A collaborative cache allocation strategy for performance and link cost in mobile edge computing

Mobile Edge Computing (MEC) represents a novel paradigm dedicated to addressing the challenge of facilitating rapid access to an immense volume of content over mobile networks. However, improper cache placement and usage, coupled with fluctuating requests for cached data at diverse timeframes, exhibits considerable variability. Despite the abundance of optimization techniques, a majority of them lack the adaptive capacities needed to navigate dynamic caching environments efficiently. Furthermore, many studies employ online deep learning methodologies, but a slow convergence speed during the training process can potentially compromise caching performance and hinder dynamic goal adjustment in alignment with realistic provider requirements. We propose an integrative utility function encapsulating the worth of cached content and the cost associated with transmission links. By dynamically modifying weight values, this function can concurrently meet the performance and link cost demands of edge computing caching systems. To enhance the real-time response of the caching policy and the efficiency of deep learning, we introduce a Collaborative two-stage Deep Reinforcement Learning (CDRL) framework for devising the caching policy model. CDRL utilizes Double Deep Reinforcement Learning (DDQN) for pre-training in the caching environment to make pre-caching decisions and employs a Deep State-Action-Reward-State-Action (SARSA) algorithm for online training and caching decision-making. Experimental results convincingly demonstrate the proposed method's efficacy in improving the cache hit rate, service latency, and link cost.