Study of Weight Importance in Neural Networks Working with Colineal Variables in Regression Problems

This paper presents a new method that can be used for symbolic knowledge extraction from neural networks, once they have been trained with the desired performance. Weights are the basis for this method. This method allows knowledge extraction from neural networks with continuous inputs and outputs, more precisely in problems dealing with the general linear regression model where exists multicolineality among the input and output. An example of the application is showed by comparison of the results between the regression and the neural networks results, concernig the estimation that gasoline yields from crudes. This example is based on detecting the most important variables when there exists multicolineality.