The Study Of Preselection Via Classification In Evolutionary Algorithms

In both scientific and engineering areas,there exist many optimization problems with different characteristics,such as large-scale,nonlinear,many objectives.Traditional gradient based optimization methods may fail to tackle such kind of complicated optimization problems,therefore many researcher resort to heuristic optimization methods.Among them,evolutionary algorithm(EA)is a promising one.EAs are characterized by the population based search property,the random search property,and the fitness value driven search property.An EA typically contains five major components,i.e,the solution representation,the population initialization,the stop condition,the offspring generation,and the selection.Although EAs have been successfully applied to deal with different complicated optimization problems,they are criticized for some disadvantages,such as the uncertainty of solutions,and low search efficiency.In EAs,many evaluated solutions are abandoned directly because of their low quality.This is one of the reasons that lead to the low search efficiency.To deal with this problem,this thesis focuses on generating high quality solutions in the offspring generation component of EAs.A classification based preselection(CPS)has been proposed for this purpose in the thesis.The main work of the thesis includes:Firstly,a conceptual classification based preselection by using a binary classification is introduced.The main idea of the proposed strategy is to deal the preselection as a binary classification problem.At first,the solutions in current population are divided into promising class and unpromising class according to the sample definition strategy.Then,using the defined samples to build the classification model.Finally,the generate candidate offspring solutions are labeled by the built model,and the one with promising label is chosen out.The experimental results suggest that the proposed strategy can significantly improve the efficiency of some state-of-the-art evolutionary algorithms.Secondly,two practical classification based preselection strategies are designed to improve the efficiency of the conceptual CPS.One is based on the one-class classification.The idea is treating all of the solutions in current population as the promising ones,and then using them to build the one-class classification model.The other one is based on the fuzzy classification.The idea is using the membership function to label the candidate offspring solutions and choose the promising one out to do following optimization procedures.Theexperimental results suggest the two practical strategies can improve the efficiency of the conceptual CPS.Thirdly,the work on preselection has been extended in two aspects.One is on applying the CPS to multiobjective optimization.In this approach,the non-dominate and dominate solutions are defined as the promising class and the unpromising class.And then use them to build the classification model.The experimental results suggest the CPS can significantly improve the efficiency of three kinds of multiobjective evolutionary algorithms.The other is on extending the classification based preselection to the classification based environmental selection.Like the offspring generation operator,the environmental selection is another major EA component.The main idea of the strategy is to use the classification model to choose the promising solutions out from the parent population and the offspring solution without evaluating them.