2426. Solving the Stochastic Dynamic k-Travelling Repairperson Problem with Deep Reinforcement Learning
Invited abstract in session WE-4: GOR Master Thesis Awards, stream PC Stream.
Wednesday, 16:30-18:00Room: H6
Authors (first author is the speaker)
| 1. | Judith Schulze
|
| Institute of Automotive Management and Industrial Production, Technische Universität Braunschweig |
Abstract
Efficient real-time dispatch of repair personnel under stochastic, dynamic demand remains a major challenge for service providers. This study addresses the stochastic dynamic k-travelling repairperson problem (SD-kTRP) in a multi-period setting, where service requests arrive over time and must be accepted for same-day service or postponed to the next day. Accepted requests must be inserted into an existing route. The objective is to minimize customer waiting times, which accumulate with each additional stop. Accepting distant or time-consuming requests may delay subsequent services, while postponing them reduces future flexibility. To support anticipative decision-making under uncertain demand, the SD-kTRP is, for the first time, modeled as a sequential decision process. Building on this formulation, a deep reinforcement learning approach is developed that combines deep Q-learning with routing heuristics and is evaluated against two rule-based policies through a numerical study. The proposed approach attains results matching those of the best-performing rule-based policy. An analysis of individual customer waiting times reveals that long service times and peripheral customer locations are primary drivers of delay. These findings underscore the importance of anticipative routing for minimizing waiting times and demonstrate the potential of deep reinforcement learning for real-time decision-making in dynamic environments.
Keywords
- Machine Learning
- Prescriptive Analytics
- Routing
Status: accepted
Back to the list of papers