Brynn Marie Evans

This is a review of a paper I read for a course on Distributed Learning (in the Education program at UCSD).

Analysing human-computer interaction as distributed cognition: The resources model
Wright, Fields, & Harrison are trying to blend theories of distributed cognition with human-computer interaction modeling. The problem is: much of HCI is task-based, but focusing on single-user task performance doesn’t help us understand multi-party cooperative work. A better unit of analysis for studying & designing for HCI would be to look at the network of people & technologies involved in some interaction. Distributed cognition theory could help here.

They propose a “model” called the distributed information resources model—or resources model for short. It’s not a model intended for making predictions, but rather for helping people think about interaction in new ways. [And in this paper, it is applied only to understanding single-user task performance, actually.]

A critical observation from the distributed cognition literature that informs their model is that external representations are critical for task performance because they (1) simplify the problem and (2) reduce the user’s cognitive workload. The problem can be made simpler by supporting recognition-based memory or perceptual judgments rather than recall, for example. But ultimately, users complete their task by coordinating both internal and external information structures. The resources model they propose tries to take this into account:
(a) that there are several types of abstract information structures in the environment that people exploit for everyday tasks or problem solving activities
(b) that we can look at the process of how people coordinate internal and external representations to understand a user-interface interaction

What is distributed?
User behaviors are distributed! Nearly all individual or cooperative work involves recruiting or exploiting resources from the environment (whether from technologies or other people).

They used the historical examples of cockpit cognition (pilots working together to land a plane) and fixing the location of a ship at sea (all from Hutchins). Laboratory tasks can be thought of as distributed cognitive events, too. Zhang and Norman studied how people perform the Tower of Hanoi problem (below), with the key observation that tasks can become greatly simplified if some of the important information for the task is represented externally in the problem apparatus (instead of staying internally in the subject’s head).

How are the distributions mediated?
Information structures are distributed and coordinated by users; illustrated by concepts in their resources model:

The resources model includes 6 abstract information structures (these six are the most obvious and they could each be represented internally or externally depending on the situation):

1. Plans: a sequence of actions, events, or states
2. Goals: will help a user accomplish a task; also thought of as the required state of the system
3. Affordances: cues about the set of next possible actions
4. States: a collection of relevant objects
5. Action-effect relations: describing the causal relation between an action and a state (or outcome of a state)
6. History: a sequence of plans, goals, states over time (may be thought of as the system’s memory)

These concepts are made clear by thinking about the process in which they may be acted upon and brought into coordination with each other. Consider the plan following procedure that illustrates how various resources are brought into coordination with each other:
Plan following:

• it requires the plan, various affordances of various states to guide you through the plan, and an interaction history to keep place of where you are in the plan
• some of these information structures are accessed through the environment /interface (e.g., states, affordances), others must be kept in a user’s head (e.g., only part of the plan is visible in any given state)

Significance and Contributions:
There are 2 key contributions from this model:

1) It can be used to compare different interfaces by talking explicitly about which information structures are being used and how they are being brought into coordination to produce action. For example, we could compare the Microsoft chart “wizard” (above) with the ClarisWorks chart dialog (below).

The ClarisWorks chart dialog combines all three steps from the Microsoft chart “wizard” into one panel. But is this better for the user? Maybe yes, maybe no.

The benefit of the ClarisWorks dialog box over the Microsoft chart “wizard” is that all parts of the plan are made explicit to the user at once. The plan is externalized; the user won’t have to keep the history of his steps in mind. But plan following is not supported: The order in which the actions are carried out is unconstrained. This may not matter, though, if the task is relatively simple or the options (buttons) in the dialog straightforward. Presenting clear meaning in the options on the interface can support goal matching instead of plan following.

The resources model can compare these two scenarios based on the resources available and whether those resources are balanced between internal and external cognitive processes.

2) The model it can be used to drive design and interaction strategies by modeling different interface states. The 8-puzzle game has been studied by researchers and found to have different outcomes depending on how you presented the game to the subject:

In version (a), the subject sees his playing board (left) and a representation of the goal state (right). Actions are made by direction clicking on the tile you want to move == direct manipulation interface.

In version (b), the subjects sees both the playing board and goal state, but his actions are made by clicking on an intermediate panel which controls which tiles move where == indirect manipulation interface.

The key finding: the indirect manipulation interface affords better performance on the game because subjects are spending their time crafting a successful strategy. In the direct manipulation interface, subjects play a trial-and-error game.

Given this, the experimenters created two novel 8-puzzle game scenarios based on their resources model. They knew that the physical layout (presentation) of the game had a large effect on subjects’ strategies and ultimate performance. How would a round layout versus a completely horizontal affect performance? The major difference between these various layouts is the action-effect representations. Clicking a tile will cause the tiles to move to a central location (“roundabout version” on the left) or the far right location (“8-bar orientation” on the right)

They found that people performed much better on the roundabout version. The round physical layout suggested the goal state at the same time that subjects were performing the task. This subtle visual cue caused them to adopt the strategy that was most successful from the first set of experiments (that only the indirect manipulation interface afforded). Additionally, when subjects were given the round version first followed by the horizontal version, they actually used the successful strategy they learned on with the round version with the 8-bar orientation.

This was an interesting study because it was based on an understanding that the experimenters gained from applying their resources model to the task design. It highlighted the fact that external representations of information, affordances (hints), and action-event outcomes completely changes user performance. What’s critical though is that it doesn’t just change performance, it can crucially guide problem solving strategy (ultimately a learning device!) which may later be applied in other situations.

Citation: P.C. Wright, R.E. Fields & M.D. Harrison (2000). Analysing human-computer interaction as distributed cognition: The resources model. Human Computer Interaction, 15(1), 1-42. [link to PDF]