Multi-agent Deep Reinforcement Learning for Dynamic Motion Control Allocation in UAV Swarm Cooperative Jamming to Netted Radar

Abstract

Netted radar has excellent anti-jamming ability because of its characteristics of cooperative detection and mutual verification. UAV swarm cooperative jamming is an effective technology against netted radar. Traditional algorithms typically cause interference to netted radar by signal layer resource allocation. But the advantages of UAV swarm cooperative jamming are not fully exploited. In this paper, a distributively executed dynamic control allocation algorithm in motion layer based on multi-agent deep reinforcement learning is proposed. We established a model of netted radar detection and UAV swarm cooperative jamming. The positions and angles of the netted radar beams will change regularly. And the UAV swarm adjusts the motion state and position of each individual to deceive the netted radar and starts up RGPO. The motion policy of UAV swarm is optimized in the model of SA-MATD3 algorithm. Novel designs of state space and action are proposed in this paper. And a new reward function is designed based on Hungarian algorithm to solve the problem of reasonable matching between netted radar and UAV swarm. Compared with the Euclidean distance algorithm, the new reward function avoids the phenomenon of “lazy agents”. The proposed algorithm achieves dynamic motion control allocation in UAV swarm cooperative jamming to netted radar. And the reward function with Hungarian algorithm performs more training effectiveness than other designs.