Research On Deep Learning-based Edge Caching In Fog Radio Access Networks

With unprecedented increase of smart devices and cellular users,data traffic grows explosively in recent years.Since the wireless resources of cellular networks are limited,massive mobile data traffic leads to the congestion of backhaul links.Considering that most of the data traffic is consumed in transmitting the same contents,the fog radio access network(F-RAN)has emerged as a promising solution to avoid the congestion problem.In F-RANs,with limited caching and computing resources,fog access points(F-APs)are able to cache some popular contents(which are at the edge of the network and close to users)during off-peak traffic periods.On account of the storage constraints,F-APs need to predict content popularity accurately to make optimal caching decisions.In this paper,we investigate the edge caching problem in F-RANs.Firstly,popularity-based content caching policy is studied.Assume content popularity is known,to guarantee the cache efficiency in the whole time period,the content caching policy includes a proactive caching scheme and a reactive caching scheme.In the proactive caching scheme,the multi-node cooperative caching is modeled as a binary integer linear programming problem to minimize the network cost.By solve this problem,the optimal content deployment scheme can be obtained.In the reactive caching scheme,to further reduce the network cost in practice,the contents in the local caches are adjusted according to the user requests.Simulation results show that the caching policy can significantly reduce the network cost.Secondly,deep learning-based content popularity prediction method is studied.Considering the spatialtemporal popularity dynamic,trends dynamic,users dynamic and limited resources in F-APs,the method learns a location prediction model for every user and a popularity prediction model for every popularity trend class by training a bidirectional long short-term memory(Bi-LSTM)network with only the number of requests.User location prediction makes popularity prediction more targeted.Specifically,in order to guarantee the popularity prediction accuracy after the classifying,deep learning and k-Nearest Neighbor(k NN)are combined to classify the popularity trends.Simultaneously,we develop a loss function to avoid overfitting and increase sensitivity to high popularity for popularity prediction models.Simulation results show that the method can predict popularity with a high accuracy and the prediction accuracy of the method is superior to that of the traditional methods.Finally,deep learning-based edge caching policy is studied.By combining the content caching policy and popularity prediction method mentioned previously,the policy can be applied in F-RANs in practice.Analytical results show that the policy can be performed with low computational complexity.Besides,the architectures of the F-AP and user equipment(UE)are designed to enable the proposed policy.Simulation results show that the edge caching policy can reduce the network cost compared to the traditional policies.