Keywords: Multi-Agent Reinforcement Learning, Coordinated Strategy, Hierarchical Multi-Agent learning, Episodic Memory
TL;DR: We present a framework which expedites and stabilizes learning of a hierarchical multi-agent reinforcement learning via episodic memory, while achieving coordinated behaviors among agents with a novel theoretic regularization.
Abstract: An agent's strategy can be considered as a subset of action spaces, specialized in certain goals. This paper introduces a coordinated Strategy Identification Multi-Agent reinforcement learning (MARL) with episodic memory, called SIMA. SIMA derives a new temporal difference (TD) target to increase the sample efficiency. The efficiency is achived by keeping the best returns and corresponding to the best joint strategies for given states. This TD target with an additive strategy mixer automatically switches between an episodic control and a conventional Q-learning according to the existence of similar memories. In addition, each agent needs to behave similarly according to its strategy trajectory for coordinated behaviors among agents and coherent evaluation of a group's joint strategies. To this end, SIMA introduces a theoretical regularization for action policies to maximize the mutual information between an agent’s trajectory and its specified strategy. We demonstrate its significant performance improvement on the StarCraft Multi-Agent Challenge benchmark.
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)