Highway Reinforcement LearningDownload PDF


22 Sept 2022, 12:32 (modified: 19 Nov 2022, 08:12)ICLR 2023 Conference Blind SubmissionReaders: Everyone
Keywords: reinforcement learning, off-policy learning, credit assignment, Bellman Equation
TL;DR: a novel adaptive multi-step Bellman Optimality Equation for efficient credit assignment that converges to the optimal value function with better contraction rate under mild assumptions
Abstract: Traditional Dynamic Programming (DP) approaches suffer from slow backward credit-assignment (CA): only a one-step search is performed at each update. A popular solution for multi-step CA is to use multi-step Bellman operators. Unfortunately, in the control settings, existing methods typically suffer from the large variance of multi-step off-policy corrections or are biased, preventing convergence. To overcome these problems, we introduce a novel multi-step Bellman optimality equation with adaptive lookahead steps. We first derive a new multi-step Value Iteration (VI) method that converges to the optimal Value Function (VF) with an exponential contraction rate but linear computational complexity. Given some trial, our so-called Highway RL performs rapid CA, by picking a policy and a possible lookahead (up to the trial end) that maximize the near-term reward during lookahead plus a DP-based estimate of the cumulative reward for the remaining part of the trial. Highway RL does not require off-policy corrections. Under mild assumptions, it achieves better convergence rates than the traditional one-step Bellman Optimality Operator. We then derive Highway Q-Learning, a convergent multi-step off-policy variant of Q-learning. We show that our Highway algorithms significantly outperform DP approaches on toy tasks. Finally, we propose a deep function approximation variant called Highway DQN. We evaluate it on visual MinAtar Games, outperforming similar multi-step methods.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics
Submission Guidelines: Yes
Please Choose The Closest Area That Your Submission Falls Into: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)
11 Replies