EURO 2025 Leeds
Abstract Submission

3017. Learning Optimization: A Machine Learning framework for predicting Vehicle Routing outcomes under demand uncertainty

Invited abstract in session TA-27: Applications of Optimization under Uncertainty, stream Stochastic and Robust optimization.

Tuesday, 8:30-10:00
Room: Maurice Keyworth G.02

Authors (first author is the speaker)

1. Lorenzo Saccucci
Department of Computer, Control and Management Engineering (DIAG), Sapienza University of Rome
2. Giuseppe Stecca
Istituto di Analisi dei Sistemi ed Informatica "Antonio Ruberti", Consiglio Nazionale delle Ricerche
3. Marco Boresta
Istituto di analisi dei sistemi ed informatica "Antonio Ruberti" (IASI), Consiglio Nazionale delle Ricerche (CNR)

Abstract

In industrial settings, optimization problems like the Vehicle Routing Problem (VRP) frequently recur with similar characteristics, where only a few parameters of previously solved instances vary while most features remain constant. This research proposes a data-driven approach using Machine Learning (ML) to predict optimal solution characteristics for recurring instances by leveraging historical data patterns, particularly when facing demand uncertainty, without requiring complete re-optimization.
A comprehensive ML pipeline has been developed to capture patterns in recurring VRP instances with time windows, encompassing the generation of artificial instances, feature extraction, and training of predictive models to forecast key aspects of optimal solutions. The methodology bridges traditional optimization methods with predictive analytics, creating a framework that predicts solution values efficiently when demand variations occur. Experimental results demonstrate high accuracy in predicting critical solution features, such as the required number of vehicles and optimal route structures.
Moreover, initial explorations indicate that these predictions can be embedded within heuristic methods, showing promise in accelerating the resolution of complex instances. This approach enables the avoidance of costly re-optimization for daily demand fluctuations while providing theoretical insights and practical benefits for industries tackling VRP challenges with variable demand.

Keywords

Status: accepted

Back to the list of papers