A Dominance With CPU+GPU Heterogeneous Computing For Many-objective Optimization

For many optimization problems,Pareto-dominated algorithms will face the problem of selective pressure loss.The core reason is that most of the individuals in the population will become non-dominated solutions with the increase of the number of objectives,which will lead to the selection of pressure loss based on the Pareto dominance relationship.The new LWM-dominance method can solve the prossure problem traditional Pareto-dominance relations in the face of high-dimensional multi-objective optimization problems.However,because solving the LWM non-dominated solution process is a linear programming problem,the calculation time is long,the efficiency is low,and extra time overhead is introduced.Usually,parallelization is the simplest and most efficient way to improve the efficiency of algorithms,Compared with CPU,GPU parallel way has low cost and high efficiency,which is very suitable for improving the efficiency of the algorithm.This paper analyzes that LWM dominating algorithm is faced with a long time-consuming problem in solving large-scale high-dimension multi-objective optimization problems.Combining the advantages of GPU parallel computing,a novel LWM high-dimension multi-objective dominating method for CPU and GPU heterogeneous computing is proposed.The utility of the LWM dominance algorithm is improved.In order to solve the problem of error accumulation in the process of solving LWM non-dominated solutions,a high precision floating-point number operation was used on the GPU side.To verify the consistency of the serial algorithm and the parallel algorithm results,a set of experiments based on random distributed uniform data was designed to vertify the results.There is a small difference in the final result of the two algorithms.Further analysis of the algorithm's acceleration efficiency in each target,it's concluded that as the number of targets increases,the speedup ratio of the proposed parallel algorithm increases.At the same time,we compared the serial and parallel LWM efficiency.In the experiments of two multi-objective benchmarks,it was shown that LWM-GA is not completely suitable for solving two-objective optimization problems.Experiments in the DTLZ multi-objective reference program show that LWM-GA is suitable for hyper scale and multi-objective optimization problems,and as the number of targets increases,the optimization efficiency of parallel algorithms is higher.In the study of the combination of LWM and multi-objective evolutionary algorithms,this topic proposes two methods of integration.LWM-GA is the use of LWM dominance relations to replace existing Pareto dominance in NSGA-?+LWM,which uses NSGA-? to solve the many-objective optimization problem,and then uses the LWM dominance relationship to optimize the solution set.In order to compare the performance of LWM-GA and NSGA-?+LWM methods,HyperVolume was used as an evaluation index of multi-objective evolution method in this project,and an experimental comparison was conducted.The experimental results show that the solution set distribution and convergence obtained by NSGA-? and LWM-GA methods are consistent.At the same time,in the comparision of NSGA-?and NSGA-?+ LWM methods,the experimental results show that the LWM dominance relationship can further reduce the non-dominated solution generated by NSGA-? on the premise of maintaining the solution set distribution,NSGA-?+LWM approach.It can help decision makers to more quickly choose the optimal solution.