DT-RL: a digital twin–driven reinforcement learning framework for proactive slice admission and predictive QoS assurance in 6G networks

In Cloud-RAN (C-RAN) infrastructures, with the dynamic growth of 6G network slicing and the volatility of multi-service traffic patterns, proactive and QoS-aware (Quality of Service) Slice Admission Control (SAC) is required. Reinforcement learning approaches are generally still reactive with conventional SAC designs, as they rely on observations and lack insight into the upcoming congestion state. In bursty or non-stationary conditions, this limitation can lead to suboptimal admission decisions. To tackle this issue, we propose DT-RL (Digital-Twin-Driven Reinforcement Learning), which is a framework that enriches the multi-horizon predictive awareness of SAC. The digital twin aims to model future resource availability, as well as produce a chain of latency evolution and the saturation risk of one layer crossing layers in the CO–LDC–RDC hierarchy when acting, by generating counterfactual rollouts (all of the possible opposite actions) to expand the experience space of each layer for the agents. RL incorporates these predictive signals into the RL state, enabling the RL agent to learn to avoid negative admission impacts and to overcome delayed-reward problems. DT-RL is on top of a high-fidelity C-RAN environment with 3-layer URLLC microbursts, log-normal eMBB bursts and mMTC spikes. Experiments reveal that sDT-RL can achieve 12–18% higher slice-acceptance rates, a 30–45% decrease in the QoS violation ratio, 46% lower URLLC P95 latency, and 55% lower eMBB P99 tail latency than the relevant state-of-the-art RL baselines. Furthermore, distributional latency analysis reveals that DT-RL reduces bursty heavy-tailed delay spikes and maintains steady performance under bursty non-stationary traffic.