Title: Conformal prediction for safe decision-making Authors:  Valentina Zangirolami - University of Milano-Bicocca (Italy) [presenting]
Matteo Borrotti - University of Milan-Bicocca (Italy)
Antonio Candelieri - University of Milano-Bicocca (Italy)
Abstract: Safety is the core feature to avoid system disruptions while aiming to develop more efficient policies that improve existing human-expert control strategies. Safe conformal constraints for decision-making in dynamic systems, leveraging split conformal prediction to provide robust uncertainty quantification for next-state predictions. By making no distributional assumptions, Conformal Prediction ensures coverage guarantees for safe constraints, formulated as no next-state violations. We employ normalized nonconformity measures to integrate conformal prediction intervals in the decision rules to obtain point-wise prediction intervals. We extend the framework with Mondrian Conformal Prediction techniques to fully adapt conformal intervals in dynamic systems while maintaining validity. The efficacy of this approach is then empirically demonstrated on an optimal control task for a water tank system, comparing the proposed safe (conformal) policy against both a baseline safe policy and a state-of-the-art alternative lacking conformal predictions. Results highlight the effectiveness of our method in balancing safety and performance.