A priori information in network design

An analysis of how a priori knowledge of relative order can be applied to train a neural network effectively, is presented. In many cases only an approximate model of a system is known. The information from this model can be used to produce a more accurate one. Often this knowledge is not available or at best is inaccurate. Under these conditions, the relative order can be determined from the structure of the trained network using the rules developed here. This analysis is demonstrated with two examples