Adaptive metamodeling-based simulation optimisation
提出自适应元模型仿真优化框架，结合机器学习与元启发式技术，在制造数字孪生和采矿厂扩容两个实际案例中，比高效全局优化方法平均提升8.1%和9.7%的解质量，且计算时间不显著增加；与遗传算法解质量相当，但计算时间分别减少83.6%和90.6%。
Simulation Optimisation is a powerful decision-making tool, but it can be challenging for complex, time-intensive models. This paper introduces Adaptive Metamodeling-based Simulation Optimisation (AMSO), a novel framework that enhances solution quality by integrating machine learning and metaheuristic techniques. AMSO combines Bagged-gradient Boosted Trees, Genetic Algorithms, Hyperparameter Optimisation, and Design of Experiments to efficiently explore promising solution areas. The study demonstrates AMSO’s application in two real-world scenarios: a resource allocation problem in a manufacturing digital twin model and a capacity expansion project at a mining plant. AMSO outperformed the Efficient Global Optimisation method, achieving solutions 8.1% and 9.7% better on average for the first and second case studies, respectively, with no significant increase in computational time. Additionally, AMSO matched the Genetic Algorithm method’s solution quality but reduced computational time by 83.6% and 90.6% in the first and second cases, respectively. AMSO is presented as a robust alternative for solving complex simulation models, complementing existing metamodeling-based methods and opening new research avenues with other machine learning models for faster, more accurate decision-making.