Stages of the Dynamic Pattern Recognition Process

Abstract

As stated in the previous chapter, the design procedure of a dynamic classifier must couple the results of the monitoring process with a mechanism for updating the classifier according to detected changes in the cluster structure. Different approaches used for establishing the monitoring process are usually based on the observation and the analysis of some characteristic values describing the performance of a classifier or the cluster structure. The temporal change of these characteristics points to some structural changes in the underlying cluster structure and to the need for adapting a classifier. According to the nature of the monitored characteristics, one can distinguish between statistical and fuzzy techniques for the monitoring process, the most important of which will be discussed in Section 3.1. In order to preserve the performance of a dynamic classifier over time, it must be adapted to temporal changes detected by the monitoring process. The updating strategies of a dynamic classifier presented in Section 3.2 depend on the type of temporal changes in the cluster structure (gradual or abrupt) and can require either the adjustment of classifier parameters or complete re-learning of a classifier. As will be shown, the most flexible adaptation law of a dynamic classifier must combine both these techniques and include additional mechanisms supporting the intelligent design of a dynamic classifier. An updating strategy represents a crucial component of a dynamic patter recognition system since it determines an adaptive capacity of a dynamic classifier and its ability to follow temporal changes in the cluster structure.