Deep Deterministic Policy Gradient-Based Edge Caching: An Inherent Performance Tradeoff

In this paper, edge caching is investigated subject to time-varying content popularity where no systematic distri-bution of content popularity can be known in advance. For achieving efficient caching updates sequentially, two optimization problems of maximizing the long-term accumulated-cache-hit-ratio (ACHR) and minimizing the long-term average-content-provision-cost (ACPC) are formulated. In order to solve these two problems, a deep deterministic policy gradient (DDPG)-based caching algorithm is proposed, which is capable of pro-cessing large-scale and continuous action space and adjusting the caching strategies based on the historical observations of users' requests. To evaluate the performance of the proposed DDPG-based caching algorithm, a real-world data set from MovieLens is adopted. Simulation results demonstrate that sig-nificant performance gains in terms of both ACHR and ACPC are achieved by the proposed algorithm over existing caching strategies. Furthermore, an inherent performance tradeoff exists between the ACHR and the ACPC, and the balance between these performance metrics requires careful system parameter selection.