Self-Improving Interference Management Based on Deep Learning With Uncertainty Quantification

TL;DR

This work targets downlink interference management for a network of $N$ single-antenna BS–user pairs with channel matrix $\boldsymbol{H}$, aiming to maximize $R(\boldsymbol{H},\boldsymbol{p})=\sum_n \log(1+\text{SINR}_n)$ under $\boldsymbol{p}\in[0,P^{\max}]^N$, where $\text{SINR}_n=\frac{|h_{nn}|^2 p_n}{\sigma_n^2+\sum_{m\neq n}|h_{nm}|^2 p_m}$. The paper proposes a self-improving interference management framework that combines deep ensembles with uncertainty quantification to predict the optimal power allocations $\boldsymbol{p}^*(\boldsymbol{H})$ and quantify predictive uncertainty via $\hat{\mu}_n(\boldsymbol{H})$ and $\hat{\sigma}^2_n(\boldsymbol{H})$. A qualifying criterion based on the uncertainty-aware performance, involving $\hat{\mathbf{p}}^*(\boldsymbol{H})$, the interval around it, and the sum-rate estimates $\hat{R},\hat{R}^U,\hat{R}^L$, decides when to trust the DL prediction or fall back to a target optimization algorithm to generate enhanced data for continual model improvement. Experimental results demonstrate that SI-DNN achieves sum-rate performance close to the WMMSE upper bound across varying topologies, with robust self-improvement rounds converging after about 10 iterations and improved generalization to unseen scenarios.

Abstract

This paper presents a groundbreaking self-improving interference management framework tailored for wireless communications, integrating deep learning with uncertainty quantification to enhance overall system performance. Our approach addresses the computational challenges inherent in traditional optimization-based algorithms by harnessing deep learning models to predict optimal interference management solutions. A significant breakthrough of our framework is its acknowledgment of the limitations inherent in data-driven models, particularly in scenarios not adequately represented by the training dataset. To overcome these challenges, we propose a method for uncertainty quantification, accompanied by a qualifying criterion, to assess the trustworthiness of model predictions. This framework strategically alternates between model-generated solutions and traditional algorithms, guided by a criterion that assesses the prediction credibility based on quantified uncertainties. Experimental results validate the framework's efficacy, demonstrating its superiority over traditional deep learning models, notably in scenarios underrepresented in the training dataset. This work marks a pioneering endeavor in harnessing self-improving deep learning for interference management, through the lens of uncertainty quantification.

Figures (6)

• Figure 1: A typical deep learning-based interference management approach.
• Figure 2: The proposed self-improving interference management framework using deep learning with uncertainty quantification.
• Figure 3: The CDFs of the sum-rates achieved by different algorithms.
• Figure 4: The predicted solutions and confidence intervals computed by SI-DNN for instances from different topologies.
• Figure 5: The performance of the algorithms varying the criterion value $\epsilon$.