A machine learning framework for surrogate modeling and benchmarking in dynamic job shop scheduling

In dynamic job shop scheduling, simulation has long been a crucial tool for analyzing complex, time-dependent production systems under uncertainty and variability, both as a standalone decision-making tool and as a component of more complex simulation–optimization methods. However, the computational demands of simulation models of complex systems may limit their practical use, especially when rapid or large-scale experimentation is required. With this in mind, the objective of this paper is twofold. First, it proposes a machine learning-based surrogate model framework that replicates scheduling dynamics and generates synthetic data, reflecting decision-making events under various dispatching rules. This surrogate model is designed to approximate the outputs of simulation models in dynamic job shop scheduling scenarios, hence significantly reducing computational effort while maintaining accuracy. Second, it benchmarks multiple supervised learning algorithms to evaluate their capability to surrogate simulation outputs effectively, considering both predictive performance and computational efficiency (in terms of both training time and time to predict). By enabling faster performance evaluation of scheduling strategies, this approach enhances simulation-driven scheduling analysis and optimization, particularly benefiting traditional heuristic and metaheuristic methods that rely heavily on extensive simulation runs. A key contribution of this work is the comprehensive benchmarking and fine-tuning of ten supervised learning algorithms. We evaluated their predictive accuracy in approximating simulation results and their computational efficiency during model training and during prediction. Our experimental analysis identifies random forests as the most effective surrogate model with an R 2 of 0.91.