A component architecture for artificial neural network systems

The main focus of the PhD thesis is about automating the implementation of artificial neural networks (ANNS) models by applying object/ and component technology. Though various ANN models exist, the aspect of how to provide reusable components in that domain for efficiently implementing adequate system architectures has been barely investigated. The prototypical component framework that was designed and implemented in the realm of the dissertation is compared to existing approaches for generically implementing ANN models and simulations. The application of ANNs faces difficulties such as limits of hardware resources and appropriate software solutions. How the ANN components harness parallelization on networked computers represents a contribution to the state-of-the-art in mobile code and distributed systems. The software architecture was defined in a way to facilitate the automated parallelization at the level of the inner execution of an ANN and at the level of the simulation of different ANNs at the same time, on one computing node or on different computing nodes in a distributed way. The Combinatorial Network Model (CNM) was chosen as case study for implementing parallelism at the level of the ANN structure. The improvement of one of the ANN models, namely the CNM, represents a contribution to the area of ANNs itself and to data mining. The original CNM algorithm could be significantly enhanced regarding the aspect how it deals with the search space, which results in a faster execution and less memory allocation. A sketch of research issues that result from the PhD work rounds out the thesis.