Answers are grounded in this paper and cite their sources.
This paper addresses AI-based cell sleep scheduling for dense 5G/6G radio access networks, where the central challenge is reducing energy consumption without degrading delay, blocking probability, or switching stability. The literature indicates growing interest in 3GPP and O-RAN energy-saving control, but many existing methods rely on local traffic thresholds or RL agents with weak spatial awareness. To address this gap, the paper proposes Graph-Aware Multi-Objective Sleep Scheduling (GMOSS), which combines graph neural encoding of inter-cell coupling with a dueling double DQN and a safety layer that filters overload-prone actions. Unlike generic GNN-RL controllers, GMOSS formulates cell sleep scheduling as a topology-aware constrained decision problem in which graph encoding, multi-objective value learning, and safety-constrained action projection are jointly used to prevent overload-prone sleep decisions before execution. A dense heterogeneous network with 19 cells is simulated under diurnal non-stationary traffic. Compared with Always-On, threshold control, independent DQN, and multi-agent DQN baselines, the proposed method achieves the best energy–QoS trade-off. Across 30 randomized seeds, GMOSS reduces average RAN power by 31.8% versus Always-On and 18.6% versus threshold control, improves energy efficiency to 1.058 Mbit/J, lowers blocking to 1.84%, and reduces switches per hour to 4.3. The gain in energy efficiency over independent DQN is 11.2% with a 95% confidence interval of [8.7%, 13.6%] and p < 0.001. The contribution of the paper lies in explicitly coupling spatial coordination, wake-up cost awareness, and practical safety constraints in a single AI controller that is well aligned with O-RAN style deployment.
No takes yet. Share an insight, caveat, or question.
Israa J. Mohammed (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: