Reinforcement-Learning-Based Network Slicing and Resource Allocation for Multi-Access Edge Computing Networks

This paper considers the network slicing and resource allocation problem in a multi-access edge computing (MEC) network. A two-level slicing model is introduced for network slicing and the considered problem is formulated as a long-term optimization problem with an objective to maximize a weighted average of the successful processing ratio of users’ computing tasks in each network slice instance by performing network slicing and resource allocation, while satisfying the service requirements of users’ computing tasks. To solve the formulated problem, we decompose the problem into two sub-problems: periodic service-level slicing (SLS) and dynamic resource allocation (DRA), and propose a reinforcement-learning based network slicing and resource allocation (RL-S&A) algorithm, which consists of a QMIX-based SLS algorithm and a PPO-based DRA algorithm, to solve the two sub-problems, respectively. To reduce the training time of an optimal policy to perform periodic SLS, an MAML-based initialization algorithm is introduced to obtain a meta-policy to initialize the neural networks used in the QMIX-based SLS algorithm. Simulation results show that the proposed RL-S&A algorithm can achieve better performance than several benchmark algorithms in terms of the successful processing ratio of users’ computing tasks in the network.