DOCSLIB.ORG

Optimization Basic Results

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Optimization Basic Results Optimization Basic Results Michel De Lara Cermics, Ecole´ des Ponts ParisTech France Ecole´ des Ponts ParisTech November 22, 2020 Outline of the presentation Magic formulas Convex functions, coercivity Existence and uniqueness of a minimum First-order optimality conditions (the case of equality constraints) Duality gap and saddle-points Elements of Lagrangian duality and Uzawa algorithm More on convexity and duality Outline of the presentation Magic formulas Convex functions, coercivity Existence and uniqueness of a minimum First-order optimality conditions (the case of equality constraints) Duality gap and saddle-points Elements of Lagrangian duality and Uzawa algorithm More on convexity and duality inf h(a; b) = inf inf h(a; b) a2A;b2B a2A b2B inf λf (a) = λ inf f (a) ; 8λ ≥ 0 a2A a2A inf f (a) + g(b) = inf f (a) + inf g(b) a2A;b2B a2A b2B Tower formula For any function h : A × B ! [−∞; +1] we have inf h(a; b) = inf inf h(a; b) a2A;b2B a2A b2B and if B(a) ⊂ B, 8a 2 A, we have inf h(a; b) = inf inf h(a; b) a2A;b2B(a) a2A b2B(a) Independence For any functions f : A !] − 1; +1] ; g : B !] − 1; +1] we have inf f (a) + g(b) = inf f (a) + inf g(b) a2A;b2B a2A b2B and for any finite set S, any functions fs : As !] − 1; +1] and any nonnegative scalars πs ≥ 0, for s 2 S, we have X X inf π f (a )= π inf f (a ) Q s s s s s s fas gs2 2 s2 As as 2As S S s2S s2S Outline of the presentation Magic formulas Convex functions, coercivity Existence and uniqueness of a minimum First-order optimality conditions (the case of equality constraints) Duality gap and saddle-points Elements of Lagrangian duality and Uzawa algorithm More on convexity and duality Convex sets Let N 2 N∗. We consider subsets of the Euclidian space RN N I The subset C ⊂ R is convex if for any x1 2 C, x2 2 C and t 2 [0; 1], we have that tx1 + (1 − t)x2 2 C I An intersection of convex sets is convex I A segment is convex I A hyperplane is convex( H ⊂ RN is a hyperplane if there exists N N y 2 R nf0g and b 2 R such that H = x 2 R hx ; yi + b = 0 ) I An affine subspace (intersection of hyperplanes) is convex Linear and affine functions Consider a function f : RN ! R N N I The function f is linear if, for any x1 2 R , x2 2 R and t1 2 R, t2 2 R, f (t1x1 + t2x2) = t1f (x1) + t2f (x2) N N I The function f is affine if, for any x1 2 R , x2 2 R and t 2 R, f (tx1 + (1 − t)x2) = tf (x1) + (1 − t)f (x2) Exercise. Show that f : RN ! R is affine if and only if g(x) = f (x) − f (0) is linear Convex functions (definitions) Let C ⊂ RN be an nonempty convex subset of RN , where N 2 N∗, and f : C ! R be a function I The function f is affine if, for any x1 2 C, x2 2 C and t 2 [0; 1], f (tx1 + (1 − t)x2) = tf (x1) + (1 − t)f (x2) I The function f is convex if, for any x1 2 C, x2 2 C and t 2 [0; 1], f (tx1 + (1 − t)x2) ≤ tf (x1) + (1 − t)f (x2) I The function f is strictly convex if, for any x1 2 C, x2 2 C, x1 6= x2, and t 2]0; 1[, f (tx1 + (1 − t)x2) < tf (x1) + (1 − t)f (x2) I The function f is strongly convex (of modulus a > 0) if, for any x1 2 C, x2 2 C and t 2 [0; 1], a f (tx + (1 − t)x ) ≤ tf (x ) + (1 − t)f (x ) − t(1 − t)kx − x k2 1 2 1 2 2 1 2 Exercises Let C ⊂ RN be an nonempty subset of RN , where N 2 N∗ I Show that both definitions of an affine function coincide when C = RN I Show that a function f : C ! R is convex if and only if its epigraph is a convex set subset of RN × R I Show that a function f : C ! R is strongly convex of modulus a > 0 a 2 if and only if g(x) = f (x) − 2 kxk is convex I If f : C ! R is convex, show that f is not strictly convex if and only there exists a nonempty convex subset C 0 ⊂ C over which f is affine Convex functions on the real line Proposition Let I ⊂ R be an nonempty interval I A C 1 function f : I ! R is convex if and only if f 0 is increasing on I I A C 2 function f : I ! R is convex if and only if f 00(x) ≥ 0, for all x 2 I I Let a > 0.A C 2 function f : I ! R is a-strongly convex if and only if f 00(x) ≥ a, for all x 2 I I A C 1 function f : I ! R is strictly convex if and only if f is convex and the set fx 2 I j f 00(x) = 0g is either empty or is a singleton Exercise. Study the family of functions fα :]0; +1[! R given by α fα(x) = x . For which values of the parameter α is the function fα convex? For a given a > 0, for which values of the parameter α is the function fα strongly convex of modulus a? Provide an example of a strictly convex function which is not strongly convex. Convexity for multivariate functions 2 N The Hessian matrix Hf (x) of a C function f : R ! R is the N × N symmetric matrix given by @2f Hf (x) = i j (x) @x @x (i;j)2f1;:::;Ng2 Proposition Let C ⊂ RN be an nonempty convex subset of RN , where N 2 N∗ I A C 2 function f : RN ! R is convex on C if and only if the symmetric Hessian matrix Hf (x) is positive for all x 2 C I A C 2 function f : RN ! R is strongly convex of modulus a > 0 on C if and only if the eigenvalues of the symmetric Hessian matrix Hf (x) are uniformly bounded below by a > 0 on C Exercise. Let Q be a N × N symmetric matrix and f (x) = 1=2x 0Qx, where x 0 is the transpose of the vector x. Give conditions on the smallest eigenvalue of Q so that the function f is convex, or strictly convex, or strongly convex of modulus a. Operations on functions preserving convexity Proposition Let (fi )i2I be a family of convex functions Then supi2I fi is a convex function Proposition Let (fi )i=1;:::;n be convex functions Let (αi )i=1;:::;n be nonnegative numbers Pm Then i=1 αi fi is a convex function Proposition Let f : RN ! R be convex Let A be a N × M matrix and b 2 RN Then y 2 RM 7! f (Ay + b) is a convex function Coercivity Definition A function f : RN ! R is coercive if lim f (x) = +1 kxk!+1 Proposition A strongly convex function is coercive Outline of the presentation Magic formulas Convex functions, coercivity Existence and uniqueness of a minimum First-order optimality conditions (the case of equality constraints) Duality gap and saddle-points Elements of Lagrangian duality and Uzawa algorithm More on convexity and duality Here are the ingredients for a general abstract optimization problem inf J(u) ad u2U I Optimization set U(= RN ) containing optimization variables u 2 U I A criterion J : U ! R [ f+1g I Constraints of the form u 2 Uad ⊂ U Examples of classes of optimization problems inf J(u) ad u2U I Linear programming N I Optimization set U = R I Criterion J is linear (affine) ad I Constraints U defined by a finite number of linear (affine) equalities and inequalities I Convex optimization I Criterion J is a convex function ad I Constraints U define a convex set I Combinatorial optimization N N I Optimization set U is discrete (binary f0; 1g , integer Z , etc.) Minimum Definition We say that u∗ 2 U is a (global) minimum of the optimization problem inf ad J(u) if u2U ∗ ad ∗ ad u 2 U and J(u ) ≤ J(u) ; 8u 2 U In this case, we write J(u∗) = min J(u) ad u2U Existence and uniqueness of a minimum We consider the optimization problem inf J(u) where ad ⊂ = N ad U U R u2U Proposition If the criterion J is continuous and the constraint set Uad is compact (bounded and closed), then there is a minimum Proposition If the constraint set Uad is closed and the criterion J is continuous and coercive, then there is a minimum Proposition If the constraint set Uad is convex, and if the criterion J is strictly convex, a minimum is necessarily unique Exercises We consider the optimization problem inf J(u) ad u2U Give an example I of continuous criterion J and of constraint set Uad for which there is no minimum I of criterion J and of compact constraint set Uad for which there is no minimum I of continuous criterion J and of unbounded and closed constraint set Uad for which there is no minimum I of convex criterion J and of constraint set Uad for which there is more than one minimum I of strictly convex criterion J and of constraint set Uad for which there is more than one minimum Local minimum Definition We say that u∗ 2 U is a local minimum of the optimization problem ∗ ad inf ad J(u) if there exists a neighborhood V of u in such that u2U U ∗ ad ∗ u 2 U and J(u ) ≤ J(u) ; 8u 2 V Proposition If the constraint set Uad is convex, and if the criterion J is convex, a local minimum is a global minimum Outline of the presentation Magic formulas Convex functions, coercivity Existence and uniqueness of a minimum First-order optimality conditions (the case of equality constraints) Duality gap and saddle-points Elements of Lagrangian duality and Uzawa algorithm More on convexity and duality Optimization under equality constraints We consider the optimization problem inf J(u) N u2R under the constraint Θ(u) = 0 where Θ is a function with values in RM N M Θ = (Θ1;:::; ΘM ): R ! R whose components are denoted by Θj , where j runs from 1 to M Sufficient condition for qualification in case of equality constraints Definition Let u∗ 2 RN .
Load more

Recommended publications
  • Turbo Decoding As Iterative Constrained Maximum Likelihood Sequence Detection John Maclaren Walsh, Member, IEEE, Phillip A
    1 Turbo Decoding as Iterative Constrained Maximum Likelihood Sequence Detection John MacLaren Walsh, Member, IEEE, Phillip A. Regalia, Fellow, IEEE, and C. Richard Johnson, Jr., Fellow, IEEE Abstract— The turbo decoder was not originally introduced codes have good distance properties, which would be relevant as a solution to an optimization problem, which has impeded for maximum likelihood decoding, researchers have not yet attempts to explain its excellent performance. Here it is shown, succeeded in developing a proper connection between the sub- nonetheless, that the turbo decoder is an iterative method seeking a solution to an intuitively pleasing constrained optimization optimal turbo decoder and maximum likelihood decoding. This problem. In particular, the turbo decoder seeks the maximum is exacerbated by the fact that the turbo decoder, unlike most of likelihood sequence under the false assumption that the input the designs in modern communications systems engineering, to the encoders are chosen independently of each other in the was not originally introduced as a solution to an optimization parallel case, or that the output of the outer encoder is chosen problem. This has made explaining just why the turbo decoder independently of the input to the inner encoder in the serial case. To control the error introduced by the false assumption, the opti- performs as well as it does very difficult. Together with the mizations are performed subject to a constraint on the probability lack of formulation as a solution to an optimization problem, that the independent messages happen to coincide. When the the turbo decoder is an iterative algorithm, which makes the constraining probability equals one, the global maximum of determining its convergence and stability behavior important.
    [Show full text]
  • A Comparative Study of Time Delay Estimation Techniques for Road Vehicle Tracking Patrick Marmaroli, Xavier Falourd, Hervé Lissek
    A comparative study of time delay estimation techniques for road vehicle tracking Patrick Marmaroli, Xavier Falourd, Hervé Lissek To cite this version: Patrick Marmaroli, Xavier Falourd, Hervé Lissek. A comparative study of time delay estimation techniques for road vehicle tracking. Acoustics 2012, Apr 2012, Nantes, France. hal-00810981 HAL Id: hal-00810981 https://hal.archives-ouvertes.fr/hal-00810981 Submitted on 23 Apr 2012 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. Proceedings of the Acoustics 2012 Nantes Conference 23-27 April 2012, Nantes, France A comparative study of time delay estimation techniques for road vehicle tracking P. Marmaroli, X. Falourd and H. Lissek Ecole Polytechnique F´ed´erale de Lausanne, EPFL STI IEL LEMA, 1015 Lausanne, Switzerland patrick.marmaroli@epfl.ch 4135 23-27 April 2012, Nantes, France Proceedings of the Acoustics 2012 Nantes Conference This paper addresses road traffic monitoring using passive acoustic sensors. Recently, the feasibility of the joint speed and wheelbase length estimation of a road vehicle using particle filtering has been demonstrated. In essence, the direction of arrival of propagated tyre/road noises are estimated using a time delay estimation (TDE) technique between pairs of microphones placed near the road.
    [Show full text]
  • Comparison of Risk Measures
    Geometry Of The Expected Value Set And The Set-Valued Sample Mean Process Alois Pichler∗ May 13, 2018 Abstract The law of large numbers extends to random sets by employing Minkowski addition. Above that, a central limit theorem is available for set-valued random variables. The existing results use abstract isometries to describe convergence of the sample mean process towards the limit, the expected value set. These statements do not reveal the local geometry and the relations of the sample mean and the expected value set, so these descriptions are not entirely satisfactory in understanding the limiting behavior of the sample mean process. This paper addresses and describes the fluctuations of the sample average mean on the boundary of the expectation set. Keywords: Random sets, set-valued integration, stochastic optimization, set-valued risk measures Classification: 90C15, 26E25, 49J53, 28B20 1 Introduction Artstein and Vitale [4] obtain an initial law of large numbers for random sets. Given this result and the similarities of Minkowski addition of sets with addition and multiplication for scalars it is natural to ask for a central limit theorem for random sets. After some pioneering work by Cressie [11], Weil [28] succeeds in establishing a reasonable result describing the distribution of the Pompeiu–Hausdorff distance between the sample average and the expected value set. The result is based on an isometry between compact sets and their support functions, which are continuous on some appropriate and adapted sphere (cf. also Norkin and Wets [20] and Li et al. [17]; cf. Kuelbs [16] for general difficulties). However, arXiv:1708.05735v1 [math.PR] 18 Aug 2017 the Pompeiu–Hausdorff distance of random sets is just an R-valued random variable and its d distribution is on the real line.
    [Show full text]
  • Direct Optimization Through $\Arg \Max$ for Discrete Variational Auto
    Direct Optimization through arg max for Discrete Variational Auto-Encoder Guy Lorberbom Andreea Gane Tommi Jaakkola Tamir Hazan Technion MIT MIT Technion Abstract Reparameterization of variational auto-encoders with continuous random variables is an effective method for reducing the variance of their gradient estimates. In the discrete case, one can perform reparametrization using the Gumbel-Max trick, but the resulting objective relies on an arg max operation and is non-differentiable. In contrast to previous works which resort to softmax-based relaxations, we propose to optimize it directly by applying the direct loss minimization approach. Our proposal extends naturally to structured discrete latent variable models when evaluating the arg max operation is tractable. We demonstrate empirically the effectiveness of the direct loss minimization technique in variational autoencoders with both unstructured and structured discrete latent variables. 1 Introduction Models with discrete latent variables drive extensive research in machine learning applications, including language classification and generation [42, 11, 34], molecular synthesis [19], or game solving [25]. Compared to their continuous counterparts, discrete latent variable models can decrease the computational complexity of inference calculations, for instance, by discarding alternatives in hard attention models [21], they can improve interpretability by illustrating which terms contributed to the solution [27, 42], and they can facilitate the encoding of inductive biases in the learning process, such as images consisting of a small number of objects [8] or tasks requiring intermediate alignments [25]. Finally, in some cases, discrete latent variables are natural choices, for instance when modeling datasets with discrete classes [32, 12, 23]. Performing maximum likelihood estimation of latent variable models is challenging due to the requirement to marginalize over the latent variables.
    [Show full text]
  • Extreme-Value Theorems for Optimal Multidimensional Pricing
    Extreme-Value Theorems for Optimal Multidimensional Pricing Yang Cai∗ Constantinos Daskalakisy Computer Science, McGill University EECS, MIT [email protected] [email protected] October 28, 2014 Abstract We provide a near-optimal, computationally efficient algorithm for the unit-demand pricing problem, where a seller wants to price n items to optimize revenue against a unit-demand buyer whose values for the items are independently drawn from known distributions. For any chosen accuracy > 0 and item values bounded in [0; 1], our algorithm achieves revenue that is optimal up to an additive error of at most , in polynomial time. For values sampled from Monotone Hazard Rate (MHR) distributions, we achieve a (1 − )-fraction of the optimal revenue in poly- nomial time, while for values sampled from regular distributions the same revenue guarantees are achieved in quasi-polynomial time. Our algorithm for bounded distributions applies probabilistic techniques to understand the statistical properties of revenue distributions, obtaining a reduction in the search space of the algorithm through dynamic programming. Adapting this approach to MHR and regular distri- butions requires the proof of novel extreme value theorems for such distributions. As a byproduct, our techniques establish structural properties of approximately-optimal and near-optimal solutions. We show that, when the buyer's values are independently distributed according to MHR distributions, pricing all items at the same price achieves a constant fraction of the optimal revenue. Moreover, for all > 0, at most g(1/) distinct prices suffice to obtain a (1 − )-fraction of the optimal revenue, where g(1/) is a quadratic function of 1/ that does not depend on the number of items.
    [Show full text]
  • Cesifo Working Paper No. 5428 Category 12: Empirical and Theoretical Methods June 2015
    A Service of Leibniz-Informationszentrum econstor Wirtschaft Leibniz Information Centre Make Your Publications Visible. zbw for Economics Aquaro, Michele; Bailey, Natalia; Pesaran, M. Hashem Working Paper Quasi Maximum Likelihood Estimation of Spatial Models with Heterogeneous Coefficients CESifo Working Paper, No. 5428 Provided in Cooperation with: Ifo Institute – Leibniz Institute for Economic Research at the University of Munich Suggested Citation: Aquaro, Michele; Bailey, Natalia; Pesaran, M. Hashem (2015) : Quasi Maximum Likelihood Estimation of Spatial Models with Heterogeneous Coefficients, CESifo Working Paper, No. 5428, Center for Economic Studies and ifo Institute (CESifo), Munich This Version is available at: http://hdl.handle.net/10419/113752 Standard-Nutzungsbedingungen: Terms of use: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Documents in EconStor may be saved and copied for your Zwecken und zum Privatgebrauch gespeichert und kopiert werden. personal and scholarly purposes. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle You are not to copy documents for public or commercial Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich purposes, to exhibit the documents publicly, to make them machen, vertreiben oder anderweitig nutzen. publicly available on the internet, or to distribute or otherwise use the documents in public. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, If the documents have been made available under an Open gelten abweichend von diesen Nutzungsbedingungen die in der dort Content Licence (especially Creative Commons Licences), you genannten Lizenz gewährten Nutzungsrechte. may exercise further usage rights as specified in the indicated licence. www.econstor.eu Quasi Maximum Likelihood Estimation of Spatial Models with Heterogeneous Coefficients Michele Aquaro Natalia Bailey M.
    [Show full text]
  • Download (543Kb)
    XOT FOR QUOTATION WITHOUT PERMISSION OF THE AUTHOR FAVORABLE CLASSES OF LIPSCHITZ CONTINUOUS FUNCTIONS IN SUBGRADIENT OPTIMIZATION R. Tyrrell Rockafellar January 1981 LQP-8 1- 1 Research supported in part by the Air Force Office of Scientific Research, Air Force Systems Command, United States Air Force, under grant number 77-3204. Working Papers are interim reports on work of the International Institute for Applied Systems Analysis and have received only limited review. Views or opinions expressed herein do not necessarily repre- sent those of the Institute or of its National Member Organizations. INTERNATIONAL INSTITUTE FOR APPLIED SYSTEMS ANALYSIS A-2361 Laxenburg, Austria ABSTRACT Clarke has given a robust definition of subgradients of arbitrary Lipschitz continuous functions f on R", but for pur- poses of minimization algorithms it seems essential that the subgradient multifunction af have additional properties, such as certain special kinds of semicontinuity, which are not auto- matic consequences of f being Lipschitz continuous. This paper explores properties of 3f that correspond to f being subdiffer- entially regular, another concept of Clarke's, and to f being a pointwise supremum of functions that are k times continuously differentiable. FAVORABLE CLASSES OF LIPSCHITZ CONTINUOUS FUNCTIONS IN SUBGRADIENT OPTIMIZATION R. Tyrrell Rockafellar 1. INTRODUCTION A function f : R~+Ris said to be locally Lipschitzian if for each x E Rn there is a neighborhood X of x such that, for some X -> 17, (1.1) If(x") -£(XI)/ -< XIXI'-xlI for all XIEX, EX . Examples include continuously differentiable functions, convex functions, concave functions, saddle functions and any linear combination or pointwise maximum of a finite collection of such functions.
    [Show full text]
  • Dual Coordinate Descent Methods for Logistic Regression and Maximum Entropy Models
    Machine Learning Journal manuscript No. (will be inserted by the editor) Dual Coordinate Descent Methods for Logistic Regression and Maximum Entropy Models Hsiang-Fu Yu · Fang-Lan Huang · Chih-Jen Lin Received: date / Accepted: date Abstract Most optimization methods for logistic regression or maximum entropy solve the primal problem. They range from iterative scaling, coordinate descent, quasi- Newton, and truncated Newton. Less efforts have been made to solve the dual problem. In contrast, for linear support vector machines (SVM), methods have been shown to be very effective for solving the dual problem. In this paper, we apply coordinate descent methods to solve the dual form of logistic regression and maximum entropy. Interest- ingly, many details are different from the situation in linear SVM. We carefully study the theoretical convergence as well as numerical issues. The proposed method is shown to be faster than most state of the art methods for training logistic regression and maximum entropy. 1 Introduction Logistic regression (LR) is useful in many areas such as document classification and natural language processing (NLP). It models the conditional probability as: 1 Pw(y = ±1jx) ≡ ; 1 + e−ywT x where x is the data, y is the class label, and w 2 Rn is the weight vector. Given l n two-class training data fxi; yigi=1; xi 2 R ; yi 2 f1; −1g, logistic regression minimizes the following regularized negative log-likelihood: l X T 1 P LR(w) = C log 1 + e−yiw xi + wT w; (1) 2 i=1 Department of Computer Science National Taiwan University Taipei 106, Taiwan fb93107, d93011, [email protected] 2 where C > 0 is a penalty parameter.
    [Show full text]
  • On Competitive Nonlinear Pricing Andrea Attar, Thomas Mariotti, Francois Salanie
    On competitive nonlinear pricing Andrea Attar, Thomas Mariotti, Francois Salanie To cite this version: Andrea Attar, Thomas Mariotti, Francois Salanie. On competitive nonlinear pricing. Theoretical Economics, Econometric Society, 2019, 14 (1), pp.297-343. 10.3982/TE2708. hal-02097209 HAL Id: hal-02097209 https://hal.archives-ouvertes.fr/hal-02097209 Submitted on 26 May 2020 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. Distributed under a Creative Commons Attribution - NonCommercial| 4.0 International License Theoretical Economics 14 (2019), 297–343 1555-7561/20190297 On competitive nonlinear pricing Andrea Attar Toulouse School of Economics, CNRS, University of Toulouse Capitole and Dipartimento di Economia e Finanza, Università degli Studi di Roma“Tor Vergata” Thomas Mariotti Toulouse School of Economics, CNRS, University of Toulouse Capitole François Salanié Toulouse School of Economics, INRA, University of Toulouse Capitole We study a discriminatory limit-order book in which market makers compete in nonlinear tariffs to serve a privately informed insider. Our model allows for general nonparametric specifications of preferences and arbitrary discrete distri- butions for the insider’s private information. Adverse selection severely restricts equilibrium outcomes: in any pure-strategy equilibrium with convex tariffs, pric- ing must be linear and at most one type can trade, leading to an extreme form of market breakdown.
    [Show full text]
  • Reinforcement Learning Via Fenchel-Rockafellar Duality
    Reinforcement Learning via Fenchel-Rockafellar Duality Ofir Nachum Bo Dai [email protected] [email protected] Google Research Google Research Abstract We review basic concepts of convex duality, focusing on the very general and supremely useful Fenchel-Rockafellar duality. We summarize how this duality may be applied to a variety of reinforcement learning (RL) settings, including policy evaluation or optimization, online or offline learning, and discounted or undiscounted rewards. The derivations yield a number of intriguing results, including the ability to perform policy evaluation and on-policy policy gradient with behavior-agnostic offline data and methods to learn a policy via max-likelihood optimization. Although many of these results have appeared previously in various forms, we provide a unified treatment and perspective on these results, which we hope will enable researchers to better use and apply the tools of convex duality to make further progress in RL. 1 Introduction Reinforcement learning (RL) aims to learn behavior policies to optimize a long-term decision-making process in an environment. RL is thus relevant to a variety of real- arXiv:2001.01866v2 [cs.LG] 9 Jan 2020 world applications, such as robotics, patient care, and recommendation systems. When tackling problems associated with the RL setting, two main difficulties arise: first, the decision-making problem is inherently sequential, with early decisions made by the policy affecting outcomes both near and far in the future; second, the learner’s knowledge of the environment is only through sampled experience, i.e., previously sampled trajectories of interactions, while the underlying mechanism governing the dynamics of these trajectories is unknown.
    [Show full text]
  • A Single-Loop Smoothed Gradient Descent-Ascent Algorithm
    A Single-Loop Smoothed Gradient Descent-Ascent Algorithm for Nonconvex-Concave Min-Max Problems Jiawei Zhang ∗ Peijun Xiao † Ruoyu Sun ‡ Zhi-Quan Luo § October 30, 2020 Abstract Nonconvex-concave min-max problem arises in many machine learning applications including minimizing a pointwise maximum of a set of nonconvex functions and robust adversarial training of neural networks. A popular approach to solve this problem is the gradient descent-ascent (GDA) algorithm which unfortunately can exhibit oscillation in case of nonconvexity. In this paper, we introduce a “smoothing” scheme which can be combined with GDA to stabilize the oscillation and ensure convergence to a stationary solution. We prove that the stabilized GDA algorithm can achieve an O(1/ǫ2) iteration complexity for minimizing the pointwise maximum of a finite collection of nonconvex functions. Moreover, the smoothed GDA algorithm achieves an O(1/ǫ4) iteration complexity for general nonconvex-concave problems. Extensions of this stabilized GDA algorithm to multi-block cases are presented. To the best of our knowledge, this is the first algorithm to achieve O(1/ǫ2) for a class of nonconvex-concave problem. We illustrate the practical efficiency of the stabilized GDA algorithm on robust training. 1 Introduction Min-max problems have drawn considerable interest from the machine learning and other engi- neering communities. They appear in applications such as adversarial learning [1–3], robust opti- mization [4–7], empirical risk minimization [8, 9], and reinforcement learning [10, 11]. Concretely speaking, a min-max problem is in the form: min max f(x,y), (1.1) x∈X y∈Y arXiv:2010.15768v1 [math.OC] 29 Oct 2020 where X Rn and Y Rm are convex and closed sets and f is a smooth function.
    [Show full text]
  • Optimization
    ECE 592 Topics in Data Science Dror Baron Associate Professor Dept. of Electrical and Computer Engr. North Carolina State University, NC, USA Optimization Keywords: linear programming, dynamic programming, convex optimization, non-convex optimization What is Optimization? What is optimization? . Wikipedia: In mathematics, computer science and operations research, mathematical optimization (alternatively, mathematical programming or simply, optimization) is the selection of a best element (with regard to some criterion) from some set of available alternatives. 4 Application #1-Classroom scheduling . Real story NCSU has classes on multiple campuses, dozens of buildings, etc. We want a “good” schedule . What’s good? – Availability of rooms – Proximity of classroom to department – Instructors have day/time preferences – Match sizes of rooms and anticipated class enrollment – Avoid conflicts between course pairs of interest to students 5 Application #2- 1 recovery . Among infinitely manyℓ solutions, seek one with smallest 1 norm (sum of absolute values) ℓ . Relation to compressed sensing recovery (later in course) . Can express x=xp-xn 1 = , + , . min + subject to (s.t. ) y=Φx -Φx , , ∑=1 p n – Also need xp, xn to be non-negative ∑=1 6 Application #3-reducing fuel consumption . Suppose gas prices increase a lot . Truck fleet company wants to save $ by reducing fuel consumption . Things are simple on flat highways . Challenges: 1) You see a hill; can push engine up hill and coast down, or accelerate before hill, then reduce speed while climbing 2) We see red light; should we coast, accelerate, slam brakes? Main point: dynamic behavior links past, present, future 7 Application #4-process design in factories . Consider factory with complicated process – Want to buy less inputs (chemicals) – Want to use less energy – Want product to be produced quickly (time) – Want robustness to surprises (e.g., power shortages) .
    [Show full text]