A Brief History of Normative Models of Human Behavior

Modeling human behavior is a vast subject and one short paper cannot come close to doing justice to the topic. There are many types of models and many reasons for modeling. One cannot pinpoint when modeling began, because that would beg the question of what is meant by modeling. In the most general sense “to model” means to represent objects or events with words, symbols, images, song, dance, or other means of expression. Naturally that goes back to the dawn of human intelligence. The great majority of such representations that we humans make are qualitative, taking the form of conjectures, theories, descriptions, or stories — typically expressed in words — of law, politics, morality, religion, physics, art, or how to do whatever. In response to the kind invitation to lecture here on models I chose to narrow the scope to quantitative models and narrow still further to normative quantitative models of human behavior. Also, I will omit the large class of models of learning, with which I have little experience. That still leaves a wide terrain that I cannot possibly encompass, so I will talk about models that seem to have been both popular and widely applicable. First some distinctions must be made: A descriptive model uses mathematics to characterize after-the-fact some set of data from experimental measurements in the form of a relation between independent and dependent variables. A predictive model seeks to predict the results of future measurements where the independent variables are similar but actual values are different from those used to develop the model. A normative model states what the dependent variables should be, given both the independent variables, and some assumptions about the mechanism of cause and effect. Seldom does the normative model fit the observed human behavioral events precisely, and one tries to get a better fit by adjusting the model parameters. To the extent that the fit is good, one can infer that the “structural mechanism” assumptions of the normative model are appropriate to characterize the cause-effect relation underlying human performance, and in this case use the normative model to predict human performance. Taxonomies, such as classes of errors, Rasmussen’s distinctions between skill, rule and knowledge, and level of automation are not quantitative normative models. Nor are theories of naturalistic decision-making quantitative normative models. Nor are the techniques of statistical analysis, such as correlation, null hypothesis tests or analysis of variance. Correlation and curve fitting are common quantitative descriptive models, and by extrapolation can become predictive models. But since they involve no cause and effect structure or “mechanism” they are not normative models. This leaves out some elegant techniques such as factor analysis and the Brunswik Lens model (which embodies a structure of stages of correlation and is being actively pursued as a descriptive model of judgment)(Kirlik, 2005). Curiously, it was the engineering development of physical weapons systems during World War 2 that led to engineering theories, that eventually, after the war, were adopted by psychologists and applied to human behavior. The same theories that underpinned the development of control systems for aircraft, ships, and guns, and that formed the basis for radar and sonar detection of enemy aircraft, were spotted by psychologists as having analogies to human control and signal detection. The same probability-cost decision theories that so-called operations research mathematicians used to develop military tactics were seen by psychologists to apply to individual human decision making. An exception was information theory, which mostly came out of Bell Laboratories well after that war. More recently, rule-based (both crisp and fuzzy) models evolved from the development of computers and software, and enabled “production rules” to be executed. Below I discuss four classes of quantitative normative models, in more or less the order in which they were developed: (1) feedback control models, (2) expected value decision models, (3) information communication models, and (4) rule-based models. Salient qualitative models begging for normative quantification are then mentioned.