The SNIF-ACT Model

The goal of our modeling effort is to develop a computer program than simulates the user in enough detail to reproduce the user data. SNIF-ACT (see figure below) is the model that we are currently developing to simulate WWW users. SNIF-ACT is an extension of the ACT-R theory and simulation environment, a general production system architecture designed to model human psychology. By using this system to model WWW behavior, we link our analysis to the same principles used to model cognitive behavior in general. ACT-R contains principles concerning: (1) knowledge representation, (2) knowledge deployment (performance), and (3) knowledge acquisition (learning). There are two major components in the ACT-R architecture: a declarative knowledge component and a procedural knowledge component. ACT-R has two kinds of memory for these two different kinds of knowledge.

Declarative Knowledge

Declarative knowledge corresponds to things that we are aware we know and that can be easily described to others, such as the content of WWW links, or the functionality of browser buttons. Declarative knowledge is represented formally as chunks in ACT-R. Declarative chunks in ACT-R have sub-symbolic activation. Activation represents the log-odds of how likely a piece of knowledge is needed at a particular time, and may be interpreted metaphorically as a kind of mental energy that drives cognitive processing. Activation spreads from the current focus of attention, including goals, through associations among chunks in declarative memory. These associations are built up from experience, and they reflect how ideas co-occur in cognitive processing. Generally, activation-based theories of memory predict that more activated knowledge structures will receive more favorable processing. Chunks with higher activation values take less time to use and have a greater chance to have an impact on behavior. Activation is a way of quantifying the degree of relevance of declarative information to the current focus of attention (mathematically, it represents the posterior probability of how likely each piece of declarative information is needed given the current focus of attention). At any point in time, there is a stack of goals encoding the user’s intentions. Goals are also represented as chunks. ACT-R is always trying to achieve the goal that is on top of that stack, and at any point in time, it is focused on a single goal.

Procedural Knowledge

Procedural knowledge is knowledge (skill) that we display in our behavior without conscious awareness, such as knowledge of how to ride a bike, or how to point a mouse to a menu item. Procedural knowledge specifies how declarative knowledge is transformed into active behavior. Procedural knowledge is represented as condition-action pairs, or productions. For instance, our SNIF-ACT simulation contains the production rule Use-Search-Engine. The production applies in situations where the user has a goal to go to a WWW site, has processed a task description, and has a browser in front of them. The production rule specifies that a subgoal will be set to use a search engine. The condition (IF) side of the production rule is matched to the current goal and the active chunks in declarative memory, and when a match is found, the action (THEN) side of the production rule will be executed.

At any point in time, a single production fires. When there is more than one match, the matching rules form a conflict set, and a mechanism called conflict resolution is used to decide which production to execute. The conflict resolution mechanism is based on a utility function. The expected utility of each matching production is calculated based on this utility function, and the one with the highest expected utility will be picked. In modeling WWW users, the utility function is provided by information foraging theory, and specifically the notion of information scent. This constitutes a major extension of the ACT-R theory and is described in greater detail below.

Utility: Information Scent

As users browse the WWW, they make judgments about the utility of different courses of action available to them. Typically, they must use local cues, such as link images and text, to make navigation decisions. Information scent refers to the local cues that users process in making such judgments. The analogy is to organisms that use local smell cues to make judgments about where to go next (for instance in pursuing some prey). The model of users’ judgments of information scent is based on spreading activation. The basic idea is that a user’s information goal activates a set of chunks in a user’s memory, and text on the display screen activates another set of chunks. Activation spreads from these chunks to related chunks in a spreading activation network. Through this spreading activation network, the amount of activation accumulating on the goal chunks and display chunks is an indicator of their mutual relevance. The spreading activation network is therefore content-based, as mutual relevance of user goals and contents are calculated each time the display changes. The amount of activation is used to evaluate and select productions. The activation of content-dependent chunks matched by production rules can be used to determine the utility of selecting those production rules dynamically.

The spread of activation from one cognitive structure to another is determined by weighting values on the associations among chunks. These weights determine the rate of activation flow among chunks. In the context of WWW browsing, we assume that activation spreads from the user’s goal, which is the focus of attention, through memory associations to words and images that the user sees on WWW pages. Associations have strengths or weights that determine the amount of activation that flows from one chunk to another. If the user reads some link text on a WWW page, and the link text is strongly associated to the user’s goal, then we expect the user to judge the link as being highly relevant to the goal.

The association strengths among words in human memory are assumed to be related to the probabilities of word occurrences and of word co-occurrences. Consequently, the spreading activation computation of information scent in SNIF-ACT requires these estimates. In past research, we derived these estimates from the Tipster corpus. This database contained statistics relevant to setting the base-level activations of 200 million word tokens and the inter-word association strengths of 55 million word pairs. Unfortunately, the Tipster corpus does not contain many of the novel words that arise in popular media such as the WWW. For instance, the movie title "Antz" does not occur in the Tipster corpus. Consequently, we augment the statistical database derived from Tipster by estimating word frequency and word co-occurrence statistics from the WWW itself using a program that calls on the AltaVista search engine to provide data. The spreading activation networks needed to perform the scent computations are stored in a scent database that is accessed when production evaluations are computed by SNIF-ACT.