Toward Adaptive Knowledge Transfer in Multifactorial Evolutionary Computation
觉得有用的话,欢迎一起讨论相互学习~
此篇文章为 [1]L. Zhou, L. Feng, K.C. Tan, J. Zhong, Z. Zhu, K. Liu, C. Chen, Toward Adaptive Knowledge Transfer in Multifactorial Evolutionary Computation, IEEE Transactions on Cybernetics. 51 (2021) 2563–2576. https://doi.org/10.1109/TCYB.2020.2974100. 的论文学习笔记,只供学习使用,不作商业用途,侵权删除。并且本人学术功底有限如果有思路不正确的地方欢迎批评指正!
Multi - factorial optimization (MFO)是Gupta等人在2016年提出的一种新的进化优化范式,用于多任务优化[1],[2]。与传统的进化算法(EA)在一次运行中只优化一个任务不同,MFO可以同时优化多个任务。在多任务优化过程中,由于任务间的隐性知识转移,MFO在收敛速度和求解质量方面都优于单任务匹配,包括连续、离散、以及连续和离散混合优化问题[3]-[7]。
在文献中,针对不同的优化问题,提出了许多具有不同搜索能力的交叉算子[17]-[20]。例如,一点和两点交叉[21],[22];均匀交叉[23],[24];模拟二进制交叉(SBX) [25], [26];等,都被引入来解决连续优化问题,而顺序交叉(OX) [27], [28];部分映射交叉(PMX) [29], [30];等,进行了组合优化设计。研究发现,不同的crossover operator对于解决不同的问题具有不同的能力。特别是在[31]中,有报道称,在具有单峰、多峰、噪声和离散搜索空间等各种属性的12个基准函数[32]上,多点和均匀交叉的性能显著优于一点和两点交叉。Michalewicz等人[33]的研究表明,在大多数无约束连续问题上,几何交叉优于算术交叉。作者还提出,几何交叉在全局最优位于可行域边界的问题上表现良好。在[25]中,我们发现SBX在困难的测试函数如双峰、不相等扩散函数和极点问题中表现优于其他实数编码的交叉,如BLX-α交叉。在这些工作中,我们可以观察到,不同的交叉具有独特的偏见产生后代,以指导进化搜索。因此,在进化多任务处理中,这些交叉点也可能具有不同的内隐知识转移能力。显然,交叉决定了任务间知识转移的质量。由于不同的交叉解在解生成过程中具有独特的偏倚性,因此利用适当的交叉解实现任务间的知识转移对进化多任务处理的性能至关重要。但值得注意的是,现有的进化多任务算法[8]-[16]采用单一且固定的交叉算子进行知识传递,当配置的交叉算子不能很好地匹配所遇到的问题时,会阻碍多任务处理。为了提高多任务处理的效率和有效性,需要使MFEA具有自适应的知识转移能力,以便针对不同的问题采用不同的知识转移crossover operator。然而,据我们所知,文献中没有关于自适应MFEA设计的研究。因此,本文试图填补这一空白。[25] R. B. Agrawal, K. Deb, and R. B. Agrawal, “Simulated binary crossover for continuous search space,” Complex Syst., vol. 9, no. 2, pp. 115–148, 1995.
Crossovers belonging to ABCOs transfer the combined genetic segments of the two parents to the offspring with an aggregation function, which is illustrated in Fig. 2. As the position of the offspring is fixed when the aggre-gation function is determined, the genetic inheritance is deterministic. Examples of ABCOs include arithmetical crossover [45], geometrical crossover [46], linear crossover (LX) [47], etc. In this article, the arithmetical crossover and the geometrical crossover are considered as the representative crossover operators in this group, which are given as follows.
2.2.2.1 算数交叉
子代决策变量是两个选择的亲本的线性组合
where λ is a predefined coefficient, which ranges from 0 to 1.
As can be observed, different crossovers have various forms that possess unique bias in generating offspring. Since crossover has also been employed in MFEA for implicit knowledge transfer across tasks, it may lead to different forms of knowledge transfer when different crossovers are config- ured, which could result in diverse multitask optimization performance
In particular, the PILS and NIMS SO multitask problems, and the CIHS and NIMS MO multitask problems are studied in this section. More details of the multitask benchmarks and configurations of the crossovers will be presented later in Section V. Fig. 4 presents the averaged convergence graphs of MFEA with six different crossover operators for knowledge transfer on the four multitask problems, over 20 independent runs. In Fig. 4(a) and (b), the y-axis denotes the averaged fitness (objective value) in log scale, and the x-axis is the generation number. Further, in Fig. 4(c) and (d), the y-axis represents the averaged IGD value in log scale, and the x-axis denotes the generation number.
As can be observed in Fig. 4, the performance of MFEA varies when different crossover operators have been employed for knowledge transfer. The performance gap in terms of objective value and IGD can be observed clearly from Fig. 4. For instance, on NIMS2 of Fig. 4(d), the BLX-α crossover converges about two times faster than the SBX crossover since the 200th generation. Furthermore, the arithmetical and geo- metrical crossover achieved the best performance on both SO PILS and SO NIMS. Particularly, on PILS [see Fig. 4(a)], while other operators are all stagnated at the early optimization stage, the arithmetical and geometrical crossover converged fast and obtained much better solutions. However, on MO CIHS, these two crossovers degraded to the worst crossover operators, as shown in Fig. 4(c). Further, on the MO prob- lems, the BLX-α crossover achieved the best IGD on MO PILS, while the geometrical crossover outperformed the others on MO NIMS2. These observations confirmed that different optimization problems may require different configurations of crossover for knowledge transfer across tasks, for efficient multitask optimization performance, and there is no single- crossover operator can perform well on all the SO and MO problems.
Toward robust and efficient multitask optimization performance when different problems are encountered, here,we thus propose a new MFEA with an adaptive configuration of crossover for knowledge transfer, called MFEA-AKT, which will be presented in the next section.
PROPOSED MFEA WITH ADAPTIVE KNOWLEDGE TRANSFER
This section presents the details of the proposed MFEA- AKT. In particular, the framework of MFEA-AKT is illustrated in Fig. 5, where the main differences between MFEA-AKT and the original MFEA are highlighted in the dashed boxes. In the proposed MFEA-AKT, we first introduce three new definitions, which are given as follows.