Practical Neural Network Recipies in C++

330

Chapter 18

the Kohonen network, not presented in detail here, learn in ways that

exhibit a striking resemblance to creature learning.

Input Normalization

One of the most serious impediments to widespread use of the

Kohonen network is the fact that its inputs are subject to serious

restrictions. Ideally, they should he within symmetric boimds, usually

taken to be [-1, 1]. The length of the input vector must be the same

for all training and test cases. This length is typically chosen to be 1.

Finally, for best performance, each input should be able to cover the

majority of its range. For example, if some of the input variables are

physicall ...