Scheduling Individualized Products with Reinforcement Learning for Variable-Length Production Plans
- This paper examines a reinforcement learning (RL) approach to solve a real-world, multi-objective scheduling problem in an automotive seat production line. The environment is dynamic and human-centered, aiming to optimize both tardiness and workload balance for workers. Building on prior work with fixed-length schedules, we introduce variable-sized production plans to better reflect real-world variability. A RL agent is trained using a reward based on relative performance against a genetic algorithm (GA) benchmark, capturing trade-offs between minimizing tardiness and maximizing workload variation. While the agent performs reliably on shorter plans, closely matching the GA baseline, performance in workload balancing deteriorates for larger plans. Results indicate that the agent tends to sacrifice workload balance to improve deadline adherence.
| Author: | Leon VogelORCiD, Maylin WartenbergORCiD |
|---|---|
| URN: | urn:nbn:de:bsz:960-opus4-37888 |
| DOI: | https://doi.org/10.25968/opus-3788 |
| Parent Title (German): | KI-Forum 2025 : KI in Forschung und Lehre an Hochschulen |
| Publisher: | HsH Applied Academics |
| Place of publication: | Hannover |
| Editor: | Hanno Homann, Cedric Rohbani, Jens Christian Will |
| Document Type: | Conference Proceeding |
| Language: | English |
| Year of Completion: | 2025 |
| Publishing Institution: | Hochschule Hannover |
| Release Date: | 2025/12/10 |
| Tag: | Artificial Intelligence; Permutation Flow Shop; Reinforcement Learning |
| GND Keyword: | Künstliche IntelligenzGND; Bestärkendes Lernen <Künstliche Intelligenz>GND |
| Page Number: | 4 |
| First Page: | 97 |
| Last Page: | 100 |
| Institutes: | Fakultät IV - Wirtschaft und Informatik |
| DDC classes: | 370 Erziehung, Schul- und Bildungswesen |
| 004 Informatik | |
| 670 Industrielle Fertigung | |
| Licence (German): |  Creative Commons - CC BY - Namensnennung 4.0 International |
