Scaling Up Multi-agent Reinforcement Learning in Complex Domains

TD-FALCON (temporal difference-fusion architecture for learning, cognition, and navigation) is a class of self-organizing neural networks that incorporates temporal difference (TD) methods for real-time reinforcement learning. In this paper, we present two strategies, i.e. policy sharing and neighboring-agent mechanism, to further improve the learning efficiency of TD-FALCON in complex multi-agent domains. Through experiments on a traffic control problem domain and the herding task, we demonstrate that those strategies enable TD-FALCON to remain functional and adaptable in complex multi-agent domains.