The strength of concrete elements under shear is a complex phenomenon, which is induced by several effective variables and governing mechanisms. Thus, each parameter’s importance depends on the values of the effective parameters and the governing mechanism. In addition, the new concrete types, including lightweight concrete and fibered concrete, add to the complexity, which is why machine learning (ML) techniques are ideal to simulate this behavior due to their ability to handle fuzzy, inaccurate, and even incomplete data. Thus, this study aims to predict the shear strength of both normal-weight and light-weight concrete beams using three well-known machine learning approaches, namely evolutionary polynomial regression (EPR), artificial neural network (ANN) and genetic programming (GP). The methodology started with collecting a dataset of about 1700 shear test results and dividing it into training and testing subsets. Then, the three considered (ML) approaches were trained using the training subset to develop three predictive models. The prediction accuracy of each developed model was evaluated using the testing subset. Finally, the accuracies of the developed models were compared with the current international design codes (ACI, EC2 & JSCE) to evaluate the success of this research in terms of enhancing the prediction accuracy. The results showed that the prediction accuracies of the developed models were 68%, 83% & 76.5% for GP, ANN & EPR, respectively, and 56%, 40% & 62% for ACI, EC2 & JSCE, in that order. Hence, the results indicated that the accuracy of the worst (ML) model is better than those of design codes, and the ANN model is the most accurate one.