EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
1889. Solving scheduling problems in real-time through deep learning methods
Invited abstract in session WA-28: Advancements of OR-analytics in statistics, machine learning and data science 8, stream Advancements of OR-analytics in statistics, machine learning and data science.
Wednesday, 8:30-10:00Room: 065 (building: 208)
Authors (first author is the speaker)
| 1. | Imanol Echeverria
|
| Smart Industry, Tecnalia | |
| 2. | Roberto Santana
|
| Department of Computer Science and Artificial Intelligence, University of the Basque Country (UPV/EHU) | |
| 3. | Maialen Murua
|
| TECNALIA |
Abstract
Recently, deep reinforcement learning (DRL) has been applied to solve scheduling problems in real-time. These methods involve generating a policy through interaction with thousands of random instances, leading to enhanced solutions by maximizing a reward. However, real-world instances often exhibit repetitive patterns and represent only a small subset of all possible scenarios. Additionally, obtaining a large number of examples is uncommon in real-world applications. In this paper, we propose a methodology to learn an optimized dispatching rule tailored to a specific set of instances. To do this, we model the scheduling problem as a Markov process, employ graph neural networks to represent instances, and utilize behavioral cloning alongside optimal solutions to simpler instances to determine the best policy. Given the challenge of training with limited instances, we suggest initially training a general scheduler using diverse instances. Subsequently, this model undergoes retraining to adapt to specific distributions, involving fine-tuning of its last layers for efficient adaptation. To validate our approach, we conduct experiments on both the job-shop scheduling problem (JSSP) and the flexible JSSP in several public scheduling benchmarks. Results demonstrate that our method, trained using behavioral cloning, outperforms DRL-trained models. Additionally, we show that the retraining strategy enables superior performance compared to a general model trained with random instances.
Keywords
- Combinatorial Optimization
- Artificial Intelligence
- Scheduling
Status: accepted
Back to the list of papers