The INNC consists of a multi-layer feed-forward
perceptron (MLP) with sigmoidal nodes in its hidden and
output layers. This MLP is trained with classic error backpropagation
(BP) learning rule [I]. An iterative, closed
loop scheme is used in the sense that the parameter settings
produced by the INNC when it is first executed (iteration
1) are used to generate wire bonds. The resulting BS and
SSD measures along with the value of the setting
themselves are input to the INNC to produce another round
of settings (iteration 2). This iterative procedure is
continued until the BS and SSD measures reach a steady
state value close to the desired values