Dynamic learning of decision trees to acquire knowledge for thediagnosis of dynamic systems

摘要

Inductive learning methods became an essential part of knowledgeacquisition tools for diagnosis components of expert systems. If dealingwith dynamic systems, not only the current state variables are relevantto the classification process of technical system's states but alsotheir histories. These requirements result in the fact, that the statespace seems to be not finite. Up to now, the way out was to base theinduction process on extended state vectors built by subsequentlyconcatenating original state vectors. On the one hand, this enlarges theamount of information to be processed enormously. On the other hand, itmay be still insufficient, if the depth of history required is not knownin advance. This article presents an approach, where the access to thehistorical values of process data immediately depends on the givenlearning data set. This dynamic access is called data-driven access tohistorical values. The use of the presented strategy results in anoptimum of efficiency without prior restrictions of the hypothesisspace. Consequently, the presented approach is able to generateclassifiers in the form of decision trees in a very effective way