Learning Heterogeneous Agent Collaboration in Decentralized Multi-Agent Systems via Intrinsic Motivation
Under Review at The 24th International Conference on Autonomous Agents and Multiagent Systems (AAMAS-25), 2024
The paper discusses the challenges and solutions in Cooperative Multi-Agent Reinforcement Learning (MARL), particularly under conditions of reward sparsity and agent heterogeneity. It introduces the CoHet algorithm, which is designed for decentralized training settings with partial observability, aiming to address these challenges. The effectiveness of CoHet is empirically validated in various environments, demonstrating its superiority over existing methods in sparse cooperative tasks that necessitate agent diversity. Read more
Download here