Role Diversity Matters: A Study of Cooperative Training Strategies for Multi-Agent RLDownload PDF


Sep 29, 2021 (edited Oct 05, 2021)ICLR 2022 Conference Blind SubmissionReaders: Everyone
  • Keywords: Multi-agent Reinforcement Learning
  • Abstract: Cooperative multi-agent reinforcement learning (MARL) is making rapid progress for solving tasks in a grid world and real-world scenarios, in which agents are given different attributes and goals. For example, in Starcraft II battle tasks, agents are initialized with the various move, defense, and attack abilities according to their unit types. Current researchers tend to treat different agents equally and expect them to form a joint policy automatically. However, ignoring the differences between agents in these scenarios may bring policy degradation. Accordingly, in this study, we quantify the agent's difference and study the relationship between the agent's role and the model performance via {\bf Role Diversity}, a metric that can describe MARL tasks. We define role diversity from three perspectives: policy-based, trajectory-based, and contribution-based to fully describe the agents' differences. Through theoretical analysis, we find that the error bound in MARL can be decomposed into three parts that have a strong relation to the role diversity. The decomposed factors can significantly impact policy optimization on parameter sharing, communication mechanism, and credit assignment strategy. Role diversity can therefore serve as a flag for selecting a suitable training strategy and helping to avoid possible bottlenecks on current tasks. The main experimental platforms are based on {\bf Multiagent Particle Environment (MPE) }and {\bf The StarCraft Multi-Agent Challenge (SMAC)}, with extensions to ensure the requirement of this study are met. Our experimental results clearly show that role diversity can serve as a robust description for the characteristics of a multi-agent cooperation task and help explain the question of why the performance of different MARL training strategies is unstable according to this description. In addition, role diversity can help to find a better training strategy and increase performance in cooperative MARL.
  • One-sentence Summary: A Study of Cooperative Training Strategies for Multi-Agent RL based on Role Diversity.
0 Replies