Cooperative caching algorithm for mobile edge networks based on multi-agent meta reinforcement learning

Edge caching reduces service latency, relieves backhaul link traffic pressure, and enhances the quality of experience by multiplexing network content. Many studies use deep reinforcement learning (DRL) methods to develop edge caching policies. However, these traditional DRL methods have limitations, such as lengthy training time, poor model generalization, and the need to relearn network parameters for new tasks. To overcome these challenges, this paper proposes a multi-agent meta reinforcement learning-based cooperative edge caching algorithm (MAMRC), which consists of an inner and an outer model. The inner model employs the multi-agent deep reinforcement learning (MADRL) algorithm to implement cooperative caching for distributed base stations (BSs). It improves the cache hit rate and reduces service latency. The outer model uses a meta-learning method to learn meta-parameters and initialize the inner model, which enhances the generalization capability of the inner model, allowing it to rapidly adapt to new tasks. Experiment results indicate that, compared to the traditional DRL method and the MADRL-based algorithm, the inner model caching performance considering edge collaboration is improved by 15.35% and 4.55% respectively. Notably, compared to traditional caching algorithms and the inner model without initialization with meta-parameters, MAMRC demonstrates superior average caching performance and stronger generalization ability when facing new tasks.