Learning-based Optimization Of Content Offloading In The Framework Of Edge Computing And D2D Communications

With the rapid development of wireless access technology and mobile devices,Internet services and applications are gradually migrating to mobile networks.The ensuing explosion of wireless content services has led to the rapid growth of network traffic load and the decline of user service quality,which poses great challenges for the development of mobile networks in the future.Research shows that caching on the edge of mobile network can significantly reduce network traffic while satisfying content requests of local mobile users.The contents can be requested from the proximity users via device-to-device(D2D)communications,and popular content can be pre-cached by local users.However,the assumption that content popularity and user preference are the same in several existing studies,which is not rigorous and invalid.At the same time,the caching strategies proposed by many studies have some limitations in practical applications.In this paper,we study the edge caching optimization of hierarchical wireless networks.Our aiming is to maximize the traffic load of content offloading by D2 D communications.In particular,based on the analysis of user mobility and social relationship,the edge caching policy with D2 D sharing model is proposed,and the multi-dimensional measurement and analysis of large-scale real D2 D data is conducted by establishing a big data process platform on Spark.Then we replaced it with the Markov Decision Process(MDP)and proposed a kind of intensive learning strategy based on the Q-learning.Additionally,combining with large-scale real-time tracking experiment and using the mobile edge calculation simulation experimental platform in our laboratory,we embed the real data and content offloading algorithm model which consisted of Least Recently Used(LRU),Least Frequently Used(LFU)caching strategies,First-In First-Out(FIFO).It was found that 25.1%,23.2% and 18.0% of performance were improve respectively in hit rate,content offloading and save delay,which verified the effectiveness of the framework we proposed.