Estimation of modified expansive soil CBR with multivariate adaptive regression splines, random forest and gradient boosting machine

Construction of flexible pavement on expansive soil subgrade relies on the safe determination of California bearing ratio (CBR) value, a critical component in flexible pavement design. However, its determination, particularly in the laboratory, often consumes sufficient man-hours. This necessitated the urgency to explore alternative procedures, such as the development of reliable models to estimate the CBR of subgrade especially modified expansive soil subgrade. In the present study, three machines learning models, which are multivariate adaptive regression splines (MARS), random forest and gradient boosting machine models, were developed to predict the CBR of expansive soil subgrade blended with sawdust ash, ordinary Portland cement and quarry dust. The performance of the models was evaluated using several error indices, and the results obtained from the evaluation showed that the random forest model has superior predictive ability when compared with the MARS and gradient boosting machine models. Specifically, the R² values for the training and testing data for the random forest model, which were, respectively, 0.84829 and 0.75282, clearly indicated that the random forest model has good predictive ability and possesses greater generalization ability than the other developed models in this study.