---

After data separation in the hyperspace, the focus of attention is reduced to a subspace where modeling methods can be applied.  The distinction of MasterMiner from other neural network tools is the fact that neural learninng is performed only in a subspace after data separation. A neural network module, based on a back propagation algorithm, is embedded in MasterMiner that can 

• train the net using up to L (e.g., L=250,000) samples (virtual samples are added if needed) until error is < 2%.
• estimate the target values given factor values
• display numbers of nodes for input, hidden and output layers
• order a dialogue box for predicting an individual target value for a set of given factor values
• display on-line the network error curve as the training process takes place.
• display a 2-color map showing the optimal operating zone
• target-factor sensitivity analysis.  Options are provided so that factors can be fixed at user defined points, center of good samples, max or min target point of original good samples, and predicted max or min of the predicated target value.

---

Fig-1 Estimation error curve by ANN

---

Fig-2 Parameter estimation for the neural network, after convergence.

---

Fig-3 Sensitivity analysis of target (t) and factors (a1, a2, a3, a4), the 4 figures above.

---

Fig-4 A 2-color map by Fisher method showing feature a2-a4 relationship, red points are for good samples and blue for bad samples.

© Copyright, Zaptron Systems, Inc.