Utilizing the Expected Gradient in Surrogate-assisted Evolutionary Algorithms
Utilizing the Expected Gradient in Surrogate-assisted Evolutionary Algorithms
复制标题
DOI：
10.1145/3583133.3590694
复制
发表时间：
2023-07
期刊：
影响因子：
--
通讯作者：
K. Nishihara;Masaya Nakata
中科院分区：
文献类型：
--
作者：
K. Nishihara;Masaya Nakata
在替代辅助进化算法（SAEAS）的领域中，高斯工艺（GP）是一种广泛使用的技术来近似目标函数。尽管GP模型可以提供要近似函数的预期梯度，但很少关注梯度信息的利用。因此，本文提出了一种预期的基于梯度的SAEA，其中使用GP模型提供的目标函数的预期梯度来进行有效的本地搜索。具体而言，所提出的算法首先通过差分进化算法进行全局搜索，以找到搜索空间的有希望的区域。然后，它为每个有前途的区域构建了GP模型，并在预期梯度的指导下在其模型上执行了准Newton方法（L-BFGS-B）。这种基于梯度的本地搜索旨在通过有效地找到各种本地最佳解决方案来充分搜索近似目标函数。实验结果表明，我们的算法在单目标优化基准套件上与最先进的SAEAS具有竞争力。
In the field of surrogate-assisted evolutionary algorithms (SAEAs), Gaussian Process (GP) is a widely used technique to approximate the objective function. Although a GP model can provide an expected gradient of a function to be approximated, little attention has been paid to the utilization of the gradient information. Thus, this paper presents an expected gradient-based SAEA, in which the expected gradient of the objective function provided by the GP models is utilized to conduct an efficient local search. Specifically, the proposed algorithm first conducts a global search with a differential evolution algorithm to find promising regions of the search space. Then, it builds a GP model for each promising region, and a quasi-Newton method (L-BFGS-B) is executed on its model with guidance from the expected gradient. This gradient-based local search intends to sufficiently search the approximate objective function, by finding various local optimal solutions in an efficient manner. Experimental results show that our algorithm is competitive with state-of-the-art SAEAs on a single-objective optimization benchmark suite.