EURO 2025 Leeds
Abstract Submission

1513. Real-time capacity management in on-demand delivery platforms: A reinforcement-learning approach

Invited abstract in session MD-38: (Deep) Reinforcement Learning for Combinatorial Optimization, stream Data Science meets Optimization.

Monday, 14:30-16:00
Room: Michael Sadler LG19

Authors (first author is the speaker)

1. Gökhan Ceyhan
Operations Research, Just Eat Takeaway.com
2. Pol Arias
3. Malte Winckler
JustEatTakeaway.com

Abstract

Just Eat Takeaway.com (JET) manages an on-demand delivery platform characterised by uncertain demand with spatiotemporal patterns and short delivery time targets. A critical component of this system is the real-time capacity management that employs various actions to keep the courier demand and supply in balance in order to maintain a satisfactory and profitable customer service level. We focus on one of the demand-side actions, dynamic delivery areas, which curtails the demand by shrinking the originally planned delivery areas of the restaurants. We model this problem as a sequential decision problem under stochastic demand and develop a reinforcement learning approach to drive these decisions. As our action space is quite large, we do not attempt to train a model to learn those decisions. Instead, we train a model to learn delivery time constraints for different hours of the day and for different zones of the city. We then solve a bi-objective mixed-integer linear program at each decision point to determine the delivery areas of each partner under those learned delivery time constraints. We train our model in a simulation environment and test our approach under different scenarios. We compare our approach against the version where a single delivery time constraint is used throughout the day and across the region. We also compare our approach to the legacy approach where delivery areas are adjusted based on a pre-defined rule-based system.

Keywords

Status: accepted

Back to the list of papers