Towards neuro-linguistic modeling: Constraints for optimization of membership functions (Q1125359)

The paper deals with fuzzy systems. It gives a survey of the field of semantic properties of membership functions. The author discusses and illustrates the fact that in fuzzy-neural systems with unconstrained optimization schemes like back-propagation, the resulting membership functions often do not represent human interpretable linguistic terms. He investigates constraints within an optimization scheme such that the membership functions are human interpretable. Especially two constraints are important: optimal interfaces (for coverage) and sigma-count of an internal variable (for distinguishability). The importance of these proposed constraints is illustrated by an example of an approximation task.