A contextual multi-armed bandit approach to caching in wireless small cell network

In this paper, we study the problem of cache content placement in the wireless small cell networks. We consider the small base station (SBS) has a cache unit and limited backhaul capacity, and the SBS cache the popular content to serve local users request while reducing the traffic load on the backhaul link. The goal of the cache unit in SBS is to maximum the traffic offloading from backhaul link, this problem can be seen as a knapsack problem when the content popularity has known in advance. However, content popularity is an index which is constantly changing, and it is difficult to obtain. Hence, we model the content popularity as a linear model based on the context information of the system, the problem becomes a contextual multi-armed bandit (CMAB) problem. We present an online learning algorithm, in which the SBS can learn content popularity by maintain a credible linear model and refresh the cache content over time. We give the regret bound of our algorithm, which prove that our algorithm can converges to the optimal SBS caching strategy. Our simulation results show that our algorithm could quickly learn the content popularity and outperform the reference algorithms.