Mutli-agent consensus under communication failure using Actor-Critic Reinforcement Learning

This paper addresses the problem of achieving multi-agent consensus under sudden total communication failure. The agents are assumed to be moving along the periphery of a circle. The proposed solution uses the actor-critic reinforcement learning method to achieve consensus, when there is no communication between the agents. A performance index is defined that take into consider the difference in angular position between the neighbouring agents. The actions of each agent while achieving consensus with full communication is learned by an actor neural network, while the critic neural network learns to predict the performance index. The proposed solution is validated by a numerical simulation with live agents moving along the periphery of a circle.