Adaptive Reward Sharing to Enhance Learning
in the Context of Multiagent Teams

Abstract:
Real-world populations include diverse social structures, such as sub-groups or teams, that create heterogeneous incentives that complicate coordination. Agents learning to cooperate in these dynamic environments must be able to adapt their internal incentives with their surroundings. We address this challenge with a decentralized multiagent reinforcement learning framework in which each agent has two policies, a low-level behavioral policy and a high-level reward-sharing policy, to adapt both its behavior and its reward function within mixed-motive environments with social structure. We demonstrate the viability of our approach by showing that agents coordinate more effectively through this simultaneous adaptation of heterogeneous reward-sharing configurations. Empirically, our framework enhances population-level outcomes, overcoming sub-optimal initializations and surpassing non-adaptive fully-cooperative baselines in two evaluation domains. Furthermore, the heterogeneous reward-sharing parameterizations that our method learns prove highly effective when applied to new agent populations, yet identifying such configurations through exhaustive or heuristic search is burdensome in this complex search space. By enabling agents to learn and adapt to heterogeneous reward-sharing schemes, our work better captures more diverse dynamics with social structures and varying incentives.