DL8.5 decision tree used in a Predict-then-Optimize framework : decision tree learning using SPO loss error function

Details

Supervisors

Nijssen, Siegfried

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée

Abstract

This thesis explores previous research done in the Predict-then-Optimize framework. More specifically, it looks at decision trees who's output is used as input in an optimization problem. In this case, the optimization problem will be a shortest path optimization. Four models are tested: the CART classifier, the CART regression tree, the DL8.5 decision tree with MSE loss function and finally, the DL8.5 decision tree with the Smart-Predict-then-Optimize loss (SPO). They are tested on situations of varying complexity. In these situations, the behavior of each method was quite similar no matter the complexity with the exception of the CART classifier. This leads to the conclusion that the CART classifier is quite accurate simpler situations, but is not capable of handling more complexity. The DL8.5 regression tree used with MSE is more accurate than the CART regression tree when observing the MSE. On the other hand, these two methods have comparable results when it comes to the extra travel time. Finally, the DL8.5 decision tree with the SPO loss function is pretty systematically the better performing method when observing the decision error. We conclude that DL8.5 and SPO loss perform well together and give consistently better results than other combinations.