Font Size: a A A

Research On Key Technologies Of Evolutionary Multiobjective Optimization Algorithm Based On Decomposition

Posted on:2020-04-03Degree:DoctorType:Dissertation
Country:ChinaCandidate:X J ChenFull Text:PDF
GTID:1360330605481305Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Multiobjective optimization problems often need to consider several con-flicting objectives simultaneously.In most cases,the improvement of one ob-jective may cause the performance of other objectives to be reduced,and it is impossible to achieve the optimal for multiple objectives at the same time,can only coordinate trade-off and compromise treatment between all objectives,makes all objectives as possible as to achieve the optimal.How to obtain the optimal solution to such problems has always been the focus of attention in academia and industry.Evolutionary algorithm is a population-based stochas-tic optimization algorithm that simulates the evolutionary process of natural organisms.Using evolutionary algorithms to solve multiobjective optimization problems has unique advantages:can solve the search problem on large-scale complex space;obtain multiple compromise solutions in one run.Due to these advantages,the evolutionary multiobjective optimization algorithms have grad-ually emerged and become one of the mainstream research directions of evolu-tionary computing.In recent years,the decomposition-based multiobjective optimization evo-lutionary algorithms(MOEA/D)have gradually become a research hotspot.The basic idea of this type of algorithm is to decompose the multiobjective optimization problem into a set of scalar subproblems.Neighboring subprob-lems cooperate with each other to generate new offspring solution,while new offspring solution not only update the solution to the corresponding subprob-lem,but also update the neighborhood solutions.In this way,all subproblems are optimized at the same time,and the whole set of approximate solutions can be obtained finally.This thesis conducts in-depth research on some key com-ponents of the decomposition-based evolutionary multiobjective optimization algorithm,including the generation generation strategy,preselection strategy and replacement strategy,and applies the research results to the data mining problem.The m ain work of the thesis includes:1.Offspring generation strategy research.Aiming at the problem that the current offspring generation strategies are difficult to generate high-quality off-spring,this paper proposes an offspring generation strategy(MOEA/D-OGS).Firstly,this paper generates a temporary solution set for each subproblem based on the superiority and inferiority between the parent solutions;then,uses the surrogate model to estimate the objective values of each temporary solution,sorts and gets the optimal temporary solution;finally,uses the differential evo-lution algorithm and the optimal temporary solution to generate the offspring solution.The experimental results suggest that the proposed offspring genera-tion strategy achieves better performance than the traditional evolutionary op-erator.2.Preselection strategy research.Aiming at the problem that the current classification-based preselection strategies are difficult to identify the optimal candidate solution,this paper gives two improvement strategies:(1)proposes a hybrid individual selection mechanism(MOEA/D-CS).Firstly,this paper gen-erates candidate solution set for each subproblem and constructs a classification model;then,uses the classification model to distinguish candidate solutions and preserves good candidate solutions;finally,uses the similarity measure method to calculate the similarity between each good solution and its neighborhood so-lution,takes the objective values of the nearest neighborhood solution as the objective values of the good solution,and sorts and gets the optimal offspring solution.(2)This paper proposes a surrogate individual selection mechanism(MOEA/D-SISM).Firstly,this paper generates a candidate solution set for each subproblem based on the offspring generation strategy;then,uses the surrogate model to estimate the objective values of each candidate solution and obtain the optimal offspring solution.The experimental results suggest that the proposed two improvement strategies achieve better performance than the classification-based preselection strategy.3.Replace the strategy study.Aiming at the problem that the current replace strategies are difficult to balance the convergence and diversity of pop-ulation,this paper gives two improvement strategies:(1)proposes a hybrid neighborhood replacement strategy and global replacement strategy method(MOEA/D-HRS).This paper uses the threshold to determine the probability that the neighborhood replacement strategy or the global replacement strategy being selected,and uses the offspring generation strategy to generate solution for each subproblem.This hybrid replace strategy can better balance the conver-gence and diversity of population.(2)To overcome the disadvantage of manual threshold adjustment in hybrid replacement strategy,this paper proposes a re-placement strategy based on adaptive method(MOEA/D-ARS).In this strategy,fitness improvement rate and decay mechanism are used to balance the proba-bility that neighborhood replacement strategy and global replacement strategy are selected.The experimental results suggest that the proposed two improve-ment strategies significantly improve the performance of the MOEA/D.4.Applied research.Extreme learning machines(ELM)use a single-layer feedforward neural network and are widely used in data mining problems.Aim-ing at the problem that the current extreme learning machine randomly gen-erates initial weight and hidden layer bias,which leads to unstable network performance,this paper proposes a multiobjective extreme learning machine parameter optimization strategy(MOEA/D-ELM).This strategy optimizes two conflicting objectives:(1)training error;(2)generalization performance.The experimental results suggest that the proposed parameter optimization strategy improves the model accuracy of the extreme learning machines.
Keywords/Search Tags:multiobjective optimization, offspring generation strategy, preselection strategy, replacement strategy
PDF Full Text Request
Related items