Nonlinear tree based regression ensemble modeling for repair cost prediction in earthquake damaged RC bridges

Data driven models are useful in various classification and regression problems in bridge engineering. Although machine learning applications in identifying the performance and characteristics of bridges has been increasing, repair cost modeling using real earthquake damage data is only sparsely reported in the literature. This study assesses various machine learning models for repair cost estimation of reinforced concrete (RC) bridges damaged by the 2015 Gorkha earthquake in Nepal. Ensemble learning interpretation is used to hierarchize bridge attributes based on their relative importance in repair cost prediction. The impact of individual features is also assessed using various combinations of bridge features. The results indicate that two categorical features, discrete damage state and foundation type, are the most important in predicting repair cost. Of the various models tested, extra trees (ET) ensemble outperformed any other ensemble as well as base learner methods. The findings indicate that, for the case-study data used here, repair cost is better estimated by a classification model for damage state in series with a regression model than just a regression model.