NetNews Usenet Archive 1992 #31/NN_1992_31.iso/spool/comp/ai/philosop/7194

Text File | 1993-01-03 | 8.8 KB | 163 lines

Newsgroups: comp.ai.philosophy
Path: sparky!uunet!think.com!spool.mu.edu!umn.edu!umeecs!quip.eecs.umich.edu!marky
From: [e-mail] (Mark Anthony Young)
Subject: Re: grounding and the entity/environment boundary
Message-ID: <[e-mail]>
Sender: [e-mail] (Mr. News)
Organization: University of Michigan EECS Dept., Ann Arbor
References: <[e-mail]> <[e-mail]> <[e-mail]>
Date: Sun, 3 Jan 1993 18:12:43 GMT
Lines: 151

%r [e-mail] (Mark Rosenfelder)
> I'm interested in what it takes to make an AI grounded-- to make its use of
> language meaningful rather than "merely syntactic"-- to have it know what
> it's talking about.  Placing Harnad, Lakoff, and my own musings into a
> blender and pressing "Liquefy", I come up with this: For an entity to be
> (statically) grounded it must have high-bandwidth sensorimotor experience
> with the real world; it's less grounded to the extent that the experience
> is of low quality or quantity, or not really its own, or is weakly
> integrated with its internal architecture.  Dynamic grounding would depend
> on a capacity to maintain this experience and function in the world.
> I trust you can see how this formulation leads to a concern with what is
> or isn't part of the system.
>
>[...]
>
>Just in case I haven't laid out enough flame-bait already... Some alternative
>conceptions of grounding avoid the boundary problem by concentrating on
>correlation of the internal world model with the real world.  The problem I
>have with these conceptions is that it's even less clear where you draw the
>line between an ungrounded system that's just playing with symbols and one
>that really knows what it's talking about; that it's not clear how many
>errors, gaps, and ambiguities are permitted in the correlation; and that the
>origin of the correlation is not taken into account.

What Mark Rosenfelder is talking about above seems to me more about how
something gets or stays grounded.  Still, just being "right" about the
world doesn't seem like enough to be grounded.  There is some notion of
"having experience" involved in that notion.  I'd like to offer the
following as a definition of grounded that uses both correlation and
experience of the world.

To give my definition of "grounded", I first have to define a couple of
auxiliary notions.  A system is said to be "correct" to the extent that its
internal world model correlates with the real world.  Note that this is a
graded feature, not binary.  Thus, one system can be said to be "more
correct" than another (though I do not claim that degree of correctness
forms a total order).

Example:  Consider two (twelve hour) clocks:  one is stopped; the other
runs slow by one minute every twelve hours.  In the first twelve hours
after they are set, the first clock is on average three hours out, while
the latter is on average thirty seconds out.  Over that limited time span,
the second clock is 360x as correct as the first (on average).  Of course,
as time goes on, the second clock becomes even more incorrect.  In the
limiting case, both clocks average three hours of error, and so are equally
correct.

Note:  it is often said that the stopped clock is right twice a day, while
the slow clock is wrong almost always.  However, unless one is prepared to
argue that the human mind actually contains a tiny scale model of the
universe, then the appropriate definition of "correctness" relies not on
"precise match", but on "amount of error".

The example of the clocks shows that "correctness" may decay.  Sensory
input can check that decay, however.  Consider having someone reset each
clock to the proper time at noon each day.  Then the stopped clock will
stay three hours out on average, while the slow clock will be 180x as
accurate, at an average of one minute error.

The measure of correctness does not depend on how the correctness is
achieved.  Thus it is irrelevent whether the clock is reset by a human or
is actually capable of detecting when an atomic clock "strikes" noon and
then resetting itself.

A system is said to "adjust" when its internal world model is compared to
the world, and is changed to a more correct model (if necessary).  Thus our
clocks above adjust at noon every day.  A system that adjusts itself is
said to be self-adjusting, while one that is adjusted by some other agent
is externally adjusted.  The clock that is reset by a human is externally
adjusted, while the clock that resets itself is self-adjusting (in our
example it is adjusting _to_ the atomic clock--an interesting question that
I will not get into here is "what if the atomic clock is incorrect?").

Note that the stopped clock in the example above is just as adjusted as the
slow clock.  When our custodian comes around, or when the atomic clock
strikes noon, the stopped clock is correct, so its model does not need to
change.  If the custodian happens by at a few minutes before or after noon,
then the stopped clock would be adjusted accordingly.  The important part
is that the model was checked and any detected inaccuracies were
eliminated.

Systems can adjust to greater or lesser extents, and more or less often.  A
self-adjusting clock that is slow by five minutes a day will adjust by five
times as much as one that is slow by only one minute a day.  Since the
amount of adjustment depends only on how fast error builds up in the
system, we are more interested in the frequency of adjustment.  If they
both adjust only at noon (standard time), then they adjust equally as often
(1ce/day).

A system is "adjusted" to the extent that it has recently adjusted its
world model.  For our example, the clock adjusts at noon, and so is fully
adjusted at noon.  As the day wears on, the clock becomes less adjusted
until, at just before noon the next day, it is 24 hours less adjusted --
half as adjusted as it was 12 hours earlier.  On average that clock is
twelve hours unadjusted.  The clock that adjusts itself every hour is 24x
as adjusted as the clock that adjusts daily.

A system is "accurate" to the extent that it can maintain correctness
without adjusting.  One system can be more accurate than another, and yet
less correct:  if it is adjusted less often, the error accumulated over a
longer time may outweigh the larger incremental errors built up over a
shorter time.

One last definition before I get to grounding proper.  A system is said to
be "complete" to the extent that its internal model of the world can
describe the world.  As an example, a clock with two faces -- say one set
to London time and the other to Tokyo -- is twice as complete as the clock
with one face.  If the two faces are however forced to read the same time
(by a mechanical connection, for example), it is no more complete than its
cousin.

Now to the definition of grounding.  A system is "grounded" (in a
particular world) to the extent that it is correct and complete and
adjusted (for that world).  If no world is mentioned, then, based on
context, either reality is assumed or we are talking about grounding in any
world.  Grounding is a graded concept, but does not admit a total order.
The extent of grounding can change over time (almost certainly will).  When
we speak of how grounded a particular system is, we mean its average
grounding over some (unstated) period of time.

A system is "self-grounded" to the extent that it is correct and complete
and self-adjusted.  This is the most interesting kind of grounding, as our
notion of intelligence relies on _self_ adjustment.

While (self-)grounding is a graded concept, it does not have a total order.
A clock that loses one minute a day and is reset every week is more correct
than one that loses ten minutes a day and is reset daily, but it is also
less adjusted.

Note that the bandwidth of the sensory experience has no direct effect on
how grounded a system is.  It will, however, have an indirect effect:  the
wider the bandwidth, the faster the entire model can be checked for errors
(until the bandwidth reaches the size of the model, anyway).  Thus, ceteris
paribus, the system with the wider sensory bandwidth will be more correct
and thus more grounded.

Also, motor abilities have no direct effect on grounding.  It seems likely
that a system that can actually experiment with the world will be able to
build better models.  Still, this is not necessarily the case.

As for the boundary problem, you can draw the boundary wherever you please.
In general, the further out you draw the boundary, the more grounded the
system is.

So, how can you build an AI as grounded as a human?  There are two
extremes: you can either give it high bandwidth sensors and effectors and
let it build its models the way people do; or you can build it more
accurate than a human and get by with more limited world experience -- the
narrower the bandwidth, the more accurate the system must be.

...mark young