Another Successful PhD Defense: Frederik Rehbach on “Enhancing Surrogate-Based-Optimization Through Parallelization“

The work introduced a rigorous benchmarking framework that allows for an in-depth comparison of parallel algorithms. It covered two distinct approaches for test function selection and generation and showed in detailed benchmark studies that existing parallel SBO methods outperform classical evolutionary algorithms and sophisticated algorithms like modern CMA-ES implementations in low budget optimization. Furthermore the work showed that the performance of SBO algorithms largely varies based on the configured kernel and infill criteria. The novel Multi-Local Expected Improvement(ML-EI) algorithm was introduced and an an automatic algorithm configuration approach based on Exploratory Landscape Analysis was presented.

The thesis was supervised by Prof. Dr. Günter Rudolph, Technische Universität Dortmund and Prof. Dr. Thomas Bartz-Beielstein, Technische Hochschule Köln.