Multi layer perception Vs SVM / Random forest

What can be the reason while Multi layer perception is not converging but SVM or Random forest is performing well on same dataset? My Data is 120 features having numeric values between -1 to 1 and need to predict any of 5 classes. Note: I am not aware about what data this is?

SVM accuracy 91% Random forest 86% 2 layer MLP acc 74% after 3500 epoch