Research Multi-objective Optimization Tasks Based On Parallel Genetic Algorithms

The main properties of real practical tasks of optimization includes the presence of many criteria and many local extremes and significant constraints,which makes it impossible to apply classical gradient optimization methods.The way out of this situation is to use adaptive stochastic algorithms,which successfully overcome these difficulties.There is a class of problems whose solution is difficult or impossible to present in a formalized form.Some tasks are not solved using traditional algorithms in an acceptable time.Using genetic algorithms can solve these problems.In addition,genetic algorithms are adaptive,can be parallelized,easily combined with neural networks and elements of fuzzy logic.The object of the research is the methods and algorithms for solving multi-objective optimization problems.The subject of the study is evolutionary algorithms of the parallel type.The aim of the work is to increase the efficiency of the process of forming the Pareto set,using parallel genetic algorithms and improving the procedures for information exchange between subpopulations.In this study to achieve goals,the following work was done:1.Analysis of methods and algorithms of multi-objective optimization problem and resolving the research task.In this chapter four most common methods,implementing various schemes of suitability and selection were investigated.The VEGA method,FFGA method,SPEA,and the NPGA method.-The VEGA method uses selection to switch target functions,that is,selection is performed on suitability of individuals for each of the K criteria separately.-The FFGA method uses a Pareto-dominant ranking procedure for individuals,where the rank of each individual is determined by the number of its dominant individuals.-The SPEA method creates an external set,where stores individuals that are not dominant in relation to other members of the population and represent in the end a non-dominant front.-In the NPGA method,the fitness assignment phase is replaced by a modified fitness separation scheme using the concept of a niche that is defined for individuals in the space of alternatives or target function and provides the ability to maintain diversity and allowing get a representative Pareto set.The effectiveness of the approaches was investigated on a representative set of test tasks,with various target functions,with different number of variables.2.Development of the parallel genetic algorithm for solving task of multi-objective optimization.In this chapter we develop an experimental scheme for the exchange of genetic information between subpopulations and leader's selection procedure.On the basis of SPEA method was developed an algorithm that includes all improvements described above.3.Program realization and investigation of the genetic algorithm properties.In this chapter described developed program for studying the properties of a genetic algorithm and program that implements the solution of the multi-objective optimization problem.All numerical results are obtained by averaging the data that was consistently recorded during the experiments with the program,with different migration scheme,different numbers of subpopulations and different numbers of individuals in populations.Conclusion.Developed a program that implements a parallel evolutionary algorithm for solution of the multi-objective optimization task with different schemes for the exchange of genetic information between subpopulations.The program allows change the number of individuals and the number of subpopulations and,thus,makes it possible to investigate the effect of the size of subpopulations and other parameters of a parallel genetic algorithms on the probability of a guaranteed finding of the global extremum.Experiments with the program showed that the parallel evolutionary algorithm is able to find the set of points which is the approximation of the Pareto set.