Prediction of Compressive Strength of Concrete: Critical Comparison of Performance of a Hybrid Machine Learning Model with Standalone Models
Prediction of Compressive Strength of Concrete: Critical Comparison of Performance of a Hybrid Machine Learning Model with Standalone Models
复制标题
DOI：
10.1061/(asce)mt.1943-5533.0002902
复制
发表时间：
2019-11-01
影响因子：
3.2
通讯作者：
Kumar, Aditya
中科院分区：
文献类型：
--
作者：
Cook, Rachel;Lapeyre, Jonathan;Kumar, Aditya
在过去的几年里，使用机器学习（ML）技术来模拟混凝土中的定量组成-性质关系受到了极大的关注。将萤火虫算法（FFA）与随机森林（RF）模型相结合，提出了一种新的混合ML模型（RF-FFA）用于混凝土抗压强度预测。萤火虫算法用于确定两个超参数的最佳值（即，树的数量和森林中每棵树的叶子数量）与数据集的性质和体积的关系。RF-FFA模型经过训练，以开发两种不同类别数据集的输入变量和输出之间的相关性;随后，模型利用这种相关性在以前未经训练的数据域中进行预测。第一类包括两个独立的数据集，其特征在于输入变量和输出之间的高度非线性和周期性关系，如三角函数所示。第二类包括两个真实世界的数据集，由混凝土的混合设计变量作为输入和其随年龄变化的抗压强度作为输出。混合RF-FFA模型的预测性能与常用的独立ML模型-支持向量机（SVM），多层感知器人工神经网络（MLP-ANN），M5 Prime模型树算法（M5 P）和RF进行了基准测试。用于评估预测准确性的指标包括五个不同的统计参数以及综合性能指数（CPI）。结果表明，无论数据集的性质和数量如何，混合RF-FFA模型在预测准确性方面始终优于独立的ML模型。
The use of machine learning (ML) techniques to model quantitative composition-property relationships in concrete has received substantial attention in the past few years. This paper presents a novel hybrid ML model (RF-FFA) for prediction of compressive strength of concrete by combining the random forests (RF) model with the firefly algorithm (FFA). The firefly algorithm is utilized to determine optimum values of two hyper-parameters (i.e., number of trees and number of leaves per tree in the forest) of the RF model in relation to the nature and volume of the dataset. The RF-FFA model was trained to develop correlations between input variables and output of two different categories of datasets; such correlations were subsequently leveraged by the model to make predictions in previously untrained data domains. The first category included two separate datasets featuring highly nonlinear and periodic relationship between input variables and output, as given by trigonometric functions. The second category included two real-world datasets, composed of mixture design variables of concretes as inputs and their age-dependent compressive strengths as outputs. The prediction performance of the hybrid RF-FFA model was benchmarked against commonly used standalone ML models-support vector machine (SVM), multilayer perceptron artificial neural network (MLP-ANN), M5Prime model tree algorithm (M5P), and RF. The metrics used for evaluation of prediction accuracy included five different statistical parameters as well as a composite performance index (CPI). Results show that the hybrid RF-FFA model consistently outperforms the standalone ML models in terms of prediction accuracy-regardless of the nature and volume of datasets.