Abstract:
Following an extensive literature review, it was established that professional subjective judgment and regression analysis were the two main techniques utilized for predicting the seismic retrofit construction cost. The study presented in this paper aims at predicting this cost by employing a more advanced modeling technique known as the artificial neural network (ANN) methodology. Using this methodology, a series of nonparametric ANN models was developed based on significant predictors of the retrofit net construction cost (RNCC). Data on these predictors, together with the RNCC, were collected from 158 earthquake-prone public school buildings, each having a framed structure. A novel systematic framework was proposed with the aim to increase the generalization ability of ANN models. Using this framework, the values of critical components involved in the design of ANN models were defined. These components included the number of hidden layers and neurons, and learning parameters in terms of learning rate and momentum. The sensitivity of the developed ANN models to these components was examined, and it was found that the predictive performance of these models was more influenced by the number of hidden neurons than by the value of learning parameters. Also, the results of this examination revealed that the overlearning problem became more serious with an increase in the number of predictors. In addition to the framework proposed for the successful development of ANN models, the primary contribution of this study to the construction industry is the introduction of building total area as the key predictor of the RNCC. This predictor enables a reliable estimation of the RNCC to be made at the early development stage of a seismic retrofit project when little information is known about the project.
