Multi-agent cross prompt mechanism based on Actor-Critic gradient strategy
Zhaoyu Zhang, Daozhi Wei, Jiong Li, Yan Jun Zhao · 2021
With the continuous integration of artificial intelligence and multi-agent system theory, using the Actor-Critic framework decision-making method suitable for multi-agent systems can better realize dynamic decision-making in multi-agent tasks. In this paper, the evaluation of a single agent in the traditional Critic framework network leads to the inability to achieve communication and cooperation between multiple agents, which leads to problems such as long system learning time or sparse rewards. This paper proposes a multi-agent cross-prompt method that combines reinforcement learning with a multi-agent system. In order to enhance the intelligent level of the system, the cross prompt function of the agent is realized, and focus on analyzing and improving the Actor-Critic action decision rules in the multi-agent system. Made on the basis of improved Actor-critic framework, to build multi-agent detection is performed tasks action decision rules to experiment, In the experiment, the central system evaluation network in the improved algorithm is compared with the traditional algorithm under the closed single-agent evaluation in the same multi-agent system and the cross-prompt framework. The experimental results prove that the improved algorithm is effective and has more application value.