Applying Artificial Neural Networks for Measuring and Predicting Construction-Labor Productivity

Abstract

Variations in labor productivity are the result of multiple influential factors. This paper attempts to develop a labor productivity model based on multilayer feedforward neural networks trained with a backpropagation algorithm by which complex mapping of factors to labor productivity is performed. To prevent networks from overfitting and improve their generalization, early stopping and Bayesian regularization are implemented and compared. The results proved a better prediction performance for Bayesian regularization than early stopping. To demonstrate the prediction performance of the presented models, the developed models are implemented at two real power plant construction projects. Moreover, in order to extract the influence rate of each factor on the predictive behavior of the neural network models and to identify the most influential factors a sensitivity analysis is conducted. This paper focuses on the work involved in installing the concrete foundations of gas, steam, and combined cycle power plant construction projects in the developing country of Iran. This study contributes to the construction project management body of knowledge by investigating the influential factors on labor productivity and developing an artificial neural network to measure and predict labor productivity in developing countries using the Bayesian regularization and early stopping methods. This approach provides insight into better ways of modeling labor productivity.