Practical Neural Network Recipies in C++

7

Eluding Local Minima I:

Simulated Annealing

This chapter will discuss a simple yet effective method for avoiding

local minima, as well as escaping from them if necessary. The

algorithm discussed is a variation of traditional simulated annealing,

optimized for the error surfaces encountered in neural network

learning. The code given here will be an implementation of that

algorithm written for general function minimization, and is primarily

for education. The specific code for learning in multiple-layer

networks can be found in the Appendix on page 429.