Predict-and-Optimize approach for the scheduling of routes for the pickup of milk from farms in Belgium

Details

Supervisors

Verhaeghe, Hélène

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée
Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enIn many real-world optimization problems, key input parameters are unknown and must be predicted from data. This thesis looks at how to combine prediction and optimization in the practical setting of milk collection logistics, where the quantities of milk to be collected from each farm are uncertain and must be predicted ahead of routing decisions. Two main paradigms are explored: Predict-then-Optimize (PtO), in which predictions are generated independently and then employed in an optimization model, and Predict-and-Optimize (PaO), in which the predictive model is trained with the optimization decision in mind. First, we evaluate the performance of the traditional PtO approach by using several regression models to predict milk production for farms. These predictions are then passed to a capacitated vehicle routing problem (CVRP) solver. Next, we explore a Predict-and-Optimize approach, in which the predictive model is trained using a decision-aware loss function. This means that rather than maximizing the accuracy of each forecast, the model will focus on making predictions that lead to better overall solutions. Our goal is to compare the performance of these two approaches. After analyzing the results of our experiments, we found that although the PaO approach was less accurate than the PtO approach at predicting farm production, it produced better overall empirical results.