Multi-Agent Trust Evaluation Model based on Reinforcement Learning

Abstract

The establishment of trust between agents is an important part of Multi-agent Reinforcement Learning (MARL). Recent studies on multi-agents have paid more attention to the game relationship between agents, ignoring trust is the basis of cooperation among agents. In this article, we propose a trust evaluation model in MARL to evaluate the trustworthiness of the target agent. The trust evaluation of the target agent is composed of direct trust and reputation. Direct trust is obtained from the historical interaction of agents. Reputation is to evaluate the impact of an agent's behavior by counterfactual action of a high-ranking agent. In addition, this paper also proposes a trust-based reward mechanism in MARL, which divides agent rewards into environmental rewards and trust rewards. Trust rewards are obtained through trust evaluation values, which drive the agents to reach a cooperative state and obtain higher long-term rewards. In this paper, we set up experiments in harvest game to verify that the trust evaluation model can effectively promote the cooperation of agents and avoid the problem of agents facing social dilemmas.