Multiagent Credit Assignment in a Team of Cooperative Q-Learning Agents with a Parallel Task

Traditionally in many multiagent reinforcement learning researches, qualifying each individual agent’s behavior is responsibility of environment’s critic. However, in most practical cases, critic is not completely aware of effects of all agents’ actions on the team performance. Using agents’ learning history, it is possible to judge the correctness of their actions. To do so, we use team common credit besides some suitable criteria. This way an internal critic distributes the environment’s reinforcement among the agents. Continuing our previous research [1], in this paper three such criteria, named Certainty, Normal Expertness and Relative Normal Expertness, for a team of agents with a parallel task are introduced and compared. It is experimentally shown that these criteria can be used to learn from a common team credit in reasonable time.