Summary

Get the entire book!

In this chapter you learned about the simulated annealing algorithm. The simulated annealing algorithm is based on the actual process of annealing. The annealing process says that a metal that is allowed to cool more slowly will form more consistent, and therefore strong, crystal structures. This is because higher temperatures result in higher energy levels for the atoms that make up the metal. At the higher energy levels the atoms that make up the metal are allowed a much greater freedom of movement. As the metal cools, this freedom of movement is curtailed. This allows the atoms to settle into consistent crystal patterns.

The process of simulated annealing is very similar to the actual annealing process. A series of input values are presented to the simulated annealing algorithm. The simulated annealing algorithm wants to optimize these input values so than an arbitrary equation can be minimized. Examples of the equation that is to be minimized might be the error function for a neural network, or the distance that the traveling salesman travels. The input values, which drive the simulated annealing algorithm, could be such things as the weight matrix of a neural network or the current route of cities that the traveling salesman is traveling.

To see a relatively simple example of how to use simulated annealing, this chapter once again turned to the traveling salesman problem. The traveling salesman problem was also used in Chapter 8 in conjunction with genetic algorithms. Reusing the traveling salesman problem allows you to easily compare the performance of genetic algorithms, when compared to simulated annealing.

In the next chapter we will see how to apply both genetic algorithms and simulated annealing to neural network training. We will develop an application that allows either of these training methods to be used.