Practical Neural Network Recipies in C++

Eluding Local

Minima I

119

Simulated annealing can be performed in optimization by

randomly perturbing the independent variables (weights in the case of

a neural network) and keeping track of the best (lowest error) function

value for each randomized set of variables.

A

relatively high standard

deviation for the random number generator is used at first. After

many tries, the set that produced the best function value is designated

to be the center about which perturbation will take place for the next

temperature. The temperature (standard deviation of the random

number generator) is then reduced, and new tries done. (Note that

this is slightl ...