Finite-Time Bounds for Fitted Value Iteration

In this paper we develop a theoretical analysis of the performance of sampling-based fitted value iteration (FVI) to solve infinite state-space, discounted-reward Markovian decision processes (MDPs) under the assumption that a generative model of the environment is available. Our main results come in the form of finite-time bounds on the performance of two versions of sampling-based FVI. The convergence rate results obtained allow us to show that both versions of FVI are well behaving in the sense that by using a sufficiently large number of samples for a large class of MDPs, arbitrary good performance can be achieved with high probability. An important feature of our proof technique is that it permits the study of weighted Lp-norm performance bounds. As a result, our technique applies to a large class of function-approximation methods (e.g., neural networks, adaptive regression trees, kernel machines, locally weighted learning), and our bounds scale well with the effective horizon of the MDP. The bounds show a dependence on the stochastic stability properties of the MDP: they scale with the discounted-average concentrability of the future-state distributions. They also depend on a new measure of the approximation power of the function space, the inherent Bellman residual, which reflects how well the function space is "aligned" with the dynamics and rewards of the MDP. The conditions of the main result, as well as the concepts introduced in the analysis, are extensively discussed and compared to previous theoretical results. Numerical experiments are used to substantiate the theoretical findings.

[1]  R. Bellman,et al.  FUNCTIONAL APPROXIMATIONS AND DYNAMIC PROGRAMMING , 1959 .

[2]  Arthur L. Samuel,et al.  Some Studies in Machine Learning Using the Game of Checkers , 1967, IBM J. Res. Dev..

[3]  W. Hoeffding.  Probability Inequalities for sums of Bounded Random Variables , 1963 .

[4]  A. L. Samuel,et al.  Some studies in machine learning using the game of checkers. II: recent progress , 1967 .

[5]  G. Wahba,et al.  Some results on Tchebycheffian spline functions , 1971 .

[6]  Vladimir Vapnik,et al.  Chervonenkis: On the uniform convergence of relative frequencies of events to their probabilities , 1971 .

[7]  Norbert Sauer,et al.  On the Density of Families of Sets , 1972, J. Comb. Theory, Ser. A.

[8]  Thomas L. Morin,et al.  COMPUTATIONAL ADVANCES IN DYNAMIC PROGRAMMING , 1978 .

[9]  C. J. Stone,et al.  Optimal Rates of Convergence for Nonparametric Estimators , 1980 .

[10]  C. J. Stone,et al.  Optimal Global Rates of Convergence for Nonparametric Regression , 1982 .

[11]  J. Tsitsiklis,et al.  An optimal multigrid algorithm for continuous state discrete time stochastic control , 1988, Proceedings of the 27th IEEE Conference on Decision and Control.

[12]  John N. Tsitsiklis,et al.  The complexity of dynamic programming , 1989, J. Complex..

[13]  P. Bougerol,et al.  Strict Stationarity of Generalized Autoregressive Processes , 1992 .

[14]  Martin L. Puterman,et al.  Markov Decision Processes: Discrete Stochastic Dynamic Programming , 1994 .

[15]  John Rust.  Using Randomization to Break the Curse of Dimensionality , 1997 .

[16]  Michael I. Jordan,et al.  Reinforcement Learning with Soft State Aggregation , 1994, NIPS.

[17]  Gerald Tesauro,et al.  Temporal Difference Learning and TD-Gammon , 1995, J. Int. Comput. Games Assoc..

[18]  Geoffrey J. Gordon.  Stable Function Approximation in Dynamic Programming , 1995, ICML.

[19]  David Haussler,et al.  Sphere Packing Numbers for Subsets of the Boolean n-Cube with Bounded Vapnik-Chervonenkis Dimension , 1995, J. Comb. Theory, Ser. A.

[20]  Andrew G. Barto,et al.  Improving Elevator Performance Using Reinforcement Learning , 1995, NIPS.

[21]  Wei Zhang,et al.  A Reinforcement Learning Approach to job-shop Scheduling , 1995, IJCAI.

[22]  Leemon C. Baird,et al.  Residual Algorithms: Reinforcement Learning with Function Approximation , 1995, ICML.

[23]  Peter L. Bartlett,et al.  Efficient agnostic learning of neural networks with bounded fan-in , 1996, IEEE Trans. Inf. Theory.

[24]  John Rust.  Numerical dynamic programming in economics , 1996 .

[25]  John N. Tsitsiklis,et al.  Neuro-Dynamic Programming , 1996, Encyclopedia of Machine Learning.

[26]  Dimitri P. Bertsekas,et al.  Reinforcement Learning for Dynamic Channel Allocation in Cellular Telephone Systems , 1996, NIPS.

[27]  R. DeVore,et al.  Nonlinear approximation , 1998, Acta Numerica.

[28]  Alexander J. Smola,et al.  Learning with kernels , 1998 .

[29]  Peter L. Bartlett,et al.  Learning in Neural Networks: Theoretical Foundations , 1999 .

[30]  Peter L. Bartlett,et al.  Neural Network Learning - Theoretical Foundations , 1999 .

[31]  Thomas G. Dietterich,et al.  Efficient Value Function Approximation Using Regression Trees , 1999 .

[32]  Federico Girosi,et al.  Generalization bounds for function approximation from scattered noisy data , 1999, Adv. Comput. Math..

[33]  Nello Cristianini,et al.  An Introduction to Support Vector Machines and Other Kernel-based Learning Methods , 2000 .

[34]  Michael I. Jordan,et al.  PEGASUS: A policy search method for large MDPs and POMDPs , 2000, UAI.

[35]  John N. Tsitsiklis,et al.  Regression methods for pricing complex American-style options , 2001, IEEE Trans. Neural Networks.

[36]  Francis A. Longstaff,et al.  Valuing American Options by Simulation: A Simple Least-Squares Approach , 2001 .

[37]  Tong Zhang,et al.  An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods , 2001, AI Mag..

[38]  Csaba Szepesvári,et al.  Efficient approximate planning in continuous space Markovian Decision Problems , 2001, AI Commun..

[39]  Xin Wang,et al.  Batch Value Function Approximation via Support Vectors , 2001, NIPS.

[40]  Tong Zhang,et al.  Covering Number Bounds of Certain Regularized Linear Function Classes , 2002, J. Mach. Learn. Res..

[41]  Shie Mannor,et al.  PAC Bounds for Multi-armed Bandit and Markov Decision Processes , 2002, COLT.

[42]  John Langford,et al.  Approximately Optimal Approximate Reinforcement Learning , 2002, ICML.

[43]  Adam Krzyzak,et al.  A Distribution-Free Theory of Nonparametric Regression , 2002, Springer series in statistics.

[44]  Martin B. Haugh,et al.  New simulation methodology for finance: duality theory and simulation in financial engineering , 2003, WSC '03.

[45]  Rémi Munos,et al.  Error Bounds for Approximate Policy Iteration , 2003, ICML.

[46]  Sham M. Kakade,et al.  On the sample complexity of reinforcement learning. , 2003 .

[47]  P. Sánchez,et al.  DUALITY THEORY AND SIMULATION IN FINANCIAL ENGINEERING , 2003 .

[48]  Michail G. Lagoudakis,et al.  Least-Squares Policy Iteration , 2003, J. Mach. Learn. Res..

[49]  Abhijit Gosavi,et al.  A Reinforcement Learning Algorithm Based on Policy Iteration for Average Reward: Empirical Results with Yield Management and Convergence Analysis , 2004, Machine Learning.

[50]  Yishay Mansour,et al.  A Sparse Sampling Algorithm for Near-Optimal Planning in Large Markov Decision Processes , 1999, Machine Learning.

[51]  Thomas Uthmann,et al.  Experiments in Value Function Approximation with Sparse Support Vector Regression , 2004, ECML.

[52]  John N. Tsitsiklis,et al.  Feature-based methods for large scale dynamic programming , 2004, Machine Learning.

[53]  William D. Smart,et al.  Interpolation-based Q-learning , 2004, ICML.

[54]  Martin A. Riedmiller.  Neural Fitted Q Iteration - First Experiences with a Data Efficient Neural Reinforcement Learning Method , 2005, ECML.

[55]  Csaba Szepesvári,et al.  Finite time bounds for sampling based fitted value iteration , 2005, ICML.

[56]  Pierre Geurts,et al.  Tree-Based Batch Mode Reinforcement Learning , 2005, J. Mach. Learn. Res..

[57]  Rémi Munos,et al.  Error Bounds for Approximate Value Iteration , 2005, AAAI.

[58]  Susan A. Murphy,et al.  A Generalization Error for Q-Learning , 2005, J. Mach. Learn. Res..

[59]  Liming Xiang,et al.  Kernel-Based Reinforcement Learning , 2006, ICIC.

[60]  Csaba Szepesvári,et al.  Learning Near-Optimal Policies with Bellman-Residual Minimization Based Fitted Policy Iteration and a Single Sample Path , 2006, COLT.

[61]  A. Antos,et al.  Value-Iteration Based Fitted Policy Iteration: Learning with a Single Trajectory , 2007, 2007 IEEE International Symposium on Approximate Dynamic Programming and Reinforcement Learning.

[62]  Dimitri P. Bertsekas,et al.  Stochastic optimal control : the discrete time case , 2007 .

[63]  Csaba Szepesvári,et al.  Learning near-optimal policies with Bellman-residual minimization based fitted policy iteration and a single sample path , 2006, Machine Learning.

[64]  Abhijit Gosavi,et al.  Self-Improving Factory Simulation using Continuous-time Average-Reward Reinforcement Learning , 2007 .