Dynamic Resource Reconfiguration for Network Slicing: An Incremental Multi-Agent Reinforcement Learning Based Approach

Network slicing is a virtualization paradigm that divides multiple isolated logical networks on the physical infrastructure to meet different service requirements. Due to the dynamic changes in service requirements, reconfiguration of slice resources is essential to ensure the performance of slicing. However, most of the existing work focuses on reconfiguring slice through the migration of virtual network functions (VNFs, the virtual nodes that constitute the slice), which leads to large reconfiguration overhead. In view of this, we concentrate on the vertical scaling of VNF and propose a dynamic slice resource reconfiguration scheme based on multi-agent deep deterministic policy gradient (MADDPG), which can flexibly adjust multidimensional slice resources, as well as reduce or avoid costly VNF migration operations. In addition, we improve the proposed scheme using incremental learning to adapt to the different number of VNFs on each physical node and accelerate the training speed. Experimental results show that our proposed scheme has significant advantages over several benchmark schemes in VNF and slice resource satisfaction ratio, and can converge quickly through incremental learning.