Statistical mixture decomposition as a method for type analysis of learning curves.

Abstract

A method for type analysis of learning curves, based on the statistical mixture decomposition, is described. Some critical points in current data-analytic techniques are discussed. The mathematical rationale of the new method is outlined in a brief sketch. The possibilities of the method are documented by two examples. In the first study, done on simulated lata of a known structure (N = 200, 2 classes), it was possible to distinguish, with an average performance of 82%, between two types, and to reproduce their original curves. In the second study data from experiments in classical eye-lid conditioning in man were analysed (N = 80). The decomposition procedure resulted into the classification into four groups, with pronounced inter-class differences in the course of respective learning curves. The variety of class curves ranges from a group with only few CRs (C1, N = 26), through a group with an initial increase and final decrease in CR frequency (C2, N = 16), a group with an apparently biphasic course of CR frequency (C3, N = 20), to a group with a rapid increase of CR and then stable course of CR frequency (C4, N = 18). The results are consistent with earlier findings concerning the existence of distinct types of learning curves. The problem of interpretation is briefly discussed. The method can be applied principally to any problems, where different types of time development trends of an alternative response are to be distinguished.