A collaborative combat decision-making method based on multi-agent deep reinforcement learning

In the complex and unknown modern warfare, drone clusters may lose communication with the command center or within the cluster. Confronted with the problem of inflexibility, and fragility when using centralized control format and the issue of complexity when utilizing decentralized format, the multi-agent deep deterministic policy gradient based (MADDPG-based) combat model with centralized training and decentralized execution is proposed to improve the collaborative ability of the drone cluster. The actor network in this model output an action that maximizes the expected rewards for the drone based on the observation, while the critic network only works during the model training to evaluate the actions. This algorithm does not require a differentiable communication channel between drones, which is suitable for the denial combat environment. The experimental results in the one-to-one interception and multiple-to-one interception scenarios demonstrate the good performance of the proposed method, and the drone could cooperate with other drones to intelligently and collaboratively accomplish the strike mission.