On challenges of simulation-based global and multiobjective optimization

Abstract

In this thesis, we address some challenges arising when solving real life simulation based optimization problems, for example, in system, device or process design. Often problems of this type are highly nonlinear, nonconvex, computationally expensive, gradient information is not available at a reasonable cost, and often there are several conflicting objectives to be considered simultaneously. These facts suggest that we should use carefully constructed optimization systems utilizing global, efficient and multiobjective approaches to solve these problems comfortably. Multiobjective optimization problems can be solved in several different ways. In this thesis, we concentrate on interactive scalarization based approaches and on evolutionary multiobjective optimization (EMO) approaches. Our main emphasis in this thesis lays on dealing with two issues, problems due to the computational complexity of objective functions (running the simulator may be very time consuming), and difficulties of choosing the final solution among a possibly large set of Pareto optimal solutions. In response to the above mentioned challenges, in this thesis we first construct a heterogenous optimization system capable of accommodating virtually any optimization algorithm and simulator combination. Then, to save in objective function evaluations, we propose an interactive approach with an adjustable solution accuracy during the process. On the algorithmic level, we propose a new and efficient single objective global optimization algorithm, which may be, for example, used in conjunction with the interactive approach to solve scalarized problems. On the other hand, to create an approximation of the Pareto optimal set in a single run, we propose a new and efficient EMO algorithm, that overcomes some drawbacks (e.g., lack of convergence) of the current approaches. In order to select the most preferred Pareto optimal solution, we propose an approach where the characteristics of the Pareto optimal set are condensed by using advanced clustering algorithms, thus reducing the cognitive burden of the decision maker.