Modified Algorithms for Fast Construction of Optimal Latin-Hypercube Design
Modified Algorithms for Fast Construction of Optimal Latin-Hypercube Design
复制标题
DOI：
10.1109/access.2020.3032122
复制
发表时间：
2020
期刊：
影响因子：
3.9
通讯作者：
Qingyu Wang;T. Nakashima;Chenguang Lai;Hidemi Mutsuda;Taiga Kanehira;Y. Konishi;Hiroyuki Okuizumi
中科院分区：
文献类型：
--
作者：
Qingyu Wang;T. Nakashima;Chenguang Lai;Hidemi Mutsuda;Taiga Kanehira;Y. Konishi;Hiroyuki Okuizumi
由于没有高质量的样本就无法保证优化的精度，因此在设计空间中进行实验的有限个评价点的分布是一个重要的问题，特别是当获得样本的实验代价很高时。最优拉丁超立方体设计（OLHD）是利用有限的评价点来表示设计空间的一种方法，具有良好的空间填充性能，被广泛应用于实验设计（DOE）中。然而，OLHD生成需要大量时间。本研究的重点是进一步提高OLHD生成的时间和拉丁超立方体设计（LHD）优化的效率。在已有算法的基础上，提出了两种改进算法，即改进的增强随机进化算法（MESE）和平移传播改进的增强随机进化算法（TPMESE）。MESE算法是从增强的随机进化（ESE）算法修改，通过使用一种新的更新方法的“温度”，而TPMESE算法优化LHD通过平移传播（TPLHD），而不是像MESE算法优化一个随机LHD。通过与几种著名的启发式算法和每一种原始算法的比较，对不同尺寸的LHD进行了优化测试，评价了它们的性能。对于所有的情况下，所提出的算法表现出更好的收敛性能比其他启发式算法参与我们的比较。对于大中型的LHD，MESE算法能更快地收敛到与原算法（ESE）相同水平的解。对于大型LHD，TPMESE算法是获得接近最优或足够接近最优设计的最省时的算法。
As accuracy of optimization can not be guaranteed without high-quality samples, the distribution of a finite number of evaluation points where experiments should be conducted in design space is an important issue, particularly when the experiment to obtain sample is expensive. To utilize limited number of evaluation points to represent the design space, optimal latin-hypercube design (OLHD), with considerable space-filling quality, is widely used as a methodology in design of experiments (DOE). However, OLHD generation requires significant time. This study focuses on further improvement of efficiency in generation of OLHD in terms of both time and latin-hypercube design (LHD) optimization. Two modified algorithms, namely the modified enhanced stochastic evolutionary (MESE) and translational propagation modified enhanced stochastic evolutionary (TPMESE) algorithms, based on existing algorithms, are proposed. The MESE algorithm is modified from the enhanced stochastic evolutionary (ESE) algorithm by using a new update method for “temperature,” while the TPMESE algorithm optimizes the LHD via translational propagation (TPLHD) instead of optimizing a random LHD like the MESE algorithm does. Their performance is evaluated by comparison with several famous heuristic algorithms and each original algorithm using optimization tests of LHDs with various sizes. For all cases, proposed algorithms show better performance of convergence than other heuristic algorithms participated in our comparison. For large and medium LHDs, the MESE algorithm faster converges to a solution with the same level as original algorithm (ESE). For large LHDs, the TPMESE algorithm is the most time efficient algorithm in obtaining near-optimal or sufficient near-optimal designs.