Hutchins [39,41,40] argues that cognition is distributed among the artifacts and individuals who participate in culturally and historically conditioned ongoing practices. These larger socio-technical systems have cognitive properties that cannot be reduced to the cognitive properties of individual persons. This larger unit of analysis provides a basis for the study of cognition [39] (p. 266):
In this paper, I will attempt to show that the classical cognitive science approach can be applied with little modification to a unit of analysis that is larger than a person. One can still ask the same questions of a larger, socio-technical system that one would ask of an individual. That is, we wish to characterize the behavioral properties of the unit of analysis in terms of the structure and the processing of presentations that are internal to the system. With the new unit of analysis many of the representations can be observed directly, so in some respects, this may be a much easier task than trying to determine the processes internal to the individual that account for the individual's behavior. Posing questions in this way reveals how systems that are larger than an individual may have cognitive properties in their own right that cannot be reduced to the cognitive properties of individual persons.If the unit of analysis is a single individual, cognitive scientists can only infer internal states of the actor from external behaviors, while with larger systems it is possible to directly observe some of the internal states of the the larger system.
Greeno [30] advocates a program for research that combines standard cognitive science with the research and methods of work on situated activity that have come out of the social sciences. For Greeno, the key is to analyze information structures in socially organized activities (p. 6):
Although these lines of research -- the study of individual cognition and of socially organized interaction -- both provide important scientific knowledge and understanding, they have developed in relative isolation from each other. Cognitive science analyzes structures of the informational contents of activity, but has little to say about the mutual interactions that people have with each other and with the material and technological resources of their environments. Interaction studies analyze patterns of coordination of activity but have little to say about the informational contents of interaction that are involved in achieving task goals and functions.Greeno recommends that both cognitive science and situative strategies of research be investigated `vigorously'.
The computational model of conventionalized behaviors that is developed in this paper fits into Greeno's proposed program of research. It details the cognition of individual members of a community of actors participating in joint activities. Individuals push through joint activities by relying on their memories of prior activities. Over time, conventions of behavior in joint activities begin to develop. But these conventions of behavior emerge only during the course of activity and are subject to the constraints of the larger context.
In general, a computational model allows the researcher to examine the interplay of the individual and distributed elements of cognition. From the perspective of individual psychology, the task for the cognitive modeler is to rework, in a manner consistent with the larger context, theories of cognition that resulted from a history of experimentation in indoor psychology. From the perspective of the distribution of cognition, the computational model shows how properties of the larger system can emerge from the interplay of individual cognitions. By integrating results across disciplinary boundaries, constraints and opportunities emerge that leads to general progress.
A computational model is a description of cognitive processes and not a demonstrative simulation [31] (p. 56). Two issues will suffice to illustrate this point. One issue concerns the selection of mechanisms to use for modeling the reasoning of the individual actor. For example, early Artificial Intelligence methods used explicit structural representations (e.g., frames: [55]) and semantic networks [61] for depicting schemata [8] as a part of memory. Within the parallel distributed processing movement (see [62]), the representation of the structure of memory became more distributed and moved closer to approximating what is known about how networks of neurons within the brain `compute'. For the cognitive modeler, the trade-off between using one or the other machinery as a part of his/her model is complicated. Neural nets may better approximate how the brain works, but as a practical matter, the current implementations of neural network computation are better for modeling low level perception than they are for characterizing larger fragments of phenomena such as everyday activity. The computational modeler can use explicit structural representations or rules for this purpose, but the caveat is that relevant features of the structured representations and rules must be eventually replaced by models that better approximate the brain activity. One could try to advocate a position that all high level theorizing is irrelevant to the eventual brain theory of behavior that will emerge, but it is hard to understand the force of this argument. In general, it is well understood that top-down filtering is a powerful problem-solving tool. In the case of combining the top-down theories of cognitive science and the situative view of cognition with the brain theories of the neuroscience community, evidence fails to support a radical neuron doctrine that would discount modeling from the top-down [28].
Another sticky issue concerns the level at which to analyze and represent the phenomena. The MW model represents the activities of the participants at a fairly course grain. Take a joint action like CARRY-TOGETHER. Actors can, and do, reason about their activity at that level. A finer grain analysis would feature the reasoning that is done as the carrying together unfolds. How is the path determined and obstructions managed? What happens when one (or both) of the actors needs to change her grip? How does the passage through a doorway proceed? At this finer grain level of analysis, there are continual adjustments that are made by the participants in order to wend their way through the activity. The decision to model at one level or another reflects a confluence of issues. A critical factor is that each finer grain level of analysis introduces additional complexity to the computational model. The level at which MW was modeled entailed the production of 25,000 lines of source code. An even finer grain analysis would have required larger amounts of code, and there is no reason to believe that the increase in code would have been linear rather than exponential. For example, at any granularity of analysis for the MW domain, one could talk about the mental parts and the action parts. It is a hotly disputed issues whether they are in fact separable. The differences of opinion pivot over claims about internal representations, interaction, and emergence. Some of these issues are addressed in the MW model, others are not. A finer grain analysis would have forced a careful examination of the constellation of issues concerning internal representations, affordances, and perception [12], but the resulting additional complexity in the modeling task would have prevented a summation on the point of convention.
In general, the selections of both a reasoning model and a level at which to analyze the phenomena are hydra-headed. Consequently, the computational modeler, at some point, needs to decide which issues should be pursued and to what depth. One way to think about a computational cognitive model is that it is a summary with an `attitude'. It is a summary because there is too much ground to cover. It has an attitude because the model that is produced is an interpretation and characterization of the relevant phenomena.