Dynamic Behavior Analysis Of Intelligent Optimization Algorithms In Phenotypic Space And Its Applications

Intelligent optimization algorithms design meta-heuristic algorithms via imitating the evolutionary or swarm behavior from nature,to solve kinds of complicated prob-lems in research and engineering.In recent twenty years,there is a rapid development in this field,where large quantities of new algorithm models have been proposed and great achievements have been obtained in real applications.However,as the number of algorithms increased,it is usually difficult to select suitable algorithm for researchers and engineers when solving their particular optimization problems.Moreover,intelli-gent optimization algorithms contain various parameters that have complex interactions.Even though the same algorithm,different configurations of parameters will generate varying optimization results.Therefore,parameter tuning is a hot while difficult topic in this research field as well.This dissertation tries to analyze the dynamic behavior of intelligent optimization algorithms for investigating their behavioral differences in phenotypic space,which is expected to better distinguish different algorithms or parameters for guiding and assist-ing the algorithm selection or parameter tuning task.Concretely,the main contents and contributions of this dissertation are as follows:[1]Proposed a quantitative behavior analysis method in phenotypic space based on the population evolvability.Concretely,we proposed three measures of popu-lation evolvability by combining behavior of algorithms and properties of prob-lems,as a new technique of dynamic fitness landscape analysis(FLA).After that,the significance and feasibility of the measures were demonstrated by theoretical and experimental studies.The proposed measures were adopted to the algorithm selection task on black-box numerical optimization problems to verify its effi-ciency,where the proposed approach achieved high accuracy from statistical re-sults,while the computational cost was obviously reduced in term of functions evaluations;[2]Proposed a parameter tuning method based on statistical racing.Concretely,the deficiencies of existing approaches for parameter tuning task were analyzed first.Then a new racing framework,i.e.KW-Race,was proposed according to the Kruskal-Wallis test.Finally,an indicator of convergence speed was embedded in the preceding KW-Race framework to form the Fast KW-Race,which consider-ably improves the efficiency of parameter tuning whist keeping high accuracy of the final results.The proposed parameter tuning framework better fulfills the re-quirements of researchers and engineers who want to solve their problems quickly and well;[3]Application of the proposed methods in an instance from engineering.Concretely,a NP-hard numerical optimization problem instance is solved by the proposed approaches to verify their utility.Good performance and stability of the results demonstrated advantages of the proposed approaches over existing methods,no matter in algorithm selection or parameter tuning task,which indicates a universal application value.In a nutshell,the defined measures of population evolvability provides a new di-rection for designing more applicable techniques of fitness landscape analysis.The proposed KW-Race and F-KW-Race framework are general frameworks for parameter tuning,which are suitable for both qualitative and quantitative parameters.This work has a promoting effect on applying intelligent optimization algorithms more reasonably and efficiently.