Finding Better Teammates in a Semi-cooperative Multi-agent System

摘要

Although in semi-cooperative systems, agents are self-interested, they have to help others to get their help in the requirement time. However, designing a distributed method that encourages agents to be truthful and cooperative is challenging. If each agent is able to find useful co-workers to team up with, they are inspired to cooperate with each other, which leads to desirable results for both individuals and the system as a whole. In this paper, a distributed mechanism on the basis of reinforcement learning (RL) is proposed which guides agents to find better teammates i.e. Agents which are cooperative and useful for a long run. A model-based RL is used to model agents' beliefs toward others as transition probabilities where agents try to influence these probabilities in a way they get benefit from. We clarify properties of our system such as Nash equilibrium by some theorems and test the mechanism by applying it to a distributed sensor network designed for target tracking. The simulation results show effectiveness of the method since RL agents gain more in comparison to selfish and random-policy agents. Experiments also indicate that mixed-strategy RL agents benefit more by taking advantage of further synergy produced by forming larger teams.