Learning Diverse Multiagent Behaviors

Deploying teams of agents for many coordination tasks such as search and rescue missions or deep ocean exploration promises effective solutions. However, these tasks generally require a team of agents to take highly coordinated joint actions, which are difficult to unearth under sparse rewards (returning the same value or zero for many joint actions). The previous works have shown that having a large coverage over behavior space via learning diverse behaviors is an effective method to address reward sparsity. Because multiagent behavior spaces are of higher-dimensions than single-agent spaces, multiagent behavior generation is often intractable. We introduce entropy seeking agents to learn diverse behaviors for multiagent systems to address reward sparsity. Our results show that the promotion of diversity among behaviors within the behavior space effectively, resulting in the discovery of collaborative behaviors.