Supporting Data and Code for RART: A Resource-Aware Recursive Transformer for Stochastic Resource-Constrained Project Scheduling

Abstract

This record provides supporting experimental data, source code, selected trained model weights, and reproducibility documentation for the manuscript “RART: A Resource-Aware Recursive Transformer for Stochastic Resource-Constrained Project Scheduling” by Cem Savas Aydin. The study evaluates RART on stochastic resource-constrained project scheduling problems using instances from the Project Scheduling Problem Library (PSPLIB). The primary comparison involves 101 J60 projects, three training seeds per method, and 1,000 common duration scenarios per project. Seven J30 projects provide secondary comparisons. The comparator is an independently trained Wheatley-based graph neural network policy. The uploaded files comprise: ESM_1.pdf: Supplementary methodological documentation, including model features, random streams, evaluation procedures, statistical reconstruction, and provenance. ESM_2.zip: Archived source code, six selected inference checkpoints, project definitions and splits, training records, scenario-level scheduling outcomes, and scripts for reconstructing the reported statistics and figures. Recorded outcomes include mean makespan, upper-tail makespan measured by CVaR90, results before and after schedule repair, and computation times. The README.md file within ESM_2.zip provides instructions for reconstructing the statistical results from the saved simulation outputs without retraining the models.