Local Structure Optimization in Evolutionary Generated Neural Network Architectures
Marko V. Borst · 1994
This thesis is an extension of the work done by Boers and Kuiper [Boers92]. In their master thesis they proposed a method to produce good modular artificial neural network structures. It was argued that modular artificial neural networks have a better performance than non-modular networks. Based on the natural process that resulted in our brain, they introduced a genetic algorithm to imitate evolution and used L-systems to model the kind of recipes nature uses in biological growth. In this research the objective was to find a local optimization method for modular neural network structures. Such a method should change the structure of the network. Since adding an additional unit to an already existing module was shown to be the most `local' way of changing the structure of a network a couple of criteria was developed to decide if and which module of the network needs an additional unit. Besides that, four, on constructive algorithms inspired, methods to install the new unit in the modu...