Decentralized Collaborative Caching in Ultra-Dense Networks

Abstract

This paper investigates the decentralized collaborative edge caching problem in ultra-dense networks (UDNs) via sharing observed information. Different from existing works which assume that each small base station (SBS) exchanges information with fixed neighboring SBSs, this paper considers that each SBS adaptively selects a set of helpful SBSs to exchange information for caching coordination. To realize the caching coordination in real time, we use bandit learning to select helpful information and adopt an actor-critic framework to learn the caching policy from its selected SBSs. In particular, each SBS makes its caching decision via the actor network, and utilizes the critic network to evaluate the caching decision based on the exchanged information. Simulation results show that the proposed algorithm outperforms other benchmark algorithms with lower information exchange overhead. It is also interesting to find that exchanging information with more nearby SBSs may not help to improve the performance when the content popularity is heterogeneously distributed in different SBSs' regions.