EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

3763. A Supervised Machine Learning Framework to Predict the Request Fit for Dynamic Dial-a-Ride Problems

Invited abstract in session TD-64: Dynamic Vehicle Routing 2, stream VeRoLog - Vehicle Routing and Logistics.

Tuesday, 14:30-16:00
Room: S16 (building: 101)

Authors (first author is the speaker)

1. Pirmin Fontaine
Ingolstadt School of Management, Catholic University of Eichstätt-Ingolstadt
2. Simon Mader
Ingolstadt School of Management, Catholic University of Eichstätt-Ingolstadt
3. Stefan Voigt
Department of Supply Chain Management & Operations, Catholic University of Eichstätt-Ingolstadt

Abstract

In the transformation towards sustainable transportation, a good public transport offer is a crucial element. However, in rural areas, public transport providers (PTPs) face challenges in introducing a good offer using classical line-based services within a restricted budget due to the low demand. One potential solution is on-demand buses, which operate on flexible schedules and routes that depend on the actual demand. Still, the low demand in these regions remains a challenge to PTPs. They want to maximize the served passengers but are forced to accept or reject customer requests immediately within the dynamic booking process and without knowing the true impact of the total number of server passengers due to potential future requests. This raises the question of predicting the customer's fit within a passenger maximizing dial-a-ride problem (PM-DARP).

To address this problem, we introduce the request fit predictor (RFP) framework that allows for fast decisions about the acceptance or rejection of customer requests. Within this framework, we model the request fit as a binary classification task and learn the fit with supervised machine-learning models from a static PM-DARP using historical data.

The framework is tested on real-life data, and the results show that the PTP can serve 7.48% more passengers using the RFP compared to a cheapest insertion benchmark and more than 25% more than the status quo. As a side effect, we also reduce the driven distance by 6.68%.

Keywords

Status: accepted

Back to the list of papers