Timezone: »
We study offline reinforcement learning (RL), which aims to learn an optimal policy based on a dataset collected a priori. Due to the lack of further interactions with the environment, offline RL suffers from the insufficient coverage of the dataset, which eludes most existing theoretical analysis. In this paper, we propose a \underline{pe}ssimistic variant of the \underline{v}alue \underline{i}teration algorithm (PEVI), which incorporates an uncertainty quantifier as the penalty function. Such a penalty function simply flips the sign of the bonus function for promoting exploration in online RL, which makes it easily implementable and compatible with general function approximators.
Without assuming the sufficient coverage of the dataset, we establish a data-dependent upper bound on the suboptimality of PEVI for general Markov decision processes (MDPs). When specialized to linear MDPs, it matches the information-theoretic lower bound up to multiplicative factors of the dimension and horizon. In other words, pessimism is not only provably efficient but also minimax optimal. In particular, given the dataset, the learned policy serves as the best effort'' among all policies, as no other policies can do better. Our theoretical analysis identifies the critical role of pessimism in eliminating a notion of spurious correlation, which emerges from the
irrelevant'' trajectories that are less covered by the dataset and not informative for the optimal policy.
Author Information
Ying Jin (Stanford University)
Zhuoran Yang (Princeton University)
Zhaoran Wang (Northwestern U)
More from the Same Authors
-
2021 : Optimistic Exploration with Backward Bootstrapped Bonus for Deep Reinforcement Learning »
Chenjia Bai · Lingxiao Wang · Lei Han · Jianye Hao · Animesh Garg · Peng Liu · Zhaoran Wang -
2021 : Randomized Least Squares Policy Optimization »
Haque Ishfaq · Zhuoran Yang · Andrei Lupu · Viet Nguyen · Lewis Liu · Riashat Islam · Zhaoran Wang · Doina Precup -
2023 Poster: Adaptive Barrier Smoothing for First-Order Policy Gradient with Contact Dynamics »
Shenao Zhang · Wanxin Jin · Zhaoran Wang -
2023 Poster: Local Optimization Achieves Global Optimality in Multi-Agent Reinforcement Learning »
Yulai Zhao · Zhuoran Yang · Zhaoran Wang · Jason Lee -
2023 Poster: Achieving Hierarchy-Free Approximation for Bilevel Programs with Equilibrium Constraints »
Jiayang Li · Jing Yu · Boyi Liu · Yu Nie · Zhaoran Wang -
2023 Poster: Enforcing Hard Constraints with Soft Barriers: Safe-driven Reinforcement Learning in Unknown Stochastic Environments »
Yixuan Wang · Simon Zhan · Ruochen Jiao · Zhilu Wang · Wanxin Jin · Zhuoran Yang · Zhaoran Wang · Chao Huang · Qi Zhu -
2022 Poster: Learning from Demonstration: Provably Efficient Adversarial Policy Imitation with Linear Function Approximation »
ZHIHAN LIU · Yufeng Zhang · Zuyue Fu · Zhuoran Yang · Zhaoran Wang -
2022 Poster: Provably Efficient Offline Reinforcement Learning for Partially Observable Markov Decision Processes »
Hongyi Guo · Qi Cai · Yufeng Zhang · Zhuoran Yang · Zhaoran Wang -
2022 Poster: Pessimistic Minimax Value Iteration: Provably Efficient Equilibrium Learning from Offline Datasets »
Han Zhong · Wei Xiong · Jiyuan Tan · Liwei Wang · Tong Zhang · Zhaoran Wang · Zhuoran Yang -
2022 Spotlight: Provably Efficient Offline Reinforcement Learning for Partially Observable Markov Decision Processes »
Hongyi Guo · Qi Cai · Yufeng Zhang · Zhuoran Yang · Zhaoran Wang -
2022 Spotlight: Learning from Demonstration: Provably Efficient Adversarial Policy Imitation with Linear Function Approximation »
ZHIHAN LIU · Yufeng Zhang · Zuyue Fu · Zhuoran Yang · Zhaoran Wang -
2022 Spotlight: Pessimistic Minimax Value Iteration: Provably Efficient Equilibrium Learning from Offline Datasets »
Han Zhong · Wei Xiong · Jiyuan Tan · Liwei Wang · Tong Zhang · Zhaoran Wang · Zhuoran Yang -
2022 Poster: Reinforcement Learning from Partial Observation: Linear Function Approximation with Provable Sample Efficiency »
Qi Cai · Zhuoran Yang · Zhaoran Wang -
2022 Poster: Adaptive Model Design for Markov Decision Process »
Siyu Chen · Donglin Yang · Jiayang Li · Senmiao Wang · Zhuoran Yang · Zhaoran Wang -
2022 Spotlight: Adaptive Model Design for Markov Decision Process »
Siyu Chen · Donglin Yang · Jiayang Li · Senmiao Wang · Zhuoran Yang · Zhaoran Wang -
2022 Spotlight: Reinforcement Learning from Partial Observation: Linear Function Approximation with Provable Sample Efficiency »
Qi Cai · Zhuoran Yang · Zhaoran Wang -
2022 Poster: Pessimism meets VCG: Learning Dynamic Mechanism Design via Offline Reinforcement Learning »
Boxiang Lyu · Zhaoran Wang · Mladen Kolar · Zhuoran Yang -
2022 Poster: Contrastive UCB: Provably Efficient Contrastive Self-Supervised Learning in Online Reinforcement Learning »
Shuang Qiu · Lingxiao Wang · Chenjia Bai · Zhuoran Yang · Zhaoran Wang -
2022 Poster: Welfare Maximization in Competitive Equilibrium: Reinforcement Learning for Markov Exchange Economy »
ZHIHAN LIU · Lu Miao · Zhaoran Wang · Michael Jordan · Zhuoran Yang -
2022 Poster: Human-in-the-loop: Provably Efficient Preference-based Reinforcement Learning with General Function Approximation »
Xiaoyu Chen · Han Zhong · Zhuoran Yang · Zhaoran Wang · Liwei Wang -
2022 Spotlight: Welfare Maximization in Competitive Equilibrium: Reinforcement Learning for Markov Exchange Economy »
ZHIHAN LIU · Lu Miao · Zhaoran Wang · Michael Jordan · Zhuoran Yang -
2022 Spotlight: Pessimism meets VCG: Learning Dynamic Mechanism Design via Offline Reinforcement Learning »
Boxiang Lyu · Zhaoran Wang · Mladen Kolar · Zhuoran Yang -
2022 Spotlight: Human-in-the-loop: Provably Efficient Preference-based Reinforcement Learning with General Function Approximation »
Xiaoyu Chen · Han Zhong · Zhuoran Yang · Zhaoran Wang · Liwei Wang -
2022 Spotlight: Contrastive UCB: Provably Efficient Contrastive Self-Supervised Learning in Online Reinforcement Learning »
Shuang Qiu · Lingxiao Wang · Chenjia Bai · Zhuoran Yang · Zhaoran Wang -
2021 Poster: Decentralized Single-Timescale Actor-Critic on Zero-Sum Two-Player Stochastic Games »
Hongyi Guo · Zuyue Fu · Zhuoran Yang · Zhaoran Wang -
2021 Spotlight: Decentralized Single-Timescale Actor-Critic on Zero-Sum Two-Player Stochastic Games »
Hongyi Guo · Zuyue Fu · Zhuoran Yang · Zhaoran Wang -
2021 Poster: Doubly Robust Off-Policy Actor-Critic: Convergence and Optimality »
Tengyu Xu · Zhuoran Yang · Zhaoran Wang · Yingbin LIANG -
2021 Poster: Randomized Exploration in Reinforcement Learning with General Value Function Approximation »
Haque Ishfaq · Qiwen Cui · Viet Nguyen · Alex Ayoub · Zhuoran Yang · Zhaoran Wang · Doina Precup · Lin Yang -
2021 Poster: Infinite-Dimensional Optimization for Zero-Sum Games via Variational Transport »
Lewis Liu · Yufeng Zhang · Zhuoran Yang · Reza Babanezhad · Zhaoran Wang -
2021 Spotlight: Infinite-Dimensional Optimization for Zero-Sum Games via Variational Transport »
Lewis Liu · Yufeng Zhang · Zhuoran Yang · Reza Babanezhad · Zhaoran Wang -
2021 Spotlight: Doubly Robust Off-Policy Actor-Critic: Convergence and Optimality »
Tengyu Xu · Zhuoran Yang · Zhaoran Wang · Yingbin LIANG -
2021 Spotlight: Randomized Exploration in Reinforcement Learning with General Value Function Approximation »
Haque Ishfaq · Qiwen Cui · Viet Nguyen · Alex Ayoub · Zhuoran Yang · Zhaoran Wang · Doina Precup · Lin Yang -
2021 Poster: Provably Efficient Fictitious Play Policy Optimization for Zero-Sum Markov Games with Structured Transitions »
Shuang Qiu · Xiaohan Wei · Jieping Ye · Zhaoran Wang · Zhuoran Yang -
2021 Poster: On Reward-Free RL with Kernel and Neural Function Approximations: Single-Agent MDP and Markov Game »
Shuang Qiu · Jieping Ye · Zhaoran Wang · Zhuoran Yang -
2021 Poster: Principled Exploration via Optimistic Bootstrapping and Backward Induction »
Chenjia Bai · Lingxiao Wang · Lei Han · Jianye Hao · Animesh Garg · Peng Liu · Zhaoran Wang -
2021 Poster: Global Convergence of Policy Gradient for Linear-Quadratic Mean-Field Control/Game in Continuous Time »
Weichen Wang · Jiequn Han · Zhuoran Yang · Zhaoran Wang -
2021 Spotlight: Global Convergence of Policy Gradient for Linear-Quadratic Mean-Field Control/Game in Continuous Time »
Weichen Wang · Jiequn Han · Zhuoran Yang · Zhaoran Wang -
2021 Oral: On Reward-Free RL with Kernel and Neural Function Approximations: Single-Agent MDP and Markov Game »
Shuang Qiu · Jieping Ye · Zhaoran Wang · Zhuoran Yang -
2021 Spotlight: Principled Exploration via Optimistic Bootstrapping and Backward Induction »
Chenjia Bai · Lingxiao Wang · Lei Han · Jianye Hao · Animesh Garg · Peng Liu · Zhaoran Wang -
2021 Oral: Provably Efficient Fictitious Play Policy Optimization for Zero-Sum Markov Games with Structured Transitions »
Shuang Qiu · Xiaohan Wei · Jieping Ye · Zhaoran Wang · Zhuoran Yang -
2021 Poster: Learning While Playing in Mean-Field Games: Convergence and Optimality »
Qiaomin Xie · Zhuoran Yang · Zhaoran Wang · Andreea Minca -
2021 Poster: Is Pessimism Provably Efficient for Offline RL? »
Ying Jin · Zhuoran Yang · Zhaoran Wang -
2021 Spotlight: Is Pessimism Provably Efficient for Offline RL? »
Ying Jin · Zhuoran Yang · Zhaoran Wang -
2021 Spotlight: Learning While Playing in Mean-Field Games: Convergence and Optimality »
Qiaomin Xie · Zhuoran Yang · Zhaoran Wang · Andreea Minca -
2021 Poster: Risk-Sensitive Reinforcement Learning with Function Approximation: A Debiasing Approach »
Yingjie Fei · Zhuoran Yang · Zhaoran Wang -
2021 Poster: Reinforcement Learning for Cost-Aware Markov Decision Processes »
Wesley A Suttle · Kaiqing Zhang · Zhuoran Yang · Ji Liu · David N Kraemer -
2021 Spotlight: Reinforcement Learning for Cost-Aware Markov Decision Processes »
Wesley A Suttle · Kaiqing Zhang · Zhuoran Yang · Ji Liu · David N Kraemer -
2021 Oral: Risk-Sensitive Reinforcement Learning with Function Approximation: A Debiasing Approach »
Yingjie Fei · Zhuoran Yang · Zhaoran Wang -
2020 Poster: Computational and Statistical Tradeoffs in Inferring Combinatorial Structures of Ising Model »
Ying Jin · Zhaoran Wang · Junwei Lu -
2020 Poster: Breaking the Curse of Many Agents: Provable Mean Embedding Q-Iteration for Mean-Field Reinforcement Learning »
Lingxiao Wang · Zhuoran Yang · Zhaoran Wang -
2020 Poster: Robust One-Bit Recovery via ReLU Generative Networks: Near-Optimal Statistical Rate and Global Landscape Analysis »
Shuang Qiu · Xiaohan Wei · Zhuoran Yang -
2020 Poster: Generative Adversarial Imitation Learning with Neural Network Parameterization: Global Optimality and Convergence Rate »
Yufeng Zhang · Qi Cai · Zhuoran Yang · Zhaoran Wang -
2020 Poster: Provably Efficient Exploration in Policy Optimization »
Qi Cai · Zhuoran Yang · Chi Jin · Zhaoran Wang -
2020 Poster: On the Global Optimality of Model-Agnostic Meta-Learning »
Lingxiao Wang · Qi Cai · Zhuoran Yang · Zhaoran Wang -
2020 Poster: Semiparametric Nonlinear Bipartite Graph Representation Learning with Provable Guarantees »
Sen Na · Yuwei Luo · Zhuoran Yang · Zhaoran Wang · Mladen Kolar -
2019 Poster: On the statistical rate of nonlinear recovery in generative models with heavy-tailed data »
Xiaohan Wei · Zhuoran Yang · Zhaoran Wang -
2019 Oral: On the statistical rate of nonlinear recovery in generative models with heavy-tailed data »
Xiaohan Wei · Zhuoran Yang · Zhaoran Wang -
2018 Poster: The Edge Density Barrier: Computational-Statistical Tradeoffs in Combinatorial Inference »
Hao Lu · Yuan Cao · Junwei Lu · Han Liu · Zhaoran Wang -
2018 Poster: Fully Decentralized Multi-Agent Reinforcement Learning with Networked Agents »
Kaiqing Zhang · Zhuoran Yang · Han Liu · Tong Zhang · Tamer Basar -
2018 Oral: Fully Decentralized Multi-Agent Reinforcement Learning with Networked Agents »
Kaiqing Zhang · Zhuoran Yang · Han Liu · Tong Zhang · Tamer Basar -
2018 Oral: The Edge Density Barrier: Computational-Statistical Tradeoffs in Combinatorial Inference »
Hao Lu · Yuan Cao · Junwei Lu · Han Liu · Zhaoran Wang -
2017 Poster: High-dimensional Non-Gaussian Single Index Models via Thresholded Score Function Estimation »
Zhuoran Yang · Krishnakumar Balasubramanian · Han Liu -
2017 Talk: High-dimensional Non-Gaussian Single Index Models via Thresholded Score Function Estimation »
Zhuoran Yang · Krishnakumar Balasubramanian · Han Liu