Adaptive Resource Management for Edge Network Slicing using Incremental Multi-Agent Deep Reinforcement Learning

• DOI: 10.48550/arxiv.2310.17523
• Publication Date: 2023-01-01
• Authors: Haiyuan Li; Yuelin Liu; Xueqing Zhou; Xenofon Vasilakos; Reza Nejabati; Shuangyi Yan; Dimitra Simenidou
• Keywords: Leverage (statistics); Benchmark (surveying); Edge device

Multi-access edge computing provides local resources in mobile networks as the essential means for meeting demands of emerging ultra-reliable low-latency communications. At edge, dynamic requests require advanced resource management adaptive network slicing, including allocations, function scaling and load balancing to utilize only necessary resource-constraint networks. Recent solutions are designed a static number slices. Therefore, painful process optimization is required again with any update on In addition, these intend maximize instant rewards, neglecting long-term scheduling. Unlike efforts, we propose an algorithmic approach based multi-agent deep deterministic policy gradient (MADDPG) optimizing slicing. Our objective two-fold: (i) maximizing slicing benefits terms delay energy consumption, (ii) adapting slice changes. Through simulations, demonstrate that MADDPG outperforms benchmark slicing-based one from literature, achieving stable high performance. Additionally, leverage incremental learning facilitate slices, enhanced performance compared pre-trained base models. Remarkably, this yields superior reward while saving approximately 90% training time costs.