Policy Gradient without Boostrapping via Truncated Value Learning

Matthew Aitchison, Penny Sweetser, Gregoire Deletang, Marcus Hutter

Primary Area: reinforcement learning

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.

Keywords: Reinforcement Learning, Policy Gradient, Discounting, Atari

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Fixed-horizon learning improves performance in high-noise policy-gradient settings.

Abstract: Reinforcement learning algorithms have typically used discounting to reduce the variance of return estimates. However, this reward transformation causes the agent to optimize an objective other than what is specified by the designer. We present a novel deep policy gradient algorithm, \textit{Truncated Value Learning} (TVL), which can learn rewards \textit{discount free} while simultaneously learning value estimates for \textit{all} summable discount functions. Moreover, unlike many other algorithms, TVL learns values without bootstrapping. We hypothesize that bootstrap-free learning improves performance in high-noise environments due to reduced error propagation. We tested TVL empirically on the challenging high-noise \textit{Procgen} benchmark and found it outperformed the previous best algorithm, Phasic Policy Gradient. We also show that our method produces state-of-the-art performance on the challenging long-horizon Atari game \env{Skiing} while using less than 1\% of the training data of the previous best result.

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.

Submission Number: 1425