Floating Gaussian Mapping: a New Model of Adaptive Systems

Wlodzislaw Duch

UMK-KMK-TR-5/93
Neural Network World 4 (1994) 645-654

Adaptive systems are usually realized by the neural network type of
algorithms. In this contribution an alternative approach, based on products
of Gaussian factors centered at the data points, and acting as feature
detectors, is proposed.  Comparing with the feedforward neural networks with
backpropagation learning is much faster because explicit construction of the
approximation to the desired mapping is performed, with fine tuning via
subsequent adaptation of the shapes and positions of the feature detectors.
Comparing with the recurrent feedback networks this approach allows for full
control of the positions and sizes of the basins of attractors of the
stationary points. Retrieving information is factorized into a series of
one-dimensional searches. The FGM (Floating Gaussian Mapping) model is
applicable to learning not only from examples but also from general laws. It
may serve as a model of associative memory or as a fuzzy expert system.
Examples of application include identification of spectra and intelligent
databases (associative memory type), analysis of simple electrical circuits
(expert system type), and classification problems (two-spirals problem).