Enhancing Simulation-Driven Optimization by Machine-Learning

Abstract

Computer simulations are being extensively used in engineering design optimization to evaluate candidate designs instead of real-world experiments. Often for some of the candidate designs the simulation will fail for an unknown reason, which leads to wasted computer resources and poor final results. To handle this challenge both effectively and efficiently this study presents an implementation in which classifiers, borrowed from the domain of machine learning, are integrated into the optimization process to predict if a candidate design is valid or not prior to evaluating it with the simulation. To further enhance the optimization effectiveness two different search methods are used. Numerical experiments show the merit of the proposed implementation. Index Terms-Expensive black-box functions, machine learning, classifiers, metamodels.Yoel Tenne received his PhD in mechanical engineering at Sydney University, Australia. Afterwards he was an australian endeavour fellow at the Korea Advanced Institute of Science and Technology (KAIST) and a Japan Society for Promotion of Science (JSPS) Fellow at Kyoto University, Japan. He currently a senior lecturer at Ariel University, Israel. His fields of research include applied optimization, computational intelligence, and systems engineering. In these respective fields he has authored over 50 publications,