Multi-Agent Confrontation Game Based on Multi-Agent Reinforcement Learning

This paper studies the multi-agent confrontation game problem, and takes unmanned aerial vehicle (UAV) offense-defense confrontation as the research object. Deep Deterministic Policy Gradient (DDPG) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm are used for policy optimization. The experimental results show that MADDPG can acquire good policy in multi-agent confrontation game environment. In order to make MADDPG suitable for large-scale multi-agent game problems and obtain robust policy, this paper improves MADDPG by introducing Mean Field Theory (MFT) and ‘Minimax’ idea. The experimental results show that the improved algorithms can deal with large-scale multi-agent game problem and obtain robust policy.