This work introduces an optimal performance model for predicting the unconfined compressive strength (UCS) of lime-stabilized soil using the machine (ensemble tree (ET), Gaussian process regression (GPR), and decision tree (DT), support vector machine (SVM)), and hybrid (relevance vector machine (RVM)) learning computational techniques. The conventional (non-optimized) and hybrid (genetic (GA) and particle swarm algorithm optimized (PSO)) RVM models have been developed and compared with machine learning models. For the first time, UCS of virgin cohesive soil has been used as input variable to predict the UCS of lime-stabilized soil. A database of 371 results of lime-stabilized soil has been compiled from the literature and used to create training, testing, and validation databases. Furthermore, the multicollinearity levels for each input variable, i.e., lime content, UCS of cohesive soil, and curing period, have been determined as weak for the overall database. The performance of built-in models has been measured by three new index performance metrics: the a20-index, the index of scatter (IOS), and the index of agreement (IOA). This research concludes that the weak multicollinearity of input variables affects the performance of the non-optimized RVM models. Also, the ensemble tree has performed better than SVM, DT, and GPR because it consists of the number of trees. The overall performance comparison concludes that the PSO-optimized Laplacian kernel–based RVM model UCS16 outperformed all models with higher a20-index (testing = 67.30, validation = 55.95), IOA (testing = 0.8634, validation = 0.7795), and IOS (testing = 0.2799, validation = 0.3506) and has been recognized as the optimal performance model. ANOVA, Z, and Anderson-darling tests reject the null hypothesis for the present research. The lime content influences the prediction of UCS of lime-stabilized soil. The computational cost and external validation results show the robustness of model UCS16.
