A comparative assessment of ANN, GP and SVM in predicting indirect tensile strength of modified bituminous mixes

The present study assesses the predictive potential of several machine learning models namely Gaussian process (GP) and support vector machines (SVM) based upon normalized polynomial Kernel function (NPKF), Pearson VII universal Kernel function (PUK) and radial basis Kernel function (RBF) and artificial neural network (ANN) based models in predicting indirect tensile strength (ITS) of modified bituminous mixes. Out of 102 experimental data set values, 71 values were selected for training the models whereas, 31 values were used for the testing the models. The proposed models uses bitumen penetration, optimum bitumen content, type of anti-stripping agent, percentage of anti-stripping agent, percentage of polypropylene fibre, Marshall stability and conditioning type to predict the ITS of modified bituminous mixes. Out of the applied models, SVM-RBF based model is found to outclass the others as established by statistical indices with highest values of coefficient of correlation, R (0.9326), Nash Sutcliffe efficiency, NSE (0.8588) and Willmott index, WI (0.957) and less values of mean absolute error, MAE (26.2769), root mean square error, RMSE (32.7389), scattering index, SI (0.063) and percent bias, PBIAS (−0.9386%) for testing stage. The findings also demonstrate that RBF Kernel function performs better than other applied Kernel functions in case of both SVM and GP models. The sensitivity analysis propose that conditioning type is the leading influencing variable affecting the prediction of ITS, using this data set.