Reinforcement distribution in a team of cooperative Q-learning agents

In a Q-learning multi-agent group, agents cooperate each other to perform their assigned task during their learning for increasing the team performance. If the role of each agent clearly specified-which is a very hard task for a supervisor agent- the team will learn more efficiently. Indeed, in this cage each agent reinforced according to its real effect on the team Performance. Assuming an identical role for all agents is the most prevalent technique of current researchers to escape the modeling complexities. But we believe this is not the optimum method for reinforcement distribution. The main goal of this research is to find an indirect evaluation method which evaluates the role of each agent in the team and distributes the reinforcement signal accordingly. The expertness of each agent is used as a criterion to estimate the effect of each agent's action on the team performance. Random and equal reinforcement signal distribution methods are also used in order to evaluate expertness-based reinforcement sharing. In addition, a new test bed, called EPIDEM, is developed to evaluate the proposed methods. The results show, the distribution of the reinforcement signals based on the proposed method improves the team learning speed.