New Approaches in Simulation-driven Optimization

This paper presents a work regarding the integration of discriminant functions (classifiers) with search algorithms to tackle the problem of failed simulation runs. The discriminant output is used to guide the search towards better solutions while minimizing adverse effects. The search is managed with a trust-region approach for convergence in the presence of prediction inaccuracies. Numerical evaluations based on engineering problem show that the approach yielded better final results in the mean and median statistics when compared to reference algorithms.