Automatic rule set modeling and change detection of systems

Abstract

An incremental inductive learning and change detection method is proposed which generates rule sets that contain general rules underling the observed data and detects changes in them. Unlike most of other algorithms, the method is an incremental algorithm that generates the rule set in the course of data observation. The method is motivated by a demand for an automatic procedure for quick modeling of systems subject to change and detecting the changes. The method starts with a most general rule and performs specialization of the rules when a conflicting piece of data is observed. Also the unification of over-specialized rules is done to enhance the generality. The change detection is based on a statistical hypothesis test of a change in error rate of the rule set assuming randomness in the data presentation order. Simulation studies have been carried out on a toy problem, where a cat observes a mouse and builds a rule set describing the mouse's behavior while the mouse disappears and a new one comes in. The results show that the method can yield a general rule set and that it can successfully detect the change of the mice.