Multi-Agent Cooperative Attacker-Defender-Target Task Decision Based on PF-MADDPG

Abstract

A novel potential function multi-agent deep deterministic policy gradient (PF-MADDPG) algorithm is proposed for the multi-agent Attacker-Defender-Target (ADT). A multi-agent continuous state space and a continuous action space are established. The potential function rewards of target and defenders are designed to accelerate the game confrontation training speed, and the MADDPG algorithm is utilized to obtain effective strategies, so as to describe the influence of different actions on attackers. Finally, simulations are given to verify the effectiveness of the proposed PF-MADDPG algorithm.