One of the problems they solve Intelligent Systems, is the estimation of functions, for which part of a finite number of data from one process and the goal is to find better role models. Among the unsupervised learning methods, optimization methods like gradient descent and conjugate gradient have been traditionally used in such problems, with advantages such as simplicity CO in the first and the speed of convergence in the latter. According to the principle of simplicity, we choose the method that is simple yet the most accurate, so that neither of the two methods can strongly considered better than another, why not simultaneously satisfy both conditions. This paper evaluates the two methods in the estimation of linear and quadratic functions and suggests improvements in its definition with the objective of proportional method that is best in terms of simplicity.