Digital filters for inductive inference applications
Roger D. Horn, J.D. Birdwell · 2003
A description is given of the selection of digital filters used to produce attributes of sequences of measurement data for an inductive inference algorithm. The selection criterion is the minimization of an attribute's conditional entropy of classification. The entropy function is constant almost everywhere in the parameter space, so the direct application of standard gradient search algorithms is not possible. A parameterized continuous and differentiable approximation to the entropy function is introduced and used to generate a family of minimization solutions. The set of local minima in this family of solutions converges to the local minima of the entropy function. An illustration of the selection method applied to a synthesized data set is presented.>