Competence Theories

CSGnet Archive CSG1993
1

[Michael Fehling 930912 12:18 PM PDT]

In re Rick Marken 930913.1030 --

Rick,

Thanks for the comments on my article. I'm very busy now, but on a quick
quick reading, your interpretations seem basically in line with what I tried
to say. I note just three things:

        1. Your use of the term "behaviorist" is unusually broad. Usually
            it is reserved for approaches seeking to describe/predict/explain
            _concrete_instances_ of behavior, rather than idealized and
            abstracted classes of instances. According to your view, _all_
            psychological theories are behaviorist, except, perhaps for one
            that denies the existence of physical actions.
        2. Regarding Chomsky, your process interpretation of Chomsky's model
            is not what one tries to capture in a competence account. Viz.,
            your (edited) diagram
                        "S --> PS -->T -->...SS",
            should _not_ represent processing (performance) steps, according
            to Chomksy. Instead, it should be read as indicating relations
            between adjacent structures defined by "generative _rules_ that
            map one structure into another. Note that "generation" is
            _not_ processing or behavior, it is a logical relationship between
            rule's input structure and its output structure. (There are some
            other, minor terminological glitches, but they aren't important
            here, so I'll pass over them for now. So, Chomsky's theory only
            qualifies as "behaviorist" if one accepts your very broad
            definition I reacted to above.
        3. In contrast, your use of the term "causal" is unusually narrow
            (and, evidentally, at odds with many passages in BCP). Here I
            have even more trouble. Clearly as you say, "cause and effect are
            arranged in a _loop_" as per PCT. But, traversals of this loop
            occur in time, Bill's static analysis notwithstanding. In fact,
            whenever loop dynamics need to be analyzed (e.g., for assessing
            covergence, stability, etc.) then the causal structure becomes
            quite importnat to understand. Based on my knowledge of learning
            control (e.g., dual-control theory) PCT's topic of
            "reorganization" will also bring causality to the surface. So,
            I think you do PCT an injustice by using a label that suggests
            no role for causal analysis.

- michael -

2

[From Rick Marken (930913.1030)]

Michael Fehling (930912 3:58 PM PDT) --

I have read your review of Newell's "Unified theories of cognition"
(AI, 59, pp 295-328) and I can see why you say that "Competence"
theories are non-behaviorist models of cognition. The goal of
competence models is to model "ideal" performance -- not what we
observe but what we would (presumably) observe from an "idealized
agent" -- one whose actual performance is not subject to "error".
This obviates the need for "predictive modelling" of behavior --
the sort recommended by Newell-- as a basis for evaluating the
success of models since all you see is real behavior -- a shadowy
reflection of the ideal. Competence theory "... focuses on des-
cribing knowledge embodied as generative processes or dispositions
that give rise to idealized performance. It does not attempt to
predict situation-specific behaviors" (p.319) So how does one
evalute a Competence theory? Apparently by comparing the model
to behavior -- but not to "situation specific" behavior. You
says "Competence theorists have no direct concern for predicting
the idiosyncratic details of actual human performance" (p.323).
But, apparently, competence models are designed to "succeed in
completing some specified tasks of interest under a plausibly
complete range of expected conditions" (p.323). So I presume
that a Competence model would be deemed a success only if it
completed the same "specified task of interest" as one completed
by a human -- even if the detailed (situation specific) behaviors
of model and human were not the same.

My interpretation of this is that Competence theories are
evaluated by comparing the model's to the human's ability to
produce goal states of environmental variables -- such as the
solution state of a water jar problem -- rather than by comparing
model and human in terms of the detailed actions (moves) that were
used to produce this result. Competence theories are still being
evaluated against behavior -- just a different aspect of behavior
(ends rather than means).

Of course, it is possible that I'm wrong and that Competence theories
are NOT evaluated on the basis of a comparison to ANY behavior
at all. Maybe this is what Avery meant when he said (930212.11258)
"Chomsky doesn't have, and never has claimed to have, any model of
behavior". Maybe Competence theories are like algebras and they are
evlauated only in terms of internal consistency. In that case they
would be truly "non-behavioristic" since they are not about behavior
AT ALL.

But my impression from reading Michael's book review is that
Competence theories are aimed at understanding human "cognitive"
behavior. But this is not what I meant when I said that cognitive
models are "behavioristic"; any model of behavior is behavioristic
in the sense that it is designed to explain something about
behavior (PCT is certainly "behavioristic" in this sense since
PCT is designed to explain purposeful behavior). So Competence
models are (hopefully) as behavioristic as PCT; otherwise they
are pure mathematics -- not that there's anything wrong with that.

Behavioristic, to me, means that the model is set up so that
inputs are lineally related to outputs. I see Chomsky's language
model as a behavioristic model because it is organized as a
cause-effect chain:

S --> PS -->T -->N-->SS

That is, the idea for a sentence,S, (what the speaker intended
to say) is turned into a phase struture, PS, which is turned
into a tranformation, T, which is turned into the "output"
sentence, SS -- the "surface structure". I have added a noise
source, N, between T and SS to account for the non-ideal nature
of the sentences that we actually hear (and see). The interface
between behaving system and environment, by the way, is somewhere
between T and SS also.

If this is a fair representation of Chomsky's Competence
model of language than it is "behavioristic" from my point
of view because it is based on the assumption that langauge
behavior (both real, SS, and ideal, T) is the end result of
a lineal causal process. The input to this process is a "thought"
in the head (S) and the result of this process is a sentence (T
or SS); the process itself is a causal chain -- with behavior
at the end.

If this is an incorrect representation of a Competence theory,
then I'd appreciate seeing how it should be corrected.

The PCT model is non-causal because cause and effect are arranged
in a LOOP. The behavior of this circular arrangement is QUITE
different than the behavior of ANY causal system (even though the
components of the loop are -- and must be -- causal elements;
arranging causal elements into a loop is not a small architectural
change -- it makes all the differnece). The behavior of such a
circular arangement of cause and effect elements is called
CONTROL or PURPOSEFUL BEHAVIOR. That's what PCT is about.

Regards

Rick