Deep reinforcement learning with emergent communication for coalitional negotiation games.

Researchers

Journal

Modalities

Models

Abstract

For tasks intractable for a single agent, agents must cooperate to accomplish complex goals. A good example is coalitional games, where a group of individuals forms coalitions to produce jointly and share surpluses. In such coalitional negotiation games, how to strategically negotiate to reach agreements on gain allocation is however a key challenge, when the agents are independent and selfish. This work therefore employs deep reinforcement learning (DRL) to build autonomous agent called DALSL that can deal with arbitrary coalitional games without human input. Furthermore, DALSL agent is equipped with the ability to exchange information between them through emergent communication. We have proved that the agent can successfully form a team, distribute the team’s benefits fairly, and can effectively use the language channel to exchange specific information, thereby promoting the establishment of small coalition and shortening the negotiation process. The experimental results shows that the DALSL agent obtains higher payoff when negotiating with handcrafted agents and other RL-based agents; moreover, it outperforms other competitors with a larger margin when the language channel is allowed.