Abstract
For the purpose of automatic and effective neural network structure design, a designing method is proposed based on Genetic Algorithms (GA). Neural Network structures are encoded on chromosomes in such a way that a number of different but related network structures result by changing a parameter in the decoding process. The GA finds an optimal chromosome that provides good network structures for a family of data sets. When a change takes place in the environment around the neural network, by changing the parameter, we can obtain a new neural network suitable for the new environment without re-running the GA again.
The proposed method is described by using, as an example, a problem of finding good neural network structures for data sets with different noise magnitudes. To obtain an accurate but non-overfitted neural network for the noisy data set, we introduce a relevant fitness function, a method for noise magnitude estimation, and a systematic way to determine the control parameter value of the decoding process. By incorporating these techniques in the network structure designing method mentioned above, we obtain a neural network which has a good generalization ability for each of data sets with different noise magnitudes.
