Content Caching Policy Based on GAN and Distributional Reinforcement Learning

To reduce content transmission power and network load pressure, content caching technology based on a large number of small base stations (SBSs) is considered to be an effective solution. However, due to the limited cache capacity and unknown content popularity, how to design an intelligent content caching policy has become a great challenge. In this paper, we propose a generative adversarial network (GAN) based on the distributional deep Q-Network (DDQN) algorithm, named QGAN, to learn the content caching policy. A content caching network that contains several cooperative SBSs is considered in the case of unknown content popularity, where each SBS fetches cached content from the adjacent SBS or cloud. Moreover, compared with three classical content caching policies and one reinforcement learning algorithm, the performance of the QGAN algorithm is verified. The simulation results show that the convergence rate is improved and the transmission cost is reduced with the proposed algorithm.