SNIF-ACT: The User-Tracing Architecture

A possible method of developing the type of computational cognitive-perceptual models we need is to make models with stochastic parameters and run them Monte Carlo style, comparing their aggregated output against aggregated groups of users. This is a good method for exploring the effect of various inventions at the browser level or at the level of design of websites or web pages, but it averages out information related to the cognitive and perceptual mechanism at work needed to develop and validate these models. Here we develop an alternative method aimed at extracting and validating information and an individual user level, depicted schematically in the following figure.

Tasks

Users are given a set of written Web tasks to do. These same written tasks will be given to the model. The user does the tasks using a browser connected to the WWW.

Instrumentation

The browser is instrumented to produce a trace of behavior and the user talks aloud while performing the task. All the Web pages accessed are also saved away. Human-friendly representations of this trace are produced to aid the theorist in building the models and in coding the spoken transcripts. The result is a set of databases containing user traces and associated data.

Cognitive-perceptual stimulation model

The SNIF-ACT user simulation model is constructed (or probably refined).

User Comparator

The model is run in the user trace architecture. On each cycle, the model makes a prediction, generating another element in the trace, which will involve accessing the saved Web pages. The user trace comparator uses a set of rules to determine whether there is a match with the protocol trace; if not, an error is scored against the model and it is set back on track. eye tracker, logging software that collects all user interactions with a WWW browser, and video recordings of think-aloud verbal protocols [23]. These data are coded by automatic means and by hand into a comprehensive trace of states and events representing the interaction of user with the WWW. Computational models of user cognition and perception are then developed to simulate—as accurately as possible—the observed user-WWW interactions.