Service Communautaire d'Information sur la Recherche et le Développement - CORDIS

Computational framework for clustering and classification of static patterns, time series prediction and classification, modelling and identification of dynamic systems

We have developed a novel and mathematically sound computational framework, namely fuzzy lattice learning framework (FLL) or FL-framework, within which we have devised various intelligent schemes for learning and decision making, especially for clustering and classification. One of them is the FLL scheme whose neural implementation, namely fuzzy lattice neural network (FLN), enhances the scope of Grossberg's adaptive resonance theory (ART).

The main advantages of the FLL/FLN are:
it learns fast, requires no tuning, and can deal jointly with disparate data such as numbers, waveforms, propositional variables, fuzzy sets, images, etc;
it can handle input-intervals instead of handling solely individual input-points;
it is easily implementable in hardware, while its neural implementation allows for fast parallel processing;
it is reliable and noise resistant.
Application of the FLL/FLN to different benchmark data clustering and classification problems showed its outstanding performance compared to various other schemes.

The predictive modular decision systems (PREMODS) is a class of algorithms applicable to time series classification and prediction and nonlinear systems identification. PREMODS consists of a number of user-defined predictive modules (linear, neural, etc) and a decision module which compares and combines the predictive module outputs using Bayesian, fuzzy, learning automata and other algorithms.

PREMODS is characterized by:
modularity (ie its components can be removed and independently retrained);
recursive, on-line operation;
high robustness to noise;
good scaling properties (the complexity of the algorithm scales as a linear function of the size of the model);
it can be handled by non-technical people.
PREMODS has been applied with good results to phoneme and enzyme classification, forecasting of electric loads and alternating current (AC) motor parameter estimation. Other potential applications include industrial fault diagnosis, medical diagnosis, prediction of financial data, structural identification, etc.

Reported by

Aristotle University of Thessaloniki
54006 Thessaloniki
Greece