UCAV Path Planning Algorithm Based on Deep Reinforcement Learning

摘要

In the field of the Unmanned Combat Aerial Vehicle (UCAV) confrontation, traditional path planning algorithms have slow operation speed and poor adaptability. This paper proposes a UCAV path planning algorithm based on deep reinforcement learning. The algorithm combines the non-cooperative game idea to build the UCAV and radar confrontation model. In the model, the UCAV must reach the target area. At the same time, in order to complete the identification of the radar communication signal based on ResNet-50 migration learning, we use the theory of Cyclic Spectrum (CS) to process the signal. With the kinematics mechanism of the UCAV, the radar detection probability and the distance between the UCAV and center of the target area are proposed as part of the reward criteria. And we make the signal recognition rate as another part of the reward criteria. The algorithm trains the Deep Q-Network (DQN) parameters to realize the autonomous planning of the UCAV path. The simulation results show that compared with the traditional reinforcement learning algorithm, the algorithm can improve the system operation speed. The accuracy reaches 90% after 300 episodes and the signal recognition rate reaches 92.59% under 0 dB condition. The proposed algorithm can be applied to a variety of electronic warfare environment. It can improve the maneuver response time of the UCAV.