基于迁移学习的NSGAⅡ算法

NSGAⅡAlgorithm Based on Transfer Learning

为解决传统进化优化算法在定义域随机产生初始种群低效率问题,基于第二代非支配排序遗传算法融入迁移学习思想,设计基于迁移学习的NSGAⅡ算法。在历史信息储存库中找出与新任务相似的历史问题,历史问题Pareto的最优解集为源域,目标函数随机产生种群得到目标域,通过迁移成分分析方法将源域和目标域映射到高维再生核希尔伯特空间,得到新的源域和目标域,计算两者之间欧几里得距离并排序,得到新的种群;最后利用NSGAⅡ常规步骤对含有历史信息的种群进行搜索。针对10个改进的多目标测试函数进行试验,结果表明该算法可以提高种群搜索效率和算法收敛性能,优化解集的均匀性和多样性。

In order to improve the low-efficiency search method of traditional evolutionary optimization algorithms that randomly gener⁃ate initial populations in the defined domain,this paper designs the NSGAⅡalgorithm based on transferning learning to solve multiobjective problems.Firstly,we can find the historical problem similar to the new task in the repository with historical information and the Pareto optimal solution set of the historical problem is the source domain.The objective function randomly generates the population to obtain the target domain.Secondly,the source domain and target domain are mapped to Reproducing Kernel Hilbert Space through the method called Transfer Component Analysis(TCA)and then we get the new source domain and target domain.Thirdly,we calcu⁃late the Euclidean distance between the new source domain and target domain and sort them to get a new population that contains histor⁃ical information.Finally,we use NSGAⅡalgorithm to calculate and search the population with historical information.To verify this idea,we test on ten improved typical functions and the results show that the designed algorithm can improve the efficiency of popula⁃tion search and the convergence performance.