Ph.D. Thesis:

The method of input selection for models of multidimensional systems using artificial intelligence methods for dimensionality reduction
(in Polish)

Author: Joanna Chimiak−Opoka

Supervisor: prof. Andrzej Piegat

University: Technical University of Szczecin
Faculty of Computer Science

Readers: prof. Danuta Rutkowska
Częstochowa University of Technology
Department of Computer Engineering

prof. Jacek Łęski
Silesian University of Technology
Faculty of Automatic Control, Electronics and Computer Science
Division of Biomedical Electronics

Defence: 2002-12-10

Key-words: dimensionality reduction, preliminary input selection, approximation of functions

Description: The goal of the thesis was designing preliminary input selection method for multidimensional models which would respect real data points distribution and use artificial intelligence. The problem of input selection for multidimensional models has practical importance for construction mathematical models of real systems. The methods for input selection are used in many branches, e.g. pattern recognition, modelling, control and decision making. The thesis is devoted to approximation of functions. A new criterion of estimating input significance was proposed. According to the criterion, significance of the input depends on directional curvature of the model. An algorithm for evaluating significance of inputs is based on cellular subdivision of input space respecting to the data points distribution. Proposed approach was theoretically analysed and the algorithm was tested on sets of artificial and real data points.

Application: The algorithm for preliminary input selection for multidimensional modelling was presented. The algorithm may be applied in many problems related to approximation of functions. It can be used in technical or economical problems. Depending on number of data points from system it uses a different accuracy (limited number of cells, subspaces analysis). The algorithm enables also inputs interconnection analysis.