Font Size: a A A

Solving Many-Objective Optimization Problems With Evolutionary Algorithm And Its Application

Posted on:2016-02-22Degree:DoctorType:Dissertation
Country:ChinaCandidate:X H ChenFull Text:PDF
GTID:1228330464459505Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
Multi-Objective Evolutionary Algorithms(MOEAs) have been recognized as efficient methods to solve Multi-objective Optimization Problems(MOPs). The most effective and frequently used MOEAs have been focused on problems with a small number of objectives. Many-objective optimization problems which refer to problems with more than 5 objectives have become a hot topic in the field of MOPs recently.This thesis aims at designing efficient and effective MOEAs for solving many-objective optimization problems. The main work of the thesis is summarized as follows:(1) A Multi-objective Shuffled Frog Leaping Algorithm with ε-indicator is proposed for solving many-objective optimization problems. The use of ε-indicator based fitness assignment method helps to conduct local evolution and to update the archiver. An improved population partitioning method based on geometrical structure is introduced, in which the closely located non-dominated solutions will be clustered in the same subpopulation while the dominated solutions are set apart according to their approximation to the non-dominated set. A dynamic selection of the global best solution is achieved by selecting the nearest non-dominated solution in the archiver to replace the current worst solution. The proposed algorithm was applied on problems with 3 to 50 objectives, and the experimental results show that the algorithm outperforms other representative algorithms on both the performance of convergence and speed of convergence to Pareto optimal front.(2) A new objective reduction algorithm based on sparse feature selection is presented to reduce the size of the original problem. It adopts the non-dominated solution set obtained by certain multi-objective optimization algorithm as sampling data to capture the geometrical structural characteristics and Pareto-dominance relationships. A sparse regression model is constructed by which the sparse projection matrix mapping the high dimensional data into low dimensional space is learned. The importance of each objective is measured in terms of the projection matrix. This can be used to realize dimension reduction, either to a specified number of objectives, or tofind a minimum objective subset satisfying certain performance criteria.(3) Online objective reduction which combines the methods of(1) and(2) is provided for the purpose of converting the original problem into the one with less fewer number of objectives. Three different structures are studied, including:(i) fixed number of objectives reduction;(ii) adaptive reduction of the number of objectives;(iii)weighted aggregation of selected objectives based on their importance measure.Experimental results show that the aggregation-based algorithm is the best among all three structures.(4) Waveform design for Pulsed Doppler Radar as a real many-objective optimization problem with nine objectives is studied. We obtained a satisfactory approximation set by ε-indicator based Multi-objective Shuffled Frog Leaping Algorithm. We analyzed the preference order of the nine objectives of this problem by the proposed objective reduction method and obtained the results agreeing with the fact.The online objective reduction based method proposed was also applied to solve the problem and the superiority over multi-objective evolutionary algorithm used alone was thus demonstrated.
Keywords/Search Tags:Many-objective Optimization Problems, Multi-Objective Evolutionary Algorithms, Shuffled Frog Leaping Algorithm, Objective Reduction, Feature Selection
PDF Full Text Request
Related items