Despite the importance and accuracy of empirical models, most of the existing models are only accurate on the collected experimental data. Adding new data, or even considering noise or variance in the data leads to loss of model accuracy. The objective of this paper is to alleviate overfitting and develop a more accurate and reliable alternative method using a decision-tree-based ensemble Machine Learning algorithm that uses a gradient boosting framework for the prediction of the ultimate shear strength of FRP-reinforced concrete beams without stirrups. To enhance the robustness of the results, make full use of training samples (without the validation set), and alleviate the randomness in selecting test samples, the K-Fold Cross Validation method is employed. Using a dataset including 205 samples, results show that the extreme gradient boosting framework (XGBoost) providing better prediction. In fact, XGBoost results have higher precision and higher generalization in comparison with the empirical equations, the current design codes of practice, Least Absolute Shrinkage and Selection Operator model (LASSO), and Random Forest model (RF).