Logarithmic Regret for Reinforcement Learning with Linear Function Approximation

Authors: Jiafan He, Dongruo Zhou, Quanquan Gu

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical In this paper, we show that logarithmic regret is attainable under two recently proposed linear MDP assumptions provided that there exists a positive sub-optimality gap for the optimal actionvalue function. More specifically, under the linear MDP assumption (Jin et al., 2020), the LSVIUCB algorithm can achieve e O(d3H5/gapmin log(T))regret; and under the linear mixture MDP assumption (Ayoub et al., 2020), the UCRL-VTR algorithm can achieve e O(d2H5/gapmin log3(T)) regret, where d is the dimension of feature mapping, H is the length of episode, gapmin is the minimal sub-optimality gap, and e O hides all logarithmic terms except log(T). To the best of our knowledge, these are the first logarithmic regret bounds for RL with linear function approximation. We also establish gap-dependent lower bounds for the two linear MDP models.
Researcher Affiliation Academia Jiafan He 1 Dongruo Zhou 1 Quanquan Gu 1 1Department of Computer Science, University of California, Los Angeles, CA 90095, USA.
Pseudocode Yes Algorithm 1 Least Square Value-iteration with UCB (LSVIUCB) (Jin et al., 2020) and Algorithm 2 UCRL with Value-Targeted Model Estimation (UCRL-VTR) (Jia et al., 2020; Ayoub et al., 2020)
Open Source Code No No statement regarding the release or availability of open-source code for the methodology described in this paper was found.
Open Datasets No This paper focuses on theoretical analysis and does not use or reference any publicly available datasets for training or evaluation.
Dataset Splits No This paper focuses on theoretical analysis and does not report empirical experiments requiring dataset splits.
Hardware Specification No This paper focuses on theoretical analysis and does not report empirical experiments, thus no hardware specifications are mentioned.
Software Dependencies No This paper focuses on theoretical analysis and algorithm design. It does not report empirical experiments that would require specific software dependencies or versions.
Experiment Setup No This paper is theoretical in nature, focusing on algorithm analysis and proofs, and therefore does not include details about an experimental setup or hyperparameters.