A REINFORCEMENT LEARNING BASED UAVS AIR COLLISION AVOIDANCE

摘要

In this paper, we propose to deal with theUAV airspace conflict resolution problem. Wepropose to search near optimal conflict freepolicies in virtue of the model-basedreinforcement learning. We first analyze theUAV airspace conflict problem and the basicconditions in ensuring collision-free planning,and then discuss the features that effect theoptimal action. We then propose thereinforcement learning based conflict resolutionalgorithm. In the model-based learningstructure, we consider the simplified dynamicsof the UAVS in the model, and employ theheuristic method to estimate the state-actionvalue. In the multi-dimension, continuous space,the optimal policy search method is utilized tofind the near optimal policy. The experiencefrom the real environment is used to criticize themodel-based learning policy. In the end, weapply simulation experiments to demonstrate theproposed algorithm.