Many real-world application problems encountered in industry have no analytical formulation, that is they are blackbox optimization problems, and often make use of expensive numerical simulations. We propose a new blackbox optimization algorithm named BOA to solve mixed-variable constrained blackbox optimization problems where the evaluations of the blackbox functions are computationally expensive. The algorithm is two-phased: in the first phase it looks for a feasible solution and in the second phase it tries to find other feasible solutions with better objective values. Our implementation of the algorithm constructs surrogates approximating the blackbox functions and defines subproblems based on these models. The open-source blackbox optimization solver NOMAD is used for the resolution of the subproblems. Experiments performed on instances stemming from the literature and two automotive applications encountered at Stellantis show promising results of BOA in particular with cubic RBF models. The latter generally outperforms two surrogate-assisted NOMAD variants on the considered problems.

工业中遇到的许多实际应用问题没有解析公式，即它们是黑盒优化问题，并且经常使用昂贵的数值模拟。我们提出了一种名为 BOA 的新黑盒优化算法来解决混合变量约束黑盒优化问题，其中黑盒函数的评估计算量很大。该算法分为两个阶段：在第一阶段，它寻找可行的解决方案，在第二阶段，它尝试找到具有更好目标值的其他可行解决方案。我们的算法实现构造了近似黑盒函数的代理，并基于这些模型定义了子问题。开源黑盒优化求解器 NOMAD 用于解决子问题。对源自文献的实例和 Stellantis 遇到的两个汽车应用进行的实验显示了 BOA 的有希望的结果，特别是对于三次 RBF 模型。在所考虑的问题上，后者通常优于两种替代辅助的 NOMAD 变体。