| Summary: | Application development of computationally intelligent systems is a demanding task. This thesis argues that this task can be facilitated through proper software technology. We propose a software architecture through which a knowledge discovery process can be embedded to final applications without large rebuilding. The thesis describes Neural Data Analysis (NDA) environment where the architecture has been implemented and tested. It is a development environment for knowledge discovery applications. NDA applies Self-Organizing Maps (SOM) and its tree-structured variant as basic methods. It also provides a number of data modeling methods as well as a large set of generic data processing and visualization methods. The thesis gives a comprehensive guideline to SOM based knowledge discovery process that is demonstrated through NDA. Within the overall framework, three topics are studied in more detail. First, we propose a new methodology for data imputation. This methodology applies Self-organising maps to the problem of poor quality data that is often one bottleneck in application development. Secondly, we introduce a framework for SOM based visualisation. This framework integrates the SOM based visualisation with data and knowledge visualisation. Thirdly, we describe the software architecture of NDA and some detailed solutions in the implementation. Finally, the software solution is evaluated through a number of real world cases. These cases verify that our idea for building embedded intelligent systems works in practice. The NDA software has helped us to apply the neural network technology in very different domains from logistics planning to process monitoring and data analysis in behaviour sciences.
|
|---|