Probabilistic Calibration in a Few-Shot Domain Adaptation Setting

Abstract

This thesis investigates the application of probabilistic machine learning models within a Few-Shot Domain Adaptation (FSDA) setting to address covariate shifts induced by new operational conditions for electric trucks. By leveraging datadriven methods instead of the truck’s physical properties, the thesis assesses the characteristics and robustness of different machine learning models in predicting energy consumption under various new conditions. The study focuses on scenarios involving uni-, bi-, and multivariate covariate shifts posed by colder temperatures, a new route type, and a new vehicle manufacturer. Utilizing real-world data from electric truck drives, the models are trained on source data, adapted using limited target domain data, and assessed for their probabilistic calibration in the target domain. The findings indicate that out of the two baseline models, Ridge regression models, the source-trained baseline model performs well under simpler shifts but struggles with multivariate shifts where the target-only baseline model excels given sufficient target domain data. Hierarchical Bayesian linear regression shows high adaptability when covariate shift affects hierarchical levels of the model. Gaussian process regression improves comparatively well with adaptation. However, the results indicate a possible sensitivity to kernel selection. Bayesian neural networks face challenges with prediction mean accuracy and high sensitivity to individual samples, further research is needed to determine the model’s feasibility in a FSDA setting. These insights provide valuable guidance for fleet management companies in improving decision-making and operational efficiency under new driving conditions through accurate probabilistic energy consumption modeling.