A review of multi-agent deep reinforcement learning for resource allocation in beyond 5G network slicing: solutions, challenges and future research directions

The rapid development of emerging technologies, such as the massive Internet of Things (IoT) and immersive applications, is driving the resource requirements of Beyond Fifth Generation (B5G) mobile networks to evolve in a more complex and dynamic direction. Network Slicing (NS) technology enables the personalized needs of different services by logically dividing the physical network. However, the resource competition between slices, dynamic traffic changes, and global optimization requirements make it difficult for traditional Resource Allocation (RA) methods to satisfy the network requirements of B5G. Deep Reinforcement Learning (DRL) offers an intelligent approach to RA of NS, leveraging its autonomous learning and adaptive capabilities. This study focused on the multi-agent approach of DRL for RA of NS optimization in B5G. It introduced the process of RA in a multi-slice environment, then summarized the key challenges of RA in B5G scenarios, including multi-domain resource coordination, adaptive resource orchestration, and joint optimization of computation and communication resources. At the same time, this study summarized the training process of Multi-Agent DRL (MADRL), then classified the recent RA methods based on DRL into value-based, policy-based and hybrid methods. Additionally, the challenges faced in deploying B5G environments by current optimization methods are highlighted, and future research directions are discussed. By analyzing the practical challenges between advanced DRL algorithms and RA optimization of NS in B5G, this study lays a theoretical foundation for designing scalable and adaptive multi-agent resource allocation optimization schemes in future communication systems.