Inference- and Optimization-based Approximated Solver for Dynamic Job-shop Scheduling Problem
Abstract: The Job-shop Scheduling Problem (JSP) is a well-known combinatorial optimization problem that arranges tasks for efficient processing. It is used in a broad range of industrial applications, such as smart manufacturing and transportation. We focus on repeatedly solving scheduling instances with variable sets of jobs for real-world applications, and we propose an inference-based model called JSPformer based on a data-driven scheme. Our main contribution lies in enabling the training and inference of schedules for variable sets of jobs by encoding input data into job-wise feature vectors and utilizing a neural network for set-structured data. Furthermore, for cases where a few minutes of additional computation is available, we propose JSPformer+Opt, a hybrid model of JSPformer and a local optimization. The local optimization is intended to make a more efficient schedule quickly from an inference solution. It uses part of the inference and optimizes the rest to improve the solution quality while reducing the problem size for fast computation. In numerical experiments, we validated that JSPformer outperformed existing inference-based models and demonstrated its ability to handle instances with variable sets of jobs.
Submission Length: Long submission (more than 12 pages of main content)
Changes Since Last Submission: We have revised our paper to address the reviewers' comments.
Assigned Action Editor: ~Stefano_Teso1
Submission Number: 2387
Loading