Personalized Collaborative Edge Caching With Federated Transfer Deep Reinforcement Learning

In this letter, collaborative edge caching is investigated in a fog radio access network (F-RAN), wherein multiple fog access points (F-APs) adjust their caching decisions under the coordination of a macro base station (MBS). Owing to the distinct user preferences across geographic areas, content popularity exhibits non-IID characteristics, which degrades the performance of traditional federated learning (FL)-based caching approaches. In view of this, a federated transfer deep reinforcement learning (FTDRL)-based personalized collaborative edge caching algorithm is proposed. To be specific, in view of both the holistic and local characteristics on user preferences, the neural networks are split into base and personalization layers, respectively. The former is aggregated periodically by the MBS to form the global model, whereas the latter is maintained by each F-AP locally. Experiments conducted on a real-world dataset demonstrate the superior performance and robustness achieved by the proposed FTDRL over the state-of-the-art.