A new approach to the formulation and testing of learning models

It is argued that current attempts to model human learning behavior commonly fail on one of two counts: either the model assumptions are artificially restricted so as to permit the application of mathematical techniques in deriving their consequences, or else the required complex assumptions are imbedded in computer programs whose technical details obscure the theoretical content of the model. The first failing is characteristic of so-called mathematical models of learning, while the second is characteristic of computer simulation models. An approach to model building which avoids both these failings is presented under the title of a black-box theory of learning. This method permits the statement of assumptions of any desired complexity in a language which clearly exhibits their theoretical content.Section II of the paper is devoted to the problem of testing and comparing alternative learning theories. The policy advocated is to abandon attempts at hypothesis testing. It is argued that, in general, we not only lack sufficient data and sufficiently powerful techniques to test hypotheses, but that the truth of a model is not really the issue of basic interest. A given model may be true in the sense that on the basis of available evidence we cannot statistically reject it, but not interesting in the sense that it provides little information about the processes underlying behavior. Rather, we should accept or reject models on the basis of how much information they provide about the way in which subjects respond to environmental structure. This attitude toward model testing is made precise by introducing a formal measure of the information content of a model. Finally, it is argued that the statistical concept of degrees-of-freedom is misleading when used in the context of model testing and should be replaced by a measure of the information absorbed from the data in estimating parameters.I would like to express my indebtedness to Ernest W. Adams and Julian Feldman for their guidance and encouragement of the research reported here.Much of the research reported here was conducted at the University of California, Berkeley, and incorporated in a Ph.D. dissertation in Logic and the Methodology of Science [15]. The research was supported in part by the Computer Institute for Social Science Research, Michigan State University.