Machine Learning Methods for Predicting Seismic Retrofit Costs
Juan F. Fung, Siamak Sattar, David Butry, Steven L. McCabe
Aging building clusters all around the world, especially in high seismic regions, will require a retrofit approach to improve the resilience of the built environment. One of the main challenges of retrofitting existing buildings is the associated cost. Fung et al.  develop a predictive modeling approach to estimating seismic retrofit costs. The predictive modeling approach uses historical data to predict retrofit costs for new buildings based on observable building characteristics ("the features"). The advantages of this approach are that it is (1) cheap, (2) fast, and (3) can be applied to a single building or a large inventory of buildings. However, the approach relies on a linear model for prediction, which assumes a restrictive relationship between retrofit cost and the features. In this paper, we consider machine learning methods that can capture more complex, potentially nonlinear relationships between retrofit cost and the features. The paper considers ensemble methods (bagging and boosting) as well as neural networks and compares their performance to that of the linear model. The results show that a neural network provides the best performance in terms of prediction error. In applications, a user faces a tradeoff between accuracy and interpretability: while ensemble methods and deep neural networks may improve accuracy, they are not as easily interpretable as the linear model.
Proceedings of the 17th World Conference on Earthquake Engineering
, Sattar, S.
, Butry, D.
and McCabe, S.
Machine Learning Methods for Predicting Seismic Retrofit Costs, Proceedings of the 17th World Conference on Earthquake Engineering, Sendai, , [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=929595
(Accessed June 7, 2023)