The Research On Large Scale Many-objective Optimization Evolutionary Algorithm Based On Variable Classification

In the filed of multi-objective evolutionary algorithms(MOEAs),most traditional optimization algorithms mainly focus on the changes on the scalablity of objectives with different optimization problems,while little work has considered the scalability to the number of decision variables.Nevertheless,many real-world problems can involve both many objectives and large-scale decision variables,Experimental results show that,despite that most existing MOEAs have been well assessed on the MOPs with a small number of decision variables,their performance degenerates dramatically on MOPs with hundreds or even thousands of decision variables,To tackle such problems,there are some methods have been proposed that can effectively solve such problems.One of the methods is based on problem transformation,the second is base on cooperative coevolutionary,the third kind is based on variable classification.In the methods of variable classification,there two famous algorithms called MOEA/DVA and LEMA,the MOEA/DVA algorithm divides variables into three categories,while the LEMA algorithm divides variables into two categories by clustering according to angle.Combining the advantages of the two algorithms above,this paper proposes a method of variable classification based on the number of dominant layers of individual population after interacting,and divides decision variables into two sets: convergence correlation and distribution correlation,then optimizes them respectively.For convergence related variables,we adopted the method of variable interaction analysis for grouping optimization.After Convergence Optimization,we use a specific MOEA for overall optimization.For diversity related variables,we first used the angle as the individual diversity evaluation standard,and finally we carry out an overall optimization again.Two high-dimensional algorithms NSGA-III and Kn EA were selected as comparisons to compare the effect of the proposed algorithm on large-scale problems with the traditional excellent high-dimensional multi-objective algorithm.It is also compared with MOEA/DVA and LEMA,two popular multi-objective large-scale algorithms based on variable classification.The experimental results show that the proposed algorithm is obviously superior to the traditional high-dimensional multi-objective algorithm,and is more effective and competitive than the two popular multi-objective large-scale algorithms in some test problems.