Self-attention Convolutional Neural Network-based SLA Decomposition for Network Slicing

Abstract

In 5G networks, a network slice is an independent end-to-end logical network running on a shared physical infrastructure, capable of guaranteeing a negotiated service level agreement (SLA) to deliver customized service. An end-to-end network slice includes three logical networks in the subnets: an access network (AN) sub-slice, a transport network (TN) sub-slice and a core network (CN) sub-slice. The network slice SLA needs to be decomposed into the corresponding sub-slice SLAs. In this paper, the SLA decomposition for network slicing is modelled as a constrained optimization problem and solved in two steps: building the sub-slice SLA acceptance probability models and solving the constrained optimization problem. The sub-slice SLA acceptance probability models are built by a self-attention convolutional neural network, and the constrained optimization problem is solved by a differential evolution algorithm based on the optimal acceptance probability of the SLA decomposition for network slicing. The simulation results show that the proposed self-attention convolutional neural network-based sub-slice SLA acceptance probability models (SACNN-SSAPMs) fit better with the ground truths, and the errors between the optimal acceptance probabilities of the SLA decomposition for network slicing based on the SACNN-SSAPMs and those based on the ground truths are smaller.