Afterstate Reinforcement Learning for Continuous Control

Norio Kosaka

Keywords: Reinforcement Learning, Actor-critic methods

TL;DR: Representing the action by the next state to reduce the complexity of learning the causality for the critic.

Abstract: Humans consider the consequence of taking action in decision-making. In par2 ticular, we imagine what will happen upon executing an option of interest. In actor-critic algorithms, the critic evaluates actions from the actor by explicitly taking the action representation as input whereas the conventional value-based methods such as Deep Q-Network [Mnih et al., 2015] do not explicitly deal with such action information. With the action being input, the critic’s task in the actor7 critic framework can be decomposed as follows; (I) learning the utility of action on the environment, (II) learning the future consequence of the action. Our work aims to enhance the critic’s imagination (I) by utilising the environment model based on the model-based RL framework. To this end, our key insight is that all actions landing on the same next state are equivalent. In continuous action space tasks, robot control and painting, we show the efficacy of our method.

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.

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

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: 5463