Multi-agents Rounding Strategy based on DeepReinforcement Learning GCMSA Algorithm

Abstract

Aiming at the problem of target rounding by multi-agents in complex environ-ments, this paper proposes a Goal Consistency Reinforcement Learning Approachbased on Multi-head Soft Attention(GCMSA). Firstly, in order to make the modelcloser to reality, the reward function when the target is at different positions andthe target escape strategy are set respectively. Then, the Multi-head soft atten-tion module is used to promote the information cognition of the target amongthe agents, so that the agents can complete the target roundup more smoothly.Finally, in the training phase, this paper introduces cognitive dissonance loss toimprove the sample utilisation. Simulation experiments show that GCMSA isable to obtain a higher task success rate and significantly better than MADDPGin terms of algorithm performance.