A framework for an automated neural network designer using evolutionary algorithms
Hlynka, Markian D.
MetadataShow full item record
One of the major stumbling blocks of neural networks is the difficulty of designing the networks. Networks must be created by experts who understand both the problem domain and the process of developing neural networks. For complex problems, the process, even for experts, can be an intuitive rather than ratiocinative process. Evolutionary and genetic algorithms are a robust, probabilistic search strategy that excel in large, complex problem spaces. Research involving the application of evolutionary algorithms to neural networks for purposes of both training and selection of an optimal network has been carried out. The focus of such research, however, has been to generate an optimal network of a given structure. No generic framework exists which allows for the automation of the network creation process--the selection and design of the architecture--for a particular problem. This thesis is concerned with the design of such an evolutionary framework. The system is subsequently evaluated with backpropagation networks on an unknown data set. A new method of evolution, probabilistic Lamarkian Learning transfer, appears to produce the desired results.