Mobile agent coordination via a distributed actor-critic algorithm

We consider a group of mobile agents operating in a given mission space with obstacles. We formulate a mission as a general reward collection problem where agents collaborate to maximize the total reward collected. The agents decide how they move but face constraints due to the geometry of the mission space, the presence of obstacles, and the need to avoid collisions with other agents. We discretize the mission space and model it as a graph. We then apply the distributed actor-critic method from modified to benefit from least squares temporal difference learning. We present one concrete application: air vehicles flying below a forest´s canopy. Our simulation results demonstrate the efficiency of the policy produced by our algorithm.