Toward Energy-Efficient AI: Learning Generic Assignment Skills with Unsupervised Reinforcement Learning for Network Slicing in Distributed 6G Networks

Since the 6th Generation (6G) of wireless networks is expected to deliver advanced services and address the evolving demands of future applications, making it a highly complex and dynamic system. Managing this complexity requires innovative Artificial Intelligence- based approaches that balance scalability, efficiency, and robustness. This work investigates decentralized, multi-level architectures for 6G networks and introduces an energy-efficient Reinforcement Learning (RL)-based automation framework for edge domain management and Network Slicing. The core objective is to mitigate network complexity by exploiting scalable and generalizable RL techniques. To this end, we propose a pre-training phase based on unsupervised RL to uncover effective resource allocation skills at the network edge. The proposed method is domain-agnostic, enabling applicability across diverse edge environments. By supporting various service types, it enhances scalability, decentralization, and operational efficiency. Our implementation demonstrates that the discovered skills comprehensively span the resource allocation space for each service category, thereby simplifying the NS in the network domains. Finally, the proposed approach is compared to other RL-based techniques and is shown to improve energy efficiency by using less computational resources.