Self-Orchestrated AP Switching Optimization via Network Uncertainty Awareness for 6G AI-RAN

Abstract

The sixth-generation (6G) mobile communication system is anticipated to provide unprecedented performance, but faces a sustainability barrier from overactivated access points (APs). This study addresses energy-efficient operation in the 6G artificial intelligence–radio access network (AI-RAN) by jointly optimizing AP switching and power allocation. The primary challenge involves making decisions at each point to achieve the depth- $\boldsymbol {N}$ optimality with respect to the wake-up time that occurs during AP switching. We formulate a long-term energy minimization problem with quality-of-service (QoS) guarantees, and recast the problem into a constrained Markov decision process (CMDP). Our solution employs a world model-based deep reinforcement learning (DRL) framework designed to enable the RAN to achieve situational awareness under network uncertainties by avoiding myopic decisions that could negatively impact future QoS. Experimental results demonstrate that the proposed framework outperforms DRL baselines in convergence rate and stability by leveraging improved situational awareness, and also achieves more than 30% power saving compared to feasible benchmarks, satisfying QoS requirements with a negligible margin in more than 90% of cases.