Archive: August 20th, 2008

Wilson & Clark: How to Situate Cognition

[Readings] (08.20.08, 10:18 pm)

Understanding the idea of “situated cognition” by comparative terms. Concludes that definitive characteristic is *cognitive extension*. Embodied cognition is a continuation of this idea.

History of cognition: Individualistic cognition, mind alone, wedged between perception and action.
Approach also suggests that cognition takes place on the features of symbols, as opposed to features of the individuals themselves.
Significant works in the individualistic thread include Fodor, Pylyshyn, and Newell and Simon. Pure example is the Cyc project, which (as we know from Alison Adam) promotes a sort of “view from nowhere”, presenting a sort of assumed identity.

Putnam and Burg
e challenged individualistic perspective, leading to perspective of taxonomic externalism, and these have been extended to more radical theories of externalism. Externalism pushes the idea that cognition extends beyond the individual thinker, even past the flesh and into the environment. When externalism is first introduced, when cognition is pushed outside of the brain (or abstract symbol system) into the body, then it becomes a slippery slope to determine where the edge of cognition stops. This naturally leads to a diversity of conflicting theories.

Extended computation (or wide computationalism, which is a synonym), is not a severe departure from computation, but simply an extension thereof. These are ideas that look at computation as taking place spread across an environment, as opposed to inside the skull of the thinker. Computationalism is not incompatible with situated or extended cognition, but those rather extend from it.

Examples of extensions are bodily extensions, technology and prosthetics. Symbolic thinkers have also used the analogy of prosthetic, but emphasized the prosthesis of the mind, eg Vannevar Bush. This idea has been continued with Don Norman with the notion of affordance. This is described in Wilson and clark as “cognitive augmentation”. Part of the issue with this is how augmentation extends, but also restricts. Augmentation fits with the planning-oriented perception of the world, but stumbles when faced with the idea of expression or limitations (a hammer encourages you to think in terms of nails).

Social structures are mentioned. A crucial example is writing (in terms of cognitive supplement).

An example given is the “task specific device” or TSD. This is something that exists either in the environment or in the actor, and is used to enable certain types of action. Like the case with prosthetics, it promotes an instrumental and intentional model of behavior. Related to TSDs are “transient extended cognitive systems” or TECSs, which seem to be ways of approaching cognition on a per-context basis. The idea of the TECS is similar to the tool oriented approach, but seems to be much more flexible and free-form.

Looking at the boundary between cognition and non-cognition. An argument against extended cognition is the “Dogma of Intrinsic Unsuitability”, which states “Certain kinds of encoding or processing are intrinsically unsuitable to act as parts of the material/computational substrate of any genuinely cognitive state or process”. At odds with Intrinsic Unsuitability is the “Tenet of Computational Promiscuity”, which is the property of computation to spread out across many parts of mind and body.

Another challenge to extended cognition is embedded cognition, which claims that cognition is embedded in things external to the body. One idea in this is memory. Other ideas are the confusion associated with examining changing thinker+tool combinations as cognitive subjects. The authors dispute this because embedding implies heirarchy and order which is missing in application.

Reading Info:
Author/EditorClark, Andy and Wilson, Robert
TitleHow to Situate Cognition: Letting Nature Take its Course
ContextGives a background and argument towards extended cognition. This notion is very useful for rationalizing contextualized situational behavior in AI based agents, expecially relating to the believability of The Sims, etc.
Tagsai, embodiment
LookupGoogle Scholar

Michael Mateas visit

[General,Talks] (08.20.08, 7:27 pm)

Michael Mateas visited today, and gave a presentation about his Expressive Intelligence Studio at USC. The project is about automated game design, which is interesting, since that was one of the original goals of my MS research, before I turned it into a space generation thing. Basically, this idea is something that would support formal game studies by exposing and finding new ways to put together mechanics. It also encourages thought about design at a meta level, reasoning about types of mechanics and how they can be put together. The existing work doesn’t do much yet, but it looks like it might yield some interesting results.

The work involves four layers of work:

  1. Game mechanics: State and state evolution. The actual mechanisms by which state is represented and can advance are part of a larger meta-model.
  2. Concrete representation: Audio, visual elements that represent the mechanics to the player.
  3. Thematic content: Real world references, common sense associations. This makes the game meaningful outside of a purely symbolic context.
  4. Control mappings: User interaction and verbs.

The starting point for the EIS lab was to look at the thematic content, which is arguably the hardest part of the problem. This bit has to make meaningful associations between game mechanics and the underlying concepts. For instance, if the game is about chasing, the player can either chase or be chased, and whatever is being chased must be something that someone would have reason to chase. The associations here were defined via Open Min, ConceptNet, or WordNet. I forget exactly which combination of these was used. The goal was to enforce consistency in game mechanics with the thematic concepts. The result of this was unfortunately rather messy and somewhat absurd in a lot of cases, due to conceptual slippage.

The more interesting area of work I found was in the reasoning about the mechanics themselves. This was done via event calculus, which has been described very effectively by Eric Mueller. The event calculus can reason about events, states, and can be used under the hood to restrict the types of states that can be reached by a given set of game mechanics. Essentially, the calculus can be used to define a suite of invariants, almost like unit tests, and test these on a given set of mechanics, allowing a designer or an automated tool to modify the mechanics quickly and find out whether the invariants are met.

Food for further investigation.