Safe RAN Slicing in O-RAN: Minimizing SLA Violations via Model-Based Reinforcement Learning

RAN slicing in Open RAN (O-RAN) architectures requires dynamic and intelligent resource allocation to ensure both high resource efficiency and strict compliance with Service Level Agreements (SLAs). However, conventional reinforcement learning (RL) methods struggle in this setting due to their unsafe exploration behavior, often resulting in frequent SLA violations. In this paper, we propose a safe and sample-efficient model-based RL (MBRL) framework tailored for SLA-aware RAN slicing. Our agent leverages an online kernelized quantile regressor to estimate the lower quantile of key performance indicators (KPIs), enabling direct modeling of SLA violation risk. To ensure safety during online adaptation, the policy incorporates an uncertainty-aware safety margin, promoting conservative decisions under high model uncertainty. Extensive simulations demonstrate that our method reduces SLA violations to under 1%, while maintaining high resource efficiency and achieving fast, stable convergence.