C++ Neural Networks and Fuzzy Logic: Preface

Other Experiments to Try

There are many other experiments you could try in order to get a better feel for how to train and use a

backpropagation neural network.

•  You could use the ASCII 8−bit code to represent each character, and try to train the network. You

could also code all of the alphabetic characters and see if it’s possible to distinguish all of them.

•  You can garble a character, to see if you still get the correct output.

•  You could try changing the size of the middle layer, and see the effect on training time and

generalization ability.

•  You could change the tolerance setting to see the difference between an overtrained and

undertrained network in generalization capability. That is, given a foreign pattern, is the network able

to find the closest match and use that particular category, or does it arrive at a new category

altogether?

We will return to the same example after enhancing the simulator with momentum and noise addition

capability.

C++ Neural Networks and Fuzzy Logic:Preface

Chapter 13 Backpropagation II

262

---