Animal-inspired optimal foraging via a distributed actor-critic algorithm

We consider a group of mobile agents operating in a mission space that collaborate to solve a general dynamic optimization problem. The agents seek to maximize the total reward collected and minimize their energy cost. We construct a control policy specifying how the agents move subject to constraints due to the geometry of the mission space, the presence of obstacles, their sensing range, their available energy, and the need to avoid collisions with other agents. The mission space is discretized and modeled as a graph. We apply the distributed actor-critic method from [1] properly modified to benefit from least squares temporal difference learning. We present one concrete application: air vehicles flying below a forest´s canopy. We demonstrate that the incorporation of bio-inspired features such as eavesdropping, spatial memory, directional sensing, and grouping into the control policy significantly improves its performance compared to a policy from [2] that included no such features.