Utilizing Confidence in Localization Predictions for Improved Spectrum Management

Transmitter localization is an important component of next-gen spectrum sharing and management systems. Recently, machine learning (ML) methods have shown promising results for localization in complex environments. However, existing ML models have significant limitations, such as the need for careful parameter tuning and the lack of accuracy on out-of-distribution (OOD) examples. Moreover, current ML models do not typically provide a “confidence” in their prediction. In this work, we propose a new training procedure based on the Earth Mover’s Distance (EMD) that improves on these limitations. The method improves OOD accuracy by up to $22 \%$ while providing more interpretable results. The EMD-trained model also produces a confidence score, which can be used to identify high-error and OOD examples. We demonstrate how model confidence serves as a guide for hybrid localization models. This includes selecting the most reliable prediction from multiple models based on confidence values or resorting to a fallback path loss technique in cases of low confidence. Our work establishes the importance of model confidence in improving the accuracy of localization and as a mechanism for effective decision making in localization applications.