Agents that Learn to Vote for a Joint Action Through Multi-Agent Reinforcement Learning

In multi-agent reinforcement learning (MARL), it is desired to achieve the overall cooperation in a decentralized manner without compromising each agent' s autonomy. In this regard, there arises a strong relation between reward design as a mechanism to incentivize for a cooperative task and assigning credit according to each agent contribution. Towards this end, we propose a framework that applies the concept of the Groves-Clarke pivotal mechanism as a reward design for MARL. We demonstrate the efficiency on maximizing the gross profit of society when voting on public good.