Abstract
This study presents a comparative analysis of individual and ensemble learning algorithms (ELAs) to predict the compressive strength (CS) and flexural strength (FS) of plastic concrete. Multilayer perceptron neuron network (MLPNN), Support vector machine (SVM), random forest (RF), and decision tree (DT) were used as base learners, which were then combined with bagging and Adaboost methods to improve the predictive performance. In addition, gene expression programming (GEP) was used to develop computational equations that can be used to predict the CS and FS of plastic concrete. An extensive database containing 357 and 125 data points was obtained from the literature, and the eight most impactful ingredients were used in the model’s development. The accuracy of all models was assessed using several statistical measures, including an error matrix, Akaike information criterion (AIC), K-fold cross-validation, and other external validation equations. Furthermore, sensitivity and SHAP analysis were performed to evaluate input variables' relative significance and impact on the anticipated CS and FS. Based on statistical measures and other validation criteria, GEP outpaces all other individual models, whereas, in ELAs, the SVR ensemble with Adaboost and RF modified with the Bagging technique demonstrated superior performance. SHapley Additive exPlanations (SHAP) and sensitivity analysis reveal that plastic, cement, water, and the age of the specimens have the highest influence, while superplasticizer has the lowest impact, which is consistent with experimental studies. Moreover, GUI and GEP-based simple mathematical correlation can enhance the practical scope of this study and be an effective tool for the pre-mix design of plastic concrete.