Prediction of UCS of fine-grained soil based on machine learning part 2: comparison between hybrid relevance vector machine and Gaussian process regression

The present research employs the models based on the relevance vector machine (RVM) approach to predict the unconfined compressive strength (UCS) of the cohesive virgin (fine-grained) soil. For this purpose, the Linear, Polynomial, Gaussian, and Laplacian kernel functions have been implemented in RVM models. Two types of RVM models have been developed: (i) single kernel function based (mentioned by SRVM) and (ii) dual kernel function-based (mentioned by DRVM). Each model has been optimized by each genetic (GA) and particle swarm optimization (PSO) algorithm. Eighty-five data points (75 training + ten testing) have been collected from the literature to train and test the SRVM and DRVM models. The data proportionality method has been used to create six training databases, i.e., 50%, 60%, 70%, 80%, 90%, and 100%, to determine the effect of the quality and quantity of training database on the performance, accuracy, and overfitting of the soft computing models. Ten conventional and three new performance parameters, i.e., a20 index, index of agreement (IOA), and index of scatter (IOS), have measured the performance of models. The present research concludes that (i) a strongly correlated pair of data points affect the performance and accuracy of the model; (ii) GA-optimized SRVM model MD119 has outperformed other SRVM and DRVM models with a20 = 100, IOA = 0.9947, and IOS = 0.0272; (iii) k-fold cross-validation test (k = 10) validates the capabilities of SRVM and DRVM models; (iv) model MD119 has predicted UCS better than GPR model MD11 developed in part 1 of this research; (v) high correlated data points increases the overfitting of the model; (vi) model MD119 has predicted UCS of lab tested soil with a confidence interval of ± 4.0%.