Switching-aware multi-agent deep reinforcement learning for target interception

This paper investigates the multi-agent interception problem under switching topology based on deep reinforcement learning. Due to communication restrictions or network attacks, the connectivity between every two intercepting agents may change during the entire tracking process before the successful interception. That is, the topology of the multi-agent system is switched, which leads to a partial missing or dynamic jump of each agent’s observation. To solve this issue, a novel multi-agent level-fusion actor-critic (MALFAC) approach is proposed with a direction assisted (DA) actor and a dimensional pyramid fusion (DPF) critic. Besides, an experience adviser (EA) function is added to the learning process of the actor. Furthermore, a reward factor is proposed to balance the relationship between individual reward and shared reward. Experimental results show that the proposed method performs better than recent algorithms in the multi-agent interception scenarios with switching topologies, which achieves the highest successful interception with the least average steps. The ablation study also verifies the effectiveness of the innovative components in the proposed method. The extensive experimental results demonstrate the scalability of our method in different scenarios.