Learning Collaboration in Reactive Agents Ensembles

Tudor Berariu, Adina Magda Florea

Abstract: In this paper we present an empirical study on using reinforcement learning techniques in reactive multi-agent systems where agents have local perception of the environment and limited communication capabilities. Agents have no a priori information about the task to be solved in the environment and no interpreted representation of the sensory input. We investigate a scenario in which agents receive a higher reward if they coordinate to solve the proposed task. We show that using a variant of Q-Learning agents can learn to value collaboration and self-organize to get higher rewards. The results are promising, but better techniques are suggested to solve the problems that arise from state space explosion.