Affordable Access

Maturing a network structure for rule extraction

STW Technology Foundation
Publication Date
  • Computer Science


Microsoft Word - WPM$7130.doc Maturing a network structure for rule extraction Berend Jan vanderZwaag1, Suleyman Malki2, and Lambert Spaanenburg2 1 . University of Twente, 2. Lund University Dept. of Electrical Engineering Dept. of Information Technology P.O.Box 217, NL-7500 AE Enschede P.O.Box 118, SE-221 00 Lund T +31.53.489 2780; E [email protected] Key words: Prediction, redundancy, industrial control, logistic process, fleet management. Aims and Scope. Neural networks are often used to adapt a generic control formulation to the process at hand. The formal model or expert rule base provides a formulation of the domain problem but a further tuning to the physical reality tends to be cumbersome, error-prone and difficult to maintain. As the process evolves in time, additional model adjustments are often required and the search for alternative techniques becomes even more pressing. In a layered control hierarchy, the generic formulation can be shaped to provide a strategy that is subsequently detailed by a set of neural networks, each reflecting part of the physical reality. This decoupling of domain and process specific effects works reasonably well, but it remains hard to guarantee the integrity of the neural activations. For proper quality monitoring, it is required that the neural knowledge can be extracted isomorphic to the original generic formulation. To achieve this goal, a novel means has to be devised to optimally match sensitivity to robustness in neural classification & clustering. Approach. The classical learning approach for neural networks is based on gradient optimization in the error space. Its operation can be explained from the Gravitational Field analogon as known from physics: by weight adaptation through the learning rule an example is moved along the energy field to the attracting cluster. However, between two clusters, there will be a point where the attraction is zero, which is the major cause for the indeterminism a

There are no comments yet on this publication. Be the first to share your thoughts.