Discrete State Models of Information Processing

Abstract

Information processing (IP) models of human behavior assume mental operations analogous to those of computers, for example, storing, maintaining, comparing, and retrieving information. Discrete state IP models assume that observable responses in an experimental task are generated probabilistically from one of a finite, and usually small, set of latent (unobservable) cognitive states. This article considers parametric discrete state IP models for categorical data, that is, probability models for situations where each response falls into one of a finite set of non-overlapping response categories. Such models were popular in the early days of mathematical psychology, especially in the areas of signal detection and association learning. The signal detection models were based on threshold assumptions and the learning models were based on Markov chains. In the 1980s, a class of discrete state models called multinomial processing tree (MPT) models was formulated. The class not only includes the discrete state signal detection and Markov learning models, it also includes many other discrete state IP models. Further, many new MPT models have been formulated in various standard experimental paradigms in cognitive psychology. A complete statistical inference has been worked out for MPT models which makes them useful tools to measure unobservable cognitive processes, and test cognitive theories in particular experimental settings.