Elastic Network Cache Control Using Deep Reinforcement Learning

Thanks to the development of virtualization technology, content service providers can flexibly lease virtualized resources from infrastructure service providers when they deploy the cache nodes in edge networks. As a result, they have two orthogonal objectives: to maximize the caching utility on the one hand and minimize the cost of leasing the cache storage on the other hand. This paper presents a caching algorithm using deep reinforcement learning (DRL) that controls the caching policy with the content time-to-live (TTL) values and elastically adjusts the cache size according to a dynamically changing environment to maximize the utility-minus-cost objective. We show that, under non-stationary traffic scenarios, our DRL-based approach outperforms the conventional algorithms known to be optimal under stationary traffic scenarios.