Multi-UAVs Air Combat Maneuver Decision Based on MADDPG Algorithm with Introduced Update Switcher

Utilizing multi-agent reinforcement learning algorithms for investigating cooperative air combat has emerged as a prominent research focus. Based on the multi-agent deep deterministic strategy gradient algorithm (MADDPG) and the curriculum learning concept, a two-stage game reinforcement learning training framework is constructed to alleviate the slow adversarial nature of the self-play framework and the exploration difficulty of directly using MADDPG for training in this paper. Furthermore, a collaborative adversarial training framework with introduced update switcher is proposed to provide a unified interpretation of several existing training frameworks. Simulation results demonstrate that the proposed enhanced collaborative adversarial algorithm yields superior performance and efficacy compared to conventional methods.