PCT Agents

[Michael_Fehling]

[Michael Fehling 930912 3:58 PM PDT]

In re Rick Marken 930911.0900:

Rick,

You ask "Is there any cognitive model that is NOT based on behaviorist ideas?"

  By now, Avery Andrews has answered this question. Also, if you reexamine my
prior post (930909 8:46 AM PDT), you'll find that I did, too. Regardless, you
seem already to be aware of the answer as your next question reflects: "In
what way does Chomsky's model of behavior differ from a behaviorist model...?"

  As for this second question, my article on Newell and unified theories of
cognition (UTCs) includes extensive discussion of Chomsky's notion of a
competence account and contrasts it with S-R behaviorism's "models" and
methods for testing them. I invite you to look at that. If that's not
enough, then read Chomsky himself or Julius Moravcsik's account in his
monograph, _Thought_and_Language_ (Routledge, Chapman & Hall, 1992).

  Finally you want to know if there are "_any_ non-behavioristic cognitive
models." Given the above, I'll take that to mean--in addition to Chomsky's.
Beyond this I am having difficulty ansering your question. This is because
you introduce some rather surprising definitions in setting up this question.
In particular, you say PCT asserts that "'behaviorist _just_ means 'causal'"
and so, "Newell's model is 'behaviorist' _because_ it is causal" and not
"...because it is S-R" (emphasis added). Later you also say that a model is
"causal" if "the dependent variable (sic) is an output."

  A lot of things seem tangled up here. In fact, it does not square with my
reading of BCP, either. First, a theory or model is causal if it discusses
cause-effect relations among the elements of the system that it describes.
This has nothing to do with whether the theory is about behavior. There are
fine examples of causal theories of mind, for example. So, I hope that PCT
doesn't really claim that "behaviorist just means causal." Having read BCP, I
don't think it does. Second, let me point out that whether a theory or model
is causal has in general nothing to do with which variables in some test
procedure are controlled or allow to vary as dependent measures. In fact,
models with absolutely _no_ cause-effect assertions--e.g., a standard
functional model electrical circuit model--can be examined in experiments with
clearly defined dependent measures. Here again, I don't think that PCT (as
per BCP at least) conflicts with my point. Third, please note that I have
claimed that Newell's model is an S-O-R model (e.g., comparable to Tolman's).
This is an important distinction to note in seeing why I compare SORA to
behaviorist models. In general, my main argument is that Newell is using
behaviorist methods and an S-O-R model to explore what is otherwise a a theory
of competence (in Chomskian terms). And, I claim that most models in
cognitive science mix oranges and apples in this way. Again, I recommend that
you look at my commentary in AIJ for details if you are interested.

- michael -

[_Avery_Andrews]

[Avery Andrews 930912.1701]

Re Michael Fehlings proposal that we make a PCT Agent, first a
short philosophical sketch, then a longer and more rambling
prospectus for a more specific project.

Philosophy:

I agree that a building a PCT Agent would be a good idea, but I
don't think it should be a head-to-head competition with SOAR
and similar projects, but something that capitalizes on the assets
that PCT already has on the ground, which means emphasizing what
might be called (vaguely after Chapman and Agre) problems of
embodiment. E.g. how do you actually get limbs to follow the trajectories
needed to achieve the things you want done.

Along these lines, I've been envisioning an Agent I call `Jerry', who's
supposed to be a PCT version of the Agre & Horswill `Toast' program,
which solves the logical problems involved in making breakfast.
Jerry (named after Fodor, who poses making breakfast without burning
the house down as a significant problem for AI) is supposed to perform
the actual motions involved in breakfast-making, and part of the aim
is to see to what extent things that look like they call for logic, etc.
can actually be done just by lashing control systems together right
(quite a lot, I suspect).

Ramble:

Jerry hasn't even begun to exist yet, but a major preliminary is to
get good at doing things with arms, which I hope to get back to pretty
soon (starting more or less from where I left off with the `naive Cartesian
transpose blues' theme of six months or so ago. So I'm thinking of
jazzing up Bill's 14deg program (an arm with 14 degrees of freedom in
its motions, but no dynamics) so that it will track things with its
finger-tip, such as bugs crawling on the opposite side of a pane of
glass.

We simulate the glass pane (without doing force dynamics, or anything
like that), by calculating a perceptual signal that is zero if the
calculated position of the fingertip is above or right at the surface
of the pane, but rises very rapidly as it goes beyond the surface
(in the direction normal thereto, of course). There will then be a control
system trying to maintain this perception at a small positive value,
which should correspond pretty well to controlling the pressure of the
fingertip on the pane. Under these circumstances, there are quite a lot
joint angles that can influence fingertip pressure in various ways in
various arm postions, so that should be a nontrivial challenge.

The actual tracking should also be. For example, when the forearm has
a vertical projection in your visual field, the `shoulder roll' and
`wrist yaw' angles provide good ways to correct errors in the
horizontal dimension, but when the forearm is horizontal, they
are good for errors in the vertical dimension. `Smart output functions'
would be one way to approach this, but recalling the Cartesian Blues
(wherein I made the easy problem of controlling two dimensions of
perception with df's of freedom into a hard problem, by using the
wrong coordinate system, equals controlling the wrong perceptions),
I suspect that smart perceptions might work better, the idea being
that these joints be used to correct errors in the direction
`normal to forearm' (there's no reason why the same error can't
be multiply represented in different coordinate systems, = different
perceptions, feeding different corrective strategies, as long as they
are all more or less consistent in their effects.

Interesting modifications would include:

  a) tracking with sticks of varying lengths (the tracking circuits are
      kept constant, but the relation between hand position and where
      the fingertip is is varied, and we see how it performs).

  b) tracking subject to constrants such as that the elbow be stuck
     way out (to confer greater ability to resist lateral disturbances,
     such as if there's an elastic cord attached to your arm which Gary
     Cziko might pull at any moment).

  c) tracking with a joint held voluntarily frozen in some position,
     by a system controlling for a specific angle at that joint (to
     illustrate control system interactions).

Well that's all for now.

Avery.Andrews@anu.edu.au