Federated Learning Based Caching Policies In Fog Radio Access Networks

The arrival of 5G era and the rapid growth of mobile data traffic pose higher requirements for the quality and speed of wireless data transmissions.Due to the capacity limitation of the fronthaul and the centralized base band unit(BBU),the traditional wireless access network will encounter the situations such as communication congestion and communication interruption in the peak hours,and it is difficult to meet the requirements of communication quality and speed.For this purpose,the fog radio access network(F-RAN)has been proposed as a new networking technology that is capable of relieving the fronthaul burden,and reducing the delay in the process of traditional transmission communication,improving efficiency of local processing and the use of computing resources.F-RAN has becoming the focus and hot spot of academia and industry.In F-RAN,it is critical to design a centralized and distributed caching and computing resource allocation scheme.In addition,federated learning(FL),as a kind of intelligent distributed analysis technology,can effectively reduce the computational complexity and reduce the communication and computing pressure of the central server compared with the centralized artificial intelligence learning.Therefore,this paper intends to carry out research on the FL-based content popularity prediction and cache policy,which is of great value and significance.The contributions of this thesis are summarized as follows:1.Aiming at the problems of large user data,high computational complexity and poor real-time performance in popularity prediction,this paper proposes a FL-based content popularity prediction method,which can improves the prediction efficiency and ensure the prediction accuracy.Specifically,the method of deep convolutional neural network(CNN)is used at each F-AP to carry out the popularity prediction training with local data,and then the model delivered by each F-AP is aggregated at the central server to achieve the popularity prediction and data sub-region processing.At the same time,a detection cycle is set up to monitor the popularity improving real-time performance.Finally,the simulation results show that the convergence rate of the proposed scheme are almost the same as the centralized machine learning scheme.By adjusting the number of communication epoch and local node,the convergence speed can be further improved,while the communication efficiency is guaranteed,thus reducing the MSE,and improving the quality of the model.2.In view of the cache resource management problem,considering the challenge that centralized cache resource management can lead to computing pressure in the cloud and heavy fronthaul burden,this thesis proposes a FL-based cache deployment policy.For those content requests received by the F-AP,a cache hit ratio maximization problem was formulated,subject to the cache capacity constraints at the F-AP.To solve this problem,a FL based content cache and update scheme is designed to make fully use of computing ability of F-AP to conduct the preliminary training of the model based on reinforcement learning locally,and to aggregate and update the model in the cloud center,ensuring that the F-AP continuously cache popular content and increase the cache hit rate.Finally,simulation results show that the proposed algorithm can achieve similar convergence performance as the centralized machine learning scheme,and has better cache hit performance than the existing schemes.In addition,it is shown that a higher cache hit rate can be achieve by our proposed algorithm in the case of smaller cache capacity at the F-AP.