SafeSCHEMA: Multi-domain Orchestration of Slices based on SafeRL for B5G Networks

The success of Software-Defined Networking and Network Function Virtualization enabled providers to partially automate the decision-making of various network operating workflows. Network slicing is an innate feature of the Fifth Generation (5G) and Beyond-5G (B5G) networks, bringing many advantages but also increasing the complexity of managing such networks. In this context, automation tools based on Artificial Intelligence (AI) have become a preeminent instrument of automation used by the industry, as it would be impossible to manually provision and manage such a vast and dynamic infrastructure. Safe interaction of the AI agents with the network is one of the predominant challenges, especially when Reinforcement Learning (RL) is used in critical environments. This is particularly important when the RL agent actions have a high or irreversible impact on the network or service. Slice management is one of the major features that operators want to automate, to offer future services at large scales with manageable complexity to the operator. However, during the exploration phase, RL agents can cause significant performance degradation during operation and possibly introduce irreversible damage to the service being offered. To address this major challenge, we propose a multi-agent, modular, SafeRL architecture for distributed slice orchestration. We study the problem of zero-touch slice management and orchestration, in the context of Ultra-Reliable Low Latency Communication services for B5G networks. Our results demonstrate improved performance over competing solutions, while ensuring the safety of the performed actions during real-time slice orchestration.