For multi-objective problems existing in the conflict between target but interrelated contradiction,many scholars and researchers have already done a very in-depth research.Particle swarm optimization algorithm is used to study not only one of the most commonly used algorithm in multi-objective optimization problem,and has also been widely used in the domains of feature selection problem,and feature selection is an important process of data preprocessing.Particle swarm optimization algorithm not only has the advantages of simple algorithm,few parameters and fast convergence,but also has the disadvantage of easily trapped in local optimization.Therefore,this thesis improves and enhances the research of particle swarm optimization in multi-objective optimization and feature selection.The main improvement work in these two aspects is divided into the following parts:(1)This thesis proposes a two-stage global optimal solution selection strategy for the research of multi-objective particle swarm optimization algorithm.The two-stage global optimal solution selection strategy is considered from the perspectives of decision space and target space respectively,and the global optimal solution is selected by similarity measurement strategy and Knee Point concept to balance the convergence and diversity of the solution set on Pareto Front.Secondly,the Cauchy mutation operation is improved to dynamically disturb the population with the change of iteration times.In the early and middle stages of the iteration,large disturbance makes the particles jump out of the local optimum,and in the later stages of the iteration,small disturbance accelerates the convergence of the population and improves the searching ability of particle.Finally,the Archive update strategy based on the crowding distance is improved.The improved method further refines the update and maintenance operations of Archive,making the distribution of particles in the Archive more uniform and dispersed(2)For multi-objective feature selection,this thesis firstly defines a set concept based on the Archive and Pareto Front,then proposes a local learning strategy based on mutual information method,and uses mutual information to calculate the correlation between features and between features and labels.In this thesis,the local learning strategy based on mutual information method is combined with the new set concept to make the particles on Pareto Front of the population learn from the particles in the Archive,which can effectively improve the quality of particles in the population.At the same time,in order to prevent the particle swarm optimization algorithm from falling into local optimum in the middle period of iteration and causing the particle falling into prematurity,this thesis proposes an adaptive disturbance strategy according to the number of iteration that the particle's historical optimal solution has not been updated.The probability of particle disturbance and the number of the particle's feature are determined dynamically based on the number of iteration that the particle's historical optimal solution has not been updated.This thesis uses the improved multi-objective particle swarm optimization algorithm with six multi-objective optimization algorithm through the contrast test,the experimental results show that the proposed algorithm converges better on the real Pareto frontier.Meanwhile,the multi-objective feature selection method base on hybrid mutual information and particle swarm optimization is proposed in this thesis was tested on 15 data sets.The experimental results show that the classification error rate is also reduced while the number of features is reduced. |