Timezone: »
Poster
Minimax-Optimal Off-Policy Evaluation with Linear Function Approximation
Yaqi Duan · Zeyu Jia · Mengdi Wang
Thu Jul 16 06:00 AM -- 06:45 AM & Thu Jul 16 05:00 PM -- 05:45 PM (PDT) @
This paper studies the statistical theory of off-policy evaluation with function approximation in batch data reinforcement learning problem. We consider a regression-based fitted Q-iteration method, show that it is equivalent to a model-based method that estimates a conditional mean embedding of the transition operator, and prove that this method is information-theoretically optimal and has nearly minimal estimation error. In particular, by leveraging contraction property of Markov processes and martingale concentration, we establish a finite-sample instance-dependent error upper bound and a nearly-matching minimax lower bound. The policy evaluation error depends sharply on a restricted $\chi^2$-divergence over the function class between the long-term distribution of target policy and the distribution of past data. This restricted $\chi^2$-divergence characterizes the statistical limit of off-policy evaluation and is both instance-dependent and function-class-dependent. Further, we provide an easily computable confidence bound for the policy evaluator, which may be useful for optimistic planning and safe policy improvement.
Author Information
Yaqi Duan (Princeton University)
Zeyu Jia (Peking University)
Mengdi Wang (Princeton University)
More from the Same Authors
-
2023 : Efficient RL with Impaired Observability: Learning to Act with Delayed and Missing State Observations »
Minshuo Chen · Yu Bai · H. Vincent Poor · Mengdi Wang -
2023 : Scaling In-Context Demonstrations with Structured Attention »
Tianle Cai · Kaixuan Huang · Jason Lee · Mengdi Wang · Danqi Chen -
2023 : Sample-Efficient Learning of POMDPs with Multiple Observations In Hindsight »
Jiacheng Guo · Minshuo Chen · Huan Wang · Caiming Xiong · Mengdi Wang · Yu Bai -
2023 : Principal-Driven Reward Design and Agent Policy Alignment via Bilevel-RL »
Souradip Chakraborty · Amrit Bedi · Alec Koppel · Furong Huang · Mengdi Wang -
2023 : Visual Adversarial Examples Jailbreak Aligned Large Language Models »
Xiangyu Qi · Kaixuan Huang · Ashwinee Panda · Mengdi Wang · Prateek Mittal -
2023 Poster: Score Approximation, Estimation and Distribution Recovery of Diffusion Models on Low-Dimensional Data »
Minshuo Chen · Kaixuan Huang · Tuo Zhao · Mengdi Wang -
2023 Poster: STEERING : Stein Information Directed Exploration for Model-Based Reinforcement Learning »
Souradip Chakraborty · Amrit Bedi · Alec Koppel · Mengdi Wang · Furong Huang · Dinesh Manocha -
2023 Poster: Provably Efficient Representation Learning with Tractable Planning in Low-Rank POMDP »
Jiacheng Guo · Zihao Li · Huazheng Wang · Mengdi Wang · Zhuoran Yang · Xuezhou Zhang -
2023 Poster: Effective Minkowski Dimension of Deep Nonparametric Regression: Function Approximation and Statistical Theories »
Zixuan Zhang · Minshuo Chen · Mengdi Wang · Wenjing Liao · Tuo Zhao -
2022 : Policy Gradient: Theory for Making Best Use of It »
Mengdi Wang -
2022 Poster: Efficient Reinforcement Learning in Block MDPs: A Model-free Representation Learning approach »
Xuezhou Zhang · Yuda Song · Masatoshi Uehara · Mengdi Wang · Alekh Agarwal · Wen Sun -
2022 Poster: Optimal Estimation of Policy Gradient via Double Fitted Iteration »
Chengzhuo Ni · Ruiqi Zhang · Xiang Ji · Xuezhou Zhang · Mengdi Wang -
2022 Poster: Off-Policy Fitted Q-Evaluation with Differentiable Function Approximators: Z-Estimation and Inference Theory »
Ruiqi Zhang · Xuezhou Zhang · Chengzhuo Ni · Mengdi Wang -
2022 Spotlight: Efficient Reinforcement Learning in Block MDPs: A Model-free Representation Learning approach »
Xuezhou Zhang · Yuda Song · Masatoshi Uehara · Mengdi Wang · Alekh Agarwal · Wen Sun -
2022 Spotlight: Off-Policy Fitted Q-Evaluation with Differentiable Function Approximators: Z-Estimation and Inference Theory »
Ruiqi Zhang · Xuezhou Zhang · Chengzhuo Ni · Mengdi Wang -
2022 Spotlight: Optimal Estimation of Policy Gradient via Double Fitted Iteration »
Chengzhuo Ni · Ruiqi Zhang · Xiang Ji · Xuezhou Zhang · Mengdi Wang -
2021 Poster: Sparse Feature Selection Makes Batch Reinforcement Learning More Sample Efficient »
Botao Hao · Yaqi Duan · Tor Lattimore · Csaba Szepesvari · Mengdi Wang -
2021 Spotlight: Sparse Feature Selection Makes Batch Reinforcement Learning More Sample Efficient »
Botao Hao · Yaqi Duan · Tor Lattimore · Csaba Szepesvari · Mengdi Wang -
2021 Poster: Bootstrapping Fitted Q-Evaluation for Off-Policy Inference »
Botao Hao · Xiang Ji · Yaqi Duan · Hao Lu · Csaba Szepesvari · Mengdi Wang -
2021 Poster: Risk Bounds and Rademacher Complexity in Batch Reinforcement Learning »
Yaqi Duan · Chi Jin · Zhiyuan Li -
2021 Spotlight: Risk Bounds and Rademacher Complexity in Batch Reinforcement Learning »
Yaqi Duan · Chi Jin · Zhiyuan Li -
2021 Spotlight: Bootstrapping Fitted Q-Evaluation for Off-Policy Inference »
Botao Hao · Xiang Ji · Yaqi Duan · Hao Lu · Csaba Szepesvari · Mengdi Wang -
2020 : QA for invited talk 7 Wang »
Mengdi Wang -
2020 : Invited talk 7 Wang »
Mengdi Wang -
2020 Workshop: Theoretical Foundations of Reinforcement Learning »
Emma Brunskill · Thodoris Lykouris · Max Simchowitz · Wen Sun · Mengdi Wang -
2020 Poster: Reinforcement Learning in Feature Space: Matrix Bandit, Kernels, and Regret Bound »
Lin Yang · Mengdi Wang -
2020 Poster: Model-Based Reinforcement Learning with Value-Targeted Regression »
Alex Ayoub · Zeyu Jia · Csaba Szepesvari · Mengdi Wang · Lin Yang -
2019 Poster: Sample-Optimal Parametric Q-Learning Using Linearly Additive Features »
Lin Yang · Mengdi Wang -
2019 Oral: Sample-Optimal Parametric Q-Learning Using Linearly Additive Features »
Lin Yang · Mengdi Wang -
2018 Poster: Estimation of Markov Chain via Rank-constrained Likelihood »
XUDONG LI · Mengdi Wang · Anru Zhang -
2018 Oral: Estimation of Markov Chain via Rank-constrained Likelihood »
XUDONG LI · Mengdi Wang · Anru Zhang -
2018 Poster: Scalable Bilinear Pi Learning Using State and Action Features »
Yichen Chen · Lihong Li · Mengdi Wang -
2018 Oral: Scalable Bilinear Pi Learning Using State and Action Features »
Yichen Chen · Lihong Li · Mengdi Wang -
2017 Poster: Strong NP-Hardness for Sparse Optimization with Concave Penalty Functions »
Yichen Chen · Dongdong Ge · Mengdi Wang · Zizhuo Wang · Yinyu Ye · Hao Yin -
2017 Talk: Strong NP-Hardness for Sparse Optimization with Concave Penalty Functions »
Yichen Chen · Dongdong Ge · Mengdi Wang · Zizhuo Wang · Yinyu Ye · Hao Yin