Science Daily July 25, 2022
Multi-agent reinforcement learning (MARL) is a promising framework for solving complex tasks with many agents. However, a key challenge in MARL is defining private utility functions that ensure coordination when training decentralized agents. This challenge is especially prevalent in unstructured tasks with sparse rewards and many agents. Researchers at the University of Illinois have shown that successor features can help address this challenge by disentangling an individual agent’s impact on the global value function from that of all other agents. They used disentanglement to compactly represent private utilities that support stable training of decentralized agents in unstructured tasks. They implemented their approach using a centralized training, decentralized execution architecture and tested it in a variety of multi-agent environments. The results showed improved performance and training time relative to existing methods and suggested that disentanglement of successor features offers a promising approach to coordination in MARL…read more. Open Access TECHNICAL ARTICLE