Validation of Human Behavioural Models

Abstract : Full validation of a model involves a number of steps. The first is to ensure that the model represents the required domain adequately (content validation). The second is to ensure that the principles underlying the model make reasonable use of current understanding of the problem space (construct validity). If the model meets these two criteria there is a requirement that the predictions of the model represent what happens in the real world to an adequate degree (predictive validity). The predictive validity of models that characterise human physiological response or low level human physical and cognitive performance can be conducted using statistical tools suitable for the analysis of interval data such as analysis of variance. When a model is developed that describes choice of course of action, an important element of human behavioural modelling, the outcomes are necessarily discrete and the volume of data available for analysis is typically smaller than desirable for validation over a broad scope. Any stream of similar decisions in a military context is likely to be aimed at maintaining the real world outcome close to a desired profile drawn up at the planning stage. In this way the process of taking decisions and monitoring their implementation is analogous to the process of tracking, embodied in such activities as driving a vehicle. The approach is applied directly to a tracking task to illustrate the interaction between a stream of decisions and outcomes and the problems of generalising the approach to more complex situations is discussed.