Predicting Undrained Shear Strength Of Soil From Cone Penetration Test Data Applying Optimized RBF Approaches

In this study, two hybrid metaheuristic models are developed to discover the applicability of using an optimized radial basis function neural network to predict the undrained shear strength (USS) from cone penetration test (CPT) records. To this aim, genetic algorithm (GA) and grey wolf optimization (GWO) were integrated with radial basis function (RBFNN), named GA-RBFNN and GWO-RBFNN, to determine the optimal determinative parameters of RBFNN. Models trained with five input variables named cone tip resistance, sleeve friction, liquid limit, plastic limit, and overburden weight. To assess the precision of developed models, four statistical performance indices (R2, RMSE, MAE, and PI) were considered. The results demonstrate powerful potential in the learning section as well as approximating in the testing phase. It means that the correlation between observed and predicted USS from hybrid models is acceptable to represent the high accuracy in the training and approximating process. Concerning developed RBFNN models, it can be observed that GWO-RBFNN has better performance than GA-RBFNN based on evaluating the performance evaluator indices. Based on the outline of this paper aimed at finding RBFNN optimal parameters with optimization algorithms, GWO is more capable and rapid than GA for determining these parameters. Therefore, GWO-RBFNN outperforms the GA-RBFNN model with a total ranking of 16, consequently being recognized as the proposed model.