The purpose of this research is to develop a self-organizing network. The network initially has only input neurons.
During the training process, neurons are selected from a pool of candidates and added to the hidden layers. These are
the polynomial and hyperbolic functions which include eleven polynomial functions. During the final stage, the proposed
system performs analysis for each model based on five error predicating measures including Maximum error, RMSE,
MSE, MAE and MAPE. Through experiments we found that all the huge databases have fixed behaviors, the best model
generated by linear of three variable functions and worst model generated by cubic of one variable or quadratic function
related to the polynomial models. The best model generated by tanh of one variable function and the worst model
generated by more than one other function is related to the hyperbolic models. Most of the small databases have
unstable behavior, the best model is generated by linear of three and two variables or quadratic of two variable functions
and the worst model is generated by cubic of one variable or quadratic functions related to the polynomial models.
Finally, the best model generated by inverse of tanh of one variable function and the worst model generated by sinh or
cosh functions are related to the hyperbolic models.