Project R-10809


Research on theme "multi-objective optimization"


Multiobjective optimization (MO) problems are common in reality; they do not only occur frequently in engineering, but also in business settings. In many real life situations, the multiple objectives cannot be determined analytically; they need to be observed through (time consuming) computer experiments, and the observed outcomes may contain noise. This project application focuses on MO stochastic simulation optimization problems. Efficient and effective methods to solve such problems are currently lacking, as most existing algorithms require many evaluations, and ignore the noise in the optimization process. This does not only mislead the search for optimal solutions, but also lead to errors in the identification of the "best" points sampled. Previous research has shown that the use of Gaussian Process Regression (a machine learning technique, also referred to as kriging in the Operations Research literature) provides a very powerful tool to analyze problems with noise. The current project aims to include user preferences in the optimization, and to provide an indifference zone approach for the identification of the Pareto optimal solutions. Novel quality indicators need to be developed to assess the quality of the resulting stochastic Pareto fronts, and a publicly available test suite will be built to facilitate the performance comparison of different stochastic MO algorithms put forward by the research community.

Period of project

01 September 2020 - 31 August 2021