Machine learning is one of the most used digital tools today. For genomic selection in plant breeding, this is no different. Several studies have already demonstrated the high potential of this tool. This study aimed to evaluate and compare the predictive performance of various machine learning methods and conventional method (G-BLUP) through GWS. For this, an F2 population formed by 1000 individuals was simulated and genotyped with 4010 SNP markers. Twelve traits were simulated from a model considering epistatic effect, with QTL numbers ranging from eight to 480 and heritability of 0.5 or 0.8. The learning methods used were: multilayer perceptron (MLP), radial basis function (RBF), decision trees (DT), bagging (BA), random forest (RF), boosting (BO) and multivariate adaptive regression splines (MARS) considering additive and non-additive model. Then, the performance of the learning methods was compared to the G-BLUP method. The results obtained were able to demonstrate the strong effect of heritability and the increase in the number of QTL on the values of selective accuracy (correlation square - R²) and predictive accuracy (root mean squared error – RMSE). The methods based on MARS presented higher R² values for traits with 8 QTLs, for both heritability scenarios. When increasing the number of QTLs (40 or more) it was observed that the RBF presented the highest R² values, followed by G-BLUP, RF, BA, and BO, and always above the general average for the traits with 50% heritability. For traits with 240 and 480 QTLs, the non-additive MARS methods presented high R² values, considering the standard error only smaller than the G-BLUP. The methods based on neural networks (MLP and RBF), similarly to G-BLUP, showed substantial improvement the higher the QTLs. The BO method demonstrated greater sensitivity to heritability and showed the best results when the scenarios were of greater heritability and 40 or more QTLs. Machine learning methods are powerful tools for predicting genetic values in traits with control of epistatic effect, to different degrees of heritability, and different numbers of control genes and should be considered in genomic selection analyses.