Performance Evaluation of Machine Learning Algorithms for Seismic Retrofit Cost Estimation Using Structural Parameters

Estimation of the seismic retrofit cost (SRC) is a complicated task in construction projects. In this study, the performance of four machine learning algorithms (MLAs), including Random Forest (RF), Extreme Learning Machine (ELM), Classification and Regression Tree (CART), and Multivariate Adaptive Regression Spline (MARS), was examined in estimating SRC values. The total floor area (TFA), number of stories (NS), seismic weight (SW), seismicity (S), soil type (ST), plan configuration (PC), and structural type (STT) were considered as structural input variables. To achieve the best performance of applied MLAs, twenty-two scenarios based on different combinations of input variables were considered. The correlation coefficient (r), Root Mean Squared Error (RMSE), Adjusted R-squared, and Nash-Sutcliffe efficiency (NSE) metrics together with the Taylor diagram were used to compare the accuracy of applied models. A sensitivity analysis using the RReliefF algorithm showed that TFA, SW, and PC are the most influential parameters, whereas the ST and STT have negative influences on SRC values. Comparison analysis results indicated that the ELM model with r of 0.896, RMSE of 0.081, and NSE of 0.758 had the best performance among other employed MLAs. Also, the RF regression achieved the second rank. In conclusion, the ELM model with single-layer feedforward neural network was superior to other data-driven models; therefore, it can be applied as an efficient tool for estimating SRC values using structural input parameters.