Simulation-Trained Explainable Policy Selection for Dynamic Production Scheduling in Manufacturing SMEs: An End-to-End Diagnostic Evaluation

Dynamic production scheduling requires decision policies that can respond to changing shop-floor conditions without imposing excessive computational demands on manufacturing SMEs. This study develops and diagnostically evaluates a simulation-trained explainable policy selector integrating discrete-event simulation (DES), supervised learning, operational reinsertion, and explainable artificial intelligence. The DES was parameterized with 1842 production orders and 9614 operation records from a high-mix, low-volume metalworking SME, and 64 scenarios were used to compare FIFO, SPT, EDD, CR, and ATC and generate state–policy labels. A classification tree and Random Forest both achieved perfect validation accuracy, macro-F1, balanced accuracy, and zero regret; Random Forest was retained by the executed pipeline tie-breaking rule. After reinsertion, however, the selector produced higher total weighted tardiness than EDD, CR, and ATC in the internal test, and its disadvantage persisted in unseen scenarios and increased under out-of-distribution conditions. Explainability analysis showed that the active-policy identifier dominated predictions, while the surrogate reproduced the selector with perfect fidelity but the same operational weakness. These findings show that predictive accuracy and explanatory fidelity do not guarantee scheduling benefit and support end-to-end validation of simulation-trained decision policies. Given mixed DES validation, the results should be interpreted as diagnostic simulation evidence rather than direct field-performance forecasts.