Distributed Training Decentralized Execution Framework of Multi-agent Learning for Cooperative Edge Caching

The edge caching technology has a great potential to improve the performance of modern wireless networks. The transmission bottleneck at the cloud server will occur when the centralized training decentralized execution (CTDE) framework is applied to find an effective caching policy running on each base station (BS) in a multi-BS cooperative caching scenario. In this paper, we propose a cooperative edge caching method based on the distributed training decentralized execution (DTDE) framework to address this issue. We further propose two DTDE based caching models, namely Single Neighbor DTDE (SN-DTDE) and Local Information Only Executing SN-DTDE (LE-SN-DTDE), to deal with the transmission bottleneck problem at BSs in DTDE. Our simulation results indicate that the models proposed can effectively address the transmission bottleneck problem and they all outperform CTDE on average user access delay in networks with training information loss. Among these new caching models, LESN-DTDE has the best performance when the network experience heavy training information loss.