Information Technology Laboratory, National Institute of Standards and Technology, Gaithersburg, MD 20899, USA
Warren Rogers Associates, Middletown, RI 20842, USA
Evol Comput. 2015 Summer;23(2):309-42. doi: 10.1162/EVCO_a_00137. Epub 2014 Sep 25.
Setting the control parameters of a genetic algorithm to obtain good results is a long-standing problem. We define an experiment design and analysis method to determine relative importance and effective settings for control parameters of any evolutionary algorithm, and we apply this method to a classic binary-encoded genetic algorithm (GA). Subsequently, as reported elsewhere, we applied the GA, with the control parameter settings determined here, to steer a population of cloud-computing simulators toward behaviors that reveal degraded performance and system collapse. GA-steered simulators could serve as a design tool, empowering system engineers to identify and mitigate low-probability, costly failure scenarios. In the existing GA literature, we uncovered conflicting opinions and evidence regarding key GA control parameters and effective settings to adopt. Consequently, we designed and executed an experiment to determine relative importance and effective settings for seven GA control parameters, when applied across a set of numerical optimization problems drawn from the literature. This paper describes our experiment design, analysis, and results. We found that crossover most significantly influenced GA success, followed by mutation rate and population size and then by rerandomization point and elite selection. Selection method and the precision used within the chromosome to represent numerical values had least influence. Our findings are robust over 60 numerical optimization problems.
设置遗传算法的控制参数以获得良好的结果是一个长期存在的问题。我们定义了一种实验设计和分析方法，以确定任何进化算法的控制参数的相对重要性和有效设置，我们将这种方法应用于经典的二进制编码遗传算法（GA）。随后，正如在其他地方报道的那样，我们将具有这里确定的控制参数设置的 GA 应用于引导一群云计算模拟器朝着表现出性能下降和系统崩溃的行为。GA 引导的模拟器可以作为一种设计工具，使系统工程师能够识别和减轻低概率、代价高昂的故障场景。在现有的 GA 文献中，我们发现了关于关键 GA 控制参数和应采用的有效设置的相互矛盾的观点和证据。因此，我们设计并执行了一项实验，以确定在从文献中提取的一组数值优化问题上应用时，七个 GA 控制参数的相对重要性和有效设置。本文描述了我们的实验设计、分析和结果。我们发现交叉操作对 GA 的成功影响最大，其次是变异率和种群大小，然后是重新随机化点和精英选择。选择方法和在染色体中用于表示数值的精度影响最小。我们的发现对于 60 多个数值优化问题是稳健的。
