Revisiting Parameter Sharing in Multi-Agent Deep Reinforcement Learning

J K Terry, Nathaniel Grammel, Ananth Hari, Luis Santos, Benjamin Black

Keywords: Reinforcement Learning, Multi-agent Reinforcement Learning

Abstract: "``Nonstationarity" is a fundamental problem in cooperative multi-agent reinforcement learning (MARL). It results from every agent's policy changing during learning, while being part of the environment from the perspective of other agents. This causes information to inherently oscillate between agents during learning, greatly slowing convergence. We use the MAILP model of information transfer during multi-agent learning to show that increasing centralization during learning arbitrarily mitigates the slowing of convergence due to nonstationarity. The most centralized case of learning is parameter sharing, an uncommonly used MARL method, specific to environments with homogeneous agents. It bootstraps single-agent reinforcement learning (RL) methods and learns an identical policy for each agent. We experimentally replicate our theoretical result of increased learning centralization leading to better performance. We further apply parameter sharing to 8 more modern single-agent deep RL methods for the first time, achieving up to 44 times more average reward in 16% as many episodes compared to previous parameter sharing experiments. We finally give a formal proof of a set of methods that allow parameter sharing to serve in environments with heterogeneous agents.

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

Reviewed Version (pdf): https://openreview.net/references/pdf?id=2QpgE2QPGb