MADDPG-empowered slice reconfiguration approach for 5G multi-tier system

Abstract

With the advancement of fifth-generation (5G) communication technology, plenty of vertical applications have emerged. These applications usually center on highly different quality of service (QoS) requirements, posing challenges for the supporting communication network. To address these challenges, network slicing has been proposed to partition the underlying network into multiple tailored logical networks, called as network slices. Each slice corresponds to a specific application. However, fluctuations in user activities for applications may lead to workload variation across corresponding slices, which compromising the consistent QoS provisioning. In this context, slice reconfiguration becomes imperative to ensure satisfactory user experiences. In this work, we introduce the multi-agent deep reinforcement learning method to tackle the slice reconfiguration problem and propose the multi-agent deep deterministic policy gradient-empowered slice reconfiguration (MADDPG-SR) algorithm. In MADDPG-SR, each network function is abstracted as an agent that learns to optimize its own migration in the multi-layer network to minimize the preference-weighted processing resource cost while adhering QoS and resource constraints. The proposed method is compared with four other benchmarks, and the comparison results demonstrate that MADDPG-SR can significantly reduce both the preference-weighted resource cost, and the running time for acquiring slice reconfiguration decisions while ensuring the QoS performance of slices.