Abstract

Artificial neural networks used to be considered only as a machine that learns using small modifications of internal parameters. Now this is changing. Such learning method do not allow to generate big neural networks for solving real world problems. This thesis defends the following three points: ffl The key word to go out of that dead-end is "modularity". ffl The tool that can generate modular neural networks is cellular encoding. ffl The optimization algorithm adapted to the search of cellular codes is the genetic algorithm. The first point is now a common idea. A modular neural network means a neural network that is made of several sub-networks, arranged in a hierarchical way. For example, the same subnetwork can be repeated. This thesis encompasses two parts. The first part demonstrates the second point. Cellular encoding is presented as a machine language for neural networks, with a theoretical basis (it is a parallel graph grammar that checks a number of properties) and a comp...

Citations