Training, Validation and Test Datasets for End-to-End Q-Learning for Operation-Fixture-Machine Scheduling in Pallet Automation Systems

This dataset accompanies the manuscript "End-to-end Q-learning for operation-fixture-machine scheduling in pallet automation systems" by Yulu Zhou, Shichang Du, and Andrea Matta. It provides synthetically generated benchmark instances for joint operation sequencing, fixture-pallet selection, and machine assignment in a pallet automation system, with makespan minimization as the scheduling objective. The release includes four separate partitions: 150 training instances and 75 validation instances spanning case-size groups 1-15, 100 original test instances spanning groups 1-20, and 100 independently generated test instances spanning groups 1-20. The training set is used for Q-table learning; the validation set is used exclusively for parameter and checkpoint selection. The original test set supports the performance evaluation and subsequent policy analysis and rule extraction reported in the manuscript. The additional independent test set is used specifically in Section 5.6.4, "Rule extraction and independent validation", and Table 16. It evaluates candidate-selection agreement and schedule-level performance of the extracted rules relative to the proposed agent after the rules have been frozen. These instances are not used for training, parameter or checkpoint selection, original testing, or rule extraction. Each test set contains 75 same-scale cases and 25 unseen larger-scale cases. Each TXT file specifies job routes, eligible fixture-pallets and machines, loading/unloading times, and machine-dependent processing times. Files have been renamed to match the manuscript's case-size groups without changing their contents. README.md explains the file format and experimental roles; file_manifest.csv provides the full original-to-release filename mapping and checksums; case_groups.csv summarizes group dimensions. This release contains 425 instance files and documentation. It does not include source code, trained policies, numerical experiment results, or the modified scenario datasets used for scenario generalization.