SURROGATE-ASSISTED ALGORITHMS FOR COMPUTATIONALLY EXPENSIVE MULTI- AND MANY-OBJECTIVE GLOBAL OPTIMIZATION PROBLEMS

Abstract

Multi-objective optimization problems often arise in real-world engineering applications, and they become more challenging when the evaluation of objective functions consumes a lot of resources. Assuming that each objective function is computationally expensive and black-box, this thesis contributes to the development of surrogate-assisted algorithms for two important variants of multi-objective optimization problems, namely 1) many-objective optimization problems where more than three objectives are considered and 2) multi-fidelity optimization problems where each objective has a computationally cheaper but less accurate representation. Different ways of modelling radial basis functions as surrogate models and using them to determine new sample points for expensive evaluation have been studied in this thesis. The newly proposed algorithms are comprehensively compared with recent state-of-the-art methods on a suite of multi-objective test problems. This thesis also investigates the performance of each proposed algorithm in calibrating hydrological simulation models built for real-world watersheds.