Published in: Computer Physics Communications 87: 341-371 (1995)
The most popular realizations of adaptive systems are based on the neural network type of algorithms, in particular feedforward multilayered perceptrons (MLP) trained by backpropagation of error procedures. In this paper an alternative approach, based on multidimensional separable, localized functions centered at the data clusters, is proposed. Comparing with the neural networks that use delocalized transfer functions this approach allows for full control of the basins of attractors of stationary points. Slow learning procedures are replaced by explicit construction of the landscape function followed by optimization of adjustable parameters using gradient techniques or genetic algorithms. Retrieving information does not require searches in multidimensional subspaces but it is factorized into a series of one-dimensional searches. FSM (Feature Space Mapping) is applicable to learning not only from facts but also from general laws and may be treated as a fuzzy expert system (neurofuzzy system). The number of nodes (fuzzy rules) is growing as the network creates new nodes for novel data but the search times are sublinear in the number of rules or data clusters stored. Such a system may work as a universal classificator, approximator and a reasoning system. Examples of application for identification of spectra (classification), intelligent database (association) and analysis of simple electrical circuits (expert system) is given.
Paper in PDF format (280 KB)
WARNING: some pages containing graphics may not be visible using ghostscript but should print well on modern postscript printers.
Projects on similar subject and BACK to the on-line publications of W. Duch.