Inter-generational architecture adaptation of neural networks

We present a neural network design method based on an inter-generational evolution process. This method is essentially a learning scheme based on a repetitive reconstruction of network architectures, in which the life span of a network is short, but offsprings can inherit the properties which were learned by its parents. At the beginning of each short-generation, the network architectures represented by chromosomes are reproduced by simple genetic operators. Then the network will be trained for short epochs, based on the conditional class entropy criteria. The simulation results showed that our method is effective in finding the fully optimized network for the given tasks.