FICO Xpress Optimization Suite Webinar

Oliver Bastert Senior Manager Xpress Product Management

September 22 2011

» Introduction to FICO » Introduction to FICO Xpress Optimization Suite » Performance » Distributed Modelling and Solving » Case Q&A

The leader in predictive analytics for decision management Founded: 1956 Profile NYSE: FICO Revenues: $605 million (fiscal 2010) Scores and related analytic models Products Analytic applications for risk management, fraud, marketing and Services Tools for decision management

Clients and 5,000+ clients in 80 countries Markets Industry focus: Banking, insurance, retail, health care

#1 in services operations analytics (IDC) Recent #7 in worldwide business analytics software (IDC) Rankings #26 in the FinTech 100 (American Banker) 20+ offices worldwide, HQ in Minneapolis, USA Offices 2,200 employees Regional Hubs: San Rafael (CA), New York, London, Birmingham (UK), Munich, Madrid, Sao Paulo, Bangalore, Beijing, Singapore

Thousands of businesses worldwide, including 9 of the top Fortune 10, rely on FICO to make every decision count.

Sharpen customer-centric Increase customer decisions profitability Predict customer needs and behavior Reduce losses from fraud and risk Pinpoint best offer and action Connect all decisions about a customer PREDICT PROFIT

ADAPT IMPROVE Change faster and Continually improve respond flexibly strategy performance Change business rules instantly Model decisions for greater control Create a test-and-learn culture Optimize strategies to grow faster

Model decisions for greater control » Identify the decision drivers and the effects of every action » Use the decision model as a planning Uses FICO optimization software to tool to test changes in the business create analytically driven decisions on environment fleet distribution and utilization Deployed across every key market in Optimize strategies to grow faster continental Europe » Create analytically derived strategies to • Benefit estimated at $19 million meet specified business objectives » Design strategies with millions of “FICO™ Xpress tells us, for example: variables – instantly On Friday morning, bring only four cars from Heathrow to Mayfair, and bring another four from Stansted Airport. The utilization of our fleet has gone up by one or two percentage points.”

FICO holds 100+ patents in analytic First credit line optimization and decision management technology, solutions with 150 more patents pending First predictive systems for insurance fraud First analytic First cross- systems for bureau credit retailers to scores optimize offers First small First adaptive business scoring analytics for First insurance systems fraud underwriting scoring systems First neural First credit network-based capacity scores First First automated First adaptive fraud solutions commercially origination control systems First score for available credit systems with for managing First cardholder prescription scoring systems analytics card accounts profiling for fraud adherence

1960s 1970s 1980s 1990s 2000s

Summarize Past and Current Behavior Predict Future Behavior and Adapt Tools Solutions Scores Tools Solutions Decision Optimization Predictive Analytics

Descriptive Decision Value Decision Business Analytics Intelligence

Analytic Capability Understand Make different Target each Automatically the trends in offers to groups decision to a take the ideal the business of customers customer’s future action on each behavior individual

For Lifecycle Specific Decision Processes Customer Collections and Fraud Marketing Origination Management Recovery Management

FICO® Precision FICO® FICO® TRIAD® FICO® Debt FICO™ Falcon® Marketing Origination Customer Manager Manager™ Fraud Manager Applications Manager Manager FICO™ Recovery FICO™ Insurance FICO® Retail Management Fraud Manager Action Manager System™

For Any Decision Process

B2B: FICO® Score  FICO® Credit Capacity Index™ Scores FICO® Insurance Risk Scores B2C: myFICO®

Business Rules Management: FICO™ Blaze Advisor® Tools Predictive Analytics: FICO™ Model Builder Optimization: FICO™ Xpress Optimization Suite  FICO™ Decision Optimizer

Custom Analytics Professional Operational Best Practices Services Strategy Design and Optimization

GUI XAD Graphical user interface development using Mosel

Programming Solver API Mosel API BCL*

Deployment Interfaces .NET/Java/C/C++/VB

Development IVE IVE-XAD Development Environment GUI development

Modelling Mosel MOdelling and Solving Environment Language

MIP MIQP MIQCQP MISLP MINLP Solvers CP LP QP QCQP SLP NLP

* Builder Component Library for modelling in a programming language

» Editor » Debugger » Profiler » Progress graphs » Visualization » Wizards » Mosel extensions » Deployment

FICO Optimization Dashboard: Debt Consolidation Module Confidential – do not copy

Features Benefits » Advanced programming languages: » Entire Mathematical Model can be stored » Algebraic modeling language in one place for rapid development and » Procedural programming language easy maintenance. » Utilize different solvers in the » From Mosel you can solve LPs, MIPs, MIQPs, same model Non-Linear problems, Stochastic problems, and Constraint problems » Decompose & parallelize a model to » Faster solve times take advantage of multiple CPUs/cores » Make full use of your computing infrastructure through distributed computing » Build a GUI exclusively within » Decreases development time, gets optimization Mosel code in front of business user quicker » Portable across operating systems » Mosel Model compiled in one OS can be deployed on all other supported Operating Systems, decreasing development time » Open, modular architecture, » User flexibility to solve the most complicated User extensible optimization problems » not limited to/by predefined language features » Compiled » Protects intellectual property » Offers a variety of APIs and data » Easy deployment and works in heterogeneous

» 26 years of experience in modelling and optimization » 24 years of experience in mixed integer optimization » 12 years of experience in nonlinear optimization » 10 years Xpress-Mosel, modelling and solving environment

» Integration of modelling and solving » Focus on (potentially) exact solution methods » Xpress-Solvers often can prove optimality of the solution » They always give you information on the quality of the solution

» Solving » 1983: LP solver running on PCs » 1992: parallel MIP (1997 on distributed PC/Linux networks) » 1995/1996 : commercial branch and cut algorithm » 1998: bound switching in dual simplex » 2003: lift-and-project cuts » 2009: parallel MIP heuristics » 2010: LP/MIP solver crosses 64-bit coefficient indexing threshold » Modelling » 1983: general purpose algebraic modelling language (mp-model) » 2001: algebraic modelling language combining modelling, solving, and programming (Mosel) » 2005: profiler and debugger for a modelling language » 2005: user-controlled parallelism at the model level » 2010: algebraic modelling language supporting distributed computing

» Unique capabilities for large scale optimization including ability to solve ultra-large problems (true64bit capabilities) and support for distributed modeling and optimization » Complete set of state-of-the-art optimization engines that are robust, reliable and faster than competing solutions » An easy-to-learn, powerful modeling and programming language, Xpress-Mosel » The premier visual development environment, IVE, for developing mathematical models » An intuitive drag-and-drop editor for creating GUIs that seamlessly integrate with the model for rapid prototyping and deployment » A partner committed to solving all of your most difficult optimization problems

Xpress 7.1 delivers (GA Nov 2010) » Solve much bigger problems » The possibilities are limitless with the enhanced optimization and modeling support for solving ultra-large-scale models where the number of coefficients can exceed 2 billion. » Solve large problems faster » Cut solution times dramatically by leveraging distributed execution of Mosel models that can now be controlled from a master model across a heterogeneous set of machines. » Get solutions faster with significantly improved solver performance » Average increase of 50% arithmetic/50% geometric for multi-threaded MIP » Average increase of 50% arithmetic/25% geometric for single-threaded MIP » Average increase of 50% arithmetic/70% geometric for MIQP » Significant improvements to speed and stability of quadratic simplex and SLP non-linear algorithms

» Improved developer usability » Developers will also enjoy greater productivity from usability improvements to the development environment and enhanced modeling functionality such as the MIIS automated modeling error/infeasibility detection. » Easier to integrate with other applications » Optimization programmers will be pleased by the addition of simplified and more robust data exchange capabilities between Mosel and applications.

Xpress 7.2 delivers (GA April 2011) » Exceptional public benchmark performance

25 © 2011 Fair Isaac Corporation. Confidential. Comparing Solver Performance » Solver performance is important but not the only decision criterion » Selection of benchmark sets » Represent client mix of problems » Solvable instances but not too simple » Feasibility and optimality check of solution » Numerically stable problems are preferred for performance benchmarks » The only public benchmarks for optimization solvers is run by Hans Mittelmann. He frequently changes the benchmarking sets » The best known collection of MIP instances is currently updated from version MIPLIB 2003 to version MIPLIB 2010 and will contain for the first time an agreed benchmarking subset. » A benchmark comparison is always a snapshot of the performance of the available software at a given point in time.

Old Best Known Obj. Xpress Improved Obj. GAIN Problem Value (*) Value (**) (|1-(**)/(*)|) atlanta-ip 95.009549704 90.00987861 5.3% msc98-ip 20980991.006 19839497.006 5.4%

protfold -30 -31 3.3%

rd-rplusc-21 171182 165395.2753 3.4% Solving Hard Mixed Integer Programming Problems with Xpress-MP: A MIPLIBsp97ar 2003 Case Study,664565103.76 Informs Journal660705646.5 on Computing, 20090.6% Optimal stp3d unknown 500.736 N/A by Richard Laundy, Michael Perregaard, Gabriel Tavares, Horia Tipi, and Alkis Vazacopoulos ds 283.4425 116.59 58.9% momentum3 370177.036 236426.335 36.1% t1717 193221 170195 11.9% liu 1172 1102 5.9%

dano3mip 691.2 687.733333 0.5% Unsolved

» Instead of comparing the overall runtime on a given matrix set (or equivalently the average runtime or arithmetic mean on that set) the accepted way of comparing optimization solver performance is by comparing the Geometric Mean » The presence of a few extremely small or large values has no considerable effect on geometric mean so it measures performance more accurately than arithmetic mean which is biased towards large outliers. » The geometric mean denotes the most likely runtime you will observe for an instance of the test set.

Standard LP Problems (Barrier, Simplex) 90

0 XPRESS CPLEX GUROBI MOSEK

Barrier on Large LP Problems 700

600

500

400

300

200

100

0 XPRESS CPLEX GUROBI MOSEK

MIQP Problems 450

400

350

300

250

200

150

100

0 XPRESS CPLEX GUROBI SCIP

300.00

250.00

200.00

150.00

100.00

50.00

0.00 XPRESS CPLEX GUROBI

1. Simple submodel run

Master start

compile/load/run submodel Submodel

User wait for results termination

results process results

c 2011 Fair Issac Corporation. Schemes of parallelization

2. Iterative sequential submodel runs (decomposition algorithms)

Master start compile/load submodel

run submodel User Submodel

wait for results termination

results process results

c 2011 Fair Issac Corporation. Schemes of parallelization

3. Independent parallel submodels

Master start start compile/load/run submodels ... User Submodel 1 Submodel n

wait for results termination

results process results

c 2011 Fair Issac Corporation. Schemes of parallelization

4. Communicating concurrent submodels

Master start start compile/load/run submodels

Submodel 1 ... Submodel n User wait for events events/results

process events broadcast updates/termination

results process results

c 2011 Fair Issac Corporation. Distributed solving

» Use all the computing power available in your local network by solving (sub)models on remote machines

Local

run master model

c 2011 Fair Issac Corporation. Distributed solving

» Use all the computing power available in your local network by solving (sub)models on remote machines

Local Remote

run master model run submodel

c 2011 Fair Issac Corporation. Distributed solving

» Use all the computing power available in your local network by solving (sub)models on remote machines

Local Remote

run master model run submodel

» Physical location of model ﬁles, input and result data depending on application

c 2011 Fair Issac Corporation. Distributed solving: Location of models and data

1. On local host » physical ﬁles or in memory (e.g. included in master model or in calling host application) » a (master) model can recursively start new instances of itself

Local Remote

run master model run submodel

load

submodel ﬁle

data, results

c 2011 Fair Issac Corporation. Distributed solving: Location of models and data

2. On remote host » conﬁgurable read/write access on remote machine

Local Remote

run master model run submodel

load submodel ﬁle

data, results

c 2011 Fair Issac Corporation. Distributed solving: Location of models and data

3. Centralised repository » eases version control in multi-user environments

Local Remote

run master model run submodel

load

submodel ﬁle

data, results

Central repository

c 2011 Fair Issac Corporation. Distributed applications

» Multi-user application with Mosel model as dispatcher User Production machine

... Mosel server ... Production User Database machine

» Example: optimization applications in ﬁnance (solving large numbers of small to medium size problems)

c 2011 Fair Issac Corporation. Distributed applications

» Decomposition with central data store

Submodel Mosel User optimization master ... Submodel

Database

» Examples: Column generation in transport or personnel planning; blockwise (Dantzig-Wolfe) decomposition in production planning

c 2011 Fair Issac Corporation. Distributed applications

» Decomposition with remote, distributed data sources

Remote model Data Mosel User optimization master ... Remote model Data

» Example: Large-scale planning in heterogeneous computing environment

c 2011 Fair Issac Corporation. Case Studies

» Modify the composition of an investment portfolio as to achieve or approach a speciﬁed investment proﬁle.

c 2010 Fair Issac Corporation. Optimization application in Mosel Standalone

IVE

start application return results

Data ﬁles Mosel model

Output ﬁles

c 2010 Fair Issac Corporation. Optimization application in Mosel XAD GUI

Conﬁgu- ration ﬁle XAD

Summary start application output return results

Data ﬁles Mosel model

Output ﬁles

c 2010 Fair Issac Corporation. Optimization application in Mosel Embedded into host application

Data ﬁles Java

Summary start application output return results

Mosel model

Output ﬁles

c 2010 Fair Issac Corporation. Optimization application in Mosel Alternative interfaces

Configu- Data files ration file XAD IVE Java

Summary start application output return results Summary output

Data ﬁles Mosel model

Output ﬁles

c 2010 Fair Issac Corporation. XAD interface: Detailed results

c 2010 Fair Issac Corporation. XAD interface: Parameter and version log

c 2010 Fair Issac Corporation. XAD interface: Multiple run summary

c 2010 Fair Issac Corporation. Some highlights

» Model: » easy maintenance through single model » deployment as BIM file: no changes to model by end-user » language extensions according to specific needs » Interfaces: » several run modes adapted to different types of usages » efficient data exchange with host application through memory » parallel model runs (Java) or repeated sequential runs (XAD)

c 2010 Fair Issac Corporation. Aircraft routing: Problem description

» For given sets of flights and aircraft, determine which aircraft services a flight. » Aircraft are not identical » they cannot all service every flight » a specific maintenance site must be used per plane » some scheduled long maintenance breaks » Starting condition: each aircraft has a starting position and a specific amount of accumulated flight minutes

c 2010 Fair Issac Corporation. Aircraft routing: Application architecture

» Master problem: route selection » Subproblems: route generation (one instance per plane) » parallel, possibly remote, execution of submodels » User interface (optional): XAD GUI

c 2010 Fair Issac Corporation. Aircraft routing: Application GUI

c 2010 Fair Issac Corporation. Aircraft routing: Visualization

c 2010 Fair Issac Corporation. Aircraft routing: User interaction

c 2010 Fair Issac Corporation. Q&A