Research On Multi-agent Communication Based On Reinforcement Learning

With the increasing application and popularization of 5G communication technology,people’s demand for communication continue to rise.It is not only necessary to achieve high speed and high reliability,but also more intelligent and humanized interaction methods,so as to better meet the communication requirements in different demand scenarios.In view of the current situation of channel capacity approaching Shannon limit in traditional communication and the huge energy consumption of communication system,this paper investigates the communication between multi-agent based on reinforcement learning,aiming to improve the transmission efficiency of communication and further improve the communication speed between agents,so as to change the behavior of agents to improve the execution efficiency of cooperative tasks.The main research work is as follows:1.A framework of pragmatic communication system is designed,in which the difference between semantics and pragmatics is investigated from the perspective of linguistics.And integrates the communication subject into the communication system,the content of the communication is the language formed by training,with target and environment construction as input,task completion as output,and action performance as an evaluation criterion define the pragmatic communication,which is used to better accomplish specific tasks in specific environments.This paper also summarizes the characteristics of pragmatic communication,especially its two advantages of lightness and confidentiality.2.Based on the framework and process of the designed pragmatic communication system,the glue neural layer is designed.As the intermediary between two agents,Glue Neural Layer connects two agents to form a deeper neural network for more efficient communication training.At the same time,a new activation function is designed,which used Softmax function instead of Logistic function to enlarge the glue neural layer,and used Gumbel to distribute random re-parameters to transfer parameters more accurately at the initial stage of training,thus faster convergence.Finally,temperature parameter is introduced,which is more conducive to lightweight communication.3.Based on the game model of the Prisoner’s Dilemma,a simulation experimental scene model,which was set as four prisoners and an interrogation room.By changing the dimension of glue neural layer and activation function to study the performance of pragmatic communication and the influence of glue neural layer.The experiment proved that Gumbel-Softmax is more effective in spreading and back-spreading messages,superior to the existing Logistic activation function,and can significantly improve the speed of communication between agents,so as to improve the execution efficiency of cooperative tasks.