Practical Neural Network Recipies in C++

Interpreting Weights

191

The author has had some success with a genetic selection

technique. Randomly choose a subset of the inputs, train a network

with this subset, and save the results. Repeat this for as many

random subsets as time and memory allow. Then use the genetic

algorithms previously described (page 135) to select pairs of superior

subsets. For each such pair, generate a pair of child subsets by

selecting inputs that were present in the parent subsets. As genera-

tions pass, useless inputs will tend to be weeded out, while valuable

inputs will appear more and more frequently. The behavior of this

method, and ways to optimize its