Reinforcement Learning and Stochastic Optimization
-15%
portes grátis
Reinforcement Learning and Stochastic Optimization
A Unified Framework for Sequential Decisions
Powell, Warren B.
John Wiley & Sons Inc
03/2022
1136
Dura
Inglês
9781119815037
15 a 20 dias
1874
Descrição não disponível.
Preface xxv
Acknowledgments xxxi
Part I - Introduction 1
1 Sequential Decision Problems 3
1.1 The Audience 7
1.2 The Communities of Sequential Decision Problems 8
1.3 Our Universal Modeling Framework 10
1.4 Designing Policies for Sequential Decision Problems 15
1.5 Learning 20
1.6 Themes 21
1.7 Our Modeling Approach 27
1.8 How to Read this Book 27
1.9 Bibliographic Notes 33
Exercises 34
Bibliography 38
2 Canonical Problems and Applications 39
2.1 Canonical Problems 39
2.2 A Universal Modeling Framework for Sequential Decision Problems 64
2.3 Applications 69
2.4 Bibliographic Notes 85
Exercises 90
Bibliography 93
3 Online Learning 101
3.1 Machine Learning for Sequential Decisions 102
3.2 Adaptive Learning Using Exponential Smoothing 110
3.3 Lookup Tables with Frequentist Updating 111
3.4 Lookup Tables with Bayesian Updating 112
3.5 Computing Bias and Variance* 118
3.6 Lookup Tables and Aggregation* 121
3.7 Linear Parametric Models 131
3.8 Recursive Least Squares for Linear Models 136
3.9 Nonlinear Parametric Models 140
3.10 Nonparametric Models* 149
3.11 Nonstationary Learning* 159
3.12 The Curse of Dimensionality 162
3.13 Designing Approximation Architectures in Adaptive Learning 165
3.14 Why Does It Work?** 166
3.15 Bibliographic Notes 174
Exercises 176
Bibliography 180
4 Introduction to Stochastic Search 183
4.1 Illustrations of the Basic Stochastic Optimization Problem 185
4.2 Deterministic Methods 188
4.3 Sampled Models 193
4.4 Adaptive Learning Algorithms 202
4.5 Closing Remarks 210
4.6 Bibliographic Notes 210
Exercises 212
Bibliography 218
Part II - Stochastic Search 221
5 Derivative-Based Stochastic Search 223
5.1 Some Sample Applications 225
5.2 Modeling Uncertainty 228
5.3 Stochastic Gradient Methods 231
5.4 Styles of Gradients 237
5.5 Parameter Optimization for Neural Networks* 242
5.6 Stochastic Gradient Algorithm as a Sequential Decision Problem 247
5.7 Empirical Issues 248
5.8 Transient Problems* 249
5.9 Theoretical Performance* 250
5.10 Why Does it Work? 250
5.11 Bibliographic Notes 263
Exercises 264
Bibliography 270
6 Stepsize Policies 273
6.1 Deterministic Stepsize Policies 276
6.2 Adaptive Stepsize Policies 282
6.3 Optimal Stepsize Policies* 289
6.4 Optimal Step sizes for Approximate Value Iteration* 297
6.5 Convergence 300
6.6 Guidelines for Choosing Stepsize Policies 301
6.7 Why Does it Work* 303
6.8 Bibliographic Notes 306
Exercises 307
Bibliography 314
7 Derivative-Free Stochastic Search 317
7.1 Overview of Derivative-free Stochastic Search 319
7.2 Modeling Derivative-free Stochastic Search 325
7.3 Designing Policies 330
7.4 Policy Function Approximations 333
7.5 Cost Function Approximations 335
7.6 VFA-based Policies 338
7.7 Direct Lookahead Policies 348
7.8 The Knowledge Gradient (Continued)* 362
7.9 Learning in Batches 380
7.10 Simulation Optimization* 382
7.11 Evaluating Policies 385
7.12 Designing Policies 394
7.13 Extensions* 398
7.14 Bibliographic Notes 409
Exercises 412
Bibliography 424
Part III - State-dependent Problems 429
8 State-dependent Problems 431
8.1 Graph Problems 433
8.2 Inventory Problems 439
8.3 Complex Resource Allocation Problems 446
8.4 State-dependent Learning Problems 456
8.5 A Sequence of Problem Classes 460
8.6 Bibliographic Notes 461
Exercises 462
Bibliography 466
9 Modeling Sequential Decision Problems 467
9.1 A Simple Modeling Illustration 471
9.2 Notational Style 476
9.3 Modeling Time 478
9.4 The States of Our System 481
9.5 Modeling Decisions 500
9.6 The Exogenous Information Process 506
9.7 The Transition Function 515
9.8 The Objective Function 518
9.9 Illustration: An Energy Storage Model 523
9.10 Base Models and Lookahead Models 528
9.11 A Classification of Problems* 529
9.12 Policy Evaluation* 532
9.13 Advanced Probabilistic Modeling Concepts** 534
9.14 Looking Forward 540
9.15 Bibliographic Notes 542
Exercises 544
Bibliography 557
10 Uncertainty Modeling 559
10.1 Sources of Uncertainty 560
10.2 A Modeling Case Study: The COVID Pandemic 575
10.3 Stochastic Modeling 575
10.4 Monte Carlo Simulation 581
10.5 Case Study: Modeling Electricity Prices 589
10.6 Sampling vs. Sampled Models 595
10.7 Closing Notes 597
10.8 Bibliographic Notes 597
Exercises 598
Bibliography 601
11 Designing Policies 603
11.1 From Optimization to Machine Learning to Sequential Decision Problems 605
11.2 The Classes of Policies 606
11.3 Policy Function Approximations 610
11.4 Cost Function Approximations 613
11.5 Value Function Approximations 614
11.6 Direct Lookahead Approximations 616
11.7 Hybrid Strategies 620
11.8 Randomized Policies 626
11.9 Illustration: An Energy Storage Model Revisited 627
11.10 Choosing the Policy Class 631
11.11 Policy Evaluation 641
11.12 Parameter Tuning 642
11.13 Bibliographic Notes 646
Exercises 646
Bibliography 651
Part IV - Policy Search 653
12 Policy Function Approximations and Policy Search 655
12.1 Policy Search as a Sequential Decision Problem 657
12.2 Classes of Policy Function Approximations 658
12.3 Problem Characteristics 665
12.4 Flavors of Policy Search 666
12.5 Policy Search with Numerical Derivatives 669
12.6 Derivative-Free Methods for Policy Search 670
12.7 Exact Derivatives for Continuous Sequential Problems* 677
12.8 Exact Derivatives for Discrete Dynamic Programs** 680
12.9 Supervised Learning 686
12.10 Why Does it Work? 687
12.11 Bibliographic Notes 690
Exercises 691
Bibliography 698
13 Cost Function Approximations 701
13.1 General Formulation for Parametric CFA 703
13.2 Objective-Modified CFAs 704
13.3 Constraint-Modified CFAs 714
13.4 Bibliographic Notes 725
Exercises 726
Bibliography 729
Part V - Lookahead Policies 731
14 Exact Dynamic Programming 737
14.1 Discrete Dynamic Programming 738
14.2 The Optimality Equations 740
14.3 Finite Horizon Problems 747
14.4 Continuous Problems with Exact Solutions 750
14.5 Infinite Horizon Problems* 755
14.6 Value Iteration for Infinite Horizon Problems* 757
14.7 Policy Iteration for Infinite Horizon Problems* 762
14.8 Hybrid Value-Policy Iteration* 764
14.9 Average Reward Dynamic Programming* 765
14.10 The Linear Programming Method for Dynamic Programs** 766
14.11 Linear Quadratic Regulation 767
14.12 Why Does it Work?** 770
14.13 Bibliographic Notes 783
Exercises 783
Bibliography 793
15 Backward Approximate Dynamic Programming 795
15.1 Backward Approximate Dynamic Programming for Finite Horizon Problems 797
15.2 Fitted Value Iteration for Infinite Horizon Problems 804
15.3 Value Function Approximation Strategies 805
15.4 Computational Observations 810
15.5 Bibliographic Notes 816
Exercises 816
Bibliography 821
16 Forward ADP I: The Value of a Policy 823
16.1 Sampling the Value of a Policy 824
16.2 Stochastic Approximation Methods 835
16.3 Bellman's Equation Using a Linear Model* 837
16.4 Analysis of TD(0), LSTD, and LSPE Using a Single State* 842
16.5 Gradient-based Methods for Approximate Value Iteration* 845
16.6 Value Function Approximations Based on Bayesian Learning* 852
16.7 Learning Algorithms and Atepsizes 855
16.8 Bibliographic Notes 860
Exercises 862
Bibliography 864
17 Forward ADP II: Policy Optimization 867
17.1 Overview of Algorithmic Strategies 869
17.2 Approximate Value Iteration and Q-Learning Using Lookup Tables 871
17.3 Styles of Learning 881
17.4 Approximate Value Iteration Using Linear Models 886
17.5 On-policy vs. off-policy learning and the exploration-exploitation problem 888
17.6 Applications 894
17.7 Approximate Policy Iteration 900
17.8 The Actor-Critic Paradigm 907
17.9 Statistical Bias in the Max Operator* 909
17.10 The Linear Programming Method Using Linear Models* 912
17.11 Finite Horizon Approximations for Steady-State Applications 915
17.12 Bibliographic Notes 917
Exercises 918
Bibliography 924
18 Forward ADP III: Convex Resource Allocation Problems 927
18.1 Resource Allocation Problems 930
18.2 Values Versus Marginal Values 937
18.3 Piecewise Linear Approximations for Scalar Functions 938
18.4 Regression Methods 941
18.5 Separable Piecewise Linear Approximations 944
18.6 Benders Decomposition for Nonseparable Approximations** 946
18.7 Linear Approximations for High-Dimensional Applications 956
18.8 Resource Allocation with Exogenous Information State 958
18.9 Closing Notes 959
18.10 Bibliographic Notes 960
Exercises 962
Bibliography 967
19 Direct Lookahead Policies 971
19.1 Optimal Policies Using Lookahead Models 974
19.2 Creating an Approximate Lookahead Model 978
19.3 Modified Objectives in Lookahead Models 985
19.4 Evaluating DLA Policies 992
19.5 Why Use a DLA? 997
19.6 Deterministic Lookaheads 999
19.7 A Tour of Stochastic Lookahead Policies 1005
19.8 Monte Carlo Tree Search for Discrete Decisions 1009
19.9 Two-Stage Stochastic Programming for Vector Decisions* 1018
19.10 Observations on DLA Policies 1024
19.11 Bibliographic Notes 1025
Exercises 1027
Bibliography 1031
Part VI - Multiagent Systems 1033
20 Multiagent Modeling and Learning 1035
20.1 Overview of Multiagent Systems 1036
20.2 A Learning Problem - Flu Mitigation 1044
20.3 The POMDP Perspective* 1059
20.4 The Two-Agent Newsvendor Problem 1062
20.5 Multiple Independent Agents - An HVAC Controller Model 1067
20.6 Cooperative Agents - A Spatially Distributed Blood Management Problem 1070
20.7 Closing Notes 1074
20.8 Why Does it Work? 1074
20.9 Bibliographic Notes 1076
Exercises 1077
Bibliography 1083
Index 1085
Acknowledgments xxxi
Part I - Introduction 1
1 Sequential Decision Problems 3
1.1 The Audience 7
1.2 The Communities of Sequential Decision Problems 8
1.3 Our Universal Modeling Framework 10
1.4 Designing Policies for Sequential Decision Problems 15
1.5 Learning 20
1.6 Themes 21
1.7 Our Modeling Approach 27
1.8 How to Read this Book 27
1.9 Bibliographic Notes 33
Exercises 34
Bibliography 38
2 Canonical Problems and Applications 39
2.1 Canonical Problems 39
2.2 A Universal Modeling Framework for Sequential Decision Problems 64
2.3 Applications 69
2.4 Bibliographic Notes 85
Exercises 90
Bibliography 93
3 Online Learning 101
3.1 Machine Learning for Sequential Decisions 102
3.2 Adaptive Learning Using Exponential Smoothing 110
3.3 Lookup Tables with Frequentist Updating 111
3.4 Lookup Tables with Bayesian Updating 112
3.5 Computing Bias and Variance* 118
3.6 Lookup Tables and Aggregation* 121
3.7 Linear Parametric Models 131
3.8 Recursive Least Squares for Linear Models 136
3.9 Nonlinear Parametric Models 140
3.10 Nonparametric Models* 149
3.11 Nonstationary Learning* 159
3.12 The Curse of Dimensionality 162
3.13 Designing Approximation Architectures in Adaptive Learning 165
3.14 Why Does It Work?** 166
3.15 Bibliographic Notes 174
Exercises 176
Bibliography 180
4 Introduction to Stochastic Search 183
4.1 Illustrations of the Basic Stochastic Optimization Problem 185
4.2 Deterministic Methods 188
4.3 Sampled Models 193
4.4 Adaptive Learning Algorithms 202
4.5 Closing Remarks 210
4.6 Bibliographic Notes 210
Exercises 212
Bibliography 218
Part II - Stochastic Search 221
5 Derivative-Based Stochastic Search 223
5.1 Some Sample Applications 225
5.2 Modeling Uncertainty 228
5.3 Stochastic Gradient Methods 231
5.4 Styles of Gradients 237
5.5 Parameter Optimization for Neural Networks* 242
5.6 Stochastic Gradient Algorithm as a Sequential Decision Problem 247
5.7 Empirical Issues 248
5.8 Transient Problems* 249
5.9 Theoretical Performance* 250
5.10 Why Does it Work? 250
5.11 Bibliographic Notes 263
Exercises 264
Bibliography 270
6 Stepsize Policies 273
6.1 Deterministic Stepsize Policies 276
6.2 Adaptive Stepsize Policies 282
6.3 Optimal Stepsize Policies* 289
6.4 Optimal Step sizes for Approximate Value Iteration* 297
6.5 Convergence 300
6.6 Guidelines for Choosing Stepsize Policies 301
6.7 Why Does it Work* 303
6.8 Bibliographic Notes 306
Exercises 307
Bibliography 314
7 Derivative-Free Stochastic Search 317
7.1 Overview of Derivative-free Stochastic Search 319
7.2 Modeling Derivative-free Stochastic Search 325
7.3 Designing Policies 330
7.4 Policy Function Approximations 333
7.5 Cost Function Approximations 335
7.6 VFA-based Policies 338
7.7 Direct Lookahead Policies 348
7.8 The Knowledge Gradient (Continued)* 362
7.9 Learning in Batches 380
7.10 Simulation Optimization* 382
7.11 Evaluating Policies 385
7.12 Designing Policies 394
7.13 Extensions* 398
7.14 Bibliographic Notes 409
Exercises 412
Bibliography 424
Part III - State-dependent Problems 429
8 State-dependent Problems 431
8.1 Graph Problems 433
8.2 Inventory Problems 439
8.3 Complex Resource Allocation Problems 446
8.4 State-dependent Learning Problems 456
8.5 A Sequence of Problem Classes 460
8.6 Bibliographic Notes 461
Exercises 462
Bibliography 466
9 Modeling Sequential Decision Problems 467
9.1 A Simple Modeling Illustration 471
9.2 Notational Style 476
9.3 Modeling Time 478
9.4 The States of Our System 481
9.5 Modeling Decisions 500
9.6 The Exogenous Information Process 506
9.7 The Transition Function 515
9.8 The Objective Function 518
9.9 Illustration: An Energy Storage Model 523
9.10 Base Models and Lookahead Models 528
9.11 A Classification of Problems* 529
9.12 Policy Evaluation* 532
9.13 Advanced Probabilistic Modeling Concepts** 534
9.14 Looking Forward 540
9.15 Bibliographic Notes 542
Exercises 544
Bibliography 557
10 Uncertainty Modeling 559
10.1 Sources of Uncertainty 560
10.2 A Modeling Case Study: The COVID Pandemic 575
10.3 Stochastic Modeling 575
10.4 Monte Carlo Simulation 581
10.5 Case Study: Modeling Electricity Prices 589
10.6 Sampling vs. Sampled Models 595
10.7 Closing Notes 597
10.8 Bibliographic Notes 597
Exercises 598
Bibliography 601
11 Designing Policies 603
11.1 From Optimization to Machine Learning to Sequential Decision Problems 605
11.2 The Classes of Policies 606
11.3 Policy Function Approximations 610
11.4 Cost Function Approximations 613
11.5 Value Function Approximations 614
11.6 Direct Lookahead Approximations 616
11.7 Hybrid Strategies 620
11.8 Randomized Policies 626
11.9 Illustration: An Energy Storage Model Revisited 627
11.10 Choosing the Policy Class 631
11.11 Policy Evaluation 641
11.12 Parameter Tuning 642
11.13 Bibliographic Notes 646
Exercises 646
Bibliography 651
Part IV - Policy Search 653
12 Policy Function Approximations and Policy Search 655
12.1 Policy Search as a Sequential Decision Problem 657
12.2 Classes of Policy Function Approximations 658
12.3 Problem Characteristics 665
12.4 Flavors of Policy Search 666
12.5 Policy Search with Numerical Derivatives 669
12.6 Derivative-Free Methods for Policy Search 670
12.7 Exact Derivatives for Continuous Sequential Problems* 677
12.8 Exact Derivatives for Discrete Dynamic Programs** 680
12.9 Supervised Learning 686
12.10 Why Does it Work? 687
12.11 Bibliographic Notes 690
Exercises 691
Bibliography 698
13 Cost Function Approximations 701
13.1 General Formulation for Parametric CFA 703
13.2 Objective-Modified CFAs 704
13.3 Constraint-Modified CFAs 714
13.4 Bibliographic Notes 725
Exercises 726
Bibliography 729
Part V - Lookahead Policies 731
14 Exact Dynamic Programming 737
14.1 Discrete Dynamic Programming 738
14.2 The Optimality Equations 740
14.3 Finite Horizon Problems 747
14.4 Continuous Problems with Exact Solutions 750
14.5 Infinite Horizon Problems* 755
14.6 Value Iteration for Infinite Horizon Problems* 757
14.7 Policy Iteration for Infinite Horizon Problems* 762
14.8 Hybrid Value-Policy Iteration* 764
14.9 Average Reward Dynamic Programming* 765
14.10 The Linear Programming Method for Dynamic Programs** 766
14.11 Linear Quadratic Regulation 767
14.12 Why Does it Work?** 770
14.13 Bibliographic Notes 783
Exercises 783
Bibliography 793
15 Backward Approximate Dynamic Programming 795
15.1 Backward Approximate Dynamic Programming for Finite Horizon Problems 797
15.2 Fitted Value Iteration for Infinite Horizon Problems 804
15.3 Value Function Approximation Strategies 805
15.4 Computational Observations 810
15.5 Bibliographic Notes 816
Exercises 816
Bibliography 821
16 Forward ADP I: The Value of a Policy 823
16.1 Sampling the Value of a Policy 824
16.2 Stochastic Approximation Methods 835
16.3 Bellman's Equation Using a Linear Model* 837
16.4 Analysis of TD(0), LSTD, and LSPE Using a Single State* 842
16.5 Gradient-based Methods for Approximate Value Iteration* 845
16.6 Value Function Approximations Based on Bayesian Learning* 852
16.7 Learning Algorithms and Atepsizes 855
16.8 Bibliographic Notes 860
Exercises 862
Bibliography 864
17 Forward ADP II: Policy Optimization 867
17.1 Overview of Algorithmic Strategies 869
17.2 Approximate Value Iteration and Q-Learning Using Lookup Tables 871
17.3 Styles of Learning 881
17.4 Approximate Value Iteration Using Linear Models 886
17.5 On-policy vs. off-policy learning and the exploration-exploitation problem 888
17.6 Applications 894
17.7 Approximate Policy Iteration 900
17.8 The Actor-Critic Paradigm 907
17.9 Statistical Bias in the Max Operator* 909
17.10 The Linear Programming Method Using Linear Models* 912
17.11 Finite Horizon Approximations for Steady-State Applications 915
17.12 Bibliographic Notes 917
Exercises 918
Bibliography 924
18 Forward ADP III: Convex Resource Allocation Problems 927
18.1 Resource Allocation Problems 930
18.2 Values Versus Marginal Values 937
18.3 Piecewise Linear Approximations for Scalar Functions 938
18.4 Regression Methods 941
18.5 Separable Piecewise Linear Approximations 944
18.6 Benders Decomposition for Nonseparable Approximations** 946
18.7 Linear Approximations for High-Dimensional Applications 956
18.8 Resource Allocation with Exogenous Information State 958
18.9 Closing Notes 959
18.10 Bibliographic Notes 960
Exercises 962
Bibliography 967
19 Direct Lookahead Policies 971
19.1 Optimal Policies Using Lookahead Models 974
19.2 Creating an Approximate Lookahead Model 978
19.3 Modified Objectives in Lookahead Models 985
19.4 Evaluating DLA Policies 992
19.5 Why Use a DLA? 997
19.6 Deterministic Lookaheads 999
19.7 A Tour of Stochastic Lookahead Policies 1005
19.8 Monte Carlo Tree Search for Discrete Decisions 1009
19.9 Two-Stage Stochastic Programming for Vector Decisions* 1018
19.10 Observations on DLA Policies 1024
19.11 Bibliographic Notes 1025
Exercises 1027
Bibliography 1031
Part VI - Multiagent Systems 1033
20 Multiagent Modeling and Learning 1035
20.1 Overview of Multiagent Systems 1036
20.2 A Learning Problem - Flu Mitigation 1044
20.3 The POMDP Perspective* 1059
20.4 The Two-Agent Newsvendor Problem 1062
20.5 Multiple Independent Agents - An HVAC Controller Model 1067
20.6 Cooperative Agents - A Spatially Distributed Blood Management Problem 1070
20.7 Closing Notes 1074
20.8 Why Does it Work? 1074
20.9 Bibliographic Notes 1076
Exercises 1077
Bibliography 1083
Index 1085
Este título pertence ao(s) assunto(s) indicados(s). Para ver outros títulos clique no assunto desejado.
Reinforcement learning; stochastic optimization; stochastic search; dynamic programming; approximate dynamic programming; adaptive dynamic programming; optimal control; stochastic programming; multiarmed bandit problems; machine learning; artificial intelligence; active learning
Preface xxv
Acknowledgments xxxi
Part I - Introduction 1
1 Sequential Decision Problems 3
1.1 The Audience 7
1.2 The Communities of Sequential Decision Problems 8
1.3 Our Universal Modeling Framework 10
1.4 Designing Policies for Sequential Decision Problems 15
1.5 Learning 20
1.6 Themes 21
1.7 Our Modeling Approach 27
1.8 How to Read this Book 27
1.9 Bibliographic Notes 33
Exercises 34
Bibliography 38
2 Canonical Problems and Applications 39
2.1 Canonical Problems 39
2.2 A Universal Modeling Framework for Sequential Decision Problems 64
2.3 Applications 69
2.4 Bibliographic Notes 85
Exercises 90
Bibliography 93
3 Online Learning 101
3.1 Machine Learning for Sequential Decisions 102
3.2 Adaptive Learning Using Exponential Smoothing 110
3.3 Lookup Tables with Frequentist Updating 111
3.4 Lookup Tables with Bayesian Updating 112
3.5 Computing Bias and Variance* 118
3.6 Lookup Tables and Aggregation* 121
3.7 Linear Parametric Models 131
3.8 Recursive Least Squares for Linear Models 136
3.9 Nonlinear Parametric Models 140
3.10 Nonparametric Models* 149
3.11 Nonstationary Learning* 159
3.12 The Curse of Dimensionality 162
3.13 Designing Approximation Architectures in Adaptive Learning 165
3.14 Why Does It Work?** 166
3.15 Bibliographic Notes 174
Exercises 176
Bibliography 180
4 Introduction to Stochastic Search 183
4.1 Illustrations of the Basic Stochastic Optimization Problem 185
4.2 Deterministic Methods 188
4.3 Sampled Models 193
4.4 Adaptive Learning Algorithms 202
4.5 Closing Remarks 210
4.6 Bibliographic Notes 210
Exercises 212
Bibliography 218
Part II - Stochastic Search 221
5 Derivative-Based Stochastic Search 223
5.1 Some Sample Applications 225
5.2 Modeling Uncertainty 228
5.3 Stochastic Gradient Methods 231
5.4 Styles of Gradients 237
5.5 Parameter Optimization for Neural Networks* 242
5.6 Stochastic Gradient Algorithm as a Sequential Decision Problem 247
5.7 Empirical Issues 248
5.8 Transient Problems* 249
5.9 Theoretical Performance* 250
5.10 Why Does it Work? 250
5.11 Bibliographic Notes 263
Exercises 264
Bibliography 270
6 Stepsize Policies 273
6.1 Deterministic Stepsize Policies 276
6.2 Adaptive Stepsize Policies 282
6.3 Optimal Stepsize Policies* 289
6.4 Optimal Step sizes for Approximate Value Iteration* 297
6.5 Convergence 300
6.6 Guidelines for Choosing Stepsize Policies 301
6.7 Why Does it Work* 303
6.8 Bibliographic Notes 306
Exercises 307
Bibliography 314
7 Derivative-Free Stochastic Search 317
7.1 Overview of Derivative-free Stochastic Search 319
7.2 Modeling Derivative-free Stochastic Search 325
7.3 Designing Policies 330
7.4 Policy Function Approximations 333
7.5 Cost Function Approximations 335
7.6 VFA-based Policies 338
7.7 Direct Lookahead Policies 348
7.8 The Knowledge Gradient (Continued)* 362
7.9 Learning in Batches 380
7.10 Simulation Optimization* 382
7.11 Evaluating Policies 385
7.12 Designing Policies 394
7.13 Extensions* 398
7.14 Bibliographic Notes 409
Exercises 412
Bibliography 424
Part III - State-dependent Problems 429
8 State-dependent Problems 431
8.1 Graph Problems 433
8.2 Inventory Problems 439
8.3 Complex Resource Allocation Problems 446
8.4 State-dependent Learning Problems 456
8.5 A Sequence of Problem Classes 460
8.6 Bibliographic Notes 461
Exercises 462
Bibliography 466
9 Modeling Sequential Decision Problems 467
9.1 A Simple Modeling Illustration 471
9.2 Notational Style 476
9.3 Modeling Time 478
9.4 The States of Our System 481
9.5 Modeling Decisions 500
9.6 The Exogenous Information Process 506
9.7 The Transition Function 515
9.8 The Objective Function 518
9.9 Illustration: An Energy Storage Model 523
9.10 Base Models and Lookahead Models 528
9.11 A Classification of Problems* 529
9.12 Policy Evaluation* 532
9.13 Advanced Probabilistic Modeling Concepts** 534
9.14 Looking Forward 540
9.15 Bibliographic Notes 542
Exercises 544
Bibliography 557
10 Uncertainty Modeling 559
10.1 Sources of Uncertainty 560
10.2 A Modeling Case Study: The COVID Pandemic 575
10.3 Stochastic Modeling 575
10.4 Monte Carlo Simulation 581
10.5 Case Study: Modeling Electricity Prices 589
10.6 Sampling vs. Sampled Models 595
10.7 Closing Notes 597
10.8 Bibliographic Notes 597
Exercises 598
Bibliography 601
11 Designing Policies 603
11.1 From Optimization to Machine Learning to Sequential Decision Problems 605
11.2 The Classes of Policies 606
11.3 Policy Function Approximations 610
11.4 Cost Function Approximations 613
11.5 Value Function Approximations 614
11.6 Direct Lookahead Approximations 616
11.7 Hybrid Strategies 620
11.8 Randomized Policies 626
11.9 Illustration: An Energy Storage Model Revisited 627
11.10 Choosing the Policy Class 631
11.11 Policy Evaluation 641
11.12 Parameter Tuning 642
11.13 Bibliographic Notes 646
Exercises 646
Bibliography 651
Part IV - Policy Search 653
12 Policy Function Approximations and Policy Search 655
12.1 Policy Search as a Sequential Decision Problem 657
12.2 Classes of Policy Function Approximations 658
12.3 Problem Characteristics 665
12.4 Flavors of Policy Search 666
12.5 Policy Search with Numerical Derivatives 669
12.6 Derivative-Free Methods for Policy Search 670
12.7 Exact Derivatives for Continuous Sequential Problems* 677
12.8 Exact Derivatives for Discrete Dynamic Programs** 680
12.9 Supervised Learning 686
12.10 Why Does it Work? 687
12.11 Bibliographic Notes 690
Exercises 691
Bibliography 698
13 Cost Function Approximations 701
13.1 General Formulation for Parametric CFA 703
13.2 Objective-Modified CFAs 704
13.3 Constraint-Modified CFAs 714
13.4 Bibliographic Notes 725
Exercises 726
Bibliography 729
Part V - Lookahead Policies 731
14 Exact Dynamic Programming 737
14.1 Discrete Dynamic Programming 738
14.2 The Optimality Equations 740
14.3 Finite Horizon Problems 747
14.4 Continuous Problems with Exact Solutions 750
14.5 Infinite Horizon Problems* 755
14.6 Value Iteration for Infinite Horizon Problems* 757
14.7 Policy Iteration for Infinite Horizon Problems* 762
14.8 Hybrid Value-Policy Iteration* 764
14.9 Average Reward Dynamic Programming* 765
14.10 The Linear Programming Method for Dynamic Programs** 766
14.11 Linear Quadratic Regulation 767
14.12 Why Does it Work?** 770
14.13 Bibliographic Notes 783
Exercises 783
Bibliography 793
15 Backward Approximate Dynamic Programming 795
15.1 Backward Approximate Dynamic Programming for Finite Horizon Problems 797
15.2 Fitted Value Iteration for Infinite Horizon Problems 804
15.3 Value Function Approximation Strategies 805
15.4 Computational Observations 810
15.5 Bibliographic Notes 816
Exercises 816
Bibliography 821
16 Forward ADP I: The Value of a Policy 823
16.1 Sampling the Value of a Policy 824
16.2 Stochastic Approximation Methods 835
16.3 Bellman's Equation Using a Linear Model* 837
16.4 Analysis of TD(0), LSTD, and LSPE Using a Single State* 842
16.5 Gradient-based Methods for Approximate Value Iteration* 845
16.6 Value Function Approximations Based on Bayesian Learning* 852
16.7 Learning Algorithms and Atepsizes 855
16.8 Bibliographic Notes 860
Exercises 862
Bibliography 864
17 Forward ADP II: Policy Optimization 867
17.1 Overview of Algorithmic Strategies 869
17.2 Approximate Value Iteration and Q-Learning Using Lookup Tables 871
17.3 Styles of Learning 881
17.4 Approximate Value Iteration Using Linear Models 886
17.5 On-policy vs. off-policy learning and the exploration-exploitation problem 888
17.6 Applications 894
17.7 Approximate Policy Iteration 900
17.8 The Actor-Critic Paradigm 907
17.9 Statistical Bias in the Max Operator* 909
17.10 The Linear Programming Method Using Linear Models* 912
17.11 Finite Horizon Approximations for Steady-State Applications 915
17.12 Bibliographic Notes 917
Exercises 918
Bibliography 924
18 Forward ADP III: Convex Resource Allocation Problems 927
18.1 Resource Allocation Problems 930
18.2 Values Versus Marginal Values 937
18.3 Piecewise Linear Approximations for Scalar Functions 938
18.4 Regression Methods 941
18.5 Separable Piecewise Linear Approximations 944
18.6 Benders Decomposition for Nonseparable Approximations** 946
18.7 Linear Approximations for High-Dimensional Applications 956
18.8 Resource Allocation with Exogenous Information State 958
18.9 Closing Notes 959
18.10 Bibliographic Notes 960
Exercises 962
Bibliography 967
19 Direct Lookahead Policies 971
19.1 Optimal Policies Using Lookahead Models 974
19.2 Creating an Approximate Lookahead Model 978
19.3 Modified Objectives in Lookahead Models 985
19.4 Evaluating DLA Policies 992
19.5 Why Use a DLA? 997
19.6 Deterministic Lookaheads 999
19.7 A Tour of Stochastic Lookahead Policies 1005
19.8 Monte Carlo Tree Search for Discrete Decisions 1009
19.9 Two-Stage Stochastic Programming for Vector Decisions* 1018
19.10 Observations on DLA Policies 1024
19.11 Bibliographic Notes 1025
Exercises 1027
Bibliography 1031
Part VI - Multiagent Systems 1033
20 Multiagent Modeling and Learning 1035
20.1 Overview of Multiagent Systems 1036
20.2 A Learning Problem - Flu Mitigation 1044
20.3 The POMDP Perspective* 1059
20.4 The Two-Agent Newsvendor Problem 1062
20.5 Multiple Independent Agents - An HVAC Controller Model 1067
20.6 Cooperative Agents - A Spatially Distributed Blood Management Problem 1070
20.7 Closing Notes 1074
20.8 Why Does it Work? 1074
20.9 Bibliographic Notes 1076
Exercises 1077
Bibliography 1083
Index 1085
Acknowledgments xxxi
Part I - Introduction 1
1 Sequential Decision Problems 3
1.1 The Audience 7
1.2 The Communities of Sequential Decision Problems 8
1.3 Our Universal Modeling Framework 10
1.4 Designing Policies for Sequential Decision Problems 15
1.5 Learning 20
1.6 Themes 21
1.7 Our Modeling Approach 27
1.8 How to Read this Book 27
1.9 Bibliographic Notes 33
Exercises 34
Bibliography 38
2 Canonical Problems and Applications 39
2.1 Canonical Problems 39
2.2 A Universal Modeling Framework for Sequential Decision Problems 64
2.3 Applications 69
2.4 Bibliographic Notes 85
Exercises 90
Bibliography 93
3 Online Learning 101
3.1 Machine Learning for Sequential Decisions 102
3.2 Adaptive Learning Using Exponential Smoothing 110
3.3 Lookup Tables with Frequentist Updating 111
3.4 Lookup Tables with Bayesian Updating 112
3.5 Computing Bias and Variance* 118
3.6 Lookup Tables and Aggregation* 121
3.7 Linear Parametric Models 131
3.8 Recursive Least Squares for Linear Models 136
3.9 Nonlinear Parametric Models 140
3.10 Nonparametric Models* 149
3.11 Nonstationary Learning* 159
3.12 The Curse of Dimensionality 162
3.13 Designing Approximation Architectures in Adaptive Learning 165
3.14 Why Does It Work?** 166
3.15 Bibliographic Notes 174
Exercises 176
Bibliography 180
4 Introduction to Stochastic Search 183
4.1 Illustrations of the Basic Stochastic Optimization Problem 185
4.2 Deterministic Methods 188
4.3 Sampled Models 193
4.4 Adaptive Learning Algorithms 202
4.5 Closing Remarks 210
4.6 Bibliographic Notes 210
Exercises 212
Bibliography 218
Part II - Stochastic Search 221
5 Derivative-Based Stochastic Search 223
5.1 Some Sample Applications 225
5.2 Modeling Uncertainty 228
5.3 Stochastic Gradient Methods 231
5.4 Styles of Gradients 237
5.5 Parameter Optimization for Neural Networks* 242
5.6 Stochastic Gradient Algorithm as a Sequential Decision Problem 247
5.7 Empirical Issues 248
5.8 Transient Problems* 249
5.9 Theoretical Performance* 250
5.10 Why Does it Work? 250
5.11 Bibliographic Notes 263
Exercises 264
Bibliography 270
6 Stepsize Policies 273
6.1 Deterministic Stepsize Policies 276
6.2 Adaptive Stepsize Policies 282
6.3 Optimal Stepsize Policies* 289
6.4 Optimal Step sizes for Approximate Value Iteration* 297
6.5 Convergence 300
6.6 Guidelines for Choosing Stepsize Policies 301
6.7 Why Does it Work* 303
6.8 Bibliographic Notes 306
Exercises 307
Bibliography 314
7 Derivative-Free Stochastic Search 317
7.1 Overview of Derivative-free Stochastic Search 319
7.2 Modeling Derivative-free Stochastic Search 325
7.3 Designing Policies 330
7.4 Policy Function Approximations 333
7.5 Cost Function Approximations 335
7.6 VFA-based Policies 338
7.7 Direct Lookahead Policies 348
7.8 The Knowledge Gradient (Continued)* 362
7.9 Learning in Batches 380
7.10 Simulation Optimization* 382
7.11 Evaluating Policies 385
7.12 Designing Policies 394
7.13 Extensions* 398
7.14 Bibliographic Notes 409
Exercises 412
Bibliography 424
Part III - State-dependent Problems 429
8 State-dependent Problems 431
8.1 Graph Problems 433
8.2 Inventory Problems 439
8.3 Complex Resource Allocation Problems 446
8.4 State-dependent Learning Problems 456
8.5 A Sequence of Problem Classes 460
8.6 Bibliographic Notes 461
Exercises 462
Bibliography 466
9 Modeling Sequential Decision Problems 467
9.1 A Simple Modeling Illustration 471
9.2 Notational Style 476
9.3 Modeling Time 478
9.4 The States of Our System 481
9.5 Modeling Decisions 500
9.6 The Exogenous Information Process 506
9.7 The Transition Function 515
9.8 The Objective Function 518
9.9 Illustration: An Energy Storage Model 523
9.10 Base Models and Lookahead Models 528
9.11 A Classification of Problems* 529
9.12 Policy Evaluation* 532
9.13 Advanced Probabilistic Modeling Concepts** 534
9.14 Looking Forward 540
9.15 Bibliographic Notes 542
Exercises 544
Bibliography 557
10 Uncertainty Modeling 559
10.1 Sources of Uncertainty 560
10.2 A Modeling Case Study: The COVID Pandemic 575
10.3 Stochastic Modeling 575
10.4 Monte Carlo Simulation 581
10.5 Case Study: Modeling Electricity Prices 589
10.6 Sampling vs. Sampled Models 595
10.7 Closing Notes 597
10.8 Bibliographic Notes 597
Exercises 598
Bibliography 601
11 Designing Policies 603
11.1 From Optimization to Machine Learning to Sequential Decision Problems 605
11.2 The Classes of Policies 606
11.3 Policy Function Approximations 610
11.4 Cost Function Approximations 613
11.5 Value Function Approximations 614
11.6 Direct Lookahead Approximations 616
11.7 Hybrid Strategies 620
11.8 Randomized Policies 626
11.9 Illustration: An Energy Storage Model Revisited 627
11.10 Choosing the Policy Class 631
11.11 Policy Evaluation 641
11.12 Parameter Tuning 642
11.13 Bibliographic Notes 646
Exercises 646
Bibliography 651
Part IV - Policy Search 653
12 Policy Function Approximations and Policy Search 655
12.1 Policy Search as a Sequential Decision Problem 657
12.2 Classes of Policy Function Approximations 658
12.3 Problem Characteristics 665
12.4 Flavors of Policy Search 666
12.5 Policy Search with Numerical Derivatives 669
12.6 Derivative-Free Methods for Policy Search 670
12.7 Exact Derivatives for Continuous Sequential Problems* 677
12.8 Exact Derivatives for Discrete Dynamic Programs** 680
12.9 Supervised Learning 686
12.10 Why Does it Work? 687
12.11 Bibliographic Notes 690
Exercises 691
Bibliography 698
13 Cost Function Approximations 701
13.1 General Formulation for Parametric CFA 703
13.2 Objective-Modified CFAs 704
13.3 Constraint-Modified CFAs 714
13.4 Bibliographic Notes 725
Exercises 726
Bibliography 729
Part V - Lookahead Policies 731
14 Exact Dynamic Programming 737
14.1 Discrete Dynamic Programming 738
14.2 The Optimality Equations 740
14.3 Finite Horizon Problems 747
14.4 Continuous Problems with Exact Solutions 750
14.5 Infinite Horizon Problems* 755
14.6 Value Iteration for Infinite Horizon Problems* 757
14.7 Policy Iteration for Infinite Horizon Problems* 762
14.8 Hybrid Value-Policy Iteration* 764
14.9 Average Reward Dynamic Programming* 765
14.10 The Linear Programming Method for Dynamic Programs** 766
14.11 Linear Quadratic Regulation 767
14.12 Why Does it Work?** 770
14.13 Bibliographic Notes 783
Exercises 783
Bibliography 793
15 Backward Approximate Dynamic Programming 795
15.1 Backward Approximate Dynamic Programming for Finite Horizon Problems 797
15.2 Fitted Value Iteration for Infinite Horizon Problems 804
15.3 Value Function Approximation Strategies 805
15.4 Computational Observations 810
15.5 Bibliographic Notes 816
Exercises 816
Bibliography 821
16 Forward ADP I: The Value of a Policy 823
16.1 Sampling the Value of a Policy 824
16.2 Stochastic Approximation Methods 835
16.3 Bellman's Equation Using a Linear Model* 837
16.4 Analysis of TD(0), LSTD, and LSPE Using a Single State* 842
16.5 Gradient-based Methods for Approximate Value Iteration* 845
16.6 Value Function Approximations Based on Bayesian Learning* 852
16.7 Learning Algorithms and Atepsizes 855
16.8 Bibliographic Notes 860
Exercises 862
Bibliography 864
17 Forward ADP II: Policy Optimization 867
17.1 Overview of Algorithmic Strategies 869
17.2 Approximate Value Iteration and Q-Learning Using Lookup Tables 871
17.3 Styles of Learning 881
17.4 Approximate Value Iteration Using Linear Models 886
17.5 On-policy vs. off-policy learning and the exploration-exploitation problem 888
17.6 Applications 894
17.7 Approximate Policy Iteration 900
17.8 The Actor-Critic Paradigm 907
17.9 Statistical Bias in the Max Operator* 909
17.10 The Linear Programming Method Using Linear Models* 912
17.11 Finite Horizon Approximations for Steady-State Applications 915
17.12 Bibliographic Notes 917
Exercises 918
Bibliography 924
18 Forward ADP III: Convex Resource Allocation Problems 927
18.1 Resource Allocation Problems 930
18.2 Values Versus Marginal Values 937
18.3 Piecewise Linear Approximations for Scalar Functions 938
18.4 Regression Methods 941
18.5 Separable Piecewise Linear Approximations 944
18.6 Benders Decomposition for Nonseparable Approximations** 946
18.7 Linear Approximations for High-Dimensional Applications 956
18.8 Resource Allocation with Exogenous Information State 958
18.9 Closing Notes 959
18.10 Bibliographic Notes 960
Exercises 962
Bibliography 967
19 Direct Lookahead Policies 971
19.1 Optimal Policies Using Lookahead Models 974
19.2 Creating an Approximate Lookahead Model 978
19.3 Modified Objectives in Lookahead Models 985
19.4 Evaluating DLA Policies 992
19.5 Why Use a DLA? 997
19.6 Deterministic Lookaheads 999
19.7 A Tour of Stochastic Lookahead Policies 1005
19.8 Monte Carlo Tree Search for Discrete Decisions 1009
19.9 Two-Stage Stochastic Programming for Vector Decisions* 1018
19.10 Observations on DLA Policies 1024
19.11 Bibliographic Notes 1025
Exercises 1027
Bibliography 1031
Part VI - Multiagent Systems 1033
20 Multiagent Modeling and Learning 1035
20.1 Overview of Multiagent Systems 1036
20.2 A Learning Problem - Flu Mitigation 1044
20.3 The POMDP Perspective* 1059
20.4 The Two-Agent Newsvendor Problem 1062
20.5 Multiple Independent Agents - An HVAC Controller Model 1067
20.6 Cooperative Agents - A Spatially Distributed Blood Management Problem 1070
20.7 Closing Notes 1074
20.8 Why Does it Work? 1074
20.9 Bibliographic Notes 1076
Exercises 1077
Bibliography 1083
Index 1085
Este título pertence ao(s) assunto(s) indicados(s). Para ver outros títulos clique no assunto desejado.
