meadors@nprdc.arpa (Tony Meadors) (12/08/88)
Following in the footsteps of Dave Peru (a netter who recently offered a
definition of intelligence) I would like to offer some comments
of my own. Of course , there are a myriad of issues
and pitfalls involved in trying to
satisfactorilly tie down the concept of intelligence.
Yet if a concept is to be useful in reasoned
discussion or in scientific investigation one must pin it down at the corners:
at least have an operational, though admittedly incomplete, definition.
The term "intelligence" has had so many varied usages one can only hope that
a definition will be somewhat central to those many meanings.
So while I will obviously ignore a host of issues, my hope is that by
providing a definition, we might move past those issues and discuss what
I see as the more interesting ones (ie. what is essential to an intelligent
system? Is a functionalist answer to the
mind-body relation workable?)
Before I offer my pithy definition :-) I should
make explicit a few assumptions that "make it work."
1) Intelligence per se, is an attributed entity, just as are stubborness
and beauty; it is a character trait. As such it is subject to
all of the subjectiveness of such labelings.
2) Thus in looking for intelligence we should direct our discussion
toward the system's actions. An analysis of what constitutes intelligence
should be a theory of intelligent action (and later of the systems
which might display them).
3) The minimum units of intelligent action are
"controlable", "goal-directed" behaviors. This rules out
much of what our bodies "do" from consideration as intelligent acts.
Specifically, it rules out autonomous functionings and reflexes, as
well as isolated parts of purposive behaviors
(say perceiving for example). At the same time
it focuses our attention upon functional acts:
acts which accomplish some intended control over the actor's situation.
4) If one wishes to assess "how intelligent" a particular system
is then one must move to the informational level of analysis:
go behind the scenes of the system's actions.
Obviously, tasks vary widely in how
demanding they are of our information processing resources
(grasping a cup vs. completing law school).
Each task has unique requirements and likely a number of
solution paths. Thus an analysis of action
efficiency would necessarily need to go behind the scenes
and assess how effectivly the system utilized
the available resources in meeting its goals.
Finally, here is the Holy Grail definition,
a) Intelligence is attributable to systems whose "actions involve
purposive coordination."
b) How intelligent a system is, determined from
1) assessing its actions in terms of how well it can use
its resources toward satisfying some goal or goals.
2) considering the range of tasks it can tackle.
Some Examples and Commentary:
In chess, one considers
alternatives and selects one; in the store one considers alternatives and
selects a matching ensemble of clothing to wear to a banquet;
one considers and selects a college to attend; one
considers and prepares the evening meal; in each there are many
simultaneous and interlocking goals to be considered.
In defining intelligence as purposive coordination I am making it very
nearly synonomous with the broad notion of skills. The skill literature
addresses itelf to the learning and execution of behaviors which have
both mental and physical components: especially the elements which affect the
rate of learning and determine one's final skill level have been at the
center of the volumous research. These concerns relate very much to the
nebulous concept of intelligence and one finds that many of those researchers
write articles concerning intelligence as well.
Bloomington's douglas Hofstater has said that the central problem to
capturing intelligence is to capture how we "recognize the letter A" in
all the various contexts and forms it may occur. I would have it that a more
appropriate centerpiece might be called the "place the square peg in the
square hole" problem. While understanding our amazing recognition ability is
essential, trying to understand it apart from its role in the fuller action
cycle is not very meaningful in terms of intelligent action: goal-directed
action is where its at.
Let's consider a modest example. Say, someone bought a new
basketball post, backboard and rim, and now
wants to assemble and erect it. We may consider then any sub-part of
behavior (such as locating the right screwdriver among my
tools) or we can consider the entire sequence; at any level there will
likely be identifiable sub-goals, weighing of alternatives and actions
carried out. Lets say he in fact did
the unbelievable as his mind was elswhere, and erected the pole while
leaving the basket on the ground. Now, his imeadiate task is to somehow
get the two attatched together properly. He chooses method X and goes
about carrying out all that is required by method X. The main concern
is in how reasonable and/or plausible is it to assess that action in
terms of intelligence? Is such an assessment always unworkable in practice?
In unconstrained situations, probably. But a more stringent measure would
set up special problems (as IQ tests normally do) and also include post-answer
evaluations or protocals to get at the processes utilized and avenues
considered.
But let me make clear that even if it proved impracticle a method,
my intention is to provide a consistent, operational definition, one
which will allow us to discuss intelligent systems.
To finish this example, method X is intelligent to the degree that he
coordinated all his resources to meet all the explicit and implicit
goals he has (many goals are implicit, for example,
to not injure oneself performing an act). Recall that the definition of
intelligence only required purposive coordinated actions:
one can now at least consider the behavior of such
systems in these terms: however, the degree
of intelligence involves all sorts of additional postulates, some
of which have been mentioned.
We desire to enter into an algorithm several factors to yield
somewhat of an efficiency rating of constraints met. Lets assume that
method X was to climb up the pole and painstakingly nail the board to
the pole. Darn the luck, he tore his pants in the process.
To assess his selection of this action we will have to know much more,
specifically, most everything about the materials available to him
(a ladder maybe) as well as the goings on of his considerations. For
example, whether it never occured to him to use a ladder, or he
mentally had attempted to recall who he had loaned his to.
To repeat, the bottom line is in how effectively one performs actions which
satisy the many simultaneous constraints one has as goals.
This resourcefulness (as solutions to situational problems) is a function
of the actor's set of goals, his
knowledge, and the environmental specifics. Thus the number and value
of the constraints satisfied in relation to the optimal solution and
the time to solution would yield a rough idea of cognitive efficiency.
Of course tasks vary
in the ratio of reasoning to movement necessary to achieve its
particular goals. The "purposive coordination" definition
is intended to include skills as disparate as
catching a hard-line drive and solving algebra story problems.
In conclusion, let me point to some related issues.
Acquisition.
Some years ago it became fashionable to refer to intelligence
as "the ability to learn." This definition persists as a popular
idea despite obvious problems; not the least of which are
cases of quite otherwise normal individuals displaying a marked
inability to learn while remaining
as "competent" in all other respects (due to brain damage).
A number of theorists choose to emphasize that intelligence is a
product of "socialization." What this is meant to convey
is often unclear. If it simply means that an individual does not,
and could not, develop intelligence without interaction with a
challanging environment, then it is obvious.
If this is indeed what is meant, then I
see no need to use the term "social"; it makes it sound as if there
is something in groups of individuals that has an effect upon how the
nervous system enacts its control.
The fact that our social organization, or civilization, has become
a bootstrapping device for each suceeding generation in no way ties
the much broader issue of intelligence to culturally
transmitted knowledge and tradition.
In sum, any intelligent system is likely to have the ability to
learn and develop, that is a part of what we expect an intelligent
system to do: to not repeatedly
make the same errors, to remember what it did
yesterday, etc. Yet this learning is but a portion
of what a system is "doing" in organizing itself to
exhibit adaptive behavior. So it is arguably necessary but not
a sufficient, or defining characteristic of intelligent systems.
The problem is really that the term learning is so general it may
be thought "responsible" for any knowledge or action which obviously
had to have been "acquired" along the experiential road.
Phenomenology:
The issues of mental experience or consciousness are simply irrelavent
to the action approach to intelligence taken here.
This is one of the main strengths of focusing upon actions.
Cognition:
It might seem that this action approach misses the more abstract reasoning
and linguistic abilities often thougth to be the essence of human
intelligence. That is, what is the status of cognitive processing
in this action-based definition of intelligence? Well, the
problem is in assuming some major cleft between
motor control and cognition. It
is explicitly assumed here that our actions are steeped in cognitive
consideration, reasoning, and the like. In sum, in an action-based
conception of intelligent systems, internal representations and
their interaction necessary will play a large role in carrying out
one's intentions regardless of the form of the goal.
Forgive this grandiose listing. My claim is only that this may be a
profitible conception of intelligence, not the best, or
the only one, by any means.
tonyM
the sci.psychology group needs some stirring up so I included them :-)
Note: if you 'f'orward anything in response to this please add
sci.psychology & philosophy.misc to the distribution line... Due to
funniness at my end I had to mail to comp.ai seperately, but all
three have received these comments.