The use of simulated annealing for the best fitting L2 regression model

In many problems we need to fit a regression model. For using effectively different adjustment criteria one of the problems to be solved is how to select the best regression model. Commonly the method of Least Squares is used assuming the normality of the errors. In this paper we suggest solving this problem by using a Simulated Annealing based heuristics. The method looks for the diminution of the residual sum of squares, using it as objective function. Algorithms are developed and an evaluation of the proposals is made by analyzing classic examples from text books. The behavior of the algorithms seems to be adequate because they identify the best fitted models.