Prediction of soaked CBR of fine-grained soils using soft computing techniques

The present research determines the effect of training data sets, correlation, and multicollinearity on the performance and overfitting of gene expression programming (GEP) and relevance vector machine (RVM) models in predicting the soaked CBR of fine-grained soil. For this purpose, one hundred and 82 training data sets have been compiled and subdivided into 50%, 60%, 70%, 80%, 90%, and 100%. In addition, 15 testing, 36 validation, and 12 laboratory-tested data sets have been compiled for trained models. The linear, polynomial, gaussian, and Laplacian kernels have been used to develop each GA and PSO optimized relevance vector machine (RVM) model, which have been trained by 50–100% training data sets. Thus, SRVM (single kernel-based) and HRVM (dual kernel-based) models have been developed and trained. The performance of models has been measured by RMSE, MAE, and R performance indicators. Based on the performance comparison, Model 21 ( R = 0.9874) & Model 39 ( R = 0.9748) of SRVM, and Model 51 ( R = 0.9606) and Model 57 ( R = 0.9701) of HRVM have been identified as better performing RVM models. However, GEP model 62 has performed ( R = 0.8847) less than RVM models. The test performance comparison shows that model 21 has outperformed models 39, 51, 57, and 62 in predicting the soaked CBR of fine-grained soil. In addition, model 21 (performance = 0.8631) has predicted soaked CBR for the validation data set better than published models. Finally, the present research concludes that model 21 (GA optimized Laplacian kernel-based SRVM model) is a robust model that can predict the soaked CBR of fine-grained soil with the least prediction error and high performance.