dc.description.abstract |
Attempts to predict construction cost represent a problem of continual concern and interest to both researchers and practitioners. Such an attempt was carried out in this study to predict the construction cost associated with seismic retrofitting of existing structures. This prediction is a fairly new field of research in the civil engineering and construction industry, which has received little attention thus far despite its increasing importance and the large investment being directed towards seismic retrofit projects. In this research, retrofit net construction cost (RNCC) and its influential variables served as the dependent and independent variables, respectively. The total number of independent variables considered was fourteen, almost half of which have never been studied in the literature. These variables, together with RNCC, were tailored to develop parametric and non-parametric retrofit cost estimating (RCE) models, using the multi-linear regression (MLR) analysis and the artificial neural network (ANN) methodologies, respectively. The database used in this development was composed of 158 data points, each pertaining to a particular earthquake-prone public school with framed structures (i.e., concrete or steel structure). Of the total 158 data points in the research database, 75% (i.e., 121 samples) were randomly separated out to constitute the development or training dataset, and the remaining 25% (i.e., 37 samples) were held to constitute the hold-out or test dataset. The former dataset was mainly used to develop RCE models and evaluate their predictive accuracy ability, while the latter dataset was mainly used to evaluate the generalization ability of the developed RCE models. The backward elimination (BE) regression technique was employed to properly explore the extent of the influence of independent variables on the RNCC, and consequently those variables that made a statistically significant contribution to the prediction of the RNCC were identified. Of the fourteen independent variables examined in this research, seven variables addressing the building characteristics in addition to the site characteristics appeared to be significant predictors of the RNCC (i.e., p<0.05). Using the BE technique, fourteen different regression models were developed. Of these regression models, the model with the inclusion of all seven statistically significant variables was found to reveal the highest predictive accuracy ability. The causal relationships between the RNCC and each of its independent variables in this regression model were described. The generalization ability of a range of regression models, each with a specific set of statistically significant variables, was also evaluated. The results of this evaluation indicated that the most parsimonious regression model having only one independent variable (i.e., building area) showed the highest generalization ability. This indication lead to the establishment of a double-log regression model that can be simply, yet, reliably used for the general purpose of predicting the retrofit construction cost. In this research, intelligent non-parametric RCE models were developed for the first time in the literature by means of the ANN methodology. Using the feed-forward multi-layer perceptron (MLP) architecture, the sigmoid logistic activation function, and the back-propagation learning algorithm, a novel two-stage procedure was proposed for the successful development of ANN models. Upon the completion of this procedure, for each of the last seven MLR models derived from the implementation of the BE regression technique, the best correspondent ANN model with the highest generalization ability was developed. The application of the best ANN models to the training and test datasets resulted in the finding that the last ANN model, including all seven statistically significant variables, maintained the highest accuracy and generalization abilities. In addition, the sensitivity of the ANN methodology with respect to the variation of different parameters, which were considered in the proposed development procedure, was investigated and practical solutions for better exploitation of this methodology were suggested. Finally, the accuracy and generalization abilities of the ANN models were respectively compared to those of their MLR counterparts, using the same sets of independent variables. The results obtained from these comparisons illustrated the general superiority of the ANN methodology over the MLR analysis. This superiority was more pronounced when a greater number of independent variables were taken into account, and when a greater number of data points were included in the analysis. |
en |