Mutli-Agent Deep Reingorcement Learning For Resource Allocation In Multi-Cell System

Global efforts to deploy the fifth generation wireless communication all over the world are well underway,and resource allocation brings terrific push to satisfy the explosively rise in user and traffic capacity in wireless mobile communications.In multi-cell multi-user systems the coverage of one base station extends to the other cells causing interference,which ultimately limits the quality of service offered to the users.A fundamental component of radio resource management is transmitter power control which satisfies the demand of users in transmission rate.As an important technique to suppress co-channel interference,the power control has attracted attentions.By performing uplink power control on inter-cell UEs,the level of inter-cell interference and the impact of interference can be effectively reduced.In order to solve the problem of inter-cell communication interference and further improve the throughput performance of multi-cell multi-user system,this paper proposes:1)A novel and efficient multi-cell multi-user communication system uplink power control method based on Multi-Agent Deep Reinforcement Learning(MADRL).In this paper,a multi-user uplink transmission power optimization problem is modeled first,the problem aims at maximizing the total throughput of multi-cells while taking the user’s quality of service as a constraint,and designs a power control optimization model as quality of service guaranteed.It is verified by simulation experiments that the training speed and convergence speed of the proposed method are faster,and an intelligent and energy-saving multi-cell multi-user power allocation method is proposed while the quality of service of users are guaranteed.2)A pruning algorithm suitable for MADRL.The pruning algorithm adopts structural iterative pruning.According to the characteristics of the MADRL algorithm,the actions chosen by different agents are processed into vectors,and the accuracy of action selection after pruning a neuron represents the importance of the neuron.The experimental results show that the designed pruning algorithm can effectively accelerate the model training and further improve the throughput performance of the proposed intelligent power allocation method,achieving the goal of maximizing the throughput while meeting the quality of service requirements of each user.