Cooperative Edge Caching via Federated Deep Deterministic Policy Gradient Learning in Fog-RANs

In this paper, the cooperative edge caching problem in fog radio access networks (F-RANs) is investigated. On account of the non-deterministic polynomial hard (NP-hard) property of this problem, a federated deep deterministic policy gradient learning (FDDPG) based caching policy is proposed. By considering the dynamic content popularity and time varying requested content information, deep deterministic policy gradient learning (DDPG) is adopted to make optimal caching decisions. As the action selection is directly generated by policy network, DDPG can effectively solve the high dimensional action space caching problem. To address the over-consumption of computational resources, slow network convergence and leakage of user sensitive data, we apply reward weighted horizontal federated learning (RWHFL) in the training of DDPG network. Simulation results show the superiority of our proposed policy compared with the baselines in reducing the average content request delay.