分组策略下的高维目标协同进化算法

Group-based Many-objective Co-evolutionary Algorithm

为平衡高维目标优化问题在进化过程中收敛性与多样性的冲突,本文提出基于两阶段分配策略的高维目标协同进化算法.首先,利用参考向量将种群进行分组,划分为若干个子种群,在进化前期,主要根据子种群中非支配解密度评估子种群优化难易程度;在进化后期,主要根据非支配解分布的广度评估子种群多样性,以此确定子种群进化潜力,为高进化潜力的子种群分配目标向量.然后,在整个目标空间内产生随机目标向量,防止其余个体的退化.本文将改进后算法与PICEAg在3、5、7、10、15维DTLZ1-7函数上进行性能对比实验.仿真实验结果表明,除DTLZ5测试问题外,改进后算法在收敛性及多样性上均优于原算法.

To balance conflicts between convergence and diversity in the optimization of many-objective problems,this paper proposes a preference inspired many-objective co-evolutionary algorithm based on tw o-stage allocation strategy.Firstly,the reference vector is introduced to divide the population into several sub-populations.Secondly,in the early stage of evolution,the evolutionary potential is evaluated mainly by the density of non-dominant solutions of sub-population.And in the later stage of evolution,the evolutionary potential is evaluated mainly by the distribution of non-dominant solutions of sub-population.Then dynamically assign the preference vector to the sub-population with high evolution potential.Lastly,a part of random preference vectors are generated in the w hole objective space to prevent degradation of other individuals.In this paper,the improved algorithm PIECAg-PE is compared with the original algorithm PIECAg on the 3,5,7,10,15-objective DTLZ1-7 test problems.Simulation results show that,except for DTLZ5 test problems,the improved algorithm is superior to the original one in the convergence and diversity.