Reward-Sharing Relational Networks in Multi-Agent Reinforcement Learning as a Framework for Emergent Behavior

Most prior works on MARL seek to implement intra-agent complex interactions by explicitly communicating agent actions. However, there have only been a few efforts that examine emergence as arising from complex 'social' interactions or relations based on individual objectives and reward functions. This study is about integrating a user-defined relational network into the MARL setup and evaluating the effects of agent-agent relations on the generation of emergent behaviors. Specifically, we propose a framework that uses the notion of Reward-Sharing Relational Networks (RSRN) to determine the relationship between agents where edge weights determine how much one agent is invested in the success of (or 'cares about') another. The preliminary results indicate that reward-sharing relational networks can effectively influence the learned behaviors towards the imposed relational network.