Research On Max-weight Scheduling Algorithm In Wireless Network Via Graph Neural Network And Reinforcement Learning

In recent years,with the development of wireless communication technology,more and more applications based on the wireless network have been born.Applications of the Internet of Things(Io Ts),such as unmanned driving,drones,and smart grids,require continuous improvement in the quality of service(Qo S)indicators such as communication bandwidth and latency.This also makes large amounts of infrastructure need to share limited channel resources(such as space,frequency domain,time domain,etc.)in the wireless network.Therefore,as the key to solving the problem of resource allocation in wireless networks,scheduling has always been a core and hot issue in wireless communication research.This paper studies the classic algorithm in wireless scheduling-Max-Weight Scheduling(MWS),which is a scheduling scheme based on the conflict graph.Traditional MWS algorithms are implemented through distributed offline algorithms,which require complicated manual settings,and with the channel fading,they cannot guarantee the scheduling performance in each time slot.The MWS problem can be seen as a combined optimization problem that solves the Maximum Weight Independent Set(MWIS)problem in each time slot.In recent years,with the development of machine learning(ML),deep learning(DL)and reinforcement learning(RL)have been used to solve complex combinatorial optimization problems.Based on this,this paper proposes an online learning scheduling method based on ML,using Graph Neural Network(GNN)and RL algorithms to solve the MWIS problem in each time slot.Specifically,this paper proposes a model named temporal port numbering graph neural network(Temporal Port Numbering Graph Neural Network,TPNGNN),which is composed of a memory module and a GNN model based on the message propagation mode.The structural features and the temporal features of the conflict graph are extracted,and the features are embedded as the probabilistic feature.Using this probability feature as the initial state,an efficient RL algorithm—Asynchronous Advantage Actor-Critic(A3C)is used to search for scheduling schemes,and the scheduling scheme is used as a label to the TPNGNN model for training.In addition,in view of the characteristics of the MWS problem,this paper proposes a weighted color feature as the input feature of the TPNGNN model,which can help the model achieve better convergence.This paper theoretically proves that the combination of the TPNGNN model and the A3C algorithm can achieve the same approximate ratio 1/?_maxas the traditional offline MWS algorithm.And by setting experiments to test the scheduling performance of the model,it proves that the model proposed in this paper has better results in scheduling performance such as throughput,network congestion,and network stability.In the case of full load intensity,this model can reduce the queue length by 19%compared with the offline algorithm.In addition,this model has better scalability.In the setting of 100 links,it can increase the throughput by 21.6%compared with the offline algorithm.