Coordination in multiagent reinforcement learning systems by virtual reinforcement signals

This paper presents a novel method for on-line coordination in multiagent reinforcement learning systems. In this method a reinforcement-learning agent learns to select its action estimating system dynamics in terms of both the natural reward for task achievement and the virtual reward for cooperation. The virtual reward for cooperation is ascertained dynamically by a coordinating agent who estimates it from the change in degree of cooperation of all agents using a separate reinforcement learning. This technique provides adaptive coordination, requires less communication and ensures agents to be cooperative. The validity of virtual rewards for convergence in learning is verified, and the proposed method is tested on two different simulated domains to illustrate its significance. The empirical performance of the coordinated system compared to the uncoordinated system illustrates its advantages for multiagent systems.