Efficient uncertainty quantification for power allocation in wireless networks

This paper investigates the problem of power allocation in wireless communication systems from the perspective of uncertainty quantification, aiming to enhance the reliability of deep learning-based approaches under practical constraints. The primary objective is to address the limitations of traditional deep models in scenarios with limited data, noise interference, and model capacity bottlenecks. To address this challenge, model uncertainty and aleatoric uncertainty are formally defined within the wireless communication context. A robust optimization framework based on Bayesian learning is introduced. Furthermore, by integrating an uncertainty quantification module with a power adaptation module, the Variational Autoencoder Power Allocation Algorithm (VAEPA) is designed, significantly mitigating the adverse effects of uncertainty on the optimisation process. Extensive experiments on communication datasets demonstrate the strong robustness of the proposed method against model misspecification and data outliers. The algorithm significantly reduces average power consumption while maintaining high communication rates, particularly in power-limited and large-scale deployment scenarios. These results suggest that the proposed approach provides a practical and reliable solution for deep learning-based wireless resource optimization.