UAV Swarm Attack-Defense Confrontation Based on Multi-agent Reinforcement Learning

This paper studies the problem of UAV swarm attack-defense confrontation, which can be viewed as an extension of defending territory game. In this problem, a swarm of intruder UAVs attempt to invade into a territory, which is guarded by a swarm of defender UAVs. This problem is a great challenge to traditional methods. To deal with it, a multi-agent deep reinforcement learning approach is proposed, which is based on the Multi-Agent Deep Deterministic Policy Gradient algorithm (MADDPG). A simulation platform is developed which takes account of UAV flight constraints and simulates a real flight environment. To study the performance of the proposed algorithm, we compare it with DDPG. Experimental results show that the UAVs using the MADDPG algorithm can learn better strategies and achieve better performance.