A Reinforcement Learning Based Uavs Air Collision Avoidance

In this paper, we propose to deal with the UAV airspace conflict resolution problem. We propose to search near optimal conflict free policies in virtue of the model-based reinforcement learning. We first analyze the UAV airspace conflict problem and the basic conditions in ensuring collision-free planning, and then discuss the features that effect the optimal action. We then propose the reinforcement learning based conflict resolution algorithm. In the model-based learning structure, we consider the simplified dynamics of the UAVS in the model, and employ the heuristic method to estimate the state-action value. In the multi-dimension, continuous space, the optimal policy search method is utilized to find the near optimal policy. The experience from the real environment is used to criticize the model-based learning policy. In the end, we apply simulation experiments to demonstrate the proposed algorithm.