Deep Reinforcement Learning for End-to-End Network Slicing: Challenges and Solutions

5G and beyond is expected to enable various emerging use cases with diverse performance requirements from vertical industries. To serve these use cases cost-effectively, network slicing plays a key role in dynamically creating virtual end-to-end networks according to specific resource demands. A network slice may have hundreds of configurable parameters over multiple technical domains that define the performance of the network slice, which makes it impossible to use traditional model-based solutions to orchestrate resources for network slices. In this article, we discuss how to design and deploy deep reinforcement learning (DRL), a model-free approach, to address the network slicing problem. First, we analyze the network slicing problem and present a standard-compliant system architecture that enables DRL-based solutions in 5G and beyond networks. Second, we provide an in-depth analysis of the challenges in designing and deploying DRL in network slicing systems. Third, we explore multiple promising techniques, i.e., safety and distributed DRL, and imitation learning, for automating end-to-end network slicing.