Dynamic Resampling for Preference-based Evolutionary Multi-objective Optimization of Stochastic Systems : Improving the efficiency of time-constrained optimization

Abstract: In preference-based Evolutionary Multi-objective Optimization (EMO), the decision maker is looking for a diverse, but locally focused non-dominated front in a preferred area of the objective space, as close as possible to the true Pareto-front. Since solutions found outside the area of interest are considered less important or even irrelevant, the optimization can focus its efforts on the preferred area and find the solutions that the decision maker is looking for more quickly, i.e., with fewer simulation runs. This is particularly important if the available time for optimization is limited, as is the case in many real-world applications. Although previous studies in using this kind of guided-search with preference information, for example, withthe R-NSGA-II algorithm, have shown positive results, only very few of them considered the stochastic outputs of simulated systems.In the literature, this phenomenon of stochastic evaluation functions is sometimes called noisy optimization. If an EMO algorithm is run without any countermeasure to noisy evaluation functions, the performance will deteriorate, compared to the case if the true mean objective values are known. While, in general, static resampling of solutions to reduce the uncertainty of all evaluated design solutions can allow EMO algorithms to avoid this problem, it will significantly increase the required simulation time/budget, as many samples will be wasted on candidate solutions which are inferior. In comparison, a Dynamic Resampling (DR) strategy can allow the exploration and exploitation trade-off to be optimized, since the required accuracy about objective values varies between solutions. In a dense, converged population, itis important to know the accurate objective values, whereas noisy objective values are less harmful when an algorithm is exploring the objective space, especially early in the optimization process. Therefore, a well-designed Dynamic Resampling strategy which resamples the solution carefully, according to the resampling need, can help an EMO algorithm achieve better results than a static resampling allocation.While there are abundant studies in Simulation-based Optimization that considered Dynamic Resampling, the survey done in this study has found that there is no related work that considered how combinations of Dynamic Resampling and preference-based guided search can further enhance the performance of EMO algorithms, especially if the problems under study involve computationally expensive evaluations, like production systems simulation. The aim of this thesis is therefore to study, design and then to compare new combinations of preference-based EMO algorithms with various DR strategies, in order to improve the solution quality found by simulation-based multi-objective optimization with stochastic outputs, under a limited function evaluation or simulation budget. Specifically, based on the advantages and flexibility offered by interactive, reference point-based approaches, studies of the performance enhancements of R-NSGA-II when augmented with various DR strategies, with increasing degrees of statistical sophistication, as well as several adaptive features in terms of optimization parameters, have been made. The research results have clearly shown that optimization results can be improved, if a hybrid DR strategy is used and adaptive algorithm parameters are chosen according to the noise level and problem complexity. In the case of a limited simulation budget, the results allow the conclusions that both decision maker preferences and DR should be used at the same time to achieve the best results in simulation-based multi-objective optimization.