In this study, it is aimed to evaluate COD removal performance of Classical-Fenton and Photo-Fenton Processes from cosmetic wastewater by different prediction models. Besides Response Surface Methodology (RSM), three neural networks were used to more reliably and effectively predict the behavior of dependent variable at different values of relevant parameters. These neural networks; multi-layer perceptron trained by Levenberg-Marquardt (MLP-LM); multi-layer perceptron and single multiplicative neuron model trained by particle swarm optimization algorithm (MLP-PSO; SMN-PSO). H2O2 doses, Fe(II) doses, and H2O2/Fe(II) rates were independent variables of prediction models to optimize both processes in batch reactors. The generated predictions for whole data set were compared with each other. The prediction performances of models were evaluated by RMSE and MAPE error criteria. Regression analysis was also applied to determine the performance of the best model. The results obtained from all prediction tools showed that the model produces the best predictive results in almost all cases is SMN-PSO model in terms of both criteria. In addition, the genetic algorithm was utilized for SMN-PSO model results to find the optimum values of the study. Thus, without the need to perform many different experiments, the optimum parameter values can be determined to get maximum removal ratios.