A Reinforcement Learning Approach to Virtual Network Embedding Problems in 5G Networks

5G network slicing is the problem of mapping requested virtual networks on the substrate network resources. Due to resource capacity constraints, the performance of network slicing depends on the number of supported requests. This challenge is a type of Virtual Network Embedding (VNE) problem in which a weighted graph is divided into multiple smaller weighted graphs according to the user's custom requirements. These problems are NP-hard, and most existing solutions have suggested using Reinforcement Learning (RL) models to solve them. However, they do not adequately represent the weighted graph to the learning model. Therefore, their learning rate is limited. This paper proposes TRL-VNE, a Two-stage RL-based VNE solution to overcome these challenges. In the first stage of this solution, an RL model is utilized for mapping the central node of each request. Novel graph-based features (G-features) are used in this model to improve its learning rate. The second stage uses a greedy algorithm to map the other components. The simulation results show that TRL-VNE improves the requests acceptance ratio and maximum supported requests by 21% and 36%, respectively, compared to existing solutions. Moreover, we have proposed a network architecture based on TRL-VNE, and emulated it in Mininet to investigate the feasibility of the proposed solution.