TABLE OF CONTENTS Back-propagation [13] (BP) neural nets as automatic correlative model-builders were qualitatively described in Chapter 3. That chapter discussed their architecture, the components forming the building blocks from which they are constructed, and how they work. These nets derive their name from the training process during which the weights, modifying the strength of the connections between their neurons, are adjusted by back propagating errors through the net structure. The BP learning process enables the nets to capture essential correlative relationships between independent variables governing the phenomena and the dependent outcomes, directly from operating data.
[13]Such nets are also known as "backprop" nets.
In a seeming contradiction of terms, BP nets are members of a family of "feedforward" neural nets. Information flows in only one (forward) direction [14] when the nets are in the predictive mode; whereas errors are propagated backwards while the nets are being trained.
The BP neural net is trained on sets of data comprising of input and corresponding target values. The input data is an n X p matrix where n is the number of input variables and p the number of different...
