Mobility-Aware Pre-Caching Based On Deep Learning

To keep up with the explosively increasing demands for mobile data traffic,the ultra-dense deployment of small cells in 5G allows the user equipmentts to communi-cate with mm Wave,which greatly expand the capacity of cellular networks but with a low-cost and low-power consumption.However,due to the short communication dis-tance of mm Wave,mobile device is more likely to move away from the coverage area of small-cell base station(SBS)and thus has to continually switch between SBSs and and reestablish the connections with the remote servers,which will severely degrade the Quality of Experience(QoE)for the users,especially for some delay-sensitive ap-plications.To address these issues,this thesis conceives a proactive caching solution where the SBSs can prefetch and pre-cache the requested content remaining to be trans-mitted in their local caches along the movement trajectory of mobile users.The traffic offloading brings contents closer to the end users so that they can fetch contents from local caches of SBSs rather than from remote server over backhaul.As a result,mobile users can experience a seamless mobility handover with a lower delay and a higher data rate.Despite of these benefits,without any prior knowledge,it is still intractable to figure out users' movements and determine the content placement in SBSs.The main contributions of this thesis are threefold.Firstly,motivated by the re-cent advances in deep learning,we propose a model-free mobility prediction approach based on the conditional variational autoencoder(CVAE)and Long Short Term Mem-ory(LSTM),which are able to infer the latent information about users' habits from the historical trajectory.Without any prior knowledge,our approach achieves a prediction accuracy of about 81.3%on real-world GPS trajectories.Secondly,we propose a user sojour time prediction approach based on the regression models of random forest and support vector regression.The experiment results show that the error of over 77.9%of the predicted sojour time is less than 20%.Thirdly,we formulate the precaching problem as a Markov decision problem,which jointly consider user mobility and user sojour time.To solve this problem,we develop a pre-caching algorithm based on deep reinforcement learning.Through the performance comparison with three benchmarks,we verify that the proposed algorithm is able to achieve the highest utility and greatly improve the QoE of mobile users.