Dynamic Resampling for Preference-based Evolutionary Multi-Objective Optimization of Stochastic Systems

In Multi-objective Optimization many solutions have to be evaluated in order to provide the decision maker with a diverse choice of solutions along the Pareto-front. In Simulation-based Optimization the number of optimization function evaluations is usually very limited due to the long execution times of the simulation models. If preference information is available however, the available number of function evaluations can be used more effectively. The optimization can be performed as a guided, focused search which returns solutions close to interesting, preferred regions of the Pareto-front. One such algorithm for guided search is the Reference-point guided Non-dominated Sorting Genetic Algorithm II, R-NSGA-II. It is a population-based Evolutionary Algorithm that finds a set of non-dominated solutions in a single optimization run. R-NSGA-II takes reference points in the objective space provided by the decision maker and guides the optimization towards areas of the Pareto-front close the reference points.In Simulation-based Optimization the modeled and simulated systems are often stochastic and a common method to handle objective noise is Resampling. Reliable quality assessment of system configurations by resampling requires many simulation runs. Therefore, the optimization process can benefit from Dynamic Resampling algorithms that distribute the available function evaluations among the solutions in the best possible way. Solutions can vary in their sampling need. For example, solutions with highly variable objective values have to be sampled more times to reduce their objective value standard error. Dynamic resampling algorithms assign as much samples to them as is needed to reduce the uncertainty about their objective values below a certain threshold. Another criterion the number of samples can be based on is a solution's closeness to the Pareto-front. For solutions that are close to the Pareto-front it is likely that they are member of the final result set. It is therefore important to have accurate knowledge of their objective values available, in order to be able to to tell which solutions are better than others. Usually, the distance to the Pareto-front is not known, but another criterion can be used as an indication for it instead: The elapsed optimization time. A third example of a resampling criterion can be the dominance relations between different solutions. The optimization algorithm has to determine for pairs of solutions which is the better one. Here both distances between objective vectors and the variance of the objective values have to be considered which requires a more advanced resampling technique. This is a Ranking and Selection problem.If R-NSGA-II is applied in a scenario with a stochastic fitness function resampling algorithms have to be used to support it in the best way and avoid a performance degradation due to uncertain knowledge about the objective values of solutions. In our work we combine R-NSGA-II with several resampling algorithms that are based on the above mentioned resampling criteria or combinations thereof and evaluate which are the best criteria the sampling allocation can be based on, in which situations.Due to the preference information R-NSGA-II has an important fitness information about the solutions at its disposal: The distance to reference points. We propose a resampling strategy that allocates more samples to solutions close to a reference point. This idea is then extended with a resampling technique that compares solutions based on their distance to the reference point. We base this algorithm on a classical Ranking and Selection algorithm, Optimal Computing Budget Allocation, and show how OCBA can be applied to support R-NSGA-II. We show the applicability of the proposed algorithms in a case study of an industrial production line for car manufacturing.