithms and Semidefinte programming deals with 10:15 AM-12:15 PM Generalizations optimization problems over the cone Room:Sunrise 10:15 AM-12:15 PM of symmetric positive semidefinite For Part 2 see MS43 Room:Sunset matrices. Copositive programming Coordinate descent methods are For Part 2 see MS44 considers linear programming over the becoming increasingly popular Linear programming, apart of its cone of copositive matrices, or its dual due to their ability to scale to big versatility to solve different types of cone of completely positive matrices. data problems. The purpose of this problems, is a fundamentally geometric This minisymposium addresses recent symposium is to bring together theory. The two standard algorithms to results on solving various optimization researchers working on developing solve it, the simplex method and the problems by using copositive or novel coordinate descent algorithms, interior point methods, take profit of semidefinite programming. A few talks analyzing their convergence/complexity, this, one from the perspective of discrete will also consider problems with special and applying the algorithms to new geometry, the other by approximating structure and/or graphs. domains. a smooth curve. This minisymposium Organizer: Renata Sotirov Organizer: Peter Richtarik is devoted to present recent progress Tilburg University, The Netherlands University of Edinburgh, United Kingdom in these and other algorithms, and to 10:15-10:40 SDP and Eigenvalue Organizer: Lin Xiao discuss other geometric ideas relevant Bounds for the Graph Partition Microsoft Research, USA for optimization. Problem Renata Sotirov and Edwin R. Van Dam, 10:15-10:40 Primal-dual Subgradient Organizer: Jesús A. De Loera Tilburg University, The Netherlands Method with Partial Coordinate University of California, Davis, USA Update 10:45-11:10 Keller’s Cube-Tiling Organizer: Francisco Santos Yurii Nesterov, Université Catholique de University of Cantabria, Spain Conjecture -- An Approach Through Louvain, Belgium Sdp Hierachies 10:15-10:40 Computing Symmetry Dion Gijswijt, Delft University of 10:45-11:10 Accelerated, Parallel and Groups of Polyhedral Cones Technology, Netherlands Proximal Coordinate Descent David Bremner, University of New Peter Richtarik and Olivier Fercoq, Brunswick, Canada; Mathieu Dutour 11:15-11:40 Completely Positive University of Edinburgh, United Kingdom Reformulations for Polynomial Sikiric, Rudjer Boskovic Institute, Croatia; Optimization 11:15-11:40 Stochastic Dual Dmitrii Pasechnik, NTU Singapore, Luis Zuluaga, Lehigh University, USA; Coordinate Ascent with ADMM Singapore; Thomas Rehn and Achill Javier Pena, Carnegie Mellon University, Taiji Suzuki, Tokyo Institute of Technology, Schuermann, University of Rostock, USA; Juan C. Vera, Tilburg University, Japan Germany The Netherlands 11:45-12:10 Recent Progress in 10:45-11:10 Alternating Projections 11:45-12:10 Title Not Available Stochastic Optimization for Machine As a Polynomial Algorithm for Linear David de Laat, Delft University of Learning Programming Technology, Netherlands Tong Zhang, Rutgers University, USA Sergei Chubanov, Universität Siegen, Germany 11:15-11:40 Optimizing over the Counting Functions of Parameterized Polytopes Nicolai Hähnle, Universität Bonn, Germany 11:45-12:10 Tropicalizing the Simplex Algorithm Xavier Allamigeon and Pascal Benchimol, CMAP, Ecole Polytechnique, France; Stephane Gaubert, INRIA and CMAP, Ecole Polytechnique, France; Michael Joswig, Technische Universität Berlin, Germany 30 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS33 MS34 MS35 On Convex Optimization Regularization and Iterative Large Scale Convex and Quadratic Methods in Optimization Optimization and Programming 10:15 AM-12:15 PM Applications 10:15 AM-12:15 PM Room:Royal Palm 1 10:15 AM-12:15 PM Room:Towne Large scale optimization frequently Room:Royal Palm 2 Convex optimization and quadratic relies on iterative techniques to solve This minisymposium will focus on optimization are two important research equations which determine the search recent advances in large-scale convex areas that are involved in various directions and employs regularization optimization and some applications, e.g. applications and in different disciplines. to improve the spectral properties of molecular imaging, phase recovery and This minisymposium presents recent matrices involved in these equations. signal recovery. This minisymposium will focus on such advances on both topics: The first two Organizer: Alexandre techniques. talks are concerned with solving specific d’Aspremont convex problems, the other two talks Organizer: Jacek Gondzio CNRS-ENS Paris, France University of Edinburgh, United Kingdom are devoted to quadratic optimization 10:15-10:40 Convex Relaxations for problems. Practical implementations are 10:15-10:40 Inexact Second- Permutation Problems presented and computational evidence order Directions in Large-scale Fajwel Fogel, ENS, France; Rodolphe is given confirming the success of the Optimization Jenatton, CRITEO, USA; Francis Bach, proposed methods. Jacek Gondzio, University of Edinburgh, Ecole Normale Superieure de Paris, United Kingdom France; Alexandre d’Aspremont, CNRS- Organizer: Angelika Wiegele ENS Paris, France Universitat Klagenfurt, Austria 10:45-11:10 Regularization Techniques in Nonlinear Optimization 10:45-11:10 Semidefinite Relaxation Organizer: Franz Rendl Paul Armand, University of Limoges, France for Orthogonal Procrustes and the Universitat Klagenfurt, Austria 11:15-11:40 Reduced Order Models Little Grothendieck Problem 10:15-10:40 A Parallel Bundle and Preconditioning Afonso S. Bandeira, Princeton University, Framework for Asynchronous Ekkehard W. Sachs, University of Trier, USA; Christopher Kennedy, University Subspace Optimization of Nonsmooth Germany and Virginia Tech, USA of Texas at Austin, USA; Amit Singer, Convex Functions Princeton University, USA Christoph Helmberg, TU Chemnitz, 11:45-12:10 Experiments with Quad 11:15-11:40 Sketching Large Scale Germany Precision for Iterative Solvers Michael A. Saunders and Ding Ma, Stanford Convex Quadratic Programs 10:45-11:10 A Feasible Active Set University, USA Mert Pilanci, Martin Wainwright, and Method for Box-Constrained Convex Laurent El Ghaoui, University of Problems California, Berkeley, USA Philipp Hungerlaender and Franz Rendl, 11:45-12:10 On Lower Complexity Universitat Klagenfurt, Austria Bounds for Large-Scale Smooth 11:15-11:40 BiqCrunch: A Convex Optimization Semidefinite Branch-and-bound Cristobal Guzman and , Method for Solving Binary Quadratic Georgia Institute of Technology, USA Problems Nathan Krislock, Northern Illinois University, USA; Frédéric Roupin, Laboratoire d’Informatique Paris-Nord, France; Jérôme Malick, INRIA, France 11:45-12:10 Local Reoptimization Via Column Generation and Quadratic Programming Lucas Letocart, Université Paris 13, France; Fabio Furini, Université Paris Dauphine, France; Roberto Wolfler Calvo, Université Paris 13, France 2014 SIAM Conference on Optimization 31

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 Lunch Break and Forward SP1 MS36 Looking Panel Discussion SIAG/OPT Prize Lecture: Stochastic Optimization 12:15 PM-1:30 PM Efficiency of the Simplex of Complex Systems - Attendees on their own and Policy Iteration Part II of II Methods for Markov 2:30 PM-4:30 PM Decision Processes Room:Golden Ballroom PD1 1:30 PM-2:15 PM For Part 1 see MS24 Forward Looking Panel Room:Golden Ballroom For Part 3 see MS48 This minisymposium seeks to showcase Discussion Chair: Mihai Anitescu, Argonne National and motivate new mathematical Laboratory, USA 12:30 PM-1:15 PM developments in the field of stochastic We prove that the classic policy-iteration Room:Golden Ballroom optimization with an eye on the unique method (Howard 1960), including the challenges posed by complex systems Chair: Juan Meza, University of simple policy-iteration (or simplex such as buildings and electrical/ California, Merced, USA method, Dantzig 1947) with the most- gas/water/transportation networks: negative-reduced-cost pivoting rule, is high dimensionality, spatio-temporal Panelists: a strongly polynomial-time algorithm Natalia Alexandrov phenomena, and behavior-driven for solving discounted Markov decision NASA Langley Research Center, USA uncertainty. processes (MDP) of any fixed discount Organizer: Victor Zavala Frank E. Curtis factor. The result is surprising since Lehigh University, USA Argonne National Laboratory, USA almost all complexity results on the Leo Liberti simplex and policy-iteration methods Organizer: Mihai Anitescu IBM T.J. Watson Research Center, USA are negative, while in practice they Argonne National Laboratory, USA Daniel Ralph are popularly used for solving MDPs, 2:30-2:55 Stochastic Model University of Cambridge, United Kingdom or linear programs at large. We also Predictive Control with Non-Gaussian Ya-Xiang Yuan present a result to show that the simplex Disturbances Tony Kelman and Francesco Borrelli, Chinese Academy of Sciences, P.R. of China method is strongly polynomial for University of California, Berkeley, USA solving deterministic MDPs regardless of discount factors. 3:00-3:25 Covariance Estimation and Relaxations for Stochastic Unit Yinyu Ye Commitment Stanford University, USA Cosmin G. Petra, Argonne National Laboratory, USA 3:30-3:55 Parallel Nonlinear Interior- Intermission Point Methods for Stochastic 2:15 PM-2:30 PM Programming Problems Yankai Cao, Purdue University, USA; Jia Kang, and Carl Laird, Texas A&M University, USA 4:00-4:25 A Two Stage Stochastic Integer Programming Method for Adaptive Staffing and Scheduling under Demand Uncertainty with Application to Nurse Management Kibaek Kim and Sanjay Mehrotra, Northwestern University, USA 32 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS37 MS38 MS39 Structural Optimization with Variational Analysis and Polynomial Optimization Composites Monotone Operator Theory: and Moment Problems - 2:30 PM-4:30 PM Modern Techniques and Part I of II Room:Pacific Salon 1 Trends 2:30 PM-4:30 PM The usage of composite material leads 2:30 PM-4:30 PM Room:Pacific Salon 3 to tools and instruments with novel Room:Pacific Salon 2 For Part 2 see MS51 properties. This field gives a lot of room Variational analysis provide key Polynomial optimization is a new for optimization. However, all relevant instruments for optimization, area in mathematical programming. optimization problems involve an theoretically and algorithmically. We It studies how to find global solutions underlying simulation model consisting will discuss recent results in variational to optimization problems defined of partial differential equations. Thus, analysis, monotone operators and by polynomials. This is equivalent challenging tasks arise involving PDE convex analysis. to a class of moment problems. The constrained optimization in combination basic techniques are semidefinite with shape optimization aspects as Organizer: Shawn Xianfu Wang programming, representation of University of British Columbia, Canada well as topology optimization aspects. nonnegative polynomials, and real This minisymposium discusses central 2:30-2:55 SI Spaces and Maximal algebraic geometry. It reaches into problem formulations as well as efficient Monotonicity theoretical and computational aspects of Stephen Simons, University of California, numerical solution strategies. optimization and moment theory. The Santa Barbara, USA Organizer: Volker H. Schulz session will present recent advances 3:00-3:25 New Techniques in Convex University of Trier, Germany made in this new area. Analysis Organizer: Heinz Zorn Jon Vanderwerff, La Sierra University, USA Organizer: Jean B. Lasserre University of Trier, Germany LAAS-CNRS, Toulouse, France 3:30-3:55 Coderivative 2:30-2:55 Optimization of the Fibre Characterizations of Maximal Organizer: Jiawang Nie Orientation of Orthotropic Material Monotonicity with Applications to University of California, San Diego, USA Heinz Zorn, University of Trier, Germany Variational Systems 2:30-2:55 On Level Sets of 3:00-3:25 A New Algorithm for the Boris Mordukhovich, Wayne State Polynomials and a Generalization of Optimal Design of Anisotropic University, USA; Nghia Tran, University the Lowner’s John Problem Materials of British Columbia, Canada Jean B. Lasserre, LAAS-CNRS, Toulouse, Michael Stingl and Jannis Greifenstein, 4:00-4:25 Resolvent Average of France Universität Erlangen-Nürnberg, Germany Monotone Operators 3:00-3:25 The CP-Matrix Completion 3:30-3:55 Multi-Phase Optimization Shawn Xianfu Wang, University of British Problem of Elastic Materials, Using a Level Set Columbia, Canada Anwa Zou and Jinyan Fan, Shanghai Method Jiaotong University, China Charles Dapogny, Laboratoire Jacques-Louis 3:30-3:55 Optimizing over Surface Lions, France Area Measures of Convex Bodies 4:00-4:25 Structural Optimization Didier Henrion, University of Toulouse, Based on the Topological Gradient France Volker H. Schulz and Roland Stoffel, 4:00-4:25 Convexity in Problems University of Trier, Germany Having Matrix Unknowns J.William Helton, University of California, San Diego, USA 2014 SIAM Conference on Optimization 33

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS40 MS41 MS42 New Approaches to Hard Convex Optimization Optimization Models for Discrete Optimization Algorithms Novel Treatment Planning Problems 2:30 PM-4:30 PM Approaches for Cancer 2:30 PM-4:30 PM Room:Pacific Salon 5 Therapies Room:Pacific Salon 4 This minisymposium highlights 2:30 PM-4:30 PM The minisymposium covers a small recent developments in the design Room:Pacific Salon 6 selection of recent results on hard of algorithms for solving convex Optimization methods for cancer optimization problems, both theoretical optimization problems that are typically (especially radiation) treatment planning and computational ones. large scale. In this case, large scale are the subject of active research. either means that the optimization Organizer: Endre Boros Commonly, optimization models are involves a large number of design Rutgers University, USA used for designing treatments that variables or that the objective function deliver high doses of radiation to 2:30-2:55 Hierarchical Cuts to measures the loss incurred by pieces Strengthen Semidefinite Relaxations cancerous tissues, while sparing healthy of data in a set that may be of extreme of NP-hard Graph Problems organs. In these models, prescription Elspeth Adams and Miguel Anjos, École scale. Typical problems of this type arise doses for tumors and dose limits for Polytechnique de Montréal, Canada; Franz in challenging applications related to healthy tissues are specified by widely Rendl and Angelika Wiegele, Universitat machine learning, compressive sensing, used protocols, and the planning Klagenfurt, Austria sparse optimization, etc. individualizes the treatment to the 3:00-3:25 Lower Bounds on the Organizer: Daniel Robinson geometry of each patient’s anatomy. In Graver Complexity of M-Fold Johns Hopkins University, USA this session, the speakers will discuss Matrices Organizer: Andreas Waechter novel treatment planning approaches Elisabeth Finhold, TU Munich, Germany; Northwestern University, USA that are further individualized, taking Raymond Hemmecke, Technische into account functional imaging and Universität München, Germany 2:30-2:55 A Stochastic Quasi-Newton Method biomarker information to design 3:30-3:55 BiqCrunch in Action: Solving , Northwestern University, biologically tuned and adaptable Difficult Combinatorial Optimization USA radiation and chemotherapy treatments. Problems Exactly Frédéric Roupin, Laboratoire d’Informatique 3:00-3:25 An Active Set Strategy for Organizer: Marina A. Epelman Paris-Nord, France; Nathan Krislock, l1 Regularized Problems University of Michigan, USA Northern Illinois University, USA; Jerome Marianna De Santis, Technische Universität 2:30-2:55 Optimal Fractionation in Malick, INRIA, France Dortmund, Germany; Stefano Lucidi, Radiotherapy 4:00-4:25 Effective Quadratization University of Rome La Sapienza, Italy; Minsun Kim, Fatemeh Saberian, and Archis of Nonlinear Binary Optimization Francesco Rinaldi, University of Padova, Ghate, University of Washington, USA Italy Problems 3:00-3:25 A Mathematical Endre Boros, Rutgers University, USA 3:30-3:55 An Iterative Reweighting Optimization Approach to Algorithm for Exact Penalty the Fractionation Problem in Subproblems Chemoradiotherapy Hao Wang, Lehigh University, USA; James Ehsan Salari, Wichita State University, V. Burke, University of Washington, USA; USA; Jan Unkelbach and Thomas Frank E. Curtis, Lehigh University, USA; Bortfeld, Massachusetts General Hospital Jiashan Wang, University of Washington, and Harvard Medical School, USA USA 3:30-3:55 Incorporating Organ 4:00-4:25 Second Order Methods for Functionality in Radiation Therapy Sparse Signal Reconstruction Treatment Planning Kimon Fountoulakis, Ioannis Dassios, and Victor Wu, Marina A. Epelman, Mary Feng, Jacek Gondzio, University of Edinburgh, Martha Matuszak, and Edwin Romeijn, United Kingdom University of Michigan, USA 4:00-4:25 A Stochastic Optimization Approach to Adaptive Lung Radiation Therapy Treatment Planning Troy Long, Marina A. Epelman, Martha Matuszak, and Edwin Romeijn, University of Michigan, USA 34 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS43 MS44 MS45 Coordinate Descent Geometry of Linear Bilevel Programs and Methods: Parallelism, Programming Part II of III: Related Topics Duality and Non- Through the interior 2:30 PM-4:30 PM smoothness- Part II of III 2:30 PM-4:30 PM Room:Towne 2:30 PM-4:30 PM Room:Sunset Bilevel programming problems are Room:Sunrise For Part 1 see MS32 hierarchical optimization problems where For Part 3 see MS56 the constraints of the upper level problem For Part 1 see MS31 For Part 3 see MS55 Linear programming, apart of its are defined in part by a second parametric Coordinate descent methods are versatility to solve different types of lower level optimization problem. becoming increasingly popular problems, is a fundamentally geometric Bilevel problems are closed related to the due to their ability to scale to big theory. The two standard algorithms to mathematical program with equilibrium data problems. The purpose of this solve it, the simplex method and the constraints (MPEC) and the equilibrium symposium is to bring together interior point methods, take profit of problem with equilibrium constraints researchers working on developing this, one from the perspective of discrete (EPEC). Although bilevel programs novel coordinate descent algorithms, geometry, the other by approximating and the related problems can be used analyzing their convergence / a smooth curve. This minisymposium to model a wide range of applications, complexity, and applying the algorithms is devoted to present recent progress they are highly challenging problems to to new domains. in these and other algorithms, and to solve both in theory and in practice due discuss other geometric ideas relevant to the intrinsical nonsmoothness and Organizer: Peter Richtarik for optimization. nonconvexity. Recently many progresses University of Edinburgh, United Kingdom Organizer: Jesús A. De Loera have been made. This minisymposia Organizer: Zhaosong Lu University of California, Davis, USA aims at presenting up to date theories and Simon Fraser University, Canada algorithms for solving bilevel programs, Organizer: Francisco Santos Organizer: Lin Xiao MPECs and EPECs. University of Cantabria, Spain Microsoft Research, USA 2:30-2:55 On the Sonnevend Organizer: Jane Ye 2:30-2:55 On the Complexity Analysis Curvature and the Klee-Walkup University of Victoria, Canada of Randomized Coordinate Descent Result Organizer: Stephan Dempe Methods Tamas Terlaky, Lehigh University, USA TU Bergakademie Freiberg, Germany Zhaosong Lu, Simon Fraser University, Canada; Lin Xiao, Microsoft Research, 3:00-3:25 On the Curvature of the 2:30-2:55 Solving Bilevel Programs by USA Central Path for Lp Smoothing Techniques Yuriy Zinchenko, University of Calgary, Jane J. Ye, University of Victoria, Canada 3:00-3:25 Universal Coordinate Canada Descent Method 3:00-3:25 An Algorithm for Solving Olivier Fercoq and Peter Richtarik, 3:30-3:55 An Efficient Affine- Smooth Bilevel Optimization Problems University of Edinburgh, United Kingdom scaling Algorithm for Hyperbolic Susanne Franke and Stephan Dempe, TU Programming Bergakademie Freiberg, Germany 3:30-3:55 Stochastic Block Mirror James Renegar, Cornell University, USA Descent Methods for Nonsmooth and 3:30-3:55 On Regular and Limiting Stochastic Optimization 4:00-4:25 Towards a Computable Coderivatives of the Normal Cone Cong D. Dang and Guanghui Lan, O(n)-Self-Concordant Barrier for Mapping to Inequality Systems n University of Florida, USA Polyhedral Sets in R Helmut Gfrerer, Johannes Kepler Universität, Kurt M. Anstreicher, University of Iowa, Linz, Austria; Jiri Outrata, Czech Academy 4:00-4:25 Iteration Complexity USA of Sciences, Czech Republic Analysis of Block Coordinate Descent Methods 4:00-4:25 A Heuristic Approach to Zhi-Quan Luo, University of Minnesota, Solve the Bilevel Toll Assignment Minneapolis, USA Problem by Making Use of Sensitivity Analysis Vyacheslav V. Kalashnikov, Tecnologico de Monterrey (ITESM), Mexico; Nataliya I. Kalashnykova, Universidad Autónoma de Nuevo León, Mexico; Aaron Arevalo Franco and Roberto C. Herrera Maldonado, Tecnologico de Monterrey (ITESM), Mexico 2014 SIAM Conference on Optimization 35

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS46 MS47 MS48 Numerical Linear Algebra Function-Value-Free Stochastic Optimization and Optimization - Optimization of Complex Systems - Part I of II 2:30 PM-4:30 PM Part III of III 2:30 PM-4:30 PM Room:Royal Palm 2 5:00 PM-7:00 PM Room:Royal Palm 1 Function-value-free (FVF) or Room:Golden Ballroom For Part 2 see MS58 comparison-based algorithms are For Part 2 see MS36 In optimization, discretization of PDEs, derivative-free optimization algorithms This minisymposium seeks to showcase least squares and other applications, that use the objective function through and motivate new mathematical linear systems sharing a common comparisons exclusively, i.e., updates developments in the field of stochastic structure arise. The efficient solution of the algorithm’s state variables use optimization with an eye on the unique of those systems naturally leads to comparisons between queried solutions challenges posed by complex systems efficient processes upstream. This only. Those algorithms are invariant to such as buildings and electrical/ minisymposium summarizes ongoing composing the objective function by gas/water/transportation networks: forays into the construction and a strictly increasing transformation. high dimensionality, spatio-temporal implementation of efficient numerical Well-known FVF are the Nelder- phenomena, and behavior-driven methods for such problems. Mead algorithm and Evolution uncertainty. Organizer: Dominique Orban Strategies (ES) like CMA-ES or Organizer: Victor Zavala École Polytechnique de Montréal, Canada xNES. This minisymposium features Argonne National Laboratory, USA recent advances on randomized FVF Organizer: Anders Forsgren algorithms for single and multi-objective Organizer: Mihai Anitescu KTH Royal Institute of Technology, Sweden Argonne National Laboratory, USA optimization. The focus is on theoretical Organizer: Annick Sartenaer concepts for algorithm design related to 5:00-5:25 A Scalable Clustering- Universite Notre Dame de la Paix and invariances, information geometry and Based Preconditioning Strategy for Stochastic Programms University of Namur, Belgium their application for constructing robust Victor Zavala, Argonne National Laboratory, 2:30-2:55 Using Partial Spectral and efficient algorithms. USA Information for Block Diagonal Preconditioning of Saddle-Point Organizer: Anne Auger 5:30-5:55 Semismooth Equation INRIA, France Systems Approaches to Structured Convex Daniel Ruiz, ENSEEIHT, Toulouse, France; 2:30-2:55 Fifty Years of Randomized Optimization Annick Sartenaer, Universite Notre Dame Function-Value-Free Algorithms: Arvind Raghunathan and Daniel Nikovski, de la Paix and University of Namur, Where Do We Stand? Mitsubishi Electric Research Laboratories, Belgium; Chalotte Tannier, University of Anne Auger, INRIA, France USA Namur, Belgium 3:00-3:25 X-NES: Exponential 6:00-6:25 Dynamic System Design 3:00-3:25 Spectral Bounds and Coordinate System Parameterization based on Stochastic Optimization Preconditioners for Unreduced for FVF Search Rui Huang and Baric Miroslav, United Symmetric Systems Arising from Tobias Glasmachers, Ruhr-Universitat Technologies Research Center, USA Bochum, Germany Interior Point Methods 6:30-6:55 Enhanced Benders Benedetta Morini, Universita’ di Firenze, 3:30-3:55 Function-Value-Free Decomposition for Stochastic Mixed- Italy; Valeria Simoncini and Mattia Tani, Continuous Optimization in High integer Linear Programs Universita’ di Bologna, Italy Dimension Emmanuel Ogbe and Xiang Li, Queen’s 3:30-3:55 A General Krylov Subspace Youhei Akimoto, Shinshu University, Japan University, Canada Method and its Connection to the 4:00-4:25 Evolutionary Multiobjective Conjugate-gradient Method Optimization and Function-Value- Tove Odland and Anders Forsgren, KTH Free Algorithms Royal Institute of Technology, Sweden Dimo Brockhoff, INRIA, France 4:00-4:25 Asqr: Interleaving Lsqr and Craig Krylov Methods for the Solution of Augmented Systems Daniel Ruiz, ENSEEIHT, Toulouse, France; Coffee Break Alfredo Buttari, CNRS, France; David 4:30 PM-5:00 PM Titley-Peloquin, CERFACS, France Room:Town & Country 36 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS49 MS50 MS51 Optimizing the Dynamics Distributed Stochastic- Polynomial Optimization of Radiation Therapy of Gradient Methods for and Moment Problems - Cancer Convex Optimization Part II of II 5:00 PM-7:00 PM 5:00 PM-7:00 PM 5:00 PM-7:00 PM Room:Pacific Salon 1 Room:Pacific Salon 2 Room:Pacific Salon 3 Previous optimization approaches Recent advances in wireless technology For Part 1 see MS39 in radiation therapy have focused have lead to an unprecedented growth Polynomial optimization is a new primarily on optimizing the spatial dose in demand for information exchange, area in mathematical programming. distribution, with the goal to bring as information storage and retrieval (e.g. It studies how to find global solutions much radiation dose to the tumor as internet applications including social to optimization problems defined possible without exceeding the tolerance networks and search engines). These by polynomials. This is equivalent limits in the surrounding normal tissues. phenomena translate to distributed to a class of moment problems. The In this minisymposium we will address optimization problems with huge basic techniques are semidefinite the question of optimal radiation dose data sets. The resulting optimization programming, representation of delivery over time. The presentations problems are characterized by nonnegative polynomials, and real will cover the dose fractionation distributed and uncertain information algebraic geometry. It reaches into problem, the question of combining necessitating computations to be theoretical and computational aspects of the schedules of different treatment done in an uncertain environment, optimization and moment theory. The modalities such as radiation and chemo with imperfect information, over a session will present recent advances therapies, certain issues with treating communication network, and most made in this new area. mobile targets, as well as the question importantly without a central entity that Organizer: Jean B. Lasserre of how to minimize the treatment time has an access to the whole information. LAAS-CNRS, Toulouse, France This session focuses on most recent through arc therapy techniques. Organizer: Jiawang Nie Organizer: Thomas Bortfeld optimization techniques dealing University of California, San Diego, USA with uncertainty, large data sets and Massachusetts General Hospital and 5:00-5:25 The Hierarchy of Local distributed computations. Harvard Medical School, USA Minimums in Polynomial Optimization 5:00-5:25 A Column Generation Organizer: Angelia Nedich Jiawang Nie, University of California, San and Routing Approach to 4π Vmat University of Illinois at Urbana-Champaign, Diego, USA USA Radiation Therapy Treatment Planning 5:30-5:55 Polynomial Optimization in Edwin Romeijn, National Science 5:00-5:25 On the Solution of Biology Foundation, USA; Troy Long, University Stochastic Variational Problems with Raphael A. Hauser, Oxford University, of Michigan, USA Imperfect Information United Kingdom Uday Shanbhag, Pennsylvania State 5:30-5:55 Adaptive and Robust 6:00-6:25 Constructive Approximation University, USA; Hao Jiang, University of Radiation Therapy in the Presence of Approach to Polynomial Optimization Illinois, USA Motion Drift Etienne De Klerk, Tilburg University, The Timothy Chan and Philip Mar, University of 5:30-5:55 Stochastic Methods for Netherlands; Monique Laurent, CWI, Toronto, Canada Convex Optimization with “Difficult” Amsterdam, Netherlands; Zhao Sun, 6:00-6:25 Mathematical Modeling Constraints Tilburg University, The Netherlands Mengdi Wang, Princeton University, USA and Optimal Fractionationated 6:30-6:55 Noncommutative Irradiation for Proneural 6:00-6:25 On Decentralized Gradient Polynomials Glioblastomas Descent Igor Klep, University of Auckland, New Kevin Leder, University of Minnesota, USA Kun Yuan and Qing Ling, University of Zealand 6:30-6:55 Combined Temporal and Science and Technology of China, China; Spatial Treatment Optimization Wotao Yin, University of California, Los Jan Unkelbach, Massachusetts General Angeles, USA Hospital and Harvard Medical School, 6:30-6:55 Distributed Optimization USA; Martijn Engelsman, TU Delft, over Time-Varying Directed Graphs Netherlands Angelia Nedich and Alexander Olshevsky, University of Illinois at Urbana- Champaign, USA 2014 SIAM Conference on Optimization 37

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS52 MS53 MS54 Optimization for Clustering Modern Unconstrained Integer Programming and and Classification Optimization Algorithms Applied Optimization 5:00 PM-7:00 PM 5:00 PM-7:00 PM 5:00 PM-7:00 PM Room:Pacific Salon 4 Room:Pacific Salon 5 Room:Pacific Salon 6 The problem of identifying similar This minisymposium highlights the Talks on: items in data is a fundamental problem recent developments in the design of 1) Integer programs with exactly k in machine learning and statistics. The algorithms for nonlinear unconstrained solutions tasks of partitioning data into clusters of optimization. This includes methods similar items and classifying data based that are derivative free, allow the use of 2) Solving a mixed-integer program on cluster membership play a significant only gradient information, or possibly with power flow equations role in modern data analysis, particularly the use of (partial) second-derivative 3) Cutting Planes for The Multi- in the areas of information retrieval, information. The methods discussed in Stage Stochastic Unit Commitment pattern recognition, computational this minisymposium advance the area Problem biology, and image processing. This either in terms of algorithm design or 4) Computational finance session will focus on the statistical and through worst-case complexity results of Organizer: Jon Lee computational challenges posed by these trust-region and related methods. University of Michigan, USA problems and the use of optimization- Organizer: Daniel Robinson based approaches to overcome these 5:00-5:25 Integer Programs with Johns Hopkins University, USA Exactly K Solutions challenges. Organizer: Andreas Waechter Jesús A. De Loera, University of California, Organizer: Brendan P. Ames Northwestern University, USA Davis, USA; Iskander Aliev, Cardiff California Institute of Technology, USA University, United Kingdom; Quentin 5:00-5:25 Alternative Quadratic Louveaux, Université de Liège, Belgium 5:00-5:25 Alternating Direction Models for Minimization Without Methods for Penalized Classification Derivatives 5:30-5:55 Pushing the Frontier Brendan Ames, California Institute of Michael Powell, University of Cambridge, (literally) with the Alpha Alignment Technology, USA United Kingdom Factor Anureet Saxena, Allianz Global Investors, 5:30-5:55 Statistical-Computational 5:30-5:55 A Nonmonotone USA Tradeoffs in Planted Models Approximate Sequence Algorithm for Yudong Chen, University of California, Unconstrained Optimization 6:00-6:25 Cutting Planes for a Multi- Berkeley, USA; Jiaming Xu, University of Hongchao Zhang, Louisiana State Stage Stochastic Unit Commitment Illinois at Urbana-Champaign, USA University, USA Problem Riuwei Jiang, University of Arizona, USA; 6:00-6:25 Regularized Spectral 6:00-6:25 The Spectral Cauchy- Yongpei Guan, University of Florida, Clustering under the Degree- Akaike Method USA; Jean-Paul Watson, Sandia National Corrected Stochastic Blockmodel Paulo J. S. Silva, University of Campinas, Laboratories, USA Tai Qin, University of Wisconsin, Madison, Brazil USA; Karl Rohe, University of Wisconsin, 6:30-6:55 A Tighter Formulation 6:30-6:55 On the Convergence and USA of a Mixed-Integer Program with Worst-Case Complexity of Trust- Powerflow Equations 6:30-6:55 Statistical and Region and Regularization Methods Quentin Louveaux, Bertrand Cornelusse, Computational Tradeoffs in for Unconstrained Optimization and Quentin Gemine, Université de Liège, Hierarchical Clustering Geovani Grapiglia and Jinyun Yuan, Federal Belgium Aarti Singh, Carnegie Mellon University, University of Paraná, Brazil; Ya-Xiang USA Yuan, Chinese Academy of Sciences, P.R. of China 38 2014 SIAM Conference on Optimization

Tuesday, May 20 Tuesday, May 20 Tuesday, May 20 MS55 MS56 MS57 Coordinate Descent Geometry of Linear Novelties from Theory and Methods: Parallelism, Optimization Part III of III: Computation in Risk Averse Duality and Non- Simplex Method and its Stochastic Programming smoothness- Part III of III Relatives 5:00 PM-7:00 PM 5:00 PM-7:00 PM 5:00 PM-7:00 PM Room:Towne Room:Sunrise Room:Sunset This minisymposium will bring together For Part 2 see MS43 For Part 2 see MS44 researchers from fairly different Coordinate descent methods are Linear programming, apart of its branches of stochastic programming becoming increasingly popular versatility to solve different types of sharing an interest in modeling, due to their ability to scale to big problems, is a fundamentally geometric analyzing, and computing risk. The talks data problems. The purpose of this theory. The two standard algorithms to concern non-traditional branches with symposium is to bring together solve it, the simplex method and the recent upswing such as semidefinite researchers working on developing interior point methods, take profit of stochastic programming, two-scale novel coordinate descent algorithms, this, one from the perspective of discrete pde-constrained models, stochastic analyzing their convergence / geometry, the other by approximating second-order cone programming, and complexity, and applying the algorithms a smooth curve. This minisymposium dominance constrained stochastic to new domains. is devoted to present recent progress programs. Organizer: Peter Richtarik in these and other algorithms, and to Organizer: Ruediger Schultz University of Edinburgh, United Kingdom discuss other geometric ideas relevant University of Duisburg-Essen, Germany for optimization. Organizer: Lin Xiao 5:00-5:25 SOCP Approach for Joint Microsoft Research, USA Organizer: Jesús A. De Loera Probabilistic Constraints University of California, Davis, USA Abdel Lisser, Michal Houda, and Jianqiang 5:00-5:25 Hydra: Distributed Cheng, Universite de Paris-Sud, France Coordinate Descent for Big Data Organizer: Francisco Santos Optimization University of Cantabria, Spain 5:30-5:55 Metric Regularity of Stochastic Dominance Constraints Martin Takac and Peter Richtarik, University 5:00-5:25 Augmentation Algorithms Induced by Linear Recourse of Edinburgh, United Kingdom for Linear and Integer Linear Matthias Claus, University of Duisburg- Programming 5:30-5:55 Parallel Coordinate Descent Essen, Germany for Sparse Regression Jesús A. De Loera, University of California, Joseph Bradley, University of California, Davis, USA; Raymond Hemmecke, 6:00-6:25 Two-Stage meets Two-Scale Berkeley, USA Technische Universität München, in Stochastic Shape Optimization Germany; Jon Lee, University of Benedict Geihe, Rheinische Friedrich- 6:00-6:25 The Rich Landscape Michigan, USA Wilhelms-Universität Bonn, Germany of Parallel Coordinate Descent Algorithms 5:30-5:55 Performance of the 6:30-6:55 Unit Commitment Under Chad Scherrer, Independent Consultant, Simplex Method on Markov Decision Uncertainty in AC Transmission USA; Ambuj Tewari, University of Processes Systems - a Risk Averse Two-Stage Michigan, USA; Mahantesh Halappanavar Ian Post, University of Waterloo, Canada Stochastic SDP Approach Tobias Wollenberg and Ruediger Schultz, and David Haglin, Pacific Northwest 6:00-6:25 An LP-Newton method for University of Duisburg-Essen, Germany National Laboratory, USA a standard form Linear Programming 6:30-6:55 An Asynchronous Parallel Problem Stochastic Coordinate Descent Shinji Mizuno and Tomonari Kitahara, Tokyo Algorithm Institute of Technology, Japan; Jianming Ji Liu and Stephen J. Wright, University of Shi, Tokyo University of Science, Japan Wisconsin, Madison, USA; Christopher 6:30-6:55 Colorful Linear Ré, Stanford University, USA; Victor Programming: Complexity and Bittorf, University of Wisconsin, Madison, Algorithms USA Frédéric Meunier, École des Ponts ParisTech, France 2014 SIAM Conference on Optimization 39

Tuesday, May 20 Tuesday, May 20 Wednesday, MS58 MS59 May 21 Numerical Linear Algebra Algorithms for Eigenvalue and Optimization - Optimization Part II of II 5:00 PM-7:00 PM Registration 5:00 PM-7:00 PM Room:Royal Palm 2 7:45 AM-5:00 PM Room:Royal Palm 1 Eigenvalue computation has been a Room:Golden Foyer For Part 1 see MS46 fundamental algorithmic component In optimization, discretization of PDEs, for solving semidefinite programming, low-rank matrix optimization, sparse least squares and other applications, Remarks linear systems sharing a common principal component analysis, sparse structure arise. The efficient solution inverse covariance matrix estimation, 8:10 AM-8:15 AM of those systems naturally leads to the total energy minimization in Room:Golden Ballroom efficient processes upstream. This electronic structure calculation, and minisymposium summarizes ongoing many other data-intensive applications forays into the construction and in science and engineering. This session implementation of efficient numerical will present a few recent advance on IP5 methods for such problems. both linear and nonlinear eigenvalue The Euclidean Distance Organizer: Dominique Orban optimization. Degree of an Algebraic Set École Polytechnique de Montréal, Canada Organizer: Zaiwen Wen 8:15 AM-9:00 AM Organizer: Anders Forsgren Peking University, China KTH Royal Institute of Technology, Sweden 5:00-5:25 Gradient Type Optimization Room:Golden Ballroom Methods for Electronic Structure Organizer: Annick Sartenaer Chair: Etienne De Klerk, Tilburg University, Calculations Universite Notre Dame de la Paix and The Netherlands Zaiwen Wen, Peking University, China University of Namur, Belgium It is a common problem in optimization 5:30-5:55 On the Convergence of the 5:00-5:25 Numerical Methods for to minimize the Euclidean distance Self-Consistent Field Iteration in Kohn- Symmetric Quasi-Definite Linear from a given data point u to some Sham Density Functional Theory Systems set X. In this talk I will consider the Xin Liu, Chinese Academy of Sciences, Dominique Orban, École Polytechnique de China; Zaiwen Wen, Peking University, situation in which X is defined by a Montréal, Canada China; Xiao Wang, University of Chinese finite set of polynomial equations. The 5:30-5:55 Block Preconditioners for Academy of Sciences, China; Michael number of critical points of the objective Saddle-Point Linear Systems Ulbrich, Technische Universität München, function on X is called the Euclidean Chen Greif, University of British Columbia, Germany distance degree of X, and is an intrinsic Canada 6:00-6:25 A Projected Gradient measure of the complexity of this 6:00-6:25 The Merits of Keeping It Algorithm for Large-scale Eigenvalue polynomial optimization problem. Smooth Calculation Algebraic geometry offers powerful Michael P. Friedlander, University of Chao Yang, Lawrence Berkeley National tools to calculate this degree in many British Columbia, Canada; Dominique Laboratory, USA situations. I will explain the algebraic Orban, École Polytechnique de Montréal, 6:30-6:55 Symmetric Low-Rank methods involved, and illustrate the Canada Product Optimization: Gauss-Newton formulas that can be obtained in several 6:30-6:55 A Primal-Dual Interior Point Method situations ranging from matrix analysis Solver for Conic Problems with the Yin Zhang, Rice University, USA; Xin Liu, to control theory to computer vision. Exponential Cone Chinese Academy of Sciences, China; Joint work with Jan Draisma, Emil Santiago Akle and Michael A. Saunders, Zaiwen Wen, Peking University, China Horobet, Giorgio Ottaviani and Bernd Stanford University, USA Sturmfels. Rekha Thomas University of Washington, USA

Coffee Break 9:00 AM-9:30 AM Room:Town & Country 40 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MT2 MS60 MS61 Mixed-integer Nonlinear Accounting for Uncertainty Variational Analysis and Optimization During Optimization Monotone Operator Theory: 9:30 AM-11:30 AM Routines Splitting Methods Room:Golden Ballroom 9:30 AM-11:30 AM 9:30 AM-11:30 AM Chair: Jeff Linderoth, University of Room:Pacific Salon 1 Room:Pacific Salon 2 Wisconsin, Madison, USA Considering the resilience and efficiency Variational Analysis and Monotone Chair: Sven Leyffer, Argonne National of modern systems characterized Operator Theory lie at the heart Laboratory, USA by growing complexity poses many of modern optimization. The Chair: James Luedtke, University of new challenges for simulation-based minisymposium will focus on recent Wisconsin, Madison, USA design and analysis. Among those theoretical advancements, applications, Mixed-integer nonlinear programming is the need to move from traditional and algorithms. problems (MINLPs) combine deterministic optimization techniques Organizer: Radu Ioan Bot the combinatorial complexity of to techniques that account for a variety University of Vienna, Austria of uncertainties in the systems being discrete decisions with the numerical 9:30-9:55 A Relaxed Projection challenges of nonlinear functions. modeled and studied. The salient Splitting Algorithm for Nonsmooth In this minitutorial, we will describe technical problems in this field are Variational Inequalities in Hilbert applications of MINLP in science and characterization and quantification of Spaces engineering, demonstrate the importance uncertainties, as well as the frequently Jose Yunier Bello Cruz, Federal University of building “good” MINLP models, prohibitive expense of uncertainty- of Goiás, Brazil, and University of British Columbia, Canada discuss numerical techniques for solving based computational analysis. In this MINLPs, and survey the forefront of minisymposium, we examine these 10:00-10:25 Forward-Backward and ongoing research topics in this important problems, both in general theoretical Tseng’s Type Penalty Schemes for and emerging area. context and in specific applications, Monotone Inclusion Problems where a variety of uncertainty Radu Ioan Bot, University of Vienna, Austria Speakers: expressions arise. Sven Leyffer, Argonne National Laboratory, 10:30-10:55 The Douglas-Rachford USA Organizer: Genetha Gray Algorithm for Nonconvex Feasibility Sandia National Laboratories, USA Problems Jeff Linderoth, University of Wisconsin, Robert Hesse, University of Goettingen, Madison, USA Organizer: Patricia D. Hough Germany Sandia National Laboratories, USA James Luedtke, University of Wisconsin, 11:00-11:25 Linear Convergence of Madison, USA Organizer: Natalia Alexandrov the Douglas-Rachford Method for NASA Langley Research Center, USA Two Nonconvex Sets 9:30-9:55 Multi-Information Source Hung M. Phan, University of British Optimization: Beyond Mfo Columbia, Canada David Wolpert, Santa Fe Institute, USA 10:00-10:25 Accounting for Uncertainty in Complex System Design Optimization Natalia Alexandrov, NASA Langley Research Center, USA 10:30-10:55 Quasi-Newton Methods for Stochastic Optimization Michael W. Trosset and Brent Castle, Indiana University, USA 11:00-11:25 Beyond Variance-based Robust Optimization Gianluca Iaccarino, Stanford University, USA; Pranay Seshadri, Rolls-Royce Canada Limited, Canada; Paul Constantine, Colorado School of Mines, USA 2014 SIAM Conference on Optimization 41

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MS62 MS63 MS64 Robust Optimization Polynomial and Tensor Polynomial and Copositive 9:30 AM-11:30 AM Optimization: Algorithms, Optimization Room:Pacific Salon 3 Theories, and Applications 9:30 AM-11:30 AM Robust Optimization has become an 9:30 AM-11:30 AM Room:Pacific Salon 5 important field in optimization. The Room:Pacific Salon 4 Polynomial optimization has become a developed theory has been successfully Multi-way data (or higher-order very active area during the past decade. applied to a wide variety of applications. tensor) naturally arises in many The first talk presents results for an Although this field is already 15 years areas including, but not limited to, extension of the nonconvex trust region old, still many new discoveries are neuroscience, chemometrics, computer problem to the case of two constraints. made. In this minisymposium some vision, machine learning, social The technique used, copositivity, is a recent research results on (adjustable) network, and graph analysis. Motivated property that is difficult to establish robust optimization will be presented. by applications and the increasing in general; the second talk introduces Organizer: Dick Den Hertog computing power, polynomial and tensor an approach based on semidefinite Tilburg University, The Netherlands optimization has become very popular in approximation. The third talk addresses 9:30-9:55 Adjustable Robust the past few years. This minisymposium the problem of checking positivity of Optimization with Decision Rules will show some most recent advances polynomials which are not sums of Based on Inexact Information and bring discussions on algorithms, squares, introducing a generalization of Frans de Ruiter, Tilburg University, The theories, and applications of the this notion. The fourth talk presents a Netherlands; Aharon Ben-Tal, Technion polynomial and tensor optimization. polynomial root optimization problem - Israel Institute of Technology, Israel; The speakers will talk about robust with two constraints on the coefficients. Ruud Brekelmans and Dick Den Hertog, nonnegative matrix/tensor factorization, Tilburg University, The Netherlands Organizer: Michael L. Overton improved convex relaxation for tensor Courant Institute of Mathematical Sciences, 10:00-10:25 Robust Growth-Optimal recovery, and parallel block coordinate New York University, USA Portfolios descent method for tensor completion. Napat Rujeerapaiboon, EPFL, Switzerland; Organizer: Henry Wolkowicz Daniel Kuhn, École Polytechnique Organizer: Yangyang Xu University of Waterloo, Canada Fédérale de Lausanne, Switzerland; Rice University, USA 9:30-9:55 Narrowing the Difficulty Wolfram Wiesemann, Imperial College 9:30-9:55 Robust Nonnegative Matrix/ Gap for the Celis-Dennis-Tapia London, United Kingdom Tensor Factorization with Asymmetric Problem 10:30-10:55 Globalized Robust Soft Regularization Immanuel Bomze, University of Vienna, Optimization for Nonlinear Uncertain Hyenkyun Woo, Korea Institute for Advanced Austria; Michael L. Overton, Courant Inequalities Study, Korea; Haesun Park, Georgia Institute of Mathematical Sciences, New Ruud Brekelmans, Tilburg University, The Institute of Technology, USA York University, USA Netherlands; Aharon Ben-Tal, Technion 10:00-10:25 Square Deal: Lower 10:00-10:25 A Nonlinear Semidefinite - Israel Institute of Technology, Israel; Bounds and Improved Relaxations for Approximation for Copositive Dick Den Hertog, Tilburg University, The Tensor Recovery Optimisation Netherlands; Jean-Philippe Vial, Ordecsys, Cun Mu, Bo Huang, John Wright, and Peter Dickinson, University of Vienna, Switzerland , Columbia University, Austria; Juan C. Vera, Tilburg University, 11:00-11:25 Routing Optimization with USA The Netherlands Deadlines under Uncertainty 10:30-10:55 Inverse Scale Space: 10:30-10:55 Real PSD Ternary Forms Patrick Jaillet, Massachusetts Institute of New Regularization Path for Sparse with Many Zeros Technology, USA; Jin Qi and Melvyn Regression Bruce Reznick, University of Illinois at Sim, National University of Singapore, Ming Yan, Wotao Yin, and Stanley J. Osher, Urbana-Champaign, USA Singapore University of California, Los Angeles, 11:00-11:25 Optimal Solutions to USA a Root Minimization Problem over 11:00-11:25 Computing Generalized a Polynomial Family with Affine Tensor Eigenpairs Constraints Tamara G. Kolda and Jackson Mayo, Sandia Julia Eaton, University of Washington, USA; National Laboratories, USA Sara Grundel, Max Planck Institute for Dynamics of Complex Systems, Germany; Mert Gurbuzbalaban, Massachusetts Institute of Technology, USA; Michael L. Overton, Courant Institute of Mathematical Sciences, New York University, USA 42 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MS65 MS66 MS67 Algorithms for Large Scale First-order Methods for Risk Quadrangles: SDP and Related Problems Convex Optimization Connections between 9:30 AM-11:30 AM 9:30 AM-11:30 AM Risk, Estimation, and Room:Pacific Salon 6 Room:Sunrise Optimization We report recent advances in first and First-order methods are currently the 9:30 AM-11:30 AM second order methods for solving some most effective class of algorithms for Room:Sunset classes of large scale SDPs and related large-scale optimization problems Quantification of risk is an integral problems. The focus will be to exploit arising in a variety of important timely part of optimization in the presence favourable underlying structures in these applications. This minisymposium of uncertainty and is now a well- classes of problems so that efficient will present some recent advances in developed area. The connections algorithms can be designed to solve this active area of research involving between risk and statistical estimation, large scale problems. novel algorithmic developments and however, have only recently come to the Organizer: Kim-Chuan Toh applications. forefront. Estimation is a prerequisite National University of Singapore, Republic Organizer: Javier Pena for risk quantification as probability of Singapore Carnegie Mellon University, USA distributions are rarely known and must 9:30-9:55 Title Not Available 9:30-9:55 An Extragradient-Based be estimated from data. But connections Renato C. Monteiro, Georgia Institute of Alternating Direction Method for exist also on a more fundamental level Technology, USA Convex Minimization as revealed through quadrangles of risk Shiqian Ma, Chinese University of Hong 10:00-10:25 Constrained Best that link measures of risk, error, regret, Kong, Hong Kong; Shuzhong Zhang, Euclidean Distance Embedding on and deviation. The minisymposium University of Minnesota, USA a Sphere: a Matrix Optimization gives an overview of these connections Approach 10:00-10:25 An Adaptive Accelerated and shows their profound influence Houduo Qi and Shuanghua Bai, University Proximal Gradient Method and Its on risk minimization, regression of Southampton, United Kingdom; Naihua Homotopy Continuation for Sparse methodology, and support vector Xiu, Beijing Jiaotong University, China Optimization machines. Qihang Lin, University of Iowa, USA; Lin 10:30-10:55 Inexact Proximal Xiao, Microsoft Research, USA Organizer: Johannes O. Royset Point and Augmented Lagrangian Naval Postgraduate School, USA Methods for Solving Convex Matrix 10:30-10:55 Decompose Composite Optimization Problems Problems for Big-data Optimization 9:30-9:55 The Fundamental Kim-Chuan Toh, Defeng Sun, and Kaifeng Guanghui Lan, University of Florida, USA Quadrangle of Risk in Optimization Jiang, National University of Singapore, and Statistics 11:00-11:25 An Acceleration Republic of Singapore Terry Rockafellar, University of Washington, Procedure for Optimal First-order USA 11:00-11:25 SDPNAL+: A Majorized Methods Semismooth Newton-CG Augmented Michel Baes and Michael Buergisser, ETH 10:00-10:25 Support Vector Machines Lagrangian Method for Semidefinite Zürich, Switzerland Based on Convex Risk Functionals Programming with Nonnegative and General Norms Constraints Jun-ya Gotoh, Chuo University, Japan; Stan Liuqin Yang, Defeng Sun, and Kim-Chuan Uryasev, University of Florida, USA Toh, National University of Singapore, 10:30-10:55 Cvar Norm in Republic of Singapore Deterministic and Stochastic Case: Applications in Statistics Stan Uryasev and Alexander Mafuslav, University of Florida, USA 11:00-11:25 Superquantile Regression: Theory and Applications Johannes O. Royset, Naval Postgraduate School, USA; R T. Rockafellar, University of Washington, USA 2014 SIAM Conference on Optimization 43

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MS68 MS69 MS70 Model-based Predictive Optimization in Signal Nonconvex Multicommodity Control for Energy Processing and Control Flow Problems Consumption in Buildings 9:30 AM-11:30 AM 9:30 AM-11:00 AM 9:30 AM-11:30 AM Room:Royal Palm 1 Room:Royal Palm 2 Room:Towne Optimization is revolutionizing This session will feature four talks Heat transfer simulation and computational methodology in on nonconvex multicommodity flow optimization are becoming instrumental signal processing and control. In this problems, which have applications for a broad range of applications. In minisymposium we wish to bring in logistics, transportation and the past years, the increase in energy together some of the leading efforts in telecommunications. Problems with costs alongside governments and this area, and understand connections piecewise linear and piecewise convex municipalities legislation have further between optimization and diverse separable (by arc) objective functions emphasized the importance of accurate problems within signal processing will be considered. Several approaches prediction and reduction of energy and control. This minisymposium will will be presented, based on mixed consumption. In this session, model- feature applications of optimization to integer programming and decomposition based predictive control for energy areas such as learning embeddings of methods. consumption in buildings will be data, decentralized control, computing Organizer: Bernard Gendron addressed. The models can either be optimal power flows in networks, Université de Montréal, Canada and learning covariance matrices. physics based, data-driven models built 9:30-9:55 Piecewise Convex using machine learning techniques or a This is intended to be a companion Multicommodity Flow Problems combination of both. Minisymposium to “Optimization for Philippe Mahey, Universite Blaise Pascal, Statistical Inference” by being proposed Organizer: Raya Horesh France by Venkat Chandrasekaran. IBM T.J. Watson Research Center, USA 10:00-10:25 Fixed-Charge 9:30-9:55 Pde-Ode Modeling, Organizer: Parikshit Shah Multicommodity Flow Problems Inversion and Optimal Control for Philips Research North America, USA Enrico Gorgone, Università della Calabria, Building Energy Demand Organizer: Venkat Italy Raya Horesh, IBM T.J. Watson Research Chandrasekaran 10:30-10:55 Piecewise Linear Center, USA California Institute of Technology, USA Multicommodity Flow Problems Bernard Gendron, Université de Montréal, 10:00-10:25 Stochastic Predictive 9:30-9:55 Using Regularization for Canada Control for Energy Efficient Buildings Design: Applications in Distributed Francesco Borrelli and Tony Kelman, Control University of California, Berkeley, USA Nikolai Matni, California Institute of 10:30-10:55 Performance Technology, USA Lunch Break Optimization of Hvac Systems 10:00-10:25 Learning Near-Optimal Andrew Kusiak, University of Iowa, USA Linear Embeddings 11:30 AM-1:00 PM 11:00-11:25 Development of Control- Richard G. Baraniuk, Rice University, Attendees on their own Oriented Models for Model Predictive USA; Chinmay Hegde, Massachusetts Control in Buildings Institute of Technology, USA; Aswin Zheng O’Neill, University of Alabama, USA Sankaranarayanan, Carnegie Mellon University, USA; Wotao Yin, University of California, Los Angeles, USA 10:30-10:55 Covariance Sketching Parikshit Shah, Gautam Dasarathy, and Badri Narayan Bhaskar, University of Wisconsin, USA; Rob Nowak, University of Wisconsin, Madison, USA 11:00-11:25 Convex Relaxation of Optimal Power Flow Steven Low, California Institute of Technology, USA 44 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 IP6 MS12 MS71 Modeling Wholesale Novel Approaches to Mixed Optimization for Statistical Electricity Markets with Integer Second Order Cone Inference Hydro Storage Optimization 2:00 PM-4:00 PM 1:00 PM-1:45 PM 2:00 PM-3:30 PM Room:Golden Ballroom Room:Golden Ballroom Room:Royal Palm 1 Optimization is one of the leading Chair: Daniel Ralph, University of Second-Order Cone Optimization frameworks for developing new Cambridge, United Kingdom (SOCO) is one of the most important statistical inference methodologies. On the conceptual side, inferential Over the past two decades most areas of conic linear optimization. objectives are naturally stated in a industrialized nations have instituted Just as in any other optimization variational manner, e.g., optimizing a markets for wholesale electricity supply. problems, integer variables naturally data fidelity criterion subject to model Optimization models play a key role occur in many applications of SOCO, complexity constraints in statistical in these markets. The understanding thus leading to Mixed Integer SOCO - estimation. On the applied side, of electricity markets has grown MISOCO. The design and computation special-purpose algorithms for convex substantially through various market of nonlinear cuts for MISOCO gained optimization are the tools of choice crises, and there is now a set of standard considerable attention in recent years. in many large-scale data analysis electricity market design principles This minisymposium reviews the problems. This minisymposium will for systems with mainly thermal plant. various disjunctive cut generation survey the latest research results in this On the other hand, markets with lots approaches, reviews how these exciting area. This minisymposium of hydro storage can face shortage disjunctive conic cuts can efficiently is a companion to the proposed risks, leading to production and pricing be used in solving MISOCO problems. “Optimization in Control and Signal arrangements in these markets that This minisymposium offers also an Processing” by Parikshit Shah, and we vary widely across jurisdictions. We opportunity to review and anticipate expect considerable synergies between show how stochastic optimization and generalizations of the methodology these two proposed minisymposia. complementarity models can be used to more general mixed ointeger conic to improve our understanding of these linear optimization problems. Organizer: Venkat systems. Organizer: Tamas Terlaky Chandrasekaran California Institute of Technology, USA Andy Philpott Lehigh University, USA University of Auckland, New Zealand 2:00-2:25 Disjunctive Conic Cuts for Organizer: Parikshit Shah Mixed Integer Second Order Cone Philips Research North America, USA Optimization 2:00-2:25 Denoising for Julio C. Goez, Lehigh University, USA; Simultaneously Structured Signals Intermission Pietro Belotti, FICO, United Kingdom; Maryam Fazel, University of Washington, 1:45 PM-2:00 PM Imre Pólik, SAS Institute, Inc., USA; USA Ted Ralphs and Tamas Terlaky, Lehigh University, USA 2:30-2:55 An Enhanced Conditional Gradient Algorithm for Atomic-Norm 2:30-2:55 Network Design with Regularization Probabilistic Capacity Constraints Nikhil Rao, Parikshit Shah, and Stephen Alper Atamturk and Avinash Bhardwaj, Wright, University of Wisconsin, USA University of California, Berkeley, USA 3:00-3:25 Learning from 3:00-3:25 Mixed Integer Programming Heterogenous Observations with p-Order Cone and Related Philippe Rigollet, Princeton University, USA Constraints Alexander Vinel and Pavlo Krokhmal, 3:30-3:55 Algorithmic Blessings of University of Iowa, USA High Dimensional Problems Arian Maleki, Columbia University, USA 2014 SIAM Conference on Optimization 45

Wednesday, May 21 Wednesday, May 21 3:00-3:25 Provable and Practical Topic Modeling Using NMF MS74 MS75 Rong Ge, Microsoft Research, USA; Sanjeev Arora, Princeton University, USA; Yoni Risk-Averse Optimization Recent Advances in Halpern, New York University, USA; 2:00 PM-4:00 PM Nonnegative Matrix David Mimno, Princeton University, USA; Ankur Moitra, Institute for Advanced Room:Pacific Salon 3 Factorization and Related Study, USA; David Sontag, New York The minisymposium will address Problems University, USA; Yichen Wu and Michael new mathematical, statistical, and 2:00 PM-4:00 PM , Princeton University, USA 3:30-3:55 Identifying K Largest computational issues associated with Room:Pacific Salon 4 risk-averse optimization. The specificity Submatrices Using Ky Fan Matrix of risk-averse problems leads to Nonnegative matrix factorization Norms new theoretical and computational (NMF) has recently attracted a lot Xuan Vinh Doan, University of Warwick, United Kingdom; Stephen A. Vavasis, challenges. Statistical estimation of of attention from researchers from University of Waterloo, Canada risk measures, which is necessary for rather different areas such as machine construction of optimization models learning, continuous and discrete based on real and simulated data, optimization, numerical linear algebra requires novel approaches. Next, using and theoretical computer science. In this risk measures in dynamic optimization minisymposium, we will review some problems creates new questions recent advances on this topic, focusing associated with time-consistency. Game- on algorithms that can guarantee like problems arise in models with recovery of underlying topics for certain inconsistent measures, which pose new structures or guarantee bounds on computational challenges. Finally, using nonnegative rank. Some applications risk measures in problem with long, (e.g., hyperspectral imaging and or possibly infinite horizon, requires document classification) will also be new approaches utilizing the Markov covered. structure of the problem. Organizer: Nicolas Gillis Universite de Mons, Belgium Organizer: Andrzej Ruszczynski Rutgers University, USA Organizer: Stephen A. Vavasis University of Waterloo, Canada 2:00-2:25 Challenges in Dynamic Risk-Averse Optimization 2:00-2:25 Semidefinite Programming Andrzej Ruszczynski, Rutgers University, Based Preconditioning for More USA Robust Near-Separable Nonnegative Matrix Factorization 2:30-2:55 Statistical Estimation of Nicolas Gillis, Universite de Mons, Belgium; Composite Risk Measures and Risk Stephen A. Vavasis, University of Optimization Problems Waterloo, Canada Darinka Dentcheva, Stevens Institute of Technology, USA 2:30-2:55 Convex Lower Bounds for Atomic Cone Ranks with Applications 3:00-3:25 Multilevel Optimization to Nonnegative Rank and cp-rank Modeling for Stochastic Programming Hamza Fawzi and Pablo A. Parrilo, with Coherent Risk Measures Massachusetts Institute of Technology, Jonathan Eckstein, Rutgers University, USA USA 3:30-3:55 Threshold Risk Measures: Dynamic Infinite Horizon and Approximations Ricardo A. Collado, Stevens Institute of Technology, USA

continued in next column 46 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MS76 MS77 MS78 Conic Programming Semidefinite Programming First-Order Methods for Approaches to Polynomial Theory and Applications Convex Optimization: Optimization and Related 2:00 PM-4:00 PM Theory and Applications Problems Room:Pacific Salon 6 2:00 PM-4:00 PM 2:00 PM-4:00 PM Semidefinite programminig (SDP) Room:Sunrise Room:Pacific Salon 5 provides a framework for an elegant The use and analysis of first-order Polynomial optimization problems theory, efficient algorithms, and methods in convex optimization has (POPs) have applications in powerful application tools. We gained a considerable amount of combinatorial optimization, finance, concentrate on the theory and sensitivity attention in recent years due to the control and various engineering analysis, as well as applications to relevance of applications (regression, disciplines. In a seminal paper in 2001, Euclidean Distance Matrix Completions boosting/classification, image Lasserre introduced an interesting and Molecular Conformation. construction, etc.), the requirement approach to solving POPs via a Organizer: Henry Wolkowicz for only moderately high accuracy hierarchy of semidefinite programs. This University of Waterloo, Canada solutions, the necessity for such simple methods for huge-scale problems, and approach combines techniques from real 2:00-2:25 Efficient Use of Semidefinite algebraic geometry, the theory of the Programming for Selection of the appeal of structural implications problem of moments, and convex conic Rotamers in Protein Conformations (sparsity, low-rank) induced by the programming. It has led to more than a Henry Wolkowicz, Forbes Burkowski, models and the methods themselves. decade of interesting research, and the and Yuen-Lam Cheung, University of The research herein includes extensions goal of this session is to present some of Waterloo, Canada of the Frank-Wolfe method, stochastic the recent developments in the field. 2:30-2:55 On Universal Rigidity of and projected gradient methods, and Tensegrity Frameworks, a Gram applications in signal processing and Organizer: Etienne De Klerk Matrix Approach statistical estimation. Tilburg University, The Netherlands Abdo Alfakih, University of Windsor, Organizer: Robert M. Freund 2:00-2:25 Approximation Quality of Canada; Viet-Hang Nguyen, G-SCOP, Massachusetts Institute of Technology, USA SOS Relaxations Grenoble, France Pablo A. Parrilo, Massachusetts Institute of 2:00-2:25 Minimizing Finite Sums with 3:00-3:25 On the Sensitivity of Technology, USA the Stochastic Average Gradient Semidefinite Programs and Second Algorithm 2:30-2:55 Dsos and Sdsos: More Order Cone Programs Mark Schmidt, Simon Fraser University, Tractable Alternatives to Sum of Yuen-Lam Cheung and Henry Wolkowicz, Canada Squares Programming University of Waterloo, Canada Amir Ali Ahmadi, IBM Research, USA; 2:30-2:55 Frank-Wolfe-like Methods 3:30-3:55 Combinatorial Certificates Anirudha Majumdar, Massachusetts for Large-scale Convex Optimization in Semidefinite Duality Institute of Technology, USA Paul Grigas and Robert M. Freund, Gabor Pataki, University of North Carolina Massachusetts Institute of Technology, 3:00-3:25 An Alternative Proof of a at Chapel Hill, USA USA PTAS for Fixed-degree Polynomial Optimization over the Simplex 3:00-3:25 Statistical Properties for Etienne De Klerk, Tilburg University, The Computation Netherlands; Monique Laurent, CWI, Sahand Negahban, Yale University, USA Amsterdam, Netherlands; Zhao Sun, 3:30-3:55 Frank-Wolfe and Greedy Tilburg University, The Netherlands Algorithms in Optimization and Signal 3:30-3:55 Lower Bounds for a Processing Polynomial on a basic Closed Martin Jaggi, University of California, Semialgebraic Set using Geometric Berkeley, USA Programming Mehdi Ghasemi, Nanyang Technological University, Singapore; Murray Marshall, University of Saskatchewan, Canada 2014 SIAM Conference on Optimization 47

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MS79 MS80 MS81 Analysis and Algorithms for Advanced Algorithms for Network Design: New Markov Decision Processes Constrained Nonlinear Approaches and Robust 2:00 PM-4:00 PM Optimization Solutions Room:Sunset 2:00 PM-4:00 PM 2:00 PM-4:00 PM Markov decision process (MDP) is the Room:Towne Room:Pacific Salon 1 most popular framework for sequential This minisymposium highlights In its classical version, the task is to decision-making problems. Theoretical recent developments in the design of assign minimum-cost capacities to the properties and solution methods of algorithms for solving constrained edges of a network such that specified MDPs have been extensively studied nonlinear optimization problems. In demands can always be routed. Due in the last several decades. However, particular, the focus is on the continuing to the NP-hardness of the problem, there are many open questions about challenge of designing methods that globally optimum solutions for large performance of those solution methods are computationally efficient, but are network design problems are often out and how to approach large-scale also backed by strong global and local of reach. In this minisymposium, we MDPs. Due to limitations in different convergence guarantees. The session introduce advanced solution concepts application domains, several extensions involves discussions of interior-point for classical design problems as well of MDPs are also studied, such as methodologies, but also renewed as new methodologies for solving the MDPs with partial observability and interests in designing efficient active- robust optimization versions. MDPs with unknown parameters that set methods for solving large-scale Organizer: Frauke Liers have to be estimated by exploring its problems. Universität Erlangen-Nürnberg, Germany environment. In this minisymposium, Organizer: Frank E. Curtis approaches to MDPs with those settings 2:00-2:25 The Recoverable Robust Lehigh University, USA Two-Level Network Design Problem and extensions are discussed. Organizer: Andreas Waechter Eduardo A. Alvarez-Miranda, Universidad Organizer: Ilbin Lee Northwestern University, USA de Talca, Chile; Ivana Ljubic, University University of Michigan, USA of Vienna, Austria; S. Raghavan, 2:00-2:25 An Interior-Point Trust-Funnel 2:00-2:25 Sensitivity Analysis of University of Maryland, USA; Paolo Toth, Algorithm for Nonlinear Optimization University of Bologna, Italy Markov Decision Processes and Daniel Robinson, Johns Hopkins University, Policy Convergence Rate of Model- USA; Frank E. Curtis, Lehigh University, 2:30-2:55 Solving Network Design Based Reinforcement Learning USA; Nicholas I.M. Gould, Rutherford Problems Via Iterative Aggregation Marina A. Epelman and Ilbin Lee, Appleton Laboratory, United Kingdom; Andreas Bärmann, Friedrich-Alexander- University of Michigan, USA Philippe L. Toint, University of Namur, Universität Erlangen-Nürnberg, Germany; 2:30-2:55 On the Use of Non- Belgium Frauke Liers and Alexander Martin, Universität Erlangen-Nürnberg, Germany; Stationary Policies for Stationary 2:30-2:55 Globalizing Stabilized SQP Maximilian Merkert, Christoph Thurner, Infinite-Horizon Markov Decision with a Smooth Exact Penalty Function and Dieter Weninger, Friedrich-Alexander- Processes Alexey Izmailov, Computing Center Universität Erlangen-Nürnberg, Germany Bruno Scherrer, INRIA, France of Russian Academy of Sciences, 3:00-3:25 Partially Observable Russia; Mikhail V. Solodov, Instituto 3:00-3:25 Robust Network Design Markov Decision Processes with de Matematica Pura E Aplicada, Brazil; for Simple Polyhedral Demand General State and Action Spaces Evgeniy Uskov, Moscow State University, Uncertainties Eugene A. Feinberg, Stony Brook Russia Valentina Cacchiani, University of Bologna, University, USA; Pavlo Kasyanov and Italy; Michael Jünger, University of 3:00-3:25 A Filter Method with Unified Michael Zgurovsky, National Technical Cologne, Germany; Frauke Liers, Step Computation University of Ukraine, Ukraine Universität Erlangen-Nürnberg, Germany; Yueling Loh, Johns Hopkins University, Andrea Lodi, University of Bologna, Italy; 3:30-3:55 Dantzig’s Pivoting Rule for USA; Nicholas I.M. Gould, Rutherford Daniel Schmidt, University of Cologne, Shortest Paths, Deterministic Markov Appleton Laboratory, United Kingdom; Germany Decision Processes, and Minimum Daniel Robinson, Johns Hopkins Cost to Time Ratio Cycles University, USA 3:30-3:55 Network Design under Thomas D. Hansen, Stanford University, Compression and Caching Rates 3:30-3:55 Polyhedral Projection USA; Haim Kaplan and Uri Zwick, Tel Uncertainty William Hager, University of Florida, Aviv University, Israel Arie Koster and Martin Tieves, RWTH USA; Hongchao Zhang, Louisiana State Aachen University, Germany University, USA 48 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 MS82 MS119 CP14 Applications of Set-Valued Perspectives from NSF Derivative-Free Analysis Program Directors Optimization 2:00 PM-4:00 PM 2:00 PM-4:00 PM 4:30 PM-6:10 PM Room:Royal Palm 2 Room:Pacific Salon 7 Room:Golden Ballroom The tools of set-valued and variational Organizer: Sheldon H. Jacobson Chair: Thomas Asaki, Washington State analysis are becoming increasingly University of Illinois at Urbana-Champaign University, USA mainstream in their application to and National Science Foundation, USA 4:30-4:45 Coupling Surrogates with fields as varied as signal and image Organizer: Edwin Romeijn Derivative-Free Continuous and processing, control, and dynamical National Science Foundation, USA Mixed Integer Optimization Methods to Solve Problems Coming from systems. This session highlights some 2:00-2:55 Funding Opportunities in the recent developments in the analysis of Industrial Aerodynamic Applications Service and Manufacturing Enterprise Anne-Sophie Crélot, University of Namur, convergence, regularity and stability of Systems Programs Belgium; Dominique Orban, École Edwin Romeijn, National Science algorithms that are fundamental to these Polytechnique de Montréal, Canada; Foundation, USA application fields. Caroline Sainvitu, CENAERO, Belgium; Organizer: Russell Luke 3:00-3:55 Funding Opportunities at Annick Sartenaer, Universite Notre Dame University of Goettingen, Germany NSF de la Paix and University of Namur, Sheldon H. Jacobson, University of Illinois Belgium 2:00-2:25 Global and Local Linear at Urbana-Champaign and National 4:50-5:05 Parallel Hybrid Convergence of Projection-Based Science Foundation, USA Algorithms for Sparse Affine Feasibility Multiobjective Derivative-Free Russell Luke, Robert Hesse, and Patrick Optimization in SAS Neumann, University of Goettingen, Steven Gardner, Joshua Griffin, and Lois Germany Coffee Break Zhu, SAS Institute, Inc., USA 2:30-2:55 Convergence, Robustness, 4:00 PM-4:30 PM 5:10-5:25 A Derivative-Free Method and Set-Valued Lyapunov Functions for Optimization Problems with Rafal Goebel, Loyola University of Chicago, Room:Town & Country General Inequality Constraints USA Phillipe R. Sampaio, University of Namur, Belgium; Philippe L. Toint, Facultés 3:00-3:25 Regularization Methods for Universitaires Notre-Dame de la Paix, Variational Inequalties Belgium Charitha Charitha, University of Goettingen, Germany; Joydeep Dutta, Indian Institute 5:30-5:45 An SQP Trust-Region of Technology, Kanpur, India; Russell Algorithm for Derivative-Free Luke, University of Goettingen, Germany Constrained Optimization Anke Troeltzsch, German Aerospace Center 3:30-3:55 Generalized Solutions for (DLR), Germany the Sum of Two Maximally Monotone Operators 5:50-6:05 Deterministic Mesh Walaa Moursi, University of British Adaptive Direct Search with Uniformly Columbia, Canada Distributed Polling Directions Thomas Asaki, Washington State University, USA 2014 SIAM Conference on Optimization 49

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 CP15 CP16 CP17 Robust Optimization Optimization with Matrices Applications of Integer 4:30 PM-6:30 PM 4:30 PM-6:30 PM Optimization Room:Pacific Salon 1 Room:Pacific Salon 2 4:30 PM-6:30 PM Chair: Jorge R. Vera, Universidad Catolica Chair: Christos Boutsidis, IBM T.J. Watson Room:Pacific Salon 3 de Chile, Chile Research Center, USA Chair: Christos Maravelias, University of 4:30-4:45 Medication Inventory 4:30-4:45 Parallel Matrix Factorization Wisconsin, USA Management in the International for Low-Rank Tensor Recovery 4:30-4:45 Power Efficient Uplink Space Station (ISS): A Robust Yangyang Xu, Rice University, USA Scheduling in SC-FDMA: Bounding Optimization Approach 4:50-5:05 Ellipsoidal Rounding for Global Optimality by Column Sung Chung and Oleg Makhnin, New Nonnegative Matrix Factorization Generation Mexico Institute of Mining and Under Noisy Separability Hongmei Zhao, Linköping University, Technology, USA Tomohiko Mizutani, Kanagawa University, Sweden 4:50-5:05 Robust Solutions for Systems Japan 4:50-5:05 A Mixed Integer of Uncertain Linear Equations 5:10-5:25 On Solving An Application Programming Approach for the Jianzhe Zhen and Dick Den Hertog, Tilburg Based Completely Positive Program Police Districting Problem University, The Netherlands of Size 3 Víctor D. Bucarey and Fernando Ordoñez, 5:10-5:25 Tractable Robust Qi Fu, University of Macau, Macao SAR, Universidad de Chile, Chile; Vladimir Counterparts of Distributionally Robust China; Chee-Khian Sim and Chung-Piaw Marianov, Pontificia Universidad Católica Constraints on Risk Measures Teo, National University of Singapore, de Chile, Chile Krzysztof Postek, Bertrand Melenberg, and Singapore 5:10-5:25 Design and Operating Dick Den Hertog, Tilburg University, The 5:30-5:45 A Semidefinite Strategy Optimization of Wind, Diesel Netherlands Approximation for Symmetric and Batteries Hybrid Systems for 5:30-5:45 Distributionally Robust Travelling Salesman Polytopes Isolated Sites Inventory Control When Demand Is a Julián Romero and Mauricio Velasco, Thibault Barbier and Gilles Savard, École Martingale University of los Andes, Colombia Polytechnique de Montréal, Canada; Linwei Xin and David Goldberg, Georgia Miguel Anjos, École Polytechnique de 5:50-6:05 Convex Sets with Institute of Technology, USA Montréal, Canada Semidefinite Representable Sections 5:50-6:05 Robust Optimization of a and Projections 5:30-5:45 Rectangular Shape Class of Bi-Convex Functions Anusuya Ghosh and Vishnu Narayanan, Management Zone Delineation in Erick Delage and Amir Ardestani-Jaafari, Indian Institute of Technology-Bombay, Agriculture Planning by Using Column HEC Montréal, Canada India Generation Linco J. Ñanco and Víctor Albornoz, 6:10-6:25 Robust Approaches for 6:10-6:25 Optimal Cur Matrix Universidad Técnica Federico Santa Stochastic Intertemporal Production Decompositions María, Chile Planning Christos Boutsidis, IBM T.J. Watson Jorge R. Vera, Universidad Catolica de Research Center, USA; David Woodruff, 5:50-6:05 Mixed Integer Linear Chile, Chile; Alfonso Lobos, Pontificia IBM Almaden Research Center, USA Programming Approach to a Rough Universidad Católica de Chile, Chile; Cut Capacity Planning in Yogurt Robert M. Freund, Massachusetts Institute Production of Technology, USA Sweety Hansuwa, Indian Institute of Technology-Bombay, India; Pratik Fadadu, IGSA Labs Private Ltd., Hyderabad, India; Atanu Chaudhuri and Rajiv Kumar Srivastava, Indian Institute of Management, Lucknow, India 6:10-6:25 Solution Methods for Mip Models for Chemical Production Scheduling Sara Velez and Christos Maravelias, University of Wisconsin, USA 50 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 CP18 CP19 CP20 Stochastic Optimization III Nonsmooth Optimization Optimal Control III 4:30 PM-6:30 PM 4:30 PM-6:30 PM 4:30 PM-6:30 PM Room:Pacific Salon 4 Room:Pacific Salon 5 Room:Pacific Salon 6 Chair: Fabian Bastin, Universite de Chair: Robert M. Freund, Massachusetts Chair: Pontus Giselsson, Stanford Montreal, Canada Institute of Technology, USA University, USA 4:30-4:45 Multiple Decomposition 4:30-4:45 A Multilevel Approach for 4:30-4:45 Moreau-Yosida and Cutting-Plane Generations L-1 Regularized Convex Optimization Regularization in Shape Optimization for Optimizing Allocation and with Application to Covariance with Geometric Constraints Scheduling with a Joint Chance Selection Moritz Keuthen and Michael Ulbrich, Constraint Eran Treister and Irad Yavneh, Technion - Technische Universität München, Yan Deng and Siqian Shen, University of Israel Institute of Technology, Israel Germany Michigan, USA 4:50-5:05 Numerical Optimization 4:50-5:05 Optimal Actuator and 4:50-5:05 Application of Variance of Eigenvalues of Hermitian Matrix- Sensor Placement for Dynamical Reduction to Sequential Sampling in Valued Functions Systems Stochastic Programming Emre Mengi, Alper Yildirim, and Mustafa Carsten Schäfer and Stefan Ulbrich, Rebecca Stockbridge, Wayne State Kilic, Koc University, Turkey Technische Universität Darmstadt, Germany University, USA; Guzin Bayraksn, Ohio 5:10-5:25 A Feasible Second Order State University, USA Bundle Algorithm for Nonsmooth, 5:10-5:25 Sensitivity-Based Multistep 5:10-5:25 A Modified Sample Nonconvex Optimization Problems Feedback MPC: Algorithm Design Approximation Method for Chance Hermann Schichl, University of Vienna, and Hardware Implementation Constrained Problems Austria; Hannes Fendl, Bank Austria, Vryan Gil Palma, Universität Bayreuth, Jianqiang Cheng, Celine Gicquel, and Abdel Austria Germany; Andrea Suardi and Eric C. Kerrigan, Imperial College London, Lisser, Universite de Paris-Sud, France 5:30-5:45 Design Optimization United Kingdom; Lars Gruene, University 5:30-5:45 Uncertainty Quantification with the Consider-Then-Choose of Bayreuth, Germany and Optimization Using a Bayesian Behavioral Model Framework Minhua Long, W. Ross Morrow, and Erin 5:30-5:45 Intelligent Optimizing Chen Liang, Vanderbilt University, USA MacDonald, Iowa State University, USA Controller: Development of a Fast Optimal Control Algorithm 5:50-6:05 A New Progressive Hedging 5:50-6:05 Finite Hyperplane Traversal Jinkun Lee and Vittal Prabhu, Pennsylvania Algorithm for Linear Stochastic Algorithms for 1-Dimensional L1pTV State University, USA Optimization Problems Minimization for 0 ≥ p ≥ 1 Shohre Zehtabian and Fabian Bastin, Heather A. Moon, St. Mary’s College of 5:50-6:05 Sliding Mode Controller Universite de Montreal, Canada Maryland, USA Design of Linear Interval Systems Via An Optimised Reduced Order Model 6:10-6:25 Optimal Scenario Set 6:10-6:25 First-Order Methods Yield Raja Rao Guntu, Anil Neerukonda Institute Partitioning for Multistage Stochastic New Analysis and Results for Boosting of Technology & Sciences, India; Programming with the Progressive Methods in Statistics/Machine Narasimhulu Tammineni, Avanthi Institute Hedging Algorithm Learning of Engineering & Technology, India Fabian Bastin, Universite de Montreal, Robert M. Freund and Paul Grigas, Canada; Pierre-Luc Carpentier, École Massachusetts Institute of Technology, 6:10-6:25 Improving Fast Dual Polytechnique de Montréal, Canada; USA; Rahul Mazumder, Columbia Gradient Methods for Repetitive Michel Gendreau, Université de Montréal, University, USA Multi-Parametric Programming Canada Pontus Giselsson, Stanford University, USA 2014 SIAM Conference on Optimization 51

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 CP21 CP22 CP23 Optimal Control IV Nonlinear Optimization II Games and Applications 4:30 PM-6:30 PM 4:30 PM-6:30 PM 4:30 PM-6:30 PM Room:Pacific Salon 7 Room:Sunrise Room:Sunset Chair: Takaaki Aoki, Kyoto University, Chair: Alfredo N. Iusem, IMPA, Rio de Chair: Philipp Renner, Stanford University, Japan Janeiro, Brazil USA 4:30-4:45 A Dual Approach on 4:30-4:45 The Block Coordinate 4:30-4:45 Strong Formulations for Solving Linear-Quadratic Control Descent Method of Multipliers Stackelberg Security Games Problems with Bang-Bang Solutions Mingyi Hong, University of Minnesota, USA Fernando Ordonez, Universidad de Chile, Christopher Schneider and Walter Alt, 4:50-5:05 A Unified Framework for Chile Friedrich Schiller Universität Jena, Subgradient Algorithms Minimizing 4:50-5:05 Multi-Agent Network Germany; Yalcin Kaya, University of Strongly Convex Functions Interdiction Games South Australia, Australia Masaru Ito and Mituhiro Fukuda, Tokyo Harikrishnan Sreekumaran and Andrew L. 4:50-5:05 Improved Error Estimates Institute of Technology, Japan Liu, Purdue University, USA for Discrete Regularization of Linear- 5:10-5:25 Optimal Subgradient 5:10-5:25 Using Optimization Quadratic Control Problems with Algorithms for Large-Scale Structured Methods for Efficient Fatigue Criterion Bang-Bang Solutions Convex Optimization Evaluation Martin Seydenschwanz, Friedrich Schiller Masoud Ahookhosh and Arnold Neumaier, Henrik Svärd, KTH Royal Institute of Universität Jena, Germany University of Vienna, Austria Technology, Sweden 5:10-5:25 Characterization of Optimal 5:30-5:45 A New Algorithm for Least 5:30-5:45 Nonlinear Programming Boundaries in Reversible Investment Square Problems Based on the in Runtime Procedure of Guidance, Problems Primal Dual Augmented Lagrangian Navigation and Control of Salvatore Federico, University of Milan, Approach Autonomous Systems Italy Wenwen Zhou and Joshua Griffin, SAS Masataka Nishi, Hitachi Ltd., Japan 5:30-5:45 Solution of Optimal Control Institute, Inc., USA 5:50-6:05 A Three-objective Problems by Hybrid Functions 5:50-6:05 A Primal-Dual Augmented Mathematical Model for One- Somayeh Mashayekhi, Mississippi State Lagrangian Method for Equality dimensional Cutting and Assortment University, USA Constrained Minimization Problems 5:50-6:05 Numerical Methods and Riadh Omheni and Paul Armand, University Nergiz Kasimbeyli, Anadolu University, Computational Results for Solving of Limoges, France Turkey Hierarchical Dynamic Optimization 6:10-6:25 The Exact Penalty Map 6:10-6:25 Studying Two Stage Games Problems for Nonsmooth and Nonconvex by Polynomial Programming Kathrin Hatz, Johannes P. Schlöder, Optimization Philipp Renner, Stanford University, USA and Hans Georg Bock, University of Alfredo N. Iusem, IMPA, Rio de Janeiro, Heidelberg, Germany Brazil; Regina Burachik, University of 6:10-6:25 Some Mathematical Southern Australia, Australia; Jefferson Properties of the Dynamically Melo, Universidade Federal de Goias, Inconsistent Bellman Equation: Brazil A Note on the Two-Sided Altruism Dynamics Takaaki Aoki, Kyoto University, Japan 52 2014 SIAM Conference on Optimization

Wednesday, May 21 Wednesday, May 21 Wednesday, May 21 CP24 CP25 CP26 Applications in Convex Optimization: Theory and Algorithms Transportation and Algorithms and Applications 4:30 PM-6:30 PM Healthcare 4:30 PM-6:30 PM Room:Royal Palm 2 4:30 PM-6:30 PM Room:Royal Palm 1 Chair: Sheldon H. Jacobson, University of Room:Towne Chair: Stephen A. Vavasis, University of Illinois at Urbana-Champaign and National Waterloo, Canada Science Foundation, USA Chair: Marleen Balvert, Tilburg University, The Netherlands 4:30-4:45 On Estimating Resolution 4:30-4:45 Primal and Dual Given Sparsity and Non-Negativity Approximation Algorithms for Convex 4:30-4:45 An Re-Optimization Keith Dillon and Yeshaiahu Fainman, Vector Optimization Problems Approach for the Train Dispatching University of California, San Diego, USA Firdevs Ulus and Birgit Rudloff, Princeton Problem (tdp) University, USA; Andreas Lohne, Martin- Boris Grimm, Torsten Klug, and Thomas 4:50-5:05 Quadratically Constrained Luther-Universität, Germany Schlechte, Zuse Institute Berlin, Germany Quadratic Programs with On/off Constraints and Applications in Signal 4:50-5:05 Regional Tests for the 4:50-5:05 Optimizing Large Scale Processing Existence of Transition States Concrete Delivery Problems Anne Philipp and Stefan Ulbrich, Technische Dimitrios Nerantzis and Claire Adjiman, Mojtaba Maghrebi, Travis Waller, and Universität Darmstadt, Germany Imperial College London, United Kingdom Claude Sammut, University of New South Wales, Australia 5:10-5:25 The Convex Hull of Graphs 5:10-5:25 Super Linear Convergence of Polynomial Functions in Inexact Restoration Methods 5:10-5:25 A Novel Framework Wei Huang and Raymond Hemmecke, Luis Felipe Bueno, UNICAMP, Brazil; Jose for Beam Angle Optimization in Technische Universität München, Mario Martinez, IMECC-UNICAMP, Radiation Therapy Germany; Christopher T. Ryan, University Brazil Hamed Yarmand, Massachusetts General of Chicago, USA Hospital and Harvard Medical School, 5:30-5:45 Preconditioned Gradient USA; David Craft, Massachusetts General 5:30-5:45 Some Insights from the Methods Based on Sobolev Metrics Hospital, USA Stable Compressive Principal Parimah Kazemi, Beloit College, USA Component Pursuit 5:30-5:45 Optimization of 5:50-6:05 Approximating the Minimum Mituhiro Fukuda and Junki Kobayashi, Radiotherapy Dose-Time Hub Cover Problem on Planar Graphs Tokyo Institute of Technology, Japan Fractionation of Proneural and Its Application Glioblastoma with Consideration of 5:50-6:05 A Tight Iteration-complexity to Query Optimization Biological Effective Dose for Early Bound for IPM via Redundant Klee- Belma Yelbay, Ilker S. Birbil, and Kerem and Late Responding Tissues Minty Cubes Bulbul, Sabanci University, Turkey; Hasan Hamidreza Badri and Kevin Leder, Murat Mut and Tamas Terlaky, Lehigh Jamil, University of Idaho, USA University of Minnesota, USA University, USA 6:10-6:25 The Balance Optimization 5:50-6:05 Stochastic Patient 6:10-6:25 A Fast Quasi-Newton Subset Selection (BOSS) Model for Assignment Models for Balanced Proximal Gradient Algorithm for Basis Causal Inference Healthcare in Patient Centered Pursuit Sheldon H. Jacobson, University of Illinois Medical Home Stephen A. Vavasis and Sahar Karimi, at Urbana-Champaign and National Issac Shams, Saeede Ajorlou, and Kai Yang, University of Waterloo, Canada Science Foundation, USA; Jason Sauppe, Wayne State University, USA University of Illinois, USA; Edward Sewell, Southern Illinois University, 6:10-6:25 Robust Optimization Edwardsville, USA Reduces the Risks Occurring from Delineation Uncertainties in HDR Brachytherapy for Prostate Cancer Marleen Balvert and Dick Den Hertog, Dinner Break Tilburg University, The Netherlands; 6:30 PM-8:00 PM Aswin Hoffmann, Maastro Clinic, The Netherlands Attendees on their own

SIAG/OPT Business Meeting 8:00 PM-8:45 PM Room:Golden Ballroom Complimentary beer and wine will be served. 2014 SIAM Conference on Optimization 53

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS83 MS84 Quadratic Programming Registration Energy Problems - and Sequential Quadratic Part I of III 7:45 AM-5:00 PM Programming Methods 9:30 AM-11:30 AM Room:Golden Foyer 9:30 AM-11:30 AM Room:Pacific Salon 1 Room:Golden Ballroom For Part 2 see MS96 Closing Remarks Sequential quadratic programming Modern energy systems provide (SQP) methods are a popular class of countless challenging issues in 8:10 AM-8:15 AM methods for nonlinear programming optimization. The increasing penetration Room:Golden Ballroom that are based on solving a sequence of of low carbon renewable sources quadratic programming subproblems. of energy introduces uncertainty in They are particularly effective for problems traditionally modeled in a solving a sequence of related problems, deterministic setting. The liberalization IP7 such as those arising in mixed integer of the electricity sector brought the need of designing sound markets, ensuring Large-Scale Optimization nonlinear programming and the optimization of functions subject to capacity investments while properly of Multidisciplinary differential equation constraints. This reflecting strategic interactions. In all Engineering Systems session concerns a number of topics these problems, hedging risk, possibly 8:15 AM-9:00 AM associated with the formulation, analysis in a dynamic manner, is also a concern. and implementation of both QP and This 3-session minisymposium, Room:Golden Ballroom SQP methods. co-organized by A. Philpott, D. Ralph Chair: Andrew R. Conn, IBM T.J. Watson Organizer: Philip E. Gill and C. Sagastizabal, includes speakers Research Center, USA University of California, San Diego, USA from academia and industry on several subjects of active current research in the There is a compelling incentive for 9:30-9:55 Regularized Sequential using optimization to design aerospace Quadratic Programming Methods fields of Optimization and Energy. systems due to large penalties Philip E. Gill, University of California, San Organizer: Claudia A. incurred by their weight. The design Diego, USA; Vyacheslav Kungurtsev, Sagastizabal of these systems is challenging due K.U. Leuven, Belgium; Daniel Robinson, IMPA, Brazil Johns Hopkins University, USA to their complexity. We tackle these Organizer: Andy Philpott challenges by developing a new view 10:00-10:25 Recent Developments in University of Auckland, New Zealand of multidisciplinary systems, and by the SNOPT and NPOPT Packages for combining gradient-based optimization, Sequential Quadratic Programming Organizer: Daniel Ralph University of Cambridge, United Kingdom efficient gradient computation, and Elizabeth Wong and Philip E. Gill, Newton-type methods. Our applications University of California, San Diego, USA; 9:30-9:55 Large Scale 2-Stage Unit- include wing design based on Navier- Michael A. Saunders, Stanford University, Commitment: A Tractable Approach Wim Van Ackooij, EDF, France; Jerome -Stokes aerodynamic models coupled USA Malick, INRIA, France to finite-element structural models, and 10:30-10:55 A Primal-Dual Active- satellite design including trajectory Set Method for Convex Quadratic 10:00-10:25 Fast Algorithms for Two- optimization. The methods used in this Programming Stage Robust Unit Commitment Anders Forsgren, KTH Royal Institute of Problem work are generalized and proposed as a Technology, Sweden; Philip E. Gill and Bo Zeng, Anna Danandeh, and Wei Yuan, new framework for solving large-scale Elizabeth Wong, University of California, University of South Florida, USA optimization problems. San Diego, USA Joaquim Martins 11:00-11:25 Steering Augmented University of Michigan, USA Lagrangian Methods Hao Jiang and Daniel Robinson, Johns Hopkins University, USA; Frank E. Curtis, Coffee Break Lehigh University, USA 9:00 AM-9:30 AM continued on next page Room:Town & Country 54 2014 SIAM Conference on Optimization

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS84 MS85 MS86 Energy Problems - Variational Analysis and Tensor and Optimization Part I of III Monotone Operator Theory: Problems - Part I of III 9:30 AM-11:30 AM Theory and Projection 9:30 AM-11:30 AM Room:Pacific Salon 1 Methods Room:Pacific Salon 3 9:30 AM-11:30 AM For Part 2 see MS98 continued Room:Pacific Salon 2 This is a stream of three sessions on tensor and optimization problems. Variational Analysis and Monotone The most recent advances in this new Operator Theory lie at the heart 10:30-10:55 Improved Mixed Integer interdisciplinary area, like tensor Linear Optimization Formulations for of modern optimization. This decomposition, low rank approximation, Unit Commitment minisymposium focuses on symmetry tensor optimization, principal Miguel F. Anjos, Mathematics and techniques, stability and projection component analysis, will be presented in Industrial Engineering & GERAD, Ecole methods. Polytechnique de Montreal, Canada; the sessions. Organizer: Heinz Bauschke James Ostrowski, University of Tennessee, Organizer: Lek-Heng Lim University of British Columbia, Canada Knoxville, USA; Anthony Vannelli, University of Chicago, USA University of Guelph, Canada 9:30-9:55 The Douglas–Rachford Algorithm for Two Subspaces Organizer: Jiawang Nie 11:00-11:25 Efficient Solution University of California, San Diego, USA of Problems with Probabilistic Heinz Bauschke, University of British Constraints Arising in the Energy Columbia, Canada 9:30-9:55 Eigenvectors of Tensors and Sector 10:00-10:25 Finite Convergence of a Waring Decomposition Luke Oeding, Auburn University, USA Claudia A. Sagastizabal, IMPA, Brazil; Wim Subgradient Projections Algorithm Van Ackooij, EDF, France; Violette Berge, Jia Xu, Heinz Bauschke, and Shawn Wang, 10:00-10:25 Structured Data Fusion ENSTA ParisTech, France; Welington de University of British Columbia, Canada; Laurent Sorber, Marc Van Barel, and Lieven Oliveira, IMPA, Brazil Caifang Wang, Shanghai Maritime De Lathauwer, K.U. Leuven, Belgium University, China 10:30-10:55 On Cones of 10:30-10:55 Full Stability in Nonnegative Quartic Forms Mathematical Programming Bo Jiang, University of Minnesota, USA; Boris Mordukhovich, Wayne State Zhening Li, University of Portsmouth, University, USA; Tran Nghia, University United Kingdom; Shuzhong Zhang, of British Columbia, Canada University of Minnesota, USA 11:00-11:25 Inheritance of Properties 11:00-11:25 Tensor Methods in Control of the Resolvent Average Optimization Sarah Moffat, Heinz Bauschke, and Xianfu Carmeliza Navasca, University of Alabama Wang, University of British Columbia, at Birmingham, USA Canada 2014 SIAM Conference on Optimization 55

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS87 MS89 MS90 Exploiting Quadratic Semidefinite Programming: Coordinate Descent Structure in Mixed Integer New Formulations and Methods: Sparsity, Non- Nonlinear Programs Fundamental Limitations convexity and Applications 9:30 AM-11:30 AM 9:30 AM-11:30 AM - Part I of III Room:Pacific Salon 4 Room:Pacific Salon 6 9:30 AM-11:30 AM Contributors in this minisymposium Semidefinite programming-based Room:Sunrise develop strong relaxations for non- methods have emerged as powerful For Part 2 see MS102 convex sets that contain both quadratic tools in combinatorial and polynomial Coordinate descent methods are inequalities and integer decision optimization, as well as application becoming increasingly popular variables. areas such as control theory and due to their ability to scale to big Organizer: Jeff Linderoth statistics. Many problems, especially data problems. The purpose of this University of Wisconsin, Madison, USA those with rich algebraic structure, symposium is to bring together have been shown to have exact SDP- 9:30-9:55 Strong Convex Nonlinear researchers working on developing Relaxations of the Pooling Problem reformulations. On the other hand, tools novel coordinate descent algorithms, James Luedtke, University of Wisconsin, for giving lower bounds on the size of analyzing their convergence / Madison, USA; Claudia D’Ambrosio, SDP formulations have recently been complexity, and applying the algorithms CNRS, France; Jeff Linderoth, University developed based on a notion of positive to new domains. of Wisconsin, Madison, USA; Andrew semidefinite rank. This minisymposium Organizer: Peter Richtarik Miller, Institut de Mathématiques de highlights both positive results and University of Edinburgh, United Kingdom Bordeaux, France fundamental limits, featuring new SDP Organizer: Zhaosong Lu 10:00-10:25 Convex Quadratic formulations and methods to simplify Simon Fraser University, Canada Programming with Variable Bounds SDP formulations, as well as new Hyemin Jeon and Jeffrey Linderoth, techniques for finding lower bounds on 9:30-9:55 Iteration Complexity University of Wisconsin, Madison, USA; PSD rank. of Feasible Descent Methods for Andrew Miller, UPS, USA Convex Optimization Organizer: James Saunderson Chih-Jen Lin, National Taiwan University, 10:30-10:55 Supermodular Massachusetts Institute of Technology, USA Taiwan Inequalities for Mixed Integer Bilinear Knapsacks 9:30-9:55 Semidefinite 10:00-10:25 Randomized Block Akshay Gupte, Clemson University, USA Representations of the Convex Hull of Coordinate Non-Monotone Gradient Rotation Matrices Method for a Class of Nonlinear 11:00-11:25 Cuts for Quadratic and James Saunderson, Pablo A. Parrilo, and Programming Conic Quadratic Mixed Integer Alan Willsky, Massachusetts Institute of Zhaosong Lu, Simon Fraser University, Programming Technology, USA Canada; Lin Xiao, Microsoft Research, Sina Modaresi, University of Pittsburgh, USA USA; Mustafa Kilinc, Carnegie Mellon 10:00-10:25 Polytopes with Minimal University, USA; Juan Pablo Vielma, Positive Semidefinite Rank 10:30-10:55 Efficient Random Massachusetts Institute of Technology, Richard Robinson, University of Coordinate Descent Algorithms for USA Washington, USA Large-Scale Structured Nonconvex 10:30-10:55 Partial Facial Reduction: Optimization Simplified, Equivalent SDPs via Inner Ion Necoara and Andrei Patrascu, University Approximations of the PSD Cone Politehnica of Bucharest, Romania Frank Permenter and Pablo A. Parrilo, 11:00-11:25 Efficient Coordinate- Massachusetts Institute of Technology, minimization for Orthogonal Matrices USA through Givens Rotations 11:00-11:25 A Criterion for Sums of Uri Shalit, Hebrew University of Jerusalem, Squares on the Hypercube Israel; Gal Chechik, Bar-Ilan University, Grigoriy Blekherman, Georgia Institute Israel of Technology, USA; Joao Gouveia, Universidade de Coimbra, Portugal; James Pfeiffer, University of Washington, USA 56 2014 SIAM Conference on Optimization

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS91 MS92 MS93 Stochastic Optimization Geometric Properties of Online Optimization and for Large-Scale Inverse Convex Optimization Dynamic Networks Problems - Theory and 9:30 AM-11:30 AM 9:30 AM-11:30 AM Applications Room:Towne Room:Royal Palm 1 9:30 AM-11:30 AM Geometric properties of convex sets Online optimization problems involve Room:Sunset have deep connections with structural making decisions with incomplete and algorithmic aspects of convex knowledge about what comes next. In this session, theoretical and applied optimization problems. For example, They arise in diverse areas including aspects of stochastic optimization it is known that ell-1 minimization inventory management, sponsored techniques applicable for large-scale is associated with sparsity. It is also search advertising, and multi-agent problems are covered. In particular, our known that the thickness of a convex dynamic networks. This minisymposium attention is centered at formulations set is closely connected with its the brings together speakers from that are pertinent for inverse problems intrinsic difficulty of finding a point discrete and continuous optimization, and their design. Exploitation of in it. This session will present several as well as dynamic networks, to the underlying structure of this results that shed light into these explore multiple facets of online class of problems is vital in order interesting connections between the optimization. In particular, problems to retain stability and robustness, geometry, properties of solutions, and with resource constraints across time while accounting for scalability and computational complexity of convex are examined and algorithms with performance optimization problems. rigorous performance guarantees under Organizer: Lior Horesh Organizer: Javier Pena uncertainty are discussed. University of Michigan, USA Carnegie Mellon University, USA Organizer: Maryam Fazel Organizer: Eldad Haber 9:30-9:55 A Geometric Theory University of Washington, USA University of British Columbia, Canada of Phase Transitions in Convex 9:30-9:55 A Dynamic Near- 9:30-9:55 NOWPAC - A Provably Optimization Optimal Algorithm for Online Linear Convergent Derivative Free Martin Lotz, University of Manchester, Programming Optimization Algorithm for Nonlinear United Kingdom Yinyu Ye, Stanford University, USA; Shipra Programming 10:00-10:25 On a Four-dimensional Agrawal, Microsoft Research, USA; Florian Augustin and Youssef M. Marzouk, Cone that is Facially Exposed but Not Zizhuo Wang, University of Minnesota, Massachusetts Institute of Technology, Nice Twin Cities, USA USA Vera Roshchina, University of Ballarat, 10:00-10:25 Online Algorithms for Ad 10:00-10:25 Approximation Methods Australia Allocation for Robust Optimization and Optimal 10:30-10:55 A Deterministic Rescaled Nikhil R. Devanur, Microsoft Research, USA Control Problems for Dynamic Perceptron Algorithm Systems with Uncertainties 10:30-10:55 Online Algorithms for Negar S. Soheili and Javier Pena, Carnegie Ekaterina Kostina, Fachbereich Mathematik Node-Weighted Network Design Mellon University, USA und Informatik, Philipps-Universität Debmalya Panigrahi, Duke University, USA Marburg, Germany 11:00-11:25 Preconditioners for 11:00-11:25 Distributed Learning on Systems of Linear Inequalities 10:30-10:55 Optimal Design and Dynamic Networks Javier Pena, Carnegie Mellon University, Model Re-specification Mehran Mesbahi, University of Washington, USA; Vera Roshchina, University of Lior Horesh, University of Michigan, USA; USA Ballarat, Australia; Negar S. Soheili, Misha E. Kilmer and Ning Hao, Tufts Carnegie Mellon University, USA University, USA 11:00-11:25 Improved Bounds on Sample Size for Implicit Matrix Trace Estimators Farbod Roosta-Khorasani and Uri M. Ascher, University of British Columbia, Canada 2014 SIAM Conference on Optimization 57

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS94 IP8 MS95 Recent Developments in A Projection Hierarchy for Quadratics in Mixed-integer Vector Optimization Some NP-hard Optimization Nonlinear Programming 9:30 AM-11:30 AM Problems 2:00 PM-4:00 PM Room:Royal Palm 2 1:00 PM-1:45 PM Room:Golden Ballroom Vector optimization is an indispensable Room:Golden Ballroom This minisymposium will present tool for decision makers in areas like Chair: Miguel Anjos, École Polytechnique new results on several topics related engineering, medicine, economics, de Montréal, Canada to MINLP in particular involving finance and the social sector. The aim There exist several hierarchies of general quadratic objectives, quadratic of the minisymposium is to give insides relaxations which allow to solve constraint, and mixed-integer variables. into recent developments in this field. NP-hard combinatorial optimization Organizer: Daniel Bienstock The talks cover such different areas as problems to optimality. The Lasserre Columbia University, USA robust multi-objective optimization, and Parrilo hierarchies are based on 2:00-2:25 Global Optimization with non-standard ordering structures, new semidefinite optimization and in each Non-Convex Quadratics application fields and novel features new level both the dimension and Marcia HC Fampa, Universidade Federal de for established areas like scalarization the number of constraints increases. Rio de Janeiro, Brazil; Jon Lee, University approaches. As a consequence even the first step of Michigan, USA; Wendel Melo, Organizer: Gabriele Eichfelder up in the hierarchy leads to problems Universidade Federal de Rio de Janeiro, Brazil Technische Universitaet Ilmenau, Germany which are extremely hard to solve 9:30-9:55 Linear Scalarizations for computationally for nontrivial problem 2:30-2:55 On the Separation of Vector Optimization Problems with a sizes. In contrast, we present a hierarchy Split Inequalities for Non-Convex Variable Ordering Structure where the dimension stays fixed and Quadratic Integer Programming Gabriele Eichfelder and Tobias Gerlach, Emiliano Traversi, Université Paris 13, only the number of constraints grows France Technische Universität Ilmenau, Germany exponentially. It applies to problems 10:00-10:25 Robust Multiobjective where the projection of the feasible set 3:00-3:25 Extended Formulations Optimization to subproblems has a ‘simple’ structure. for Quadratic Mixed Integer Margaret Wiecek, Clemson University, USA Programming We consider this new hierarchy for Juan Pablo Vielma, Massachusetts Institute 10:30-10:55 Variational Analysis in Max-Cut, Stable-Set and Graph- of Technology, USA Psychological Modeling Coloring. We look at some theoretical Bao Q. Truong, Northern Michigan properties, discuss practical issues 3:30-3:55 Finding Low-Rank Solutions University, USA; Boris Mordukhovich, to Qcqps and provide computational results, Daniel Bienstock, Columbia University, USA Wayne State University, USA; Antoine comparing the new bounds with the Soubeyran, Aix-Marseille Université, current state-of-the-art. France Franz Rendl 11:00-11:25 Comparison of Different Universitat Klagenfurt, Austria Scalarization Methods in Multi- objective Optimization Refail Kasimbeyli, Zehra Kamisli Ozturk, Nergiz Kasimbeyli, Gulcin Dinc Yalcin, Intermission and Banu Icmen, Anadolu University, Turkey 1:45 PM-2:00 PM

Lunch Break 11:30 AM-1:00 PM Attendees on their own 58 2014 SIAM Conference on Optimization

Thursday, May 22 2:30-2:55 Computing Closed Loop Thursday, May 22 Equilibria of Electricity Capacity MS96 Expansion Models MS97 Sonja Wogrin, Comillas Pontifical Energy Problems - University, Spain Variational Analysis and Part II of III 3:00-3:25 Risk in Capacity Energy Monotone Operator 2:00 PM-4:00 PM Equilibria Theory: Various First Order Daniel Ralph, University of Cambridge, Algorithms Room:Pacific Salon 1 United Kingdom; Andreas Ehrenmann and For Part 1 see MS84 Gauthier de Maere, GDF-SUEZ, France; 2:00 PM-4:00 PM Yves Smeers, Universite Catholique de For Part 3 see MS108 Room:Pacific Salon 2 Modern energy systems provide Louvain, Belgium We look at various algorithms and countless challenging issues in 3:30-3:55 Optimization (and Maybe their analysis. Specifically, first order optimization. The increasing penetration Game Theory) of Demand Response algorithms for large nonsmooth of low carbon renewable sources Golbon Zakeri and Geoff Pritchard, nonconvex problems, ε-subgradient of energy introduces uncertainty in University of Auckland, New Zealand methods for bilevel convex optimization, problems traditionally modeled in a and supporting halfspaces and quadratic deterministic setting. The liberalization programming algorithms for convex of the electricity sector brought the need feasibility problems. of designing sound markets, ensuring Organizer: C.H. Jeffrey Pang capacity investments while properly National University of Singapore, Singapore reflecting strategic interactions. In all 2:00-2:25 A First Order Algorithm for these problems, hedging risk, possibly a Class of Nonconvex-nonsmooth in a dynamic manner, is also a concern. Minimization This 3-session minisymposium, Shoham Sabach, Universität Goettingen, co-organized by A. Philpott, D. Ralph Germany and C. Sagastizabal, includes speakers 2:30-2:55 ε-Subgradient Methods for from academia and industry on several Bilevel Convex Optimization subjects of active current research in the Elias Salomão S. Helou Neto, Universidade fields of Optimization and Energy. de Sao Paulo, Brazil Organizer: Golbon Zakeri 3:00-3:25 An Algorithm for Convex University of Auckland, New Zealand Feasibility Problems Using Supporting Halfspaces and Dual Quadratic Organizer: Andy Philpott Programming University of Auckland, New Zealand C.H. Jeffrey Pang, National University of Organizer: Daniel Ralph Singapore, Singapore University of Cambridge, United Kingdom 3:30-3:55 Fast Variational Methods for Organizer: Claudia A. Matrix Completion and Robust PCA Sagastizabal Aleksandr Aravkin, IBM T.J. Watson IMPA, Brazil Research Center, USA 2:00-2:25 Numerical Methods for Stochastic Multistage Optimization Problems Applied to the European Electricity Market Nicolas Grebille, Ecole Polytechnique, France

continued in next column 2014 SIAM Conference on Optimization 59

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS98 MS99 MS100 Tensor and Optimization Modern Sequential Recent Progress on Problems - Part II of III Quadratic Algorithms for Theoretical Aspects of 2:00 PM-4:00 PM Nonlinear Optimization Conic Programming Room:Pacific Salon 3 2:00 PM-4:00 PM 2:00 PM-4:00 PM For Part 1 see MS86 Room:Pacific Salon 4 Room:Pacific Salon 5 For Part 3 see MS110 This minisymposium highlights Due to the recent explosion of This is a stream of three sessions on recent developments in the design of application of semidefinite programming tensor and optimization problems. algorithms for solving constrained (SDP) and second-order cone The most recent advances in this new nonlinear optimization problems. In programming (SOCP), now conic interdisciplinary area, like tensor particular, the focus is on methods of programming, a general extension of decomposition, low rank approximation, the sequential quadratic optimization SDP and SOCP, is recognized as one tensor optimization, principal (SQP) variety. Method of this type of the most important research subjects component analysis, will be presented in have been the subject of research for in the field of optimization. This the sessions. decades, but this minisymposium minisymposium will focus on theoretical Organizer: Lek-Heng Lim emphasize the “new frontiers” of SQP aspects of conic programming, University of Chicago, USA methods, such as the incorporation of such as new duality results, new Organizer: Jiawang Nie inexactness in the subproblem solves decompositions, and others. These talks University of California, San Diego, USA and the application of SQP to optimize will provide fresh viewpoints on the 2:00-2:25 On Symmetric Rank and nonsmooth functions. The discussions theory of conic programming. Approximation of Symmetric Tensors in the minisymposium will emphasize Organizer: Masakazu Muramatsu Shmuel Friedland, University of Illinois, the strengths of active-set methods for University of Electro-Communications, Chicago, USA constrained nonlinear optimization. Japan 2:30-2:55 Algorithms for Organizer: Frank E. Curtis Organizer: Akiko Yoshise Decomposition Lehigh University, USA University of Tsukuba, Japan Elias Tsigaridas, INRIA Paris- Organizer: Andreas Waechter Rocquencourt, France 2:00-2:25 An LP-based Algorithm to Northwestern University, USA Test Copositivity 3:00-3:25 Some Extensions of 2:00-2:25 An Inexact Trust-Region Akihiro Tanaka and Akiko Yoshise, Theorems of the Alternative Algorithm for Nonlinear Programming University of Tsukuba, Japan Shenglong Hu, Hong Kong Polytechnic Problems with Expensive General University, China 2:30-2:55 On the rolls of optimality Constraints conditions for polynomial 3:30-3:55 Semidefinite Relaxations for Andrea Walther, Universität Paderborn, optimization Best Rank-1 Tensor Approximations Germany; Larry Biegler, Carnegie Mellon Yoshiyuki Sekiguchi, Tokyo University of Jiawang Nie and Li Wang, University of University, USA Marine Science and Technology, Japan California, San Diego, USA 2:30-2:55 A BFGS-Based SQP 3:00-3:25 Gauge optimization, duality Method for Constrained Nonsmooth, and applications Nonconvex Optimization Michael P. Friedlander, Macedo Ives, and Frank E. Curtis, Lehigh University, USA; Ting Kei Pong, University of British Tim Mitchell and Michael L. Overton, Columbia, Canada Courant Institute of Mathematical 3:30-3:55 A Geometrical Analysis Sciences, New York University, USA of Weak Infeasibility in Semidefinite 3:00-3:25 A Two-phase SQO Method Programming and Related Issues for Solving Sequences of NLPs Bruno F. Lourenço, Tokyo Institute of Jose Luis Morales, ITAM, Mexico Technology, Japan; Masakazu Muramatsu, 3:30-3:55 An Inexact Sequential University of Electro-Communications, Quadratic Optimization Method for Japan; Takashi Tsuchiya, National Nonlinear Optimization Graduate Institute for Policy Studies, Frank E. Curtis, Lehigh University, USA; Japan Travis Johnson, University of Washington, USA; Daniel Robinson, Johns Hopkins University, USA; Andreas Waechter, Northwestern University, USA 60 2014 SIAM Conference on Optimization

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS101 MS102 MS103 Recent Development in Coordinate Descent Efficient Optimization Tools for Building Conic Methods: Sparsity, Non- Techniques for Chaotic and Optimization Models convexity and Applications Deterministic Large Scale 2:00 PM-4:00 PM - Part II of III Time-Dependent Problems - Room:Pacific Salon 6 2:00 PM-4:00 PM Part I of II Conic optimization which includes Room:Sunrise 2:00 PM-4:00 PM linear, quadratic and semi-definite conic For Part 1 see MS90 Room:Sunset optimization as special cases has gained For Part 3 see MS114 For Part 2 see MS115 of lot of attention the last 10 years. The Coordinate descent methods are Time dependent PDE-constrained purpose of this minisymposium is to becoming increasingly popular problems require optimization algorithms present recent developments in tools for due to their ability to scale to big to handle the respective problem size. build conic optimization models. data problems. The purpose of this Many strategies such as the one-shot Organizer: Erling D. Andersen symposium is to bring together approach are not directly applicable to MOSEK ApS, Denmark researchers working on developing such problems. This is especially true in novel coordinate descent algorithms, 2:00-2:25 The Cvx Modeling applications whose trajectory is sensitive analyzing their convergence / Framework: New Development to small perturbations. Such problems are Michael C. Grant, California Institute of complexity, and applying the algorithms prototypical for optimal control in fluid Technology, USA to new domains. dynamics, often also having the added 2:30-2:55 Quadratic Cone Modeling Organizer: Peter Richtarik difficulty of involving shape optimization. Language University of Edinburgh, United Kingdom Because the handling of shapes, time- Eric Chu, Stanford University, USA Organizer: Zhaosong Lu dependence and chaotic dynamics are 3:00-3:25 Recent Devlopments in Simon Fraser University, Canada essential for many such problems, this Picos 2:00-2:25 Coordinate Descent minisymposium is meant to bring together Guillaume Sagnol, Zuse Institute Berlin, Type Methods for Solving Sparsity new ideas from each of those fields, Germany Constrained Problems constructing new algorithms for shape 3:30-3:55 Building Large-scale Conic Amir Beck, Technion - Israel Institute of optimization of time-dependent and Optimization Models using MOSEK Technology, Israel chaotic problems. Fusion 2:30-2:55 Coordinate descent Organizer: Stephan Schmidt Erling D. Andersen, MOSEK ApS, Denmark Imperial College London, United Kingdom methods for l0-regularized optimization problems Organizer: Qiqi Wang Andrei Patrascu and Ion Necoara, University Massachusetts Institute of Technology, USA Politehnica of Bucharest, Romania 2:00-2:25 Optimization in Chaos 3:00-3:25 Efficient Randomized Qiqi Wang and Patrick Blonigan, Coordinate Descent for Large Scale Massachusetts Institute of Technology, USA Sparse Optimization 2:30-2:55 Efficient Estimation of Xiaocheng Tang and Katya Scheinberg, Selecting and Locating Actuators and Lehigh University, USA Sensors for Controlling Compressible, 3:30-3:55 GAMSEL: A Penalized Viscous Flows Likelihood Approach to Model Daniel J. Bodony and Mahesh Natarajan, Selection for Generalized Additive University of Illinois at Urbana-Champaign, Models USA Alexandra Chouldechova and Trevor Hastie, 3:00-3:25 New Trends in Aerodynamic Stanford University, USA Shape Optimization using the Continuous Adjoint Method Francisco Palacios, Thomas Economon, and Juan J. Alonso, Stanford University, USA 3:30-3:55 Shape Analysis for Maxwell’s Equations Maria Schütte, Universität Paderborn, Germany; Stephan Schmidt, Imperial College London, United Kingdom; Andrea Walther, Universität Paderborn, Germany 2014 SIAM Conference on Optimization 61

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS104 MS105 MS106 Infinite Dimensional Fast Algorithms for Large- Nonsmooth PDE-constrained Optimization: Theory and Scale Inverse Problems and optimization - Part I of II Applications - Part I of II Applications 2:00 PM-4:00 PM 2:00 PM-4:00 PM 2:00 PM-4:00 PM Room:Royal Palm 2 Room:Towne Room:Royal Palm 1 For Part 2 see MS118 For Part 2 see MS116 Many real-world problems reduce to In recent years, the field of PDE- We explore recent advancements in optimization problems that are solved by constrained optimization has provided infinite and semi-infinite optimization, iterative algorithms. This minisymposium a rich class of nonsmooth optimization both theory and applications. Infinite focuses on recently developed efficient problems, where the nonsmooth structure dimensional optimization can be used algorithms for solving large scale is given either by cost terms enforcing to study Markov decision processes, inverse problems that arise in image a desired structure of the solution (such stochastic programming, continuous reconstruction, sparse optimization, and as sparse or bang-bang controls) or by location and routing problems, and compressed sensing. variational inequality constraints, e.g., control problems. Current research representing pointwise constraints. Organizer: William Hager Such problems arise in a wide variety directions have uncovered new insights University of Florida, USA into duality theory using projection and of applications from fluid flow and Riesz space techniques, advancements Organizer: Maryam Yashtini elastoplasticity to image denoising, University of Florida, USA in algorithms for infinite dimensional microelectromechanical systems. The linear programs, and novel applications. Organizer: Hongchao Zhang main challenges concerning these Issues such as identifying the Louisiana State University, USA problems involve the development of presence of nonzero duality gaps and 2:00-2:25 Adaptive BOSVS Algorithm analytical techniques to derive sharp showing finiteness of algorithms have for Ill-Conditioned Linear Inversion optimality conditions, and the design of traditionally been important challenges with Application to Partially Parallel efficient algorithms for their numerical in this field, and this minisymposium Imaging solution. describes work which may help Maryam Yashtini and William Hager, University of Florida, USA; Hongchao Organizer: Christian Clason overcome these challenges. Zhang, Louisiana State University, USA Universität Graz, Austria Organizer: Matt Stern 2:30-2:55 Analysis and Design of Fast Organizer: Juan Carlos De los University of Chicago, USA Graph Based Algorithms for High Reyes Organizer: Chris Ryan Dimensional Data Escuela Politécnica Nacional, Ecuador University of Chicago, USA Andrea L. Bertozzi, University of California, 2:00-2:25 Optimal Control of Free Los Angeles, USA 2:00-2:25 The Slater Conundrum: Boundary Problems with Surface Duality and Pricing in Infinite 3:00-3:25 GRock: Greedy Coordinate- Tension Effects Dimensional Optimization Block Descent Method for Sparse Harbir Antil, George Mason University, USA Matt Stern, Chris Ryan, and Kipp Martin, Optimization 2:30-2:55 Fast and Sparse Noise University of Chicago, USA Zhimin Peng, Ming Yan, and Wotao Yin, Learning via Nonsmooth PDE- University of California, Los Angeles, USA 2:30-2:55 Global Optimization in constrained Optimization Infinite Dimensional Hilbert Spaces 3:30-3:55 Fast Algorithms for Luca Calatroni, University of Cambridge, Boris Houska, Shanghai Jiaotong University, Multichannel Compressed Sensing United Kingdom China; Benoit Chachuat, Imperial College with Forest Sparsity 3:00-3:25 A Primal-dual Active Set London, United Kingdom Junzhou Huang, University of Texas at Algorithm for a Class of Nonconvex Arlington, USA 3:00-3:25 Equitable Division of a Sparsity Optimization Geographic Region Yuling Jiao, Wuhan University, China; Bangti John Gunnar Carlsson, University of Jin, University of California, Riverside, Minnesota, USA USA; Xiliang Lu, Wuhan University, China 3:30-3:55 Structured Convex Infinite- 3:30-3:55 Multibang Control of Elliptic dimensional Optimization with Equations Double Smoothing and Chebfun Christian Clason and Karl Kunisch, Olivier Devolder, Francois Glineur, and Universität Graz, Austria Yurii Nesterov, Université Catholique de Louvain, Belgium Coffee Break 4:00 PM-4:30 PM Room:Town & Country 62 2014 SIAM Conference on Optimization

Thursday, May 22 Thursday, May 22 5:30-5:55 Islanding of Power Networks Andreas Grothey and Ken McKinnon, MS107 MS108 University of Edinburgh, United Kingdom; Paul Trodden, University of New Directions in Large- Energy Problems - Sheffield, United Kingdom; Waqquas Scale and Sparse Conic Part III of III Bukhsh, University of Edinburgh, United Kingdom Programming 4:30 PM-6:30 PM 6:00-6:25 Topology Control for 4:30 PM-6:30 PM Room:Pacific Salon 1 Economic Dispatch and Reliability in Room:Golden Ballroom For Part 2 see MS96 Electric Power Systems Shmuel Oren, University of California, A variety of recent large-scale data Modern energy systems provide countless challenging issues in Berkeley, USA; Kory Hedman, Arizona analysis problems have sparked the State University, USA interest of the optimization community optimization. The increasing penetration in conic optimization based algorithms of low carbon renewable sources that exploit various types of structure of energy introduces uncertainty in arising from notions of “simplicity” problems traditionally modeled in a in large-scale optimization problems. deterministic setting. The liberalization In this session, some of the leading of the electricity sector brought the need researchers in the area will present new of designing sound markets, ensuring and exciting directions the field is taking. capacity investments while properly These include a unified framework for reflecting strategic interactions. In all geometric and semidefinite programming, these problems, hedging risk, possibly techniques for reducing the size of in a dynamic manner, is also a concern. large-scale linear and convex quadratic This 3-session minisymposium, programs arising in machine learning, co-organized by A. Philpott, D. Ralph improvements on the nuclear norm and C. Sagastizabal, includes speakers relaxation for low-rank optimization, and from academia and industry on several generalizations of sparse optimization to subjects of active current research in the settings where the signal to be recovered fields of Optimization and Energy. lives in a continuous domain. Organizer: Shmuel Oren University of California, Berkeley, USA Organizer: Amir Ali Ahmadi IBM Research, USA Organizer: Andy Philpott University of Auckland, New Zealand 4:30-4:55 Conic Geometric Programming Organizer: Daniel Ralph Venkat Chandrasekaran, California Institute University of Cambridge, United Kingdom of Technology, USA; Parikshit Shah, Organizer: Claudia A. Philips Research North America, USA Sagastizabal 5:00-5:25 Sketching the Data Matrix of IMPA, Brazil Large Linear and Convex Quadratic 4:30-4:55 A Strong Relaxation for the Programs Alternating Current Optimal Power Laurent El Ghaoui, University of California, Flow (acopf) Problem Berkeley, USA; Riadh Zorgati, EDF, France Chen Chen, Alper Atamturk, and Shmuel 5:30-5:55 Multi-Stage Convex Oren, University of California, Berkeley, Relaxation Approach for Low-Rank USA Structured Psd Optimization Problems 5:00-5:25 Real-Time Pricing in Smart Defeng Sun, National University of Singapore, Grid: An ADP Approach Republic of Singapore Andrew L. Liu, Purdue University, USA; 6:00-6:25 Convex Programming for Jingjie Xiao, Reddwerks Corporation, Continuous Sparse Optimization USA Ben Recht, University of California, Berkeley, USA

continued in next column 2014 SIAM Conference on Optimization 63

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS109 MS110 MS111 Variational Analysis and Tensor and Optimization Relaxations and solution Optimization Problems - Part III of III techniques for MINLP 4:30 PM-6:30 PM 4:30 PM-6:30 PM 4:30 PM-6:30 PM Room:Pacific Salon 2 Room:Pacific Salon 3 Room:Pacific Salon 4 Originating from classical convex For Part 2 see MS98 This session will address problems in analysis and the calculus of variations, This is a stream of three sessions on mixed integer nonlinear optimization “variational analysis” has become a tensor and optimization problems. (MINLP). These problems arise in powerful toolkit comprising not only The most recent advances in this new many application areas and require convex but also nonsmooth and set- interdisciplinary area, like tensor branch-and-bound (BB) based methods valued analysis among other things. decomposition, low rank approximation, to converge to global optimum. Tight Applications of and motivation for tensor optimization, principal polyhedral and convex relaxations variational analysis are areas such component analysis, will be presented in play a crucial role in accelerating as the calculus of variations, control the sessions. performance of BB. New results on theory, equilibrium problems, monotone Organizer: Lek-Heng Lim convex relaxations and enhancements in operator theory, variational inequalities, University of Chicago, USA BB for nonconvex QCQP and general and most importantly optimization. This MINLP will be presented. Sometimes Organizer: Jiawang Nie minisymposium aims at highlighting University of California, San Diego, USA MINLPs also have indicator variables this intimate relationship between to turn on/off some of the variables. 4:30-4:55 On Low-Complexity Tensor optimization and variational analysis. Stronger relaxations, obtained by taking Approximation and Optimization Organizer: Tim Hoheisel Shuzhong Zhang and Bo Jiang, University into account the role played by these University of Würzburg, Germany of Minnesota, USA; Shiqian Ma, Chinese indicators, will be derived. 4:30-4:55 Epi-Convergent Smoothing University of Hong Kong, Hong Kong Organizer: Akshay Gupte with Applications to Convex 5:00-5:25 On Generic Nonexistence Clemson University, USA Composite Functions of the Schmidt-Eckart-Young 4:30-4:55 Decomposition Techniques Tim Hoheisel, University of Würzburg, Decomposition for Tensors in Global Optimization Germany; James V. Burke, University of Nick Vannieuwenhoven, Johannes Nicaise, Mohit Tawarmalani, Purdue University, Washington, USA Raf Vandebril, and Karl Meerbergen, USA; Jean-Philippe Richard, University of 5:00-5:25 Some Variational Properties Katholieke Universiteit Leuven, Belgium Florida, USA of Regularization Value Functions 5:30-5:55 The Rank Decomposition 5:00-5:25 Recent Advances in James V. Burke, University of Washington, and the Symmetric Rank Branch-and-Bound Methods for USA; Aleksandr Aravkin, IBM T.J. Decomposition of a Symmetric Tensor MINLP Watson Research Center, USA; Michael Xinzhen Zhang and Zhenghai Huang, Tianjin Pietro Belotti, FICO, United Kingdom P. Friedlander, University of British University, China; Liqun Qi, Hong Kong 5:30-5:55 Computing Strong Bounds Columbia, Canada Polytechnic University, China for Convex Optimization with 5:30-5:55 Optimization and Intrinsic 6:00-6:25 Title Not Available Indicator Variables Geometry Bart Vandereycken, Princeton University, Hongbo Dong, Washington State University, Dmitriy Drusvyatskiy, University of Waterloo, USA USA Canada; Adrian Lewis, Cornell University, USA; Alexander Ioffe, Technion - Israel 6:00-6:25 A Hybrid Lp/nlp Paradigm Institute of Technology, Israel; Aris for Global Optimization Daniildis, University of Barcelona, Spain Aida Khajavirad, IBM T.J. Watson Research Center, USA; Nick Sahinidis, Carnegie 6:00-6:25 Epi-splines and Exponential Mellon University, USA Epi-splines: Pliable Approximation Tools Roger J.-B. Wets, University of California, Davis, USA; Johannes O. Royset, Naval Postgraduate School, USA 64 2014 SIAM Conference on Optimization

Thursday, May 22 Thursday, May 22 Thursday, May 22 MS112 MS113 MS114 Optimal Routing on Exploiting Problem Structure Coordinate Descent Stochastic Networks with in First-order Convex Methods: Sparsity, Non- Transportation Applications Optimization convexity and Applications 4:30 PM-6:30 PM 4:30 PM-6:30 PM - Part III of III Room:Pacific Salon 5 Room:Pacific Salon 6 4:30 PM-6:30 PM Online optimization on large-scale Huge-scale optimization methos are Room:Sunrise graphs models of urban networks, increasingly using problem structure to For Part 2 see MS102 motivated by the recent explosion achieve faster convergence rates than Coordinate descent methods are of real-time sensing capabilities, black-box methods. Recent successes becoming increasingly popular is paramount to a variety of urban include smoothing, proximal gradient, due to their ability to scale to big computing applications. In particular, stochastic gradient, and coordinate- data problems. The purpose of this the problem of online decision making descent methods. This minisymposium symposium is to bring together under uncertainty on a graph is one of focuses on important recent advances researchers working on developing the keys for improving the efficiency in this area. Volkan Cevher will novel coordinate descent algorithms, of transportation networks. Several discuss using self-concordance in the analyzing their convergence / uncertainty models can be considered, context of first-order methods, Stephen complexity, and applying the algorithms and the resulting optimization problem Becker will talk about a non-diagonal to new domains. can be formulated as a stochastic yet efficient projected quasi-Newton Organizer: Peter Richtarik programming problem, chance- method, Mehrdad Mahdavi will University of Edinburgh, United Kingdom constrained optimization problem, discuss comibining both stochastic 4:30-4:55 Efficient Accelerated or a robust optimization problem. In and deterministic gradients, and Julien Coordinate Descent Methods and this symposium, we present in-depth Mairal will present a framework that Faster Algorithms for Solving Linear theoretical and numerical results for admits fast rates in a variety of non- Systems these approaches. standard scenarios. Yin Tat Lee and Aaron Sidford, Organizer: Samitha Organizer: Mark Schmidt Massachusetts Institute of Technology, USA Samaranayake Simon Fraser University, Canada University of California, Berkeley, USA 4:30-4:55 Composite Self-Concordant 5:00-5:25 Applications of Coordinate Descent in Combinatorial Prediction Organizer: Sebastien Blandin Minimization Markets IBM Research, Singapore Quoc Tran Dinh, Kyrillidis Anastasios, and Volkan Cevher, École Polytechnique Miroslav Dudik, Microsoft Research, USA 4:30-4:55 Speed-Up Algorithms for the Fédérale de Lausanne, Switzerland 5:30-5:55 A Comparison of PCDM Stochastic on-Time Arrival Problem and DQAM for Big Data Problems Samitha Samaranayake, University of 5:00-5:25 A Quasi-Newton Proximal Rachael Tappenden, Peter Richtarik, and California, Berkeley, USA; Sebastien Splitting Method Burak Buke, University of Edinburgh, Blandin, IBM Research, Singapore Stephen Becker, IBM Research, USA United Kingdom 5:00-5:25 Stochastic Route Planning 5:30-5:55 Mixed Optimization: On the 6:00-6:25 Direct Search Based on in Public Transport Interplay between Deterministic and Probabilistic Descent Kristof Berczi, Alpár Jüttner, and Mátyás Stochastic Optimization Serge Gratton and Clément Royer, Korom, Eötvös Loránd University, Mehrdad Mahdavi, Michigan State ENSEEIHT, Toulouse, France; Luis N. ; Marco Laumanns and University, USA Vicente and Zaikun Zhang, Universidade Tim Nonner, IBM Research-Zurich, 6:00-6:25 Incremental and Stochastic de Coimbra, Portugal Switzerland; Jacint Szabo, Zuse Institute Majorization-Minimization Algorithms Berlin, Germany for Large-Scale Optimization 5:30-5:55 Risk-averse Routing Julien Mairal, INRIA, France Albert Akhriev, Jakub Marecek, and Jia Yuan Yu, IBM Research, Ireland 6:00-6:25 Robust Formulation for Online Decision Making on Graphs Arthur Flajolet, Massachusetts Institute of Technology, USA; Sebastien Blandin, IBM Research, Singapore 2014 SIAM Conference on Optimization 65

Thursday, May 22 Thursday, May 22 5:30-5:55 Efficient Techniques for Optimal Active Flow Control MS115 Nicolas R. Gauger, Stefanie Guenther, MS116 Efficient Optimization Anil Nemili, and Emre Oezkaya, RWTH Infinite Dimensional Techniques for Chaotic and Aachen University, Germany; Qiqi Wang, Optimization: Theory and Massachusetts Institute of Technology, Deterministic Large Scale USA Applications -Part II of II Time-Dependent Problems - 6:00-6:25 Improved Monte-Carlo 4:30 PM-6:30 PM Part II of II Methods for Optimization Problem Room:Towne Bastian Gross, University of Trier, Germany 4:30 PM-6:30 PM For Part 1 see MS104 Room:Sunset We explore recent advancements in infinite and semi-infinite optimization, For Part 1 see MS103 Time dependent PDE-constrained both theory and applications. Infinite problems require optimization dimensional optimization can be used to algorithms to handle the respective study Markov decision processes, stochastic problem size. Many strategies such programming, continuous location and as the one-shot approach are not routing problems, and control problems. directly applicable to such problems. Current research directions have uncovered This is especially true in applications new insights into duality theory using whose trajectory is sensitive to small projection and Riesz space techniques, perturbations. Such problems are advancements in algorithms for infinite prototypical for optimal control in dimensional linear programs, and novel fluid dynamics, often also having the applications. Issues such as identifying added difficulty of involving shape the presence of nonzero duality gaps and optimization. Because the handling of showing finiteness of algorithms have shapes, time-dependence and chaotic traditionally been important challenges dynamics are essential for many such in this field, and this minisymposium problems, this minisymposium is describes work which may help overcome meant to bring together new ideas from these challenges. each of those fields, constructing new Organizer: Matt Stern algorithms for shape optimization of University of Chicago, USA time-dependent and chaotic problems. Organizer: Chris Ryan Organizer: Stephan Schmidt University of Chicago, USA Imperial College London, United Kingdom 4:30-4:55 Projection: A Unified Approach Organizer: Qiqi Wang to Semi-Infinite Linear Programs Massachusetts Institute of Technology, USA Chris Ryan, University of Chicago, USA; Amitabh Basu, Johns Hopkins University, 4:30-4:55 Time-dependent Shape USA; Kipp Martin, University of Chicago, Derivative in Moving Domains – USA Morphing Wings Adjoint Gradient- based Aerodynamic Optimisation 5:00-5:25 On the Sufficiency of Finite Nicusor Popescu and Stephan Schmidt, Support Duals in Semi-infinite Linear Imperial College London, United Programming Kingdom Amitabh Basu, Johns Hopkins University, USA; Kipp Martin and Chris Ryan, University of 5:00-5:25 Gradient Projection Type Chicago, USA Methods for Multi-material Structural Topology Optimization using a Phase 5:30-5:55 Countably Infinite Linear Field Ansatz Programs and Their Application to Luise Blank, Universität Regensburg, Nonstationary MDPs Germany Archis Ghate, University of Washington, USA; Robert Smith, University of Michigan, USA continued in next column 6:00-6:25 A Linear Programming Approach to Markov Decision Processes with Countably Infinite State Space Ilbin Lee, Marina A. Epelman, Edwin Romeijn, and Robert Smith, University of Michigan, USA 66 2014 SIAM Conference on Optimization

Thursday, May 22 Thursday, May 22 5:30-5:55 Boundary Concentrated Finite Elements for Optimal Control MS117 MS118 Problems and Application to Bang- Bang Controls Recent Advances in Robust Nonsmooth PDE- Jan-Eric Wurst and Daniel Wachsmuth, Discrete Choice Modelling constrained optimization - University of Würzburg, Germany; Sven Beuchler and Katharina Hofer, Universität 4:30 PM-6:30 PM Part II of II Bonn, Germany Room:Royal Palm 1 4:30 PM-6:30 PM 6:00-6:25 On the Use of Second Order Discrete choice models are widely used Room:Royal Palm 2 Information for the Numerical Solution of L1-Pde-Constrained Optimization in the analysis of individual choice For Part 1 see MS106 Problems behavior in areas including economics, In recent years, the field of PDE- Juan Carlos De los Reyes, Escuela transportation, and marketing. Despite constrained optimization has provided Politécnica Nacional, Ecuador their broad application, the evaluation a rich class of nonsmooth optimization of choice probabilities is a challenging problems, where the nonsmooth task. The minisymposium will include structure is given either by cost terms novel models for calculating choice enforcing a desired structure of the probabilities in challenging environments solution (such as sparse or bang-bang such as large scale data sets from real controls) or by variational inequality world applications, with missing entries constraints, e.g., representing pointwise and behavioral bias, or route choice constraints. Such problems arise in problems on large scale networks. While a wide variety of applications from some papers focus more on theory and fluid flow and elastoplasticity to image algorithms, others put emphasis on denoising, microelectromechanical the application of choice modelling in systems. The main challenges practice such as for revenue management concerning these problems involve the and product assortment. development of analytical techniques to Organizer: Selin D. Ahipasaoglu derive sharp optimality conditions, and Singapore University of Technology & the design of efficient algorithms for Design, Singapore their numerical solution. 4:30-4:55 On Theoretical and Organizer: Christian Clason Empirical Aspects of Marginal Universität Graz, Austria Distribution Choice Models Organizer: Juan Carlos De los Xiaobo Li, University of Minnesota, USA; Reyes Vinit Kumar Mishra, University of Sydney, Escuela Politécnica Nacional, Ecuador Australia; Karthik Natarajan, Singapore University of Technology & Design, 4:30-4:55 Finite Element Error Singapore; Dhanesh Padmanabhan, Estimates for Elliptic Optimal Control General Motors R&D, India; Chung-Piaw Problems with Varying Number of Teo, National University of Singapore, Finitely Many Inequality Constraints Singapore Ira Neitzel, Technical University of Munich, Germany; Winnifried Wollner, University 5:00-5:25 A Convex Optimization of Hamburg, Germany Approach for Computing Correlated Choice Probabilities 5:00-5:25 PDE Constrained Selin D. Ahipasaoglu, Singapore University Optimization with Pointwise Gradient- of Technology & Design, Singapore; State Constraints Xiaobo Li, University of Minnesota, USA; Michael Hintermüller, Humboldt University Rudabeh Meskarian and Karthik Natarajan, Berlin, Germany; Anton Schiela, Singapore University of Technology & Technische Universität Hamburg, Design, Singapore Germany; Winnifried Wollner, University of Hamburg, Germany 5:30-5:55 Modeling Dynamic Choice Behavior of Customers Srikanth Jagabathula and Gustavo Vulcano, New York University, USA 6:00-6:25 A New Compact Lp for Network Revenue Management with continued in next column Customer Choice Kalyan Talluri, Universitat Pompeu Fabra, Spain; Sumit Kunnumkal, Indian School of Business (ISB), India 2014 SIAM Conference on Optimization 67

OP14 Abstracts

Abstracts are printed as submitted by the author. 68 OP14 Abstracts

IP1 of polyhedra. Universal Gradient Methods Francisco Santos In Convex Optimization, numerical schemes are always University of Cantabria developed for some specific problem classes. One of the [email protected] most important characteristics of such classes is the level of smoothness of the objective function. Methods for nons- IP4 mooth functions are different from the methods for smooth ones. However, very often the level of smoothness of the Combinatorial Optimization for National Security objective is difficult to estimate in advance. In this talk Applications we present algorithms which adjust their behavior in ac- cordance to the actual level of smoothness observed during National-security optimization problems emphasize safety, the minimization process. Their only input parameter is cost, and compliance, frequently with some twist. They the required accuracy of the solution. We discuss also the may merit parallelization, have to run on small, weak plat- abilities of these schemes in reconstructing the dual solu- forms, or have unusual constraints or objectives. We de- tions. scribe several recent such applications. We summarize a massively-parallel branch-and-bound implementation for a Yurii Nesterov core task in machine classification. A challenging spam- CORE classification problem scales perfectly to over 6000 pro- Universite catholique de Louvain cessors. We discuss math programming for benchmark- [email protected] ing practical wireless-sensor-management heuristics. We sketch theoretically justified algorithms for a scheduling problem motivated by nuclear weapons inspections. Time IP2 permitting, we will present open problems. For exam- ple, can modern nonlinear solvers benchmark a simple lin- Optimizing and Coordinating Healthcare Networks earization heuristic for placing imperfect sensors in a mu- and Markets nicipal water network?

Healthcare systems pose a range of major access, quality Cynthia Phillips and cost challenges in the U.S. and globally. We survey Sandia National Laboratories several healthcare network optimization problems that are [email protected] typically challenging in practice and theory. The challenge comes from network affects, including the fact that in many cases the network consists of selfish and competing players. IP5 Incorporating the complex dynamics into the optimization The Euclidean Distance Degree of an Algebraic Set is rather essential, but hard to models and often leads to computationally intractable models. We focus attention to It is a common problem in optimization to minimize the computationally tractable and practical policies, and an- Euclidean distance from a given data point u to some set alyze their worst-case performance compared to optimal X. In this talk I will consider the situation in which X is de- policies that are computationally intractable and concep- fined by a finite set of polynomial equations. The number tually impractical. of critical points of the objective function on X is called the Euclidean distance degree of X, and is an intrinsic measure Retsef Levi of the complexity of this polynomial optimization prob- Massachusetts Institute of Technology lem. Algebraic geometry offers powerful tools to calculate [email protected] this degree in many situations. I will explain the algebraic methods involved, and illustrate the formulas that can be obtained in several situations ranging from matrix analy- IP3 sis to control theory to computer vision. Joint work with Recent Progress on the Diameter of Polyhedra and Jan Draisma, Emil Horobet, Giorgio Ottaviani and Bernd Simplicial Complexes Sturmfels.

The Hirsch conjecture, posed in 1957, stated that the graph Rekha Thomas of a d-dimensional polytope or polyhedron with n facets University of Washington cannot have diameter greater than n − d. The conjec- [email protected] ture itself has been disproved (Klee-Walkup (1967) for un- bounded polyhedra, Santos (2010) for bounded polytopes), but what we know about the underlying question is quite IP6 scarce. Most notably, no polynomial upper bound is known Modeling Wholesale Electricity Markets with Hy- for the diameters that were conjectured to be linear. In dro Storage contrast, no polyhedron violating the Hirsch bound by more than 25% is known. In this talk we review several re- Over the past two decades most industrialized nations have cent attempts and progress on the question. Some of these instituted markets for wholesale electricity supply. Opti- work in the world of polyhedra or (more often) bounded mization models play a key role in these markets. The un- polytopes, but some try to shed light on the question by derstanding of electricity markets has grown substantially generalizing it to simplicial complexes. In particular, we through various market crises, and there is now a set of show that the maximum diameter of arbitrary simplicial standard electricity market design principles for systems complexesisinnΘ(d), we sketch the proof of Hirsch’s bound with mainly thermal plant. On the other hand, markets for “flag’ polyhedra (and more general objects) by Adipr- with lots of hydro storage can face shortage risks, leading to asito and Benedetti, and we summarize the main ideas in production and pricing arrangements in these markets that the polymath 3 project, a web-based collective effort try- vary widely across jurisdictions. We show how stochastic ing to prove an upper bound of type nd for the diameters optimization and complementarity models can be used to OP14 Abstracts 69

improve our understanding of these systems. method is strongly polynomial for solving deterministic MDPs regardless of discount factors. Andy Philpott University of Auckland Yinyu Ye [email protected] Stanford University [email protected] IP7 Large-Scale Optimization of Multidisciplinary En- CP1 gineering Systems Superlinearly Convergent Smoothing Continuation Algorithms for Nonlinear Complementarity Prob- There is a compelling incentive for using optimization to lems over Definable Convex Cones design aerospace systems due to large penalties incurred by their weight. The design of these systems is challeng- We consider the superlinear convergence of smoothing con- ing due to their complexity. We tackle these challenges by tinuation algorithms without Jacobian consistency. In our developing a new view of multidisciplinary systems, and approach, we use a barrier based smoothing approxima- by combining gradient-based optimization, efficient gradi- tion, which is defined for every closed convex cone. When ent computation, and Newton-type methods. Our applica- the barrier has definable derivative, we prove the superlin- tions include wing design based on Navier–Stokes aerody- ear convergence of a smoothing continuation algorithm for namic models coupled to finite-element structural models, solving nonlinear complementarity problems over the cone. and satellite design including trajectory optimization. The Such barriers exist for cones definable in the o-minimal ex- methods used in this work are generalized and proposed as pansion of globally analytic sets by power functions with a new framework for solving large-scale optimization prob- real algebraic exponents. lems. Chek Beng Chua Joaquim Martins Nanyang Technological University University of Michigan [email protected] [email protected]

CP1 IP8 On the Quadratic Eigenvalue Complementarity A Projection Hierarchy for Some NP-hard Opti- Problem mization Problems A new sufficient condition for the existence of solutions to There exist several hierarchies of relaxations which allow the Quadratic Eigenvalue Complementarity Problem (QE- to solve NP-hard combinatorial optimization problems to iCP) is established. This condition exploits the reduction optimality. The Lasserre and Parrilo hierarchies are based of QEiCP to a normal EiCP. An upper bound for the num- on semidefinite optimization and in each new level both the ber of solutions of the QEiCP is presented. A new strategy dimension and the number of constraints increases. As a for QEiCP is analyzed which solves the resulting EiCP by consequence even the first step up in the hierarchy leads to an equivalent Variational Inequality Problem. Numerical problems which are extremely hard to solve computation- experiments with a projection VI method illustrate the in- ally for nontrivial problem sizes. In contrast, we present terest of this methodology in practice. a hierarchy where the dimension stays fixed and only the number of constraints grows exponentially. It applies to Joaquim J. Jdice problems where the projection of the feasible set to sub- Instituto de Telecomunica˜oes problems has a ’simple’ structure. We consider this new hi- [email protected] erarchy for Max-Cut, Stable-Set and Graph-Coloring. We look at some theoretical properties, discuss practical is- Carmo Br´as sues and provide computational results, comparing the new Universidade Nova de Lisboa bounds with the current state-of-the-art. Portugal Franz Rendl [email protected] Alpen-Adria Universitaet Klagenfurt Institut fuer Mathematik Alfredo N. Iusem [email protected] IMPA, Rio de Janeiro [email protected]

SP1 SIAG/OPT Prize Lecture: Efficiency of the Sim- CP1 plex and Policy Iteration Methods for Markov De- Complementarity Formulations of 0-Norm Opti- cision Processes mization Problems

We prove that the classic policy-iteration method (Howard There has been interest recently in obtaining sparse solu- 1960), including the simple policy-iteration (or simplex tions to optimization problems, eg in compressed sensing. method, Dantzig 1947) with the most-negative-reduced- Minimizing the number of nonzeroes (also known as the 0- cost pivoting rule, is a strongly polynomial-time algorithm norm) is often approximated by minimizing the 1-norm. for solving discounted Markov decision processes (MDP) In contrast, we give an exact formulation as a mathematical of any fixed discount factor. The result is surprising since program with complementarity constraints; our formula- almost all complexity results on the simplex and policy- tion does not require a big-M term. We discuss properties iteration methods are negative, while in practice they are of the exact formulation such as stationarity conditions, popularly used for solving MDPs, or linear programs at and solution procedures for determining local and global large. We also present a result to show that the simplex optimality. We compare our solutions with those from an 70 OP14 Abstracts

1-norm formulation. [email protected], [email protected]

John E. Mitchell Rensselaer Polytechnic Institute CP1 110, 8th Street, Troy, NY, 12180 Computing a Cournot Equilibrium in Integers [email protected] We give an efficient algorithm for computing a Cournot Mingbin Feng equilibrium when the producers are confined to inte- Northwestern University gers, the inverse demand function is linear, and costs are [email protected] quadratic. The method also establishes existence, which follows in much more generality because the problem can Jong-Shi Pang be modelled as a potential game. University of Southern California [email protected] Michael J. Todd Cornell University Xin Shen Sch of Oper Rsch & Indust Eng Rensselaer Polytechnic Institute [email protected] [email protected]

Andreas Waechter CP2 Northwestern University Newton’s Method for Solving Generalized Equa- [email protected] tions Using Set-Valued Approximations

Results on stability of both local and global metric regu- CP1 larity under set-valued perturbations will be presented. As an application, we study (super-)linear convergence of a Competitive Equilibrium Relaxations in General Newton-type iterative process for solving generalized equa- Auctions tions. We will investigate several iterative schemes such as the inexact Newton’s method, the non-smooth Newton’s In practice, many auctions do not possess a competitive method for semi-smooth functions and the inexact proxi- equilibrium. For this reason, we provide two relaxations mal point algorithm. which compute solutions similar to a competitive equilib- rium. Both relaxations determine one price per commodity Samir Adly by solving a non-convex optimization problem. The first Universit´edeLimoges,LaboratoireXLIM model (an MPEC) allows for a fast heuristic and an ex- [email protected] act decomposition algorithm. The second model ensures that no participant can be made better off without making Radek Cibulka another one worse off. University of West Bohemia Pilsen, Czech Republic Johannes C. M¨uller [email protected] University of Erlangen-N¨urnberg [email protected] Huynh Van Ngai University of Qui Nhon, Vietnam [email protected] CP1 Global Convergence of Smoothing and Scaling Con- jugate Gradient Methods for Systems of Nons- CP2 mooth Equations Polynomiality of Inequality-Selecting Interior- Point Methods for Linear Programming The systems of nonsmooth equations is important class of problems, because it has many applications, for example, We present a complexity analysis for interior-point frame- the variational inequality and the complementarity prob- works that solve linear programs by dynamically select- lems. In this talk, we treat numerical methods for solving ing constraints from a known but possibly large set of systems of nonsmooth equations. Systems of nonsmooth inequalities. Unlike conventional cutting-plane methods, equations are reformulated as least squares problems by our method predicts constraint violations and integrates using the smoothing technique. We propose scaling conju- needed inequalities using new types of corrector steps that gate gradient methods for solving this least squares prob- maintain feasibility at each major iteration. Our analysis lems. Moreover, we show the global convergence of the provides theoretical insight into conditions that may cause proposed methods. such algorithms to jam or that guarantee their convergence in polynomial time. Yasushi Narushima Yokohama National University Alexander Engau Department of Manegement System Science University of Colorado, Denver, USA. [email protected] [email protected]

Hiroshi Yabe, Ryousuke Ootani Miguel F. Anjos Tokyo University of Science Mathematics and Industrial Engineering & GERAD Department of Mathematical Information Science Ecole´ Polytechnique de Montr´eal OP14 Abstracts 71

[email protected] Helder Inacio Georgia Institute of Technology [email protected] CP2 FA IPA SDP: A Feasible Arc Interior Point Algo- Sergio Assuncao Monteiro rithm for Nonlinear Semidefinite Programming Federal University of Rio de Janeiro [email protected] FAIPA SDP deals with the minimization of nonlinear func- tions with constraints on nonlinear symmetric matrix- valued functions, in addition to standard nonlinear con- CP2 straints. FAIPA SDP generates a decreasing sequence of A Quadratic Cone Relaxation-Based Algorithm for feasible points. At each iteration, a feasible descent arc Linear Programming is obtained and a line search along this arc is then per- formed. FAIPA SDP merely requires the solution of three We present and analyze a linear programming algorithm linear systems with the same matrix. We prove global and based on replacing the non-negative orthant with larger superlinear convergence. Several numerical examples were quadratic cones. For each quadratic relaxation that has solved very efficiently. an optimal solution, there naturally arises a parameter- ized family of quadratic cones for which the optimal so- Jose Herskovits lutions create a path leading to the linear programs op- COPPE - Federal University of Rio de Janeiro timal solution. We show that this path can be followed [email protected] efficiently, thereby resulting in an algorithm whose com- plexity matches the best bounds proven for interior-point methods. Jean-Rodolphe Roche Institut Elie´ Cartan Mutiara Sondjaja University of Lorraine, France Cornell University [email protected] School of Operations Research and Information Engineering Elmer Bazan [email protected] COPPE Federal University of Rio de Janeiro [email protected] CP3 An Analytic Center Algorithm for Semidefinite Jose Miguel Aroztegui Programming Federal University of Paraiba, In this work, we study the method of analytic centers, Brazil which was developed with the Newton method, applied [email protected] to semidefinite programming. In this method, we com- pute a descent direction using either Newton’s method. We present the complexity analysis and computational exper- CP2 iments performed with the semidefinite programming test An Efficient Algorithm for the Projection of a Point problems available on SDPLIB. on the Intersection of M Hyperplanes with Non- negative Coefficients and The Nonnegative Orthant Sergio Assuncao Monteiro Federal University of Rio de Janeiro In this work, we present an efficient algorithm for the pro- [email protected] jection of a point on the intersection of an arbitrary num- ber of nonnegative hyperplanes and the nonnegative or- Claudio Prata Santiago thant. Interior point methods are the most efficient algo- Center for Applied Scientific Computing rithm in the literature to solve this problem. While ef- Lawrence Livermore National Laboratory ficient in practice, the complexity of interior point meth- [email protected] ods is bounded by a polynomial in the dimension of the problem and in the accuracy of the solution. In addition, Paulo Roberto Oliveira their efficiency is highly dependent on a series of param- COPPE/UFRJ eters depending on the specific method chosen (especially [email protected] for nonlinear problems), such as step size, barrier param- eter, accuracy, among others. We propose a new method Nelson Maculan based on the KKT optimality conditions. In this method, Programa de Engenharia de Sistemas de Computao we write the problem as a function of the Lagrangian mul- COPPE Universidade Federal do Rio de Janeiro tipliers of the hyperplanes and seek to find an m-tuple of [email protected] multipliers that corresponds to the optimal solution. We present the numerical experiments for m = 2, 5, 10. CP3 Claudio Santiago A Low-Rank Algorithm to Solve SDP’s with Block Lawrence Livermore National Laboratory Diagonal Constraints via Riemannian Optimization [email protected] We solve Nelson Maculan min f(X) such that X 0andXii = Id for all i, Programa de Engenharia de Sistemas de Computao X COPPE Universidade Federal do Rio de Janeiro [email protected] where f is convex, Id denotes the d×d identity matrix and 72 OP14 Abstracts

th Xii is the i diagonal block of X. These appear in ap- [email protected] proximation algorithms for some NP-hard problems, where the solution is expected to have low rank. To fully exploit this structure, we demonstrate that the search space ad- CP3 mits a Riemannian geometry when rank(X)isfixed,and Gradient Methods for Least-Squares Over Cones show how to obtain a solution by gradually incrementing the rank, using Riemannian optimization. Nesterov’s excessive gap method is extended to more gen- eral problems. Fast gradient methods are used to solve Nicolas Boumal, P.-A. Absil the least squares problem over convex cones – include the Universit´e catholique de Louvain semidefinite least squares problem. Numerical examples [email protected], [email protected] are given. Yu Xia School of Mathematics CP3 University of Birmingham [email protected] The Simplex Method and 0-1 Polytopes

We will derive two results of the primal simplex method by CP3 constructing simple LP instances on 0-1 polytopes. One of Augmented Lagrangian-Type Solution Methods for the results is that, for any 0-1 polytope and any its two Second Order Conic Optimization vertices, we can construct an LP instance for which the simplex method finds a path between them, whose length We consider the minimization of a nonlinear function is at most the dimension of the polytope. This proves a partly subjected to nonlinear second order conic con- well-known result that the diameter of any 0-1 polytope is straints. Augmented Lagrangian-type functions for the bounded by its dimension. Next we show that an upper problem are discussed, with focus on the Log-Sigmoid ap- bound for the number of distinct solutions generated by proach. We then respectively apply the Newton and Quasi- the simplex method is tight. We prove these results by Newton methods to solve the resulting unconstrained using a basic property of the simplex method. model. In this talk, we present the method and numerical experiments done on some problems from trust topology.

Tomonari Kitahara, Shinji Mizuno Alain B. Zemkoho,MichalKocvara Tokyo Institute of Technology School of Mathematics [email protected], [email protected] University of Birmingham [email protected], [email protected]

CP3 CP4 On a Polynomial Choice of Parameters for Interior Nonmonotone Grasp Point Methods A Greedy Randomized Adaptive Search Procedure In this work we solve Linear optimization problems on a (GRASP) is an iterative multistart metaheuristic for dif- Interior Point Method environment by combining different ficult combinatorial optimization. Each GRASP iteration directions, such as predictor, corrector or centering ones, to consists of two phases: a construction phase and a local produce a better direction. We measure how good the new search phase that applies iterative improvement until a lo- direction by using a polynomial merit function on variables cally optimal solution is found. During this phase, starting (α, μ, σ), where α is the step length, μ defines the central from the current solution only improving neighbor solu- path and σ models the weight that a corrector directions tions are accepted and considered as new current solution. should have in a predictor-corrector method. Some nu- In this talk, we propose a new variant of the GRASP frame- merical test show that this approach is competitive when work that explores the neighborhood of the current solu- compared to more well established solvers as PCx,usingthe tion in a nonmonotone fashion resembling a nonmonotone Netlib test set. line search technique proposed by Grippo et al. in 1986 for Newton’s method in nonlinear optimization. We prove the convergence of the resulting Nonmonotone GRASP (NM- Luiz-Rafael Santos GRASP) and illustrate its effectiveness on the Maximum Federal Institute Catarinense Cut Problem (MAX-CUT), a classical hard combinatorial IMECC/Unicamp optimization problem. [email protected] Paola Festa Fernando Villas-Bˆoas Department of Mathematics and Applications ”R. IMECC/Unicamp Caccioppoli” Brazil University of Napoli FEDERICO II [email protected] [email protected]

Aurelio Oliveira Marianna De Santis Departamento de Matem´atica Aplicada Dipartimento di Informatica e Sistemistica Universidade Estadual de Campinas La Sapienza University of Rome, Italy. [email protected] [email protected]

Clovis Perin Giampaolo Liuzzi IMECC/Unicamp Istituto di Analisi dei Sistemi ed Informatica OP14 Abstracts 73

CNR, Rome, Italy [email protected] [email protected]

Stefano Lucidi, Francesco Rinaldi CP4 University of Rome The Discrete Ordered Median Problem Revisited [email protected], [email protected] This paper compares classical and new formulations for the discrete ordered median location problem. Starting with CP4 three already known formulations that use different spaces of variables, we introduce two new ones based on new prop- A New Exact Appoach for a Generalized Quadratic erties of the problem. The first formulation is based on Assignment Problem an observation about how to use scheduling constraints to model sorting. The second one is a novel approach based on a new mixed integer non-linear programming paradigm. The quadratic matching problem (QMP) asks for a match- Some of the new formulations decrease considerably the ing in a graph that optimizes a quadratic objective in the number of constraints to define the problem with respect edge variables. The QMP generalizes the quadratic as- to some previously known formulations. Furthermore, the signment problem. In our approach we strengthen the lin- lower bounds provided by their continuous relaxations im- earized IP-formulation by cutting planes that are derived prove the ones obtained with previous formulations in the from facets of the corresponding matching problem with literature even when strengthened is not applied. Exten- only one quadratic term in the objective function. We sive computational results are presented. present methods to strengthen these new inequalities for the general QMP and report computational results. Justo Puerto Universidad de Sevilla Lena Hupp [email protected] Friedrich-Alexander-Universitaet Erlangen-Nuernberg (FAU) Chair of EDOM CP4 [email protected] The Graph Partitioning into Constrained Spanning Sub-Trees Laura Klein Technische Universitaet Dortmund Lukes (1974) suggested a heuristic of partitioning a general Fakultaet fuer Mathematik - LSV graph by taking a maximum spanning tree of the graph and [email protected] then partitioning the tree based on his pseudo-polynomial algorithm to solve a uniformly constrained tree partition Frauke Liers problem. We generalize the heuristic to the maximum Institut f¨ur Informatik graph partitioning into k constrained spanning sub-trees Universit¨at zu K¨oln and develop a tree updating heuristic to solve the general- [email protected] ized problem. As a sub-problem of the generalized prob- lem, the constrained tree partition problem is tackled by Lagrangian relaxation method based on the branch-and- cut algorithm which Lee,Chopra and Shim (2013) gave to CP4 solve the unconstrained tree partition problem. Binary Steiner Trees and Their Application in Bi- ological Sequence Analysis Sangho Shim, Sunil Chopra Kellogg School of Management Northwestern University Advances in molecular bioinformatics have led to increased [email protected], information about biological sequences. Alignments and [email protected] phylogentic trees of these sequences play an important role in inferring evolutionary history. The problem of finding Diego Klabjan tree alignments or phylogentic trees can be modeled as Industrial Engineering and Management Sciences a binary Steiner tree problem. We study the computa- Northwestern University tional complexity of this problem and introduce the binary [email protected] Steiner tree polytop and valid inequalities. An approach for computing binary Steiner trees and computational re- sults are presented. Dabeen Lee POSTECH, Korea [email protected] Susanne Pape FAU Erlangen-N¨urnberg, Germany Discrete Optimization CP4 [email protected] Effects of the Lll Reduction on Integer Least Squares Estimation Frauke Liers Institut f¨ur Informatik Solving an integer least squares problem is often needed Universit¨at zu K¨oln to estimate an unknown integer parameter vector from a [email protected] linear model. It has been observed that the LLL reduction can improve the success probability of the Babai point, a Alexander Martin suboptimal solution, and decrease the cost of the sphere University of Erlangen-N¨urnberg decoders by simulations. In this talk, we will theoretically 74 OP14 Abstracts

show that these two observations are indeed true. lems solving such than an objective function is maximized in the domain of a feasible region subject to time consis- Jinming Wen tent first- and second- order Stochastic Dominance Con- Dept. of Mathematics and Statistics, McGill University straints integer-recourse. We also present an extension of [email protected] our Branch-and-Fix Coordination Algorithm, where a spe- cial treatment is given to cross scenario constraints that Xiao-Wen Chang link variables from different scenarios. McGill University School of Computer Science Araceli Garin [email protected] University of the Basque Country UPV/EHU [email protected] Xiaohu Xie School of Computer Science, McGill University Laureano F. Escudero [email protected] Dept. de Estadistica e Investigacion Operativa Universidad Rey Juan Carlos [email protected] CP5 Parallelized Branch-and-Fix Coordination Algo- Maria Merino rithm (P-Bfc) for Solving Large-Scale Multistage Dept de Matematica. Aplicada, Estadistica e IO Stochastic Mixed 0-1 Problems Universidad del Pais Vasco [email protected] The parallelization is performed in two levels. The inner one parallelizes the optimization of the set of MIP submod- Gloria Perez els attached to the set of scenario clusters in our Branch- Dept. de Matematica. Aplicada, Estadistica e IO and-Fix Coordination methodology. The outer level defines Universidad del Pais Vasco a set of 0-1 variables where the combinations of their 0-1 [email protected] values, so named paths allow independent models to be op- timized in parallel. Computational results are reported on a broad testbed. CP5 Laureano F. Escudero Distributed Generation for Energy-Efficient Build- Dept. de Estadistica e Investigacion Operativa ings: A Mixed-Integer Multi-Period Optimization Universidad Rey Juan Carlos Approach [email protected] We study a two-stage mixed-integer program with appli- cation in distributed generation for energy-efficient build- Unai Aldasoro ings. This challenging problem is beyond the capacity of Dept. de Matematica Aplicada current solvers, because the second-stage problem contains Universidad del Pais Vasco a large number of binary variables. By exploiting periodic- [email protected] ity structure in daily, weekly, seasonal, and yearly demand profiles, we develop a column generation approach that Maria Merino significantly reduces the number of binary variables, con- Dept de Matematica. Aplicada, Estadistica e IO sequently, rendering computationally tractable problems. Universidad del Pais Vasco Furthermore, our approach provides bounds with provable [email protected] performance guarantees.

Gloria Perez Fu Lin Dept. de Matematica. Aplicada, Estadistica e IO Mathematics and Computer Science Division Universidad del Pais Vasco Argonne National Laboratory [email protected] [email protected]

Sven Leyffer CP5 Argonne National Laboratory Robust Decision Analysis in Discrete Scenario leyff[email protected] Spaces In this paper, we introduce an efficient method for find- Todd Munson ing optimal dynamic policies in decision making problems Argonne National Laboratory with discrete scenario space. We assume the probability Mathematics and Computer Science Division distribution belongs to an arbitrary box uncertainty set. [email protected]

Saman Eskandarzadeh Sharif University of Technology CP5 [email protected] Linear Programming Twin Support Vector Regres- sion Using Newton Method

CP5 In this paper, a new linear programming formulation of a 1- On Solving Multistage Mixed 0-1 Optimization norm twin support vector regression is proposed whose so- Problems with Some Risk Averse Strategies lution is obtained by solving a pair of dual exterior penalty problems as unconstrained minimization problems using We extend to the multistage case two recent risk averse Newton method. The idea of our formulation is to reformu- measures defined for two-stage stochastic mixed 0-1 prob- late TSVR as a strongly convex problem by incorporated OP14 Abstracts 75

regularization technique and then derive a new 1-norm lin- [email protected] ear programming formulation for TSVR to improve robust- ness and sparsity. Our approach has the advantage that a pair of matrix equation of order equals to the number of CP6 input examples is solved at each iteration of the algorithm. Two-Stage Portfolio Optimization with Higher- The algorithm converges from any starting point and can Order Conditional Measures of Risk be easily implemented in MATLAB without using any opti- mization packages. The efficiency of the proposed method We describe a study of novel risk modeling and optimiza- is demonstrated by experimental results on a number of tion techniques to daily portfolio management. First, we interesting synthetic and real-world datasets. develop and compare specialized methods for scenario gen- eration and scenario tree construction. Second, we con- Mohammad Tanveer struct a two-stage stochastic programming problem with Jawaharlal Nehru University New Delhi conditional measures of risk, which is used to re-balance tanveer [email protected] the portfolio on a rolling horizon basis, transaction costs included in the model. Third, we present an extensive sim- ulation study on real-world data. CP5 An Integer L-Shaped Algorithm for the Single- Sitki Gulten Vehicle Routing Problem with Simultaneous De- Rutgers University livery and Stochastic Pickup Rutgers Business School [email protected] Stochastic vehicle routing with simultaneous delivery and pickup is addressed. Quantities to be delivered are fixed, Andrzej Ruszczynski whereas pick-up quantities are uncertain. For route fail- Rutgers University ures, i.e., arriving at a customer with insufficient vehicle MSIS Department capacity, compensation strategies are developed. In the [email protected] single vehicle case, a recourse stochastic program is formu- lated and solved by the integer L-shaped method. Algo- rithm efficiency is improved by strengthening lower bounds CP6 for recourse cost associated to partial routes encountered Semi-Stochastic Gradient Descent Methods throughout the solution process. Nadine Wollenberg In this paper we study the problem of minimizing the aver- University of Duisburg-Essen age of a large number (n) of smooth convex loss functions. [email protected] We propose a new method, S2GD (Semi-Stochastic Gra- dient Descent), which runs for one or several epochs in each of which a single full gradient and a random number Michel Gendreau of stochastic gradients is computed, following a geometric Universit´edeMontr´eal law. The total work needed for the method to output an [email protected] ε-accurate solution in expectation, measured in the num- ber of passes over data, or equivalently, in units equiva- Walter Rei lent to the computation of a single gradient of the loss, is Universit´eduQu´ebec `aMontr´eal O((κ/n)log(1/ε)), where κ is the condition number. This [email protected] is achieved by running the method for O(log(1/ε)) epochs, with a single gradient evaluation and O(κ) stochastic gra- R¨udiger Schultz dient evaluations in each. The SVRG method of Johnson University of Duisburg-Essen and Zhang arises as a special case. If our method is lim- [email protected] ited to a single epoch only, it needs to evaluate at most O((κ/ε)log(1/ε)) stochastic gradients. In contrast, SVRG requires O(κ/ε2) stochastic gradients. To illustrate our CP6 theoretical results, S2GD only needs the workload equiva- Optimal Low-Rank Inverse Matrices for Inverse lent to about 2.1 full gradient evaluations to find an 10−6- Problems accurate solution for a problem with n =109 and κ =103.

Inverse problems arise in scientific applications such as Jakub Konecny, Peter Richtarik biomedical imaging, computational biology, and geo- University of Edinburgh physics, and computing accurate solutions to inverse prob- [email protected], [email protected] lems can be both mathematically and computationally challenging. By incorporating probabilistic information, we reformulate the inverse problem as a Bayes risk opti- CP6 mization problem. We present theoretical results for the Risk Averse Computational Stochastic Program- low-rank Bayes optimization problem and discuss efficient ming methods for solving associated empirical Bayes risk opti- mization problems. We demonstrate our algorithms on ex- This work proposed a (general purpose) framework to amples from image processing. compute an approximate optimal non-stationary policy for finite-horizon, discrete-time Markov decision problems, Matthias Chung, Julianne Chung when the optimality criteria explicitly includes the ex- Virginia Tech pected value and a given risk measure of the total cost. The [email protected], [email protected] methodology relies on a parametric decision rule, direct policy search, spline approximation, and a nested parallel Dianne P. O’Leary derivative free optimization. A Java implementation of the University of Maryland, College Park approach is provided in a software package, called Smart 76 OP14 Abstracts

Parallel Policy Search (Smart-PPS). Tractability and effi- condition on a suitable critical cone. We conclude with ciency of the devised method are illustrated by an energy first numerical results. system risk management problem. Stefanie Bott Somayeh Moazeni TU Darmstadt School of Computer Science Graduate School CE University of Waterloo [email protected] [email protected] Stefan Ulbrich Warren Powell Technische Universitaet Darmstadt Princeton University Fachbereich Mathematik [email protected] [email protected]

CP6 CP7 Support Vector Machines Via Stochastic Optimiza- Optimal Decay Rate for the Indirect Stabilization tion of Hyperbolic Systems

We discuss the problem of identifying two classes of records We consider the indirect stabilization of systems of by means of a classifier function. We propose a new ap- hyperbolic-type equations, such as wave and plate equa- proach to determine a classifier using techniques of stochas- tions with different boundary conditions. By energy tic programming and risk-averse optimization. The ap- method, we show that a single feedback allows to stabi- proach is particularly relevant when the observations of lize the full system at a polynomial rate. Furthermore, we one of the classes of interest are substantially less than exploit refined resolvent estimates deducing the optimal the data available for the other class. We compare the decay rate of the energy of the system. Numerical simula- proposed new methods to existing ones and analyze their tions show resonance effect between the components of the classification power. system and confirm optimality of the proven decay rate.

Konstantin Patev Roberto Guglielmi Rutgers, The State University of New Jersey University of Bayreuth [email protected] [email protected]

Darinka Dentcheva Department of Mathematical Sciences CP7 Stevens Institute of Technology Recent Advances in Optimum Experimental De- [email protected] sign for PDE Models

Mathematical models are of great importance in engineer- CP6 ing and natural sciences. Usually models contain un- Decomposition Algorithms for the Optimal Power known parameters which must be estimated from exper- Flow Problem under Uncertainty imental data. Often many expensive experiments have to be performed for estimating the parameters in order to get A two-stage non-linear stochastic formulation for the op- enough information for parameter estimation. The number timal power flow problem under renewable-generation un- of experiments can be drastically reduced by computing op- certainty is investigated. Certain generation decisions are timal experiments. Our talk presents an efficient method made only in the first stage and fixed for the second stage, for computing optimal experiments for processes described where the actual renewable generation is realized. The un- by partial differential equations (PDE). certainty in renewable output is captured by a finite num- ber of scenarios. We present two outer-approximation algo- Gregor Kriwet,EkaterinaKostina rithms to solve the large-scale non-convex problem, where Fachbereich Mathematik und Informatik a global solution is obtained under some suitable assump- Philipps-Universitaet Marburg tions. [email protected], [email protected] Dzung Phan IBM T.J. Watson Research Center [email protected] CP7 Optimal Control of Index 1 PDAEs

CP7 We consider optimal control problems constrained by Par- Adaptive Multilevel Sqp Method for State Con- tial Differential-Algebraic Equations, which we interpret strained Optimization with Navier-Stokes Equa- as abstract DAEs on Banach spaces. The latter trans- tions lates to coupled systems of each a parabolic/hyperbolic- and a quasi-stationary elliptic equation for two unknown We consider an adaptive multilevel method for optimal functions. We set up the general optimality theory and control problems with state constraints. More precisely, we explore implications of the so-called index of a (P)DAE, combine an SQP algorithm with Moreau-Yosida regulariza- i.e., an invertibility supposition for the deriviative of the tion where adaptive mesh refinements and the penalty pa- differential-operator for the elliptic equation. rameter update are appropriately connected. Afterwards, convergence results without and with second-order suffi- Hannes Meinlschmidt cient optimality condition are presented. In order to apply TU Darmstadt our results to flow control problems, we present an (SSC) [email protected] OP14 Abstracts 77

Stefan Ulbrich constraints. Technische Universitaet Darmstadt Fachbereich Mathematik Charkaz A. Aghayeva [email protected] Azerbaijan National of Science, Baku; Anadolu University, Tu Anadolu University, Turkey CP7 [email protected] Optimal Control of Nonlinear Hyperbolic Conser- vation Laws with Switching CP8 A Semismooth Newton-Cg Method for Full Wave- Entropy solutions of hyperbolic conservation laws gener- form Seismic Tomography ally contain shocks, which leads to non-differentiability of the solution operator. The notion of shift-differentiability We present a semismooth Newton-PCG method for full turned out to be useful in order to show Fr´echet- waveform seismic inversion governed by the elastic wave differentiabilty of the reduced objective for Initial- equation. We establish results on the differentiability of (Boundary-) Value Problems. In this talk we consider the solution operator and analyze a Moreau-Yosida regu- problems motivated by traffic flow modeling, that involve larization to handle constraints on the material parameters. conservation laws with additional discontinuities intro- Numerical results are shown for an example of geophysical duced by switching. We discuss how the ideas from the exploration on reservoir scale. Additionally, we outline the I(B)VP-case can be transfered to such problems. application of the semismooth Newton method to a nons- mooth sparsity regularization in curvelet space. Sebastian Pfaff TU Darmstadt Christian Boehm Nonlinear Optimization Technische Universit¨at M¨unchen pfaff@mathematik.tu-darmstadt.de [email protected]

Stefan Ulbrich Michael Ulbrich Technische Universitaet Darmstadt Technische Universitaet Muenchen Fachbereich Mathematik Chair of Mathematical Optimization [email protected] [email protected]

CP7 CP8 A Numerical Method for Design of Optimal Exper- A New Semi-Smooth Newton Multigrid Method iments for Model Discrimination under Model and for Control-Constrained Semi-Linear Elliptic PDE Data Uncertainties Problems

The development and validation of complex nonlinear dif- In this paper a new multigrid algorithm is proposed to ferential equation models is a difficult task that requires accelerate the semi-smooth Newton method that is applied the support by numerical methods for sensitivity analy- to the first order necessary optimality systems arising from sis, parameter estimation, and optimal design of experi- a class of semi-linear control-constrained elliptic optimal ments. The talk presents an efficient method for design of control problems. Under admissible assumptions on the optimal experiments for model discrimination by lack-of-fit nonlinearity, the discretized Jacobian matrix is proved to test. Mathematically this leads to highly structured, multi- have an uniformly bounded inverse with respect to mesh ple experiments, multiple models optimal control problems size. Different from current available approaches, a new with integer controls. Special emphasis is placed on robus- numerical implementation that leads to a robust multigrid tification of optimal designs against uncertainties. solver is employed to coarsen the grid operator. Numerical simulations are provided to illustrate the efficiency of the Hilke Stibbe proposed method, which shows to be computationally more University of Marburg efficient than the full-approximation-storage multigrid in [email protected] current literature. Jun Liu Ekaterina Kostina Southern Illinois University Fachbereich Mathematik und Informatik [email protected] Philipps-Universitaet Marburg [email protected] Mingqing Xiao Southern Illinois University at Carbondale [email protected] CP8 Stochastic Optimal Control Problem for Delayed Switching Systems CP8 Iterative Solvers for Time-Dependent Pde- This paper provides necessary condition of optimality, in Constrained Optimization Problems the form of a maximum principle, for optimal control prob- lem of switching systems with constraints. Dynamics of A major area of research in continuous optimization, as the constituent processes take the form of stochastic dif- well as applied science more widely, is that of developing ferential equations with delay on state and control terms fast and robust solution techniques for PDE-constrained . The restrictions on the transitions or switches between optimization problems. In this talk we construct precondi- operating mode, are described by collections of functional tioned iterative solvers for a range of time-dependent prob- 78 OP14 Abstracts

lems of this form, using the theory of saddle point systems. β,p > 0, where x∗ ∈ X∗,theoptimalset. Our solvers are efficient and flexible, involve storing rela- tively small matrices, can be parallelized, and are tailored Clovis C. Gonzaga towards a number of scientific applications. Federal Univ of Sta Catarina [email protected] John W. Pearson University of Edinburgh [email protected] CP9 A Flexible Inexact Restoration Method and Appli- CP8 cation to Multiobjective Constrained Optimization Second Order Conditions for Strong Minima in the We introduce a new flexible Inexact-Restoration (IR) al- Optimal Control of Partial Differential Equations gorithm and an application to Multiobjective Constrained Optimization Problems (MCOP) under the weighted-sum I this talk we summarize some recent advances on the the- scalarization approach. In IR methods each iteration has ory of optimality conditions for the optimization of par- two phases. In the first phase one aims to improve fea- tial differential equations. We focus our attention on the sibility and, in the second phase, one minimizes a suit- concept of strong minima for optimal control problems able objective function. In the second phase we also im- governed by semi-linear elliptic and parabolic equations. pose bounded deterioration of the feasibility obtained in Whereas in the field of calculus of variations this notion has the first phase. Here we combine the basic ideas of the been deeply investigated, the study of strong solutions for Fischer-Friedlander approach for IR with the use of ap- optimal control problems of partial differential equations proximations of the Lagrange multipliers. We present a has been addressed only recently. We first revisit some new option to obtain a range of search directions in the well-known results coming from the calculus of variations optimization phase and we employ the sharp Lagrangian that will highlight the subsequent results. We then present as merit function. Furthermore, we introduce a flexible way a characterization of strong minima satisfying quadratic to handle sufficient decrease requirements and an efficient growth for optimal control problems of semi-linear equa- way to deal with the penalty parameter. We show that tions and we end by describing some current investigations. with the IR framework there is a natural way to explore the structure of the MCOP in both IR phases. Global conver- gence of the proposed IR method is proved and examples Francisco J. Silva of the numerical behavior of the algorithm are reported. Univesit´e de Limoges [email protected] Luis Felipe Bueno, Gabriel Haeser Federal University of Sao Paulo [email protected], [email protected] CP8 Branch and Bound Approach to the Optimal Con- Jos´eMarioMart´ınez trol of Non-Smooth Dynamic Systems Department of Applied Mathematics State University of Campinas We consider optimal control problems, given in terms of [email protected] a set of ordinary differential and algebraic equations. We allow for non-smooth functions in the model equations, in particular for piecewise linear functions, commonly used to CP9 represent linearly interpolated data. In order to solve these problems, we reformulate them into mixed-integer nonlin- The Alpha-BB Cutting Plane Algorithm for ear optimization problems, and apply a tailored branch- Semi-Infinite Programming Problem with Multi- and-bound approach. The talk will present the approach Dimensional Index Set as well as numerical results for test instances. In this study, we propose the alpha-BB cutting plane al- Ingmar Vierhaus gorithm for semi-infinite program (SIP) in which the in- Zuse Institute Berlin dex set is multi-dimensional. When the index set is one- [email protected] dimensional, the “refinement step’ in the algorithm is just to divide a closed interval into the left and the right. How- Armin F¨ugenschuh ever, such a division is not unique in the multi-dimensional Helmut Schmidt University Hamburg case. We also show that the sequence generated by the [email protected] algorithm converges to the SIP optimum under mild as- sumptions.

CP9 Shunsuke Hayashi Graduate School of Information Sciences, Tohoku On the Full Convergence of Descent Methods for University Convex Optimization s [email protected] A strong descent step h from x for a convex function f satisfies f(x + h) − f(x) ≤−αF (x)h,whereF (x)= Soon-Yi Wu min{γ|γ ∈ ∂f(x)} and α ∈ (0, 1). Algorithms like steep- National Cheng-Kung University est descent, trust region and quasi-Newton using Armijo [email protected] descent conditions are strong. We show that under a mild growth condition on f these methods generate convergent sequences of iterates. The condition, which is implied by CP9 sub-analyticity, is f(x) − f(x∗) ≥ βd(x, X∗)p for some Optimality Conditions for Global Minima of Non- OP14 Abstracts 79

convex Functions on Riemannian Manifolds Learning principles are a main key of the procedure.

A version of Lagrange multipliers rule for locally Lipschitz Umberto Dellepiane, Alberto De Santis functions is presented. Using Lagrange multipliers, a suffi- Department of Computer, Control, and Management cient condition for x to be a global minimizer of a locally Engineering Lipschitz function defined on a Riemannian manifold, is Antonio Ruberti at Sapienza University of Rome proved. Then a necessary and sufficient condition for fea- [email protected], [email protected] sible point x to be a global minimizer of a concave function on a Rimennian manifold is obtained. Stefano Lucidi University of Rome Seyedehsomayeh Hosseini [email protected] Seminar for Applied Mathematic, ETH University Stefania Renzi [email protected] Department of Computer, Control, and Management Engineering Antonio Ruberti at Sapienza University of Rome CP9 [email protected] A Multiresolution Proximal Method for Large Scale Nonlinear Optimisation CP10 The computational efficiency of optimisation algorithms Globie: An Algorithm for the Deterministic Global depends on both the dimensionality and the smoothness Optimization of Box-Constrained Nlps of the model. Several algorithms have been proposed to take advantage of lower dimensional models (e.g. multi A novel algorithm for the deterministic global optimization grid methods). In addition several methods smooth the of box-constrained NLPs is presented. A key feature of the model and apply a smooth optimisation algorithm. In this proposed algorithm is mapping the objective function to a talk we present an algorithm that simultaneously smooths modified subenergy function. This mapping can be used out and reduces the dimensions of the model. We discuss to impose arbitrary bounds on the negative eigenvalues of the convergence of the algorithm and discuss its application the Hessian matrix of the new objective function. This to the problem of computing index-1 saddle points. property is exploited to identify sub-domains that cannot contain the global solution, using the aBB algorithm. Panos Parpas Massachusetts Institute of Technology Nikolaos Kazazakis, Claire Adjiman [email protected] Imperial College London [email protected], [email protected] CP9 Global Convergence Analysis CP10 of Several Riemannian Conjugate Gradient Meth- ParDYCORS for Global Optimization ods ParDYCORS (Parallel DYnamic COordinate search using We present a globally convergent conjugate gradient Response Surface models) is based on DYCORS [Regis and method with the Fletcher-Reeves β on Riemannian man- Shoemaker (2013)], in which the next evaluation points are ifolds. The notion of “scaled” vector transport plays an selected from random trial solutions obtained by perturb- important role to guarantee the global convergence prop- ing only a subset of the coordinates of the best solution. erty without any artificial assumptions. We also discuss To reduce the computation time, selected evaluation points a Riemannian conjugate gradient method with another β are evaluated in parallel. Several numerical results are il- and show that the method has global convergence prop- lustrated, demonstrating that ParDYCORS makes a fast erty if a Wolfe step size is chosen in each iteration, while decrease in f(x) and is well suited for a high-dimensional a strong Wolfe step size has to be chosen in the Fletcher- problem. Reeves type method. Tipaluck Krityakierne Hiroyuki Sato Center for Applied Mathematics, Cornell University Department of Applied Mathematics and Physics [email protected] Graduate School of Informatics, Kyoto University [email protected] Christine A. Shoemaker Cornell University [email protected] CP10 Approaching Forex Trading Through Global Opti- mization Techniques CP10 Derivative-Free Multi-Agent Optimization In applying optimization techniques to trading in Forex, given a strategy and defined the optimization problem We present an algorithm for the cooperative optimization properly, a reliable global optimization method must be of multiple agents connected by a network. We attempt to chosen. This study is aimed to highlight possible rela- optimize a global objective which is the sum of the agents’ tionships between the behavior of the particular security objectives when each agent can only query his objective. and the performance of the global optimization technique. Since each agent does not have reliable derivative informa- Moreover, comparing the results on different trading strate- tion, we show that agents building local models of their gies, the best strategy could be identified on a given item. objectives can achieve their collective goal. We lastly show 80 OP14 Abstracts

how our algorithm performs on benchmark problems. [email protected]

Jeffrey M. Larson, Euhanna Ghadimi, Mikael Johansson KTH - Royal Institute of Technology CP11 jeff[email protected], [email protected], [email protected] Optimal Unit Selection in Batched Liquid Pipelines with DRA CP10 An Interval Algorithm for Global Optimization un- Petroleum products with different physical properties are der Linear and Box Constraints often transported as multiple discrete batches in a cross- country pipeline. Tens of millions of dollars per year are We describe a method based on interval analysis and typically spent on chemical drag reducing agents (DRA) branch and bound to find at least one global minimum of and electric power for pump stations with multiple inde- a differentiable nonconvex function under linear and box pendent units. An important issue is that DRA is heavily constraints. We use KKT conditions to bound Lagrange degraded by certain operational conditions. We consider multipliers and to apply constraint propagation. Our ap- hybrid algorithms that handle both the discrete and con- proach needs only the first derivative of objective function tinuous aspects of this problem very effectively. and can also find solutions on the boundary of box. We implement our ideas in C++ and compare our results with Richard Carter Baron GL-Group [email protected] Tiago M. Montanher University of So Paulo Institute of Mathematics and Statistics CP11 [email protected] Polynomial Optimization for Water Distribution Networks CP10 Pressure management in water networks can be modeled Complexity and Optimality of Subclasses of Zero- as a polynomial optimization problem. The problem is Order Global Optimization challenging to solve due to the non-linearities in the hy- draulic and energy conservation equations that define the The problem class of Zero-order Global Optimization is way water flows through a network. We propose an itera- split into subclasses according to a proposed “complexity tive scheme for improving semidefinite-based relaxations of measure”. For each subclass, Complexity and Branch-and- the pressure management problem. The approach is based Bound methods’ Laboriousness are rigorously estimated. on a recently proposed dynamic approach for generating For conventional “Cubic” BnB upper and lower laborious- valid inequalities to polynomial programs. Computational ness estimates are obtained. A new method based on the ∗ results are presented. An lattice is presented with upper laboriousness bound smaller than the lower bound of the conventional method by factor O(3n/2). All results are extended to Adaptive Bissan Ghaddar Covering problems. University of Waterloo [email protected] Serge L. Shishkin, Alan Finn United Technologies Corp., Research Center Mathieu Claeys [email protected], fi[email protected] University of Cambridge [email protected]

CP11 Martin Mevissen Stochastic On-time Arrival Problem for Public IBM Research Transportation Networks [email protected]

We formulate the stochastic on-time arrival problem on public transportation networks. Link travel times are mod- CP11 eled using a route-based headway diffusion model, which al- lows the computation of associated waiting times at transit Pump Scheduling in Water Networks nodes. We propose a label-setting algorithm for comput- ing the solution to this problem, and show that this for- We present a mixed integer nonlinear program for the mulation leads to efficient computations tractable for large pump scheduling problem in water networks. The proposed networks, even in the case of time-varying distributions. model uses a quadratic approximation for the nonlinear Algorithm performance is illustrated using publicly avail- hydraulic equations. We propose a Lagrangian decomposi- able and synthetic transit data. tion coupled with a simulation based heuristic as a solution methodology and present results on water networks from Sebastien Blandin the literature. Research Scientist IBM Research Collaboratory – Singapore Joe Naoum-Sawaya [email protected] IBM Research [email protected] Samitha Samaranayake Systems Engineering Bissan Ghaddar UC Berkeley University of Waterloo OP14 Abstracts 81

[email protected] Columbia University New York, NY [email protected] CP11 Generation of Relevant Elementary Flux Modes in Xiaocheng Tang Metabolic Networks Lehigh University Bethlehem, PA Elementary flux modes (EFMs) are a set of pathways in a [email protected] metabolic reaction network that can be combined to define any feasible pathway in the network. It is in general pro- Amit Chakraborty hibitive to enumerate all EFMs for complex networks. We Siemens Corporation, Corporate Technology present a method based on a column generation technique Princeton, NJ that finds EFMs in a dynamic fashion, while minimizing [email protected] the norm of the difference between extracellular rate mea- surements and calculated flux through the EFMs in the network. CP12 Hildur sa Oddsd´ottir A Parallel Optimization Algorithm for Nonnega- KTH Royal Institute of Technology. tive Tensor Factorization [email protected] We present a new algorithm for bound constrained opti- mization. The algorithm uses multiple points to build a Erika Hagrot, V´eronique Chotteau representative model of the objective function. Using the KTH Royal Institute of Technology new information gathered from those multiple points; a lo- [email protected], [email protected] cal step is gradually improved by updating its direction as well as its length. The algorithm scales well on a shared- Anders Forsgren memory system. We provide parallel implementation de- KTH Royal Institute of Technology tails accompanied with numerical results on nonnegative Dept. of Mathematics tensor factorization models. [email protected] Ilker S. Birbil Sabanci University CP11 [email protected] Discrete Optimization Methods for Pipe Routing Figen Oztoprak Our problem comes from the planning of a power plant Istanbul Technical University which involves the routing of a high-pressure pipe. Opti- fi[email protected] mizing the life cycle costs of the pipe not only depends on the length but also on its bendings, that can be incorpo- rated with an extended graph. An approach based on a Ali Taylan Cemgil convex reformulation that yields a mixed-integer-second- Bogazici University order-cone problem, and a decomposition algorithm that [email protected] can handle non-convex constraints that arise from regula- tions, are proposed. CP12 Jakob Schelbert,LarsSchewe Topology Optimization of Geometrically Nonlinear FAU Erlangen-N¨urnberg, Discrete Optimization Structures Department of Mathematics [email protected], [email protected] We present a Sequential Piecewise Linear Programming algorithm for topology optimization of geometrically non- linear structures, that require the solution of a nonlinear CP12 system each time the objective function is evaluated. The Hipad - A Hybrid Interior-Point Alternating Di- method relies on the solution of convex piecewise linear rection Algorithm for Knowledge-Based Svm and programming subproblems that include second order infor- Feature Selection mation. To speed up the algorithm, these subproblems are converted into linear programming ones. The new method We propose a new hybrid optimization algorithm that is globally convergent to stationary points and also shows solves the elastic-net SVM through an alternating direc- good numerical performance. tion method of multipliers in the first phase, followed by an interior-point method in the second. Our algorithm ad- Francisco M. Gomes dresses three challenges: a. high optimization accuracy, University of Campinas - Brazil b. automatic feature selection, and c. algorithmic flexi- Department of Applied Mathematics bility for including prior knowledge. We demonstrate the [email protected] effectiveness and efficiency of our algorithm and compare it with existing methods on a collection of synthetic and Thadeu Senne real-world datasets. Universidade Federal do Triangulo Mineiro Departamento de Matematica Aplicada Ioannis Akrotirianakis [email protected] Research Scientist at Siemens Corporate Research [email protected] CP12 Zhiwei Qin Nonlinear Material Architecture Using Topology 82 OP14 Abstracts

Optimization [email protected]

This paper proposes a computational topology optimiza- tion algorithm for tailoring the nonlinear response of pe- CP13 riodic materials through optimization of unit cell archi- Riemannian Optimization in Multidisciplinary De- tecture. Material design via topology optimization has sign Optimization been achieved in the past for elastic properties. We dis- cuss on the different nonlinear mechanisms in the optimiza- Standard gradient-based optimization algorithms are de- tion problem to compare results and suggest more powerful rived for Euclidean spaces. However, Multidisciplinary De- mechanisms. The challenges of topology optimization un- sign Optimization (MDO) problems exist on Riemannian der nonlinear constitutive equations include singularities manifolds defined by the state equations. As such, Rieman- and excessive distortions in the low stiffness space of the nian Optimization (RO) algorithms may be particularly ef- topology optimized design domain. Low stiffness and void fective for MDO; our differential geometry framework for elements are meant to represent holes in the final topol- MDO now makes it possible to apply these algorithms to ogy. We propose extending the logic of HPM to nonlin- MDO problems. Following a review of RO and our frame- ear mechanics by eliminating modeling of low-stiffness ele- work, we make qualitative and quantitative comparisons ments. Therefore we can remove the element from analysis between standard optimization algorithms and their Rie- part without removing that from sensitivity analysis. This mannian counterparts. method helps avoid numerical instability in the analysis in the optimization process. Craig K. Bakker, Geoffrey Parks Engineering Design Centre, Department of Engineering Reza Lotfi University of Cambridge Johns Hopkins University [email protected], [email protected] Johns Hopkins University reza.lotfi@gmail.com CP13 A Unified Approach for Solving Continuous Multi- CP12 facility Ordered Median Location Problems A Barycenter Method for Direct Optimization In this paper we propose a general methodology for solving a broad class of continuous, multifacility location problems The purpose of this paper is to advocate a direct optimiza- ≥ tion method based on the computation of the barycenter of with τ -norms (τ 1) proposing two different methodolo- a sequence of evaluations or measurements of a given func- gies: 1) by a new second order cone mixed integer pro- −μf(x ,t ) gramming formulation and 2) by formulating a sequence tion. The weight given to each point xi is e i i ,where · of semidefinite programs that converges to the solution of f( ) is the function to be optimized, xi is a point probed the problem; each of these relaxed problems solvable with at time ti,andμ is a suitable constant. We develop formu- SDP solvers in polynomial time. We apply dimensionality las that relate the derivative-free barycenter update rule reductions of the problems by sparsity and symmetry in to gradient-type methods. Extensions of note include opti- order to be able to solve larger problems. mization on Riemannian manifolds, and the use of complex weightings, deriving intuition from Richard Feynman’s in- Victor Blanco terpretation of quantum electrodynamics. Universidad de Granada [email protected] Felipe M. Pait, Diego Col´on Univ Sao Paulo Justo Puerto, Safae ElHaj BenAli Brazil Universidad de Sevilla [email protected], [email protected] [email protected], [email protected]

CP12 CP13 Distributed Benders Decomposition Method A Near-Optimal Dynamic Learning Algorithm for Online Matching Problems with Concave Returns The combined network design and (distributed) traffic under Random Permutations Model routing problem can be formulated as a large-scale multi- period mixed integer optimization problem. Its resolution We propose a near-optimal dynamic learning algorithm for on realistic instances is intractable and unscalable with online matching problem with concave objective functions. state-of-the-art solvers which ignore the distributed na- In this problem, the input data arrives sequentially and ture of routing. We decompose the global optimization decisions have to be made in an online fashion. Our algo- problem by means of a distributed version of the Benders rithm is of primal-dual type, which dynamically uses the decomposition method. Using this method, the initial op- optimal solutions to selected partial problems in assisting timization problem subdivides into a distributed master future decisions. This problem has important application problem solved at each node and several subproblems of in online problems such as the online advertisement allo- tractable size involving only local decisions when nodes cation problem. compute routing path(s) online. Xiao Chen,ZizhuoWang Dimitri Papadimitriou University of Minnesota, Twin Cities Bell Labs [email protected], [email protected] [email protected]

Bernard Fortz CP13 ULB A Scenario Based Optimization of Resilient Supply OP14 Abstracts 83

Chain with Dynamic Fortification Cost rect search (MADS) algorithms based on equal area hy- persphere partitionings introduced by Van Dyke (2013). Considering all possible disruption scenarios, the developed These nearly-orthogonal directional direct search instances stochastic optimization model looks for the most efficient of MADS provide better uniformity of polling directions solution to allocate the fortification budget and emergency relative to other MADS instances. We consider the rela- inventory while minimizing the expected shortage costs. tive the performance among several MADS instances on a The reliability of each supplier depends on the assigned variety of test problems. fortification budget and varies from no reliability to fully reliable. As the reliability of the supplier increases the Thomas Asaki fortification cost to improve its level of reliability will grow Washington State University dynamically. [email protected] Amirhossein Meisami Student CP14 [email protected] Coupling Surrogates with Derivative-Free Contin- uous and Mixed Integer Optimization Methods to Nima Salehi Solve Problems Coming from Industrial Aerody- PhD student at West Virginia University namic Applications [email protected] We focus on two DFO frameworks, Minamo and NO- MAD, developed to solve costly black-box (continuous or CP13 mixed-integer) problems coming from industrial applica- Linear Regression Estimation under Different Cen- tions. We compare the numerical performance of the soring Models of Survival Data surrogate-assisted evolutionary algorithm Minamo, a solver developed at Cenaero, with that of NOMAD, an implemen- In survival analysis, data are often censored and differ- tation of the mesh-adaptive pattern-search method. Our ent censoring patterns can be utilized to model the cen- goal is to improve the efficiency of both frameworks by ap- sored data. During this talk we will propose the methods propriately combining their attractive features in the con- of estimating parameters in the linear regression analysis text of aerodynamic applications. when the data are under different censoring models includ- ing the general right censoring model, the Koziol-Green Anne-Sophie Cr´elot model, and partial Koziol-Green model. Simulations to University of Namur compare the efficiencies of the estimators under these cen- Department of mathematics, naXys soring models are described. [email protected] Ke Wu Dominique Orban California State University, Fresno GERAD and Dept. Mathematics and Industrial [email protected] Engineering Ecole Polytechnique de Montreal CP13 [email protected] PDE-Constrained Optimization Using Hyper- Reduced Models Caroline Sainvitu Minamo team Leader, Cenaero The large computational cost associated with high-fidelity, [email protected] physics-based simulations has limited their use in many practical situations, particularly PDE-constrained opti- Annick Sartenaer mization. Replacing the high-fidelity model with a hyper- University of Namur reduced (physics-based, surrogate) model has been shown Department of Mathematics to generate a 400x speedup. A novel approach for in- [email protected] corporating hyper-reduced models in the context of PDE- constrained optimization will be presented and the technol- ogy demonstrated on a standard shape optimization prob- CP14 lem from computational fluid dynamics, shape design of a Parallel Hybrid Multiobjective Derivative-Free simplified rocket nozzle. Optimization in SAS

Matthew J. Zahr We present enhancements to a SAS high performance pro- Stanford University cedure for solving multiobjective optimization problems University of California, Berkeley in a parallel environment. The procedure, originally de- [email protected] signed as a derivative-free solver for mixed-integer nonlin- ear black-box single objective optimization, has now been Charbel Farhat extended for multiobjective problems. In the multiobjec- Stanford University tive case the procedure returns an approximate Pareto- [email protected] optimal set of nondominated solutions to the user. We will discuss the software architecture and algorithmic changes made to support multiobjective optimization and provide CP14 numerical results. Deterministic Mesh Adaptive Direct Search with Uniformly Distributed Polling Directions Steven Gardner SAS Institute Inc. We consider deterministic instances of mesh adaptive di- [email protected] 84 OP14 Abstracts

Joshua Griffin Oleg Makhnin SAS Institute, Inc. New Mexico Institute of Mining and Technology joshua.griffi[email protected] [email protected]

Lois Zhu SAS Institute Inc. CP15 [email protected] Robust Optimization of a Class of Bi-Convex Func- tions

CP14 Robust optimization is a methodology that has gain a lot A Derivative-Free Method for Optimization Prob- of attention in the recent years. This is mainly due to lems with General Inequality Constraints the simplicity of the modeling process and ease of resolu- tion even for large scale models. Unfortunately, the second In this work, we develop a new derivative-free algorithm to property is usually lost when the function that needs to solve optimization problems with general inequality con- be ”robustified” is not concave (or linear) with respect to straints. An active-set method combined with trust-region the perturbed parameters. In this work, we propose a new techniques is proposed. This method is based on polyno- scheme for constructing safe tractable approximations for mial interpolation models, making use of a self-correcting the robust optimization of a class of bi-convex functions; geometry scheme to maintain the quality of geometry of specifically, those that decompose as the sum of maximum the sample set. Initial numerical results are also presented of bi-affine functions (with respect to the decisions and the at last. perturbation). Such functions arise very naturally in a wide range of application, including news-vendor, inventory, and Phillipe R. Sampaio classification problems and multi-attribute utility theory. University of Namur We will present conditions under which our approximation [email protected] scheme is exact. Some empirical results will illustrate the performance that can be achieved. Philippe L. Toint Fac Univ Notre Dame de la Paix Erick Delage Department of Mathematics HEC Montr´eal [email protected] [email protected] Amir Ardestani-Jaafari CP14 GERAD, HEC Montr´eal An SQP Trust-Region Algorithm for Derivative- [email protected] Free Constrained Optimization

We want to propose a new trust-region model-based al- CP15 gorithm for solving nonlinear generally constrained opti- Tractable Robust Counterparts of Distributionally mization problems where derivatives of the function and Robust Constraints on Risk Measures constraints are not available. To handle the general con- straints, an SQP method is applied. The bound constraints In optimization problems risk measures of random vari- are handled by an active-set approach as was proposed in ables often need to be kept below some fixed level, whereas [Gratton et al., 2011]. Numerical results are presented on a knowledge about the underlying probability distribution is set of CUTEr problems and an application from engineer- limited. In our research we apply Fenchel duality for ro- ing design optimization. bust optimization to derive computationally tractable ro- bust counterparts of such constraints for commonly used Anke Troeltzsch risk measures and uncertainty sets for probability distribu- CERFACS tions. Results may be used in various applications such as Toulouse, France portfolio optimization, economics and engineering. [email protected] Krzysztof Postek, Bertrand Melenberg Tilburg University CP15 [email protected], [email protected] Medication Inventory Management in the Interna- tional Space Station (ISS): A Robust Optimization Dick Den Hertog Approach Tilburg University Faculty of Economics and Business Administration Medication inventory management in space is highly com- [email protected] plex, performed in resource-constrained, highly-uncertain settings. Shortage of drugs during exploration missions may result in detrimental impacts. In space, the shelf space CP15 is limited; replenishment may not be possible; and the shelf Robust Approaches for Stochastic Intertemporal life and stability of medications are uncertain due to the Production Planning unique space environment. This research explores the po- tential of the Robust Optimization approach to provide the In many industries, production planning is done in different robust inventory (no or minimal drugs shortages) strategy. time horizons: tactical and operational, and several incon- sistencies between them arise due to data aggregation and Sung Chung uncertainty. We model this problem as 2-stages stochas- New Mexico Institute of Mining and Technology tic problem, for which we have used Robust Optimization, New Mexico Tech to estimate the degree of robustness needed and also some [email protected] modifications to a Stochastic Mirror Descent method, as OP14 Abstracts 85

well as some other first order methods. We show results [email protected] from a real industrial problem.

Jorge R. Vera,AlfonsoLobos CP16 Pontificia Universidad Cat´olica de Chile Optimal Cur Matrix Decompositions [email protected], [email protected] Given a matrix A ∈ Rm×n and integers c

CP15 Vishnu Narayanan Assistant professor, Industrial Engineering and Robust Solutions for Systems of Uncertain Linear Operations Equations Research, Indian Institute of Technology Bombay [email protected] We develop a new way for obtaining robust solutions of a system of uncertain linear equations, where the parame- ter uncertainties are column-wise. Firstly, we construct a CP16 convex representation of the solution set. We then employ Ellipsoidal Rounding for Nonnegative Matrix Fac- adjustable robust optimization to approximate the robust torization Under Noisy Separability solution, i.e., the center of the maximum volume inscribed ellipsoid only with respect to a subset of variables. Finally, We present a numerical algorithm for nonnegative matrix our method is applied to find robust versions of Google’s factorization (NMF) problems under noisy separability. An PageRank and Colley’s Rankings. NMF problem under separability can be stated as one of finding all vertices of the convex hull of data points. The Jianzhe Zhen interest of this study is to find the vectors as close to the Tilburg University vertices as possible in a situation where noise is added to [email protected] the data points. We show that the algorithm is robust to noise from theoretical and practical perspectives. Dick Den Hertog Tilburg University Tomohiko Mizutani Faculty of Economics and Business Administration Kanagawa University 86 OP14 Abstracts

[email protected] problems. Though our model is highly non-convex, I will show that our method performs reliably well on a wide range of low-rank tensors including both synthetic data CP16 and real-world 3D images and videos. A Semidefinite Approximation for Symmetric Travelling Salesman Polytopes Yangyang Xu Rice University A. Barvinok and E. Veomett had introduced a hierarchy [email protected] of approximations of a convex set via projections of spec- trahedra. The main results of this talk are quantitative measurements on the quality of these approximations for CP17 the symmetric travelling salesman polytopes Tn and Tn,n Design and Operating Strategy Optimization of associated to the complete graph Kn and the complete bi- Wind, Diesel and Batteries Hybrid Systems for Iso- partite graph Kn,n respectively. By a result of Veommett it lated Sites is known that scaling the k−th approximation by the factor n/k + O(1/n) contains the polytope Tn for k ≤n/2.We Simultaneous optimization of hybrid electricity production systems (HEPS) design and operating strategy for isolated show that these metric bounds can be improved by a fac- 1 k−1 2 sites is a real challenge. Our approach: Off Grid Electricity tor of 3 n −1 + 3 . Moreover, we give new metric bounds 2 Supply Optimization (OGESO) is based on a mixed inte- for the k-th spectrahedral approximation of the travelling ger programing model to find HEPS optimal design and salesman polytope Tn,n. These results are joint work with optimal one year hourly power dispatch with deterministic M. Velasco. data. This hourly dispatch is then analyzed by data min- ing to find an appropriate operating strategy. First results Juli´an Romero show benefits of using OGESO rather than simulation. Student [email protected] Thibault Barbier Ecole´ Polytechnique de Montr´eal Mauricio Velasco [email protected] University of los Andes [email protected] Gilles Savard Ecole Polytechnique Montreal [email protected] CP16 On Solving An Application Based Completely Pos- Miguel Anjos itive Program of Size 3 Ecole´ Polytechnique de Montr´eal [email protected] In this talk, I will discuss a (dual) completely positive pro- gram, with matrices involved to be of size 3, that arises from a supply chain problem. This (dual) completely pos- CP17 itive program is solved completely to obtain a closed form A Mixed Integer Programming Approach for the solution. The supply chain problem will also be discussed Police Districting Problem in the talk. We have developed a multiobjective mixed integer pro- Qi Fu gramming approach for a districting problem based on the University of Macau Carabineros de Chile’s case. This approach has shape con- [email protected] siderations and balance constraints between the districts. This problem is difficult to solve for small instances (50 Chee-Khian Sim nodes). We solve this problem with a Location Allocation National University of Singapore Heuristic and we can found a lot of feasibles solutions with [email protected] good properties in few seconds for realistics size instances (over 400 nodes). Chung-Piaw Teo Department of Decision Sciences V´ıctor D. Bucarey, Fernando Ordo˜nez NUS Business School Universidad de Chile [email protected] Departamento de Ingenieria Industrial [email protected], [email protected]

CP16 Vladimir Marianov Parallel Matrix Factorization for Low-Rank Tensor Pontificia Universidad Cat´olica de Chile Recovery [email protected]

Multi-way data naturally arises in applications, among which there are many with missing values or compressed CP17 data. In this talk, I will present a parallel matrix factor- Mixed Integer Linear Programming Approach to a ization method for low-rank recovery. Existing methods ei- Rough Cut Capacity Planning in Yogurt Produc- ther unfold the tensor to a matrix and then apply low-rank tion matrix recovery, or employ nuclear norm minimization to each mode unfolding of the tensor. The former methods A rough cut capacity plan is important to meet weekly are often inefficient since they only utilize one mode low- demand of multiple stock-keeping units (SKU) of perish- rankness of the tensor, while the latter ones are usually able products like yogurt. We propose a mixed integer slow due to expensive SVD. We address both these two linear programming (MILP) model, that create a capacity OP14 Abstracts 87

plan considering the relevant process parameters, capac- [email protected] ity and processing times of the different production stages while incorporating relevant constraints in the filling and packaging processes. The model helps identifying the bot- CP17 tleneck processes and in exploring possible alternatives to Rectangular Shape Management Zone Delineation meet demand. in Agriculture Planning by Using Column Genera- tion Sweety Hansuwa Research Assistant, Industrial Engineering and We consider an optimization model for the delineation of a Operations field into rectangular-shape site-specific management zones Research, Indian Institute of Technology Bombay and propose a column generation method for solving it. [email protected] At each iteration of the algorithm, the subproblem verifies the optimality condition of the solution proposed by the Pratik Fadadu reduced master problem, or provides a new set of feasible An operations research modeler, zones. This master problem gives a partition of the field IGSA Labs Pvt. Ltd, Hyderabad minimizing the number of zones, while maximizing the ho- [email protected] mogeneity within this zones. Linco J. anco,V´ıctor Albornoz Atanu Chaudhuri Universidad T´ecnica Federico Santa Mara Assistant Professor, [email protected], vic- Indian Institute of Management, Lucknow [email protected] [email protected]

Rajiv Kumar Srivastava CP18 Professor, Optimal Scenario Set Partitioning for Multistage Indian Institute of Management, Lucknow Stochastic Programming with the Progressive [email protected] Hedging Algorithm

We propose a new approach to construct scenario bun- CP17 dles that aims to enhance the convergence rate of the pro- gressive hedging algorithm. We apply a heuristic to par- Solution Methods for Mip Models for Chemical tition the scenario set in order to minimize the number Production Scheduling of non-anticipativity constraints on which an augmented Lagrangian relaxation must be applied. The proposed We discuss three solution methods for MIP models for method is tested on an hydroelectricity generation schedul- chemical production scheduling. First, we develop a prop- ing problem covering a 52-week planning horizon. agation algorithm to calculate bounds on the number of batches needed to meet demand and valid inequalities Fabian Bastin based on these bounds. Second, we generate models that University of Montreal employ multiple time grids, thereby substantially reduc- [email protected] ing the number of binaries. Third, we propose reformula- tions which lead to more efficient branching. The proposed Pierre-Luc Carpentier methods reduce solution times by several orders of magni- Ecole Polytechnique de Montr´eal tude. [email protected]

Sara Velez, Christos Maravelias Michel Gendreau University of Wisconsin Universit´edeMontr´eal [email protected], [email protected] [email protected]

CP17 CP18 A Modified Sample Approximation Method for Power Efficient Uplink Scheduling in SC-FDMA: Chance Constrained Problems Bounding Global Optimality by Column Genera- tion In this talk, we present a modified sample approximation (SA) method to solve the chance constrained problems. We study resource allocation in cellular systems and con- We show that the modified SA problem has similar con- sider the problem of finding a power efficient scheduling in vergence properties as the previous SA approximation one. an uplink single carrier frequency division multiple access Moreover, the modified SA problem is convex under some (SC-FDMA) system with localized allocation of subcarri- conditions. Finally, numerical experiments are conducted ers. An integer linear programming and column-oriented to illustrate the strength of the modified SA approach. formulation is provided. The computational evaluation demonstrates that the column generation method produces Jianqiang Cheng, Celine Gicquel, Abdel Lisser very high-quality subcarrier allocations that either coin- LRI, University of Paris-Sud, France cide with the global optimum or enable an extremely sharp [email protected], [email protected], [email protected] bounding interval.

Hongmei Zhao CP18 Mathematics Department, Optimisation Division Multiple Decomposition and Cutting-Plane Gen- Linkoping University, Sweden erations for Optimizing Allocation and Scheduling 88 OP14 Abstracts

with a Joint Chance Constraint associated with nonanticipativity constraints. In this work, we investigate its connection with the developments in aug- We study a class of stochastic resource planning problems mented Lagrangian methods. We consider linear penalty where we integrate resource allocation and job scheduling terms in order to preserve linearity when present in the decisions under uncertain job processing time. We min- original problem, and evaluate the numerical performance imize the number of resource units allocated, subject to on various test problems. a low risk of resource use overtime modeled as a joint chance constraint. We formulate a mixed-integer program Shohre Zehtabian based on discrete samples of the random processing time. Computer Science and Operations Research Department We develop a multi-decomposition algorithm that first de- University of Montreal composes allocation and scheduling into two stages. We [email protected] strengthen the first stage by adding a joint chance con- straint related to the original one, and implement a cutting- Fabian Bastin plane algorithm using scenario-wise decomposition. At the University of Montreal second stage, we implement a resource-wise decomposition [email protected] to verify the feasibility of the current first-stage solution. We test instances of operating room allocation and schedul- ing generated based on real hospital data, and show very CP19 promising results of our multi-decomposition approach. First-Order Methods Yield New Analysis and Re- Yan Deng, Siqian Shen sults for Boosting Methods in Statistics/Machine Department of Industrial and Operations Engineering Learning University of Michigan [email protected], [email protected] Using mirror descent and other first-order convex optimiza- tion methods, we present new analysis, convergence re- sults, and computational guarantees for several well-known CP18 boosting methods in statistics, namely AdaBoost, Incre- Uncertainty Quantification and Optimization Us- mental Forward Stagewise Regression (FSε), and their vari- ing a Bayesian Framework ants. For AdaBoost and FSε, we show that a minor variant of these algorithms corresponds to the Frank-Wolfe method This paper presents a methodology that combines uncer- on a constraint-regularized problem, whereby the modified tainty quantification, propagation and robustness-based procedures converge to optimal solutions at an O(1/k)rate. design optimization. Epistemic uncertainty regarding model inputs/parameters is efficiently propagated using an auxiliary variable approach. The uncertainty quantifica- Robert M. Freund tion result is used to inform design optimization to im- MIT prove the systems robustness and reliability. The proposed Sloan School of Management methodology is demonstrated using a challenge problem [email protected] developed by NASA.

Chen Liang Paul Grigas Vanderbilt University M.I.T. [email protected] Operations Research Center [email protected]

CP18 Rahul Mazumder Application of Variance Reduction to Sequential Columbia University Sampling in Stochastic Programming Department of Statistics [email protected] We apply variance reduction techniques, specifically anti- thetic variates and Latin hypercube sampling, to optimal- ity gap estimators used in sequential sampling algorithms. CP19 We discuss both theoretical and computational results on a range of test problems. Design Optimization with the Consider-Then- Choose Behavioral Model Rebecca Stockbridge Wayne State University The consider-then-choose model describes a decision- [email protected] making process in which consumers first eliminate a large number of product alternatives with non-compensatory Guzin Bayraksn screening rules, subsequently performing careful tradeoff The Ohio State University evaluation over the remaining alternatives. Modeling con- [email protected] sideration introduces discontinuous choice probabilities to optimization problems using choice probabilities as a de- mand model. We compare several treatments of this CP18 discontinuous optimization problem: genetic algorithms, A New Progressive Hedging Algorithm for Linear smoothing, and relaxation via complementary constraints. Stochastic Optimization Problems The application of methods is illustrated through a vehicle design example. Progressive Hedging Algorithm, while introduced more than twenty years ago, remains a popular method deal- Minhua Long, W. Ross Morrow, Erin MacDonald ing with multistage stochastic problems. The performance Iowa State University can be poor due to the number of quadratic penalty terms [email protected], [email protected], erin- OP14 Abstracts 89

[email protected] [email protected]

CP19 CP19 A Multilevel Approach for L-1 Regularized Con- Numerical Optimization of Eigenvalues of Hermi- vex Optimization with Application to Covariance tian Matrix-Valued Functions Selection

We present numerical approaches for (i) unconstrained We present a multilevel framework for solving l-1 regular- optimization of a prescribed eigenvalue of a Hermitian ized convex optimization problems, which are widely pop- matrix-valued function depending on a few parameters an- ular in the fields of signal processing and machine learning. alytically, and (ii) an optimization problem with linear ob- Such l-1 regularization is used to find sparse minimizers of jective and an eigenvalue constraint on the matrix-valued the convex functions. The framework is used for solving function. The common theme for numerical approaches the Covariance Selection problem, where a sparse inverse is the use of global under-estimators, called support func- covariance matrix is estimated from a only few samples of tions, for eigenvalue functions, that are built on the ana- a multivariate normal distribution. Numerical experiments lyticity and variational properties of eigenvalues over the demonstrate the potential of this approach, especially for space of Hermitian matrices. large-scale problems.

Emre Mengi Eran Treister,IradYavneh Koc University, Turkey Computer Science Department, Technion [email protected] [email protected], [email protected]

Alper Yildirim, Mustafa Kilic CP20 Koc University [email protected], [email protected] Improving Fast Dual Gradient Methods for Repet- itive Multi-Parametric Programming We show how to combine the low iteration cost of first or- CP19 der methods with the fast convergence rate of second order Finite Hyperplane Traversal Algorithms for 1- methods, when solving the dual to composite optimization Dimensional L1pT V Minimization for 0

Heather A. Moon CP20 Washington State University Sliding Mode Controller Design of Linear Interval [email protected] Systems Via An Optimised Reduced Order Model This paper presents a method of designing the Sliding CP19 Mode Controller for large scale Interval systems. The Con- A Feasible Second Order Bundle Algorithm for troller is designed via an optimised stable reduced order Nonsmooth, Nonconvex Optimization Problems model. The proposed reduction method makes use of Inter- polation criteria. It has been shown that the Sliding Mode Controller designed for the reduced order model, when ap- We extend the SQP-approach of the well-known bundle- plied to the higher order system, improves the performance Newton method for nonsmooth unconstrained minimiza- of the controlled system. Some numerical examples are tion to the nonlinearly constrained case. Instead of using a tested and found satisfactory. penalty function or a filter or an improvement function to deal with the presence of constraints, the search direction is Raja Rao Guntu determined by solving a convex quadratically constrained Anil Neerukonda Institute of Technology & Sciences quadratic program to obtain good iteration points. Issues [email protected] arising with this procedure concerning the bundling process as well as in the line search are discussed. Furthermore, we Narasimhulu Tammineni show global convergence of the method under certain mild Avanthi Institute of Engineering & Technology assumptions. [email protected] Hermann Schichl University of Vienna CP20 Faculty of Mathematics Moreau-Yosida Regularization in Shape Optimiza- [email protected] tion with Geometric Constraints Hannes Fendl We employ the method of mappings to obtain an optimal Bank Austria control problem with nonlinear state equation on a refer- 90 OP14 Abstracts

ence domain. The Lagrange multiplier associated with the [email protected] geometric shape constraint has low regularity, which we cir- cumvent by penalization and a continuation scheme. We employ a Moreau-Yosida-type regularization and assume CP20 a second-order condition. We study the properties of the Optimal Actuator and Sensor Placement for Dy- regularized solutions, which we obtain by a semismooth namical Systems Newton method. The theoretical results are supported by numerical tests. Vibrations occur in many areas of industry and produce un- desirable side effects. To avoid respectively suppress these Moritz Keuthen effects actuators and sensors are attached to the structure. Technische Universit¨at M¨unchen The appropriate positioning of actuators and sensors is of [email protected] significant importance for the controllability and observ- ability of the structure. In this talk, a method for deter- Michael Ulbrich mining the optimal actuator and sensor placement is pre- Technische Universitaet Muenchen sented, which leads to an optimization problem with binary Chair of Mathematical Optimization and continuous variables and linear matrix inequalities. [email protected] Carsten Sch¨afer TU Darmstadt CP20 [email protected]

Intelligent Optimizing Controller: Development of Stefan Ulbrich a Fast Optimal Control Algorithm Technische Universitaet Darmstadt Fachbereich Mathematik In this paper we propose a novel algorithm to solve a trajec- [email protected] tory optimization problem. The proposed algorithm does not need an explicit dynamic model of the system but com- putes partial derivatives of dynamic function numerically CP21 from observation history to estimate the adjoint variable and the Hamiltonian. A candidate optimal control tra- Some Mathematical Properties of the Dynamically jectory is computed such that the resulting Hamiltonian Inconsistent Bellman Equation: A Note on the is constant, which is a necessary condition for optimality. Two-Sided Altruism Dynamics Simulations are used to test the algorithm. This article describes some dynamic aspects on dynas- tic utility incorporating two-sided altruism with an OLG Jinkun Lee,VittalPrabhu setting. The dynamically inconsistent Bellman equation The Pennsylvania State University proves to be reduced to one-sided dynamic problem, but [email protected], [email protected] the effective discount factor is different only in the current generation. It is shown that a contraction mapping result of value function cannot be achieved in general, and that CP20 there can locally exist an infinite number of self-consistent policy functions with distinct steady states. Sensitivity-Based Multistep Feedback MPC: Algo- rithm Design and Hardware Implementation Takaaki Aoki Kyoto University, Institute of Economic Research In model predictive control (MPC), an optimization prob- [email protected] lem is solved every sampling instant to determine an op- timal control for a physical system. We aim to accelerate this procedure for fast systems applications and address CP21 the challenge of implementing the resulting MPC scheme on an embedded system. We present the sensitivity-based Characterization of Optimal Boundaries in Re- multistep feedback MPC a strategy which significantly re- versible Investment Problems duces the time, cost and energy consumption in computing control actions while fulfilling performance expectations. This paper studies a reversible investment problem where a social planner aims to control its capacity production in Vryan Gil Palma order to fit optimally the random demand of a good. Our University of Bayreuth, Germany model allows for general diffusion dynamics on the demand [email protected] as well as general cost functional. The resulting optimiza- tion problem leads to a degenerate two-dimensional singu- lar stochastic control problem, for which explicit solution Andrea Suardi is not available in general and the standard verification Imperial College London approach can not be applied a priori. We use a direct Electrical and Electronic Engineering viscosity solutions approach for deriving some features of [email protected] the optimal free boundary function, and for displaying the structure of the solution. In the quadratic cost case, we are Eric C. Kerrigan able to prove a smooth-fit C2 property, which gives rise to Imperial College London a full characterization of the optimal boundaries and value [email protected] function.

Lars Gruene Salvatore Federico Mathematical Institute, University of Bayreuth University of Milan OP14 Abstracts 91

[email protected] ization of Linear-Quadratic Control Problems with Bang-Bang Solutions

CP21 We consider linear–quadratic “state regulator problems’ Numerical Methods and Computational Results for with “Bang–Bang’ solutions. In order to solve these prob- Solving Hierarchical Dynamic Optimization Prob- lems numerically, we analyze a combined regularization– lems discretization approach. By choosing the regularization pa- rameter α with respect to the mesh size h of the discretiza- We present numerical methods for hierarchical dynamic tion we approximate the optimal solution. Under the as- optimization problems with a regression objective on the sumption that the optimal control has bang–bang struc-√ upper level and a nonlinear optimal control problem in or- ture we improve existing error estimates of order O( h) dinary differential equations with mixed control-state con- to O(h). Moreover we will show that the discrete and the straints and discontinuous transitions on the lower level. continuous controls coincide except on a set of measure Numerical results from a recent benchmarking study are O(h). provided. Furthermore, our hierarchical dynamic optimiza- tion methods are used to identify gait models from real- Martin Seydenschwanz world data of the Orthopaedic Hospital Heidelberg that Friedrich Schiller University of Jena are used for treatment planning. Germany [email protected] Kathrin Hatz,JohannesP.Schl¨oder Interdisciplinary Center for Scientific Computing (IWR) University of Heidelberg, Germany CP22 [email protected], Optimal Subgradient Algorithms for Large-Scale [email protected] Structured Convex Optimization

Hans Georg Bock Optimal subgradient algorithms for solving wide classes of IWR, University of Heidelberg large-scale structured convex optimization are introduced. [email protected] Our framework considers the first-order oracle and pro- vides a novel fractional subproblem which can be explicitly solved by employing appropriate prox-functions. Conver- CP21 gence analysis certifies attaining first-order optimal com- Solution of Optimal Control Problems by Hybrid plexity for smooth and nonsmooth convex problems. Com- Functions prehensive numerical experiments on problems in signal and image processing and machine learning including com- In this talk, a numerical method for solving the nonlinear parisons with state-of-the-art solvers are reported. constrained optimal control with quadratic performance in- dex is presented. The method is based upon hybrid func- Masoud Ahookhosh tions approximation. The properties of hybrid functions College assistant, Faculty of Mathematics, consisting of block-pulse functions and Bernoulli polyno- University of Vienna, mials are presented. The numerical solutions are compared [email protected] with available exact or approximate solutions in order to assess the accuracy of the proposed method. Arnold Neumaier University of Vienna Somayeh Mashayekhi Department of Mathematics Mississippi state university [email protected] [email protected]

CP22 CP21 The Block Coordinate Descent Method of Multi- A Dual Approach on Solving Linear-Quadratic pliers Control Problems with Bang-Bang Solutions In this paper, we consider a (possibly nonsmooth) con- We use Fenchel duality to solve a class of linear-quadratic vex optimization problem with multiple blocks of variables optimal control problems for which we assume a bang-bang and a linear constraint coupling all the variables. We pro- solution. We determine that strong duality holds in this pose an algorithm called block coordinate descent method case. In numerical experiments we then compare the solu- of multipliers (BCDMM) to solve this family of problems. tion of both the dual and the original problem and figure The BCDMM is a primal-dual type of algorithm. It inte- out if computational savings can be obtained. The dis- grates the traditional block coordinate decent (BCD) algo- cretization of these problems will also be discussed. rithm and the alternating direction method of multipliers (ADMM), in which it optimizes the (approximated) aug- Christopher Schneider,WalterAlt mented Lagrangian of the original problem one block vari- Friedrich-Schiller-University of Jena able each time, followed by a gradient update for the dual [email protected], [email protected] variable. Under certain regularity conditions, and when the order for which the block variables are either updated Yalcin Kaya in a deterministic way or in a random fashion, we show that University of South Australia the BCDMM converges, under either a diminishing step- [email protected] size rule or a small enough stepsize for the dual update.

CP21 Mingyi Hong Improved Error Estimates for Discrete Regular- University of Minnesota 92 OP14 Abstracts

[email protected] Paul Armand Universit´edeLimoges Laboratoire XLIM CP22 [email protected] A Unified Framework for Subgradient Algorithms Minimizing Strongly Convex Functions CP22 We propose a unified framework for (sub)gradient algo- rithms for both non-smooth and smooth convex optimiza- A New Algorithm for Least Square Problems Based tion problems whose objective function may be strongly on the Primal Dual Augmented Lagrangian Ap- convex. Our algorithms have as particular cases the mirror- proach descent, Nesterov’s dual-averaging and variants of Tseng’s accelerated gradient methods. We exploit Nesterov’s idea We propose a new algorithm for least square problems to analyze those methods in a unified way. This frame- based on the Primal Dual Augmented Lagrangian approach work also provides some variants of the conditional gradi- of Gill and Robinson. This talk will explore the effects of ent method. the regularization term on constant objective problems and suggest strategies for improving convergence to the nonlin- Masaru Ito ear constraint manifold in the presence of general objec- Tokyo Institute of Technology tives. Theoretical and numerical results will be given. [email protected] Wenwen Zhou Mituhiro Fukuda SAS Institute Inc. Tokyo Institute of Technology [email protected] Department of Mathematical and Computing Sciences [email protected] Joshua Griffin SAS Institute, Inc. CP22 joshua.griffi[email protected] The Exact Penalty Map for Nonsmooth and Non- convex Optimization CP23 Augmented Lagrangian duality provides zero duality gap A Three-objective Mathematical Model for One- and saddle point properties for nonconvex optimization. dimensional Cutting and Assortment Problems Hence, subgradient-like methods can be applied to the (convex) dual of the original problem, recovering the op- timal value of the problem, but possibly not a primal so- In this talk, a new three-objective linear integer program- lution. We prove that the recovery of a primal solution ming model without the use of a set of cutting patterns by such methods can be characterized in terms of the dif- is developed. The objectives are related to the trim loss ferentiability properties of the dual function and the exact amount, the total number of different standard lengths penalty properties of the primal-dual pair. used, and the production amount that exceeds the given demand for each cutting order. The advantages of the pro- Alfredo N. Iusem posed mathematical model are demonstrated on test prob- IMPA, Rio de Janeiro lems. [email protected] Nergiz Kasimbeyli Regina Burachik Anadolu University University of Southern Australia [email protected] [email protected]

Jefferson Melo CP23 Universidade Federal de Goias Nonlinear Programming in Runtime Procedure of jeff[email protected] Guidance, Navigation and Control of Autonomous Systems CP22 We redefine a runtime procedure of guidance, navigation A Primal-Dual Augmented Lagrangian Method for and control of autonomous systems as constraint program- Equality Constrained Minimization ming. We build a list of nonlinear equality constraints We propose a primal-dual method for solving equality con- that corresponds to nonlinear dynamics of the system un- strained minimization problems. This is a Newton-like der control, and a list of linear inequality constraints that method applied to a perturbation of the optimality system specifies action commands subject to safety constraints and that follows from a reformulation of the initial problem operational capability. Violations to the safety constraints by introducing an augmented Lagrangian to handle equal- and the constraints regarding the action are observed as ity constraints. An important aspect of this approach is hazardous action and loss of control, respectively. We re- that the algorithm reduces asymptotically to a regularized port an issue of execution time that primal-dual interior Newton method applied to KKT conditions of the original point method IPOPT determines infeasible when there ex- problem. The rate of convergence is quadratic. ists any input vectors that satisfy all constraints.

Riadh Omheni Masataka Nishi University of Limoges (France) - XLIM Laboratory Hitachi Research Laboratory [email protected] Hitachi Ltd OP14 Abstracts 93

[email protected] Criterion Evaluation

In many industrial fields, computer simulations of dura- CP23 bility are increasing in importance. In such simulations, critical plane criteria are often used to determine the prob- Strong Formulations for Stackelberg Security ability of fatigue crack initiation in a point. This type of Games criterion involves searching, in each material point, for the Recent work in security applications uses Stackelberg plane where the combination of stresses is the most dam- games to formulate the problem of protecting critical in- aging. This talk concerns new efficient ways to evaluate frastructure from strategic adversaries. In this work we ex- this type of criterion, which are based on the use of opti- plore different integer programming formulations for these mization methods, shortening the lead time and increasing Stackelberg security games. Our results show that 1) there the accuracy of the results. is a tight linear programming formulation in the case of Henrik Sv¨ard one adversary; 2) this formulation exhibits a smaller inte- Department of Mathematics grality gap in general; 3) this formulation does not work KTH Royal Institute of Technology with payoff structures common in security applications. [email protected] Fernando Ordonez USC, Industrial and Systems Engineering CP24 [email protected] Optimization of Radiotherapy Dose-Time Fraction- ation of Proneural Glioblastoma with Considera- tion of Biological Effective Dose for Early and Late CP23 Responding Tissues Studying Two Stage Games by Polynomial Pro- gramming We examined the optimal fractionation problem in the con- text of a novel model for radiation response in brain tumors We are investigating two stage games, such as those stud- recently developed in Leder et. al. (in press Cell), where ied by Kreps and Scheinkman (1983). Two profit maximiz- the radiation response of stem-like and tumor bulk cells ing players choose simultaneously their capacity in a first are considered separately. We find the minimal number stage. In the second stage there is a Bertrand-like price of tumor cells at the conclusion of treatment while keep- competition. In the original paper this produced a unique ing the toxicity in early and late responding tissues at an equilibrium, namely the Cournot outcome. We alter some acceptable level. This is carried out using non-linear pro- of the functions used to specify the problem. Doing so we gramming methodology loose many of the nice properties exhibited by the origi- nal problem. This makes a numerical treatment necessary. Hamidreza Badri,KevinLeder Using the first order conditions for the second stage game University of Minnesota does not work in this case, since the final system of KKT [email protected], [email protected] conditions might not even have the real equilibrium as a solution. We use a Positivstellensatz from real algebraic geometry to reformulate the second stage game. We then CP24 solve the resulting game with standard methods form non- Robust Optimization Reduces the Risks Oc- linear optimization. curring from Delineation Uncertainties in HDR Brachytherapy for Prostate Cancer Philipp Renner Department of Business Administration Delineation inaccuracy is a widely recognized phenomenon Universitaet Zuerich in HDR brachytherapy, though with unknown effects on [email protected] dosimetry. We show that these uncertainties yield a high risk of undesirably low target coverage, resulting in insuffi- cient tumor control. The classical solution, using a margin, CP23 results in overdosage. Therefore, robust optimization tech- niques are used for taking these uncertainties into account Multi-Agent Network Interdiction Games when optimizing treatment plans. The model shows a cru- cial improvement in target coverage, without the risk of Network interdiction games involve multiple leaders who overdosage. utilize limited resources to disrupt the network operations of adversarial parties. We formulate a class of such games Marleen Balvert,DickDenHertog as generalized Nash equilibrium problems. We show that Tilburg University these problems admit potential functions, which enables us [email protected], to design algorithms based on best response mechanisms [email protected] that converge to equilibria. We propose to study the in- efficiency of equilibria in such games using a decentralized Aswin Hoffmann algorithm that enables efficient computation. Maastro Clinic aswin.hoff[email protected] Harikrishnan Sreekumaran, Andrew L. Liu School of Industrial Engineering Purdue University CP24 [email protected], [email protected] An Re-Optimization Approach for the Train Dis- patching Problem (tdp) CP23 The TDP is to schedule trains through a network in a cost Using Optimization Methods for Efficient Fatigue optimal way. Due to disturbances during operation exist- 94 OP14 Abstracts

ing track schedules often have to be re-scheduled and inte- in Radiation Therapy grated into the timetable. This has to be done in seconds and with minimal timetable changes to guarantee conflict Two important elements of a radiation therapy treatment free operation. We present an integrated modeling ap- plan are quality, which depends on dose to organs at risk proach for the re-optimization task using Mixed Integer (OARs), and delivery time, which depends on the num- Programming and provide computational results for sce- ber of beams. We propose a technique to find the plan narios deduced from real world data. with the minimum number of beams and a predetermined maximum deviation from the ideal plan, which uses all Boris Grimm,TorstenKlug,ThomasSchlechte candidate beams. We use mixed integer programming for Zuse Institute Berlin (ZIB) optimization and two heuristics to reduce the computation [email protected], [email protected], [email protected] time.

Hamed Yarmand CP24 Massachusetts General Hospital Optimizing Large Scale Concrete Delivery Prob- Harvard Medical School lems [email protected]

In Ready Mixed Concrete (RMC) dispatching, common David Craft practice is to rely on human experts for any necessary real- Massachusetts General Hospital time decision-making. This is due to the perceived com- [email protected] plexity of RMC dispatching and the general lack of highly applicable optimization tools for the task. Critically, the accuracy of expert decisions (as compared to optimization CP25 approaches) has not been comprehensively examined in the On Estimating Resolution Given Sparsity and Non- literature. To address the question of current-practice ex- Negativity pert accuracy in the context of optimized outcomes, this paper first models the RMC dispatching problem mathe- We consider imaging problems where additional informa- matically according to the introduced methods in the lit- tion such as sparsity or non-negativity may be used to re- erature. Two approaches are taken, which are integer pro- construct the image to a higher resolution than possible gramming (without time windows) and mixed integer pro- without such information. Just as optimization techniques gramming (with time windows). Further, the constructed are required to find the improved image estimate, opti- models are tested with field data and compared with the mization must also be used to describe the true system decisions as made by experts. The results show that on performance. We discuss linear and conic programming average experts decisions are 90% accurate in comparison problems that estimate the data-dependent performance of to the examined optimization models. a system given a particular image, and provide insight into the interpretation of the result in underdetermined cases. Mojtaba Maghrebi The University of New SOuth Wales(UNSW) Keith Dillon, Yeshaiahu Fainman [email protected] University of California, San Diego [email protected], [email protected] Travis Waller, Claude Sammut The University of New South Wales(UNSW) [email protected], [email protected] CP25 Some Insights from the Stable Compressive Prin- cipal Component Pursuit CP24 Stochastic Patient Assignment Models for Bal- We propose a new problem related to compressive princi- anced Healthcare in Patient Centered Medical pal component pursuit, i.e., a low rank and sparse matrix Home decomposition from a partial measurement, which addi- tionally takes into account some noise. The computational Recently, patient centered medical home (PCMH) has be- results for randomly generated problems suggest that we come a popular model for providing healthcare services. can have a perfect recovery of a low rank matrix for small PCMH is a team based service and each team consists of perturbations. a group of professionals. In order to make balance be- tween healthcare supply and demand portfolios, we develop Mituhiro Fukuda, Junki Kobayashi stochastic optimization models to allocate patients to all Tokyo Institute of Technology PCMH teams to have an equitable level of workload for Department of Mathematical and Computing Sciences each team member. [email protected], [email protected] Issac Shams Wayne State University CP25 Wayne State University The Convex Hull of Graphs of Polynomial Func- [email protected] tions

Saeede Ajorlou, Kai Yang The convex hull of the graph of a polynomial function over Wayne State University a polytope is the intersection of all closed half-spaces con- [email protected], [email protected] taining the graph. We give a description of these half- spaces using semi-algebraic sets. This gives a finite algo- rithm to compute the convex hull. For polynomials in low CP24 dimension and degree (related to real applications), a poly- A Novel Framework for Beam Angle Optimization hedral relaxation can be computed quickly by an algorithm OP14 Abstracts 95

which can be extended to a spatial branch-and-bound al- chosen, IMRO can be regarded as either a generalization gorithm for MINLPs. of linear conjugate gradient or a generalization of FISTA. According to our testing, IMRO often outperforms other Wei Huang, Raymond Hemmecke published algorithms for BPDN. Technische Universitaet Muenchen [email protected], [email protected] Stephen A. Vavasis University of Waterloo Christopher T. Ryan Dept of Combinatorics & Optimization University of Chicago [email protected] [email protected] Sahar Karimi University of Waterloo CP25 [email protected] A Tight Iteration-complexity Bound for IPM via Redundant Klee-Minty Cubes CP26 We consider two curvature integrals for the central path of Super Linear Convergence in Inexact Restoration a polyhedron. Redundant Klee-Minty cubes (Nematollahi Methods et al., 2007) have central path whose geometric curvature is exponential in the dimension of the cube. We prove an Inexact Restoration methods are optimization methods analogous result for the curvature integral introduced by that address feasibility and optimality in different phases Sonnevend et al. 1990. For the Mizuno-Todd-Ye predictor- of each iteration. In 2005 Birgin and Martinez propose a corrector algorithm, we prove that the iteration-complexity method, within the Inexact Restoration framework, which upper bound for the Klee-Minty cubes is tight. has good local convergence properties. However the au- thors did not prove that the method is well defined, they Murat Mut only suggest an alternative to attempt to complete an iter- Department of Industrial and Systems Engineering ation of the method. Our contributions are to present suffi- Lehigh University cient conditions to ensure that the iteration is well defined [email protected] and to show a Newtonian way to complete the iteration under these assumptions. Tamas Terlaky Lehigh University Luis Felipe Bueno Department of industrial and Systems Engineering Unicamp [email protected] [email protected]

Jose Mario Martinez CP25 IMECC-UNICAMP Quadratically Constrained Quadratic Programs [email protected] with On/off Constraints and Applications in Sig- nal Processing CP26 Many applications from signal processing can be modeled The Balance Optimization Subset Selection as (non-convex) quadratically constrained quadratic pro- (BOSS) Model for Causal Inference grams (QCQP) featuring ”on/off” constraints. Our solu- tion strategy for solving the underlying QCQPs includes Matching is used to estimate treatment effects in observa- the consideration of a semidefinite programming relaxation tional studies. One mechanism for overcoming its restric- as well as the sequential second-order cone programming tiveness is to relax the exact matching requirement to one algorithm. We deal with the ”on/off” structure within a of balance on the covariate distributions for the treatment SDP-based Branch-and-Bound algorithm. Moreover, con- and control groups. The Balance Optimization Subset Se- vergence results and some first numerical results are pre- lection (BOSS) model is introduced to identify a control sented. group featuring optimal covariate balance. This presenta- tion discusses the relationship between the matching and Anne Philipp BOSS models, providing a comprehensive picture of their TU Darmstadt relationship. [email protected] Sheldon H. Jacobson University of Illinois Stefan Ulbrich Dept of Computer Science Technische Universitaet Darmstadt [email protected] Fachbereich Mathematik [email protected] Jason Sauppe University of Illinois CP25 [email protected] A Fast Quasi-Newton Proximal Gradient Algo- rithm for Basis Pursuit Edward Sewell Southern Illinois University Edwardsville We propose a proximal-gradient algorithm in which the ap- [email protected] proximate Hessian is the identity minus a rank-one (abbre- viated IMRO) matrix. IMRO is intended to solve the basis- pursuit denoising (BPDN) variant of compressive sensing CP26 signal recovery. Depending on how the rank-one matrix is Preconditioned Gradient Methods Based on 96 OP14 Abstracts

Sobolev Metrics [email protected]

We give a description of the application of the Sobolev gra- dient method to continuous optimization problems, such as CP26 the the minimization of the Ginzburg-Landau and Gross- Approximating the Minimum Hub Cover Problem Pitaevskii energy functionals governing superconductivity on Planar Graphs and Its Application to Query and superfluidity. The Sobolev gradient method leads to Optimization operator preconditioning of the Euler-Lagrange equations. A significant gain in computational efficiency is obtained We introduce a new combinatorial optimization problem for the above problems. We end with an extension of the which is NP-hard on general graphs. The objective is to Sobolev gradient method to include quasi-Newton meth- cover the edges of a graph by selecting a set of hub nodes. ods. A hub node covers its incident edges and also the edges between its adjacent neighbors. As an application, we dis- Parimah Kazemi cuss query processing over graph databases. We introduce Beloit College an approximation algorithm for planar graphs arising in Department of Mathematics object recognition and biometric identification and inves- [email protected] tigate its empirical performance. Belma Yelbay Sabanci University, Manufacturing Systems CP26 and Industrial Engineering Regional Tests for the Existence of Transition [email protected] States Ilker S. Birbil Aiming to locate the transition states (first-order saddle Sabanci University points) of general nonlinear functions, we introduce re- [email protected] gional tests for the identification of hyper-rectangular ar- eas that do not contain a transition state. These tests are Kerem Bulbul based on the interval extensions of theorems from linear Sabanci University, Manufacturing Systems and algebra and can be used within the framework of global Industrial Engineering optimization methods which would otherwise expend com- [email protected] putational resources locating all stationary points.

Dimitrios Nerantzis Hasan Jamil Imperial College London University of Idaho, Department of Centre for Process Systems Engineering Computer Science [email protected] [email protected]

Claire Adjiman MS1 Imperial College London Dynamic Extensible Bin Packing with Applications [email protected] to Patient Scheduling We present a multistage stochastic programming model CP26 formulation for a dynamic version of the stochastic exten- Primal and Dual Approximation Algorithms for sible bin packing problem. Motivated by applications to Convex Vector Optimization Problems health services, we discuss a special case of the problem that can be solved using decomposition methods. Results Two approximation algorithms for solving convex vector for a series of practical tests cases are presented. Perfor- optimization problems (CVOPs) are provided. Both al- mance guarantees for fast approximations are developed gorithms solve the CVOP and its geometric dual problem and their expected performance is evaluated. simultaneously. The first algorithm is an extension of Ben- Bjorn Berg son’s outer approximation algorithm, and the second one George Mason University is a dual variant of it. Both algorithms provide an inner Department of Systems Engineering and Operations as well as an outer approximation of the (upper and lower) Research images. Only one scalar convex program has to be solved in [email protected] each iteration. We allow objective and constraint functions that are not necessarily differentiable, allow solid pointed polyhedral ordering cones, and relate the approximations Brian Denton to an appropriate -solution concept. Department of Industrial and Operations Engineering University of Michigan Firdevs Ulus [email protected] PhD Student at Princeton University ORFE [email protected] MS1 Online Stochastic Optimization of Radiotherapy Birgit Rudloff Patient Scheduling Princeton University brudloff@princeton.edi The effective management of a cancer treatment facility for radiation therapy depends mainly on optimizing the Andreas Lohne use of the linear accelerators. In this project, we sched- Martin-Luther-Universit¨at ule patients on these machines taking into account their OP14 Abstracts 97

priority for treatment, the maximum waiting time before [email protected] the first treatment, and the treatment duration. We col- laborate with the Centre Intgr de Cancrologie de Laval to determine the best scheduling policy. Furthermore, we in- MS2 tegrate the uncertainty related to the arrival of patients Multiple Optimization Problems with Equilibrium at the center. We develop a hybrid method combining Constraints (MOPEC) stochastic optimization and online optimization to better meet the needs of central planning. We use information We present a mechanism for describing and solving col- on the future arrivals of patients to provide an accurate lections of optimization problems that are linked by equi- picture of the expected utilization of resources. Results librium conditions. The general framework of MOPEC based on real data show that our method outperforms the captures many example applications that involve indepen- policies typically used in treatment centers. dent decisions coupled by shared resources. We describe this mechanism in the context of energy planning, en- Antoine Legrain vironmental and land-use problems. We outline several Ecole´ Polytechnique de Montr´eal algorithms and investigate their computational efficiency. [email protected] Some stochastic extensions will also be outlined. Michael C. Ferris Marie-Andr´ee Fortin University of Wisconsin Centre Int´egr´edeCanc´erologie Department of Computer Science [email protected] [email protected] Nadia Lahrichi, Louis-Martin Rousseau Ecole´ Polytechnique de Montr´eal MS2 [email protected], louis- Methods of Computing Individual Confidence In- [email protected] tervals for Solutions to Stochastic Variational In- equalities

MS1 Stochastic variational inequalities provide a means for Scheduling Chemotherapy Patient Appointments modeling various optimization and equilibrium problems where model data are subject to uncertainty. Often the This talk reports our four year effort to improve chemother- true form of these problems cannot be analyzed and some apy appointment scheduling practices at the BC Cancer approximation is used. This talk considers the use of a Agency (BCCA). It describes our initial work at the Van- sample average approximation (SAA) and the question of couver Clinic including our process review, optimization building individual confidence intervals for components of model, software development, implementation, evaluation the true solution computable only from the sample data. and modifications to address system changes. It then dis- Two methods will be contrasted to illustrate the benefits of cusses the challenges encountered implementing the tool at incorporating the SAA solution more directly in the com- three other BCCA clinic locations and more broadly. putation of interval half-widths.

Martin L. Puterman Michael Lamm Sauder School of Business, UBC University of North Carolina [email protected] [email protected]

Amarjit Budhiraja MS1 Department of Statistics and Operations Research Chance-Constrained Surgery Planning under Un- University of North Carolina at Chapel Hill certain or Ambiguous Surgery Time [email protected]

In this paper, we consider surgery planning problems un- Shu Lu der uncertain operating time of surgeries. We decide which University of North Carolina at Chapel Hill operating rooms (ORs) to open, allocation of surgeries to [email protected] ORs, as well as the sequence and time to start each surgery. We formulate a binary integer program with individual and joint chance constraints to restrict the risk of having MS2 surgery delays and overtime ORs, respectively. We fur- Semismooth Newton Method for Differential Vari- ther analyze a distributionally robust problem variant by ational Inequalities assuming ambiguous distributions of random surgery time, for which we build a confidence set using statistical diver- We study a semismooth Newton method for differential gence functions. The variant restricts the maximum risk of variational inequalities (DVIs). Such problems comprise surgery delay and OR overtime for any probability function the solution of an ODE and a variational inequality (VI) in the confidence set, and becomes equivalent to a chance- and have various applications in engineering as well as in constrained program evaluated on an empirical probability economic sciences. The method we propose is based on function but with smaller risk tolerances. We compare dif- a suitable time discretization scheme of the underlying ferent models, approaches and derive insights of surgery ODE and a reformulation of the resulting finite dimen- planning under surgery time uncertainty or ambiguity. sional problem as a system of nonlinear, nonsmooth equa- tions. We will theoretically analyze the resulting method Yan Deng, Siqian Shen, Brian Denton and finish with some numerical results. Department of Industrial and Operations Engineering University of Michigan Sonja Steffensen [email protected], [email protected], btden- RWTH Aachen 98 OP14 Abstracts

steff[email protected] [email protected]

Sanjay Mehrotra MS2 IEMS On the Optimal Control of a Class of Variational Northwestern University Inequalities of the Second Kind [email protected] We consider optimal control problems in which the fea- sible set is governed by a certain class of variational in- MS3 equalities of the second kind. The class encompasses a number of important applications in mechanics and image Distributionally Robust Modeling Frameworks and processing. As a means of comparison to known results, Results for Least-Squares we consider both discretized and continuous forms of the problem. The existence of a hierarchy of optimality condi- Abstract not available. tions (S-/M-/C-stationarity) similar to MPCC and MPEC models is shown. Our approach uses sensitivity results for Sanjay Mehrotra the control-to-state mapping of the variational inequality. IEMS Finally, a numerical method is developed and its perfor- Northwestern University mance is illustrated by a few examples. [email protected]

Thomas M. Surowiec Department of Mathematics MS3 Humboldt University of Berlin A Cutting Surface Algorithm for Semi-infinite Con- [email protected] vex Programming and Moment Robust Optimiza- tion

MS3 We present a novel algorithm for distributionally robust A Robust Additive Multiattribute Preference optimization problems with moment uncertainty. First a Model using a Nonparametric Shape-Preserving new cutting surface algorithm is presented, that is applica- Perturbation ble to problems with non-differentiable semi-infinite con- straints indexed by an infinite-dimensional set. Our sec- We develop a multiattribute preference ranking rule in the ond ingredient is an algorithm to find approximately op- context of utility robustness. A nonparametric perturba- timal discrete probability distributions subject to moment tion of a given additive reference utility function is speci- constraints. The combination of these algorithms yields a fied to solve the problem of ambiguity and inconsistency in solution to a general family of distributionally robust op- utility assessments, while preserving the additive structure timization problems. and the decision maker’s risk preference under each crite- rion. A concept of robust preference value is defined using David Papp the worst expected utility of an alternative incurred by the Harvard perturbation, and we rank alternatives by comparing their [email protected] robust preference values. Sanjay Mehrotra Jian Hu IEMS University of Michigan, Dearborn Northwestern University [email protected] [email protected] Yung-wen Liu University of Michigan - Dearborn MS4 [email protected] The Two Facility Splittable Flow Arc Set

Sanjay Mehrotra We study a generalization of the single facility splittable IEMS flow arc set which was studied by Magnanti et. al. (1993), Northwestern University who obtained the facial structure of the convex hull of this [email protected] set. A simple separation algorithm for this set was later given by Atamturk and Rajan. The set is defined by a sin- MS3 gle capacity constraint, one nonnegative integer variable and any number of bounded continuous variables. We gen- Distributionally-Robust Support Vector Machines eralize this set by considering two integer variables and give for Machine Learning its convex hull. We propose distributionally-robust Support Vector Ma- Sanjeeb Dash chines (DR-SVMs) and present an algorithm to solve the IBM T.J. Watson Research Center DR-SVMs. We assume that the training data are drawn [email protected] from an unknown distribution and consider the standard SVM as a sample average approximation (SAA) of a stochastic version of SVM with the unknown distribution. Oktay Gunluk Using the ambiguity set based on the Kantorovich distance, IBM, Watson, New York we find the robust-counterpart of the stochastic SVM. [email protected]

Changhyeok Lee Northwestern University University Catholic de Louvain OP14 Abstracts 99

[email protected] Sven Wiese University of Bologna [email protected] MS4 How Good Are Sparse Cutting-Planes? MS5 Sparse cutting-planes are often used in mixed-integer pro- Rethinking MDO for Dynamic Engineering System gramming solvers, since they help in solving linear pro- Design grams more efficiently. However, how well can we approx- imate the integer hull by just using sparse cutting-planes? Dynamic engineering systems present distinctive design In order to understand this question better, given a polyope challenges, such as addressing time-varying behavior ex- P (e.g. the integer hull of a MIP), let P k be its best ap- plicitly while managing the fundamentally connected ac- proximation using cuts with at most k non-zero coefficients. tivities of physical and control system design. Whole- We present various results on how well P k approximates system design strategies, such as multidisciplinary design P . optimization (MDO), are a promising solution. Existing MDO formulations, however, do not address explicitly the Santanu S. Dey, Marco Molinaro, Qianyi Wang unique needs of dynamic systems. Recent results from the Georgia Institute of Technology emerging area of multidisciplinary dynamic system design School of Industrial and System Engineering optimization (MDSDO) are presented that begin to ad- [email protected], [email protected], dress these gaps. [email protected] James T. Allison University of Illinois MS4 [email protected] Finitely Convergent Decomposition Algorithms for Two-Stage Stochastic Pure Integer Programs MS5 We study a class of two-stage stochastic integer programs Using Graph Theory to Define Problem Formula- with general integer variables in both stages and finitely tion in Openmdao many realizations of the uncertain parameters. We pro- pose decomposition algorithms, based on Benders method, Graph theory is applied to develop a formal specification that utilize Gomory cuts in both stages. The Gomory cuts for optimization problem formulation. A graph can be for the second-stage scenario subproblems are parameter- constructed that uniquely defines the problem formulation. ized by the first-stage decision variables. We prove the This graph syntax is implemented as the internal represen- finite convergence of the proposed algorithms and report tation for problem formulation in the OpenMDAO frame- our computations that illustrate their effectiveness. work. Combining the graph with standard graph traversal algorithms enables the framework compute system level Simge Kucukyavuz analytic coupled derivatives very efficiently via Hwang and Dept of Integrated Systems Engineering Martins general derivatives method. We demonstrate this Ohio State Univ. capability on a large MDO problem with over 20 disciplines [email protected] and 25000 design variables. Justin Gray Minjiao Zhang NASA University of Alabama [email protected] [email protected]

MS5 MS4 A Matrix-free Trust-region SQP Method for Water Networks Optimization: MILP vs MINLP Reduced-space PDE-constrained Optimization: Insights Enabling Modular Multidisciplinary Design We will consider some important applications in the design Optimization of energy networks, where energy is considered at large In the reduced-space formulation of PDE-constrained op- and includes the distribution of natural resources as for timization problems, the state variables are treated as im- example water and gas. We will discuss the challenges of plicit functions of the control variables via the PDE con- these applications and the good mathematical tools to deal straint. Such reduced-space formulations remain attrac- with them, where often the difficulty arises in the decisions tive in many applications, especially those with complex of when, where and what should be linearized. nonlinear physics that demand specialized globalization al- Cristiana Bragalli gorithms for the PDE. However, constraints (beyond the University of Bologna PDE) present a challenge for reduced-space algorithms, be- [email protected] cause the Jacobian is prohibitively expensive to form. This precludes many traditional optimization strategies — like null-space methods — that leverage factorizations of the Claudia D’Ambrosio Jacobian. This motivates the present work, which seeks a CNRS, LIX, France matrix-free trust-region SQP method to solve constrained [email protected] optimization problems. In this talk, we will discuss the pro- posed algorithm and illustrate its use on a multidisciplinary Andrea Lodi design optimization problem in which the disciplines are University of Bologna coupled using constraints. DEIS [email protected] Jason E. Hicken 100 OP14 Abstracts

Rensselaer Polytechnic Institute [email protected] Assistant Professor [email protected] MS6 Alp Dener An Exterior-Point Method for Support Vector Ma- Rensselaer Polytechnic Institute chines Department of Mechanical, Aerospace, and Nuclear Engineering We present an exterior-point method (EPM) for solving [email protected] convex quadratic programming problems (QP) required to train support vector machines. The EPM allows iterates to approach the solution to the QP from the exterior of MS5 the feasible set. This feature helps design a simple active- passive strategy for reducing the size of linear systems A General Theory for Coupling Computational solved at each EPM iteration and thus accelerating the al- Models and their Derivatives gorithm. We present numerical results and discuss future directions for the EPM development. The computational solution of multidisciplinary optimiza- tion problems can be challenging due to the coupling be- Igor Griva tween disciplines and the high computational cost. We George Mason University have developed a general theory and framework that ad- [email protected] dress these issues by decomposing the problem into vari- ables managed by a central framework. The problem im- plementation then becomes a systematic process of defining MS6 each variable sequentially, and the solution of the multi- L1 Regularization Via the Parametric Simplex disciplinary problem and the coupled derivatives can be Method automated. An l1 penalty term is often used to coerce sparsity in least- John T. Hwang, Joaquim Martins squares problems. The associated weighting parameter re- University of Michigan quires tuning to get the “right’ answer. If the least-squares [email protected], [email protected] problem is changed to least-absolute-deviations (LAD), then the problem can be reduced to a linear programming problem that can be solved using the parametric simplex MS6 method thereby resolving the tuning question. We will re- port some comparisons between the least-squares approach Cubic Regularization in Interior-Point Methods and the LAD method.

We present several algorithms for nonlinear optimization, Robert J. Vanderbei all employing cubic regularization. The favorable theo- Operations Research and Financial Engineering retical results of Griewank (1981), Nesterov and Polyak Department (2006), and Cartis et.al. (2011) motivate the use of cubic Princeton University regularization, but its application at every iteration of the [email protected] algorithm, as proposed by these papers, may be compu- tationally expensive. We propose some modifications, and numerical results are provided to illustrate the robustness MS7 and efficiency of the proposed approaches on both uncon- On Subspace Minimization Conjugate Gradient strained and constrained problems. Method

Hande Y. Benson The proposition of limited-memory BFGS method and Drexel University Barzilai-Borwein gradient method heavily restrict the use Department of Decision Sciences of conjugate gradient method in large-scale optimization. [email protected] This is, to the great extent, due to the requirement of a relatively exact line search at each iteration and the loss David Shanno of conjugacy property of the search directions in various Rutgers University (Retired) occasions. In this talk, I shall propose a new variant of [email protected] the subspace minimization conjugate gradient method by Yuan and Stoer (1995). The new method shares with the Barzilai-Borwein gradient method a numerical advantage MS6 that the trial stepsize can be accepted by the corresponding line search when the iteration tends to the solution. Some A Stochastic Optimization Algorithm to Solve theoretical properties of the method will be explored and Sparse Additive Model some numerical results will be provided. Consequently, we can see that the proposed algorithm can become a promis- We present a new stochastic optimization algorithm to ing candidate for large-scale optimization. solve the sparse additive model problem. The problem is of the form f(E[g(x)]). We prove that our algorithm con- Yuhong Dai verges almost surely to the optimal solution. This is the AMSS, first algorithm solves the sparse additive model problem Chinese Academy of Sciences with a provably convergence result. [email protected]

Ethan Fang, Han Liu, Mengdi Wang Princeton University MS7 [email protected], [email protected], Iteration-Complexity with Averaged Nonexpansive OP14 Abstracts 101

Operators: Application to Convex Programming methods as smoothers or regularizers is examined as well.

In this paper, we establish the iteration-complexity bounds Kees van den Doel (pointwise and ergodic) for relaxed fixed point iteration University of British Columbia built from the nonexpansive monotone operators, and ap- [email protected] ply them to analyze the convergence speed of various meth- ods proposed in literature. Furthermore, for the general- Uri M. Ascher ized forward-backward splitting algorithm, recently pro- University of British Columbia posed by Raguet, Fadili and Peyr´e for finding a zero of Department of Computer Science asumofn>0 maximal monotone operators and a co- [email protected] coercive operator on a Hilbert space, we develop an easily verifiable termination criterion for finding an approximate solution, which is a generalization of the termination cri- MS8 terion for classic gradient descent methods. We then illus- A Two-Phase Augmented Lagrangian Filter trate the usefulness of the above results by applying them Method to a large class of composite convex optimization problems n of the form f + i=1 hi,wheref has a Lipschitz-continuous We present a new two-phase augmented Lagrangian filter gradient and the hi’s are simple (i.e. whose proximity oper- method. In the first phase, we approximately minimize ator is easily computable). Various experiments on signal the augmented Lagrangian until a filter acceptable point is and image recovery problem are shown. found. In the second phase, we use the active-set predic- tion from the first phase to an solve an equality-constrained Jingwei Liang QP. The algorithm also incorporates a feasibility restora- Ecole Nationale Sup´erieure d’Ing´enieurs de Caen, France tion phase to allow fast convergence for infeasible problem. [email protected] We show that the filter provides a nonmonotone global convergence mechanism, and present numerical results. Jalal Fadili CNRS, ENSICAEN-Univ. of Caen, France Sven Leyffer [email protected] Argonne National Laboratory leyff[email protected] Gabriel Peyre CNRS, CEREMADE, Universit´e Paris-Dauphine MS8 [email protected] Adaptive Observations And Multilevel Optimiza- tion In Data Assimilation MS7 We propose to use a decomposition of large-scale incremen- 2. An Efficient Gradient Method Using the Yuan tal 4D-Var data assimilation problems in order to make Steplength their numerical solution more efficient. It is based on ex- ploiting an adaptive hierarchy of the observations. Start- We present a new gradient method for quadratic program- ing with a low-cardinality set and the solution of its cor- ming, which exploits the asymptotic spectral behaviour of responding optimization problem, observations are adap- the Yuan steplength to foster a selective elimination of tively added based on a posteriori error estimates. The the components of the gradient along the eigenvectors of particular structure of the sequence of associated linear sys- the Hessian matrix. Numerical experiments show that this tems allows the use of an efficient variant of the conjugate method tends to outperform other efficient gradient meth- gradient algorithm. The method is justified by deriving the ods, especially as the Hessian condition number and the relevant error estimates at different levels of the hierarchy accuracy requirement increase. and a practical computational technique is then derived. The algorithm is tested on a 1D-wave equation and on the Roberta D. De Asmundis Lorenz-96 system, which is of special interest because of Dpt. of Statistical Sciences, its similarity with Numerical Weather Prediction (NWP) University of Rome ”La Sapienza”, Italy systems. [email protected] Philippe L. Toint Daniela di Serafino Fac Univ Notre Dame de la Paix Second University of Naples Department of Mathematics Department of Mathematics [email protected] daniela.diserafi[email protected] Serge Gratton Gerardo Toraldo ENSEEIHT, Toulouse, France University of Naples ”Federico II” [email protected] Department of Agricultural Engineering [email protected] Monserrat Rincon-Camacho CERFACS, Toulouse, France [email protected] MS7 1. Faster Gradient Descent MS8 How fast can practical gradient descent methods be? On Multilevel Methods for Very Large Scale NLPs Re- average we observe cond( (A)), like for the conjugate gra- sulting from PDE Constrained Optimization dient method. We will explain this, and also propose a new method related to Leja points. The performance of such We discuss recent developments for multilevel optimization 102 OP14 Abstracts

methods applied to PDE constrained problems. We pro- entiablity of the optimal value function and derive a limit pose a multilevel optimization approach that generates a theorem for the optimal values of empirical (sample aver- hierarchy of discretizations during the optimization itera- age) approximations of dominance constrained optimiza- tion using adaptive discrete approximations and reduced tion models (joint work with Darinka Dentcheva). order models such as POD. The adaptive refinement strat- egy is based on a posteriori error estimators for the PDE- Werner Roemisch constraint, the adjoint equation and the criticality mea- Humboldt University Berlin sure. We demonstrate the efficiency of the approach by Department of Mathematics numerical examples. [email protected]

Stefan Ulbrich Technische Universitaet Darmstadt MS9 Fachbereich Mathematik Stochastic Dominance in Shape Optimization un- [email protected] der Uncertain Loading

Shape optimization under linearized elasticity and uncer- MS8 tain loading is considered from the viewpoint of two-stage A Class of Distributed Optimization Methods with stochastic programming. Emphasizing risk aversion we Event-Triggered Communication minimize volume under stochastic dominance (stochastic order) constraints involving compliance benchmark distri- We present distributed optimization methods with event- butions. Numerical results for different types of stochastic triggered communication that keep data local and private orders, i.e.,the ‘usual’ stochastic order and the increasing as much as possible. Nesterov’s first order scheme is ex- convex order, are presented. tended to use event-triggered communication in networked environments. Then, this approach is combined with the Ruediger Schultz proximal center algorithm by Necoara and Suykens. We University of Duisburg-Essen use dual decomposition and apply our event-triggered ver- [email protected] sion of Nesterov’s scheme to update the multipliers. Exten- sions to problems with LMI constraints are also discussed and numerical results are presented. MS9 Numerical Methods for Problems with Multivari- Michael Ulbrich,MartinMeinel ate Stochastic Dominance Constraints Technische Universitaet Muenchen Chair of Mathematical Optimization We consider risk-averse stochastic optimization problems [email protected], [email protected] involving the use of linear stochastic dominance of the sec- ond order. Various primal and dual methods have been proposed to solve these types of problems. The crux of MS9 any solution method involves solving a non-convex, non- Regularization Methods for Stochastic Order Con- linear global optimization problem which verifies whether strained Problems the linear stochastic dominance constraint is satisfied. Var- ious solution techniques for this problem will be discussed We consider stochastic optimization problems with along with their relevant numerical experience. stochastic order constraints. We develop two numerical methods with regularization for their numerical solution. Darinka Dentcheva Our methods utilize the characterization of the stochastic Department of Mathematical Sciences order by integrated survival or integrated quantile func- Stevens Institute of Technology tions to progressively approximate the feasible set. Con- [email protected] vergence of the methods is proved in the case of discrete distributions. Eli Wolfhagen Stevens Institute of Technology Gabriela Martinez Mathematical Sciences Cornell University [email protected] Department of Biological and Environmental Engineering [email protected] MS10 Darinka Dentcheva On a Reduction of Cardinality to Complementarity Department of Mathematical Sciences in Sparse Optimization Stevens Institute of Technology [email protected] A reduction of cardinality-constrained problems (CardCP) to complementarity constrained problems is presented. We prove their equivalence in the sense that they have the MS9 same global minimizers. A relation between their local Stability of Optimization Problems with Stochastic minimizers is also discussed. Local optimality conditions Dominance Constraints for CardCPs are derived on the base of presenting car- dinality constraints in a disjunctive form. A continuous We consider convex optimization problems with kth order reformulation of portfolio optimization problem with semi- stochastic dominance constraints for k ≥ 2. We establish continuous variables and cardinality constraint is given. quantitative stability results for optimal values and primal- dual solution sets of the optimization problems in terms Oleg Burdakov of a suitably selected probability metric. Moreover, we Linkoping University provide conditions ensuring Hadamard-directional differ- [email protected] OP14 Abstracts 103

Christian Kanzow oracle calls for the proposed methods. Furthermore, meth- University of Wuerzburg ods using only stochastic zeroth-order stochastic informa- [email protected] tion are proposed and analysed.

Alexandra Schwartz Xiao Wang University of W¨urzburg University of Chinese Academy of Sciences Institute of Mathematics [email protected] [email protected] Shiqian Ma The Chinese University of HongKong MS10 [email protected] Quasi-Newton methods for the Trust-Region Sub- problem Ya-Xiang Yuan Chinese Academy of Sciences This talk will discuss consider various quasi-Newton sub- Inst of Comp Math & Sci/Eng, AMSS problem solvers. Numerical results will be presented. [email protected] Jennifer Erway Wake Forest University [email protected] MS11 Plan-C, A C-Package for Piecewise Linear Models RoummelF.Marcia in Abs-Normal Form University of California, Merced [email protected] Abstract not available.

Andreas Griewank MS10 HU Berlin, Germany Sum Rate Maximization Algorithms for Mimo Re- [email protected] lay Networks in Wireless Communications

Sum rate maximization problem is always of great inter- MS11 ests in the field of wireless communications. For MIMO relay networks, we propose a new approach to approxi- Evaluating Generalized Derivatives of Dynamic mate sum rate maximization, and prove it is a lower bound Systems with a Linear Program Embedded of achievable sum rate. To solve the nonlinear noncon- vex optimization problem, we change the fraction function Dynamic systems with a linear program embedded can be into a non-fraction function in the objective function, and found in control and optimization of detailed bioreactor show that the optimization problems share the same sta- models. The optimal value of the embedded linear pro- tionary points. By applying the alternating minimization gram, which is parameterized by the dynamic states, is first method, we decompose the complex problem into noncon- not continuously differentiable and second not directionally vex quadratic constraint quadratic programming subprob- differentiable on the boundary of its domain. For evaluat- lems. We propose an efficient algorithm to solve such sub- ing generalized derivatives, these are two challenges that problems. Numerical rsults show that our new model and are overcome by computing elements of the lexicographic algorithm perform well. subdifferential based on a forward sensitivity system and determining strictly feasible trajectories. Cong Sun Beijing University of Posts and Telocommunications Kai Hoeffner [email protected] MIT hoeff[email protected] Eduard Jorswieck Dresden University of Technology [email protected] MS11

Yaxiang Yuan Evaluating Generalized Derivatives for Nonsmooth Laboratory of Scientific and Engineering Computing Dynamic Systems Chinese Academy of Science [email protected] Established numerical methods for nonsmooth problems typically require evaluation of generalized derivatives such as the Clarke Jacobian. However, existing methods for MS10 evaluating generalized derivatives for nonsmooth dynamic Penalty Methods with Stochastic Approximation systems are limited either in scope or in quality of the eval- for Stochastic Nonlinear Programming uated derivatives. This presentation asserts that Nesterovs lexicographic derivatives are as useful as the Clarke Ja- We propose a unified framework of penalty methods with cobian in numerical methods, and describes lexicographic stochastic approximation for stochastic nonlinear program- derivatives of a nonsmooth dynamic system as the unique ming. In each iteration we solve a nonconvex stochas- solution of an auxiliary dynamic system. tic composite optimization problem with structured nons- mooth term. A new stochastic algorithm is presented for Kamil Khan,PaulI.Barton solving this particular class of subproblems. We also anal- Massachusetts Institute of Technology yse the worst-case complexity on the stochastic first-order Department of Chemical Engineering 104 OP14 Abstracts

[email protected], [email protected] Lehigh University [email protected]

MS11 Pietro Belotti Adjoint Mode Computation of Subgradients for FICO McCormick Relaxations pietrobelotti@fico.com Subgradients of McCormick relaxations of factorable func- Imre P´olik tions can naturally be computed in tangent (or forward) SAS Institute mode Algorithmic Differentiation (AD). Subgradients are [email protected] natural extensions of “usual” derivatives which allow the application of derivative-based methods to possibly non- differentiable convex and concave functions. In this talk Ted Ralphs an adjoint (reverse mode AD) method for the computation Department of Industrial and Systems Engineering of subgradients for McCormick relaxations is presented. A Lehigh University corresponding implementation by overloading in Fortran is [email protected] provided. The calculated subgradients are used in a deter- ministic global optimization algorithm based on a simple Tamas Terlaky branch-and-bound method. The potential superiority of Lehigh University adjoint over tangent mode AD is discussed. Department of industrial and Systems Engineering [email protected] Markus Beckers German Research School for Simulation Sciences LuFG Informatik 12: STCE MS12 [email protected] Mixed Integer Programming with p-Order Cone and Related Constraints Viktor Mosenkis Software and Tools for Computational Engineering Abstract Not Available RWTH Aachen University Alexander Vinel, Pavlo Krokhmal [email protected] University of Iowa Department of Mechanical and Industrial Engineering Uwe Naumann [email protected], [email protected] RWTH Aachen University Software and Tools for Computational Engineering [email protected] MS13 Information Relaxations, Duality, and Convex Stochastic Dynamic Programs MS12 Network Design with Probabilistic Capacity Con- We consider the information relaxation approach for cal- straints culating performance bounds for stochastic dynamic pro- grams. We study convex DPs and consider penalties based We consider a network design problem with probabilis- on linear approximations of approximate value functions. tic capacity constraints and the corresponding separation We show that these “gradient penalties” in theory pro- problem. We derive valid combinatorial inequalities for im- vide tight bounds and can improve on bounds provided by proving solution times and present computational results. other relaxations, e.g. Lagrangian relaxations. We apply the method to a network revenue management problem and find that some relatively easy-to-compute heuristic policies Alper Atamturk are nearly optimal. U.C. Berkeley [email protected] David Brown,JamesSmith Duke University Avinash Bhardwaj [email protected], [email protected] University of California-Berkeley [email protected] MS13 High Dimensional Revenue Management MS12 Disjunctive Conic Cuts for Mixed Integer Second Motivated by online advertising, we discuss the problem Order Cone Optimization of optimal revenue management in settings with a large, potentially exponential, number of customer types. We We investigate the derivation and use of disjunctive conic present a conceptually simple learning algorithm for this cuts for solving MISOCO problems. We first describe the problem that has shown empirical promise in prior work. derivation of these cuts. Then, we analyze their effective- We show that when the ’revenue’ functions arise from a ness in a branch and cut framework. Various criteria are practically interesting class of functions, the sample com- explored to select the disjunctions to build the cuts, as plexity of learning a near optimal control policy scales loga- well as for node selection and branching rules. These ex- rithmically in the number of customer types. A long stream periments help to understand the impact these cuts may of antecedent work on this problem has achieved a linear have on the size of the search tree and the solution time of dependence. the problems. Vivek Farias Julio C. Goez Massachusetts Institute of Technology OP14 Abstracts 105

[email protected] more general than the methods prevalent in the literature in that it is applicable to a broad range of problems charac- Dragos Ciocan terized by a variety of action polytopes and generic arrival MIT and departure processes. In order to access the perfor- [email protected] mance of our ADP methods, we illustrate computationally tractable upper bounds on the performance of optimal poli- cies in our setting. We apply our methodology to a series MS13 of kidney matching problems calibrated to realistic kidney Dynamic Robust Optimization and Applications exchange statistics, where we obtain a significant perfor- mance improvement over established benchmarks and, via Robust optimization has recently emerged as a tractable upper bounds, illustrate that our approach is near optimal. and scalable paradigm for modeling complex decision prob- lems under uncertainty. Its applications span a wide range Nikhil Bhat, Ciamac C. Moallemi of challenging problems in finance, pricing, supply chain Graduate School of Business management or network design. In the context of dynamic Columbia University robust decision models, a typical approach is to look for [email protected], [email protected] control policies that are directly parameterized in the un- certainties affecting the model. This has the advantage of leading to simple convex optimization problems for find- ing particular classes of rules (e.g., affine). However, the MS15 approach typically yields suboptimal policies, and is hard On Some Sensitivity Results in Stochastic Optimal to analyze. In this talk, we seek to bridge this paradigm Control Theory: A Lagrange Multiplier Point of with the classical Dynamic Programming (DP) framework View for solving decision problems. We provide a set of unifying conditions – based on the interplay between the convexity and supermodularity of the DP value functions, and the In this joint work with Francisco Silva, we provide a lattice structure of the uncertainty sets – which guarantee functional framework for classical stochastic optimal con- the optimality of the class of affine decision rules, and fur- trol problems and justify rigorously that the adjoint state thermore allow such rules to be found very efficiently. Our (p, q) appearing in the stochastic Pontryagin principle cor- results suggest new modeling paradigms for dynamic ro- responds to a Lagrange multiplier when the controlled bust optimization, and our proofs bring together ideas from stochastic differential equation is viewed as a constraint. three areas of optimization typically studied separately: ro- Under some convexity assumptions the processes (p, q)cor- bust, combinatorial (lattice programming and supermodu- respond to the sensitivity of the value function wrt. the larity), and global (the theory of concave envelopes). We drift and the volatility of the equation. We apply this in exemplify our findings in a class of applications concern- concrete examples. ing the design of flexible production processes, where a firm seeks to optimally compute a set of strategic decisions Julio Backhoff (before the start of a selling season), as well as in-season Department of Mathematics-Applied Financial replenishment policies. Our results show that – when the Mathematics costs incurred are convex – replenishment policies that de- Humboldt Universit¨at zu Berlin pend linearly on the realized demands are optimal. When backhoff@math.hu-berlin.de the costs are also piece-wise affine, all the optimal decisions (strategic and tactical) can be found by solving a single lin- ear program of small size. MS15 Dan A. Iancu Some Irreversible Investment Models with Finite Stanford University Fuel and Random Initial Time for Real Options Graduate School of Business Analysis on Local Electricity Markets [email protected] We analyse a stochastic model for local, physical trading Mayank Sharma, Maxim Sviridenko of electricity formulated as a mixed control/stopping prob- IBM T.J. Watson Research Center lem. We derive the Real Option value for two-party con- [email protected], [email protected] tracts where one party (A) decides (optimally) at time τ to accept an amount P0 from the other party (B)andcom- mits to supplying a unit of electricity to B at a random MS13 time T ≥ τ. The aim is hedging B’s exposure to real-time An Approximate Dynamic Programming Approach market prices of electricity. to Stochastic Matching

We consider a class of stochastic control problems where Tiziano De Angelis the action space at each time can be described by a class of School of Mathematics matching or, more generally, network flow polytopes. Spe- The University of Manchester cial cases of this class of dynamic matching problems in- [email protected] clude many problems that are well-studied in the literature, such as: (i) online keyword matching in Internet advertis- Giorgio Ferrari ing (the ‘adwords’ problem); (ii) the bipartite matching of Center for Mathematical Economics donated kidneys from cadavers to recipients; and (iii) the Bielefeld University allocation of donated kidneys through exchanges over cy- [email protected] cles of live donor-patient pairs. We provide an approximate dynamic program (ADP) algorithm for dynamic matching John Moriarty with stochastic arrivals and departures. Our framework is School of Mathematics The University of Manchester 106 OP14 Abstracts

[email protected] Lehigh University [email protected]

MS15 A Stochastic Reversible Investment Problem on a MS16 Finite-Time Horizon: Free-Boundary Analysis Sampling Within Algorithmic Recursions

We study a continuous-time, finite horizon optimal Abstract not available. stochastic reversible investment problem of a firm. The production capacity is a one-dimensional diffusion con- Raghu Pasupathy trolled by a bounded variation process representing the ISE Department cumulative investment-disinvestment strategy. We study Virginia Tech the associated zero-sum optimal stopping game and char- [email protected] acterize its value function through a free-boundary problem with two moving boundaries. The optimal control is then shown to be a diffusion reflected at the two boundaries. MS16 Convergence Rates of Line-Search and Trust Re- Giorgio Ferrari gion Methods Based on Probabilistic Models Center for Mathematical Economics Bielefeld University Abstract not available. [email protected] Katya Scheinberg Tiziano De Angelis Lehigh University School of Mathematics [email protected] The University of Manchester [email protected] MS16 MS15 On the Use of Second Order Methods in Stochastic Optimization Probabilistic and Smooth Solutions to Stochastic Optimal Control Problems with Singular Terminal Estimating curvature with inherently noisy information is Values with Application to Portfolio Liquidation difficult, but could lead to improvements in the perfor- Problems mance of stochastic approximation methods. A few al- gorithms that update Hessian approximations have been We review recent existence, uniqueness and characteriza- proposed in the literature, but most lose the attractive tion of solutions results for PDEs and BSPDEs with sin- properties of their deterministic counterparts. We propose gular terminal values. Such equations arise in portfolio two methods based on Polyak-Ruppert averaging, which liquidation problems under price-sensitive market impact. collect second-order information along the way, and only Markovian control problems arise when the cost function sparingly use it in a disciplined manner. depends only on the current price; non-Markovian prob- lems arise if, for instance, the investors trading strategy is Stefan Solntsev benchmarked against volume-weighted average prices. The EECS Department former gives rise to a PDE with singular terminal value for Northwestern University which we establish existence of a smooth solution using [email protected] analytic methods; the latter gives rise to a novel class of BSPDEs which we solve using probabilistic methods. MS17 Ulrich Horst Institut f¨ur Mathematik Discounted Referral Pricing Between Healthcare Humboldt-Universit at zu Berlin Networks [email protected] Academic medical centers (AMC) receive referrals from low acuity networks, and usually use a fee for service contract. MS16 We consider referrals from a capitated provider (HMO), Alternating Linearization Methods for Quadratic and characterize a two-tier discount contract. Our con- Least-Square Problem tract is mutually beneficial based on both parties risk pref- erences: it increases the AMCs expected revenue while de- Second-order least square problems have arisen widely in creasing the HMOs conditional value at risk of the cost. We binary optimization, portfolio optimization, etc.. In this extend the analysis to the case of asymmetric information talk, we propose an alternating linearization framework to about referral volume. solve this set of problems which are potentially nonconvex. We show the effectiveness of our technique in terms of both Fernanda Bravo, Angela King, Retsef Levi theory and numerical experiments in the application of risk Massachusetts Institute of Technology parity optimization in portfolio management. [email protected], [email protected], [email protected]

Xi Bai Georgia Perakis ISE Department MIT Lehigh University [email protected] [email protected] Gonzalo Romero Yanez Katya Scheinberg Massachusetts Institute of Technology OP14 Abstracts 107

[email protected] Retsef Levi Massachusetts Institute of Technology [email protected] MS17 Efficient Resource Allocation in Hospital Networks Georgia Perakis MIT Over the past several years the healthcare industry wit- [email protected] nessed the consolidation of multiple community hospitals into larger care organizations. In order for large care Wei Sun providers to best utilize their growing networks, it is critical IBM Watson to understand not only system-wide demand and capacity, [email protected] but also how the deployment of limited resources can be improved. We build an optimization model that allows de- ciding how to allocate network resources efficiently in order MS18 to offer additional services and to recapture leaked demand Reduced-Complexity Semidefinite Relaxations Via within the network-owned hospitals. The model was cali- Chordal Decomposition brated using real data from a network that consists of two community hospitals and one academic medical center. We propose a new method for generating semidefinite re- laxations of sparse quadratic optimization problems. The Marcus Braun, Fernanda Bravo, Vivek Farias, Retsef Levi method is based on chordal conversion techniques: by drop- Massachusetts Institute of Technology ping some equality constraints in the conversion, we obtain [email protected], [email protected], [email protected], ret- semidefinite relaxations that are computationally cheaper, [email protected] but potentially weaker, than the standard semidefinite re- laxation. Our numerical results show that the new re- laxations often produce the same results as the standard MS17 semidefinite relaxation, but at a lower computational cost. The Effectiveness of Uniform Subsidies in Increas- ing Market Consumption Martin S. Andersen We analyze a subsidy allocation problem with endogenous University of California, Los Angeles market response under a budget constraint. The central Electrical Engineering Department planner allocates subsidies to heterogeneous producers to [email protected] maximize the consumption of a commodity. We identify marginal cost functions such that uniform subsidies are Anders Hansson optimal, even under market state uncertainty. This is pre- Division of Automatic Control cisely the policy usually implemented in practice. More- Link¨oping University over, we show that uniform subsidies have a guaranteed [email protected] performance in important cases where they are not opti- mal. Lieven Vandenberghe University of California Retsef Levi Los Angeles Massachusetts Institute of Technology [email protected] [email protected]

Georgia Perakis MS18 MIT Domain Decomposition in Semidefinite Program- [email protected] ming with Application to Topology Optimization

Gonzalo Romero Yanez We investigate several approaches to the solution of topol- Massachusetts Institute of Technology ogy optimization problems using decomposition of the com- [email protected] putational domain. We will use reformulation of the origi- nal problem as a semidefinite optimization problem. This formulation is particularly suitable for problems with vi- MS17 bration or global buckling constraints. Standard semidefi- On Efficiency of Revenue-sharing Contracts in nite optimization solvers exercise high computational com- Joint Ventures in Operations Management plexity when applied to problems with many variables and a large semidefinite constraint. To avoid this unfavourable We study capacity planning problems with resource pool- situation, we use results from the graph theory that allow ing in joint ventures under uncertainties. When resources us to equivalently replace the original large-scale matrix are heterogeneous, there exists a unique efficient revenue constraint by several smaller constraints associated with sharing contract under proportional fairness. This opti- the subdomains. Alternatively, we directly decompose the mal contract rewards every player proportionally to her computational domain for the underlying PDE, in order to marginal cost. When resources are homogeneous, there formulate an equivalent optimization problem with many does not exist an efficient revenue sharing contract. We small matrix inequalities. This leads to a significant im- propose a provably good contract that rewards each player provement in efficiency, as will be demonstrated by numer- inversely proportional to her marginal cost. ical examples.

Cong Shi Michal Kocvara University of Michigan School of Mathematics [email protected] University of Birmingham 108 OP14 Abstracts

[email protected] [email protected]

Daniela di Serafino MS18 Second University of Naples Decomposition and Partial Separability in Sparse Department of Mathematics Semidefinite Optimization daniela.diserafi[email protected] We discuss a decomposition method that exploits chordal Germana Landi graph theorems for large sparse semidefinite optimization. Department of Mathematics, Bologna University The method is based on a Douglas-Rachford splitting al- [email protected] gorithm and uses a customized interior-point algorithm to evaluate proximal operators. In many applications, and depending on the structure of the linear constraints, these MS19 proximal operators can be evaluated in parallel. We give Gradient Descent Approaches to Image Registra- results showing the scalability of this algorithm for large tion semidefinite programs, compared against standard interior point methods. Image registration is an essential step in the processing of planetary satellite images, and involves the computation Yifan Sun of an optimal low-order rigid, or high-order deformation, University of California, Los Angeles transformation between two or more images. The problem Electrical Engineering is complicated by the wide range of image data from dif- [email protected] ferent satellite systems (LANDSAT, SPOT, many others) and sensor systems (multi-band, hyperspectral, and oth- Lieven Vandenberghe ers). The presentation will look at alternative approaches University of California including basic and stochastic gradient methods. Los Angeles [email protected] Roger Eastman Department of Computer Science Loyola College in Maryland, Baltimore, Maryland 21210 MS18 [email protected] Preconditioned Newton-Krylov Methods for Topology Optimization Arlene Cole-Rhodes Electrical & Computer Engineering Departmen When modelling structural optimization problems, there is Morgan State University a perpetual need for increasingly accurate conceptual de- [email protected] signs. This impacts heavily on the overall computational effort required by a computer, and it is therefore natural to consider alternative possibilities such as parallel comput- MS19 ing. This talk will discuss the application of domain de- Adaptive Hard Thresholding Algorithm For Con- composition to a typical problem in topology optimization. strained l0 Problem Our aim is to consider the formation of a Newton-Krylov type approach, with an appropriate preconditioning strat- The hard thresholding algorithm has proved very efficient egy for the resulting interface problem. for unconstrained l0 sparse optimization. In this paper, we propose an adaptive hard thresholding (AHT) algorithm James Turner for finding sparse solutions of box constrained l0 sparse School of Mathematics optimization problems and prove that it can find a solu- University of Birmingham, UK tion in a finite number of iterations. The equality and box [email protected] constrained sparse optimization can be transformed box constrained l0 sparse optimization problems by penalty and Michal Kocvara, Daniel Loghin then can be solved by the AHT algorithm. Some numerical School of Mathematics results are reported for compressed sensing problems with University of Birmingham equality constraints and index tracking problems, which [email protected], [email protected] demonstrate that the algorithm is promising and time- saving. It can find the s-sparse solution in the former case and can guarantee out-of-sample prediction performance in MS19 the latter case. On the Regularizing Effect of a New Gradient Method with Applications to Image Processing Fengmin Xu Dpt. of Mathematics, Xi’an Jaotong University, We investigate the regularizing properties of a recently pro- Xi’an 710049, P.R. China posed gradient method, named SDC, in the solution of dis- [email protected] crete ill-posed inverse problems formulated as linear least squares problems. An analysis of the filtering features of SDC is presented, based on its spectral behaviour. Nu- MS19 merical experiments show the advantages offered by SDC, First Order Methods for Image Deconvolution in with respect to other gradient methods, in the restoration Microscopy of noisy and blurred images. High-performance deconvolution algorithms become cru- Roberta D. De Asmundis cial to avoid long delays in data analysis pipelines, es- Dpt. of Statistical Sciences, pecially in case of large-scale imaging problems, such as University of Rome ”La Sapienza”, Italy those arising in astronomy and microscopy. In this work OP14 Abstracts 109

we present effective deconvolution approaches obtained by the Conditional Value-At-Risk following two main directions: using acceleration strategies for first order methods and exploiting Graphics Processing I discuss primal and dual approaches for the application Units (GPUs). Numerical experiments on large-scale im- of the conditional value-at-risk in PDE optimization un- age deconvolution problems are performed for evaluating der uncertainty. For the primal approach, I introduce a the effectiveness of the proposed optimization algorithm. smoothing and quadrature-based discretization. I prove well-posedness of the smoothed-primal formulation as well Giuseppe Vicidomini as error bounds. For the dual approach, I derive rigorous Nanophysics Istituto Italiano di Tecnologia, Genoa optimality conditions and present a derivative-based solu- [email protected] tion technique constructed from these conditions. I con- clude with a numerical comparison of these approaches. Riccardo Zanella Department of Mathematics, Drew P. Kouri University of Modena e Reggio Emilia,41100 Modena, Optimization and Uncertainty Quantification Italy Sandia National Laboratories [email protected] [email protected]

Gaetano Zanghirati MS20 Department of Mathematics and Computer Science A-optimal Sensor Placement for Large-scale Non- University of Ferrara linear Bayesian Inverse Problems [email protected] We formulate an A-optimal criterion for the optimal place- Luca Zanni ment of sensors in infinite-dimensional nonlinear Bayesian University of Modena and Reggio Emilia inverse problems. I will discuss simplifications required Italy to make the problem computationally feasible and present [email protected] a formulation as bi-level PDE-constrained optimization problem. Numerical results for the inversion of the per- meability in ground water flow are used to illustrate the MS20 feasibility of the approach for large-scale inverse problems. Adaptive Reduced Basis Stochastic Collocation for PDE Optimization Under Uncertainty Alen Alexanderian I discuss the numerical solution of optimization prob- University of Texas at Austin lems governed by PDEs with uncertain parameters us- [email protected] ing stochastic collocation methods. The structure of this method is explored in gradient and Hessian computations. Noemi Petra A trust-region framework is used to adapt the collocation Institute for Computational Engineering and Sciences points based on the progress of the algorithm and structure (ICES) of the problem. Reduced order models are constructed to The University of Texas at Austin replace the PDEs by low dimensional equations. Conver- [email protected] gence results and numerical results are presented. Georg Stadler Matthias Heinkenschloss University of Texas at Austin Department of Computational and Applied Mathematics [email protected] Rice University [email protected] Omar Ghattas The University of Texas at Austin MS20 [email protected] Generalized Nash Equilibrium Problems in Banach Spaces MS21 Title not available A class of non-cooperative Nash equilibrium problems is presented, in which the feasible set of each player is per- Abstract not available. turbed by the decisions of their competitors via an affine constraint. For every vector of decisions, the affine con- Claire Adjiman straint defines a shared state variable. The existence of Imperial College an equilibrium for this problem is demonstrated, first or- [email protected] der optimality conditions are derived under a constraint qualification, and a numerical method is proposed. A new path-following strategy based in part on the Nikaido- Isoda MS21 function is proposed to update the path parameter. Prospects for Reduced-Space Global Optimization

Michael Hintermueller Reduced-space branch-and-bound, in which only a subset Humboldt-University of Berlin of the variables are branched on, appears a promising ap- [email protected] proach to mitigate the computational cost of global op- timization. However, results to date have only exhibited limited success. This talk highlights the importance of the MS20 convergence order of any constraint propagation technique Risk-Averse PDE-Constrained Optimization using used to ensure convergence in a reduced-space approach. 110 OP14 Abstracts

In particular, a new second-order convergent method for planning and operation of natural gas networks. The talk bounding the parametric solutions of nonlinear equations describes mixed-integer programming approaches to solve will be demonstrated. the difficult mixed-integer nonlinear programming prob- lems that arise in the planning and operation of such net- Paul I. Barton works. In this talk we mainly discuss stationary models Massachusetts Institute of Technology for gas transport and show computational results for real- Department of Chemical Engineering world networks. We also discuss how the underlying tech- [email protected] niques can be used for more general MINLPs.

Harry Watson, Achim Wechsung Lars Schewe MIT Friedrich-Alexander-Universit¨at Erlangen-N¨urnberg [email protected], [email protected] Department Mathematik, EDOM [email protected]

MS21 New Classes of Convex Underestimators for Global MS22 Optimization Nonlinear Programming Techniques for Gas Trans- port Optimization Problems We introduce the theoretical development of a new class of convex underestimators based on edge-concavity. The Since our goal is to replace simulation by optimization tech- linear facets of the convex envelop of the edge-concavity niques in gas transport, nonlinear programming plays an based underestimator, which are generated using an effi- important role for solving real-world optimization prob- cient algorithm for up to seven dimensions, are used to re- lems in order to achieve a physical and technical accu- lax a nonconvex problem. We present computational com- racy comparable with standard simulation software. In parison of different variants of aBB and the edge-concave this talk, we address the involvement of discrete aspects in based underestimator, and show under which condition the NLP models of gas transport and propose techniques for proposed underestimator is tighter than aBB. solving highly detailed NLP models that are comparable with today’s simulation models. Christodoulos A. Floudas Princeton University Martin Schmidt fl[email protected] Leibniz Universit¨at Hannover Institute of Applied Mathematics [email protected] MS21 On Feasibility-based Bounds Tightening MS22 Feasibility-based Bounds Tightening (FBBT) is used to Hard Pareto Optimization Problems for Deter- tighten the variable ranges at the nodes of spatial Branch- minig Gas Network Capacities and-Bound (sBB) algorithms for Mixed-Integer Nonlinear Programs (MINLP). FBBT may not converge finitely to Recent regulation of the German gas market has funda- its limit ranges, even in the case of linear constraints. mentally changed the contractual relationship of network Tolerance-based termination criteria may not yield a poly- operators and customers. A key concept are so called tech- time behaviour. We model FBBT by using fixed-point nical capacities of network entries and exits: network op- equations in terms of the variable ranges. This yields an erators must publish future capacities on a web platform auxiliary linear program, which can be solved efficiently. where transport customers can book them independently. We demonstrate the usefulness of our approach by improv- The talk will discuss hard optimization problems that arise ing the open-source sBB solver Couenne. from mid-term planning in this context, addressing math- ematical modeling, theoretical analysis, computational as- Leo Liberti pects and practical consequences. IBM T.J. Watson Research Lab [email protected] Marc C. Steinbach Leibniz Universit¨at Hannover Pietro Belotti [email protected] FICO pietrobelotti@fico.com Martin Schmidt, Bernhard Willert Leibniz Universit¨at Hannover Sonia Cafieri Institute of Applied Mathematics ENAC, Toulouse, France [email protected], [email protected] sonia.cafi[email protected] hannover.de

Jon Lee IBM T.J. Watson Research Center MS22 [email protected] Nonlinear Optimization in Gas Networks for Stor- age of Electric Energy

MS22 To ensure security of energy supply in the presence of Mixed-Integer Nonlinear Programming Problems highly volatile generation of renewable electric energy ex- in the Optimization of Gas Networks – Mixed- tensive storage is required. One possibility is to use electric Integer Programming Approaches compressors to store electric energy in terms of pressure increase in existing gas transport networks. We present Optimization method can play an important role in the a transient optimization model that incorporates gas dy- OP14 Abstracts 111

namics and technical network elements. Direct discretiza- [email protected] tion leads to an NLP which is solved by an interior point method. First results for a realistic gas pipeline are pre- sented. MS23 Subspace Decomposition Methods Jan Thiedau Leibniz Universit¨at Hannover Abstract not available. Institute of Applied Mathematics [email protected] Serge Gratton ENSEEIHT, Toulouse, France Marc C. Steinbach [email protected] Leibniz Universit¨at Hannover [email protected] MS23 On the Use of Iterative Methods in Adaptive Cu- MS23 bic Regularization Algorithms for Unconstrained Optimization Accelerated Gradient Methods for Nonconvex Nonlinear and Stochastic Programming We consider adaptive cubic regularization (ARC) algo- rithms for unconstrained optimization recently investi- In this talk, we present a generalization of Nesterov’s AG gated in many papers. In order to reduce the cost required method for solving general nonlinear (possibly nonconvex by the trial step computation, we focus on the employment and stochastic) optimization problems. We show that the of Hessian-free methods for minimizing the cubic model AG method employed with proper stepsize policy possesses preserving the same worst-case complexity count as ARC. the best known rate of convergence for solving smooth non- We present both a monotone and a nonmonotone version convex problems. We also show that this algorithm allows of ARC and we show the results of numerical experiments us to have a uniform treatment for solving a certain class obtained using different methods as inexact solver. of composite optimization problems no matter it is convex or not. Tommaso Bianconcini Universit`a degli Studi di Firenze Saeed Ghadimi [email protected] University of Florida Department of Industrial and Systems Engineeing Giampaolo Liuzzi sghadimi@ufl.edu Istituto di Analisi dei Sistemi ed Informatica CNR, Rome, Italy Guanghui Lan [email protected] Department of Industrial and Systems University of Florida Benedetta Morini [email protected]fl.edu Universita’ di Firenze benedetta.morini@unifi.it MS23 Marco Sciandrone Global Convergence and Complexity of a Universit`a degli Studi di Firenze Derivative-Free Method for Composite Nons- [email protected]fi.it mooth Optimization

A derivative-free trust-region algorithm is proposed for MS24 minimizing the composite function Φ(x)=f(x)+h(c(x)), Stochastic Day-Ahead Clearing of Energy Markets where f and c are smooth and h is convex but may be nonsmooth. Global convergence results are given and a Abstract not available. worst-case complexity bound is obtained. The complex- ity result is then specialized to the case when the com- Mihai Anitescu posite function is an exact penalty function, providing a Argonne National Laboratory worst-case complexity bound for equality-constrained op- Mathematics and Computer Science Division timization problems when the solution is computed using [email protected] a derivative-free exact penalty algorithm.

Geovani Grapiglia MS24 Federal University of Parana, Brazil Cutting Planes for the Multi-Stage Stochastic Unit Brazil Commitment Problem geovani [email protected] Abstract not available. Jinyun Yuan Federal University of Parana Yongpei Guan Brazil University of Florida [email protected] [email protected]fl.edu

Ya-Xiang Yuan Chinese Academy of Sciences MS24 Inst of Comp Math & Sci/Eng, AMSS Modeling Energy Markets: An Illustration of the 112 OP14 Abstracts

Four Classes of Policies Constrained Global Optimization

The modeling of independent system operators such as In the field of global optimization many efforts have been PJM requires capturing the complex dynamics of energy devoted to globally solving bound constrained optimiza- generation, transmission and consumption. Classical algo- tion problems without using derivatives. In this talk we rithmic strategies associated with stochastic programming, consider global optimization problems where also general dynamic programming or stochastic search will not work nonlinear constraints are present. To solve this problem in isolation. We provide a unified framework for stochas- we propose the combined use of a DIRECT-type algo- tic optimization that encompasses all of these fields, rep- rithm with derivative-free local minimizations of a nons- resented using four fundamental classes of policies. This mooth exact penalty function. In particular, we define a framework is demonstrated in a model of the PJM grid new DIRECT-type strategy to explore the search space by and energy markets. taking into account the two-fold nature of the optimization problems, i.e. the global optimization of both the objec- Warren Powell,HugoSimao tive function and of a feasibility measure. We report an Princeton University extensive experimentation on hard test problems. [email protected], [email protected] Gianni Di Pillo University of Rome ”La Sapienza” MS24 Dept. of Computer and Systems Science [email protected] New Lower Bounding Results for Scenario-Based Decomposition Algorithms for Stochastic Mixed- Giampaolo Liuzzi Integer Programs CNR - IASI [email protected] Non-convexity and/or a restrictions on computation time may cause the progressive hedging (PH) algorithm to ter- minate with a feasible solution and a cost upper bound. A Stefano Lucidi tight lower bound can provide reassurance on the quality University of Rome of the PH solution. We describe a new method to ob- [email protected] tain lower bounds using the information prices generated by PH. We show empirically, using these lower bounds, Veronica Piccialli that for large-scale non-convex stochastic unit commitment Universit`a degli Studi di Roma Tor Vergata problems, PH generates very high-quality solutions. [email protected]

Jean-Paul Watson Francesco Rinaldi Sandia National Laboratories University of Padova Discrete Math and Complex Systems [email protected] [email protected]

Sarah Ryan MS25 Iowa State University A World with Oil and Without Derivatives [email protected] The unstoppable increase in world energy consumption has led to a global concern regarding fossil fuel reserves and David Woodruff their production. This in part motivates oil and gas compa- University of California, Davis nies to use complex modeling and optimization algorithms. dlwoodruff@ucdavis.edu However, these algorithms are very often designed inde- pendently, and, in general, derivative information in the optimization cannot be computed efficiently. This talk il- MS25 lustrates through a number of examples the use and impor- Using Probabilistic Regression Models in Stochas- tance of derivative-free optimization in oil field operations. tic Derivative Free Optimization

We consider the use of probabilistic regression models in a David Echeverria Ciaurri classical trust region framework for optimization of a de- Environmental Science and Natural Resources, T. J. terministic function f when only having access to noise- Watson corrupted function values f˜. Contrasting to traditional re- [email protected] quirements on the poisedness of the sample set, our models are constructed using randomly selected points while pro- MS25 viding sufficient quality of approximation with high proba- bility. We will discuss convergence proofs of our proposed Formulations for Constrained Blackbox Optimiza- algorithm based on error bounds from machine learning tion Using Statistical Surrogates literature. This presentation introduces alternative formulations for using statistical surrogates and the Mesh Adaptive Direct Ruobing Chen, Katya Scheinberg Search (MADS) algorithm for constrained blackbox opti- Lehigh University mization. The surrogates that we consider are global mod- [email protected], [email protected] els that provide diversification capabilities to escape local optima. We focus on different formulations of the surrogate problem considered at each search step of the MADS algo- MS25 rithm using a statistical toolbox called dynaTree. The for- DIRECT-Type Algorithms for Derivative-Free mulations exploit different concepts such as interpolation, OP14 Abstracts 113

classification, expected improvement and feasible expected Computer Science Department improvement. Numerical examples are presented for both University of Wisconsin Madison analytical and simulation-based engineering design prob- [email protected] lems.

Sebastien Le Digabel MS26 GERAD - Polytechnique Montreal [email protected] Vulnerability Analysis of Power Systems: A Bilevel Optimization Approach Bastien Talgorn We consider a transmission line attack problem to iden- GERAD tify the most vulnerable lines of an electric power grid. [email protected] The attack is performed by increasing the impedances of each transmission lines and the disruption of the system Michael Kokkolaras is measured by the amount of load that needs to be shed McGill University to recover feasibility of the grid. We propose a continu- [email protected] ous bilevel programming model based on AC power flow equations and a combination of algorithmic and heuristics techniques from optimization will be discussed. MS26 Stochastic Variational Inequalities: Analysis and Taedong Kim Algorithms Computer Science Department University of Wisconsin at Madison Variational inequality problems find wide applicability in [email protected] modeling a range of optimization and equilibrium prob- lems. We consider the stochastic generalization of such a problem wherein the mapping is pseudomonotone. We Stephen Wright make three sets of contributions in this paper. First, we University of Wisconsin provide sufficiency conditions for the solvability of such Dept. of Computer Sciences problems that do not require evaluating the expectation. [email protected] Second, we introduce an extragradient variant of stochas- tic approximation for the solution of such problems. Under suitable conditions, it is shown that this scheme produces MS26 iterates that converge in an almostsure sense. Furthermore, A SLPEC/EQP Method for Mathematical Pro- we present a rate of convergence analysis in a monotone grams with Variational Inequality Constraints regime under a weak-sharpness requirement. The paper concludes with some preliminary numerics. I will discuss a new method for mathematical programs with complementarity constraints that is globally conver- Aswin Kannan gent to B-stationary points. The method solves a linear Penn State University program with complementarity constraints to obtain an [email protected] estimate of the active set. It then fixes the activities and solves an equality-constrained quadratic program to ob- Uday Shanbhag tain fast convergence. The method uses a filter to pro- Department of Industrial and Manufacturing Engineering mote global convergence. We establish convergence to B- Pennsylvania State University stationary points. [email protected] Todd Munson Argonne National Laboratory MS26 Mathematics and Computer Science Division Polyhedral Variational Inequalities [email protected] PathVI is a Newton-based solver for variational inequali- Sven Leyffer ties (VIs) defined over a polyhedral convex set. It com- Argonne National Laboratory putes a solution by finding a zero of the normal map asso- leyff[email protected] ciated with a given VI, whose linearization in this case is an affine VI. At each iteration, we solve an affine VI using a pivotal method and do a path search to make a Newton MS27 step. A complementary pivoting method which follows a piecewise linear manifold for solving an affine VI will be Stochastic Nonlinear Programming for Natural introduced and a path search algorithm for a Newton step Gas Networks will be presented. Implementation issues regarding initial basis setup and resolving numerical difficulties will be dis- A robust and efficient parallel nonlinear solver is alway a cussed. Some experimental results comparing PathVI to challenge for large-scale stochastic NLP problems. We dis- other codes based on reformulations as a complementary cuss the details of our implementation PIPS-NLP, which problem will be given. is a parallel nonlinear interior-point solver. The parallel strategy is designed by decomposing the problem struc- MichaelC.Ferris ture and building the Schur Complement. We also present University of Wisconsin some numerical results of the stochastic natural gas net- Department of Computer Science work problem. [email protected] Naiyuan Chiang,VictorZavala Youngdae Kim Argonne National Laboratory 114 OP14 Abstracts

[email protected], [email protected] [email protected]

MS27 MS28 A Sequential Linear-Quadratic Programming Elementary Closures in Nonlinear Integer Pro- Method for the Online Solution of Mixed-Integer gramming Optimal Control Problems The elementary closure of a Mixed Integer Programming (MIP) is obtained by augmenting the continuous relaxation We are interested in mathematical programming methods of the MIP with all cuts in a give family. Because the num- for optimizing control of nonlinear switched systems. Com- ber of cuts in a family is usually infinite, the elementary putational evidence has shown that many problems of this closure is not necessarily a polyhedron even if the continu- class are amenable to MPEC or MPVC reformulations. In ous relaxation of the MIP is a rational polyhedron (i.e. for this talk, we present a framework for mixed-integer non- linear MIP). However, most elementary closures for linear linear model-predictive control of switched systems. Our MIP have been shown to be rational polyhedron. In this method employs a direct and all-at-once method to ob- talk we study when the elementary closure of a nonlin- tain a nonlinear programming representation of the control ear MIP (i.e. those with non-polyhedral continuous relax- problem. Complementarity and vanish constraints are used ations) is a rational polyhedron. We pay special attention to model restrictions on switching decisions. A sequential to the case of nonlinear MIPs with continuous relaxations linear-quadratic programming algorithm is developed and with unbounded feasible regions. used in a real-time iteration scheme to repeatedly compute mixed-integer feedback controls. Daniel Dadush New York University Christian Kirches [email protected] Interdisciplinary Center for Scientific Computing University of Heidelberg, Germany [email protected] MS28 Understanding Structure in Conic Mixed Integer Programs: From Minimal Linear Inequalities to MS27 Conic Disjunctive Cuts An Optimization Algorithm for Nonlinear Prob- lems with Numerical Noise In this talk, we will cover some of the recent developments in Mixed Integer Conic Programming. In particular, we We present a numerical algorithm for solving nonlinear will study nonlinear mixed integer sets involving a general optimization problems where the value and the gradient regular (closed, convex, full dimensional, and pointed) cone of objective function are distorted by numerical noise. K such as the nonnegative orthant, the Lorentz cone or The method is based on regression models and transitions the positive semidefinite cone, and introduce the class of smoothly from the regime in which noise can be ignored K-minimal valid linear inequalities. Under mild assump- to the regime in which methods fail that do not explicitly tions, we will show that these inequalities together with address the noise. the trivial cone-implied inequalities are sufficient to de- scribe the convex hull. We study the characterization of Andreas Waechter, Alvaro Maggiar, Irina Dolinskaya K-minimal inequalities by identifying necessary, and suf- Northwestern University ficient conditions for an inequality to be K-minimal; and IEMS Department establish relations with the support functions of sets with [email protected], al- certain structure, which leads to efficient ways of showing [email protected], a given inequality is K-minimal. This framework natu- [email protected] rally generalizes the corresponding results for Mixed Inte- ger Linear Programs (MILPs), which have received a lot of interest recently. In particular, our results recover that MS27 the minimal inequalities for MILPs are generated by sub- linear (positively homogeneous, subadditive and convex) A Taxonomy of Constraints in Grey-Box Optimiza- functions that are also piecewise linear. However our study tion also reveals that such a cut generating function view is not possible for the conic case even when the cone involved is The types of constraints encountered in simulation-based the Lorentz cone. Finally, we will conclude by introducing or black-box optimization problems differ significantly from a new technique on deriving conic valid inequalities for sets the constraints traditionally treated in nonlinear program- involving a Lorentz cone via a disjunctive argument. This ming theory. We introduce a characterization of con- new technique also recovers a number of results from the straints to address this shortcoming. We provide formal recent literature on deriving split and disjunctive inequal- definitions of these constraint classes and provide illustra- ities for the mixed integer sets involving a Lorentz cone. tive examples for each type of constraint in the resulting The last part is joint work with Sercan Yildiz and Gerard taxonomy. These constraint classes present challenges and Cornuejols. opportunities for developing nonlinear optimization algo- rithms and theory. Fatma Kilinc-Karzan Tepper School of Business Stefan Wild Carnegie Mellon University Argonne National Laboratory [email protected] [email protected]

Sebastien Le Digabel MS28 GERAD - Polytechnique Montreal Valid Inequalities and Computations for a Nonlin- OP14 Abstracts 115

ear Flow Set with several application. It is also one of few combinatorial problems whose complexity is still unknown. In this talk Many engineering applications concern the design and op- we present a new approach to the graph isomorphism prob- eration of a network on which flow must be balanced. We lem by formulating the problem as a copositive program. present analysis of a single node flow set that has been aug- Then, using several hierarchies of cones approximating the mented with variables to capture nonlinear network behav- copositive cone, we will attempt to solve the problem. ior. Water networks, gas networks, and the electric power grid are applications that could benefit from our analysis. Luuk Gijben New classes of valid inequalities are given, and we plan to Department of Mathematics provide computational results demonstrating the utility of University of Groningen the new inequalities. [email protected]

Jeff Linderoth Mirjam D¨ur University of Wisconsin-Madison Department of Mathematics Dept. of Industrial and Sys. Eng University of Trier [email protected] [email protected]

James Luedtke University of Wisconsin-Madison MS29 Department of Industrial and Systems Engineering The Order of Solutions in Conic Programming [email protected] The order of optimal solutions is connected with the con- Hyemin Jeon vergence rate of discretization methods. Geometrically, it University of Wisconsin-Madison is related to the curvature of the feasible set around the so- [email protected] lution. We study this concept for conic optimization prob- lems and formulate conditions for first order solutions as well as higher order solutions. We also discuss stability of MS28 these properties. MINOTAUR Framework: New Developments and Bolor Jargalsaikhan an Application Department of Mathematics MINOTAUR is an open-source software framework for University of Groningen solving MINLPs. We will describe some recent additions [email protected] and improvements in the framework and demonstrate their impact on the performance on benchmark instances. We Mirjam Duer also show how this framework can be taiored to solve spe- University of Trier cific applications. We consider as an example a difficult [email protected] nonlinear combinatorial optimization problem arising in the design of power-delivery networks in electronic com- Georg Still ponents. University of Twente [email protected] Ashutosh Mahajan Argonne National Laboratory [email protected] MS29 Combinatorial Proofs of Infeasibility, and of Posi- tive Gaps in Semidefinite Programming MS29 Cutting Planes for Completely Positive Conic Farkas lemma in linear programming provides a certificate Problems to easily verify the infeasibility of a linear system of in- equalities. We prove that a similar certificate exists for Many combinatorial problems as well as nonconvex semidefinite systems: the reader can verify the infeasibility quadratic problems can be reformulated as completely pos- using only elementary linear algebra. The main tool we itive problems, that is, linear problems over the cone of use is a generalized Ramana type dual. We also present completely positive matrices. unfortunately, this reformu- similar proofs for positive duality gaps in SDPs. lation does not solve the difficulty of the problem because the completely positive cone is intractable. a tractable re- Minghui Liu laxation lies in substituting it by the semidefinite cone. In Dept of Statistics and Operations Research this talk, we show how we can construct cutting planes UNC Chapel Hill that will sharpen the relaxation. [email protected]

Mirjam Duer Gabor Pataki University of Trier University of North Carolina, at Chapel Hill [email protected] [email protected]

MS29 MS30 A Copositive Approach to the Graph Isomorphism Keller’s Cube-Tiling Conjecture – An Approach Problem Through Sdp Hierachies

The Graph Isomorphism Problem is the problem of decid- Kellers cube-tiling conjecture (1930) states that in any ing whether two graphs are the same or not. It is a problem tiling of Euclidean space by unit hypercubes, some two hy- 116 OP14 Abstracts

percubes share a full facet. The origin is tied to Minkowskis points on the unit sphere to minimize the Coulomb energy conjecture and the geometry of numbers. Dimension 7 is (the sum of the reciprocals of the pairwise distances). I will unresolved. For fixed dimension, Kellers conjecture reduces show how techniques from harmonic analysis and polyno- to an upper bound on the stability number of a highly sym- mial optimization can be used to compute these bounds. metric graph. Exploiting explicit block-diagonalizations of the invariant algebras associated with the Lasserre hierar- David de Laat chy, this problem can be approached using SDP. Delft University of Technology [email protected] Dion Gijswijt Delft University of Technology [email protected] MS31 Primal-dual Subgradient Method with Partial Co- ordinate Update MS30 SDP and Eigenvalue Bounds for the Graph Parti- In this talk we consider a primal-dual method for solv- tion Problem ing nonsmooth constrained optimization problem. This method consists in alternating updating strategies for pri- In this talk we present a closed form eigenvalue-based mal and dual variables, such that one of them can be bound for the graph partition problem. Our result is a seen as a coordinate descent scheme. We show that such generalization of a well-known result in spectral graph the- a method can be applied to the problems of very big ory for the 2-partition problem to any k-partition problem. size, keeping nevertheless the worst-case optimal complex- Further, we show how to simplify a known matrix-lifting ity bounds. SDP relaxation for different classes of graphs, and aggre- gate additional triangle and independent set constraints. Yurii Nesterov Our approach leads to interesting theoretical results. CORE Universite catholique de Louvain Renata Sotirov [email protected] Tilburg University [email protected] MS31 Edwin R. Van Dam Accelerated, Parallel and Proximal Coordinate De- Tilburg University scent Dept. Econometrics and Operations Research [email protected] We propose a new stochastic coordinate descent method for minimizing the sum of convex functions each of which de- pends on a small number of coordinates only. Our method MS30 (APPROX) is simultaneously Accelerated, Parallel and Completely Positive Reformulations for Polyno- PROXimal; this is the first time such a method is pro- mial Optimization posed. In the special case when the number of processors is equal to the number of coordinates, the method con- We consider the convex reformulation of polynomial opti- verges at the rate 2¯ωLR¯ 2/(k +1)2,wherek is the iteration mization (PO) problems that are not necessarily quadratic. counter,ω ¯ is an average degree of separability of the loss For this, we use a natural extension of the cone of CP ma- function, L¯ is the average of Lipschitz constants associ- trices; namely, the cone of completely positive tensors. We ated with the coordinates and individual functions in the provide a characterization of the class of PO problems that sum, and R is the distance of the initial point from the can be formulated as a conic program over the cone of CP minimizer. We show that the method can be implemented tensors. As a consequence, it follows that recent results without the need to perform full-dimensional vector oper- for quadratic problems can be further strengthened and ations, which is considered to be the major bottleneck of generalized to higher order PO problems. accelerated coordinate descent. The fact that the method depends on the average degree of separability, and not on Luis Zuluaga the maximum degree of separability, can be attributed to Lehigh University the use of new safe large stepsizes, leading to improved [email protected] expected separable overapproximation (ESO). These are of independent interest and can be utilized in all existing Javier Pena parallel stochastic coordinate descent algorithms based on Carnegie Mellon University the concept of ESO. [email protected] Peter Richtarik Juan C. Vera University of Edinburgh Tilburg School of Economics and Management [email protected] Tilburg University [email protected] Olivier Fercoq INRIA Saclay and CMAP Ecole Polytechnique [email protected] MS30 Title Not Available MS31 In this talk I will discuss a hierarchy of conic optimization Stochastic Dual Coordinate Ascent with ADMM problems which can be used to compute the minimal poten- tial energy of a system of repelling particles. For instance, We present a new stochastic dual coordinate ascent tech- in the Thomson problem one distributes a fixed number of nique that can be applied to a wide range of composite OP14 Abstracts 117

objective functions appearing in machine learning tasks, rithm is strongly polynomial for linear optimization prob- that is, regularized learning problems. In particular, our lems of the form min{cx|Ax = b, x ≥ 0} having 0-1 optimal method employs alternating direction method of multipli- solutions. The algorithm is also applicable to more general ers (ADMM) to deal with complicated regularization func- convex optimization problems. tions. Although the original ADMM is a batch method in the sense that it observes all samples at each iteration, Sergei Chubanov the proposed method offers a stochastic update rule where Universit¨at Siegen each iteration requires only one or few sample observations. [email protected] Moreover, our method can naturally afford mini-batch up- date and it gives speed up of convergence. We show that, under some strong convexity assumptions, our method con- MS32 verges exponentially. Optimizing over the Counting Functions of Param- Taiji Suzuki eterized Polytopes Tokyo Institute of Technology [email protected] Let Pb be the family of polytopes defined by Ax ≤ b. The number of integer points in Pb is a piece-wise step- polynomial in the parameter b that can be computed in MS31 polynomial time if the dimension of Pb is fixed. We in- Recent Progress in Stochastic Optimization for vestigate the problem of minimizing or maximizing this Machine Learning polynomial subject to constraints on b. We show that this problem is NP-hard even in two dimensions, but approxi- It has been demonstrated recently that for machine learn- mation schemes exist. ing problems, modern stochastic optimization techniques such as stochastic dual coordinate ascent have significantly Nicolai H¨ahnle better convergence behavior than traditional optimization Universit¨at Bonn methods. In this talk I will present a broad view of this [email protected] class of methods including some new algorithmic develop- ments. I will also discuss algorithms and practical consid- erations in their parallel implementations. MS32 Tong Zhang Tropicalizing the Simplex Algorithm Rutgers [email protected] Based on techniques from tropical geometry we show that special pivoting rules for the classical simplex algorithm MS32 admit tropical analogues which can be used to solve mean payoff games. In this way, any such pivoting rule with a Computing Symmetry Groups of Polyhedral Cones strongly polynomial complexity (the existence of such an Knowing the symmetries of a polyhedron can be very useful algorithm is open) would provide a strongly polynomial algorithm solving mean payoff games. The latter problem for the analysis of its structure as well as for the practical ∩ solution of linear constrained optimization (and related) is known to be NP co-NP. problems. In this talk I discuss symmetry groups preserv- ing the linear, projective and combinatorial structure of a Xavier Allamigeon polyhedron. In each case I give algorithmic recipes to com- CMAP, Ecole Polytechnique pute the corresponding group and mention some practical [email protected] experiences. Pascal Benchimol David Bremner INRIA and CMAP, Ecole´ Polytechnique MITACS n/a [email protected] Stephane Gaubert Mathieu Dutour Sikiric INRIA and CMAP, Ecole Polytechnique Rudjer Boskovic Institute [email protected] [email protected] Michael Joswig Dmitrii Pasechnik Technische Universit¨at Berlin NTU Singapore [email protected] [email protected]

Thomas Rehn, Achill Schuermann MS33 University of Rostok [email protected], [email protected] A Parallel Bundle Framework for Asynchronous Subspace Optimization of Nonsmooth Convex Functions MS32 Alternating Projections As a Polynomial Algo- Abstract not available. rithm for Linear Programming Christoph Helmberg We consider a polynomial algorithm for linear program- TU Chemnitz ming on the basis of alternating projections. This algo- Germany 118 OP14 Abstracts

[email protected] tion approaches. The goal becomes to optimize the original objective function plus the quadratic gains collected by re- producing as much as possible the reference solution. MS33 A Feasible Active Set Method for Box-Constrained Lucas Letocart Convex Problems Universit´e Paris 13 [email protected] A primal-dual active set method for quadratic problems with bound constraints is presented which extends the Fabio Furini infeasible active set approach of [K. Kunisch and F. LAMSADE Rendl. An infeasible active set method for convex prob- Universit´e Paris Dauphine lems with simple bounds. SIAM Journal on Optimization, [email protected] 14(1):3552, 2003]. Based on a guess on the active set, a primal-dual pair (x,a) is computed that satisfies the first Roberto Wolfler Calvo order optimality condition and the complementary condi- LIPN tion. If x is not feasible, the variables connected to the Universit´e Paris 13 infeasibilities are added to the active set and a new primal- roberto.wolfl[email protected] dual pair (x,a) is computed. This process is iterated until a primal feasible solution is generated. Then a new active set is guessed based on the feasibility information of the MS34 dual variable a. Strict convexity of the quadratic problem Regularization Techniques in Nonlinear Optimiza- is sufficient for the algorithm to stop after a finite number tion of steps with an optimal solution. Computational expe- rience indicates that this approach also performs well in We present two regularization techniques for nonlinear op- practice. timization. The first one is based on an inexact proxi- mal point method, the second one on an augmented La- Philipp Hungerlaender, Franz Rendl grangian reformulation of the original problem. We dis- Alpen-Adria Universitaet Klagenfurt cuss the strong global convergence properties of both these Austria approaches. [email protected], [email protected] Paul Armand Universit´edeLimoges MS33 Laboratoire XLIM BiqCrunch: A Semidefinite Branch-and-bound [email protected] Method for Solving Binary Quadratic Problems

BiqCrunch is a branch-and-bound solver for finding exact MS34 solutions of any 0-1 quadratic problem, such as many com- Inexact Second-order Directions in Large-scale Op- binatorial optimization problems in graphs. We present timization the underlying theory of the semidefinite bounding proce- dure of BiqCrunch, and how to use BiqCrunch to solve Many large-scale problems defy optimization methods some general numerical examples. Note that F. Roupin’s which rely on exact directions obtained by factoring ma- talk “BiqCrunch in action: solving difficult combinatorial trices. I will argue that second-order methods (includ- optimization problems exactly” presents the performance ing interior point algorithms) which use inexact directions of BiqCrunch on a variety of well-known combinatorial op- computed by iterative techniques and run in a matrix- timization problems. free regime offer an attractive alternative to fashionable first-order methods. I will address a theoretical issue of Nathan Krislock how much of inexactness is allowed in directions and sup- INRIA Grenoble Rhˆone-Alpes port the findings with computational experience of solving [email protected] large-scale optimization problems.

Fr´ed´eric Roupin Jacek Gondzio Laboratoire d’Informatique Paris-Nord The University of Edinburgh, United Kingdom Institut Galil´ee [email protected] [email protected]

J´erˆome Malick MS34 INRIA Reduced Order Models and Preconditioning [email protected] Reduced order models are a very successful tool in solving highly complex problems in engineering and otehr appli- MS33 cations. In this talk we point out several ways how these Local Reoptimization Via Column Generation and models can be used in the context of preconditioning. Quadratic Programming Ekkehard W. Sachs We introduce a new reoptimization framework called Lo- University of Trier cal Reoptimization for dealing with uncertainty in data. [email protected] The framework aims to construct a solution as close as possible to a reference one not anymore feasible. Local Reoptimization is conceived for generic Set Partitioning MS34 formulations which are typically solved by column genera- Experiments with Quad Precision for Iterative OP14 Abstracts 119

Solvers exactly by a spectral algorithm in the noiseless case and we produce a convex relaxation for the 2-SUM problem Implementing conjugate-gradient-type methods in quad to improve the robustness of solutions in a noisy setting. precision is a plausible alternative to reorthogonalization This relaxation also allows us to impose additional struc- in double precision. Since the data for Ax = b is likely tural constraints on the solution, to solve semi-supervised to be double, not quad, products y = Av (with quad y seriation problems. We present numerical experiments on and v) need to take advantage of A being double; for ex- archeological data, Markov chains and gene sequences. ample, by splitting v = v1 + v2 (with v1 and v2 double) and forming y = Av1 + Av2. This needs dot-products Fa jwel Fogel with double-precision data and quad-precision accumula- DI,ENS,Paris tion, but there is no such intrinsic function in Fortran 90. [email protected] We explore ways of using double-precision floating-point to achieve quad-precision products. Rodolphe Jenatton CRITEO Michael A. Saunders [email protected] Systems Optimization Laboratory (SOL) Dept of Management Sci and Eng, Stanford [email protected] Francis Bach Ecole Normale Sup´erieure, Paris [email protected] Ding Ma Stanford University [email protected] Alexandre d’Aspremont ORFE, Princeton University [email protected] MS35 Semidefinite Relaxation for Orthogonal Procrustes and the Little Grothendieck Problem MS35 On Lower Complexity Bounds for Large-Scale The little Grothendieck problem over the Orthogonal Smooth Convex Optimization T T Group consists of maximizing ij Tr(CijOiOj )overvari- ables Oi restricted to take values in the orthogonal group We prove lower bounds on the black-box oracle com- O(d), where Cij denotes the (ij)-th d x d block of a pos- plexity of large-scale smooth convex minimization prob- itive semidefinite matrix C. Several relevant problems are lems. These lower bounds work for unit balls of normed encoded by the little Grothendieck over the orthogonal spaces under a technical smoothing condition, and arbi- group, including Orthogonal Procrustes. The Orthogonal trary smoothness parameter of the objective with respect to this norm. As a consequence, we show a unified frame- Procrustes problem consists of, given N point clouds in d- p dimensional space, finding N orthogonal transformations work for the complexity of convex optimization on -balls, that best simultaneously align the point clouds. We pro- for 2 ≤ p ≤∞. In particular, we prove that the T -step pose an approximation algorithm, to solve this version of Conditional Gradient algorithm as applied to minimizing the Grothendieck problem, based on a natural semidefi- smooth convex functions over the n-dimensional box with nite relaxation. For each dimension d, we show a constant T ≤ n is nearly optimal. On the other hand, we prove lower bounds for the complexity of convex optimization approximation ratio with matching integrality gaps. In p comparison with the previously known approximation al- over -balls, for 1 ≤ p<2, by combining a random sub- gorithms for this problem, our semidefinite relaxation is spacemethodwiththep = ∞ lower bounds. In particular, smaller and the approximation ratio is better. we establish the complexity of problem classes that contain the sparse recovery problem. Afonso S. Bandeira Math. Dept. Cristobal Guzman Princeton University ISYE, Georgia Tech [email protected] [email protected]

Christopher Kennedy Arkadi Nemirovski Department of Mathematics School of Industrial and Systems Engineering University of Texas at Austin Georgia Tech [email protected] [email protected]

Amit Singer Princeton University MS35 [email protected] Sketching Large Scale Convex Quadratic Programs

We consider a general framework for randomly projecting MS35 data matrices of convex quadratic programs, and analyze Convex Relaxations for Permutation Problems the approximation ratio of the resulting low-dimensional program. Solving a much lower dimensional problem when Seriation seeks to reconstruct a linear order between vari- is very desirable in a computation/memory limited setting ables using unsorted similarity information. It has direct while it’s also possible that the nature of the sensors might applications in archeology and shotgun gene sequencing require random projections regardless of dimensions. We for example. We prove the equivalence between the seri- show that with a sub-gaussian sketching matrix, the data ation and the combinatorial 2-SUM problem (a quadratic matrices can be projected down to the statistical dimen- minimization problem over permutations) over a class of sion of the tangent cone at the original solution, which similarity matrices. The seriation problem can be solved we show to be substantially smaller than original dimen- 120 OP14 Abstracts

sions. Similar results also hold for partial Hadamard and Method for Adaptive Staffing and Scheduling Fourier projections. We discuss Lasso and interval con- under Demand Uncertainty with Application to strained Least Squares, Markowitz portfolios and Support Nurse Management Vector Machines as examples of our general framework. We present a two stage stochastic integer programming Mert Pilanci method with application of workforce planning. We con- EECS, U.C. Berkeley vexify the recourse function by adding valid inequalities to [email protected] the second stage. As a result, we present a tight formula- tion in which the integrality of the recourse can be relaxed. Martin Wainwright The model is solved by an integer L-shaped method with University of California a novel multicut approach and a new hyperplane branch- Berkeley ing strategy. We also conduct computational experiments [email protected] using real patient census data. Kibaek Kim Laurent El Ghaoui Northwestern University EECS, U.C. Berkeley [email protected] [email protected] Sanjay Mehrotra MS36 IEMS Parallel Nonlinear Interior-Point Methods for Northwestern University Stochastic Programming Problems [email protected]

The dominant computational cost in a nonlinear interior- point algorithms like IPOPT is solution of the linear KKT MS36 system at each iteration. When solving nonlinear stochas- Covariance Estimation and Relaxations for tic programming problems, the linear system can be solved Stochastic Unit Commitment in parallel with a Schur-complement decomposition. In this presentation, we show the parallel performance of a PCG- Abstract not available. based Schur-complement decomposition using an L-BFGS preconditioner that avoids the cost of forming and factor- Cosmin G. Petra ing the Schur-complement. Argonne National Laboratory Mathematics and Computer Science Division Yankai Cao [email protected] School of Chemical Engineering Purdue University [email protected] MS37 Multi-Phase Optimization of Elastic Materials, Us- Jia Kang ing a Level Set Method Department of Chemical Engineering Texas A&M University Abstract not available. [email protected] Charles Dapogny Laboratoire Jacques-Louis Lions Carl Laird Paris, France Assistant Professor [email protected] Texas A&M University [email protected] MS37 MS36 Structural Optimization Based on the Topological Stochastic Model Predictive Control with Non- Gradient Gaussian Disturbances Abstract not available. We present a stochastic model predictive control algorithm designed to handle non-Gaussian disturbance distributions. Volker H. Schulz The method relies on piecewise polynomial parameteri- University of Trier zation of the distribution, for which fast exact convolu- Department of Mathematics tion is tractable. The problem formulation minimizes ex- [email protected] pected cost subject to linear dynamics, polyhedral hard constraints on inputs and chance constraints on states. Roland Stoffel Joint chance constraints are approximated using Boole’s in- University of Trier equality, and conservatism in this approximation is reduced Trier, Germany by optimal risk allocation and affine disturbance feedback. stoff[email protected] Tony Kelman, Francesco Borrelli University of California, Berkeley MS37 [email protected], [email protected] A New Algorithm for the Optimal Design of Anisotropic Materials

MS36 A new algorithm for the solution of optimal design prob- A Two Stage Stochastic Integer Programming lems with control in parametrized coefficients is discussed. OP14 Abstracts 121

The algorithm is based on the sequential convex program- Nghia Tran ming idea, however, in each major iteration a model is es- Department of Mathematics tablished on the basis of the parametrized material tensor. University of British Columbia The potentially nonlinear parametrization is then treated [email protected] on the level of the sub-problem, where, due to block separa- bility of the model, global optimization techniques can be applied. Theoretical properties of the algorithm like global MS38 convergence are discussed. The effectiveness of the algo- New Techniques in Convex Analysis rithm is demonstrated by a series of numerical examples from the area of parametric and two-scale material design Abstract not available. involving optimal orientation problems. Jon Vanderwerff Michael Stingl Department of Mathematics Institut f¨ur Angewandte Mathematik La Sierra University Universit¨at Erlangen-N¨urnberg, Germany [email protected] [email protected] MS38 Jannis Greifenstein Universit¨at Erlangen-N¨urnberg, Germany Resolvent Average of Monotone Operators [email protected] Monotone operators are important in optimization. We provide a new technique to average two or more monotone MS37 operators. The new maximal monotone operator inherits many nice properties of given monotone operators. Optimization of the Fibre Orientation of Or- thotropic Material Shawn Xianfu Wang Department of Mathematics Abstract not available. University of British Columbia Okanagan Heinz Zorn [email protected] University of Trier [email protected] MS39 The CP-Matrix Completion Problem MS38 A symmetric matrix C is completely positive (CP) if there SI Spaces and Maximal Monotonicity exists an entrywise nonnegative matrix B such that C = T We introduce SI spaces, which include Hilbert spaces, neg- BB . The CP-completion problem is to study whether we ative Hilbert spaces and spaces of the form E×E∗,whereE can assign values to the missing entries of a partial ma- is a nonzero reflexive real Banach space. We introduce L– trix (i.e., a matrix having unknown entries) such that the positive subsets, which include monotone subsets of E×E∗, completed matrix is completely positive. CP-completion and BC–functions, which include Fitzpatrick functions of has wide applications in probability theory, the nonconvex monotone multifunctions. We show how convex analysis quadratic optimization, etc. In this talk, we will propose an canbecombinedwithSIspacetheorytoobtainandgen- SDP algorithm to solve the general CP-completion prob- eralize various results on maximally monotone multifunc- lem and study its properties. Computational experiments tions on a reflexive Banach space, such as the significant di- will also be presented to show how CP-completion prob- rection of Rockafellar’s surjectivity theorem, and sufficient lems can be solved. conditions for the sum of maximally monotone multifunc- tions to be maximally monotone. If time permits, we will Anwa Zou also give an abstract Hammerstein theorem. Department of Mathematics Shanghai Jiaotong University Stephen Simons [email protected] ¡[email protected]¿ Department of Mathematics University of California Santa Barbara Jinyan Fan [email protected] Department of Mathematics, Shanghai Jiaotong University Shanghai 200240, China MS38 [email protected] Coderivative Characterizations of Maximal Mono- tonicity with Applications to Variational Systems MS39 In this talk we provide various characterizations of maxi- Convexity in Problems Having Matrix Unknowns mal monotonicity in terms of regular and limiting coderiva- tives, which are new in finite-dimensional and infinite- Problems in classical system engineering have unknowns dimensional frameworks. We also develop effective applica- which are matrices appearing in a polynomial determined tions to obtaining Lipschitzian continuity of solution maps by the system’s “signal flow diagram”. Recent results show to conventional models of variational systems and varia- that if such a situation is convex, then there is also an LMI tional conditions in nonlinear programming. associated to it- a rather negative result to an engineer who is desperate for convexity. The talk will discuss asso- Boris Mordukhovich ciated topics, eg. , matrix convex hulls and consequences Wayne State University for polynomial inequalities where convexity is involved. Department of Mathematics [email protected] J.William Helton 122 OP14 Abstracts

University of California, San Diego [email protected] [email protected]

MS40 MS39 Effective Quadratization of Nonlinear Binary Op- Optimizing over Surface Area Measures of Convex timization Problems Bodies We consider large nonlinear binary optimization problems We are interested in shape optimization problems under arising in computer vision applications. Numerous recent convexity constraints. Using the notion of surface area techniques transform high degree binary functions into measure introduced by H. Minkowski and refined by A. equivalent quadratic ones, using additional variables. In D. Alexandrov, we show that several well-known problems this work we study the space of ”quadratizations,” give of this class (constant width body of maximum volume, sharp lower and upper bounds on the number of additional Newton’s body of minimum resistance) can be formulated variables needed, and provide new algorithms to quadra- as linear programming (LP) problems on Radon measures. tize high degree functions. (Joint work with Martin An- Then we show that these measure LPs can be solved ap- thony (LSE), Yves Crama (University of Liege) and Ari- proximately with a hierarchy of semidefinite programming tanan Gruber (Rutgers University).) (SDP) relaxations. The support function of the almost op- timal convex body is then recovered from its surface area Endre Boros measure by solving the Minkowski problem with the SDP Rutgers University hierarchy. Joint work with Terence Bayen (University of [email protected] Montpellier, France). MS40 Didier Henrion University of Toulouse, France Lower Bounds on the Graver Complexity of M-Fold [email protected] Matrices In this talk, we present a construction that turns certain MS39 relations on Graver basis elements of an M-fold matrix A(M) into relations on Graver basis elements of A(M+1). On Level Sets of Polynomials and a Generalization In doing so, we give a lower bound on the Graver com- of the Lowner’s John Problem (M) (M) plexity g(A ) for general M-fold matrices A and we strengthen the bound on the Graver complexity of A3×M We investigate some properties of level sets of polynomials M−3 from g(A3×M ) ≥ 17 · 2 − 7 (Berstein and Onn) to and solve a generalization of the Lowner’s John problem. M−3 g(A3×M ) ≥ 24 · 2 − 21. Jean B. Lasserre LAAS CNRS and Institute of Mathematics Elisabeth Finhold [email protected] Zentrum Mathematik, M9 Technische Universit¨at M¨unchen fi[email protected] MS40 Hierarchical Cuts to Strengthen Semidefinite Re- Raymond Hemmecke laxations of NP-hard Graph Problems Dept. of Mathematics TU Munich, Germany The max-cut problem can be closely approximated us- [email protected] ing the basic semidefinite relaxation iteratively refined by adding valid inequalities. We propose a projection poly- MS40 tope as a new way to improve the relaxations. These cuts are based on requiring the solution to be valid for smaller BiqCrunch in Action: Solving Difficult Combina- cut polytopes. Finding new cuts creates a hierarchy that torial Optimization Problems Exactly iteratively tightens the semidefinite relaxation in a con- trolled manner. Theoretical and computational results will We present different combinatorial optimization problems be presented. that can be solved by BiqCrunch, a semidefinite-based branch-and-bound solver: MaxCut, Max-k-Cluster, Max- Elspeth Adams imum Independent Set, k-Exact Assignment, and the Ecole Polytechnique de Montreal Quadratic Assignment Problem. With each problem, we [email protected] describe some of the techniques used in formulating them, such as adding reinforcing constraints, and making special heuristics. We show numerical results that demonstrate Miguel Anjos ´ the efficiency of BiqCrunch on these problems. Note that Ecole Polytechnique de Montr´eal N. Krislock’s talk ”BiqCrunch: a semidefinite branch-and- [email protected] bound method for solving binary quadratic problems” ex- plains in details the underlying theory of BiqCrunch. Franz Rendl Alpen-Adria Universitaet Klagenfurt Fr´ed´eric Roupin Austria Laboratoire d’Informatique Paris-Nord [email protected] Institut Galil´ee [email protected] Angelika Wiegele Alpen-Adria-Universit{¨a}t Klagenfurt Nathan Krislock Institut f. Mathematik INRIA Grenoble Rhˆone-Alpes OP14 Abstracts 123

[email protected] scale learning applications demonstrate the efficiency and robustness of the approach. Jerome Malick CNRS Jorge Nocedal [email protected] Department of Electrical and Computer Engineering Northwestern University [email protected] MS41 An Active Set Strategy for 1 Regularized Prob- MS41 lems An Iterative Reweighting Algorithm for Exact The problem of finding sparse solutions to underdeter- Penalty Subproblems mined systems of linear equations arises from several ap- Matrix free alternating direction and reweighting meth- plication problems (e.g. signal and image processing, com- ods are proposed for solving exact penalty subproblems. pressive sensing, statistical inference). A standard tool The methods attain global convergence guarantees under for dealing with sparse recovery is the 1-regularized least- mild assumptions. Complexity analyses for both meth- squares approach that has been recently attracting the at- ods are also discussed. Emphasis is placed on their ability tention of many researchers. Several ideas for improving to rapidly find good approximate solutions, and to han- the Iterative Shrinkage Thresholding (IST) algorithm have dle large-scale problems. Numerical results are presented been proposed. We present an active set estimation that for elastic QP subproblems arising in NLP algorithms and can be included within IST-algorithm and its variants. We 1-SVM problems. report numerical results on some test problems showing the efficiency of the proposed strategy. Hao Wang Industrial and Systems Engineering Marianna De Santis, Stefano Lucidi Lehigh University University of Rome [email protected] [email protected], [email protected] James V. Burke University of Washington Francesco Rinaldi Department of Mathematics University of Padova [email protected] [email protected] Frank E. Curtis MS41 Industrial and Systems Engineering Lehigh University Second Order Methods for Sparse Signal Recon- [email protected] struction

In this talk we will discuss efficient second order meth- Jiashan Wang ods for a family of 1-regularized problems from the University of Washington field sparse signal reconstruction. The problems include Department of Mathematics isotropic total-variation, 1-analysis, the combinations of [email protected] those two and 1-regularized least-squares. Although first- order methods have dominated these fields, we will argue that specialized second order methods offer a viable alter- MS42 native. We will provide theoretical analysis and computa- Optimal Fractionation in Radiotherapy tional evidence to illustrate our findings. The prevalent practice in radiotherapy is to first obtain a Kimon Fountoulakis radiation intensity profile by solving a fluence-map opti- University of Edinburgh mization problem and then deliver the resulting dose over [email protected] a fixed number of equal-dosage treatment sessions. We will present convex programming methods that simultaneously optimize the number of treatment sessions and the inten- Ioannis Dassios sity profile using the linear-quadratic dose-response model. Research assistant (PostDoc) Our models and methods will also be extended to allow idassios@staffmail.ed.ac.uk non-stationary dosing schedules.

Jacek Gondzio Minsun Kim, Fatemeh Saberian, Archis Ghate The University of Edinburgh, United Kingdom University of Washington [email protected] [email protected], [email protected], [email protected] MS41 A Stochastic Quasi-Newton Method MS42 A Stochastic Optimization Approach to Adaptive We propose a stochastic approximation method based on Lung Radiation Therapy Treatment Planning the BFGS formula. The algorithm is of the type proposed by Robbins-Monro, but scaled by a full quasi-Newton ma- Intensity Modulated Radiation Therapy is a common tech- trix (that is defined implicitly). We show how to control nique used to treat cancer. Treatment planners optimize noise in the construction of the curvature correction pairs a physician’s treatment planning goals using patient infor- that determined the BFGS matrix. Results on three large mation obtained pre-treatment while remaining within the 124 OP14 Abstracts

constraints of the treatment modality. Biomarker data ob- ality information from perfusion maps to avoid delivering tained during treatment (post-planning) has been shown dose to well-perfused areas. We use liver cancer cases to to be predictive of a patient’s predisposition to radiation- validate our model. induced toxicity. We develop a two-stage stochastic treat- ment plan optimization model to explicitly consider future Victor Wu patient-specific biomarker information. Several adaptive University of Michigan treatment strategies are studied. [email protected]

Troy Long Marina A. Epelman University of Michigan University of Michigan [email protected] Industrial and Operations Engineering [email protected] Marina A. Epelman University of Michigan Mary Feng Industrial and Operations Engineering University of Michigan [email protected] Radiation Oncology [email protected] Martha Matuszak University of Michigan Martha Matuszak [email protected] University of Michigan [email protected] Edwin Romeijn University of Michigan Edwin Romeijn Industrial and Operations Engineering University of Michigan [email protected] Industrial and Operations Engineering [email protected] MS42 A Mathematical Optimization Approach to the MS43 Fractionation Problem in Chemoradiotherapy Stochastic Block Mirror Descent Methods for Non- smooth and Stochastic Optimization In concurrent chemoradiotherapy (CRT) chemotherapeutic agents are administered during the course of radiotherapy In this paper, we present a new stochastic algorithm, to enhance the primary tumor control. That often comes namely the stochastic block mirror descent (SBMD) at the expense of increased risk of normal-tissue complica- method for solving large-scale nonsmooth and stochas- tions. The additional biological damage is attributed to the tic optimization problems. The basic idea of this algo- drug cytotoxic activity and its interactive cooperation with rithm is to incorporate the block-coordinate decomposition radiation, which is dose and time dependent. We develop and an incremental block averaging scheme into the clas- a mathematical optimization framework to determine ra- sic (stochastic) mirror-descent method, in order to signifi- diation and drug administration schedules that maximize cantly reduce the cost per iteration of the latter algorithm. the therapeutic gain for concurrent CRT. We establish the rate of convergence of the SBMD method along with its associated large-deviation results for solving Ehsan Salari general nonsmooth and stochastic optimization problems. Wichita State University We also introduce different variants of this method and [email protected] establish their rate of convergence for solving strongly con- vex, smooth, and composite optimization problems, as well Jan Unkelbach as certain nonconvex optimization problems. Department of Radiation Oncology Massachusetts General Hospital and Harvard Medical Cong D. Dang School University of Florida [email protected] congdd@ufl.edu

Thomas Bortfeld Guanghui Lan Massachusetts General Hospital and Harvard Medical Department of Industrial and Systems School University of Florida Department of Radiation Oncology [email protected]fl.edu [email protected]

MS43 MS42 Universal Coordinate Descent Method Incorporating Organ Functionality in Radiation Therapy Treatment Planning We design and analyze universal coordinate descent meth- ods, that is methods that can efficiently solve smooth as Goals of radiotherapy treatment planning include (i) erad- well as nonsmooth problems without knowing the level of icating tumor cells and (ii) sparing critical structures to smoothness. We propose a line search strategy that does ultimately preserve functionality. A good indicator of lo- not involve function evaluations and derive iteration com- cal and global function, perfusion (blood flow) shows that plexity bounds. We also study a parallel version of the organ function is non-homogenous. Thus, spatial distribu- method. The key technical tool is the introduction of non- tion of dose in critical structures matters. We propose an quadratic separable overapproximations that allow us to optimization model that explicitly incorporates function- define and compute in parallel the update of each coordi- OP14 Abstracts 125

nate. [email protected]

Olivier Fercoq INRIA Saclay and CMAP Ecole Polytechnique MS44 [email protected] An Efficient Affine-scaling Algorithm for Hyper- bolic Programming Peter Richtarik University of Edinburgh Hyperbolic programming (HP) is a generalization of [email protected] semidefinite programming. Hyperbolicity cones, however, in general are not symmetric, limiting the number of interior-point methods that can be applied. For example, MS43 the variants of Dikin’s affine-scaling algorithm that have Iteration Complexity Analysis of Block Coordinate been shown to run in polynomial time depend heavily on Descent Methods the underlying cone being symmetric, and thus do not ap- ply to HP. Until now, that is. We present a natural variant We provide a unified iteration complexity analysis for a of Dikin’s algorithm that applies to HP, and establish a family of block coordinate descent (BCD)-type algorithms, complexity bound that is no worse than√ the best bound covering both the classic Gauss-Seidel coordinate update known for interior-point methods (“O( n) iterations to rule and the randomized update rule. We show that for a halve the duality gap”). Interestingly, the algorithm is pri- broad class of nonsmooth convex problems, the BCD-type mal, but the analysis is almost entirely dual (and provides algorithms, including the classic BCD, the block coordi- new geometric insight into hyperbolicity cones). nate gradient descent and the block coordinate proximal gradient methods, have an iteration complexity of O(1/r), James Renegar where r is the iteration index. Cornell University [email protected] Zhi-Quan Luo Univerisity of Minnesota Minneapolis, Minnesota MS44 [email protected] On the Sonnevend Curvature and the Klee-Walkup Result MS43 We consider Sonnevend’s (1991) curvature integral of the On the Complexity Analysis of Randomized Coor- centralpath(CP)ofLO.Ourmainresultstatesthatfor dinate Descent Methods establishing an upper bound for the total Sonnevend cur- vature, it is sufficient to consider the case n =2m.Our We give refined analysis of the randomized block- construction yields an Ω(m) worst-case lower bound for coordinate descent (RBCD) method proposed by Richtarik Sonnevend’s curvature. Our results are analogous to Deza and Takac for minimizing the sum of a smooth convex func- et al.’ (2008) results for the geometric curvature and to tion and a block-separable convex function. We obtain the Klee-Walkup result about a polytope’s diameter. sharper bounds on its convergence rate, both in expected value and in high probability. For unconstrained smooth Tamas Terlaky convex optimization, we develop a randomized estimate Lehigh University sequence technique to establish a sharper expected value Department of industrial and Systems Engineering type of convergence rate for Nesterov’s accelerated RBCD [email protected] method.

Zhaosong Lu MS44 Department of Mathematics Simon Fraser University On the Curvature of the Central Path for Lp [email protected] Similarly to the diameter of a polytope, one may define its Lin Xiao curvature based on the worst-case central path associated Microsoft Research with solving an LP posed over the polytope. Furthermore, [email protected] a continuous analogue of the Hirsch conjecture and a dis- crete analogue of the average curvature result of Dedieu, Malajovich and Shub may be introduced. A continuous MS44 analogue of the result of Holt and Klee –a polytope con- struction that attains a linear order largest total curvature– Towards a Computable O(n)-Self-Concordant Bar- n and a continuous analogue of a d-step equivalence result for rier for Polyhedral Sets in the diameter of a polytope may also be proved. We survey It is known that for a polyhedral set in n determined by the recent progress towards better understanding of the m linear inequalities, there is a self-concordant barrier (the curvature. universal barrier) whose parameter is O(n), as opposed to m for the usual logarithmic barrier. However there is no Yuriy Zinchenko known computable barrier whose parameter is O(n). We University of Calgary describe one possible approach for construcing such a bar- Department of Mathematics and Statistics rier based on maximum volume inscribing ellipsoids. [email protected]

Kurt M. Anstreicher University of Iowa MS45 Department of Management Sciences An Algorithm for Solving Smooth Bilevel Opti- 126 OP14 Abstracts

mization Problems Tecnologico de Monterrey (ITESM), Campus Monterrey [email protected] Our basis is a standard nonsmooth, nonconvex optimistic bilevel programmig problem with the upper level feasible Roberto C. Herrera Maldonado set not depending on the lower level variable. We use the Dept. Systems & Industrial Engineering (IIS) optimal value reformulation and apply partial calmness to Tecnologico de Monterrey (ITESM), Campus Monterrey the resulting problem which allows us to formulate suit- [email protected] able constraint qualifications. The bundle algorithm by K.C. Kiwiel, A proximal bundle method with approximate subgradient linearizations, SIAM Journal on Optimization MS45 16 (2006), 1007–1023 is extended to the nonconvex case. Solving Bilevel Programs by Smoothing Techniques

Susanne Franke, Stephan Dempe We propose to solve bilevel programs by reformulating it TU Bergakademie Freiberg as a combined program where both the value function con- [email protected], [email protected] straint and the first order condition are present. Using freiberg.de smoothing technique, we approximate the value function by a smoothing function, use certain penalty based meth- MS45 ods to solve the smooth problem and drive the smoothing parameter to infinity. We prove the convergence of the al- On Regular and Limiting Coderivatives of the Nor- gorithm under a weak extended generalized Mangasarian- mal Cone Mapping to Inequality Systems Fromovitz constraint qualification.

In modern variational analysis, generalized derivatives are Jane J. Ye the essential ingredients for deriving stability results for University of Victoria solutions maps of generalized equations as well as stating [email protected] optimality conditions for mathematical programs. For the special case of generalized equations involving the normal 2 cone mapping to a system of C inequalities, we give work- MS46 able formulas for the graphical derivative and the regu- Spectral Bounds and Preconditioners for Unre- lar respectively limiting coderivative, when LICQ does not duced Symmetric Systems Arising from Interior hold at the reference point. We use neighborhood based Point Methods first order constraint qualifications, which, however, can be characterized by point based second order conditions. The focus of this talk is on the linear systems that arise from the application of Primal-Dual Interior Point methods Helmut Gfrerer to convex quadratic programming problems in standard Johannes Kepler Universitaet Linz form. The block structure of the matrices allows for for- Austria mulations differing in their dimensions and spectral prop- [email protected] erties. We consider the unreduced symmetric 3x3 block formulation of such systems, provide new spectral bounds Jiri Outrata and discuss the impact of the formulation used on precon- Czech Academy of Sciences ditioning techniques. [email protected] Benedetta Morini Universita’ di Firenze MS45 benedetta.morini@unifi.it A Heuristic Approach to Solve the Bilevel Toll As- signment Problem by Making Use of Sensitivity Valeria Simoncini, Mattia Tani Analysis Universita’ di Bologna [email protected], [email protected] Toll Optimization Problem is considered as a bilevel pro- gramming model. At the upper level, a public regulator or a private company manages the toll roads to increase MS46 its profit. At the lower level, several users try to satisfy A General Krylov Subspace Method and its Con- the existing demand for transportation of goods and/or nection to the Conjugate-gradient Method passengers, and simultaneously, to select the routes so as to minimize their travel costs. To solve this bilevel pro- Krylov subspace methods are used for solving systems of gramming problem, a direct algorithm based on sensitivity linear equations Ax = b. We present a general Krylov analysis is proposed. subspace method that can be used when the matrix A is indefinite. We show that the approximate solution in each Vyacheslav V. Kalashnikov step, xk, is given by the chosen scaling of the orthogo- Universidad Aut´onoma de Nuevo Le´on nal vectors that successively make the Krylov subspaces Mexico available. Our framework gives an intuitive way to see the [email protected] conjugate-gradient method and why it sometimes fails if A is not positive definite. Nataliya I. Kalashnykova Dept. of Physics & Maths (FCFM), Tove Odland Universidad Autonoma de Nuevo Leon Department of Mathematics [email protected] Stockholm Royal Institute of Technology [email protected] Aaron Arevalo Franco Dept. Systems & Industrial Engineering Anders Forsgren OP14 Abstracts 127

KTH Royal Institute of Technology [email protected] Dept. of Mathematics [email protected] MS47 Function-Value-Free Continuous Optimization in MS46 High Dimension Asqr: Interleaving Lsqr and Craig Krylov Methods Information geometric optimization provide a generic for the Solution of Augmented Systems framework of function-value-free stochastic search algo- rithms for arbitrary optimization, including the state-of- We consider the solution of standard augmented systems, the-art stochastic search algorithm for continuous opti- which can actually be split into the sum of a least-squares mization, the CMA-ES. From this framework, we derive a solution and a minimum-norm solution. We thus propose novel algorithm for high dimensional continuous optimiza- to interleave both left and right Krylov sequences from tion, whose time and space complexity for each iteration is LSQR and Craig methods, with corresponding starting linear in the dimension of the problem. vectors, and to compute at the same time an approxima- tion of the least-squares solution and an approximation of Youhei Akimoto the minimum norm solution, both by means of appropri- Shinshu University, Japan ate restrictions onto these interleaved Krylov basis, and to y [email protected] sum them for an approximation of the solution at each it- eration. This interleaving procedure leads to a tridiagonal- ization procedure of the constraint rectangular sub-matrix, MS47 which can be factorized in QR form with Givens rotations. Fifty Years of Randomized Function-Value-Free Al- Putting this altogether yields the method that we denote gorithms: Where Do We Stand? as ASQR, the acronym standing for “Augmented-system Solution with QR factorization’. We shall discuss and il- The first randomized function-value-free (FVF) or lustrate numerically the different aspects of this method. comparison-based algorithms were already introduced in the 60’s. There is nowadays a renewal of interests for those algorithms in the context of derivative-free optimization. We will discuss important design principles of randomized Daniel Ruiz FVF algorithms like invariances and emphasize on the the- ENSEEIHT, Toulouse, France oretical properties that follow from those principles, for [email protected] example linear convergence or learning of second order in- formation. Alfredo Buttari CNRS-IRIT, France Anne Auger [email protected] INRIA Saclay [email protected] David Titley-Peloquin CERFACS - France [email protected] MS47 Evolutionary Multiobjective Optimization and Function-Value-Free Algorithms MS46 Evolutionary Multiobjective Optimization Algorithms Using Partial Spectral Information for Block Diag- (EMOAs) are derivative-free, stochastic optimization algo- onal Preconditioning of Saddle-Point Systems rithms to approximate the Pareto front of multiobjective optimization problems. The approximations are thereby We focus on KKT systems with very badly conditioned typically of a fixed, pre-determined size, the algorithms’ symmetric positive definite (1,1) blocks due to few very so-called population size. Interestingly, when compared to small eigenvalues. Assuming availability of these eigenval- the single-objective case, all state-of-the-art EMOAs are ues and associated eigenvectors, we consider different ap- currently not function-value free and this talk discusses the proximations of the block diagonal preconditioner of Mur- main reasons behind and possible remedies for it. phy, Golub and Wathen that appropriately recombine the available spectral information through a particular Schur Dimo Brockhoff complement approximation built at little extra cost. We Inria Lille - Nord Europe, France also discuss under which circumstances these eigenvalues dimo.brockhoff@inria.fr effectively spoil the convergence of MINRES. MS47 Daniel Ruiz X-NES: Exponential Coordinate System Parame- ENSEEIHT, Toulouse, France terization for FVF Search [email protected] Online adaptation of mean vector and covariance matrix is Annick Sartenaer a key step in randomized FVF optimization with Gaussian University of Namur search distributions. Adaptation of the covariance matrix Department of Mathematics must handle the positive definiteness constraint. This talk [email protected] presents a solution proposed in the exponential natural evo- lution strategy algorithm, namely a parameterization of the Chalotte Tannier covariance matrix with the exponential map. The result- University of Namur ing multiplicative update blends smoothly with step size 128 OP14 Abstracts

adaptation rules. Its tangent space decomposes naturally advantage of the enhanced BD. into meaningful, independent components. Emmanuel Ogbe, Xiang Li Tobias Glasmachers Queen’s University Ruhr-Universit¨at Bochum, Germany [email protected], xi- [email protected] [email protected]

MS48 MS48 Semismooth Equation Approaches to Structured Dynamic System Design based on Stochastic Opti- Convex Optimization mization Structured convex optimization problems arise in the con- The common design practice in industry has traditionally text of resource allocation problems, network constrained been based on steady-state requirement for nominal and optimization, machine learning, stochastic programming, fully specified conditions considering economic and various optimization-based control to name a few. Popular ap- constraints. Afterwards control systems are deisgned to proach to solving structured convex optimization utilizes remedify the operability bottlenects and problems caused dual decomposition of complicating constraints to yield by the dynamics not considered in the design phase. In smaller subproblems. A subgradient algorithm, which is this sequential process, overdeign based on engineering in- extremely slow to converge, is typically employed to ob- sight and knowledge is usually exercised. Nevertheless, tain convergence of dualized constraints. We propose a it has been realized in many applications that the steady novel semismooth equation approach to solving the stan- state design may yield systems which are difficult to con- dard dual decomposition formulation. The approach relies trol due to the lack of consideration of operability, con- on the ability to compute sensitivity of the subproblem so- trollability, flexibility and stability in the deisgn phase. In lutions to the multipliers of the dualized constraints. We addition, this sequential approach may also lead to more show that under certain assumptions the approach con- expensive or less efficient design options. With the recent verges locally superlinearly to the solution of the origi- advances in simulation and computational algorithms, de- nal problem. Globalization of the proposed algorithm us- sign problems include controllability and control system ing a linesearch is also described. Numerical experiments design have emerged in the literature [Flores-Tlacuahuac, compare the performance against state-of-the-art nonlinear A and Biegler, L. Simultaneous control and process de- programming solver. sign during optimal grade transition operations. Comput. Chem. Eng., 32:2823, 2008]. In addition, stochastic and ro- Arvind Raghunathan bust optimization formulations have been proposed to deal United Technologies Research Center with parametric uncertainty and disturbances. However, [email protected] robust optimization formulation may lead to conservative solutions and the size of the stochastic optimization prob- Daniel Nikovski lem maybe limited due to its sampling nature. This work Mitsubishi Electric Research Laboratories presents a case study of optimal dynamic system design [email protected] considering uncertainty based on Generalized Polynomial Chaos. It represents the effects of uncertainties in second order random proceses with guarrantted convergence. As MS48 a result, a stochastic optimziation problem is formulated A Scalable Clustering-Based Preconditioning considering the economic, design, controllability and flexi- Strategy for Stochastic Programms bility objectives based on a dynamic model of the system of interest, including path constraints, end point constri- We present a scalable implementation of a multi-level ant, and generalized polynomial chaos model of uncertainty clustering-based preconditioning strategy for structured parameters. KKT systems arising in stochastic programs. We demon- strate that the strategy can achieve compression rates of over 80% while retaining good spectral properties. Case Rui Huang studies arising from power grid systems are presented. United Technologies [email protected] Victor Zavala Argonne National Laboratory Baric Miroslav [email protected] United Technologies Research Center [email protected] MS49 Adaptive and Robust Radiation Therapy in the MS48 Presence of Motion Drift

Enhanced Benders Decomposition for Stochastic We present an adaptive, multi-stage robust optimization Mixed-integer Linear Programs approach to radiation therapy treatment planning for a lung cancer case subject to tumor motion. It was previ- Stochastic mixed-integer linear programs often hold a de- ously shown that this approach is asymptotically optimal composable structure that can be exploited by Benders de- if the uncertain tumor motion distribution stabilizes over composition (BD). However, BD may suffer from the tail- the treatment course. We show computationally that this ing effect, i.e., the lower and upper bounds tend to con- approach maintains good performance even if the motion verge very slowly after the optimal objective value has been distribution does not stabilize but instead exhibits drifts reached. We propose to enhance the BD bounds via solving within the probability simplex. extra subproblems that come from Dantzig-Wolfe decom- position strategy, and case study results demonstrate the Timothy Chan OP14 Abstracts 129

Department of Mechanical & Industrial Engineering For proton therapy, it is shown that delivering different University of Toronto, Canada dose distributions in subsequent fractions may yield a ther- [email protected] apeutic advantage.

Philip Mar Jan Unkelbach Department of Mechanical and Industrial Engineering Department of Radiation Oncology University of Toronto Massachusetts General Hospital and Harvard Medical [email protected] School [email protected]

MS49 Martijn Engelsman Mathematical Modeling and Optimal Fractiona- TU Delft, The Netherlands tionated Irradiation for Proneural Glioblastomas [email protected]

Glioblastomas (GBM) are the most common and ma- lignant primary tumors of the brain and are commonly MS50 treated with radiation therapy. Despite modest advances Distributed Optimization over Time-Varying Di- in chemotherapy and radiation, survival has changed very rected Graphs little over the last 50 years. Radiation therapy is one of the pillars of adjuvant therapy for GBM but despite treatment, We consider distributed optimization by a collection of recurrence inevitably occurs. Here we develop a mathemat- nodes, each having access to its own convex function, whose ical model for the tumor response to radiation that takes collective goal is to minimize the sum of the functions. into account the plasticity of the hierarchical structure of The communications between nodes are described by a the tumor population. Based on this mathematical model time-varying sequence of directed graphs, which is uni- we develop an optimized radiation delivery schedule. This formly strongly connected. For such communications, as- strategy was validated to be superior in mice and nearly suming that every node knows its out-degree, we develop a doubled the efficacy of each Gray of radiation administered. broadcast-based algorithm, termed the subgradient-push, This is based on joint work with Ken Pitter, Eric Holland, which steers every node to an optimal value under a and Franziska Michor.** standard assumption of subgradient boundedness. The subgradient-push requires no knowledge of either the num- Kevin Leder ber of agents or the graph sequence to implement. Our University of Minnesota analysis shows that the subgradient-push algorithm con- Industrial and Systems Enginering verges at a rate of O(ln t/t), where the constant depends [email protected] on the initial values at the nodes, the subgradient norms, and, more interestingly, on both the consensus speed and the imbalances of influence among the nodes. MS49 A Column Generation and Routing Approach to Angelia Nedich 4π Vmat Radiation Therapy Treatment Planning Industrial and Enterprise Systems Engineering University of Illinois at Urbana-Champaign Volumetric Modulated Arc Therapy (VMAT) is rapidly [email protected] emerging as a method for delivering radiation therapy treatments to cancer patients that is of comparable quality Alexander Olshevsky to IMRT but much more efficient. Since VMAT only uses University of Illinois at Urbana-Champaign coplanar beam orientations, the next step is to consider [email protected] non-coplanar orientations as well. This greatly complicates the radiation therapy treatment plan optimization problem by introducing a routing component. We propose a con- MS50 structive approach that employs both column generation On the Solution of Stochastic Variational Problems and routing heuristics. with Imperfect Information

Edwin Romeijn We consider the solution of a stochastic convex optimiza- National Science Foundation tion problem E[f(x; θ∗,ω)] over a closed and convex set X [email protected] in a regime where θ∗ is unavailable. Instead, may be learnt by minimizing a suitable stochastic metric in θ over a closed Troy Long and convex set . We present a coupled stochastic approx- University of Michigan imation scheme for the associated stochastic optimization [email protected] problem with imperfect information. It is shown that the resulting schemes are shown to be equipped with almost sure convergence properties in regimes where the function MS49 f is both strongly convex as well as merely convex. We Combined Temporal and Spatial Treatment Opti- show that such schemes display the optimal rate of con- mization vergence in convergence and quantify the degradation in convex regimes by a modest amount. We examine exten- We consider the interdependence of optimal fractionation sions to variational regimes and provide some numerics to schemes and the spatial dose distribution in radiotherapy illustrate the workings of the scheme. planning. This is approached by simultaneously optimiz- ing multiple distinct treatment plans for different fractions Uday Shanbhag based on the biologically equivalent dose (BED) model. Department of Industrial and Manufacturing Engineering This leads to a large-scale non-convex optimization prob- Pennsylvania State University lem, arising from the quadratic relation of BED and dose. [email protected] 130 OP14 Abstracts

Hao Jiang [email protected] University of Illinois [email protected] MS51 Constructive Approximation Approach to Polyno- MS50 mial Optimization Stochastic Methods for Convex Optimization with In this talk we explore the links between some construc- ”Difficult” Constraints tive approximation approaches and polynomial optimiza- tion. In particular, we revisit some approximation results Convex optimization problems involving large-scale data or for polynomial optimization over a simplex by exploring the expected values are challenging, especially when these dif- links with Bernstein approximation and other constructive ficulties are associated with the constraint set. We propose approximation schemes. random algorithms for such problems, and focus on special Etienne De Klerk structures that lend themselves to sampling, such as when Tilburg University the constraint is the intersection of many simpler sets, in- [email protected] volves a system of inequalities, or involves expected values, and/or the objective function is an expected value, etc. We propose a class of new methods that combine elements of Monique Laurent successive projection, stochastic gradient descent and prox- CWI, Amsterdam imal point algorithm. This class of methods also contain and Tilburg University as special cases many known algorithms. We use a unified [email protected] analytic framework to prove their almost sure convergence and the rate of convergence. Our framework allows the Zhao Sun random algorithms to be applied with various sampling Tilburg University schemes (e.g., i.i.d., Markov, sequential, etc), which are [email protected] suitable for applications involving distributed implementa- tion, large data set, computing equilibriums, or statistical learning. MS51 Polynomial Optimization in Biology Mengdi Wang The problem of inferring the phylogenic tree of n extant Department of Operations Research and Financial species from their genome is treated via mainly two ap- Engineering proaches, the maximum parsimony and the maximum like- Princeton University lihood approach. While the latter is thought to produce [email protected] more meaningful results, it is harder to solve computation- ally.We show that under a molecular clock assumption, the maximum likelihood model has a natural formulation that MS50 can be approached via branch-and-bound and polynomial optimization. Using this approach, we produce a coun- On Decentralized Gradient Descent terexample to a conjecture relating to the reconstruction of an ancestral genome under the maximum parsimony and n Consider the consensus problem of minimizing i=1 fi(x) maximum likelihood approaches. where each fi is only known to one individual agent in a connected network of many agents. All the agents shall col- Raphael A. Hauser laboratively solve this problem subject to data exchange Oxford University Computing Laboratory restriction to between neighbor agents. Such algorithms [email protected] avoids the need of a fusion center and long-distance com- munication. It offers better network load balance and im- proves data privacy. We study the decentralized gradient MS51 method in which each agent i updates its copy of vari- Noncommutative Polynomials able x(i) by averaging of its neighbors’ variables x(j) and moving along the negative gradient −α∇fi(x(i).Thatis, Abstract not available. x(i) ← (i,j)∈E wij x(j) − α∇fi(x(i)). We show that if the ∗ Igor Klep problem has a solution x and all fi’s are proper closed convex with Lipschitz continuous gradients, then provided Department of Mathematics The University of Auckland stepsize α<(1 + λminW )/2, the function value at the av- erage, f(¯x(k)) − f ∗, reduces at a speed of O(1/k) until it [email protected] reaches O(1/α). If fi are further (restricted) strongly con- vex, then each x(i)(k) converges to the global minimizer at MS51 a linear rate until reaching an O(1/α)-neighborhood of x∗. The Hierarchy of Local Minimums in Polynomial Optimization Kun Yuan, Qing Ling University of Science and Technology of China This paper studies the hierarchy of local minimums of a Department of Automation polynomial in the space. For this purpose, we first compute [email protected], [email protected] H-minimums, for which the first and second order optimal- ity conditions are satisfied. To compute each H-minimum, Wotao Yin we construct a sequence of semidefinite relaxations, based Department of Mathematics on optimality conditions. We prove that each constructed University of California at Los Angeles sequence has finite convergence, under some generic con- OP14 Abstracts 131

ditions. A procedure for computing all local minimums is a common feature of empirical networks–can be explained given. When there are equality constraints, we have sim- by the Degree-Corrected Stochastic Blockmodel and the ilar results for computing the hierarchy of critical values Extended Planted Partition model, two statistical models and the hierarchy of local minimums. Several extensions that allow for highly heterogeneous degrees. Throughout, are discussed. the paper characterizes and justifies several of the varia- tions of the spectral clustering algorithm in terms of these Jiawang Nie models. University of California, San Diego [email protected] Tai Qin University of Wisconsin, Madison Department of Statistics MS52 [email protected] Alternating Direction Methods for Penalized Clas- sification Karl Rohe University of Wisconsin, Madison Department of Fisher Linear Discriminant Analysis is a classical technique Statistics for feature reduction in supervised classification. How- [email protected] ever, this technique relies on the fact that the data being processed contains fewer features than observations, which typically fails to hold in a high-dimensional setting. In this MS52 talk, we present a modification, based on 1-regularization Statistical and Computational Tradeoffs in Hierar- and the alternating direction method of multipliers, for per- chical Clustering forming Fisher Linear Discriminant Analysis in the high- dimensional setting. Recursive spectral clustering is robust to noise in similar- ities between the objects to be clustered, however it has Brendan Ames cubic computational complexity. I will present an active California Institute of Technology spectral clustering algorithm that uses selective similari- Department of Computing + Mathematical Sciences ties, allowing one to trade off noise tolerance for compu- [email protected] tational efficiency. For one setting, it reduces to spectral clustering and for another it yields nearly linear compu- tational complexity while being more robust than greedy MS52 hierarchical clustering methods such as single linkage. Statistical-Computational Tradeoffs in Planted Models Aarti Singh Machine Learning Department We study the statistical-computational tradeoffs in the Carnegie Mellon University planted clustering model, where edges are randomly placed [email protected] between nodes according to their cluster memberships and the goal is to recover the clusters. When the number of clusters grows, we show that the space of model parame- MS53 ters can be partitioned into four regions: impossible, where Alternative Quadratic Models for Minimization all algorithms fail; hard, where an exponential-time algo- Without rithm succeeds; easy, where a polynomial-time algorithm Derivatives succeeds; simple, where a simple counting algorithm suc- ceeds. Let quadratic models be updated by an algorithm for derivative-free optimization, using a number of values of Yudong Chen the objective function that is fewer than the amount of University of California, Berkeley freedom in each model. The freedom may be taken up Department of Electrical Engineering and Computer by minimizing the Frobenius norm of either the change Sciences to or the new second derivative matrix of the model. An [email protected] automatic way of choosing between these alternatives is described. Its advantages are illustrated by two numerical Jiaming Xu examples. University of Illinois at Urbana-Champaign [email protected] Michael Powell University of Cambridge Department of Applied Maths and Theoretical Physics MS52 [email protected] Regularized Spectral Clustering under the Degree- Corrected Stochastic Blockmodel MS53 Spectral clustering is a fast and popular algorithm for find- The Spectral Cauchy-Akaike Method ing clusters in networks. Recently, Chaudhuri et al. and Amini et al. proposed inspired variations on the algorithm We introduce a variation of Cauchy method to minimize a that artificially inflate the node degrees for improved sta- convex quadratic function. It explores the fact, proved by tistical performance. The current paper extends the pre- Akaike, that the steepest descent method will be trapped vious statistical estimation results to the more canonical in a two dimensional subspace associated to extreme eigen- spectral clustering algorithm in a way that removes any values. The algorithm then estimates the largest eigenvalue assumption on the minimum degree and provides guid- and performs a spectral step to scape this subspace. We ance on the choice of the tuning parameter. Moreover, will compare the new algorithm to the Barzilai-Borwein our results show how the “star shape” in the eigenvectors– and Cauchy-Barzilai-Bowein methods and discuss its ex- 132 OP14 Abstracts

tensions to box constrained problems. alization of the famous Doignon-Bell-Scarf theorem: Given an integer k, we prove that there exists a constant c(k, n), PauloJ.S.Silva depending only on the dimension n and k, such that if a University of Sao Paulo polyhedron {x : Ax ≤ b} contains exactly k integer solu- [email protected] tions, then there exists a subset of the rows of cardinality no more than c(k, n), defining a polyhedron that contains exactly the same k integer solutions. MS53 On the Convergence and Worst-Case Complexity Jess A. De Loera of Trust-Region and Regularization Methods for Department of Mathematics Unconstrained Optimization Univ. of California, Davis [email protected] A nonlinear stepsize control framework for unconstrained optimization was recently proposed by Toint (2013), pro- Iskander Aliev viding a unified setting in which the global convergence Cardiff University can be proved for trust-region algorithms and regulariza- iskander aliev ¡alievi@cardiff.ac.uk¿ tion schemes. The original analysis assumes that the Hes- sians of the models are uniformly bounded. In this paper, the global convergence of the nonlinear stepsize control al- Quentin Louveaux gorithm is proved under the assumption that the norm of Universite Liege, Belgium the Hessians can grow by a constant amount at each iter- [email protected] ation. The worst-case complexity is also investigated. The results obtained for unconstrained smooth optimization are extended to some algorithms for composite nonsmooth op- MS54 timization and unconstrained multiobjective optimization Cutting Planes for a Multi-Stage Stochastic Unit as well. Commitment Problem

Geovani Grapiglia Motivated by the intermittency of renewable energy, we Federal University of Parana, Brazil propose a multi-stage stochastic integer programming Brazil model in this talk to address unit commitment problems geovani [email protected] under uncertainty, for which we construct several classes of strong inequalities by lifting procedures to strengthen Jinyun Yuan the original formulation. Our preliminary computational Federal University of Parana experiments show encouraging results. Brazil [email protected] Riuwei Jiang University of Arizona Ya-Xiang Yuan [email protected] Chinese Academy of Sciences Inst of Comp Math & Sci/Eng, AMSS Yongpei Guan [email protected] University of Florida [email protected]fl.edu

MS53 Jean-Paul Watson A Nonmonotone Approximate Sequence Algorithm Sandia National Laboratories for Unconstrained Optimization Discrete Math and Complex Systems [email protected] A new nonmonotone algorithm will be presented for uncon- strained nonlinear optimization. Under proper searching direction assumptions, this algorithm has global conver- MS54 gence for minimizing general nonlinear objective function with Lipschitz continuous derivatives. For convex objec- A Tighter Formulation of a Mixed-Integer Program tive function, this algorithm would maintain the optimal with Powerflow Equations convergence rate of convex optimization. Some preliminary numerical results will be also presented. We consider a problem arising in power systems where a distribution network operator needs to decide on the acti- Hongchao Zhang vation of a set of load modulations in order to avoid con- Department of Mathematics and CCT gestion in its network. The problem can be modeled as Louisiana State University a mixed-integer nonlinear program where the nonconvex [email protected] nonlinearities come from the powerflow equations. We first show that the Lagrangian and the SDP relaxation provide loose lower bounds for such a problem. We propose a for- MS54 mulation based on a linear network flow strengthened by Integer Programs with Exactly K Solutions the presence of loss variables. We show with computational experiments that we can provide tighter dual bounds with In this work we study a generalization of the classical fea- our formulation if we add some lower and upper bounds on sibility problem in integer linear programming, where an the loss variables. ILP needs to have a prescribed number of solutions to be considered solved. (e.g., exactly k-solutions). We devel- Quentin Louveaux oped the structure theory of such problems and prove some Universite Liege, Belgium generalization of known results: E.g., we provide a gener- [email protected] OP14 Abstracts 133

Bertrand Cornelusse, Quentin Gemine linear convergence rate on functions that satisfy an essen- Universite de Liege tial strong convexity property and a sublinear rate (1/K) [email protected], [email protected] on general convex functions. Near-linear speedup on a mul- ticore system can be expected if the number of processors is O(n1/2) in unconstrained optimization and O(n1/4)inthe MS54 separable-constrained case, where n is the number of vari- Pushing the Frontier (literally) with the Alpha ables. We describe results from implementation on 40-core Alignment Factor processors.

Construction of optimized portfolios entails complex inter- Ji Liu, Stephen J. Wright action between three key entities, namely, the risk factors, University of Wisconsin-Madison the alpha factors and the constraints. The problems that [email protected], [email protected] arise due to mutual misalignment between these three en- tities are collectively referred to as Factor Alignment Prob- Christopher R´e lems (FAP). Examples of FAP include risk-underestimation Stanford University of optimized portfolios, undesirable exposures to factors [email protected] with hidden and unaccounted systematic risk, consistent failure in achieving ex-ante performance targets, and in- Victor Bittorf ability to harvest high quality alphas into above-average University of Wisconsin-Madison IR. In this paper, we give a detailed analysis of FAP and [email protected] discuss solution approaches based on augmenting the user risk model with a single additional factor y. For the case of unconstrained MVO problems, we develop a parametric MS55 non-linear optimization model to analyze the ex-post util- Hydra: Distributed Coordinate Descent for Big ity function of the corresponding optimal portfolios, derive Data Optimization a closed form expression of the optimal factor volatility value and compare the solutions for various choices of y Hydra: HYbriD cooRdinAte descent method for solving culminating with a closed form expression for the optimal loss minimization problems with big data. We initially par- choice of y. Ultimately, we show how the Alpha Align- tition the coordinates and assign each partition to a differ- ment Factor (AAF) approach emerges as a natural and ent node of a cluster. At every iteration, each node picks effective remedy to FAP. AAF not only corrects for risk a random subset of the coordinates from those it owns, underestimation bias of optimal portfolios but also pushes independently from the other computers, and in parallel the ex-post efficient frontier upwards thereby empowering computes and applies updates to the selected coordinates a PM to access portfolios that lie above the traditional based on a simple closed-form formula. We give bounds on risk-return frontier. We provide extensive computational the number of iterations sufficient to approximately solve results to corroborate our theoretical findings. the problem with high probability, and show how it de- pends on the data and on the partitioning. We perform Anureet Saxena numerical experiments with a LASSO instance described Allianz Global Investors by a 3TB matrix. [email protected] Martin Takac, Peter Richtarik University of Edinburgh MS55 [email protected], [email protected] Parallel Coordinate Descent for Sparse Regression

This talk discusses parallelization of coordinate descent- MS55 based methods, targeted at sparse regression. Sev- The Rich Landscape of Parallel Coordinate De- eral works have demonstrated theoretical and empirical scent Algorithms speedups from parallelizing coordinate descent, in both the multicore and distributed settings. We examine these Large-scale 1-regularized loss minimization problems arise works, highlighting properties of sparse regression prob- in high-dimensional applications such as compressed sens- lems which permit effective parallel coordinate descent. We ing and high-dimensional supervised learning, including then generalize this discussion by asking: How well can a classification and regression problems. High-performance local solution (single or block coordinate descent updates) algorithms and implementations are critical to efficiently approximate a global solution? Exploring this question solving these problems. Building upon previous work on leads to new parallel block coordinate descent-based algo- coordinate descent algorithms for 1-regularized problems, rithms which are well-suited for the distributed setting and we introduce a novel family of algorithms that includes, use limited communication. as special cases, several existing algorithms such as SCD, Greedy CD, Shotgun, and Thread-Greedy. We give a uni- Joseph Bradley fied convergence analysis for the family of block-greedy al- UC Berkeley gorithms. We hope that algorithmic approaches and con- [email protected] vergence analysis we provide will not only advance the field, but will also encourage researchers to systematically ex- plore the design space of parallel coordinate descent algo- MS55 rithms for solving large-scale 1-regularization problems. An Asynchronous Parallel Stochastic Coordinate Descent Algorithm Chad Scherrer Independent Consultant We describe an asynchronous parallel stochastic coordinate [email protected] descent algorithm for minimizing smooth unconstrained or separably constrained functions. The method achieves a Ambuj Tewari 134 OP14 Abstracts

University of Michigan [email protected] Fr´ed´eric Meunier Mahantesh Halappanavar Ecole´ des Ponts ParisTech Pacific Northwest National Laboratory [email protected] [email protected]

David Haglin MS56 PNNL [email protected] An LP-Newton method for a standard form Linear Programming Problem

MS56 Fujishige et al. propose the LP-Newton method, a new algorithm for solving linear programming problems (LPs). Augmentation Algorithms for Linear and Integer They address LPs which have a lower and an upper bound Linear Programming for each variable. They reformulate the problem by in- troducing a related zonotope. Their algorithm solves the Separable convex IPs can be solved via polynomially many problem by repeating projections to the zonotope. In this augmentation steps if best augmenting steps along Graver paper, we develop the LP-Newton method for solving stan- basis directions are performed. Using instead augmen- dard form LPs. We recast the LP by introducing a related tation along directions with best ratio of cost improve- convex cone. Our algorithm solves the problem by iterating ment/unit length, we show that for linear objectives the projections to the convex cone. number of augmentation steps is bounded by the number of elements in the Graver basis of the problem matrix, giv- Shinji Mizuno, Tomonari Kitahara ing strongly polynomial-time algorithms for the solution of Tokyo Institute of Technology N-fold LPs and ILPs. [email protected], [email protected]

Jess A. De Loera Jianming Shi Department of Mathematics Tokyo University of Science Univ. of California, Davis [email protected] [email protected]

Raymond Hemmecke MS56 Dept. of Mathematics TU Munich, Germany Performance of the Simplex Method on Markov [email protected] Decision Processes

Jon Lee Markov decision processes (MDPs) are a particularly in- IBM T.J. Watson Research Center teresting class of linear programs that lie just beyond the [email protected] point where our ability to solve LPs in strongly polynomial time stops, and number of recent results have studied the performance of the simplex method on MDPs and proven MS56 both upper and lower bounds, most famously Friedmann, Hansen, and Zwick’s sub-exponential lower bound for ran- Colorful Linear Programming: Complexity and Al- domized pivoting rules [STOC 2011]. This talk will discuss gorithms the known upper bounds for solving MDPs using the sim- plex method. d Let S1,...,Sk be k sets of points in Q .Thecolorful linear programming problem, defined by B´ar´any and Onn (Math- Ian Post ematics of Operations Research, 22 (1997), 550–567) aims University of Waterloo k at deciding whether there exists a T ⊆ i=1 Si such that [email protected] |T ∩ Si|≤1fori =1,...,k and 0 ∈ conv(T ). They proved in their paper that this problem is NP-complete when k = d. They leave as an open question the com- MS57 plexity status of the problem when k = d + 1. Contrary to the case k = d, this latter case still makes sense when Metric Regularity of Stochastic Dominance Con- the points are in a generic position. We solve the question straints Induced by Linear Recourse by proving its NP-completeness while exhibiting some rela- tionships between colorful linear programming and comple- The introduction of stochastic dominance constraints is one mentarity programming. In particular, we prove that any option of handling risk aversion in linear programming un- polynomial time algorithm solving the case with k = d +1 der (stochastic) uncertainty. This approach leads to an op- and |Si|≤2fori =1,...,d+ 1 allows to compute easily timization problem with infinitely many probabilistic con- Nash equilibrium in bimatrix games in polynomial time. straints. Metric regularity is of crucial importance, e.g. On our track, we found a new way to prove that a com- for optimality conditions and stability of optimal solutions plementarity problem belongs to the PPAD class with the when the underlying probability measure is subjected to help of Sperner’s lemma. We also show that the algorithm perturbations. The talk addresses sufficient conditions for proposed by B´ar´any and Onn for computing a feasible so- metric regularity including their verification. lution T in a special case can be interpreted as a “Phase I” simplex method, without any projection or distance com- Matthias Claus putation. This is joint work with Pauline Sarrabezolles. University of Duisburg-Essen, Germany OP14 Abstracts 135

[email protected] Ruediger Schultz University of Duisburg-Essen [email protected] MS57 Two-Stage meets Two-Scale in Stochastic Shape Optimization MS58 A Primal-Dual Interior Point Solver for Conic In structural elastic shape optimization one faces the prob- Problems with the Exponential Cone lem of distributing a given solid material over a working domain exposed to volume and surface loads. The result- Primal-dual interior point methods for linear programming ing structure should be as rigid as possible while meeting a have been successfully extended to the realm of conic pro- constraint on the volume spent. Without any further reg- gramming with self-dual cones. However, other useful and ularization the problem is ill-posed and on a minimizing interesting conic constraints are ?not self-dual and the ex- sequence the onset of microstructures can be observed. We tensions cannot be used. The exponential cone is one such are interested in approximating these optimal microstruc- non-self-dual cone that can be used to define geometric tures by simple, constructible geometries within a two-scale programming problems in the language of conic program- model based on boundary elements for the elastic problem ing. We describe the development and implementation of on the microscale and finite elements on the macroscale. a predictor-corrector algorithm suitable for linear, second- Geometric details described by a small set of parameters order cone, and exponential cone problems. thereby enter the macroscopic scale via effective material properties and lead to a finite dimensional constraint opti- Santiago Akle mization problem. The performance of resulting shapes in Institute for Computational and Mathematical the presence of uncertain loading will be compared to the Engineering one of classical single scale designs by applying stochastic Stanford University dominance constraints. [email protected]

Benedict Geihe Michael A. Saunders Rheinische Friedrich-Wilhelms-Universit¨at Bonn Systems Optimization Laboratory (SOL) Institut f¨ur Numerische Simulation Dept of Management Sci and Eng, Stanford [email protected] [email protected]

MS57 MS58 SOCP Approach for Joint Probabilistic Con- The Merits of Keeping It Smooth straints For optimization problems with general nonlinear con- In this talk, we investigate the problem of linear joint prob- straints, exact merit functions, which are used to en- abilistic constraints. We assume that the rows of the con- sure convergence from arbitrary starting points, typically straint matrix are dependent. Furthermore, we assume the come in two varieties: primal-dual merit functions that distribution of the constraint rows to be elliptically dis- are smooth, and primal-only merit functions that are non- tributed. Under some conditions, we prove the convexity smooth. Primal-only smooth merit functions for equality of the investigated set of feasible solutions. We also de- constraints have been proposed, though they are generally velop an approximation scheme for this class of stochastic considered too expensive to be practical. We describe the programming problems based on SOCP. Numerical results properties of a smooth exact merit function first proposed are presented. by Fletcher (1970), extensions that can handle inequality constraints, and our efforts to use it efficiently. Abdel Lisser, Michal Houda University of Paris Sud, LRI MichaelP.Friedlander [email protected], [email protected] Dept of Computer Science University of British Columbia [email protected] Jianqiang Cheng LRI, University of Paris-Sud, France [email protected] Dominique Orban Ecole Polytechnique de Montreal [email protected] MS57 Unit Commitment Under Uncertainty in AC MS58 Transmission Systems - a Risk Averse Two-Stage Block Preconditioners for Saddle-Point Linear Sys- Stochastic SDP Approach tems The talk addresses unit commitment under uncertainty Symmetric indefinite 2x2 block matrices of the form of load and power infeed from renewables in alternating-   current (AC) power systems. Beside traditional unit- ABT commitment constraints, the physics of power flow are B 0 included. This leads to risk averse two-stage stochastic semidefinite programs whose structure is analyzed, and for feature prominently in the solution of constrained opti- which a decomposition algorithm is presented. mization problems, and in the solution of partial differ- ential equations with constraints. When the systems are Tobias Wollenberg large and sparse, modern iterative methods are often very University of Duisburg-Essen, Germany effective. These methods require the use of precondition- [email protected] ers, and for block-structured matrices of the above form it 136 OP14 Abstracts

is often desirable to design block diagonal preconditioners, tronic Structure Calculations based primarily on the primal or the dual Schur comple- ment. In this talk we present a few block precondition- We present gradient type methods for electronic structure ing approaches and discuss the spectral properties of the calculations by constructing new iterations along the gra- associated preconditioned matrices. We are particularly dient on the Stiefel manifold. The main costs of our ap- interested in situations where the leading block (typically proaches arise from the assembling of the total energy func- representing the Hessian) is singular and has a high nul- tional and its gradient and the projection onto the mani- lity. In such situations one often has to resort to Schur fold. These tasks are cheaper than eigenvalue computation complements of a regularized version of the block matrix, and they are often more suitable for parallelization as long and the resulting preconditioners have interesting spectral as the evaluation of the total energy functional and its gra- properties. In particular, the high nullity seems to promote dient is efficient. clustering of the eigenvalues, which in turn may accelerate the convergence of a typical minimum residual Krylov sub- Zaiwen Wen space solver. The analytical results are accompanied by Peking University numerical illustrations. Beijing International Center For Mathematical Research [email protected] Chen Greif Department of Computer Science The University of British Columbia MS59 [email protected] A Projected Gradient Algorithm for Large-scale Eigenvalue Calculation

MS58 We examine a projected gradient algorithm for computing Numerical Methods for Symmetric Quasi-Definite a relatively large number of lowest eigenvalues of a Hermi- Linear Systems tian matrix. The algorithm performs fewer Rayleigh-Ritz calculations than some of the existing algorithms, thus has Abstract not available. better parallel scalability. It is relatively easy to imple- ment. We will discuss a number of practical issues for Dominique Orban implementing this algorithm, and demonstrate its perfor- GERAD and Dept. Mathematics and Industrial mance in an electronic structure application. Engineering Ecole Polytechnique de Montreal Chao Yang [email protected] Lawrence Berkeley National Lab chao yang [email protected]

MS59 MS59 On the Convergence of the Self-Consistent Field Symmetric Low-Rank Product Optimization: Iteration in Kohn-Sham Density Functional Theory Gauss-Newton Method We view the SCF iteration as a fixed point iteration with Emerging applications demand higher capacities in large- respect to the potential and formulate the simple mixing scale eigenspace computation, among which are higher par- schemes accordingly. The convergence of the simple mixing allel scalability and better warm-start ability that many schemes are established. A class of approximate Newton existing eigensolvers often lack. We propose to study a approaches and their convergence properties are presented nonlinear least squares formulation for symmetric matrix and discussed respectively. Preliminary numerical exper- eigenspace calculation. We derive a class of Guass-Newton imentations show the efficiency of our proposed approxi- type algorithms from this formulation, and present conver- mate Newton approaches. gence and numerical results.

Xin Liu Yin Zhang Academy of Mathematics and Systems Science Rice University Chinese Academy of Sciences Dept of Comp and Applied Math [email protected] [email protected]

Zaiwen Wen Xin Liu Peking University Academy of Mathematics and Systems Science Beijing International Center For Mathematical Research Chinese Academy of Sciences [email protected] [email protected]

Xiao Wang Zaiwen Wen University of Chinese Academy of Sciences Peking University [email protected] Beijing International Center For Mathematical Research [email protected] Michael Ulbrich Technische Universitaet Muenchen Chair of Mathematical Optimization MS60 [email protected] Accounting for Uncertainty in Complex System Design Optimization

MS59 Uncertainty manifests itself in complex systems in a variety Gradient Type Optimization Methods for Elec- of ways. In airspace traffic, a major source of uncertainty OP14 Abstracts 137

is the potential inability to compute a safe trajectory in of a function f(x). However often one has several informa- real time, i.e., to solve a large nonlinear optimization prob- tion sources concerning the relation between x and f(x) that lem. One way to mitigate the uncertainty is to evaluate the can be sampled, each with a different sampling cost. Multi- airspace and anticipate approach to phase transitions from Information Source Optimization is the problem of opti- controllable (where trajectory computation is feasible) to mally choosing which information source to sample during uncontrollable (where trajectories cannot be computed). the overall optimization of f, by trading off information the We discuss such an approach. samples of the different sources could provide with the cost of generating those samples. Natalia Alexandrov NASA Langley Research Center David Wolpert [email protected] Santa Fe Institute Los Alamos National Laboratory [email protected] MS60 Beyond Variance-based Robust Optimization MS61 In optimization under uncertainty of a wing, maximization A Relaxed Projection Splitting Algorithm for Non- of aerodynamic performance (in expectation) is comple- smooth Variational Inequalities in Hilbert Spaces mented by minimization of the variance induced by man- ufacturing tolerances. This typically leads to stiff multi- We introduce a relaxed-projection splitting algorithm for objective optimization; moreover, the variance is a sym- solving variational inequalities in Hilbert spaces for the metric measure and therefore penalizes both better- and sum of nonsmooth maximal monotone operators, where worse-than-average designs. We propose to use the over- the feasible set is defined by a nonlinear and nonsmooth all probability density of the wing performance and define continuous convex function inequality. In our scheme, the the optimization goal as the minimization of the ”distance” orthogonal projections onto the feasible set are replaced between this and a target distribution. by projections onto separating hyperplanes. Furthermore, each iteration of the proposed method consists of simple Gianluca Iaccarino subgradient-like steps, which does not demand the solution Stanford University of a nontrivial subproblem, using only individual operators, Mechanical Engineering which explores the structure of the problem. Assuming [email protected] monotonicity of the individual operators and the existence of solutions, we prove that the generated sequence con- Pranay Seshadri verges weakly to a solution. RollsRoyce [email protected] Jose Yunier Bello Cruz University of British Columbia [email protected] Paul Constantine Colorado School of Mines Applied Mathematics and Statistics MS61 [email protected] Forward-Backward and Tseng’s Type Penalty Schemes for Monotone Inclusion Problems

MS60 We propose a forward-backward penalty algorithm for Quasi-Newton Methods for Stochastic Optimiza- monotone inclusion problems of the form 0 ∈ Ax + Dx + tion NC (x) in real Hilbert spaces, where A is a maximally monotone operator, D a cocoercive operator and C the We describe a class of algorithms for stochastic optimiza- nonempty set of zeros of another cocoercive operator. tion, i.e., for problems in which repeated attempts to eval- Weak ergodic convergence is shown under the use of the uate the objective function generate a random sample from Fitzpatrick function associated to the operator that de- a probability distribution. These algorithms exploit a con- scribes the set C. A forward-backard-forward penalty al- nection between ridge analysis in response surface method- gorithm is also proposed by relaxing the cocoercivity as- ology and trust-region methods for numerical optimization. sumptions to monotonicity and Lipschtz continuity. First-order information is obtained by designed regression experiments; second-order information is obtained by se- Radu Ioan Bot cant updates. A convergence analysis is borrowed from University of Vienna stochastic approximation. [email protected] Michael W. Trosset Department of Statistics, Indiana University MS61 [email protected] The Douglas-Rachford Algorithm for Nonconvex Feasibility Problems Brent Castle Indiana University Projection algorithms for solving (nonconvex) feasibility [email protected] problems in Euclidean spaces are considered. Of special interest is the Douglas-Rachford or Averaged Alternating Reflection Algorithm (AAR). In the case of convex feasi- MS60 bility, firm nonexpansiveness of projection mappings is a Multi-Information Source Optimization: Beyond global property that yields global convergence. A relaxed Mfo local version of firm nonexpansiveness is introduced. To- gether with a coercivity condition that relates to the reg- Conventional optimization concerns finding the extremizer ularity of the intersection, this yields local linear conver- 138 OP14 Abstracts

gence for a wide class of nonconvex problems for consistent part. Examples show that this new approach yields much feasibility problems. better solutions than using the existing approach.

Robert Hesse Frans de Ruiter University of G¨ottingen Tilburg University [email protected] [email protected]

MS61 Aharon Ben-Tal Technion - Israel Institue of Technology Linear Convergence of the Douglas-Rachford Israel Method for Two Nonconvex Sets [email protected] In this talk, we study the Douglas-Rachford method for two closed (possibly nonconvex) sets in Euclidean spaces. Ruud Brekelmans In particular, we show that under suitable assumptions, Dept. of Econometrics & Operations Research, CentER, the method converges locally with linear rate. In convex Tilburg University settings, we prove that the linear convergence is global. [email protected] Our study recovers recent results on the same topic. Dick Den Hertog Hung M. Phan Tilburg University University of British Columbia Faculty of Economics and Business Administration [email protected] [email protected]

MS62 MS62 Globalized Robust Optimization for Nonlinear Un- certain Inequalities Robust Growth-Optimal Portfolios

Globalized Robust Optimization (GRO) is an extension of The log-optimal portfolio is known to outperform any other Robust Optimization (RO) where controlled constraint vio- portfolio in the long run if stock returns are i.i.d. and fol- lations are allowed outside the primary uncertainty region. low a known distribution. In this talk, we establish similar The traditional GRO approach only deals with linear in- guarantees for finite investment horizons where the dis- equalities and a secondary uncertainty region that is re- tribution of stock returns is ambiguous. By focusing on stricted to a special form. We extend GRO to nonlinear fixed-mix portfolios, we exploit temporal symmetries to inequalities and secondary regions that are general convex formulate the emerging distributionally robust optimiza- regions. Furthermore, we propose a method to optimize tion problems as tractable conic programs whose sizes are the controlled violation outside the primary region based independent of the investment horizon. on intuitive criteria. Napat Rujeerapaiboon Ruud Brekelmans EPFL Dept. of Econometrics & Operations Research, CentER, napat.rujeerapaiboon@epfl.ch Tilburg University [email protected] Daniel Kuhn Imperial College London Aharon Ben-Tal daniel.kuhn@epfl.ch Technion - Israel Institue of Technology Israel Wolfram Wiesemann [email protected] Department of Computing Imperial College London Dick Den Hertog [email protected] Tilburg University Faculty of Economics and Business Administration [email protected] MS62 Routing Optimization with Deadlines under Un- Jean-Philippe Vial certainty Ordecsys [email protected] We consider a class of routing optimization problems on networks with deadlines imposed at a subset of nodes, and MS62 with uncertain arc travel times. The problems are static in the sense that routing decisions are made prior to the Adjustable Robust Optimization with Decision realization of uncertain travel times. The goal is to find Rules Based on Inexact Information optimal routing solution such that arrival times at nodes Adjustable robust optimization, a powerful technique to respect deadlines “as much as possible’. solve multistage optimization problems, assumes that the revealed information on uncertain parameters is exact. Patrick Jaillet However, in practice such information often contain inac- MIT curacies. We therefore extend the adjustable robust opti- [email protected] mization approach to the more practical situation that the revealed information is inexact. We construct a new model, Jin Qi, Melvyn Sim and develop a computationally tractable robust counter- NUS Business School OP14 Abstracts 139

[email protected], [email protected] with Asymmetric Soft Regularization

In this talk, we introduce ∞-norm based new low rank MS63 promoting regularization framework, that is, soft asymmet- ric regularization (SAR) framework for robust nonnegative Computing Generalized Tensor Eigenpairs matrix factorization (NMF). The main advantage of the proposed low rank enforcing SAR framework is that it is Several tensor eigenpair definitions have been put forth in less sensitive to the rank selecting regularization param- the past decade which can be unified under the generalized eters since we use soft regularization framework, instead tensor eigenpair framework introduced by Chang, Pear- of using the conventional hard constraints such as nuclear son, and Zhang (2009). Given an mth-order, n-dimensional norm, γ2-norm, or rank itself in matrix factorization. The real-valued symmetric tensors A and B, the goal is to find numerical results show that, although we fixed all parame- a real-valued scalar λ and a real-valued n-vector x with m−1 m−1 ters of the proposed SAR framework for robust NMF, the x =0suchthat Ax = λBx .Tosolvetheproblem, proposed method recover low-rank structure better than we present our generalized eigenproblem adaptive power that of the state-of-the-art nuclear norm based robust prin- method (GEAP) method, which uses interesting proper- cipal component analysis (PCA) and other robust NMF ties of convex optimization on a sphere, even though our models. Moreover, the basis generated by the proposed objective function is not convex. method is more interpretable than that of the robust PCA. Tamara G. Kolda, Jackson Mayo Sandia National Laboratories Hyenkyun Woo, Haesun Park [email protected], [email protected] Georgia Institute of Technology [email protected], [email protected] MS63 Square Deal: Lower Bounds and Improved Relax- MS63 ations for Tensor Recovery Inverse Scale Space: New Regularization Path for Sparse Regression Recovering a low-rank tensor from incomplete information is a recurring problem in signal processing and machine The mostly used regularization path for sparse regression is learning. The most popular convex relaxation of this prob- minimizing an 1-regularization term. However it has many lem minimizes the sum of the nuclear norms of the unfold- disadvantages such as bias and introducing more features. ings of the tensor. We show that this approach can be sub- We introduce inverse scale space (ISS)-a new regulariza- stantially suboptimal: reliably recovering a K-way tensor tion path for sparse regression, which is unbiased and can of length n and Tucker rank r from Gaussian measurements remove unrelated features quickly. We show why ISS is K−1 requires Ω(rn ) observations. In contrast, a certain (in- better than 1 minimization, and the comparison of both tractable) nonconvex formulation needs only O(rK + nrK) methods is done on synthetic and real data. In addition, observations. We introduce a simple, new convex relax- we develop a fast algorithm for ISS. ation, which partially bridges this gap. Our new formu- lation succeeds with O(rK/2nK/2) observations. While Ming Yan these results pertain to Gaussian measurements, numerical University of California, Los Angeles experiments on both synthetic and real data sets strongly Department of Mathematics suggest that the new convex regularizer also outperforms [email protected] the sum of nuclear norms for tensor completion from a random subset of entries. The lower bound for the sum-of- Wotao Yin nuclear-norms model follows from our new result on recov- Department of Mathematics ering signals with multiple sparse structures (e.g. sparse, University of California at Los Angeles low rank), which demonstrates the significant suboptimal- [email protected] ity of the commonly used recovery approach via minimiz- ing the sum of individual sparsity inducing norms (e.g. l1, Stanley J. Osher nuclear norm). Our new formulation for low-rank tensor University of California recovery however opens the possibility in reducing the sam- Department of Mathematics ple complexity by exploiting several structures jointly. [email protected]

Cun Mu, Bo Huang Columbia University MS64 [email protected], [email protected] A Nonlinear Semidefinite Approximation for Copositive Optimisation John Wright Department of Electrical Engineering Copositive optimisation has numerous applications, e.g. an Columbia University exact formulation of the stability number. Due to the [email protected] difficulty of copositive optimisation, we often consider ap- proximations, e.g. the Parrilo-cones. The Parrilo-cones are not invariant under scaling, but in this talk a method is Donald Goldfarb presented to optimise over all scalings at once. This is Columbia University through nonlinear semidefinite optimisation, which is in [email protected] general intractable. However for some problems, including the stability number, quasiconvexity in our method makes it tractable. MS63 Robust Nonnegative Matrix/Tensor Factorization Peter Dickinson 140 OP14 Abstracts

University of Vienna [email protected] Austria [email protected] MichaelL.Overton New York University Juan C. Vera Courant Instit. of Math. Sci. Tilburg School of Economics and Management [email protected] Tilburg University [email protected] MS64 Real PSD Ternary Forms with Many Zeros MS64 Optimal Solutions to a Root Minimization Problem We are concerned with real psd ternary forms p(x, y, z)of over a Polynomial Family with Affine Constraints degree 2d. Choi, Lam and Reznick proved in 1980 that a psd ternary sextic with more than 10 zeros has infinitely Consider the system y = A(x)y,whereA is a depending many zeros and is sos. We will discuss the situation for on a parameter x ∈ Ω ⊂ Ck. ThissystemisHurwitz- ternary octics, in which 10 must be replaced by at least 17. stable if the eigenvalues of A(x) lie in the left half of the This is, in part, joint work with Greg Blekherman. complex plane and Schur-stable if the eigenvalues of A(x) lie in the unit disk. A related topic is to consider poly- Bruce Reznick nomials whose coefficients lie in a parameter set. In 2012, University of Illinois at Urbana-Champaign Blondel, G¨urb¨uzbalaban, Megretski and Overton investi- [email protected] gate the Schur and Hurwitz stability of monic polynomi- als whose coefficients lie in an affine hyperplane of dimen- MS65 sion n − 1inR and C, respectively. They provide explicit global solutions to the radius minimization problem and Title Not Available closely related results for the abscissa minimization prob- Abstract Not Available lem for a family of polynomials with one affine constraint. In addition to their theoretical results, the authors pro- Renato C. Monteiro vide Matlab implementations of the algorithms they de- Georgia Institute of Technology rive. A major question that is left open is: suppose there School of ISyE ∈{ − } are ν 2,...,n 1 constraints on the coefficients, not [email protected] just one. Our current work is to extend results on the poly- nomial radius and abscissa minimization problems to this more general case. MS65 Constrained Best Euclidean Distance Embedding Julia Eaton on a Sphere: a Matrix Optimization Approach University of Washington Tacoma The problem of data representation on a sphere of unknown [email protected] radius arises from various disciplines. The best representa- tion often needs to minimize a distance function of the data Sara Grundel on a sphere as well as to satisfy some Euclidean distance MPI Magdeburg constraints. In this paper, we reformulate the problem as [email protected] an Euclidean distance matrix optimization problem with a low rank constraint. We then propose an iterative algo- Mert Gurbuzbalaban rithm that uses a quadratically convergent Newton method Massachusetts Institute of Technology at its each step. We use some classic examples from the [email protected] spherical multidimensional scaling to demonstrate the flex- ibility of the algorithm in incorporating various constraints. Michael L. Overton New York University Houduo Qi Courant Instit. of Math. Sci. University of Southampton [email protected] [email protected]

Shuanghua Bai MS64 School of Mathematics Narrowing the Difficulty Gap for the Celis-Dennis- University of Southampton Tapia Problem [email protected]

We study the Celis-Dennis-Tapia problem: minimize a non- Naihua Xiu convex quadratic function over the intersection of two ellip- Department of Applied Mathematics soids, a generalization of the well-known trust region prob- Beijing Jiaotong University lem. Our main objective is to narrow the difficulty gap that naihua [email protected] occurs when the Hessian of the Lagrangian is indefinite at Karush-Kuhn-Tucker points. We prove new sufficient and necessary conditions both for local and global optimality, MS65 based on copositivity, giving a complete characterization Inexact Proximal Point and Augmented La- in the degenerate case. grangian Methods for Solving Convex Matrix Op- timization Problems Immanuel Bomze University of Vienna Convex matrix optimization problems (MOPs) arise in a OP14 Abstracts 141

wide variety of applications. In this talk, we present the Defeng Sun framework of designing algorithms based on the inexact Dept of Mathematics proximal point and augmented Lagrangian methods for National University of Singapore solving MOPs. In order to solve the subproblem in each [email protected] iteration efficiently and robustly, we propose an inexact block coordinate descent method coupled with a semis- Kim-Chuan Toh mooth Newton-CG method for the purpose. A concrete National University of Singapore illustration of the methodology is given to solve the class of Department of Mathematics and Singapore-MIT Alliance linearly constrained semidefinite matrix least squares prob- [email protected] lems with nuclear norm regularization.

Kim-Chuan Toh MS66 National University of Singapore Department of Mathematics and Singapore-MIT Alliance An Acceleration Procedure for Optimal First-order [email protected] Methods We introduce a first-order method for smooth convex prob- Defeng Sun lems that allows an easy evaluation of the optimal value Dept of Mathematics of the local Lipschitz constant of the objective’s gradi- National University of Singapore ent, without even using any backtracking strategies. We [email protected] show that our method is optimal. Numerical experiments on very large-scale eigenvalue minimization problems show Kaifeng Jiang that our method reduces computation times by orders of Department of Mathematics magnitude over standard optimal first-order methods. National University of Singapore, Singapore [email protected] Michel Baes, Michael Buergisser ETH Zurich [email protected], [email protected] MS65 SDPNAL+: A Majorized Semismooth Newton-CG Augmented Lagrangian Method for Semidefinite MS66 Programming with Nonnegative Constraints Decompose Composite Problems for Big-data Op- We present a majorized semismooth Newton-CG aug- timization mented Lagrangian method, called SDPNAL+, for semidefinite programming (SDP) with partial or full non- Composite problems play an important role in analyzing negative constraints on the matrix variable. SDPNAL+ big datasets. This talk presents new optimization proce- is a much enhanced version of SDPNAL introduced by dures that can benifit from the decoupling of data fidelity Zhao, Sun and Toh [SIAM Journal on Optimization, 20 components from regularization terms constituting these (2010), pp. 1737–1765] for solving generic SDPs. SDP- data analysis models. NAL works very efficiently for nondegenerate SDPs and may encounter numerical difficulty for degenerate ones. Guanghui Lan Here we tackle this numerical difficulty by employing a Department of Industrial and Systems majorized semismooth Newton-CG method coupled with University of Florida a block coordinate descent method to solve the inner prob- [email protected]fl.edu lems. Numerical results for various large scale SDPs with or without nonnegative constraints show that the proposed method is not only fast but also robust in achieving accu- MS66 rate solutions. It outperforms, by a significant margin, An Adaptive Accelerated Proximal Gradient two other competitive codes: (1) an alternating direction Method and Its Homotopy Continuation for Sparse based solver called SDPAD by Wen et al. [Mathematical Optimization Programming Computation, 2 (2010), pp. 203–230] and (2) a two-easy-block-decomposition hybrid proximal extragra- We first propose an adaptive accelerated proximal gradi- dient method called 2EBD-HPE by Monteiro et al. [Math- ent (APG) method for minimizing strongly convex com- ematical Programming Computation, (2013), pp. 1–48]. In posite functions with unknown convexity parameters. This contrast to these two codes, we are able to solve all the 95 method incorporates a restarting scheme to automatically difficult SDP problems arising from QAP problems tested estimate the strong convexity parameter and achieves a in SDPNAL to an accuracy of 10−6 efficiently, while SD- nearly optimal iteration complexity. Then we consider the PAD and 2EBD-HPE successfully solve 30 and 16 prob- l1- regularized least-squares (l1-LS) problem in the high- lems, respectively. It is also noted that SDPNAL+ appears dimensional setting. Although such an objective function to be the only viable method currently available to solve is not strongly convex, it has restricted strong convexity large scale SDPs arising from rank-1 tensor approxima- over sparse vectors. We exploit this property by combining tion problems constructed by Nie and Li [arXiv preprint the adaptive APG method with a homotopy continuation arXiv:1308.6562, (2013)]. The largest rank-1 tensor ap- scheme, which generates a sparse solution path towards op- proximation problem solved is nonsym(21,4),inwhichits timality. This method obtains a global linear rate of con- resulting SDP problem has matrix dimension n =9, 261 vergence and its complexity is the best among known re- and the number of equality constraints m =12, 326, 390. sults for solving the l1-LS problem in the high-dimensional setting. Liuqin Yang National University of Singapore Qihang Lin Department of Mathematics University of Iowa [email protected] [email protected] 142 OP14 Abstracts

Lin Xiao into a larger scheme in which they are complemented by Microsoft Research measures of deviation, error and regret. The interactions [email protected] among these four different quantifications of loss, as cor- ner points of a quadrangle of relationships, provide rich modeling opportunities for risk management, not only in MS66 finance. Furthermore they reveal surprising connections An Extragradient-Based Alternating Direction between the ways an optimization problem is set up and Method for Convex Minimization how the data associated with it ought to be handled sta- tistically. In this talk, we consider the problem of minimizing the sum of two convex functions subject to linear linking con- Terry Rockafellar straints. The classical alternating direction type methods University of Washington usually assume that the two convex functions have rela- Department of Mathematics tively easy proximal mappings. However, many problems [email protected] arising from statistics, image processing and other fields have the structure that only one of the two functions has easy proximal mapping, and the other one is smoothly con- MS67 vex but does not have an easy proximal mapping. There- Superquantile Regression: Theory and Applica- fore, the classical alternating direction methods cannot be tions applied. For solving this kind of problems, we propose in this paper an alternating direction method based on ex- The presentation presents a generalized regression tech- tragradients. Under the assumption that the smooth func- nique centered on a superquantile (also called conditional tion has a Lipschitz continuous gradient, we prove that the value-at-risk) that is consistent with that coherent measure proposed method returns an eps-optimal solution within of risk and yields more conservatively fitted curves than O(1/eps) iterations. We test the performance of different classical least-squares and quantile regression. In contrast variants of the proposed method through solving the basis to other generalized regression techniques that approxi- pursuit problem arising from compressed sensing. We then mate conditional superquantiles by various combinations apply the proposed method to solve a new statistical model of conditional quantiles, we directly and in perfect ana- called fused logistic regression. Our numerical experiments log to classical regression obtain superquantile regression show that the proposed method performs very well when functions as optimal solutions of certain error minimiza- solving the test problems. tion problems. We discuss properties of this new regression technique, computational methods, and applications. Shiqian Ma The Chinese University of Hong Kong Johannes O. Royset [email protected] Operations Research Department Naval Postgraduate School Shuzhong Zhang [email protected] Department of Industrial & Systems Engineering University of Minnesota R T. Rockafellar [email protected] University of Washington Department of Mathematics

MS67 Support Vector Machines Based on Convex Risk MS67 Functionals and General Norms Cvar Norm in Deterministic and Stochastic Case: Applications in Statistics This paper studies unified formulations of support vector machines (SVMs) for binary classification on the basis of CVaR norm for a random variable is CVaR of absolute convex analysis. Using the notion of convex risk function- value of this random variable. L1 and L-infinity norms als, a pair of primal and dual formulations of the SVMs are are limiting cases of the CVaR norm. Several properties described in a general manner, and duality results and op- for CVaR norm, as a function of confidence level were timality conditions are established. The formulation uses proved. CVaR norm, as a Measure of Error, generates arbitrary norms for regularizers and incorporates new fami- a Regular Risk Quadrangle. We discuss several statistical lies of norms (in place of p-norms) into SVM formulations. applications of CVaR norm: 1) Robust linear regression disregarding some percentage of outliers; 2) Estimation of Jun-ya Gotoh probabilistic distributions. Chuo University [email protected] Stan Uryasev University of Florida Stan Uryasev uryasev@ufl.edu University of Florida uryasev@ufl.edu Alexander Mafuslav PhD student mafusalov@ufl.edu MS67 The Fundamental Quadrangle of Risk in Optimiza- tion and Statistics MS68 Stochastic Predictive Control for Energy Efficient Measures of risk, which assign a numerical risk value to Buildings random variable representing loss, are well recognized as a basic tool in financial management. But they also fit We present a stochastic model predictive control (SMPC) OP14 Abstracts 143

approach to building heating, ventilation, and air condi- [email protected] tioning (HVAC) systems. The building uncertain load is described by finitely supported probability density func- tions identified from historical data. The SMPC is de- MS68 signed with the goal of minimizing expected energy con- Development of Control-Oriented Models for sumption while bounding thermal comfort violations at Model Predictive Control in Buildings each time instance. The SMPC uses the predictive knowl- edge of building loads described by stochastic models. The Model Predictive Control (MPC) has gained attention in presentation focuses on the trade off between the compu- recent years for application to building automation and tational tractability and the conservatism of the resulting controls because of significant potential for energy cost. SMPC scheme. The large number of zones in a commercial MPC utilizes dynamic building and HVAC equipment building requires proper handling of system nonlinearities, models and input forecasts to estimate future energy usage chance constraints, and robust constraints in order to allow and employs optimization to determine control inputs that real-time computation. We will present an approach which minimize an integrated cost function or a specified predic- combines feedback linearization, tailored tightening offset tion horizon. A dynamic model with reasonable prediction for chance constraints, and tailored sequential quadratic performance (e.g., accuracy and simulation speed) is cru- programming. The proposed SMPC approach, existing cial for a practical implementation of MPC. One modeling SMPC approaches and current industrial control logics are approach is to use whole-building energy simulation pro- compared to demonstrate the link between controller com- grams such as EnergyPlus, TRNSYS and ESP-r etc. How- plexity, energy savings and comfort violations. ever, the computational and set up costs for these models are significant and they do not appear to be suitable for Francesco Borrelli,TonyKelman on-line implementation. This talk will present the devel- University of California, Berkeley opment of control-oriented models for the thermal zones [email protected], [email protected] in buildings. A simple linear ARX (Auto-Regressive with exogenous input) model and a low-order state-space model are identified from the designed input-output responses of thermal zones with disturbances from ambient conditions MS68 and internal heat gains. A high-fidelity TRNSYS model Pde-Ode Modeling, Inversion and Optimal Control of an office building was used as a virtual testbed to gen- for Building Energy Demand erate data for system identification, parameter estimation and validation of the two model structures. This talk will be concluded with comparisons of the ARX model and the Development of an accurate heat transfer model of build- state-space model in terms of model accuracy for a predic- ings is of high importance. Such a model can be used tive control design and the results of applying MPC. for analyzing energy efficiency of buildings, predicting en- ergy consumption and providing decision support for en- Zheng O’Neill ergy efficient operation of buildings. In this talk, we will Mechanical Engineering present the proposed PDE-ODE hybrid model to describe University of Alabama heat transfer through building envelope as well as heat [email protected] evolution inside building. A inversion procedure is pre- sented to recover parameters of equations from sensor data and building characteristic so that the model represents a MS69 specific building with current physical condition. We will Learning Near-Optimal Linear Embeddings present how this model is being used in MPC for building heating, ventilation, and air conditioning (HVAC) systems. We introduce a new framework for the deterministic con- struction of linear, near-isometric embeddings of a fi- Raya Horesh nite set of data points. Our formulation is based on an IBM T.J. Watson Research Center affine rank minimization problem that we relax into a [email protected] tractable semidefinite program (SDP). The resulting Nu- clear norm minimization with Max-norm constraints (Nu- Max) framework outperforms both principal components MS68 analysis (PCA) and random projections in a range of ap- plications in machine learning and signal processing, which Performance Optimization of Hvac Systems we demonstrate via a range of experiments on large-scale datasets. Performance of an HVAC system installed in a multi-story office building is studied. A model minimizing the energy Richard G. Baraniuk consumption and room temperature ramp rate of is pre- Rice University sented. The relationship between the input and output Electrical and Computer Engineering Department parameters, energy consumption, and temperature of five [email protected] rooms is developed based on data. As the energy opti- mization model includes nonparametric component mod- Chinmay Hegde els, computational intelligence algorithms are applied to MIT solve it. Experiments are designed to analyze performance [email protected] of the computational intelligence algorithms. The com- putational experiments have confirmed significant energy Aswin Sankaranarayanan savings. CMU [email protected] Andrew Kusiak Department of Mechanical and Industrial Engineering Wotao Yin University of Iowa Department of Mathematics 144 OP14 Abstracts

University of California at Los Angeles Univ. Wisconsin [email protected] [email protected], [email protected]

Rob Nowak MS69 Department of Electrical and Computer Engineering Convex Relaxation of Optimal Power Flow University of Wisconsin-Madison [email protected] The optimal power flow (OPF) problem is fundamental in power systems as it underlies many applications such as economic dispatch, unit commitment, state estimation, MS70 stability and reliability assessment, volt/var control, de- Piecewise Linear Multicommodity Flow Problems mand response, etc. OPF seeks to optimize a certain ob- jective function, such as power loss, generation cost and/or Abstract Not Available user utilities, subject to Kirchhoff’s laws, power balance as well as capacity, stability and security constraints on the Bernard Gendron voltages and power flows. It is a nonconvex quadratically CIRRELT and DIRO, Universit´edeMontr´eal constrained quadratic program. This is a short survey of [email protected] recent advances in the convex relaxation of OPF.

Steven Low MS70 Caltech Fixed-Charge Multicommodity Flow Problems [email protected] Abstract Not Available

MS69 Enrico Gorgone Using Regularization for Design: Applications in Dipartimento di Elettronica Informatica e Sistemistica Distributed Control Universit`a della Calabria [email protected] Both distributed optimal control and sparse reconstruction have seen tremendous success recently. Although the ob- MS70 jectives of these fields are very different – the former is concerned with the synthesis and design of controllers, the Piecewise Convex Multicommodity Flow Problems latter with recovering some underlying signal – they are united through their quest for structure. We show that Abstract Not Available ideas from sparse approximation can be used for the co- Philippe Mahey design of communication networks that are well-suited for LIMOS, Universit´e Blaise-Pascal, Clermont-Ferrand distributed optimal control. [email protected] Nikolai Matni California Institute of Technologu MS71 [email protected] Denoising for Simultaneously Structured Signals

Signals or models exhibiting low dimensional behavior play MS69 an important role in statistical signal processing and sys- Covariance Sketching tem identification. We focus on signals that have multiple structures simultaneously; e.g., matrices that are both low Learning covariance matrices from high-dimensional data rank and sparse, arising in phase retrieval, sparse PCA, and is an important problem that has received a lot of atten- cluster detection in social networks. We consider the esti- tion recently. For example, network links can be detected mation of such signals when corrupted by additive Gaus- by identifying strongly correlated variables. We are partic- sian noise, and provide tight upper and lower bounds on ularly interested in the high-dimensional setting, where the the mean squared error (MSE) of a denoising program that number of samples one has access to is much fewer than uses a combination of convex regularizers to induce multi- the number of variates. Fortunately, in many applications ple structures. In the case of low rank and sparse matrices, of interest, the underlying covariance matrix is sparse and we quantify the gap between the error of this convex pro- hence has limited degrees of freedom. In most existing gram and the best achievable error. work however, it is assumed that one can obtain samples of all the variates simultaneously. This could be very ex- Maryam Fazel pensive or physically infeasible in some applications. As University of Washington a means of overcoming this limitation, we propose a new Electrical Engineering framework whereby: (a) one can pool information about [email protected] the covariates by forming sketches of the samples and (b) reconstruct the original covariance matrix from just these sample sketches. We show theoretically that this is indeed MS71 possible and we discuss efficient algorithms to solve this Algorithmic Blessings of High Dimensional Prob- problem. lems

Parikshit Shah The curse of dimensionality is a phrase used by scientists Philips Research/University of Wisconsin to indicate the following observation: as the dimension of [email protected] the data exceeds the number of observations, many infer- ence tasks become seemingly more complicated. The last Gautam Dasarathy, Badri Narayan Bhaskar decade has witnessed the efforts of many researchers who OP14 Abstracts 145

proposed a variety of algorithms that can cope with the zon and Approximations high-dimensionality of the data. In this line of work, high- dimension is often viewed as a curse and the effort is to Threshold risk measures is a class of non-coherent risk propose schemes that work well even when this curse hap- measures that quantifies the risk of being above fixed pens. In this talk, I take a different approach and show the thresholds. These measures appear naturally in the dy- algorithmic blessings of high dimensions. As a concrete ex- namic optimization of energy systems where crucial eco- ample, I discuss the iterative thresholding algorithms for nomic thresholds are set. In this talk we introduce the solving the 1-regularized least squares problem. Then, I infinite horizon risk-averse dynamic optimization method show that the high dimensionality of the data enables us to with threshold risk measures. Tho solve these, we develop obtain the optimal value of parameters and linear conver- variants of the policy and value iteration algorithms, and gence. These goals cannot be achieved for low-dimensional an approximate dynamic programming algorithm with per- problems. This is based on a joint work with Ali Mousavi formance guarantees. and Richard Baraniuk. Ricardo A. Collado Arian Maleki IOMS-Operations Management Group Columbia University Stern School of Business, NYU [email protected] [email protected]

MS71 MS74 Learning from Heterogenous Observations Statistical Estimation of Composite Risk Measures and Risk Optimization Problems The aggregation of various sources of data to construct massive databases does not always come as a benefit to Motivated by estimation of law-invariant coherent mea- statistical methods. We study a canonical model: sparse sures of risk, we deal with statistical estimation of com- linear regression when the observed sample may come from posite functionals. Additionally, the optimal values of different, unidentified models. This question raises not composite risk functionals are analysed when they are only statistical but also algorithmic questions. parametrized by vectors from a deterministic convex set. We establish central limit formulae and characterize the Philippe Rigollet limiting distribution of the corresponding empirical estima- Princeton University tors. The results are applied to characterize the asymptotic [email protected] distribution of estimators for the mean-semi-deviation risk measures and higher order inverse risk measures. Central limit formulae for the optimal value function of problems MS71 with these risk measures in the objective are established as An Enhanced Conditional Gradient Algorithm for well. Atomic-Norm Regularization Darinka Dentcheva In many applications in signal and image processing, com- Department of Mathematical Sciences munications, system identification, one aims to recover a Stevens Institute of Technology signal that has a simple representation, as a linear combi- [email protected] nation of a modest number of elements from a certain basis. Atoms and atomic norms are key devices for finding and ex- pressing such representations. Fast and efficient algorithms MS74 have been proposed to solve reconstruction problems in im- Multilevel Optimization Modeling for Stochastic portant special cases (such as compressed sensing), but a Programming with Coherent Risk Measures framework that handles the general atomic-norm setting would be a further contribution to unifying the area. We Although there is only one decision maker, we propose a propose a method that combines greedy selection of atoms multilevel optimization modeling scheme that allows sim- with occasional reduction of the basis. Without the reduc- ple, law-invariant coherent risk measures to be used in tion step, the approach reduces to the conditional gradient stochastic programs with more than two stages without or “Frank-Wolfe’ scheme for constrained convex optimiza- running afoul of time inconsistency. We motivate the needs tion. for such an approach, but also show that the resulting mod- els are NP-hard even in the simplest cases. On the other Nikhil Rao hand, we also present empirical evidence that some non- Dept. of ECE trivial practical instances may not be difficult to solve. University of Wisconsin [email protected] Jonathan Eckstein Rutgers University [email protected] Parikshit Shah Philips Research/University of Wisconsin [email protected] MS74 Challenges in Dynamic Risk-Averse Optimization Stephen Wright University of Wisconsin We shall discuss modern challenges in dynamic risk-averse Dept. of Computer Sciences optimization. In particular, we shall address the issue of [email protected] time-consistency and its relations to the dynamic program- ming principle. We shall consider time-consistent approx- imations of time-inconsistent models. Next, we shall dis- MS74 cuss Markov consistency, which determines the usefulness Threshold Risk Measures: Dynamic Infinite Hori- of a dynamic risk measure for as an objective functional 146 OP14 Abstracts

in control of Markov models. We shall derive dynamic Sanjeev Arora programming equations for Markov decision problems with Princeton University Markov consistent measures, and discuss methods for their Computer Science Dept. solution. [email protected]

Andrzej Ruszczynski Yoni Halpern Rutgers University New York University MSIS Department Computer Science Dept. [email protected] [email protected]

MS75 David Mimno Princeton University Identifying K Largest Submatrices Using Ky Fan Computer Science Dept. Matrix Norms [email protected] We propose a convex relaxation using Ky Fan matrix norms for the problem of identifying k largest approximately rank- Ankur Moitra one submatrices of a nonnegative data matrix. This prob- IAS lem is related to nonnegative matrix factorization and has [email protected] important applications in data mining. We show that un- der some certain randomized model, then the k largest David Sontag blocks can be recovered via the proposed convex relaxation. New York University Computer Science Dept. [email protected] Xuan Vinh Doan University of Warwick Yichen Wu, Michael Zhu Warwick Business School Princeton University [email protected] Computer Science Dept. [email protected], [email protected] Stephen A. Vavasis University of Waterloo Dept of Combinatorics & Optimization MS75 [email protected] Semidefinite Programming Based Preconditioning for More Robust Near-Separable Nonnegative Ma- trix Factorization MS75 Convex Lower Bounds for Atomic Cone Ranks with Nonnegative matrix factorization under the separability as- Applications to Nonnegative Rank and cp-rank sumption can be solved efficiently, even in the presence of noise. This problem is referred to as near-separable NMF We propose new lower bounds using semidefinite program- and requires that there exists a cone spanned by a subset of ming on a class of atomic rank functions defined on a con- the columns of the input matrix containing all columns. In vex cone. We focus on two important special cases which this talk, we propose a preconditioning based on semidefi- are the nonnegative rank and the cp-rank and we show that nite programming which can improve significantly the per- our lower bound has interesting connections with existing formance of near-separable NMF algorithms. We illustrate combinatorial bounds. Our lower bound also inherits many our result on some hyperspectral images. of the structural properties satisfied by these rank functions such as invariance under diagonal scaling and subadditiv- Nicolas Gillis ity. Universit´e catholique de Louvain Center for Operation Research and Econometrics Hamza Fawzi, Pablo A. Parrilo [email protected] Massachusetts Institute of Technology [email protected], [email protected] Stephen A. Vavasis University of Waterloo MS75 Dept of Combinatorics & Optimization [email protected] Provable and Practical Topic Modeling Using NMF

Topic modeling is a classical application of NMF. Recently, MS76 [Arora et al.] showed NMF can be solved in polynomial time under separability. This assumption naturally trans- Dsos and Sdsos: More Tractable Alternatives to lates to anchor-words assumption for topic modeling. In Sum of Squares Programming this talk I will address some problems that arise when ap- plying NMF to topic modeling, and how to redesign NMF We present linear and second order inner approximations to algorithms to solve those problems. The final algorithm has the sum of squares cone and new hierarchies for polynomial performance close to classical Gibbs sampling algorithms optimization problems that are amenable to LP and SOCP but is much faster. and considerably more scalable than the SDP-based sum of squares approaches. Rong Ge Princeton University Amir Ali Ahmadi Computer Science Department IBM Research [email protected] a a [email protected] OP14 Abstracts 147

Anirudha Majumdar Zhao Sun MIT Tilburg University [email protected] [email protected]

MS76 MS77 Lower Bounds for a Polynomial on a basic Closed On Universal Rigidity of Tensegrity Frameworks, a Semialgebraic Set using Geometric Programming Gram Matrix Approach A tensegrity framework (G, p)inRr is a graph G,where Let f,g1,...,gm be polynomials in R[x1,...,xn]. We deal each edge is labeled as either a bar, a cable, or a strut; and with the general problem of computing a lower bound for i r n where each node i is mapped to a pint p in R . Connelly f on the subset of R defined by the inequalities gi ≥ 0, proved that if an r-dimensional tensegrity framework (G, p) i =1,...,m. We show that there is an algorithm for r computing such a lower bound, based on geometric pro- on n vertices in R admits a proper semidefinite stress ma- − − 1 n gramming, which applies in a large number of cases. The trix of rank n r 1, and if configuration p =(p ,...,p ) bound obtained is typically not as good as the bound ob- is generic, then (G, p) is universally rigid. In this talk, we tained using semidefinite programming, but it has the ad- show that this result continues to hold under the much weaker assumption that in configuration p, each point and vantage that it is computable rapidly, even in cases where r the bound obtained by semidefinite programming is not its neighbors in G affinely span R . computable. Abdo Alfakih University of Windsor Mehdi Ghasemi [email protected] NTU, Singapore [email protected] Viet-Hang Nguyen G-SCOP, Grenoble, France Murray Marshall viet-hang [email protected] University of Saskatchewan [email protected] MS77 MS76 On the Sensitivity of Semidefinite Programs and Second Order Cone Programs Approximation Quality of SOS Relaxations Given a feasible conic program with finite optimal value Sums of squares (SOS) relaxations provide efficiently com- that does not satisfy strong duality, a small feasible per- putable lower bounds for minimization of multivariate turbation of the problem data may lead to a relatively big polynomials. Practical experience has shown that these change in the optimal value. We quantify the notion of bounds usually outperform most other available tech- big change in the cases of semidefinite programs and of niques, but a fully satisfactory theoretical justification is second order cone programs. If there is a nonzero duality still lacking. In this talk, we discuss several results (new gap, then one can always find an arbitrarily small feasible and old) about the approximation quality of these SOS perturbation so that the change in optimal value is no less bounds, focusing on the case of polynomial optimization than the duality gap. If there is a zero duality gap, then the on the sphere. change in optimal value due to any sufficiently small and feasible right-hand side perturbation S is bounded above Pablo A. Parrilo 1/2d by κS ,whereκ is a fixed constant and d is the degree Massachusetts Institute of Technology of singularity of the optimization problem, i.e., the number [email protected] of facial reduction iterations required to find the minimal face of the optimization problem. We will also discuss some MS76 applications of these results to the solution of semidefinite and second order cone programs in the absence of strictly An Alternative Proof of a PTAS for Fixed-degree complementary solutions. Polynomial Optimization over the Simplex Yuen-Lam Cheung The problem of minimizing a polynomial over the stan- Dept. of Comb. & Opt. dard simplex is a well-known NP-hard nonlinear optimiza- University of Waterloo tion problem. It is known that the problem allows a [email protected] polynomial-time approximation scheme (PTAS) for poly- nomials of fixed degree. We provide an alternative proof of Henry Wolkowicz the PTAS property for one simple scheme that only evalu- University of Waterloo ates the polynomial on a regular grid, and the proof relies Department of Combinatorics and Optimization on the properties of Bernstein approximation on the sim- [email protected] plex.

Etienne De Klerk MS77 Tilburg University Combinatorial Certificates in Semidefinite Duality [email protected] A semidefinite system is called badly behaved,ifforsome Monique Laurent objective function the dual does not attain, or has a pos- CWI, Amsterdam itive gap with the primal; well-behaved, if not badly be- and Tilburg University haved. We show that combinatorial type certificates exist [email protected] to easily verify the badly/well behaved nature of a semidef- 148 OP14 Abstracts

inite system, using only elementary linear algebra. This ing, and sparse optimization over arbitrary atomic norms. work continues the work presented in Bad semidefinite pro- grams: they all look the same. Martin Jaggi Berkeley Gabor Pataki [email protected] University of North Carolina, at Chapel Hill [email protected] MS78 MS77 Statistical Properties for Computation Efficient Use of Semidefinite Programming for Se- lection of Rotamers in Protein Conformations Abstract Not Available

In this paper we study a semidefinite programming relax- Sahand Negahban ation of the (NP-hard) side chain positioning problem. We MIT show that the Slater constraint qualification (strict feasi- [email protected] bility) fails for the SDP relaxation. We then show the advantages of using facial reduction to regularize the SDP. In fact, after applying facial reduction, we have a smaller MS78 problem that is more stable both in theory and in practice. Minimizing Finite Sums with the Stochastic Aver- age Gradient Algorithm

Henry Wolkowicz We propose the stochastic average gradient (SAG) method University of Waterloo for optimizing the sum of a finite number of smooth con- Department of Combinatorics and Optimization vex functions. Like stochastic gradient (SG) methods, the [email protected] SAG method’s iteration cost is independent of the number of terms in the sum. However, by incorporating a mem- Forbes Burkowski ory of previous gradient values the SAG method achieves a Dept. of Computer Science faster convergence rate than black-box SG√ methods. The University of Waterloo convergence rate is improved from O(1/ k)toO(1/k)in [email protected] general, and when the sum is strongly-convex the conver- gence rate is improved from the sub-linear O(1/k) to a lin- Yuen-Lam Cheung ear convergence rate of the form O(ρk)forρ<1. Further, Dept. of Comb. & Opt. in many cases the convergence rate of the new method is University of Waterloo also faster than black-box deterministic gradient methods, [email protected] in terms of the number of gradient evaluations. Numerical experiments indicate that the new algorithm often dramat- ically outperforms existing SG and deterministic gradient MS78 methods. Further, we argue that the performance may be Frank-Wolfe-like Methods for Large-scale Convex further improved through the use non-uniform sampling Optimization strategies.

We develop and analyze variants of the Frank-Wolfe Mark Schmidt method, particularly designed to be mindful of large-scale Simon Fraser University problems over domains with favorable sparsity properties. [email protected] We analyze the methods presented in terms of the trade- offs between solution accuracy and sparsity guarantees. We also adapt our methods and analysis to large-scale parallel MS79 and/or distributed computing environments. Sensitivity Analysis of Markov Decision Processes Paul Grigas and Policy Convergence Rate of Model-Based Re- M.I.T. inforcement Learning Operations Research Center [email protected] Consider a finite-state discounted Markov Decision Prob- lem (MDP) where the transition probabilities have to be Robert M. Freund statistically estimated through observations, as is done, MIT e.g., in reinforcement learning. We analyze the MDP Sloan School of Management through its linear programming (LP) representation. This [email protected] allows us to use old-fashioned LP post-optimality sensitiv- ity analysis to provide sufficient conditions on the differ- ence between the true and estimated transition probabili- MS78 ties that ensure that the optimal policy can be discovered Frank-Wolfe and Greedy Algorithms in Optimiza- by solving the approximate MDP. tion and Signal Processing Marina A. Epelman The Frank-Wolfe algorithm is one of the earliest first-order University of Michigan optimization algorithms, and has recently seen a late re- Industrial and Operations Engineering vival in several large-scale machine learning and signal pro- [email protected] cessing applications. We discuss some recent new insights for such methods, highlighting in particular the close con- Ilbin Lee nection to popular sparse greedy methods in signal process- University of Michigan OP14 Abstracts 149

[email protected] optimal, which constitutes a significant improvement in the usual situation when γ is close to 1. Surprisingly, this shows that the problem of “computing near-optimal non- MS79 stationary policies’ is much simpler than that of “comput- Partially Observable Markov Decision Processes ing near-optimal stationary policies’. with General State and Action Spaces Bruno Scherrer For Partially Observable Markov Decision Processes INRIA (POMDPs) with Borel state, observation, and action sets [email protected] and with the expected total costs, this paper provides suf- ficient conditions for the existence of optimal policies and validity of other optimality properties including that op- MS80 timal policies satisfy optimality equations and value iter- Polyhedral Projection ations converge to optimal values. Action sets may not be compact and one-step functions may not be bounded. A fast algorithm is developed for computing the projection Since POMDPs can be reduced to Completely Observ- onto a polyhedron. The algorithm solves the dual projec- able Markov Decision Processes (COMDPs), whose states tion problem in two phases. In one phase a nonmonotone are posterior state distributions, this paper focuses on the SpaRSA algorithm is used to handle nonsmoothness in the validity of the above mentioned optimality properties for dual problem, while the second phase uses active set tech- COMDPs. niques. The second phase could be implemented with a sparse linear solver and update/downdate techniques or Eugene A. Feinberg with the conjugate gradient method. The projection algo- Stony Brook University rithm can exploit a warm starting point. [email protected] William Hager University of Florida Pavlo Kasyanov, Michael Zgurovsky hager@ufl.edu National Technical University of Ukraine [email protected], [email protected] Hongchao Zhang Department of Mathematics and CCT MS79 Louisiana State University [email protected] Dantzig’s Pivoting Rule for Shortest Paths, Deter- ministic Markov Decision Processes, and Minimum Cost to Time Ratio Cycles MS80 Orlin showed that the simplex algorithm with Dantzig’s A Filter Method with Unified Step Computation pivoting rule solves the shortest paths problem using 2 We will present a new filter line search method for solving at most O(mn log n) pivoting steps for graphs with m general nonlinear nonconvex optimization problems. This edges and n vertices. Post and Ye recently showed that 2 3 2 3 5 2 filter method avoids the restoration phase required in tra- O(m n log n)andO(m n log n)stepssufficetosolve ditional filter methods by using a penalty mode instead. deterministic Markov decision processes with uniform and The same step computation procedure is used at each iter- varying discount factors, respectively. We improve these ation, and the generated trial step always incorporates in- bounds by a factor of n, assuming in the last case that the formation from both the objective function and constraint discounts are close to 1. violation. We will present convergence results and report on numerical experiments. Thomas D. Hansen Stanford University Yueling Loh [email protected] Johns Hopkins University [email protected] Haim Kaplan Tel-Aviv University Nicholas I.M. Gould [email protected] Rutherford Appleton Laboratory UK Uri Zwick [email protected] Tel Aviv University [email protected] Daniel Robinson Northwestern University [email protected] MS79 On the Use of Non-Stationary Policies for Station- ary Infinite-Horizon Markov Decision Processes MS80 An Interior-Point Trust-Funnel Algorithm for Non- I will consider infinite-horizon stationary γ-discounted linear Optimization Markov Decision Processes, for which it is known that there exists a stationary optimal policy. Using Value and Policy We present an inexact barrier-SQP trust-funnel algorithm Iteration with some error at each iteration, it is well- for solving large-scale nonlinear optimization problems. known that one can compute stationary policies that are 2γ Our method, which is designed to solve problems with both (1−γ)2 -optimal. After arguing that this guarantee is tight, equality and inequality constraints, achieves global conver- I will describe variations of Value and Policy Iteration for gence guarantees by combining a trust-region methodology 2γ computing non-stationary policies that can be up to 1−γ - with a funnel mechanism. The prominent features of our 150 OP14 Abstracts

algorithm are that (i) the subproblems that define each computations of MIP solvers. search direction may be solved approximately, (ii) critical- ity measures for feasibility and optimality aid in determin- Arie Koster ing which subset of computations will be performed dur- Zuse Institute Berlin ing each iteration, (iii) no merit function or filter is used, [email protected] (iv) inexact sequential quadratic optimization steps may be computed when advantageous, and (v) it may be im- Martin Tieves plemented matrix-free so that derivative matrices need not RWTH Aachen University be formed or factorized so long as matrix-vector products [email protected] with them can be performed. Preliminary numerical tests will be given. MS81 Daniel Robinson Solving Network Design Problems Via Iterative Northwestern University Aggregation [email protected] We present an exact approach for solving network design Frank E. Curtis problems (NDP). Motivated by applications in transporta- Industrial and Systems Engineering tion and energy optimization, the instances may contain Lehigh University preexisting capacities such that some percentage of the de- [email protected] mand can already be routed. Starting with an initial net- work aggregation, we solve a sequence of NDP over increas- Nicholas I.M. Gould ingly fine-grained representations until a globally optimum Rutherford Appleton Laboratory solution is determined. Computational results on realis- UK tic networks show a drastic improvement over solving the [email protected] original problem from scratch. Andreas B¨armann Philippe L. Toint Friedrich-Alexander-Universit¨at Erlangen-N¨urnberg University of Namur, Belgium [email protected] [email protected] Frauke Liers MS80 Institut f¨ur Informatik Universit¨at zu K¨oln Globalizing Stabilized SQP with a Smooth Exact [email protected] Penalty Function Alexander Martin Identifying a suitable merit/penalty function for which the University of Erlangen-N¨urnberg direction given by the so-called stabilized sequencial pro- [email protected] gramming algorithm is that of descent, proved to be quite a challenge. In this work, for an equality-contrained prob- lem, we propose for the task a smooth two-parameter ex- Maximilian Merkert, Christoph Thurner, Dieter Weninger act penalty function consisting of the the augmented La- Friedrich-Alexander-Universit¨at Erlangen-N¨urnberg grangian plus a penalty for violating Lagrangian station- [email protected], arity. [email protected], [email protected] Alexey Izmailov Computing Center of the Russian Academy of Sciences [email protected] MS81 The Recoverable Robust Two-Level Network De- Mikhail V. Solodov sign Problem Inst de Math Pura e Applicada [email protected] We consider a network design application which is modeled as the two level network design problem under uncertainty. In this problem, one of the two available technologies can Evgeniy Uskov be installed on each edge and all customers of the network MSU, Moscow need to be served by at least the lower level (secondary) [email protected] technology. The decision maker is confronted with uncer- tainty regarding the set of primary customers, i.e., the set of nodes that need to be served by the higher level (pri- MS81 mary) technology. A set of discrete scenarios associated to Network Design under Compression and Caching the possible realizations of primary customers is available. Rates Uncertainty The network is built in two stages. In the first-stage the network topology must be determined. One may decide Energy-aware routing can significantly reduce energy con- to install the primary technology on some of the edges in sumption in a communication network. Compression and the first stage, or one can wait to see which scenario will caching techniques can be activated at the nodes to reduce be realized, in which case, edges with the installed sec- the required bandwidths further. Given demands and com- ondary technology may be upgraded, if necessary to pri- pression rates, our aim is to find a minimum energy net- mary technology, but at higher recovery cost. The overall work configuration. In this talk, we discuss the generaliza- goal then is to build a spanning tree in the first stage that tion where the compression rates are uncertain. We derive serves all customers by at least the lower level technology, a robust formulation and cutting planes to speed-up the and that minimizes the first stage installation cost plus the OP14 Abstracts 151

worst-case cost needed to upgrade the edges of the selected eral varia- tional inequality problem. Based of the D-gap tree, so that the primary customers of each scenario can be function, we propose se- quential inexact method for solv- served using the primary technology. Using the recently in- ing the the general variational inequality and discuss the troduced concept of recoverable robustness, we address this convergence properties. We extend the concept of exact problem of importance in the design of telecommunication regularization to generalized variational inequality prob- and distribution networks, and provide mixed integer pro- lems and establish error bounds to the solution sets of the gramming models and a branch-and-cut algorithm to solve regularized problems. In the case when exact regulariza- it. tion fails to exist, we give an estimate for the distance between the solution sets of regularized and original prob- Eduardo A. Alvarez-Miranda lems. Universidad de Talca, Merced 437 [email protected] Charitha Charitha University of G¨ottingen Ivana Ljubic [email protected] University of Vienna [email protected] Joydeep Dutta ITT Kanpur, India S. Raghavan [email protected] The Robert H. Smith School of Business University of Maryland, USA Russell Luke [email protected] Institute for Numerical and Applied University of Goettingen Paolo Toth [email protected] University of Bologna [email protected] MS82 Convergence, Robustness, and Set-Valued Lya- MS81 punov Functions Robust Network Design for Simple Polyhedral De- mand Uncertainties Every optimization algorithm leads to a discrete-time dy- namical system, which can be analyzed through control We consider a basic network design problem. Given a net- theory tools. Minimizers correspond to equilibria, conver- work (V,E)andasetD of supply/demand vectors, find gence and robustness can be studied through Lyapunov minium cost capacities such that any d ∈Dcan be bal- techniques. This talk presents a robustness result applica- anced. We extend an exact approach by [Buchheim, Liers ble to a class of algorithms, obtained through the use of a and Sanit, INOC 2011]. The uncertainty set D can be a set-valued Lyapunov mapping. The technique is motivated finite set or a specific polyhedron and we give an integer by problems of convergence to a consensus in a multi-agent programming formulation with O(|E|) variables and a sep- system and applies to systems with a continuum of equi- aration algorithm for both cases. libria.

Valentina Cacchiani Rafal Goebel University of Bologna Loyola University Chicago DEI [email protected] [email protected] MS82 Michael J¨unger Universit¨at zu K¨oln Global and Local Linear Convergence of Institut f¨ur Informatik Projection-Based Algorithms for Sparse Affine [email protected] Feasibility The problem of finding a vector with the fewest nonzero Frauke Liers elements that satisfies an underdetermined system of lin- Institut f¨ur Informatik ear equations is an NP-complete problem that is typically Universit¨at zu K¨oln solved numerically via convex heuristics or nicely-behaved [email protected] nonconvex relaxations. In this work we consider elemen- tary methods based on projections for solving a sparse fea- Andrea Lodi sibility problem without employing convex heuristics. In University of Bologna a recent paper Bauschke, Luke, Phan and Wang (2013) DEIS showed that, locally, the fundamental method of alternat- [email protected] ing projections (AP) must converge linearly to a solution to the sparse feasibility problem with an affine constraint. Daniel Schmidt In this paper we apply different analytical tools that allow Institut f¨ur Informatik us to show global linear convergence of AP and a relaxation Universit¨at zu K¨oln thereof under familiar constraint qualifications. These an- [email protected] alytical tools can also be applied to other algorithms. This is demonstrated with the prominent DouglasRachford al- gorithm where we establish local linear convergence of this MS82 method applied to the sparse affine feasibility problem. Regularization Methods for Variational Inequalties Russell Luke In this paper we study regularization methods for the gen- Institute for Numerical and Applied 152 OP14 Abstracts

University of Goettingen Lagrangian. Each iteration involves the solution of a reg- [email protected] ularized quadratic program (QP) that is equivalent to a strictly convex bound-constrained QP based on minimiz- Robert Hesse, Patrick Neumann ing a quadratic model of the augmented Lagrangian. The University of G¨ottingen solution of the QP subproblem will be discussed in the con- [email protected], pneumann@uni- text of applying active-set and interior methods that are math.gwdg.de themselves regularized versions of conventional methods. Philip E. Gill MS82 University of California, San Diego Generalized Solutions for the Sum of Two Maxi- Department of Mathematics mally Monotone Operators [email protected]

A common theme in mathematics is to define generalized Vyacheslav Kungurtsev solutions to deal with problems that potentially do not Electical Engineering Department, KU Leuven have solutions. A classical example is the introduction of [email protected] least squares solutions via the normal equations associated with a possibly infeasible system of linear equations. In Daniel Robinson this talk, we introduce a ’normal problem” associated with Northwestern University finding a zero of the sum of two maximally monotone op- [email protected] erators. If the original problem admits solutions, then the normal problem returns this same set of solutions. The nor- mal problem may yield solutions when the original problem MS83 does not admit any; furthermore, it has attractive varia- Steering Augmented Lagrangian Methods tional and duality properties. Several examples illustrate our theory. We propose an enhanced augmented Lagrangian algorithm for solving large-scale optimization problems with both Walaa Moursi equality and inequality constraints. The novel feature of University of British Columbia the algorithm is an adaptive update for the penalty pa- walaa [email protected] rameter motivated by recently proposed techniques for ex- act penalty methods. We provide convergence results from MS83 remote starting points and illustrate by a set of numeri- cal experiments that our method outperforms traditional A Primal-Dual Active-Set Method for Convex augmented Lagrangian methods in terms of critical perfor- Quadratic Programming mance measures.

We discuss active-set methods for convex quadratic pro- Hao Jiang gram with general equality constraints and simple lower Department of Applied Mathematics and Statistics bounds on the variables. In the first part of the talk, two Johns Hopkins University methods are proposed, one primal and one dual. In the [email protected] second part of the talk, a primal-dual method is proposed that solves a sequence of quadratic programs created from the original by simultaneously shifting the simple bound Daniel Robinson constraints and adding a penalty term to the objective Northwestern University function. [email protected]

Anders Forsgren Frank E. Curtis KTH Royal Institute of Technology Industrial and Systems Engineering Dept. of Mathematics Lehigh University [email protected] [email protected]

Philip E. Gill, Elizabeth Wong University of California, San Diego MS83 Department of Mathematics Recent Developments in the SNOPT and NPOPT [email protected], [email protected] Packages for Sequential Quadratic Programming

We discuss recent developments in the nonlinear optimiza- MS83 tion packages SNOPT and NPOPT. Both packages are de- Regularized Sequential Quadratic Programming signed to solve nonlinear programming problems. SNOPT Methods is best suited for sparse large-scale problem, while NPOPT is best suited for dense medium-scale problems. New de- Regularized and stabilized sequential quadratic program- velopments include the incorporation of second derivatives ming (SQP) methods are two classes of methods designed and the implementation of a new quadratic solver. Numer- to resolve the numerical and theoretical difficulties asso- ical results will be presented. ciated with ill-posed or degenerate nonlinear optimization problems. Recently, a regularized SQP method has been Elizabeth Wong, Philip E. Gill proposed that provides a strong connection between aug- University of California, San Diego mented Lagrangian methods and stabilized SQP meth- Department of Mathematics ods. The method is formulated as a regularized SQP [email protected], [email protected] method with an implicit safeguarding strategy based on minimizing a bound-constrained primal-dual augmented Michael A. Saunders OP14 Abstracts 153

Systems Optimization Laboratory (SOL) [email protected] Dept of Management Sci and Eng, Stanford [email protected] MS84 Large Scale 2-Stage Unit-Commitment: A MS84 Tractable Approach

Improved Mixed Integer Linear Optimization For- In energy management, generation companies have to sub- mulations for Unit Commitment mit a generation schedule to the grid operator for the com- ing day. Computing such an optimal schedule is known We present two ways to improve the mixed-integer linear as the ”unit-commitment” problem. In electrical systems optimization formulation for the unit commitment prob- wherein renewable generation has overall high generation lem. The first is a new class of inequalities that give a capacity, uncertainty is strongly present. Generation com- tighter description of the feasible generator schedules. The panies can therefore occasionally submit a change to the second is a modified orbital branching technique that ex- originally submitted schedule. These changes can be seen ploits the symmetry created by identical generators. Com- as intra-daily recourse actions. Recourse is incomplete be- putational results show that these approaches can signifi- cause of technical constraints on generation. It is of inter- cantly reduce overall solution times for realistic instances est to investigate the impact of uncertainty and recourse of unit commitment. on the originally submitted schedule. In this work we will investigate a two stage formulation of unit-commitment Miguel F. Anjos wherein both the first and second stage problems are full Mathematics and Industrial Engineering & GERAD unit-commitment problems. We propose a primal-dual de- Ecole´ Polytechnique de Montr´eal composition approach, whose computational core makes [email protected] extensive use of warmstarted bundle methods. Results are demonstrated on typical unit-commitment problems. James Ostrowski University of Tennessee Knoxville Wim Van Ackooij [email protected] EDF R&D Department OSIRIS [email protected] Anthony Vannelli University of Guelph [email protected] Jerome Malick CNRS [email protected] MS84 Efficient Solution of Problems with Probabilistic MS84 Constraints Arising in the Energy Sector Fast Algorithms for Two-Stage Robust Unit Com- mitment Problem For problems with constraints involving random parame- ters due to (market, weather) uncertainty, if a decision x To hedge against randomness in power systems, we employ has to be taken “here and now”, no matter how carefully x robust optimization method to construct two-stage robust is chosen, there is no guarantee that the random constraint unit commitment models. To solve those challenging prob- will be satisfied for all possible realizations of the uncertain lems, we investigate their structural properties and develop parameters. A chance-constrained model declares x feasi- advanced algorithm strategies for fast computation. Nu- ble if the constraint probability is higher than certain safety merical results on typical IEEE testbeds will be presented. level. To solve this type of problems a dual approach was recently proposed by D. Dentcheva and M. G. Mart´ınez, using the so-called p-efficient points. We discuss the bene- Bo Zeng, Anna Danandeh, Wei Yuan fits of applying an inexact bundle algorithm in such a set- USF ting and assess the interest of the proposal on a problem [email protected], [email protected], arising in unit-commitment to optimally manage a hydro- [email protected] valley, with several hydropower plants cascaded along the same basin. MS85 Claudia A. Sagastizabal The DouglasRachford Algorithm for Two Sub- IMPA spaces Visiting Researcher [email protected] I will report on recent joint work (with J.Y. Bello Cruz, H.M. Phan, and X. Wang) on the DouglasRachford algo- rithm for finding a point in the intersection of two sub- Wim Van Ackooij spaces. We prove that the method converges strongly to EdF France the projection of the starting point onto the intersection. [email protected] Moreover, if the sum of the two subspaces is closed, then the convergence is linear with the rate being the cosine Violette Berge of the Friedrichs angle between the subspaces. Our results ENSTA ParisTech improve upon existing results in three ways: First, we iden- [email protected] tify the location of the limit and thus reveal the method as a best approximation algorithm; second, we quantify the Welington de Oliveira rate of convergence, and third, we carry out our analysis in IMPA general (possibly infinite-dimensional) Hilbert space. We 154 OP14 Abstracts

also provide various examples as well as a comparison with Algorithm the classical method of alternating projections. Reference: http://arxiv.org/pdf/1309.4709v1.pdf This research focuses on finite convergence of a subgradient projections algorithm for solving convex feasibility problem Heinz Bauschke in the Euclidean space. I will present properties of cutters University of British Columbia and subgradient projections and applications to subgradi- Mathematics ent projection algorithms. Our convergence results relate [email protected] to previous work by Crombez and by Polyak.

Jia Xu MS85 University of British Columbia [email protected] Inheritance of Properties of the Resolvent Average

Maximally monotone operators play a key role in many op- Heinz Bauschke timization problems. It is well known that the sum of two University of British Columbia maximally monotone operators is not always maximally Mathematics monotone; however the resolvent average always maintains [email protected] maximal monotonicity. In this talk, we discuss which other desirable properties the resolvent average inherits from its Shawn Wang component operators, including rectangularity and para- University of British Columbia monotonicity. [email protected]

Sarah Moffat Caifang Wang University of British Columbia Shanghai Maritime University sarah.moff[email protected] [email protected]

Heinz Bauschke University of British Columbia MS86 Mathematics On Cones of Nonnegative Quartic Forms [email protected] In this talk we study the class of nonnegative polynomials. Xianfu Wang Six fundamentally important convex cones of nonnegative University of British Columbia quartic functions will be presented. It turns out that these [email protected] convex cones coagulate into a chain in decreasing order. The complexity status of these cones is sorted out as well. We further consider the polynomial sized representation of MS85 a very specific nonnegative polynomial, and this represen- tation enables us to address an open question asserting that Full Stability in Mathematical Programming the computation of the matrix 2 → 4normisNP-hardin general. The talk is devoted to the study of fully stable local minimizers of general optimization problems in finite- Bo Jiang dimensional spaces and its applications to classical nonlin- University of Minnesota ear programs with twice continuously differentiable data. [email protected] The importance of full stability has been well recognized from both theoretical and numerical aspects of optimiza- Zhening Li tion, and this notion has been extensively studied in the Department of Mathematics, University of Portsmouth literature. Based on advanced tools of second-order varia- Portsmouth PO1 3HF, United Kingdom tional analysis and generalized differentiation, we develop a [email protected] new approach to full stability, which allows us to derive not only qualitative but also quantitative characterizations of fully stable minimizers with calculating the corresponding Shuzhong Zhang moduli. The implementation of this approach and general Department of Industrial & Systems Engineering results in the classical framework of nonlinear program- University of Minnesota ming provides complete characterizations of fully stable [email protected] minimizers under new second-order qualification and op- timality conditions. MS86 Boris Mordukhovich Tensor Methods in Control Optimization Wayne State University Department of Mathematics We present some tensor based methods for solving opti- [email protected] mal control problems. In particular, we formulate high- dimensional nonlinear optimal control problems in a tensor framework and compute polynomial solutions using tensor Tran Nghia based approach. In addition, we look at stability and con- University of British Columbia trollability of ensemble control in the tensor framework. [email protected]

Carmeliza Navasca MS85 Department of Mathematics Finite Convergence of a Subgradient Projections University of Alabama at Birmingham OP14 Abstracts 155

[email protected] ear Knapsacks

We address the problem of developing new valid inequali- MS86 ties for S = {(x, y): i,j aij xiyj + i a0ixi + j bj yj ≤ r, xi ∈ [0,ui],yj ∈{0, 1}∀i, j} in the (x, y)-space. We Eigenvectors of Tensors and Waring Decomposition { study the convex hull of X = (x, y): j aj xyj + a0x + ≤ ∈ ∈{ }∀} In recent work with Ottaviani, we have developed algo- j bj yj r, x [0,u],yj 0, 1 j ,whichisare- rithms for symmetric tensor decomposition. I will explain laxation of S upon aggregation of variables. The set X these algorithms, their connection to algebraic geometry, also appears when maximizing hyperbolic functions and and possible applications. discretizing variables in a single bilinear term. Exploit- ing the supermodular structure in X and using disjunctive Luke Oeding programming techniques leads to an exponential family of University of California Berkeley inequalities, which define convX for certain values of a, b, u. [email protected] These valid inequalities from X can be strengthened when the coefficient matrix in S has rank one and/or x is dis- crete. Computational results with a branch-and-cut algo- MS86 rithm will also be discussed. Structured Data Fusion Akshay Gupte Mathematical Sciences We present structured data fusion (SDF) as a framework Clemson University for the rapid prototyping of coupled tensor factorizations. [email protected] In SDF, each data set—stored as a dense, sparse or incom- plete multiway array—is factorized with a tensor decompo- sition. Factorizations can be coupled with each other by in- MS87 dicating which factors should be shared between data sets. Convex Quadratic Programming with Variable At the same time, factors may be imposed to have any type Bounds of structure that can be constructed as an explicit function of some underlying variables. The scope of SDF reaches We aim to obtain a strong relaxation for the convex hull far beyond factor analysis alone, encompassing nearly the of a mixed binary set defined by convex non-separable full breadth of applications resulting from matrix factoriza- quadratic functions that appears in many important ap- tions and tensor decompositions. It subsumes, for example, plications. Our approach starts by reformulation through tasks based on dimensionality reduction such as feature Cholesky factorization. Several classes of linear and nonlin- extraction, subspace learning and model order reduction ear valid inequalities are derived. Computational results on and tasks related to machine learning such as regression, different formulations are compared, and we demonstrate classification, clustering, and imputation of missing data. that the derived inequalities are helpful in the solution pro- Tensorlab v2.0 offers a domain specific language (DSL) for cess of the application problems. modelling SDF problems. The three key ingredients of an SDF model are (1) defining variables, (2) defining factors Hyemin Jeon, Jeffrey Linderoth as transformed variables and (3) defining the data sets and University of Wisconsin-Madison their factorizations based on these factors. Tensor decom- [email protected], [email protected] positions and factor structure may be chosen in a modular way. Currently, Tensorlab comes with the choice of the Andrew Miller canonical polyadic decomposition (CPD), low multilinear UPS rank approximation (LMLRA) and block term decomposi- [email protected] tion (BTD) and ships with a library of 32 predefined fac- tor structures such as nonnegativity, orthogonality, Han- kel, Toeplitz, Vandermonde and the matrix inverse. By MS87 selecting the right combination of tensor decompositions and factor structures, even classical matrix factorizations Strong Convex Nonlinear Relaxations of the Pool- such as the eigenvalue decomposition, singular value de- ing Problem composition and QR factorization can be computed with SDF. We investigate relaxations for the non-convex pooling prob- lem, which arises in production planning problems in which products with are mixed in intermediate ”pools” in order to Laurent Sorber meet quality targets at their destinations. We derive valid KU Leuven nonlinear convex inequalities, which we conjecture define [email protected] the convex hull of this continuous non-convex set for some special cases. Numerical illustrations of the results will be Marc Van Barel presented. Katholieke Universiteit Leuven Department of Computer Science James Luedtke [email protected] University of Wisconsin-Madison Department of Industrial and Systems Engineering Lieven De Lathauwer [email protected] KU Leuven Kulak [email protected] Claudia D’Ambrosio CNRS,LIX,France [email protected] MS87 Supermodular Inequalities for Mixed Integer Bilin- Jeff Linderoth 156 OP14 Abstracts

University of Wisconsin-Madison Kim-Chuan Toh Dept. of Industrial and Sys. Eng National University of Singapore [email protected] Department of Mathematics and Singapore-MIT Alliance [email protected] Andrew Miller Institut de Math´ematiques de Bordeaux Sunyoung Kim France Ewha W. University [email protected] Department of Mathematics [email protected]

MS87 Cuts for Quadratic and Conic Quadratic Mixed In- MS88 teger Programming Mixed Integer Second-Order Cone Programming Formulations for Variable Selection We study algorithms and formulas for constructing the con- vex hull of the sets defined by two quadratic inequalities. Variable selection is to select the best subset of explanatory We also study the generalization of split and intersection variables in a multiple linear regression model. To evaluate a subset regression model, some goodness-of-fit measures, cuts from Mixed Integer Linear Programming to Mixed ¯2 Integer Conic Quadratic Programming. Constructing such such as AIC, BIC and R , are generally employed. A step- cuts requires characterizing the convex hull of the differ- wise regression method, which is frequently used, does not ence of two convex sets with specific geometric structures. always provide the best subset of variables according to these measures. In this talk, we introduce mixed integer second-order cone programming formulations for selecting Sina Modaresi the best subset of variables. MIT [email protected] Yuichi Takano Tokyo Institute of Technology Department of Industrial Engineering and Management Mustafa Kilinc [email protected] Department of Chemical Engineering Carnegie Mellon University [email protected] Ryuhei Miyashiro Institute of Symbiotic Science and Technology Tokyo University of Agriculture and Technology Juan Pablo Vielma [email protected] Sloan School of Management Massachusetts Institute of Technology [email protected] MS88 Improved Implementation of Positive Matrix Com- pletion Interior-Point Method for Semidefinite MS88 Programs A Lagrangian-Dnn Relaxation: a Fast Method for Computing Tight Lower Bounds for a Class of Exploiting sparsity is the key to solving SDPs in a short Quadratic Optimization Problems time. An important feature of the completion method in SDPA-C is that it decomposes the variable matrices to re- We propose an efficient computational method for linearly veal the structural sparsity. We implement a new decom- constrained quadratic optimization problems (QOPs) with position formula focusing the inverse of variable matrices. complementarity constraints based on their Lagrangian Numerical results show that this new formula and multiple- and doubly nonnegative (DNN) relaxation and first-order threaded parallel computing reduce the computation time algorithms. The simplified Lagrangian-CPP relaxation of for some SDPs that have structural sparsity. such QOPs proposed by Arima, Kim, and Kojima in 2012 takes one of the simplest forms, an unconstrained conic Makoto Yamashita linear optimization problem with a single Lagrangian pa- Tokyo Institute of Technology rameter in a completely positive (CPP) matrix variable Dept. of Mathematical and Computing Sciences with its upper-left element fixed to 1. Replacing the CPP [email protected] matrix variable by a DNN matrix variable, we derive the Lagrangian-DNN relaxation, and establish the equivalence Kazuhide Nakata between the optimal value of the DNN relaxation of the Tokyo Institute of Technology original QOP and that of the Lagrangian-DNN relaxation. Department of Industrial Engineering and Management We then propose an efficient numerical method for the [email protected] Lagrangian-DNN relaxation using a bisection method com- bined with the proximal alternating direction multiplier and the accelerated proximal gradient methods. Numeri- MS88 cal results on binary QOPs, quadratic multiple knapsack Dual Approach Based on Spectral Projected Gra- problems, maximum stable set problems, and quadratic as- dient Method for Log-Det Sdp with L1 Norm signment problems illustrate the superior performance of the proposed method for attaining tight lower bounds in An SDP problem combined with log-determinant and L1 shorter computational time. norm terms in its objective function has attracted more attention by a connection to the sparse covariance selec- Masakazu Kojima tion. Lu adapted an adaptive spectral projected gradient Chuo University method to this SDP. To apply this method to the SDP [email protected] with linear constraints, we focus its dual formulation and OP14 Abstracts 157

develop the dual method. Numerical results indicate that tope P to be the size of the smallest cone of psd matrices the dual method is more efficient than the primal method. that admits a lift of P. PSD minimal polytopes have PSD rank equal to dim(P )+1. We will discuss characterizations and properties of these polytopes. Makoto Yamashita Tokyo Institute of Technology Richard Robinson Dept. of Mathematical and Computing Sciences University of Washington [email protected] [email protected]

Mituhiro Fukuda Tokyo Institute of Technology MS89 Department of Mathematical and Computing Sciences Semidefinite Representations of the Convex Hull of [email protected] Rotation Matrices

Takashi Nakagaki The convex hull of SO(n), the set of n × n orthogonal ma- Yahoo Japan Corporation trices with determinant one, arises naturally when dealing 9-7-1 Akasaka, Midtown Tower, Minato-ku, Tokyo, Japan with optimization problems over rotations, e.g. in satellite [email protected] attitude estimation. In this talk we describe explicit con- structions showing that the convex hull of SO(n) is doubly spectrahedral, i.e. both it and its polar have a description MS89 as the intersection of a cone of positive semidefinite matri- A Criterion for Sums of Squares on the Hypercube ces with an affine subspace. This allows us to solve certain problems involving rotations using semidefinite program- Optimizing a polynomial over the {0, 1}-cube by sums of ming. squares is a standard problem in combinatorial optimiza- tion. We will present a criterion showing the limitations James Saunderson, Pablo A. Parrilo, Alan Willsky of this approach, even when a weaker version of the cer- Massachusetts Institute of Technology tificates is considered, improving results of Laurent on the [email protected], [email protected], [email protected] strength of Lassere’s hierarchy for maxcut. As a byprod- uct, we construct a family of globally nonnegative poly- nomials that need an increasing degree of sums of squares MS90 multipliers to be certified nonnegative. Iteration Complexity of Feasible Descent Methods for Convex Optimization Grigoriy Blekherman Georgia Tech Abstract Not Available [email protected] Chih-Jen Lin Joao Gouveia NTU Universidade de Coimbra [email protected] [email protected]

James Pfeiffer MS90 University of Washington Randomized Block Coordinate Non-Monotone jpfeiff@math.washington.edu Gradient Method for a Class of Nonlinear Pro- gramming

MS89 We propose a randomized block coordinate non-monotone Partial Facial Reduction: Simplified, Equivalent gradient (RBCNMG) method for minimizing the sum of SDPs via Inner Approximations of the PSD Cone a smooth (possibly nonconvex) function and a block- separable (possibly nonconvex nonsmooth) function. We We develop practical semidefinite programming (SDP) fa- show that the solution sequence generated by this method cial reduction procedures that utilize computationally ef- is arbitrarily close to an approximate stationary point with ficient inner approximations of the positive semidefinite high probability. When the problem under consideration cone. The proposed methods simplify SDPs with no is convex, we further establish that the sequence of ex- strictly feasible solution by solving a sequence of easier pected values generated converges to the optimal value of optimization problems. We relate our techniques to pars- the problem. ing algorithms for polynomial nonnegativity decision prob- lems, demonstrate their effectiveness on SDPs arising in Zhaosong Lu practice, and describe our publicly-available software im- Department of Mathematics plementation. Simon Fraser University [email protected] Frank Permenter, Pablo A. Parrilo Massachusetts Institute of Technology [email protected], [email protected] Lin Xiao Microsoft Research [email protected] MS89 Polytopes with Minimal Positive Semidefinite MS90 Rank Efficient Random Coordinate Descent Algorithms We define the positive semidefinite (PSD) rank of a poly- for Large-Scale Structured Nonconvex Optimiza- 158 OP14 Abstracts

tion Youssef M. Marzouk Massachusetts Institute of Technology We analyze several new methods for solving nonconvex op- [email protected] timization problems with the objective function formed as a sum of two terms: one is nonconvex and smooth, and an- other is convex but simple and its structure is known. Fur- MS91 ther, we consider both cases: unconstrained and linearly Optimal Design and Model Re-specification constrained nonconvex problems. For optimization prob- lems of the above structure, we propose random coordinate Improved characterization and assimilation of prior infor- descent algorithms and analyze their convergence proper- mation is essential in the context of large-scale inverse ties. For the general case, when the objective function is problems, yet, for realistic problems, the ability to do so nonconvex and composite we prove asymptotic convergence is rather limited. Complementary to such endeavors; it for the sequences generated by our algorithms to stationary is instrumental to attempt to maximize the extraction of points and sublinear rate of convergence in expectation for measureable information. This can be performed through some optimality measure. We also present extensive nu- improved prescription of experiments, or through improved merical experiments for evaluating the performance of our specification of the observation model. Conventionally, the algorithms in comparison with state-of-the-art methods. latter is achieved through first principles approaches, yet, in many situations, it is possible to learn a supplement for Ion Necoara the observation operator from the data. Such an approach University Politehnica Bucharest may be advantageous when the modeler is agnostic to the [email protected] principle sources of model-misspecification as well as when the development effort of revising the observation model Andrei Patrascu explicitly is not cost-effective. Bucharest [email protected] Lior Horesh Business Analytics and Mathematical Sciences IBM TJ Watson Research Center MS90 [email protected] Efficient Coordinate-minimization for Orthogonal Matrices through Givens Rotations Misha E. Kilmer Tufts University Optimizing over the orthogonal-matrices set is central [email protected] to many problems including eigenvalue problems, sparse- PCA, and tensor decomposition. Such optimization is hard Ning Hao since operations on orthogonal matrices easily break or- Tufts University thogonality, and reorthogonalization is usually costly. We Department of Mathematics propose a framework for orthogonal matrix optimization [email protected] that is parallel to coordinate-minimization in Euclidean spaces. It is based on Givens-rotations, fast-to-compute linear operations that modify few matrix entries and pre- MS91 serve orthogonality. We apply it to orthogonal tensor de- Approximation Methods for Robust Optimization composition and sparse-PCA. and Optimal Control Problems for Dynamic Sys- tems with Uncertainties Uri Shalit Hebrew University of Jerusalem All models contain systematic errors such as statistical un- [email protected] certainties of state and parameter estimates, model plant mismatch and discretization errors provided by simulation Gal Chechik methods. That is why application of optimization to real- Bar Ilan University life processes demands taking into account model and data [email protected] uncertainties. One of the possibilities is robust optimiza- tion which leads to problems with extremely high degree of computational complexity. The talk discusses efficient MS91 approximative robust optimization methods and their ap- NOWPAC - A Provably Convergent Derivative plication in optimal control problems. Free Optimization Algorithm for Nonlinear Pro- gramming Ekaterina Kostina Fachbereich Mathematik und Informatik We present the derivative free trust-region algorithm Philipps-Universitaet Marburg NOWPAC for computing local solutions of nonlinear con- [email protected] strained optimization problems. NOWPAC is designed to work solely with black box evaluations of the objective and the constraints, a setting which is of particular inter- MS91 est, for example, in robust optimization with chance con- Improved Bounds on Sample Size for Implicit Ma- straints. Constraints are handled using a path-augmented trix Trace Estimators inner boundary of the feasible domain, guaranteeing strict feasibility of all intermediate designs as well as convergence This talk is concerned with Monte-Carlo methods for the to a first order critical point. estimation of the trace of an implicitly given matrix A whose information is only available through matrix-vector Florian Augustin products. Such a method approximates the trace by an MIT average of N expressions of the form wt (Aw), with ran- [email protected] dom vectors w drawn from an appropriate distribution. We OP14 Abstracts 159

prove, discuss and experiment with bounds on the number Burgisser. of realizations N required in order to guarantee a proba- bilistic bound on the relative error of the trace estimation Javier Pena upon employing Rademacher (Hutchinson), Gaussian and Carnegie Mellon University uniform unit vector (with and without replacement) prob- [email protected] ability distributions. In total, one necessary and six suf- ficient bounds are proved, improving upon and extending Vera Roshchina similar estimates obtained in the seminal work of Avron University of Ballarat and Toledo (2011) in several dimensions. We first improve [email protected] their bound on N for the Hutchinson method, dropping a term that relates to rank(A) and making the bound com- parable with that for the Gaussian estimator. We further Negar S. Soheili prove new sufficient bounds for the Hutchinson, Gaussian Carnegie Mellon University and the unit vector estimators, as well as a necessary bound [email protected] for the Gaussian estimator, which depend more specifically on properties of the matrix A. As such they may suggest for what type of matrices one distribution or another provides MS92 a particularly effective or relatively ineffective stochastic On a Four-dimensional Cone that is Facially Ex- estimation method. posed but Not Nice Farbod Roosta-Khorasani Computer Science A closed convex cone is nice if the Minkowski sum of its UBC dual with the orthogonal complement of each of its faces [email protected] is a closed set. This property is useful in the analysis of a range of conic problems. We show that while all three- dimensional facially exposed cones are nice, there exists UriM.Ascher a four-dimensional cone that is facially exposed but is not University of British Columbia nice. We will also demonstrate some tricks that allow deal- Department of Computer Science ing with 4D objects using lower-dimensional argument and [email protected] 3D graphics.

MS92 Vera Roshchina University of Ballarat A Geometric Theory of Phase Transitions in Con- [email protected] vex Optimization

Convex regularization has become a popular approach to solve large scale inverse or data separation problems. A MS92 prominent example is the problem of identifying a sparse A Deterministic Rescaled Perceptron Algorithm signal from linear samples my minimizing the l1 norm un- der linear constraints. Recent empirical research indicates that many convex regularization problems on random data The classical perceptron algorithm is a separation-based exhibit a phase transition phenomenon: the probability algorithm for solving conic-convex feasibility problems in of successfully recovering a signal changes abruptly from a number of iterations that is bounded by the reciprocal zero to one as the number of constraints increases past of the square of the cone thickness. We propose a modi- a certain threshold. We present a rigorous analysis that fied perceptron algorithm that leverages periodic rescaling explains why phase transitions are ubiquitous in convex for exponentially faster convergence, where the iteration optimization. It also describes tools for making reliable bound is proportional to the logarithm of the reciprocal of predictions about the quantitative aspects of the transi- the cone thickness and another factor that is polynomial tion, including the location and the width of the transi- in the problem dimension. tion region. These techniques apply to regularized linear inverse problems, to demixing problems, and to cone pro- Negar S. Soheili,JavierPena grams with random affine constraints. These applications Carnegie Mellon University depend on a new summary parameter, the statistical di- [email protected], [email protected] mension of cones, that canonically extends the dimension of a linear subspace to the class of convex cones. Joint work with Dennis Amelunxen, Mike McCoy and Joel Tropp. MS93

Martin Lotz Online Algorithms for Ad Allocation School of Mathematics The University of Manchester As an important component of any ad serving system, on- [email protected] line capacity (or budget) planning is a central problem in online ad allocation. Various models are used to formally capture these problems and various algorithmic techniques MS92 such as primal-dual and sampling based algorithms have Preconditioners for Systems of Linear Inequalities been applied to these problems. The talk will survey the different models and techniques, theoretical guarantees and We show that a combination of two simple preprocess- practical evaluations of these algorithms on real data sets. ing steps improves the conditioning of a system of lin- ear inequalities. Our approach is based on a comparison among three different but related notions of conditioning Nikhil R. Devanur due to Renegar, Goffin-Cheung-Cucker, and Amelunxen- Microsoft Research, Redmond 160 OP14 Abstracts

[email protected] [email protected]

Shipra Agrawal MS93 Microsoft Distributed Learning on Dynamic Networks [email protected] In this talk, I will explore a data-driven online optimization framework for the synthesis of certain classes of networks Zizhuo Wang that are resistive to an external influence. The setup is in University of Minnesota, Twin Cities line with performance measures often adopted in control [email protected] theory, tailored for diffusion-type protocols. The approach provides means of embedding a distributed adaptive mech- MS94 anism on networks that modify the way the network col- lectively responds to an external input. Along the way, Linear Scalarizations for Vector Optimization we will explore connections between online distributed op- Problems with a Variable Ordering Structure timization, limits of influence on diffusion-type networks, spectral graph theory, electrical networks, and linear equa- In recent applications the preferences in the considered tion solving on graphs. multi-objective optimization problems turned out to be variable and to depend on the current function values. This Mehran Mesbahi is modeled by a vector optimization problem with an order- University of Washington ing structure defined by a cone-valued map. We discuss in [email protected] this talk how the well-known linear scalarizations based on elements from the dual cone can be generalized to variable orderings. We present characterization results for optimal MS93 elements including proper optimal elements. Online Algorithms for Node-Weighted Network Design Gabriele Eichfelder, Tobias Gerlach Institute of Mathematics In recent years, an online adaptation of the classical primal- Technische Universit¨at Ilmenau dual paradigm has been successfully used to obtain new on- [email protected], tobias.gerlach@tu- line algorithms for node-weighted network design problems, ilmenau.de and simplify existing ones for their edge-weighted counter- parts. In this talk, I will give an outline of this emerg- ing toolbox using three fundamental problems in this cate- MS94 gory for illustration: the Steiner tree (Naor-P.-Singh, 2011) Comparison of Different Scalarization Methods in and Steiner forest (Hajiaghayi-Liaghat- P., 2013) problems, Multi-objective Optimization and their prize-collecting versions (Hajiaghayi-Liaghat-P., 2014). This talk presents the comparison of different scalarization methods in multi-objective optimization. The methods are Debmalya Panigrahi compared with respect to the ability to consider the pref- Duke University erences of decision maker, the ability to generate differ- [email protected] ent kinds of efficient solutions such as weak efficient, effi- cient and proper efficient solutions. Theorems establishing relations between different scalarization methods are pre- MS93 sented. The computational skills and dependence on the A Dynamic Near-Optimal Algorithm for Online parameters sets for different scalarization methods are dis- Linear Programming cussed and demonstrative examples are presented.

A natural optimization model that formulates many online Refail Kasimbeyli resource allocation and revenue management problems is Anadolu University the online linear program (LP) where the constraint ma- [email protected] trix is revealed column by column along with the objective function. We provide a near-optimal algorithm for this Zehra Kamisli Ozturk surprisingly general class of online problems under the as- Anadolu Universiy sumption of random order of arrival and some mild con- [email protected] ditions on the size of the LP right-hand-side input. Our algorithm has a feature of ”learning while doing” by dy- namically updating a threshold price vector at geometric Nergiz Kasimbeyli, Gulcin Dinc Yalcin, Banu Icmen time intervals, where the dual prices learned from revealed Anadolu University columns in the previous period are used to determine the [email protected], [email protected], bic- sequential decisions in the current period. In particular, [email protected] our algorithm doesn’t assume any distribution information on the input itself, thus is robust to data uncertainty and variations due to its dynamic learning capability. Applica- MS94 tions of our algorithm include many online multi-resource Variational Analysis in Psychological Modeling allocation and multi-product revenue management prob- lems such as online routing and packing, online combina- This talk presents some mathematical models arising in torial auctions, adwords matching, inventory control and psychology and some other areas of behavioral sciences yield management. that are formalized via general variable preferences. In the mathematical framework, we derive a new extension of the Yinyu Ye Ekeland variational principle to the case of set-valued map- Stanford University pings with variable cone-valued ordering structures. Such OP14 Abstracts 161

a general setting provides an answer to the striking ques- [email protected] tion: in the world, where all things change what can stay fixed for a while? MS95 Bao Q. Truong On the Separation of Split Inequalities for Non- Mathematics and Computer Science. Department Convex Quadratic Integer Programming Northern Michigan University [email protected] We investigate the computational potential of split inequal- ities for non-convex quadratic integer programming, first Boris Mordukhovich introduced by Letchford and further examined by Burer Wayne State University and Letchford. These inequalities can be separated by solv- Department of Mathematics ing convex quadratic integer minimization problems. For [email protected] small instances with box-constraints, we show that the re- sulting dual bounds are very tight; they can close a large Antoine Soubeyran percentage of the gap left open by both the RLT- and Aix-Marseille University the SDP-relaxations of the problem. The gap can be fur- [email protected] ther decreased by separating so-called non-standard split inequalities, which we examine in the case of ternary vari- ables. MS94 Emiliano Traversi Robust Multiobjective Optimization LIPN Universite Paris 13 We consider multiobjective programs (MOPs) with uncer- [email protected] tainty in objective and constraint functions, and in the pa- rameters converting MOPs into single-objective programs (SOPs). We propose several types of robust counterpart MS95 problems (RCs) that involve infinitely many MOPs, in- finitely many objective functions in a single MOP, and Extended Formulations for Quadratic Mixed Inte- MOPs or SOPs with infinitely many constraints. In each ger Programming case, we reduce the RC into a computationally tractable deterministic MOP (or SOP) and examine the relationship An extended formulation for Mixed Integer Programming between their efficient sets. (MIP) is a formulation that uses a number of auxiliary vari- ables in addition to the original or natural variables of a Margaret Wiecek MIP. Extended formulations for linear MIP have been ex- Mathematical Sciences, Clemson University tensively used to construct small, but strong formulations [email protected] for a wide range of problems. In this talk we consider the use of extended formulations in quadratic MIP and show how they can be used to improve the strength cutting plane MS95 procedures. Finding Low-Rank Solutions to Qcqps Juan Pablo Vielma Sloan School of Management We describe continuing work to find low-rank solutions Massachusetts Institute of Technology to QCQPs, especially as they arise in optimal power flow [email protected] (OPF) problems.

Daniel Bienstock MS96 Columbia University IEOR and APAM Departments IEOR Department Numerical Methods for Stochastic Multistage Op- [email protected] timization Problems Applied to the European Elec- tricity Market

MS95 Abstract Not Available Global Optimization with Non-Convex Quadratics Nicolas Grebille Ecole Polytechnique We present results on a new spatial B&B solver, iquad, for France optimization problems in which all of the non-convexity [email protected] is quadratic. Our approach emphasizes pre-processing to effectively exploit any convexity. MS96 Marcia HC Fampa Risk in Capacity Energy Equilibria UFRJ [email protected] We look at a risky investment equilibrium that combines a game of investments by risk averse firms in electricity Jon Lee plants, a risk market in financial products, and a stochas- IBM T.J. Watson Research Center tic electricity market. In a complete risk market, all risks [email protected] can be priced and the equilibrium model collapses, eg, to convex optimization problem when the energy market is Wendel Melo perfectly competitive. We also look at equilibria in incom- UFRJ plete markets which are computationally and theoretically 162 OP14 Abstracts

more challenging. (sparse optimization) and matrix recovery (matrix com- pletion and robust principle component analysis). While Daniel Ralph very general in principle, the efficacy of the approach relies University of Cambridge on efficient first order solvers. For matrix completion and Dept. of Engineering & Judge Institute of Management robust PCA, we incorporate new ideas in factorized and [email protected] accelerated first order methods into the variational frame- work to develop state of the art approaches for large scale Andreas Ehrenmann, Gauthier de Maere applications, and illustrate with both synthetic and real GDF-SUEZ examples. [email protected], [email protected] Aleksandr Aravkin Yves Smeers IBM T.J. Watson Research Center Universit´e catholique de Louvain [email protected] CORE [email protected] MS97 -Subgradient Methods for Bilevel Convex Opti- MS96 mization Computing Closed Loop Equilibria of Electricity Capacity Expansion Models Existing techniques used in the solution of the convex bilevel optimization problem include Cabot’s idea of us- We propose a methodology to compute closed loop capac- ing the proximal method applied to a varying objective ity equilibria using only open loop capacity equilibrium function and Solodov’s bundle variation of the same tech- models and apply this approximation scheme to large-scale nique. Continuing this trend we introduce an -subgradient generation expansion planning models in liberalized elec- method for the bilevel problem. Our algorithm presents tricity markets. This approximation scheme allows us to two main advantages over current technique: cheap itera- solve the closed loop model reasonably well by smartly em- tion cost and ease of implementation. The main drawback ploying open loop models which reduces the computational is slow asymptotic convergence rate. time by two orders of magnitude. These results are con- firmed in a large-scale, multi-year, multi-load period and Elias Salomo S. Helou Neto multi-technology numerical example. Instituto de Cincias Matem´aticas e de Computao Universidade de So Paulo Sonja Wogrin [email protected] Universidad de Comillas Spain [email protected] MS97 An Algorithm for Convex Feasibility Problems Us- MS96 ing Supporting Halfspaces and Dual Quadratic Optimization (and Maybe Game Theory) of De- Programming mand Response A basic idea of finding a point in the intersection of convex In this talk we will briefly outline the market clearing mech- sets is to project iterates onto these convex sets to obtain anism for the NZ electricity market (NZEM) which co- halfspaces containing the intersection, and then to project optimizes energy and reserve simultaneously. We will then the iterates onto the polyhedron generated. We discuss present a tool designed for the demand response of major theoretical issues and implementation strategies. users of electricity. This tool embeds the change to the distribution of future prices as a function of the actions of C.H. Jeffrey Pang the major electricity users and optimizes the consumption Mathematics, National University of Singapore level as well as the reserve offer for a major user of elec- [email protected] tricity. We will also discuss extensions of this to schedules for longer time horizons. MS97 Golbon Zakeri A First Order Algorithm for a Class of Nonconvex- Auckland University nonsmooth Minimization New Zealand [email protected] We introduce a new first order algorithm for solving a broad class of nonconvex and nonsmooth problems. The re- Geoff Pritchard sulting scheme involves simple iterations and is particularly University of Auckland adequate for solving large scale nonsmooth and nonconvex [email protected] problems arising in a wide variety of fundamental appli- cations. We outline a self contained convergence analysis framework describing the main tools and methodology to MS97 prove that the sequence generated by the proposed scheme Fast Variational Methods for Matrix Completion globally converges to a critical point. Our results are illus- and Robust PCA trated on challenging sparse optimization models arising in data analysis applications. This is a joint work with We present a general variational framework that allows effi- Jerome Bolte and Marc Teboulle. cient algorithms for denoising problems (where a functional is minimized subject to a constraint on fitting the observed Shoham Sabach data). This framework is useful for both vector recovery Technion - Israel institute of Technology OP14 Abstracts 163

[email protected] Newton-CG augmented Lagrangian method by Zhao, Sun and Toh is suitable for solving these semidefinite relax- ations. Extensive numerical experiments are presented to MS98 show that this approach is practical in getting best rank-1 On Symmetric Rank and Approximation of Sym- approximations. metric Tensors JiawangNie,LiWang In this talk we will discuss two problems: The first one is University of California, San Diego when the rank of real symmetric tensor is equal to its sym- [email protected], [email protected] metric rank. The second problem is when a best k-border rank approximation can be chosen symmetric. Some par- tial results are given in a joint paper with M. Stawiska: MS99 arXiv:1311.1561. An Inexact Sequential Quadratic Optimization Method for Nonlinear Optimization Shmuel Friedland Department of Mathematics We present a sequential quadratic optimization algorithm University of Illinois, Chicago for solving nonlinear constrained optimization problems. [email protected] The unique feature of the algorithm is that inexactness is allowed when solving the arising quadratic subproblems. Easily implementable inexactness criteria for the subprob- MS98 lem solver are established that ensure global convergence Some Extensions of Theorems of the Alternative guarantees for the overall algorithm. This work represents a step toward a scalable active-set method for large-scale In this talk, some extensions of theorems of the alterna- nonlinear optimization. tive to the tensor setting will be discussed. We will talk about the difficulties and impossibility of the general case. Frank E. Curtis Nevertheless, there are classes of structured tensors which Industrial and Systems Engineering have suitable generalisations. Specially, we will present a Lehigh University tractable extension of Yuan’s theorem of the alternative [email protected] with application to a class of nonconvex polynomial opti- misation problems. The optimal solution of this class prob- lems can be recovered from the corresponding solution of Travis Johnson the convex conic programming problem. University of Washington [email protected] Shenglong Hu The Hong Kong Polytechnic University Daniel Robinson [email protected] Northwestern University [email protected]

MS98 Andreas Waechter Algorithms for Decomposition Northwestern University IEMS Department We present an algorithm for decomposing a symmetric [email protected] tensor, of dimension n and order d,asasumofrank-1 symmetric tensors, based on an extension of Sylvester’s algorithm for binary forms. We also present an efficient MS99 variant for the case where the factor matrices enjoy some A BFGS-Based SQP Method for Constrained Non- structure, such as block-Hankel, triangular, band, etc. Fi- smooth, Nonconvex Optimization nally, we sketch the bit complexity analysis for the case of binary forms. The talk is based on common works with We consider constrained optimization problems where both J. Brachat, P. Comon, P. Hub´a˘cek, B. Mourrain, V. Pastro, the objective and constraints may be nonsmooth and non- S. K. Stiil Frederiksen, and M. Sorensen. convex. In 2012, Curtis and Overton presented a gradient- Elias Tsigaridas sampling-based SQP method, proving convergence results, Inria Paris-Rocquencourt but in the unconstrained nonsmooth case, Lewis and Over- [email protected] ton contrastingly argue that BFGS is a far more efficient approach than gradient-sampling. We extend BFGS using SQP to dynamically adapt to nonsmooth objectives and MS98 constraint violations and show promising results on chal- lenging applied problems from feedback control. Semidefinite Relaxations for Best Rank-1 Tensor Approximations Frank E. Curtis We study the problem of finding best rank-1 approxima- Industrial and Systems Engineering tions for both symmetric and nonsymmetric tensors. For Lehigh University symmetric tensors, this is equivalent to optimizing homo- [email protected] geneous polynomials over unit spheres; for nonsymmetric tensors, this is equivalent to optimizing multi-quadratic Tim Mitchell forms over multi-spheres. We propose semidefinite relax- Courant Institute of Mathematical Sciences ations, based on sum of squares representations, to solve New York University these polynomial optimization problems. Their properties [email protected] and structures are studied. In applications, the resulting semidefinite programs are often large scale. The recent MichaelL.Overton 164 OP14 Abstracts

New York University University of Electro-Communications Courant Instit. of Math. Sci. [email protected] [email protected] Takashi Tsuchiya National Graduate Institute for Policy Studies MS99 [email protected] A Two-phase SQO Method for Solving Sequences of NLPs MS100 SQO methods are fundamental tools while solving more Gauge optimization, duality and applications complex problems. Usually the problem one wants to solve adopts the form of a sequence of closely related nonlinear Gauge optimization seeks the element of a convex set that programming problems. Therefore, warm starts an inexact is minimal with respect to a gauge function. It can be solutions, among other desirable features, play an impor- used to model a large class of useful problems, including tant role in the overall performance of the final method. a special case of conic optimization, and various problems In this talk we explore the potential of SQP+ for solving that arise in machine learning and signal processing. In sequences of NLPs. SQP+ is a two phase method that this talk, we explore the duality framework proposed by combines an inequality QP step for detecting the active Freund, and discuss a particular form of the problem that set, followed by an equality QP phase that promotes fast exposes some useful properties of the gauge optimization convergence. framework.

Jose Luis Morales MichaelP.Friedlander Departamento de Matematicas. Dept of Computer Science ITAM. MEXICO. University of British Columbia [email protected] [email protected]

Macedo Ives MS99 Dept of Computer Science An Inexact Trust-Region Algorithm for Nonlinear [email protected] Programming Problems with Expensive General Constraints Ting Kei Pong Department of Mathematics This talk presents an inexact SQP algorithm for nonlinear University of Washington optimization problems with equality and inequality con- [email protected] straints. The proposed method does not require the ex- act evaluation of the constraint Jacobian. Instead, only approximations of these Jacobians are required. Corre- MS100 sponding accuracy requirements for the presented first- order global convergence result can be verified easily during On the rolls of optimality conditions for polynomial the optimization process to adjust the approximation qual- optimization ity of the constraint Jacobian. First numerical results for A polynomial optimization problem is a problem to find this new approach are discussed. a minimum of a polynomial function over the set de- fined by polynomial equations and inequalities. There is Andrea Walther an algorithm which obtains the global minimum,bysolv- Universit¨at Paderborn ing a sequence of semidefinite programming. We study [email protected] how optimality conditions, or other preferable properties of polynomial optimization problems affect the generated Larry Biegler semidefinite programming. This involves attainability of Carnegie Mellon University semidefinite programming, finite convergence properties of [email protected] Lasserre’s hierarchy and representability of nonnegative polynomials.

MS100 Yoshiyuki Sekiguchi A Geometrical Analysis of Weak Infeasibility in The University of Marine Science and Technology Semidefinite Programming and Related Issues [email protected]

In this talk, we present a geometrical analysis of Semidefi- nite Feasibility Problems (SDFPs). We show how to di- MS100 vide an SDFP into smaller subproblems in a way that An LP-based Algorithm to Test Copositivity the feasibility properties are mostly preserved. This is ac- complished by introducing an object called “hyper feasible A symmetric matrix is called copositive if it generates a partition”. We use our techniques to study weakly infea- quadratic form taking no negative values over the posi- sible problems in a systematic way and to provide some tive orthant, and the linear optimization problem over the insight into how weakly infeasibility arises in semidefinite set of copositive matrices is called the copositive program- programming. ming problem. Recently, many studies have been done on the copositive programming problem which (cf, Dur Bruno F. Louren¸co (2010)). Among others, several blanch and bound type Tokyo Institute of Technology algorithms have been provided to test copositivity since it fl[email protected] is known that the problem for deciding whether a given ma- trix is copositive is co-NP-complete (cf. Murty and Kabadi Masakazu Muramatsu (1987)). In this talk, we propose a new blanch and bound OP14 Abstracts 165

type algorithm for this testing problem. Our algorithm [email protected] is based on solving linear optimization problems over the nonnegative orthant, repeatedly. Numerical experiments suggest that our algorithm is promising for determining MS101 upper bounds of the maximum clique problem. Recent Devlopments in Picos

PICOS is a user-friendly modelling language written in Akihiro Tanaka python which interfaces several conic and integer program- University of Tsukuba ming solvers, similarly to YALMIP or CVX under MAT- [email protected] LAB. In this talk we will present some recent developments of PICOS, in particular concerning complex semidefi- Akiko Yoshise nite programming (with Hermitian matrices) and ro- University of Tsukuba bust optimization. Graduate School of Systems and Information Engineering [email protected] Guillaume Sagnol ZIB, Berlin [email protected] MS101 Building Large-scale Conic Optimization Models MS102 using MOSEK Fusion Coordinate Descent Type Methods for Solving Sparsity Constrained Problems A modern trend in convex optimization is to pose convex problem in conic form, which is a surprisingly versatile and We consider the problem of minimizing a smooth function expressive description. MOSEK Fusion is an object orien- over the intersection of a closed convex set and a spar- tated tool for building conic optimization models. In this sity constraint. We begin by showing how the orthogonal talk we present Fusion with an emphasis on techniques that projection operator can be computed efficiently when the can be used to vectorize the model construction. When underlying convex set has certain symmetry properties. A applicable such vectorization techniques can speed up the hierarchy of optimality conditions for the problem is pre- model construction by more than one order of magnitude. sented, and the theoretical superiority of coordinate-wise type conditions is established. Finally, based on the de- rived optimality conditions, we present several coordinate descent type algorithms for solving the sparsity constrained Erling D. Andersen problem, and illustrate their performance on several nu- MOSEK ApS merical examples. This is joint work with Nadav Hallak. Copenhagen, Denmark [email protected] Amir Beck TECHNION - Israel Institute of technology [email protected] MS101

Quadratic Cone Modeling Language MS102 GAMSEL: A Penalized Likelihood Approach to The quadratic cone modeling language (QCML) is a Model Selection for Generalized Additive Models lightweight convex optimization modeling language imple- mented in Python. It is a code generator that produces We introduce GAMSEL (Generalized Additive Model SE- lightweight Python, Matlab, or C code for matrix stuffing. Lection), a method for fitting generalized additive models This resulting code can then be called to solve a particular in high dimensions. Our method borrows ideas from the problem instance, in which the parameters and dimensions overlap group lasso of Jacob et al (2009) to retain, when of the original description are fixed to given values. The possible, the interpretability advantages of a simple linear resulting cone program is solved with an external solver. fit while allowing a more flexible additive fit when neces- sitated by the data. We present a blockwise coordinate Eric Chu descent procedure for optimizing the penalized likelihood Stanford University objective. [email protected] Alexandra Chouldechova Standord [email protected] MS101 The Cvx Modeling Framework: New Development Trevor Hastie Professor, Dept of Statistics; Health Research and Policy Stanford University CVX is a well known modeling framework for convex op- [email protected] timization. Initially targeting just MATLAB, CVX is now available on Octave, and work is actively ongoing in bring- ing CVX to other computational platforms as well. In this MS102 talk, we will provide an overview of CVX and provide an Coordinate descent methods for 0-regularized op- update on the status of these recent developments. timization problems MichaelC.Grant The recent interest in sparse optimization represents a CVX Research strong motivation for deriving complexity results for 0- California Institute of Technology regularized problems. We present new results regarding 166 OP14 Abstracts

the classification and quality of local optimal points for [email protected] 0-regularized problems. In order to provide efficient algo- rithms that converge to particular classes of optimal points, we combine coordinate descent framework with iterative MS103 hard thresholding methods. For all the new algorithms we New Trends in Aerodynamic Shape Optimization provide estimates on the convergence rate that are superior using the Continuous Adjoint Method to the existing results. In this presentation, the authors will review their research Andrei Patrascu on high-fidelity Computational Fluid Dynamics (CFD) Bucharest tools and continuous adjoint-based techniques for the op- [email protected] timal aerodynamic shape design. While there is a well- consolidated theory for the application of adjoint methods Ion Necoara to optimal shape design, the use of these techniques in re- University Politehnica Bucharest alistic industrial problems is challenging due to a range of [email protected] technical and mathematical issues that will be discussed in this presentation. In particular, different examples (rotor- craft, free-surface, or supersonic) will be used to introduce MS102 original contributions to the state-of-the-art in CFD-based Efficient Randomized Coordinate Descent for optimal shape design: a systematic approach to shape de- Large Scale Sparse Optimization sign, ALE formulation using continuous adjoint, design with discontinuities, and robust grid adaptation, among Recently several methods were proposed for sparse opti- other topics. mization which make careful use of second-order informa- tion to improve local convergence rates. These methods Francisco Palacios construct a composite quadratic approximation using Hes- Stanford University sian information, optimize this approximation using a first- [email protected] order method, such as coordinate descent and employ a line search to ensure sufficient descent. Here we propose a gen- Thomas Economon, Juan J. Alonso eral method, which improves upon these ideas to improve Department of Aeronautics and Astronautics the practical performance and prove a global convergence Stanford University analysis in the spirit of proximal gradient methods. mailto:[email protected], [email protected] Xiaocheng Tang Lehigh MS103 [email protected] Shape Analysis for Maxwell’s Equations

Katya Scheinberg In the context of the project HPC-FLiS, we formulate the Lehigh University problem of identifying unknown objects in predefined do- [email protected] mains as an inverse electromagnetic scattering problem. Here, we consider it as a shape optimization problem. A MS103 typical design cycle consists of a forward simulation, ad- joint solution, shape gradient calculation and shape&grid Efficient Estimation of Selecting and Locating Ac- modification. This talk will cover the theoretical derivation tuators and Sensors for Controlling Compressible, of the shape derivatives, the realization using the tool FEn- Viscous Flows iCS and results obtained for various test configurations. A new method is developed to estimate optimal actuator Maria Sch¨utte types and locations for controlling compressible, viscous Universit¨at Paderborn flows using linear feedback. Based on an analysis of the [email protected] structural sensitivity of the linearized compressible Navier- Stokes operator about a time-steady baseflow, the forward and adjoint global modes are used to optimize eigenvalue Stephan Schmidt placement which leads to an estimate of where the con- Imperial College London troller should be placed, and what type of controller (mass, [email protected] momentum, energy, etc.) it should be. The method is demonstrated using direct numerical simulations of a sep- Andrea Walther arated boundary layer in a Mach 0.65 diffuser at different Universit¨at Paderborn Reynolds numbers and whether the baseflow is taken as [email protected] the true steady solution or the time-averaged flow. For sufficiently low Reynolds numbers global stabilization of the flow is achieved; only partial stabilization is achieved MS103 at higher Reynolds numbers. Preliminary results for con- Optimization in Chaos trolling supersonic jet noise will also be given. Turbulent flow is chaotic. So can be flow-structure coupled Daniel J. Bodony oscillations. High fidelity simulations of both systems of- University of Illinois at Urbana-Champaign ten inherit their chaotic dynamics, which brings fascinating [email protected] mathematical questions into their optimization. What is butterfly effect? How does it affect gradient computation? Mahesh Natarajan What is sampling error? How does it affect gradient-based University of Illinois at Urbana-Champaign and gradient-free optimization? Can ideas in stochastic Department of Aerospace Engineering programming apply to optimization in chaos? What if our OP14 Abstracts 167

problem is both chaotic and uncertain? optimization problems. We also discuss best- and worst- case run-time bounds for this algorithm. Qiqi Wang Massachusetts Institute of Technology Boris Houska [email protected] Shanghai Jiao Tong University [email protected] Patrick Blonigan MIT Benoit Chachuat [email protected] Imperial College London [email protected]

MS104 Equitable Division of a Geographic Region MS104 The Slater Conundrum: Duality and Pricing in In- In many practical problems, a geographic region must be finite Dimensional Optimization subdivided into smaller regions in a fair or balanced fash- ion. Partitioning a region is generally a difficult problem In convex optimization, the Slater constraint qualification because the underlying optimization variables are infinite- is perhaps the most well known interior point condition dimensional and because shape constraints may be difficult that ensures a zero duality gap. Using an algebraic ap- to impose from within a standard optimization context. proach, we examine the infinite dimensional vector spaces Here we present several natural formulations of region par- that admit Slater points and show that the resulting dual titioning problems and discuss how to solve them, with variable space contains singular functionals. These dual applications to vehicle routing and facility location. variables lack the standard pricing interpretation desired for many modeling applications. We then present sufficient John Gunnar Carlsson conditions that ensure these functionals are not optimal so- University of Minnesota lutions to the dual program. [email protected] Matt Stern University of Chicago MS104 Booth School of Business Structured Convex Infinite-dimensional Optimiza- [email protected] tion with Double Smoothing and Chebfun Chris Ryan, Kipp Martin We propose an efficient approach to minimize a convex University of Chicago function over a simple infinite-dimensional convex set un- A ∈ A · [email protected], der an additional constraint (x) T where ( ) is linear [email protected] and T finite-dimensional. We apply a fast gradient method to a doubly smoothed dualized problem, while relying on the chebfun toolbox to perform numerical computations MS105 on the primal side. We present results for optimal control Analysis and Design of Fast Graph Based Algo- problems where the trajectory is forced to visit some sets rithms for High Dimensional Data at certain moments in time. Geometric methods have revolutionized the field of image Olivier Devolder processing and image analysis. Recently some of these con- Center for Operations Research and Econometrics cepts have shown promise for problems in high dimensional Universit´e catholique de Louvain data analysis and machine learning on graphs. I will briefly [email protected] review the methods from imaging and then focus on the new methods for high dimensional data and network data. Francois Glineur These ideas include diffuse interface methods and threshold Universit´e Catholique de Louvain (UCL) dynamics for total variation minimization. Center for Operations Research and Econometrics (CORE) Andrea L. Bertozzi [email protected] UCLA Department of Mathematics [email protected] Yurii Nesterov CORE Universite catholique de Louvain MS105 [email protected] Fast Algorithms for Multichannel Compressed Sensing with Forest Sparsity

MS104 In this paper, we propose a new model called forest spar- Global Optimization in Infinite Dimensional sity for multichannel compressive sensing. It is an exten- Hilbert Spaces sion of standard sparsity when the support set of the data is consisted of a series of mutually correlated trees. For- This talk is about complete search methods for solving est sparsity exists in many practical applications such as non-convex infinite-dimensional optimization problems to multi-contrast MRI, parallel MRI, multispectral image and global optimality. The optimization variables are assumed color image recovery. We theoretically prove its benefit, to be bounded and in a given Hilbert space. We present that much less measurements are required for successful an algorithm and analyze its convergence under certain recovery in compressive sensing. Moreover, efficient algo- regularity conditions for the objective and constraint func- rithms are proposed and applied on several applications tions, which hold for a large class of practically-relevant with forest sparsity. All experimental results validate the 168 OP14 Abstracts

superiority of forest sparsity. [email protected]

Junzhou Huang Department of Computer Science and Engineering MS106 University of Texas at Arlington Optimal Control of Free Boundary Problems with [email protected] Surface Tension Effects

We will give a novel proof for the second order order suffi- MS105 cient conditions for an optimal control problem where the GRock: Greedy Coordinate-Block Descent state system is composed of Laplace equation in the bulk Method for Sparse Optimization and Young-Laplace on the free boundary. We will use piecewise linear finite elements to discretize the control Modern datasets usually have a large number of features or problem and prove the optimal a priori error estimates. training samples, and stored in a distributed manner. Mo- Finally, we will discuss a novel analysis for an extrusion tivated by the need of solving sparse optimization problems process where the bulk equations are Stokes equations. with large datasets, we propose GRock, a parallel greedy coordinate-block descent method. We also establish the Harbir Antil asymptotic linear convergence of GRock and explain why George Mason University it often performs exceptionally well for sparse optimization. Fairfax, VA Numerical results on a computer cluster and Amazon EC2 [email protected] demonstrate the efficiency and elasticity of our algorithms. MS106 Zhimin Peng Fast and Sparse Noise Learning via Nonsmooth University of California. Los Angeles PDE-constrained Optimization Department of Mathematics We consider a nonlinear PDE constrained optimization [email protected] approach to learn the optimal weights for a generic TV- denoising model featuring different noise distributions pos- Ming Yan sibly present in the data. To overcome the high compu- University of California, Los Angeles tational costs needed to compute the numerical solution, Department of Mathematics we use dynamical sampling schemes. We will consider also [email protected] spatially dependent weights combined with a sparse regu- larisation on the parameter vector. This is joint work with Wotao Yin J. C. De Los Reyes and C.-B. Schoenlieb. Department of Mathematics University of California at Los Angeles Luca Calatroni [email protected] CambridgeCentreforAnalysis University of Cambridge [email protected] MS105 Adaptive BOSVS Algorithm for Ill-Conditioned MS106 Linear Inversion with Application to Partially Par- allel Imaging Multibang Control of Elliptic Equations Multi-bang control refers to pde-constrained optimization This paper proposes a new adaptive Bregman opera- problems where a distributed control should only take on tor splitting algorithm with variable stepsize (Adaptive values from a given discrete set, which can be promoted by BOSVS) for solving large scale and ill-conditioned inverse 2 0 a combination of L and L -type control costs. Although problems that arise in partially parallel magnetic resonance the resulting functional is nonconvex and lacks weak lower- imaging. The original BOSVS algorithm uses a line search semicontinuity, application of Fenchel duality yields a for- to achieve efficiency, while a proximal parameter is ad- mal primal-dual optimality system that admits a unique justed to ensure global convergence whenever a monotonic- solution and can be solved numerically via a regularized ity condition is violated. The new Adaptive BOSVS uses semismooth Newton method. a simpler line search than that used by BOSVS, and the monotonicity test can be skipped. Numerical experiments Christian Clason based on partially parallel image reconstruction compare University of Graz the performance of BOSVS and Adaptive BOSVS scheme. Institute for Mathematics and Scientific Computing [email protected] Maryam Yashtini Department of Mathematics Karl Kunisch University of Florida Karl-Franzens University Graz myashtini@ufl.edu Institute of Mathematics and Scientific Computing [email protected] William Hager University of Florida MS106 hager@ufl.edu A Primal-dual Active Set Algorithm for a Class of Nonconvex Sparsity Optimization Hongchao Zhang Department of Mathematics and CCT We develop an algorithm of primal-dual active set type Louisiana State University for a class of nonconvex sparsity-promoting penalties. OP14 Abstracts 169

We derive a novel necessary optimality condition for the mization global minimizer using the associated thresholding opera- tor. The solutions to the optimality system are necessarily Abstract Not Available coordinate-wise minimizers, and under minor conditions, they are also local minimizers. Upon introducing the dual Ben Recht variable, the active set can be determined from the pri- UC Berkeley mal and dual variables. This relation lends itself to an [email protected] iterative algorithm of active set type which at each step in- volves updating the primal variable only on the active set and then updating the dual variable explicitly. Numerical MS107 experiments demonstrate its efficiency and accuracy. Multi-Stage Convex Relaxation Approach for Low- Rank Structured Psd Optimization Problems Yuling Jiao This work is concerned with low-rank structured positive Wuhan University semidefinite (PSD) matrix optimization problems. We [email protected] start from reformulating them as mathematical programs with PSD equilibrium constraints and their exact penalty Bangti Jin forms. Then, we propose a multi-stage convex relax- Department of Mathematics ation approach that solves at each iteration a weighted University of California, Riverside trace semi-norm minimization problem. Under conditions [email protected] weaker than the RIP, we establish an error bound for the optimal solution of the kth sub-problem, and show that the Xiliang Lu error bound in the second stage is strictly less than that of Wuhan University the first stage with the reduction rate explicitly quantifi- [email protected] able. Numerical results are provided to verify the efficiency of the proposed approach. [Joint work with Shujun Bi and Shaohua Pan] MS107 Defeng Sun Conic Geometric Programming Dept of Mathematics National University of Singapore We introduce conic geometric programs (CGPs), which [email protected] are convex optimization problems that unify geometric programs (GPs) and conic optimization problems such as semidefinite programs (SDPs). Computing global optima MS108 of CGPs is not much harder than solving GPs and SDPs. A Strong Relaxation for the Alternating Current The CGP framework facilitates a range of new applications Optimal Power Flow (acopf ) Problem that fall outside the scope of SDPs and GPs: permanent maximization, hitting-time estimation in dynamical sys- ACOPF is an electric generation dispatch problem that is tems, computation of quantum capacity, and robust opti- nonlinear and nonconvex. Fortunately, its Lagrangian dual mization formulations of GPs. can be solved with Semidefinite Programming; the problem can thus be solved in polynomial time when zero duality Venkat Chandrasekaran gap occurs. Establishing global optimality with duality California Institute of Technology gap can be problematic; in this case we strengthen the dual [email protected] relaxation by adding valid inequalities based on a convex envelope. Our relaxation motivates a globally convergent Parikshit Shah algorithm that exploits sparse network topology. University of Wisconsin/Philips Research [email protected] Chen Chen Berkeley [email protected] MS107 Alper Atamturk Sketching the Data Matrix of Large Linear and U.C. Berkeley Convex Quadratic Programs [email protected]

Abstract Not Available Shmuel Oren University of California, Berkeley Laurent El Ghaoui Industrial Engineering and Operations Research EECS, U.C. Berkeley [email protected] [email protected]

Riadh Zorgati MS108 EDF R&D Islanding of Power Networks Department OSIRIS [email protected] In the past there have been multiple high-profile cases of cascading blackouts of power systems. In this talk we will present a MINLP model for optimal islanding of a power MS107 system network in order to prevent an imminent cascading blackout. We give a tractable reformulation of the orig- Convex Programming for Continuous Sparse Opti- inal problem based on a new linear approximation of the 170 OP14 Abstracts

transmission constraints. We give computational results on [email protected] networks of various sizes and show that our method scales well. MS109 Andreas Grothey Some Variational Properties of Regularization School of Mathematics Value Functions University of Edinburgh [email protected] Regularization plays a key role in a variety of optimiza- tion formulations of inverse problems. A recurring theme Ken McKinnon in regularization approaches is the selection of regulariza- University of Edinburgh tion parameters, and their effect on the solution and on [email protected] the optimal value of the optimization problem. The sensi- tivity of the value function to the regularization parameter Paul Trodden can be linked directly to the Lagrange multipliers. In this University of Sheffield talk we show how to characterize the variational properties p.trodden@sheffield.ac.uk of the value functions for a broad class of convex formu- lations. An inverse function theorem is given that links the value functions of different regularization formulations Waqquas Bukhsh (not necessarily convex). These results have implications University of Edinburgh for the selection of regularization parameters, and the de- [email protected] velopment of specialized algorithms. Numerical examples illustrate the theoretical results.

MS108 James V. Burke Real-Time Pricing in Smart Grid: An ADP Ap- University of Washington proach [email protected]

To realize real-time pricing (RTP) of electricity, demand- Aleksandr Aravkin level control automation devices are essential. Aided by IBM T.J. Watson Research Center such devices, we propose an approximate dynamic pro- [email protected] gramming (ADP)-based modeling and algorithm frame- work to integrate wholesale markets dispatch operation MichaelP.Friedlander with demand response under RTP. With controllable Dept of Computer Science charging/discharging of plug-in electric vehicles, our nu- University of British Columbia merical results show that RTP can lower the expected val- [email protected] ues of wholesale electricity prices, increase capacity factors of renewable resources, and reduce system-wide CO2 emis- sions. MS109 Andrew L. Liu Optimization and Intrinsic Geometry Purdue Modern optimization - both in its driving theory and in [email protected] its classical and contemporary algorithms - is illuminated by geometry. I will present two case studies of this ap- Jingjie Xiao proach. The first example - seeking a common point of Reddwerks Corporation two convex sets by alternately projecting onto each - is [email protected] classical and intuitive, and widely applied (even without convexity) in applications like image processing and low- order control. I will sketch some nonconvex cases, and MS108 relate the algorithm’s convergence to the intrinsic geome- Topology Control for Economic Dispatch and Re- try of the two sets. The second case study revolves around liability in Electric Power Systems “partly smooth manifolds” - a geometric generalization of the active set notion fundamental to classical Nonlinear Transmission topology control through line switching has optimization. I emphasize examples from eigenvalue opti- been recognized as a valuable mechanism that can produce mization. Partly smooth geometry opens the door to ac- economic and reliability benefits through reduced conges- celeration strategies for first-order methods, and is central tion cost and corrective actions in response to contingen- to sensitivity analysis. Reassuringly, given its power as a cies. Such benefits are due to the inherent physical laws, tool, this geometry is present generically in semi-algebraic known as Kirchhoffs Laws, governing the flow of power in optimization problems. electricity networks. This paper provides an overview de- scribing the underlying motivation, current industry prac- Dmitriy Drusvyatskiy tices and the computational challenges associated with University of Waterloo topology control in power systems. [email protected]

Shmuel Oren Adrian Lewis University of California, Berkeley Cornell University Industrial Engineering and Operations Research Oper. Res. and Indust. Eng. [email protected] [email protected]

Kory Hedman Alexander Ioffe Arizona State University Technion - Israel Institute of Technology OP14 Abstracts 171

Department of Mathematics [email protected] ioff[email protected]

Aris Daniildis MS110 Universitsy of Barcelona [email protected] On Generic Nonexistence of the Schmidt-Eckart- Young Decomposition for Tensors

MS109 The Schmidt-Eckart-Young theorem for matrices states Epi-Convergent Smoothing with Applications to that the optimal rank-r approximation to a matrix in the Convex Composite Functions Euclidean topology is obtained by retaining the first r terms from the singular value decomposition of that ma- In this talk we synthesize and extend recent results due to trix. In this talk, we consider a generalization of this opti- Beck and Teboulle on infimal convolution smoothing for mal truncation property to the CANDECOMP/PARAFAC convex functions with those of X. Chen on gradient con- decomposition of tensors and establish a necessary orthog- sistency for nonconvex functions. We use epi-convergence onality condition. We prove that this condition is not sat- techniques to define a notion of epi-smoothing that allows isfied at least by an open set of positive Lebesgue measure. us to tap into the rich variational structure of the subdiffer- We prove, moreover, that for tensors of small rank this ential calculus for nonsmooth, nonconvex, and nonfinite- orthogonality condition can be satisfied on only a set of valued functions. As an illustration of the versatility of tensors of Lebesgue measure zero. epi-smoothing techniques, the results are applied to the general constrained optimization problem. Nick Vannieuwenhoven Tim Hoheisel Department of Computer Science University of W¨urzburg KU Leuven 97074 W¨urzburg, Germany [email protected] [email protected] Johannes Nicaise James V. Burke Department of Mathematics University of Washington KU Leuven [email protected] [email protected]

Raf Vandebril MS109 Katholieke Universiteit Leuven Epi-splines and Exponential Epi-splines: Pliable Department of Computer Science Approximation Tools [email protected] Epi-splines are primordially approximation tools rather Karl Meerbergen than their splines ‘cousins’ which are mostly interpolation K. U. Leuven (or smoothing) tools. This allows them to accommodate [email protected] more easily a richer class of information about the function, or system, that we are trying to approximate yielding in some instances stunning results. The theory that supports their properties is to a large extent anchored in the up-to- MS110 date version of Variational Analysis and implementations rely almost always on the availability of optimization rou- On Low-Complexity Tensor Approximation and tines that have been so successfully developed in the last Optimization half century. Exponential epi-spline are composition of the exponential function with an epi-spline and are particularly In this talk we shall discuss the issues related to the low- well suited to specific types of applications. This lecture rank approximation for general tensors. Attention will be will provide a quick overview of the basic framework and paid to the relationships between the classical CANDE- describe a few applications. COMP/PARAFAC (abbreviated as CP) rank, the multi- linear (or Tucker) rank, the n-mode rank, and a newly in- Roger J.-B. Wets troduced matrix rank. We then discuss the solution meth- University of California, Davis ods for the optimal rank-1 approximation model, which [email protected] leads to the so-called tensor PCA problem. Finally, we introduce the notion of low-order robust co-clustering for Johannes O. Royset tensors, and discuss its applications in bioinformatics. Operations Research Department Naval Postgraduate School Shuzhong Zhang [email protected] Department of Industrial & Systems Engineering University of Minnesota [email protected] MS110 Title Not Available Bo Jiang Abstract Not Available University of Minnesota [email protected] Bart Vandereycken Department of Mathematics Shiqian Ma Princeton University The Chinese University of Hong Kong 172 OP14 Abstracts

[email protected] tion

After their introduction in global optimization by Tawar- MS110 malani and Sahinidis over a decade ago, polyhedral relax- The Rank Decomposition and the Symmetric Rank ations have been incorporated in a variety of solvers for Decomposition of a Symmetric Tensor the global optimization of nonconvex NLPs and MINLPs. In this talk, we introduce a new relaxation paradigm for For a symmetric tensor, we show that its rank decompo- global optimization. The proposed framework combines sition must be its symmetric rank decomposition when its polyhedral and convex nonlinear relaxations, along with rank is less than its order. Furthermore, for a symmet- fail-safe techniques, convexity identification at each node ric tensor whose rank is equal to its order, we have that of the search tree, and learning heuristics for automatic its symmetric rank is equal to its order. As a corollary, sub-problem algorithm selection. We report computational for a symmetric tensor, its rank is equal to its symmetric experiments on widely-used test problem collections, in- rank when its rank is not greater than its order. This par- cluding 369 problems from GlobalLib, 250 problems from tially gives a positive answer to the conjecture proposed in MINLPLib and 980 problems from PrincetonLib. Results [Comon, Golub, Lim and Murrain, Symmetric tensors and show that incorporating the proposed techniques in the symmetric tensor rank, SIAM Journal on Matrix Analysis BARON software leads to significant improvements in ex- and Applications, 30 (2008) pp. 1254-1279]. ecution time, and increases the number of problems that are solvable to global optimality by 30%. Xinzhen Zhang, Zhenghai Huang Department of Mathematics Aida Khajavirad, Nick Sahinidis Tianjin University Carnegie Mellon university [email protected], [email protected] [email protected], [email protected]

Liqun Qi MS111 Department of Applied Mathematics The Hong Kong Polytechnic University Decomposition Techniques in Global Optimization [email protected] In this paper, we develop new decomposition techniques for convexification of constraints. First, we provide con- MS111 ditions under which convex hull of an equality constraint Recent Advances in Branch-and-Bound Methods is obtained by convexifying inequality constraints. Sec- for MINLP ond, we discuss situations where the convex hull can be written as a disjunctive hull of orthogonal sets. Third, we We present recent advances in the solution of mixed in- show that positively-homogenous functions can often be teger quadratically constrained quadratic programming replaced with a new variable. Finally, we apply these tech- (MIQCQP) problems. This very general class of prob- niques to develop closed-form expressions and optimization lems comprises several practical applications in Finance formulations. and Chemical Engineering. We describe a solution tech- nique that is currently implemented in Xpress and that Mohit Tawarmalani allows for efficiently solving various types of MIQCQP, in- Krannert School of Management cluding problems whose quadratic constraints are Lorentz [email protected] (or second-order) cones. Jean-Philippe Richard Pietro Belotti Department of Industrial and Systems Engineering FICO University of Florida pietrobelotti@fico.com [email protected]fl.edu

MS111 MS112 Computing Strong Bounds for Convex Optimiza- Robust Formulation for Online Decision Making on tion with Indicator Variables Graphs

We propose an iterative procedure for computing strong We consider the problem of routing in a stochastic network relaxations of convex quadratic programming with indica- with a probabilistic objective function when knowledge on tor variables. An SDP relaxation that is equivalent to the arc cost distributions is restricted to low-order statistics. A optimal perspective relaxation is approximately solved by stochastic dynamic programming equation with an inner a sequence of quadratic subprograms and “simple” SDPs. Moment Problem is developed to simultaneously capture Each subproblem is solved by tailored first order meth- the randomness of arc costs and cope with limited infor- ods. Our algorithm readily extends to general strongly mation on their distributions. We design algorithms with convex case. Numerical experiments and comparisons are an emphasis on tractability and run numerical experiments reported. to illustrate the benefits of a robust strategy.

Hongbo Dong Arthur Flajolet Assistant Professor Operations Research Center Washington State University MIT [email protected] fl[email protected]

Sebastien Blandin MS111 Research Scientist A Hybrid Lp/nlp Paradigm for Global Optimiza- IBM Research Collaboratory – Singapore OP14 Abstracts 173

[email protected] [email protected]

MS112 MS112 Risk-averse Routing Stochastic Route Planning in Public Transport We consider a routing problem in a transportation network, There are plenty of public transport planners that help pas- where historical delay information is available for various sangers to find a route satisfying their needs, most of them road segments. We model future delays as random vari- assuming a deterministic environment. However, vehicles ables whose distribution can be estimated. Given a source, in public transport are typically behind or before time, a destination, and a functional, we optimize the functional hence making the availability of a given journey in real life of the delay distributions along routes. This allows us to questionable. We present an algorithm that overcomes this minimize risk measures of the total delay and the proba- uncertainty by using a stochastic model for departure and bility of missing a deadline, instead of the expected total travel times. The output of the algorithm is not a single delay. route but a policy for each node that defines which services to take at a given time. Albert Akhriev, Jakub Marecek, Jia Yuan Yu IBM Research Ireland Kristof Berczi albert [email protected], [email protected], E¨otv¨os Lor´and University [email protected] Hungary [email protected] MS113 Alp´ar J¨uttner A Quasi-Newton Proximal Splitting Method E¨otv¨os Lor´and University Projected and proximal quasi-Newton methods occasion- [email protected] ally resurface in the literature as an elegant and natural method to handle constraints and non-smooth terms, but M´aty´as Korom these methods rarely become popular because the projec- E¨otv¨os Lor´and University tion step needs to be solved by an iterative non-linear Hungary solver. We show that with the correct choice of quasi- [email protected] Newton scheme and for special classes of problems (no- tably, those with non-negativity constraints, l1 penalties Marco Laumanns or hinge-loss penalties), the projection/proximal step can Research Scientist be performed almost for free. This leads to a method that IBM Zurich Research Laboratory has very few parameters and is robust, and is consistently [email protected] faster than most state-of-the-art alternatives. The second part of the talk discusses extending this approach to solve Tim Nonner general non-linear programs. Joint work with Jalal Fadili. IBM Research - Zurich [email protected] Stephen Becker UPMC Universite Paris 6 [email protected] Jacint Szabo Zuse Insitut Berlin +41-76-2125215 MS113 Composite Self-Concordant Minimization

MS112 We propose a variable metric framework for minimizing Speed-Up Algorithms for the Stochastic on-Time the sum of a self-concordant function and a possibly non- Arrival Problem smooth convex function endowed with a computable prox- imal operator. We theoretically establish the convergence We consider the stochastic on-time arrival (SOTA) problem of our framework without relying on the usual Lipschitz of finding a routing strategy that maximizes the probability gradient assumption on the smooth part. An important of reaching a destination within a pre-specified time bud- highlight of our work is a new set of analytic step-size se- get in a road network with probabilistic link travel-times. lection and correction procedures based on the structure of In this work, we provide a theoretical understanding of the the problem. We describe concrete algorithmic instances SOTA problem and present efficient computational tech- of our framework for several interesting large-scale applica- niques that can enable practical stochastic routing appli- tions and demonstrate them numerically on both synthetic cations. Experimental speedups achieved by these tech- and real data. niques are demonstrated via numerical results on real and Quoc Tran Dinh, Kyrillidis Anastasios, Volkan Cevher synthetic networks. EPFL quoc.trandinh@epfl.ch, anastasios.kyrillidis@epfl.ch, Samitha Samaranayake volkan.cevher@epfl.ch Systems Engineering UC Berkeley [email protected] MS113 Mixed Optimization: On the Interplay between Sebastien Blandin Deterministic and Stochastic Optimization Research Scientist IBM Research Collaboratory – Singapore Many machine learning algorithms follow the framework 174 OP14 Abstracts

of empirical risk minimization which involve optimizing existing coordinate descent algorithms. an objective which is in most cases a convex function or replaced by a convex surrogate. This formulation makes Miroslav Dudik an intimate connection between learning and convex op- Microsoft Research timization. In optimization for empirical risk minimiza- [email protected] tion method, there exist three regimes in which popular algorithms tend to operate: deterministic (full gradient or- acle) in which whole training data are used to compute MS114 the gradient at each iteration, stochastic (stochastic oracle) Efficient Accelerated Coordinate Descent Methods which samples a small data set per iteration to perform and Faster Algorithms for Solving Linear Systems descent towards the optimum, and more recently hybrid regime which is a combination of stochastic and determin- In this paper we show how to accelerate randomized co- istic regimes. We introduced a new setup for optimization ordinate descent methods and achieve faster convergence termed as mixed optimization, which allows to access rates without paying per-iteration costs in asymptotic run- both a stochastic oracle and a full gradient oracle. The ning time. To highlight the computational power of this proposed setup is an alternation between stochastic and algorithm, - We show how this method achieves a faster deterministic steps while trying to minimize the number asymptotic runtime than conjugate gradient. - We give an of calls to the full gradient oracle. In minimizing smooth ˜ 3/2 −1 functions, we propose the MixedGrad algorithm which has O m log (n)log(ε ) time algorithm for solving Sym- an O(ln T ) calls to the full gradient oracle and an O(T ) metric Diagonally Dominant (SDD) system. - We im- calls to the stochastic oracle, and is able to achieve an op- prove the best known asymptotic convergence guarantees timization error of O√(1/T ) which is significantly improves for Kaczmarz methods. on the known O(1/ T ) rate. This result shows an intri- cate interplay between stochastic and deterministic convex Yin Tat Lee, Aaron Sidford optimization. MIT [email protected], [email protected] Mehrdad Mahdavi Michigan State University [email protected] MS114 A Comparison of PCDM and DQAM for Big Data Problems MS113 Incremental and Stochastic Majorization- In this talk we study and compare two algorithms for Minimization Algorithms for Large-Scale Op- big data problems: the Diagonal Quadratic Approxima- timization tion method (DQAM) of Mulvey and Ruszczy´nski and the Parallel Coordinate Descent method (PCDM) of Richt´arik Majorization-minimization algorithms consist of iteratively and Tak´aˇc. These algorithms are applied to optimization minimizing a majorizing surrogate of an objective function. problems with a convex composite objective. We demon- Because of its simplicity and its wide applicability, this strate that PCDM has better theoretical guarantees (itera- principle has been very popular in statistics and in signal tion complexity results) than DQAM, and performs better processing. In this paper, we intend to make this prin- in practice than DQAM. ciple scalable. We introduce incremental and stochastic majorization-minimization schemes that are able to deal Rachael Tappenden with large-scale data sets. We compute convergence rates Edinburgh of our methods for convex and strongly-convex problems, [email protected] and show convergence to stationary points for non-convex ones. We develop several efficient algorithms based on Peter Richtarik, Burak Buke our framework. First, we propose new incremental and University of Edinburgh stochastic proximal gradient methods, which experimen- [email protected], [email protected] tally matches state-of-the-art solvers for large-scale l1 and l2 logistic regression. Second, we develop an online DC programming algorithm for non-convex sparse estimation. MS114 Finally, we demonstrate the effectiveness of our approach for solving large-scale non-convex structured matrix fac- Direct Search Based on Probabilistic Descent torization problems. Direct-search methods are a class of derivative-free algo- Julien Mairal rithms based on evaluating the objective function along INRIA - LEAR Project-Team directions in positive spanning sets. We study a more [email protected] general framework where the directions are only required to be probabilistic descent, meaning that with a signifi- cantly positive probability at least one of them is descent. MS114 This framework enjoys almost-sure global convergence and Applications of Coordinate Descent in Combinato- a global rate of 1/sqrt(k) (like in gradient methods) with rial Prediction Markets overwhelmingly high probability.

The motivating application of my talk is the problem of Serge Gratton, Cl´ement Royer pricing of interrelated securities in a prediction market. ENSEEIHT, Toulouse, France The key step is the Bregman projection on a convex set, [email protected], [email protected] which lends itself naturally to a solution by the coordinate descent. I will discuss the application domain and high- Luis N. Vicente, Zaikun Zhang light some of the specifics, which require modifications of University of Coimbra OP14 Abstracts 175

[email protected], [email protected] RWTH Aachen University [email protected]

MS115 Emre Oezkaya Gradient Projection Type Methods for Multi- RWTH Aachen University material Structural Topology Optimization using [email protected] a Phase Field Ansatz Qiqi Wang A phase field approach for structural topology optimization Massachusetts Institute of Technology with multiple materials is numerically considered. First the [email protected] problem formulation is introduced and the choice of the involved potential is discussed. This reasons our choice of the obstacle potential and leads to an optimization problem MS115 with mass constraints and inequality constraints. Moreover Improved Monte-Carlo Methods for Optimization we motivate the choice of the interpolation of the elasticity 1 Problem tensor on the interface. Then, an H -gradient projection type method in function spaces is deduced, where we can The class of problems dealt with here is a (time depen- prove global convergence. The gradient is not required but dent) parameter identification calibration problem with only far less expensive directional derivatives. In addition SDE constraints. As known, Monte-Carlo method is used it turns out that the scaling of the H1-gradient with respect to solve such problems. Nevertheless this method is very to the interface thickness is important to obtain a drastic costly and may be inefficient for large-scale problems. speed up of the method. With computational experiments Therefore some improved Monte-Carlo Techniques are pre- we demonstrate the independence of the mesh size and of sented to significantly enhance the computational efficiency the interface thickness in the number of iterations as well of an optimization algorithm. Also an adjoint approach for as its efficency in time. the Monte-Carlo method is also presented. Since this ap- proach is even successful for more sophisticated discretiza- Luise Blank tion schemes, like Milstein scheme and stochastic predictor- NWF I - Mathematik corrector schemes, we present an extended version with this Universit¨at Regensburg schemes of higher order of convergence. Finally numerical [email protected] results are presented and discussed.

Bastian Gross MS115 University of Trier Efficient Techniques for Optimal Active Flow Con- [email protected] trol

For efficient optimal active control of unsteady flows, the MS115 use of adjoint approaches is a first essential ingredient. Time-dependent Shape Derivative in Moving Do- We compare continuous and discrete adjoint approaches mains Morphing Wings Adjoint Gradient-based in terms of accuracy, efficiency and robustness. For the Aerodynamic Optimisation generation of discrete adjoint solvers, we discuss the use of Automatic Differentiation (AD) and its combination In the field of aerodynamics the focus is partly put on re- with checkpointing techniques. Furthermore, we discuss ducing drag and making the airplanes greener and eco- so-called one-shot methods. Here, one achieves simultane- nomically efficient. Nowadays the developments reached ously convergence of the primal state equation, the adjoint a point where highly-detailed shape control is needed to state equation as well as the design equation. The direction obtain further improvements. In this talk, we explain and size of the one-shot optimization steps are determined a methodology that can be used to design an airplane by a carefully selected design space preconditioner. The with morphing wings. Such aircraft has the capability one-shot method has proven to be very efficient in opti- to adapt and optimize its shape in real time in order to mization with steady partial differential equations (PDEs). achieve multi-objective mission roles efficiently and effec- Applications of the one-shot method in the field of aerody- tively. We present the general approach of this methodol- namic shape optimization with steady Navier-Stokes equa- ogy which can be applied to a number of time-dependent tions have shown, that the computational cost for an opti- PDE-constrained optimisation problems. Special attention mization, measured in runtime as well as iteration counts, is given to the derivation of the shape derivative in moving is only 2 to 8 times the cost of a single simulation of the domains. For primal and adjoint flow solution we used an governing PDE. We present a framework for applying the in-house flow solver built in FEniCS environment. We il- one-shot approach also to optimal control problems with lustrate the effectiveness of this approach using numerical unsteady Navier-Stokes equations. Straight forward appli- results. cations of the one-shot method to unsteady problems have shown, that its efficiency depends on the resolution of the Nicusor Popescu, Stephan Schmidt physical time domain. In order to dissolve this dependency, Imperial College London we consider unsteady model problems and investigate an [email protected], adaptive time scaling approach. [email protected] Nicolas R. Gauger, Stefanie Guenther RWTH Aachen University MS116 [email protected], On the Sufficiency of Finite Support Duals in Semi- [email protected] infinite Linear Programming

Anil Nemili We consider semi-infinite linear programs with countably Department of Mathematics and CCES many constraints indexed by the natural numbers. When 176 OP14 Abstracts

the constraint space is the vector space of all real valued [email protected] sequences, we show that the finite support (Haar) dual is equivalent to the algebraic Lagrangian dual of the linear program. This settles a question left open by Anderson MS116 and Nash. This result implies that if there is a duality gap Projection: A Unified Approach to Semi-Infinite between the primal linear program and its finite support Linear Programs dual, then this duality gap cannot be closed by considering the larger space of dual variables that define the algebraic We extend Fourier-Motzkin elimination to semi-infinite lin- Lagrangian dual. However, if the constraint space corre- ear programs. Applying projection leads to new character- sponds to certain subspaces of all real-valued sequences, izations of important properties for primal-dual pairs of there may be a strictly positive duality gap with the finite semi-infinite programs such as zero duality gap. Our ap- support dual, but a zero duality gap with the algebraic proach yields a new classification of variables that is used to Lagrangian dual. determine the existence of duality gaps. Our approach has interesting applications in finite-dimensional convex opti- Amitabh Basu mization. Johns Hopkins University [email protected] Chris Ryan University of Chicago Kipp Martin, Chris Ryan [email protected] University of Chicago [email protected], [email protected] Amitabh Basu Johns Hopkins University [email protected] MS116 Countably Infinite Linear Programs and Their Ap- Kipp Martin plication to Nonstationary MDPs University of Chicago [email protected] Well-known results in finite-dimensional linear programs do not extend in general to countably infinite linear pro- grams (CILPs). We will discuss conditions under which MS117 a basic feasible characterization of extreme points, weak duality, complementary slackness, and strong duality hold A Convex Optimization Approach for Computing in CILPs. We will also present simplex-type algorithms for Correlated Choice Probabilities CILPs. Nonstationary Markov decision processes and their The recently proposed Cross Moment Model (CMM) by variants will be used to illustrate our results. Mishra, Natarajan and Teo (IEEE Transactions and Au- Archis Ghate tomatic Control, 2012) utilize a semidefinite programming University of Washington approach to compute choice probabilities for the joint dis- [email protected] tribution of the random utilities that maximizes expected agent utility given only the mean, variance and covariance information. We develop an efficient first order method to Robert Smith compute choice probabilities in the CMM model. Numer- University of Michigan ical results demonstrate that the new approach can calcu- [email protected] late choice probabilities of a large number of alternatives. We also discuss an application related to route choice in MS116 transportation networks. A Linear Programming Approach to Markov Deci- Selin D. Ahipasaoglu sion Processes with Countably Infinite State Space Assistant Professor We consider Markov Decision Processes (MDPs) with Singapore University of Technology and Design countably infinite state spaces. Previous solution meth- [email protected] ods were limited to solving finite-state truncations, or esti- mating the reward function for a finite subset of states. Xiaobo Li Our proposed approach considers the countably-infinite PhD Candidate linear programming (CILP) formulations of MDPs (both University of Minnesota the number of variables and constraints are countably infi- [email protected] nite). We extend the theoretical results for finite LPs such as duality to the resulting CILPs and suggest ideas for a Rudabeh Meskarian simplex-like algorithm. Singapore University of Technology and Design [email protected] Ilbin Lee University of Michigan Karthik Natarajan [email protected] Engineering Systems and Design Singapore University of Technology and Design Marina A. Epelman, Edwin Romeijn natarajan [email protected] University of Michigan Industrial and Operations Engineering [email protected], [email protected] MS117 Modeling Dynamic Choice Behavior of Customers Robert Smith University of Michigan Traditionally, choice-based demand models in operations OP14 Abstracts 177

used transaction data to learn aggregate user preferences. various simulated and empirical datasets to test the per- However, there is now an abundance of panel data: trans- formance of this approach. We evaluate the performance action data tagged with customer ids, which is obtained against the classical Multinomial Logit, Mixed Logit, and by tracking customers through credit cards, loyalty cards, a machine learning approach developed by Evgeniou et etc. Given this, we consider the problem of constructing al. (2007) (for partworth heterogeneity). Our numeri- individual user preferences from their purchase paths in or- cal results indicate that MDM provides a practical semi- der to predict demand, and personalize recommendations parametric alternative to choice modeling. and promotions. We do not assume access to demographic information about the customers or any attribute infor- Xiaobo Li mation about the products. In order to accomplish this, PhD Candidate we propose a dynamic model for how customers construct University of Minnesota preference lists over time as a function of all products that [email protected] were purchased and were on offer in the past. The model accounts for several ”behavioral” biases in choice that have Vinit Kumar Mishra been empirically established. Due to scarcity of data, the Department of Business Analytics individual preference lists we construct are often sparse. The University of Sydney Business School Hence, we ”cluster” the constructed preference lists in or- [email protected] der to make demand predictions. Our methods were tested on real-world IRI dataset. We also provide statistical and Karthik Natarajan computational guarantees for our methods, in the process Engineering Systems and Design of which we make important theoretical contributions to Singapore University of Technology and Design the study of preference theory in statistics and computer natarajan [email protected] science.

Srikanth Jagabathula,GustavoVulcano Dhanesh Padmanabhan Leonard N. Stern School of Business General Motors R&D - India Science Lab NYU Bangalore, India [email protected], [email protected] [email protected]

Chung-Piaw Teo MS117 Department of Decision Sciences A New Compact Lp for Network Revenue Manage- NUS Business School ment with Customer Choice [email protected]

The choice network revenue management model incorpo- rates customer purchase behavior as a function of the of- MS118 fered products. We derive a new compact LP formulation On the Use of Second Order Information for the for the choice network RM problem. Our LP gives an up- Numerical Solution of L1-Pde-Constrained Opti- per bound that is provably between the choice LP value mization Problems and the affine relaxation, and often coming close to the latter in numerical experiments. We present a family of algorithms for the numerical solu- tion of PDE-constrained optimization problems, which in- Kalyan Talluri volve an L1-cost of the design variable in the objective func- Universitat Pompeu Fabra tional. It is well-known that this non-differentiable term [email protected] leads to a sparse structure of the optimal control, which acts on ”small” regions of the domain. In order to cope Sumit Kunnumkal with the non-differentiability, we consider a Huber regu- 1 Indian School of Business larization of the L -term, which approximates the original sumit [email protected] problem by a family of parameterized differentiable prob- lems. The main idea of our method consists in computing descent directions by incorporating second order informa- MS117 tion. An orthantwise-direction strategy is also used in the On Theoretical and Empirical Aspects of Marginal spirit of OW-algorithms in order to obtain a fast identifi- Distribution Choice Models cation of the active sets. We present several experiments to illustrate the efficiency of our approach. In this paper, we study the properties of a recently pro- posed class of semiparametric discrete choice models (re- Juan Carlos De los Reyes ferred to as the Marginal Distribution Model), by optimiz- Escuela Polit´ecnica Nacional Quito ing over a family of joint error distributions with prescribed Quito, Ecuador marginal distributions. Surprisingly, the choice probabili- [email protected] ties arising from the family of Generalized Extreme Value models can be obtained from this approach, despite the dif- ference in assumptions on the underlying probability dis- MS118 tributions. We use this connection to develop flexible and Finite Element Error Estimates for Elliptic Op- general choice models to incorporate respondent and prod- timal Control Problems with Varying Number of uct/attribute level heterogeneity in both partworths and Finitely Many Inequality Constraints scale parameters in the choice model. Furthermore, the ex- tremal distributions obtained from the MDM can be used We derive a priori error estimates for the finite element to approximate the Fisher’s Information Matrix to obtain discretization of an elliptic optimal control problem with reliable standard error estimates of the partworth param- finitely many pointwise inequality constraints on the state. eters, without having to bootstrap the method. We use In particular, following ideas of Leykekhman, Meidner, and 178 OP14 Abstracts

Vexler, who considered a similar problem with equality [email protected], [email protected] constraints, we obtain an optimal rate of convergence for the state variable. In contrast to earlier works, the number of inequality constraints may vary on refined meshes. MS119 Funding Opportunities at NSF Ira Neitzel Technische Universit¨at M¨unchen, This session discusses funding opportunities at the Na- Department of Mathematics tional Science Foundation. [email protected] Sheldon H. Jacobson University of Illinois Winnifried Wollner Dept of Computer Science University of Hamburg, Department of Mathematics [email protected] [email protected]

MS119 MS118 Funding Opportunities in the Service and Manu- PDE Constrained Optimization with Pointwise facturing Enterprise Systems Programs Gradient-State Constraints We provide an overview of the MES and SES programs, which support research on strategic decision making, de- In this talk, we will review several recent result in PDE sign, planning, and operation of manufacturing and service constrained optimization with pointwise constraints on the enterprises, that is both grounded in an interesting and gradient of the state. This includes barrier and penalty relevant application and requires the development of novel methods in a function space setting to eliminate the con- methodologies. straint on the gradient. Convergence of such methods is discussed. Further, we will consider the discretization of Edwin Romeijn such problems in particular for non smooth domains, where National Science Foundation the control to state mapping does not assert the gradient [email protected] to be Lipschitz.

Michael Hinterm¨uller PP1 Humboldt-University of Berlin Design of a Continuous Bioreactor Via Multiobjec- [email protected] tive Optimization Coupled with Cfd Modeling

Anton Schiela In this paper a multiobjective design optimization was ap- Technische Universit¨at Hamburg plied in a continuous bioreactor. The aim of the procedure [email protected] is to find the geometry most favorable to simultaneously optimize the fluid dynamics parameters that can influence Winnifried Wollner the biochemical process involved in the bioreactor, such University of Hamburg, Department of Mathematics as shear stress and residence time distribution. An open [email protected] source computer package,called pyCFD-O, was developed to perform the CFD and the optimizations processes in a fully automatic manner.

MS118 Jonas L. Ansoni Boundary Concentrated Finite Elements for Opti- University of Sao Paulo mal Control Problems and Application to Bang- Escola de Engenharia de So Carlos - EESC USP Bang Controls [email protected]

We consider the discretization of an optimal boundary con- Paulo Seleghim Jr. trol problem with distributed observation by the bound- University of Sao Paulo ary concentrated finite element method. With an H1+δ(Ω) Escola de Engenharia de Sao Carlos - EESC USP regular elliptic PDE on two-dimensional domains as con- [email protected] straint, we prove that the discretization error u∗ − ∗  −δ uh L2(Γ) decreases like N ,whereN denotes the total number of unknowns. We discuss the application to prob- PP1 lems whose optimal control displays bang-bang character. Interior Point Algorithm As Applied to the Trans- mission Network Expansion Planning Problem

Jan-Eric Wurst The main purpose of this poster is to study the long term Universit¨at W¨urzburg transmission network expansion planning problem, which [email protected] consists in finding an optimal synthesis of a power network that minimizes the capital cost of circuits construction un- der some technical restrictions. Regarding the robustness Daniel Wachsmuth and efficiency when solving large scale real-world problems, Institute of Mathematics we use the IP algorithm proposed by [Vanderbei & Shanno, University of W¨urzburg 1999] to the IEEE-118 nodes test systems to get the opti- [email protected] mum cost of construction.

Sven Beuchler, Katharina Hofer Carlos Becerril Universitaet Bonn Instituto Polit´ecnico Nacional OP14 Abstracts 179

Escuela Superior de Ingeniera Mec´anica y El´ectrica [email protected] [email protected]

Ricardo Mota, Mohamed Badaoui PP1 National Polytechnic Institute, IPN A Minimization Based Krylov Method for Ill-Posed ESIME Problems and Image Restoration [email protected], [email protected] We are concerned with the computation of accurate ap- proximate solutions of linear discrete ill-posed problems. PP1 These are systems of equations with a very ill-conditioned matrix and an error-contaminated right-hand side vector. Meshless Approximation Minimizing Divergence- Their solution requires regularization such as a Tikhonov free Energy regularization or other regularization techniques based on Krylov subspace methods such as conjugate gradient- or In this work, we propose a meshless approximation of a vec- GMRES-type methods. It can be difficult to determine tor field in multidimensional space minimizing quadratic when to truncate. We will be interested in a minimiza- forms related to the divergence of a vector field. Our ap- tion method namely the reduced rank extrapolation. The proach guarantees the conservation of the divergence-free, present work describes a novel approach to determine a which is of great importance in applications. For instance, suitable truncation index by comparing the original and divergence-free vector fields correspond to incompressible extrapolated vector sequences. Applications to image fluid flows. Our construction is based on the meshless ap- restoration is presented. proximation by the pseudo-polyharmonic functions which are a class of radial basis functions minimizing an appro- Khalide Jbilou priate energy in an adequate functional space. We provide University, ULCO Calais France an HelmholtzHodge decomposition of the used native func- [email protected] tional space. Numerical examples are included to illustrate our approach. PP1 Abderrahman Bouhamidi Introducing Penlab, a Matlab Code for Nonlinear University Littoral Cˆote d’Opale (and) Semidefinite Optimization [email protected] We will introduce a new code PENLAB, an open Matlab implementation and extension of our older code PENNON. PP1 PENLAB can solve problems of nonconvex nonlinear opti- Robust Security-Constrained Unit Commitment mization with standard (vector) variables and constraints, with Dynamic Ratings as well as matrix variables and constraints. We will demon- strate its functionality using several nonlinear semidefinite The actual capacity of overhead transmission lines and gas examples. generators vary by their ambient temperatures. To pro- Michal Kocvara vide efficient and also reliable generation and dispatch de- School of Mathematics cisions, we develop a two stage robust unit commitment University of Birmingham formulation considering both dynamic line and generation [email protected] ratings, subject to the uncertainty in weather conditions. Our model is demonstrated on standard IEEE testbeds as well as real industrial data. Michael Stingl Institut f¨ur Angewandte Mathematik Anna Danandeh,BoZeng Universit¨at Erlangen-N¨urnberg, Germany USF [email protected] [email protected], [email protected] Jan Fiala Brent Caldwell The Numerical Algorithms Group, Ltd Tampa Electric Company [email protected] [email protected] PP1 PP1 Aircraft Structure Design Optimization Via a Matrix-Free Approach Preventive Maintenance Scheduling of Multi- Component Systems with Interval Costs When applying gradient-based optimization algorithms to structural optimization problems, it is the computation of We consider the preventive maintenance scheduling of a the constraint gradients that requires the most effort be- multi-component system with economic dependencies. The cause repeated calls to the finite element solver are needed. corresponding decision problem is modeled as an integer We aim to reduce this effort by using a matrix-free op- linear program (ILP) where maintenance of a set of compo- timizer that estimates the gradient information based on nents are to be scheduled over a discretized time horizon. quasi-Newton methods. We demonstrate our approach us- The main contributions are a theoretical analysis of the ing two problems arising from aircraft design. We show properties of the ILP model and three case studies demon- that our method can result in many fewer calls to the ex- strating the usefulness of the model. pensive finite element solver.

Emil Gustavsson Andrew Lambe Chalmers University of Technology University of Toronto 180 OP14 Abstracts

[email protected] other than transforming each equality into two inequali- ties, which does not work very well in practice. We present Joaquim Martins various extensions to solve problems with linear equality University of Michigan constraints. Strategies combining these extensions are pre- [email protected] sented. The convergence properties of MADS are main- tained. Numerical results on a subset of CUTEr problems Sylvain Arreckx are reported. Ecole Polytechnique de Montreal Mathilde Peyrega [email protected] Ecole´ Polytechnique de Montr´eal, Canada [email protected] Dominique Orban GERAD and Dept. Mathematics and Industrial Engineering Charles Audet ´ Ecole Polytechnique de Montreal Ecole Polytechnique de Montr´eal - GERAD [email protected] [email protected]

Sebastien Le Digabel PP1 GERAD - Polytechnique Montreal Damage Detection Using a Hybrid Optimization [email protected] Algorithm

This paper presents a hybrid stochastic/deterministic op- PP1 timization algorithm to solve the target optimization prob- Nonlinear Model Reduction in Optimization with lem of vibration-based damage detection. It employs the Implicit Constraints representation formula of Pincus to locate the region of the optimum and the Nelder-Mead algorithm as the local We consider nonlinear optimization problems in which the optimizer. A series of numerical examples with different evaluation of the objective function requires the solution damage scenarios and noise levels was performed under im- of time-dependent PDEs of large dimension. We design a pact and ambient vibrations. The proposed algorithm was surrogate management framework that relies on the POD more accurate and efficient than well-known heuristic al- and DEIM nonlinear model reduction techniques to con- gorithms. struct surrogates for the constraints. We present results for two case studies: an optimal control problem with a Rafael H. Lopez, Leandro Miguel simplified flow equation (Burgers’ equation), and model Universidade Federal de Santa Catarina calibration (history matching) of an oil reservoir model. Civil Engineering Department This work is the result of collaborations with Manuel Bau- [email protected], [email protected] mann, Slawomir Szklarz, and Martin van Gijzen. Marielba Rojas Andr´eTorii Delft University of Technology Universidade Federal da Paraba [email protected] Civil Engineering Department [email protected] PP1 PP1 An Optimization Model for Truck Tyres’ Selection Social Welfare Comparisons with Fourth-Order To improve the truck tyre selection process at Volvo Group Utility Derivatives Trucks Technology an optimization model has been devel- oped to determine an optimal set of tyres for each vehicle In this paper, we provide intuitive justifications of norma- and operating environment specification. The overall pur- tive restrictions based on the signs of fourth-order deriva- pose is to reduce the fuel consumption while preserving the tives of utilities in the context of multidimensional wel- levels of other tyre dependent features. The model - cur- fare analysis. For this, we develop a new notion of welfare rently being improved with respect to accuracy - is based shock sharing. This allows us to derive new characteriza- on vehicle dynamics equations and surrogate models of ex- tions for symmetric and asymmetric conditions on the signs pensive simulation based functions. of fourth-order derivatives of utility functions. Then, we use these restrictions to derive new stochastic dominance Zuzana Sabartova criteria for multidimensional welfare comparisons. Chalmers University of Technology [email protected] Christophe Muller Aix-Marseille University [email protected] PP1 Analysis of Gmres Convergence Via Convex Opti- mization PP1 The Mesh Adaptive Direct Search Algorithm for In this poster, we will give new bounds for the GMRES Blackbox Optimization with Linear Equalities method of Saad and Schultz, for solving linear system. The explicit formula of the residual norm of the GMRES when The Mesh Adaptive Direct Search (MADS) algorithm is applied to normal matrix, which is a rational function, is designed to solve blackbox optimization problems with given in terms of eigenvalues and of the components of the bounds under general inequality constraints. Currently eigenvector decomposition of the initial residual. By mini- the MADS theory does not support equality constraints mizing this rational function over a convex subset, we ob- OP14 Abstracts 181

tain the sharp bound of the residual norm of the GMRES presumption of device ordering, we calculate maximally ef- method applied to normal matrix, even if the spectrum ficient designs for wet cycles and integrated solid-oxide fuel contains complex eigenvalues. Hence we completely char- cell, gas turbine systems. acterize the worst case GMRES for normal matrices. We use techniques from constrained optimization rather than Rebecca Zarin Pass,ChrisEdwards solving the classical min-max problem (problem in polyno- Stanford University mial approximation theory) [email protected], n/a

Hassane Sadok Universit´e du Littoral Calais Cedex, France [email protected]

PP1 In-Flight Trajectory Redesign Using Sequential Convex Programming

Onboard trajectory redesign capabilities for autonomous orbiters is being explored as part of the NASA Space Tech- nology Research Fellowship. Development has combined system discretization, heuristic search, sequential convex programming, and software architecture. The onboard sys- tem focus has motivated the use of techniques with prov- able convergence properties and feasible iterations in case of system interrupts or time constraints. Examples of local correction and full replanning will be shown for the case of an orbiter at Phobos.

Eric Trumbauer, Benjamin F. Villac University of California, Irvine [email protected], [email protected]

PP1 Riemannian Pursuit for Big Matrix Recovery

Low rank matrix recovery (MR) is a fundamental task in many real-world applications. We propose a new model for the MR problem and then relax it as a convex quadrati- cally constrained quadratic program (QCQP) problem. To solve this problem, we propose an efficient Convex Rieman- nian Pursuit (CRP) method which addresses the MR prob- lem by iteratively solving a series of fixed-rank problems. Experimental results on matrix completion tasks show that, compared with state-of-the-art methods, the pro- posed method can achieve superior scalability and main- tain comparable or even better performance in terms of recovery error on both large-scale synthetic and real-world datasets.

Li Wang University of California, San Diego [email protected]

MingkuiTan,IvorTsang Nanyang Technology Univeristy [email protected], [email protected]

PP1 Stochastic Optimization of the Topology of a Combustion-based Power Plant for Maximum En- ergy Efficiency

Much research has been done to improve the efficiency of internal combustion-based power plants. These plants are not limited by the Carnot efficiency and the topology of the optimal power plant using existing technologies is un- known. This work uses genetic algorithms to determine the optimal structure and parameters of such a plant. With no 18268 2014 SIAM Conference on Optimization

Notes 2014 SIAM Conference on Optimization 183

OP14 Speaker and Organizer Index

Italicized names indicate session organizers. 184 2014 SIAM Conference on Optimization

A B Bott, Stefanie, CP7, 4:50 Mon Adams, Elspeth, MS40, 2:30 Tue Backhoff, Julio, MS15, 2:30 Mon Bouhamidi, Abderrahman, PP1, 8:00 Mon Adjiman, Claire, MS21, 2:00 Mon Badri, Hamidreza, CP24, 5:30 Wed Boumal, Nicolas, CP3, 5:50 Mon Adly, Samir, CP2, 5:30 Mon Baes, Michel, MS66, 11:00 Wed Boutsidis, Christos, CP16, 6:10 Wed Aghayeva, Charkaz A., CP8, 5:30 Mon Bai, Xi, MS16, 2:30 Mon Bradley, Joseph, MS55, 5:30 Tue Ahipasaoglu, Selin D., MS117, 4:30 Thu Bakker, Craig K., CP13, 4:30 Mon Bravo, Fernanda, MS17, 3:00 Mon Ahipasaoglu, Selin D., MS117, 5:00 Thu Balvert, Marleen, CP24, 6:10 Wed Brekelmans, Ruud, MS62, 10:30 Wed Ahmadi, Amir Ali, MS76, 2:30 Wed Bandeira, Afonso S., MS35, 10:45 Tue Bremner, David, MS32, 10:15 Tue Ahmadi, Amir Ali, MS107, 4:30 Thu Baraniuk, Richard G., MS69, 10:00 Wed Brockhoff, Dimo, MS47, 4:00 Tue Brown, David, MS13, 2:00 Mon Ahookhosh, Masoud, CP22, 5:10 Wed Barbier, Thibault, CP17, 5:10 Wed Akimoto, Youhei, MS47, 3:30 Tue Barton, Paul I., MS11, 9:30 Mon Brown, David, MS13, 2:00 Mon Akle, Santiago, MS58, 6:30 Tue Barton, Paul I., MS21, 2:30 Mon Bucarey, Víctor D., CP17, 4:50 Wed Akrotirianakis, Ioannis, CP12, 5:30 Mon Bastin, Fabian, CP18, 6:10 Wed Bueno, Luis Felipe, CP26, 5:10 Wed Alexandrov, Natalia, PD1, 12:30 Tue Basu, Amitabh, MS116, 5:00 Thu Burdakov, Oleg, MS10, 9:30 Mon Alexandrov, Natalia, MS60, 9:30 Wed Bauschke, Heinz, MS85, 9:30 Thu Burke, James V., MS109, 5:00 Thu Alexandrov, Natalia, MS60, 10:00 Wed Bauschke, Heinz, MS85, 9:30 Thu Alfakih, Abdo, MS77, 2:30 Wed C Becerril, Carlos, PP1, 8:00 Mon Calatroni, Luca, MS106, 2:30 Thu Allison, James T., MS5, 11:00 Mon Beck, Amir, MS102, 2:00 Thu Cao, Yankai, MS36, 3:30 Tue Ames, Brendan, MS52, 5:00 Tue Becker, Stephen, MS113, 5:00 Thu Carlsson, John Gunnar, MS104, 3:00 Thu Ames, Brendan P., MS52, 5:00 Tue Bello Cruz, Jose Yunier, MS61, 9:30 Carter, Richard, CP11, 5:30 Mon Andersen, Erling D., MS101, 2:00 Thu Wed Cartis, Coralia, MS23, 2:00 Mon Andersen, Erling D., MS101, 3:30 Thu Belotti, Pietro, MS111, 5:00 Thu Cevher, Volkan, MS113, 4:30 Thu Andersen, Martin S., MS18, 2:00 Mon Benson, Hande Y., MS6, 9:30 Mon Chan, Timothy, MS49, 5:30 Tue Anitescu, Mihai, MS24, 10:15 Tue Benson, Hande Y., MS6, 10:30 Mon Chandrasekaran, Venkat, MS69, 9:30 Wed Anitescu, Mihai, MS24, 10:45 Tue Berg, Bjorn, MS1, 10:00 Mon Chandrasekaran, Venkat, MS71, 2:00 Wed Anitescu, Mihai, MS36, 2:30 Tue Bertozzi, Andrea L., MS105, 2:30 Thu Chandrasekaran, Venkat, MS107, 4:30 Anitescu, Mihai, MS48, 5:00 Tue Bienstock, Daniel, MS95, 2:00 Thu Thu Anjos, Miguel F., MS84, 10:30 Thu Bienstock, Daniel, MS95, 3:30 Thu Charitha, Charitha, MS82, 3:00 Wed Ansoni, Jonas L., PP1, 8:00 Mon Birbil, Ilker S., CP12, 5:10 Mon Chen, Chen, MS108, 4:30 Thu Anstreicher, Kurt M., MS44, 4:00 Tue Blanco, Victor, CP13, 4:50 Mon Chen, Ruobing, MS25, 10:45 Tue Antil, Harbir, MS106, 2:00 Thu Blandin, Sebastien, CP11, 4:30 Mon Chen, Xiao, CP13, 5:30 Mon Aoki, Takaaki, CP21, 6:10 Wed Blandin, Sebastien, MS112, 4:30 Thu Chen, Yudong, MS52, 5:30 Tue Aravkin, Aleksandr, MS97, 3:30 Thu Blank, Luise, MS115, 5:00 Thu Cheng, Jianqiang, CP18, 5:10 Wed Armand, Paul, MS34, 10:45 Tue Bodony, Daniel J., MS103, 2:30 Thu Cheung, Yuen-Lam, MS77, 3:00 Wed Asaki, Thomas, CP14, 5:50 Wed Boehm, Christian, CP8, 4:50 Mon Chiang, Naiyuan, MS27, 10:45 Tue Assuncao Monteiro, Sergio, CP3, 6:10 Boros, Endre, MS40, 2:30 Tue Chouldechova, Alexandra, MS102, 3:30 Mon Boros, Endre, MS40, 4:00 Tue Thu Atamturk, Alper, MS12, 2:30 Wed Borrelli, Francesco, MS68, 10:00 Wed Chu, Eric, MS101, 2:30 Thu Auger, Anne, MS47, 2:30 Tue Bortfeld, Thomas, MS49, 5:00 Tue Chua, Chek Beng, CP1, 5:10 Mon Auger, Anne, MS47, 2:30 Tue Bot, Radu Ioan, MS61, 9:30 Wed Chubanov, Sergei, MS32, 10:45 Tue Augustin, Florian, MS91, 9:30 Thu Bot, Radu Ioan, MS61, 10:00 Wed Chung, Matthias, CP6, 5:10 Mon

continued on next page 2014 SIAM Conference on Optimization 185

Chung, Sung, CP15, 4:30 Wed Deng, Yan, CP18, 4:30 Wed Ferris, Michael C., MS2, 2:00 Mon Clason, Christian, MS106, 2:00 Thu Dentcheva, Darinka, MS9, 9:30 Mon Festa, Paola, CP4, 5:10 Mon Clason, Christian, MS106, 3:30 Thu Dentcheva, Darinka, MS74, 2:30 Wed Finhold, Elisabeth, MS40, 3:00 Tue Clason, Christian, MS118, 4:30 Thu Devanur, Nikhil R., MS93, 10:00 Thu Flajolet, Arthur, MS112, 6:00 Thu Claus, Matthias, MS57, 5:30 Tue Dey, Santanu S., MS4, 9:30 Mon Floudas, Christodoulos A., MS21, 2:00 Mon Collado, Ricardo A., MS74, 3:30 Wed Di Pillo, Gianni, MS25, 10:15 Tue Floudas, Christodoulos A., MS21, 3:30 Conn, Andrew R., MS25, 10:15 Tue di Serafino, Daniela, MS7, 10:00 Mon Mon Crélot, Anne-Sophie, CP14, 4:30 Wed Dickinson, Peter, MS64, 10:00 Wed Fogel, Fajwel, MS35, 10:15 Tue Curtis, Frank E., MS8, 9:30 Mon Dillon, Keith, CP25, 4:30 Wed Forsgren, Anders, MS46, 2:30 Tue Curtis, Frank E., MS27, 10:15 Tue Doan, Xuan Vinh, MS75, 3:30 Wed Forsgren, Anders, MS58, 5:00 Tue Curtis, Frank E., PD1, 12:30 Tue Dong, Hongbo, MS111, 5:30 Thu Forsgren, Anders, MS83, 10:30 Thu Curtis, Frank E., MS80, 2:00 Wed Drusvyatskiy, Dmitriy, MS109, 5:30 Fountoulakis, Kimon, MS41, 4:00 Tue Curtis, Frank E., MS99, 2:00 Thu Thu Franke, Susanne, MS45, 3:00 Tue Curtis, Frank E., MS99, 3:30 Thu Dudik, Miroslav, MS114, 5:00 Thu Freund, Robert M., MS78, 2:00 Wed Duer, Mirjam, MS29, 10:15 Tue Freund, Robert M., CP19, 6:10 Wed D Duer, Mirjam, MS29, 11:45 Tue Friedland, Shmuel, MS98, 2:00 Thu Dadush, Daniel, MS28, 10:45 Tue Friedlander, Michael P., MS58, 6:00 Tue Dai, Yuhong, MS7, 11:00 Mon E Fukuda, Mituhiro, MS88, 9:30 Mon Danandeh, Anna, PP1, 8:00 Mon Eastman, Roger, MS19, 3:00 Mon Fukuda, Mituhiro, CP25, 5:30 Wed Dang, Cong D., MS43, 3:30 Tue Eaton, Julia, MS64, 11:00 Wed Dapogny, Charles, MS37, 3:30 Tue Echeverria Ciaurri, David, MS25, 11:45 G Tue Dash, Sanjeeb, MS4, 10:00 Mon Garin, Araceli, CP5, 5:10 Mon Eckstein, Jonathan, MS74, 3:00 Wed d’Aspremont, Alexandre, MS35, 10:15 Tue Gauger, Nicolas R., MS115, 5:30 Thu Eichfelder, Gabriele, MS94, 9:30 Thu De Angelis, Tiziano, MS15, 3:00 Mon Ge, Rong, MS75, 3:00 Wed Eichfelder, Gabriele, MS94, 9:30 Thu De Asmundis, Roberta D., MS19, 2:30 Geihe, Benedict, MS57, 6:00 Tue Mon El Ghaoui, Laurent, MS107, 5:00 Thu Gendron, Bernard, MS70, 9:30 Wed De Klerk, Etienne, MS51, 6:00 Tue Engau, Alexander, CP2, 5:50 Mon Gendron, Bernard, MS70, 10:30 Wed De Klerk, Etienne, MS76, 2:00 Wed Epelman, Marina A., MS42, 2:30 Tue Gfrerer, Helmut, MS45, 3:30 Tue de Laat, David, MS30, 11:45 Tue Epelman, Marina A., MS79, 2:00 Wed Ghaddar, Bissan, CP11, 6:10 Mon De Loera, Jesús A., MS32, 10:15 Tue Erway, Jennifer, MS10, 10:30 Mon Ghadimi, Saeed, MS23, 2:30 Mon De Loera, Jesús A., MS44, 2:30 Tue Escudero, Laureano F., CP5, 4:50 Mon Ghasemi, Mehdi, MS76, 3:30 Wed De Loera, Jesús A., MS56, 5:00 Tue Eskandarzadeh, Saman, CP5, 5:50 Mon Ghate, Archis, MS116, 5:30 Thu De Loera, Jesús A., MS54, 5:00 Tue Ghosh, Anusuya, CP16, 5:50 Wed De los Reyes, Juan Carlos, MS106, 2:00 Thu F Fan, Jinyan, MS39, 3:00 Tue Gijben, Luuk, MS29, 11:15 Tue De los Reyes, Juan Carlos, MS118, 4:30 Thu Fang, Ethan, MS6, 9:30 Mon Gijswijt, Dion, MS30, 10:45 Tue De los Reyes, Juan Carlos, MS118, 6:00 Gill, Philip E., MS83, 9:30 Thu Thu Farias, Vivek, MS13, 3:00 Mon Gill, Philip E., MS83, 9:30 Thu De Santis, Marianna, MS41, 3:00 Tue Fawzi, Hamza, MS75, 2:30 Wed Gillis, Nicolas, MS75, 2:00 Wed Delage, Erick, CP15, 5:50 Wed Fazel, Maryam, MS71, 2:00 Wed Gillis, Nicolas, MS75, 2:00 Wed Dellepiane, Umberto, CP10, 4:50 Mon Fazel, Maryam, MS93, 9:30 Thu Giselsson, Pontus, CP20, 6:10 Wed Dempe, Stephan, MS45, 2:30 Tue Federico, Salvatore, CP21, 5:10 Wed Glasmachers, Tobias, MS47, 3:00 Tue Den Hertog, Dick, MS62, 9:30 Wed Feinberg, Eugene A., MS79, 3:00 Wed Glineur, Francois, MS104, 3:30 Thu Den Hertog, Dick, MS62, 9:30 Wed Fercoq, Olivier, MS43, 3:00 Tue Ferrari, Giorgio, MS15, 3:30 Mon

continued on next page 186 2014 SIAM Conference on Optimization

Goebel, Rafal, MS82, 2:30 Wed Hähnle, Nicolai, MS32, 11:15 Tue Huang, Junzhou, MS105, 3:30 Thu Goez, Julio C., MS12, 2:00 Wed Hansen, Thomas D., MS79, 3:30 Wed Huang, Rui, MS48, 6:00 Tue Gomes, Francisco M., CP12, 4:50 Mon Hansuwa, Sweety, CP17, 5:50 Wed Huang, Wei, CP25, 5:10 Wed Gondzio, Jacek, MS34, 10:15 Tue Hatz, Kathrin, CP21, 5:50 Wed Hungerlaender, Philipp, MS33, 10:45 Gondzio, Jacek, MS34, 10:15 Tue Hauser, Raphael A., MS51, 5:30 Tue Tue Gonzaga, Clovis C., CP9, 4:30 Mon Hayashi, Shunsuke, CP9, 5:10 Mon Hupp, Lena, CP4, 4:50 Mon Gorgone, Enrico, MS70, 10:00 Wed Heinkenschloss, Matthias, MS20, 2:00 Hwang, John T., MS5, 9:30 Mon Gotoh, Jun-ya, MS67, 10:00 Wed Mon Helmberg, Christoph, MS33, 10:15 Tue I Gouveia, Joao, MS89, 11:00 Thu Iaccarino, Gianluca, MS60, 11:00 Wed Helou Neto, Elias Salomão S., MS97, Grant, Michael C., MS101, 2:00 Thu Iancu, Dan A., MS13, 3:30 Mon 2:30 Thu Grapiglia, Geovani, MS23, 3:30 Mon Ito, Masaru, CP22, 4:50 Wed Helton, J.William, MS39, 4:00 Tue Gratton, Serge, MS23, 3:00 Mon Iusem, Alfredo N., CP22, 6:10 Wed Hemmecke, Raymond, MS56, 5:00 Tue Gray, Genetha, MS60, 9:30 Wed Henrion, Didier, MT1, 9:30 Mon Gray, Justin, MS5, 10:30 Mon J Henrion, Didier, MT1, 9:30 Mon Jacobson, Sheldon H., MS119, 2:00 Wed Grebille, Nicolas, MS96, 2:00 Thu Henrion, Didier, MS39, 3:30 Tue Jacobson, Sheldon H., MS119, 3:00 Greif, Chen, MS58, 5:30 Tue Herskovits, Jose, CP2, 5:10 Mon Wed Griewank, Andreas, MS11, 9:30 Mon Hesse, Robert, MS61, 10:30 Wed Jacobson, Sheldon H., CP26, 6:10 Wed Griffin, Joshua, CP14, 4:50 Wed Hicken, Jason E., MS5, 10:00 Mon Jagabathula, Srikanth, MS117, 5:30 Thu Grigas, Paul, MS78, 2:30 Wed Hintermueller, Michael, MS2, 2:00 Mon Jaggi, Martin, MS78, 3:30 Wed Grimm, Boris, CP24, 4:30 Wed Hintermueller, Michael, MS20, 3:30 Jargalsaikhan, Bolor, MS29, 10:45 Tue Griva, Igor, MS6, 10:00 Mon Mon Jbilou, Khalide, PP1, 8:00 Mon Gross, Bastian, MS115, 6:00 Thu Hoeffner, Kai, MS11, 10:30 Mon Jeon, Hyemin, MS87, 10:00 Thu Grothey, Andreas, MS108, 5:30 Thu Hoheisel, Tim, MS109, 4:30 Thu Jiang, Bo, MS86, 10:30 Thu Guan, Yongpei, MS24, 11:15 Tue Hoheisel, Tim, MS109, 4:30 Thu Jiang, Hao, MS83, 11:00 Thu Guglielmi, Roberto, CP7, 5:50 Mon Hong, Mingyi, CP22, 4:30 Wed Jiang, Riuwei, MS54, 6:00 Tue Gulten, Sitki, CP6, 4:30 Mon Horesh, Lior, MS91, 9:30 Thu Jin, Bangti, MS106, 3:00 Thu Guntu, Raja Rao, CP20, 5:50 Wed Horesh, Lior, MS91, 10:30 Thu Joswig, Michael, MS32, 11:45 Tue Gupte, Akshay, MS87, 10:30 Thu Horesh, Raya, MS68, 9:30 Wed Júdice, Joaquim J., CP1, 6:10 Mon Gupte, Akshay, MS111, 4:30 Thu Horesh, Raya, MS68, 9:30 Wed Gustavsson, Emil, PP1, 8:00 Mon Horst, Ulrich, MS15, 2:00 Mon K Kalashnikov, Vyacheslav V., MS45, 4:00 Guzman, Cristobal, MS35, 11:45 Tue Hosseini, Seyedehsomayeh, CP9, 5:50 Tue Mon H Hough, Patricia D., MS60, 9:30 Wed Kannan, Aswin, MS26, 11:15 Tue Haber, Eldad, MS91, 9:30 Thu Houska, Boris, MS104, 2:30 Thu Kasimbeyli, Nergiz, CP23, 5:50 Wed Haeser, Gabriel, CP9, 4:50 Mon Hu, Jian, MS3, 10:30 Mon Kasimbeyli, Refail, MS94, 11:00 Thu Hager, William, MS80, 3:30 Wed Hu, Shenglong, MS98, 3:00 Thu Kazemi, Parimah, CP26, 5:30 Wed Hager, William, MS105, 2:00 Thu Kelman, Tony, MS36, 2:30 Tue

continued on next page 2014 SIAM Conference on Optimization 187

Keuthen, Moritz, CP20, 4:30 Wed Laumanns, Marco, MS112, 5:00 Thu Liu, Andrew L., MS108, 5:00 Thu Khajavirad, Aida, MS111, 6:00 Thu Laurent, Monique, MT1, 9:30 Mon Liu, Ji, MS55, 6:30 Tue Khan, Kamil, MS11, 10:00 Mon Laurent, Monique, MT1, 9:30 Mon Liu, Jun, CP8, 5:10 Mon Kilinc-Karzan, Fatma, MS28, 10:15 Tue Le Digabel, Sebastien, MS25, 11:15 Tue Liu, Minghui, MS29, 10:15 Tue Kim, Kibaek, MS36, 4:00 Tue Leder, Kevin, MS49, 6:00 Tue Liu, Xin, MS59, 5:30 Tue Kim, Minsun, MS42, 2:30 Tue Lee, Changhyeok, MS3, 11:00 Mon Ljubic, Ivana, MS81, 2:00 Wed Kim, Sunyoung, MS88, 10:30 Mon Lee, Ilbin, MS79, 2:00 Wed Lodi, Andrea, MS4, 9:30 Mon Kim, Taedong, MS26, 11:45 Tue Lee, Ilbin, MS116, 6:00 Thu Lodi, Andrea, MS4, 11:00 Mon Kim, Youngdae, MS26, 10:45 Tue Lee, Jinkun, CP20, 5:30 Wed Loh, Yueling, MS80, 3:00 Wed Kirches, Christian, MS27, 11:15 Tue Lee, Jon, MS54, 5:00 Tue Long, Minhua, CP19, 5:30 Wed Kitahara, Tomonari, CP3, 4:50 Mon Lee, Jon, MS95, 2:00 Thu Long, Troy, MS42, 4:00 Tue Klep, Igor, MS51, 6:30 Tue Lee, Yin Tat, MS114, 4:30 Thu Lopez, Rafael H., PP1, 8:00 Mon Kocvara, Michal, MS18, 2:00 Mon Letocart, Lucas, MS33, 11:45 Tue Lotfi, Reza, CP12, 4:30 Mon Kocvara, Michal, MS18, 3:00 Mon Levi, Retsef, IP2, 1:00 Mon Lotz, Martin, MS92, 9:30 Thu Kocvara, Michal, PP1, 8:00 Mon Levi, Retsef, MS17, 2:00 Mon Lourenço, Bruno F., MS100, 3:30 Thu Kolda, Tamara G., MS63, 11:00 Wed Levi, Retsef, MS17, 2:30 Mon Louveaux, Quentin, MS54, 6:30 Tue Konecny, Jakub, CP6, 5:50 Mon Leyffer, Sven, MS8, 11:00 Mon Low, Steven, MS69, 11:00 Wed Koster, Arie, MS81, 3:30 Wed Leyffer, Sven, MT2, 9:30 Wed Lu, Zhaosong, MS43, 2:30 Tue Kostina, Ekaterina, MS91, 10:00 Thu Leyffer, Sven, MT2, 9:30 Wed Lu, Zhaosong, MS90, 9:30 Thu Kouri, Drew P., MS20, 3:00 Mon Li, Xiang, MS48, 6:30 Tue Lu, Zhaosong, MS90, 10:00 Thu Krislock, Nathan, MS33, 11:15 Tue Li, Xiaobo, MS117, 4:30 Thu Lu, Zhaosong, MS102, 2:00 Thu Krityakierne, Tipaluck, CP10, 4:30 Mon Liang, Chen, CP18, 5:30 Wed Luedtke, James, MT2, 9:30 Wed Kriwet, Gregor, CP7, 5:30 Mon Liang, Jingwei, MS7, 10:30 Mon Luedtke, James, MT2, 9:30 Wed Krokhmal, Pavlo, MS12, 3:00 Wed Liberti, Leo, MS21, 3:00 Mon Luedtke, James, MS87, 9:30 Thu Kucukyavuz, Simge, MS4, 10:30 Mon Liberti, Leo, PD1, 12:30 Tue Luke, Russell, MS82, 2:00 Wed Kunnumkal, Sumit, MS117, 6:00 Thu Liers, Frauke, MS81, 2:00 Wed Luke, Russell, MS82, 2:00 Wed Kusiak, Andrew, MS68, 10:30 Wed Liers, Frauke, MS81, 2:30 Wed Luo, Zhi-Quan, MS43, 4:00 Tue Lim, Lek-Heng, MS86, 9:30 Thu Lim, Lek-Heng, MS98, 2:00 Thu M L Ma, Shiqian, MS66, 9:30 Wed Lahrichi, Nadia, MS1, 9:30 Mon Lim, Lek-Heng, MS110, 4:30 Thu Maghrebi, Mojtaba, CP24, 4:50 Wed Lahrichi, Nadia, MS1, 9:30 Mon Lin, Chih-Jen, MS90, 9:30 Thu Mahajan, Ashutosh, MS28, 11:45 Tue Lambe, Andrew, PP1, 8:00 Mon Lin, Fu, CP5, 6:10 Mon Mahdavi, Mehrdad, MS113, 5:30 Thu Lamm, Michael, MS2, 3:30 Mon Lin, Qihang, MS66, 10:00 Wed Mahey, Philippe, MS70, 9:30 Wed Lan, Guanghui, MS66, 10:30 Wed Linderoth, Jeff, MS28, 11:15 Tue Mairal, Julien, MS113, 6:00 Thu Larson, Jeffrey M., CP10, 5:50 Mon Linderoth, Jeff, MT2, 9:30 Wed Maleki, Arian, MS71, 3:30 Wed Lasserre, Jean B., MS39, 2:30 Tue Linderoth, Jeff, MT2, 9:30 Wed Lasserre, Jean B., MS39, 2:30 Tue Linderoth, Jeff, MS87, 9:30 Thu Maravelias, Christos, CP17, 6:10 Wed Lasserre, Jean B., MS51, 5:00 Tue Lisser, Abdel, MS57, 5:00 Tue Martinez, Gabriela, MS9, 10:30 Mon

continued on next page 188 2014 SIAM Conference on Optimization

Martins, Joaquim, IP7, 8:15 Thu N P Martins, Joaquim, MS5, 9:30 Mon Nakata, Kazuhide, MS88, 10:00 Mon Pait, Felipe M., CP12, 5:50 Mon Mashayekhi, Somayeh, CP21, 5:30 Wed Ñanco, Linco J., CP17, 5:30 Wed Palacios, Francisco, MS103, 3:00 Thu Matni, Nikolai, MS69, 9:30 Wed Naoum-Sawaya, Joe, CP11, 5:50 Mon Palma, Vryan Gil, CP20, 5:10 Wed Mehrotra, Sanjay, MS3, 9:30 Mon Narushima, Yasushi, CP1, 5:30 Mon Pang, C.H. Jeffrey, MS97, 2:00 Thu Mehrotra, Sanjay, MS3, 9:30 Mon Naumann, Uwe, MS11, 11:00 Mon Pang, C.H. Jeffrey, MS97, 3:00 Thu Meinlschmidt, Hannes, CP7, 5:10 Mon Navasca, Carmeliza, MS86, 11:00 Thu Panigrahi, Debmalya, MS93, 10:30 Thu Meisami, Amirhossein, CP13, 5:50 Mon Necoara, Ion, MS90, 10:30 Thu Papadimitriou, Dimitri, CP12, 6:10 Mon Mengi, Emre, CP19, 4:50 Wed Nedich, Angelia, MS50, 5:00 Tue Pape, Susanne, CP4, 4:30 Mon Mesbahi, Mehran, MS93, 11:00 Thu Nedich, Angelia, MS50, 6:30 Tue Papp, David, MS3, 10:00 Mon Meunier, Frédéric, MS56, 6:30 Tue Negahban, Sahand, MS78, 3:00 Wed Parpas, Panos, CP9, 6:10 Mon Meza, Juan, PD1, 12:30 Tue Neitzel, Ira, MS118, 4:30 Thu Parrilo, Pablo A., MS76, 2:00 Wed Mitchell, John E., CP1, 4:50 Mon Nerantzis, Dimitrios, CP26, 4:50 Wed Pasupathy, Raghu, MS16, 3:00 Mon Mitchell, Tim, MS99, 2:30 Thu Nesterov, Yurii, IP1, 8:15 Mon Pataki, Gabor, MS77, 3:30 Wed Miyashiro, Ryuhei, MS88, 11:00 Mon Nesterov, Yurii, MS31, 10:15 Tue Patev, Konstantin, CP6, 5:30 Mon Mizuno, Shinji, MS56, 6:00 Tue Nie, Jiawang, MS39, 2:30 Tue Patrascu, Andrei, MS102, 2:30 Thu Mizutani, Tomohiko, CP16, 4:50 Wed Nie, Jiawang, MS51, 5:00 Tue Pearson, John W., CP8, 5:50 Mon Moallemi, Ciamac C., MS13, 2:30 Mon Nie, Jiawang, MS51, 5:00 Tue Pena, Javier, MS66, 9:30 Wed Moazeni, Somayeh, CP6, 4:50 Mon Nie, Jiawang, MS86, 9:30 Thu Pena, Javier, MS92, 9:30 Thu Modaresi, Sina, MS87, 11:00 Thu Nie, Jiawang, MS98, 2:00 Thu Pena, Javier, MS92, 11:00 Thu Moffat, Sarah, MS85, 11:00 Thu Nie, Jiawang, MS110, 4:30 Thu Peng, Zhimin, MS105, 3:00 Thu Montanher, Tiago M., CP10, 5:30 Mon Nishi, Masataka, CP23, 5:30 Wed Permenter, Frank, MS89, 10:30 Thu Monteiro, Renato C., MS65, 9:30 Wed Nocedal, Jorge, MS41, 2:30 Tue Petra, Cosmin G., MS36, 3:00 Tue Moon, Heather A., CP19, 5:50 Wed O Peyrega, Mathilde, PP1, 8:00 Mon Morales, Jose Luis, MS99, 3:00 Thu Oddsdóttir, Hildur Æsa, CP11, 4:50 Pfaff, Sebastian, CP7, 4:30 Mon Mordukhovich, Boris, MS85, 10:30 Thu Mon Phan, Dzung, CP6, 6:10 Mon Morini, Benedetta, MS46, 3:00 Tue Odland, Tove, MS46, 3:30 Tue Phan, Hung M., MS61, 11:00 Wed Moursi, Walaa, MS82, 3:30 Wed Oeding, Luke, MS86, 9:30 Thu Philipp, Anne, CP25, 4:50 Wed Mu, Cun, MS63, 10:00 Wed Omheni, Riadh, CP22, 5:50 Wed Phillips, Cynthia, IP4, 9:00 Tue Muller, Christophe, PP1, 8:00 Mon O’Neill, Zheng, MS68, 11:00 Wed Philpott, Andy, IP6, 1:00 Wed Müller, Johannes C., CP1, 5:50 Mon Orban, Dominique, MS46, 2:30 Tue Philpott, Andy, MS84, 9:30 Thu Munson, Todd, MS26, 10:15 Tue Orban, Dominique, MS58, 5:00 Tue Philpott, Andy, MS96, 2:00 Thu Munson, Todd, MS26, 10:15 Tue Orban, Dominique, MS58, 5:00 Tue Philpott, Andy, MS108, 4:30 Thu Muramatsu, Masakazu, MS100, 2:00 Thu Ordonez, Fernando, CP23, 4:30 Wed Pilanci, Mert, MS35, 11:15 Tue Mut, Murat, CP25, 5:50 Wed Oren, Shmuel, MS108, 4:30 Thu Pong, Ting Kei, MS100, 3:00 Thu Oren, Shmuel, MS108, 6:00 Thu Popescu, Nicusor, MS115, 4:30 Thu Overton, Michael L., MS64, 9:30 Wed Post, Ian, MS56, 5:30 Tue Overton, Michael L., MS64, 9:30 Wed Postek, Krzysztof, CP15, 5:10 Wed

continued on next page 2014 SIAM Conference on Optimization 189

Powell, Michael, MS53, 5:00 Tue Romeijn, Edwin, MS49, 5:00 Tue Santos, Francisco, MS44, 2:30 Tue Powell, Warren, MS24, 10:15 Tue Romeijn, Edwin, MS119, 2:00 Wed Santos, Francisco, MS56, 5:00 Tue Puerto, Justo, CP4, 6:10 Mon Romeijn, Edwin, MS119, 2:00 Wed Santos, Luiz-Rafael, CP3, 5:30 Mon Puterman, Martin L., MS1, 10:30 Mon Romero, Julián, CP16, 5:30 Wed Sartenaer, Annick, MS46, 2:30 Tue Romero Yanez, Gonzalo, MS17, 2:00 Sartenaer, Annick, MS58, 5:00 Tue Q Mon Sato, Hiroyuki, CP9, 5:30 Mon Qi, Houduo, MS65, 10:00 Wed Roosta-Khorasani, Farbod, MS91, 11:00 Saunders, Michael A., MS34, 11:45 Tue Qin, Tai, MS52, 6:00 Tue Thu Saunderson, James, MS89, 9:30 Thu R Roshchina, Vera, MS92, 10:00 Thu Saunderson, James, MS89, 9:30 Thu Raghunathan, Arvind, MS48, 5:30 Tue Roupin, Frédéric, MS40, 3:30 Tue Saxena, Anureet, MS54, 5:30 Tue Royset, Johannes O., MS67, 9:30 Wed Ralph, Daniel, PD1, 12:30 Tue Schäfer, Carsten, CP20, 4:50 Wed Ralph, Daniel, MS84, 9:30 Thu Royset, Johannes O., MS67, 11:00 Wed Scheinberg, Katya, MS16, 2:00 Mon Ralph, Daniel, MS96, 2:00 Thu Ruiz, Daniel, MS46, 4:00 Tue Scheinberg, Katya, MS16, 2:00 Mon Ralph, Daniel, MS96, 3:00 Thu Rujeerapaiboon, Napat, MS62, 10:00 Schelbert, Jakob, CP11, 5:10 Mon Wed Ralph, Daniel, MS108, 4:30 Thu Scherrer, Bruno, MS79, 2:30 Wed Ruszczynski, Andrzej, MS74, 2:00 Wed Recht, Ben, MS107, 6:00 Thu Schewe, Lars, MS22, 2:00 Mon Ruszczynski, Andrzej, MS74, 2:00 Wed Rendl, Franz, IP8, 1:00 Thu Schewe, Lars, MS22, 2:00 Mon Ryan, Chris, MS104, 2:00 Thu Rendl, Franz, MS33, 10:15 Tue Schichl, Hermann, CP19, 5:10 Wed Ryan, Chris, MS116, 4:30 Thu Renegar, James, MS44, 3:30 Tue Schmidt, Daniel, MS81, 3:00 Wed Ryan, Chris, MS116, 4:30 Thu Renner, Philipp, CP23, 6:10 Wed Schmidt, Mark, MS78, 2:00 Wed Reznick, Bruce, MS64, 10:30 Wed S Schmidt, Mark, MS113, 4:30 Thu Richtarik, Peter, MS31, 10:15 Tue Sabach, Shoham, MS97, 2:00 Thu Schmidt, Martin, MS22, 2:30 Mon Richtarik, Peter, MS31, 10:45 Tue Sabartova, Zuzana, PP1, 8:00 Mon Schmidt, Stephan, MS103, 2:00 Thu Richtarik, Peter, MS43, 2:30 Tue Sachs, Ekkehard W., MS34, 11:15 Tue Schmidt, Stephan, MS115, 4:30 Thu Richtarik, Peter, MS55, 5:00 Tue Sadok, Hassane, PP1, 8:00 Mon Schneider, Christopher, CP21, 4:30 Wed Richtarik, Peter, MS90, 9:30 Thu Sagastizabal, Claudia A., MS84, 9:30 Thu Schultz, Ruediger, MS9, 9:30 Mon Richtarik, Peter, MS102, 2:00 Thu Sagastizabal, Claudia A., MS84, 11:00 Schultz, Ruediger, MS57, 5:00 Tue Richtarik, Peter, MS114, 4:30 Thu Thu Schulz, Volker H., MS37, 2:30 Tue Rigollet, Philippe, MS71, 3:00 Wed Sagastizabal, Claudia A., MS96, 2:00 Thu Schulz, Volker H., MS37, 4:00 Tue Robinson, Daniel, MS8, 9:30 Mon Sagastizabal, Claudia A., MS108, 4:30 Thu Schütte, Maria, MS103, 3:30 Thu Robinson, Daniel, MS27, 10:15 Tue Sagnol, Guillaume, MS101, 3:00 Thu Sciandrone, Marco, MS23, 2:00 Mon Robinson, Daniel, MS41, 2:30 Tue Salari, Ehsan, MS42, 3:00 Tue Sekiguchi, Yoshiyuki, MS100, 2:30 Thu Robinson, Daniel, MS53, 5:00 Tue Samaranayake, Samitha, MS112, 4:30 Thu Seydenschwanz, Martin, CP21, 4:50 Robinson, Daniel, MS80, 2:00 Wed Samaranayake, Samitha, MS112, 4:30 Wed Robinson, Richard, MS89, 10:00 Thu Thu Shah, Parikshit, MS69, 9:30 Wed Rockafellar, Terry, MS67, 9:30 Wed Sampaio, Phillipe R., CP14, 5:10 Wed Shah, Parikshit, MS69, 10:30 Wed Roemisch, Werner, MS9, 10:00 Mon Santiago, Claudio, CP2, 4:30 Mon Shah, Parikshit, MS71, 2:00 Wed Rojas, Marielba, PP1, 8:00 Mon Santos, Francisco, IP3, 8:15 Tue Shalit, Uri, MS90, 11:00 Thu Santos, Francisco, MS32, 10:15 Tue

continued on next page 190 2014 SIAM Conference on Optimization

Shams, Issac, CP24, 5:50 Wed T V Shanbhag, Uday, MS50, 5:00 Tue Takac, Martin, MS55, 5:00 Tue Van Ackooij, Wim, MS84, 9:30 Thu Shen, Siqian, MS1, 11:00 Mon Tanaka, Akihiro, MS100, 2:00 Thu van den Doel, Kees, MS7, 9:30 Mon Shi, Cong, MS17, 3:30 Mon Tang, Xiaocheng, MS102, 3:00 Thu Vanderbei, Robert J., MS6, 11:00 Mon Shim, Sangho, CP4, 5:30 Mon Tannier, Chalotte, MS46, 2:30 Tue Vandereycken, Bart, MS110, 6:00 Thu Shishkin, Serge L., CP10, 5:10 Mon Tanveer, Mohammad, CP5, 5:30 Mon Vanderwerff, Jon, MS38, 3:00 Tue Silva, Francisco J., MS15, 2:00 Mon Tappenden, Rachael, MS114, 5:30 Thu Vannieuwenhoven, Nick, MS110, 5:00 Silva, Francisco J., CP8, 6:10 Mon Tawarmalani, Mohit, MS111, 4:30 Thu Thu Silva, Paulo J. S., MS53, 6:00 Tue Terlaky, Tamas, MS44, 2:30 Tue Vavasis, Stephen A., MS75, 2:00 Wed Sim, Chee-Khian, CP16, 5:10 Wed Terlaky, Tamas, MS12, 2:00 Wed Vavasis, Stephen A., CP25, 6:10 Wed Sim, Melvyn, MS62, 11:00 Wed Tewari, Ambuj, MS55, 6:00 Tue Vera, Jorge R., CP15, 6:10 Wed Simons, Stephen, MS38, 2:30 Tue Thiedau, Jan, MS22, 3:00 Mon Vielma, Juan Pablo, MS28, 10:15 Tue Singh, Aarti, MS52, 6:30 Tue Thomas, Rekha, IP5, 8:15 Wed Vielma, Juan Pablo, MS95, 3:00 Thu Soheili, Negar S., MS92, 10:30 Thu Todd, Michael J., CP1, 4:30 Mon Vierhaus, Ingmar, CP8, 4:30 Mon Solntsev, Stefan, MS16, 3:30 Mon Toh, Kim-Chuan, MS65, 9:30 Wed W Solodov, Mikhail V., MS80, 2:30 Wed Toh, Kim-Chuan, MS65, 10:30 Wed Waechter, Andreas, MS27, 10:15 Tue Sondjaja, Mutiara, CP2, 4:50 Mon Toint, Philippe L., MS8, 9:30 Mon Waechter, Andreas, MS41, 2:30 Tue Sorber, Laurent, MS86, 10:00 Thu Toraldo, Gerardo, MS7, 9:30 Mon Waechter, Andreas, MS53, 5:00 Tue Sotirov, Renata, MS30, 10:15 Tue Toraldo, Gerardo, MS19, 2:00 Mon Waechter, Andreas, MS80, 2:00 Wed Sotirov, Renata, MS30, 10:15 Tue Tran, Nghia, MS38, 3:30 Tue Waechter, Andreas, MS99, 2:00 Thu Sreekumaran, Harikrishnan, CP23, 4:50 Traversi, Emiliano, MS95, 2:30 Thu Walther, Andrea, MS99, 2:00 Thu Wed Treister, Eran, CP19, 4:30 Wed Wang, Hao, MS41, 3:30 Tue Stadler, Georg, MS20, 2:30 Mon Troeltzsch, Anke, CP14, 5:30 Wed Wang, Li, PP1, 8:00 Mon Steffensen, Sonja, MS2, 3:00 Mon Trosset, Michael W., MS60, 10:30 Wed Wang, Li, MS98, 3:30 Thu Steinbach, Marc C., MS22, 3:30 Mon Trumbauer, Eric, PP1, 8:00 Mon Wang, Mengdi, MS50, 5:30 Tue Stern, Matt, MS104, 2:00 Thu Truong, Bao Q., MS94, 10:30 Thu Wang, Qiqi, MS103, 2:00 Thu Stern, Matt, MS104, 2:00 Thu Tsigaridas, Elias, MS98, 2:30 Thu Wang, Qiqi, MS103, 2:00 Thu Stern, Matt, MS116, 4:30 Thu Turner, James, MS18, 3:30 Mon Wang, Qiqi, MS115, 4:30 Thu Stibbe, Hilke, CP7, 6:10 Mon Wang, Shawn Xianfu, MS38, 2:30 Tue Stingl, Michael, MS18, 2:00 Mon U Wang, Shawn Xianfu, MS38, 4:00 Tue Ulbrich, Michael, MS8, 10:00 Mon Stingl, Michael, MS37, 3:00 Tue Wang, Xiao, MS10, 11:00 Mon Ulbrich, Michael, MS20, 2:00 Mon Stockbridge, Rebecca, CP18, 4:50 Wed Watson, Jean-Paul, MS24, 11:45 Tue Ulbrich, Michael, MS2, 2:00 Mon Sun, Cong, MS10, 10:00 Mon Wen, Jinming, CP4, 5:50 Mon Ulbrich, Stefan, MS8, 10:30 Mon Sun, Defeng, MS107, 5:30 Thu Wen, Zaiwen, MS59, 5:00 Tue Ulbrich, Stefan, MS20, 2:00 Mon Sun, Yifan, MS18, 2:30 Mon Wen, Zaiwen, MS59, 5:00 Tue Ulus, Firdevs, CP26, 4:30 Wed Sun, Zhao, MS76, 3:00 Wed Wets, Roger J.-B., MS109, 6:00 Thu Unkelbach, Jan, MS49, 6:30 Tue Surowiec, Thomas M., MS2, 2:30 Mon Wiecek, Margaret, MS94, 10:00 Thu Uryasev, Stan, MS67, 10:30 Wed Suzuki, Taiji, MS31, 11:15 Tue Svärd, Henrik, CP23, 5:10 Wed

continued on next page 2014 SIAM Conference on Optimization 191

Wiegele, Angelika, MS33, 10:15 Tue Yashtini, Maryam, MS105, 2:00 Thu Zinchenko, Yuriy, MS44, 3:00 Tue Wild, Stefan, MS27, 11:45 Tue Ye, Jane, MS45, 2:30 Tue Zorn, Heinz, MS37, 2:30 Tue Wogrin, Sonja, MS96, 2:30 Thu Ye, Jane J., MS45, 2:30 Tue Zorn, Heinz, MS37, 2:30 Tue Wolfhagen, Eli, MS9, 11:00 Mon Ye, Yinyu, SP1, 1:30 Tue Zuluaga, Luis, MS30, 11:15 Tue Wolkowicz, Henry, MS64, 9:30 Wed Ye, Yinyu, MS93, 9:30 Thu Wolkowicz, Henry, MS77, 2:00 Wed Yelbay, Belma, CP26, 5:50 Wed Wolkowicz, Henry, MS77, 2:00 Wed Yin, Wotao, MS50, 6:00 Tue Wollenberg, Nadine, CP5, 4:30 Mon Yoshise, Akiko, MS100, 2:00 Thu Wollenberg, Tobias, MS57, 6:30 Tue Yu, Jia Yuan, MS112, 5:30 Thu Wollner, Winnifried, MS118, 5:00 Thu Yuan, Ya-Xiang, MS10, 9:30 Mon Wolpert, David, MS60, 9:30 Wed Yuan, Ya-Xiang, PD1, 12:30 Tue Wong, Elizabeth, MS83, 10:00 Thu Yuan, Ya-Xiang, MS53, 6:30 Tue Woo, Hyenkyun, MS63, 9:30 Wed Wright, Stephen, MS71, 2:30 Wed Z Zahr, Matthew J., CP13, 6:10 Mon Wu, Ke, CP13, 5:10 Mon Zakeri, Golbon, MS96, 2:00 Thu Wu, Victor, MS42, 3:30 Tue Zakeri, Golbon, MS96, 3:30 Thu Wurst, Jan-Eric, MS118, 5:30 Thu Zanella, Riccardo, MS19, 2:00 Mon X Zanni, Luca, MS7, 9:30 Mon Xia, Yu, CP3, 4:30 Mon Zanni, Luca, MS19, 2:00 Mon Xiao, Lin, MS31, 10:15 Tue Zarin Pass, Rebecca, PP1, 8:00 Mon Xiao, Lin, MS43, 2:30 Tue Zavala, Victor, MS24, 10:15 Tue Xiao, Lin, MS43, 2:30 Tue Zavala, Victor, MS36, 2:30 Tue Xiao, Lin, MS55, 5:00 Tue Zavala, Victor, MS48, 5:00 Tue Xin, Linwei, CP15, 5:30 Wed Zavala, Victor, MS48, 5:00 Tue Xu, Fengmin, MS19, 3:30 Mon Zehtabian, Shohre, CP18, 5:50 Wed Xu, Jia, MS85, 11:00 Thu Zemkoho, Alain B., CP3, 5:10 Mon Xu, Yangyang, MS63, 9:30 Wed Zeng, Bo, MS84, 10:00 Thu Xu, Yangyang, CP16, 4:30 Wed Zhang, Hongchao, MS53, 5:30 Tue Zhang, Hongchao, MS105, 2:00 Thu Y Zhang, Shuzhong, MS110, 4:30 Thu Yamashita, Makoto, MS88, 9:30 Mon Zhang, Tong, MS31, 11:45 Tue Yamashita, Makoto, MS88, 9:30 Mon Zhang, Xinzhen, MS110, 5:30 Thu Yan, Ming, MS63, 10:30 Wed Zhang, Yin, MS59, 6:30 Tue Yang, Chao, MS59, 6:00 Tue Zhang, Zaikun, MS114, 6:00 Thu Yang, Liuqin, MS65, 11:00 Wed Zhao, Hongmei, CP17, 4:30 Wed Yao, Liangjin, MS73, 2:00 Wed Zhen, Jianzhe, CP15, 4:50 Wed Yarmand, Hamed, CP24, 5:10 Wed Zhou, Wenwen, CP22, 5:30 Wed Yashtini, Maryam, MS105, 2:00 Thu 192 2014 SIAM Conference on Optimization

Notes 2014 SIAM Conference on Optimization 193

OP14 Budget

Conference Budget SIAM Conference on Optimization May 19 - 22, 2014 San Diego, CA

Expected Paid Attendance 550

Revenue Registration Income $178,300 Total $178,300

Expenses Printing $4,600 Organizing Committee $4,500 Invited Speakers $12,800 Food and Beverage $38,800 AV Equipment and Telecommunication $23,900 Advertising $8,000 Proceedings $0 Conference Labor (including benefits) $50,861 Other (supplies, staff travel, freight, misc.) $12,800 Administrative $18,184 Accounting/Distribution & Shipping $8,928 Information Systems $15,966 Customer Service $5,913 Marketing $9,183 Office Space (Building) $5,014 Other SIAM Services $5,684 Total $225,133

Net Conference Expense -$46,833

Support Provided by SIAM $46,833 $0

Estimated Support for Travel Awards not included above: Students and Early Career 56 $39,600

Town and Country Resort & Convention Center

FSC logo text box indicating size & layout of logo. Conlins to insert logo.