Research On Group Game Technologies Based On Attention Mechanism And Entropy Regularization

The intelligence of group games is one of the key points in the field of artificial intelligence nowadays.This problem has unique research prospects and application values in practical scenarios such as the control of drone swarms and the intelligent control of robots.As an important solution to the group game problem,the field of multi-agent reinforcement learning is currently booming.However,the instability of multi-agent reinforcement learning algorithms and the dimensional explosion of information cannot be ignored.Based on the analysis of the current research methods in the field of multi-agent reinforcement learning,this paper aims to strengthen the effect of the algorithm in group game scenarios,and aims at the stability of the algorithm and the effective use of information by the algorithm.Multi-agent reinforcement learning algorithm for group game scenarios.First of all,this paper implements the basic network architecture of the algorithm in this paper based on the basic theories of the existing multi-agent reinforcement learning field,combined with the advantages of the current existing algorithms.Specifically,based on the actor-critic framework,combining the ideas of multi-agent deep deterministic policy gradient algorithm and asynchronous advantage actor-critic algorithm,the overall network structure of the algorithm is realized,including policy networks,value networks,and so on.Secondly,on the basis of the realized network framework,as the information dimension in the multi-agent reinforcement learning algorithm increases,the algorithm lacks attention to information utilization,which leads to the problem of low information utilization efficiency.The attention mechanism is adopted to optimize the networks of the agents.The introduction of the attention mechanism can strengthen the algorithm's use of effective information.Moreover,in view of the insufficient exploration of information in the multiagent reinforcement learning algorithm and the unstable strategy changes during training,combined with the basic principle of entropy,the loss function of the algorithm is modified.This modification not only enhances the use of information by the algorithm,but also improves the stability of the algorithm,making it easier for the algorithm to find the optimal strategy.Compared with the original algorithm,the effect of the final algorithm in this paper has been improved to a certain extent in the researched scenarios,which verifies the rationality and effectiveness of the algorithm.In the intuitive visualization effect display,it can be seen that the algorithm can achieve the expected goal in the studied group game scenarios.