Work

Multi-objective parallelization of efficient global optimization

Year: 2016
Type: dissertation
Abstract: Design optimization is a subject field where mathematical algorithms are used to improve \ndesigns. Analyses of designs using computational techniques often require significant \ncomputing resources, and for these problems, an efficient optimization method \nis needed. Efficient Global Optimization (EGO), first proposed by Jones et al. [25] is \nan optimization method which aims to use few function evaluations when optimizing a \ndesign problem. In this study, we use a multi-objective strategy to parallelize EGO. \nEGO is part of a set of algorithms called surrogate optimization methods. A set of initial \ndesigns are analyzed and then a response surface is fitted to the evaluated designs. In \neach iteration, EGO selects the set of design variables for which the next analysis will \nbe performed. It makes this decision based on two opposing criteria. EGO will either \ndecide to sample where the predicted objective function value is low, an exploitation \napproach, or where there is high uncertainty, an exploration approach. \nIn each iteration, the classical EGO only selects one design per iteration. This selected \ndesign vector is either a result of exploitation or exploration based on a measure referred \nto as maximum Expected Improvement (EI). However, the modern day computing environment \nis capable of running multiple different analyses in parallel. Thus, it would \nbe advantageous if EGO would be able to select multiple designs to evaluate in each \niteration. \nIn this research, we treat EGO?s inherent selection criteria to either exploit or explore as \na multi-objective optimization problem, since each criterion can be defined by a separate \nobjective function. In general multi-objective optimization problems don?t only have one \nsolution, but a set of solutions called a Pareto optimal set. In our proposed strategy \nmultiple designs from this Pareto optimal set are selected by EGO to be analyzed in the \nsubsequent iteration. This proposed strategy is referred to as Simple Intuitive Multiobjective \nParalLElization of Efficient Global Optimization (SIMPLE-EGO). We start our study by investigating the behaviour of classical EGO. During each iteration \nof EGO, a new design is selected to be evaluated. This is performed by finding \nthe maximum of the Expected Improvement (EI) function. Maximizing this function \ninitially proved challenging. However, by exploiting information regarding the nature \nof the EI function, the maximization problem is simplified significantly, and the robustness \nof finding the maximum is enhanced. More importantly, solving this maximization \nproblem robustly, dramatically improves the convergence behaviour once a local basin \nhas been found. \nWe compare our SIMPLE-EGO method to a multi-objective optimization algorithm \n(EGO-MO) published by Feng et al. [16]. We first investigate the behaviour of EGO, \nEGO-MO, and SIMPLE-EGO. Thereafter the convergence performance of these methods \nis quantified. \nAs expected the parallelization of both SIMPLE-EGO and EGO-MO lead to faster \nconvergence on a range of test functions compared to classical EGO, which only sampled \none point per iteration. The convergence characteristics of SIMPLE-EGO and EGOMO \nare also markedly different. We conclude with a discussion on the advantages and \ndisadvantages of the investigated methods.
Language: English
Open Access status: green
OpenAlex ID: W2559940143
Cites: 42
Cited by: 1
Related to: 10