一种结合 MADDPG 和对比学习的无人机追逃博弈方法

To solve the pursuit and evasion game problem of unmanned aerial vehicles in complex combat environments, a Markov model is established, and reward functions for both pursuer and evader are designed under the zero-sum game concept. A centralized training with distributed execution framework is constructed for multi-agent deep deterministic policy gradient (MADDPG) to solve the Nash equilibrium of the pursuit-evasion game. To address the difficult issue of analytically representing the high-dimensional capture(escape)regions characterized by initial positions of the pursuers and evaders, a deep contrastive learning algorithm based on the MADDPG network is built to indirectly represent the high-dimensional capture (escape) regions through the construction and training of Siamese Network. Simulation results show that the Nash equilibrium solution of the pursuit-evasion game of UAVs under given conditions can be gotten by the MADDPG algorithm, and the accuracy rate of representing high-dimensional capture(escape)regions achieves 95% by the combination of contrastive learning algorithm and the converged MADDPG network.