Research On Interactive Evolutionary Multi-Objective Optimization Via Preference Learning

In multi-objective optimization problems(MOP),the improvement of one sub-objective may cause the performance of another or several sub-objectives to decrease.That is to say,it is impossible to achieve the optimal value of multiple objectives at the same time.Partially due to the population-based characteristics,evolutionary algorithms(EA)have been widely recognized to be effective for multi-objective optimization.Due to the negli-gence of decision maker(DM)'s preference in the loop,it is not guaranteed to identify the solutions most relevant to the DM's aspiration.This is further aggravated when the num-ber of objectives becomes large given that the PF approximation is too sparse to cover the solutions of interest(SOI),letting alone the cognitive burden for understanding the high-dimensional data.Most of the algorithms are designed to approximate the entire Pareto-optimal front(Pareto-optimal front,PF).In reality,the decision maker may only be interested in his/her region of interest(ROI),that is,a part of the PF.Solutions in other parts of PF may be useless.Therefore,this thesis develops an interactive framework for dominance-based and decomposition-based evolutionary multi-objective optimization(EMO)algorithm to lead a DM to the preferred solutions of her/his choice.This interactive framework consists of three modules,i.e.,consultation,preference elicitation and optimization modules.The consultation module progressively learns the value function representing the preference information of the DM during the loop through the holistic pairwise comparisons of so-lutions.Once the value function is learned,the preference elicitation module converts it into a form that the EMO algorithms can use.There are different forms between the dominance-based EMO algorithms and decomposition-based EMO algorithms.Finally,the optimization module is used to optimize the population.In the consultation module,this thesis proposes to use Gaussian process and learning-to-rank(LTR)to learn the value function through the holistic pairwise comparisons of the solution.After every several generations,the DM is asked to score a few candidate solutions in a consultation session.EMO algorithms combine with the value function to guide the population toward the SOI.For 13 test problems,a large number of experiments with 2 to 10 objectives fully prove the effectiveness of the interactive framework,and the combination of preference learning with pairwise-based algorithms can help two widely used EMO algorithms(NSGA-II and MOEA/D)find the preferred solutions that the DM are satisfied with.What's more,we found that LTR is expected to break the possibility of higher target dimensional space in interactive EMO algorithms.