How to Design Neural Networks Using Genetic Algorithms
- 1). Describe the structure of the neural network you want to have when the design process is complete. This description will include the number of inputs, the number of outputs and a "design string" that defines the internal structure of the neural network. A typical neural network will have three layers: the input layer, the hidden layer and the output layer. The design string could define the number of elements in the hidden layer and which elements in each layer are connected to which elements in the adjacent layers.
- 2). Write the program that controls the genetic algorithm. The algorithm does to the design strings what nature does to chromosomes. Starting with a random population of design strings, the genetic algorithm will select pairs of strings to blend. From each blending, a neural network will be constructed then tested. The strings that were used to create neural networks that preform well will be kept for the next generation. Design strings that were used to construct neural networks that performed poorly will be deleted from the population. After a few generations, some of the design strings will produce neural networks that are good enough to do the job you are interested in.
- 3). Try and try again. Like the natural phenomena you are simulating, using the genetic algorithm to design neural networks is not an exact science. A lot depends on luck. Sometimes you can run such a program 10 times and fail each time. On the 11th trial, you could succeed. Ever time you start the project, you start with a random population, so every time the outcome will be slightly different. Of course if your layout of the design string focuses on the wrong issues and you are changing the wrong parameters of the neural network, the process will never succeed. Even in nature, some evolutionary lines are dead ends.