Interpretable and Uncertainty-Aware by Construction Decisions with Fuzzy Theory

Autonomous systems increasingly use Machine-Learning (ML) components to perceive their environment and guide the decision making process. However, the confidence scores of these components are often miscalibrated and fail to represent true predictive reliability. While existing calibration methods reduce this issue, they still compress predictions and uncertainty into a single value, losing significant information. In this work, we propose a method to have decision-making “uncertainty aware by construction” by applying fuzzy theory to represent and manipulate uncertainty in ML confidence scores. Our approach extends a standard perception-decision pipeline with two new modules: an Estimator Performance component that collects diverse calibration and error indicators, and a Fuzzifier that transforms these indicators, together with the raw confidence score, into a fuzzy number encoding second-order uncertainty. This fuzzy confidence is then consumed by a fuzzy decisionmaker capable of adapting its interpretation of uncertainty to the system’s current goals. Experiments on a classification dataset show that fuzzy confidence retains uncertainty information discarded by traditional calibration, enabling better-aligned decisions under changing goals.