Comparison of Machine-Learning Algorithms for Estimating Cost of Conventional and Accelerated Bridge Construction Methods during Early Design Phase

Abstract

The use of accelerated bridge construction methods such as prefabricated bridge elements, lateral slide, and self-propelled modular transporter has increased in recent years to minimize on-site construction time and related traffic disruptions, and to improve safety, quality, and sustainability. This paper presents the development and evaluation of six novel machine-learning models for estimating the cost of conventional and accelerated bridge construction methods during the early design phase. The models were developed in four phases that focused on (1) collecting a data set of 413 conventional and accelerated bridge projects; (2) preprocessing the collected data to ensure its quality and reliability by identifying predicted and predictor variables, classifying predictor variables, cleaning data, transforming predictor variables, and splitting data into training and testing data sets; (3) training the models using ordinary least squares, least absolute shrinkage and selection operator (LASSO) regression, ridge regression, random forest, gradient boosting, and extreme gradient boosting; and (4) evaluating and validating the performance of the developed models. The outcome of the validation phase showed that the extreme gradient boosting model outperformed the other machine-learning models in terms of the metrics mean absolute percentage error, mean absolute error, and median absolute error; and the gradient boosting model outperformed the other models in the metric root mean square error. The developed machine-learning models and their improved cost estimating accuracy are expected to provide much-needed support to bridge planners and enable them to accurately estimate, compare, and select the most cost-effective construction method for their planned bridge construction projects during the early design phase.