Deep Reinforcement Learning Based Two-phase Proactive Caching for Collaborative Edge Networks

In cache-assisted wireless edge networking, the proactive content caching policy is acknowledged as an effective way for data traffic relief. In this paper, we introduce an online content caching scheme for a mobile edge computing (MEC) system, which consists of multiple edge servers equipped with limited storage capacity. We formulate the collaborative content caching problem as the minimization of long-term average cost of the system under the uncertainties of users' demands and dynamic content popularity. Then, a two-phase proactive caching algorithm based on deep reinforcement learning (RL) is proposed, which successfully break the curse of high dimensionality by redesigning the output layer of the neural network, and adaptively updates the caching decisions for edge servers. The numerical results show that the proposed proactive caching algorithm is robust to large-scale caching scenarios, able to predict the users' future requests with a high accuracy and collaboratively update the caching policies. Compared to four well-known caching schemes, the proposed scheme outperforms on various performance metrics including the average system cost and overall cache hit rate.