DPRG mailing list archives: [DPRG] SRS thread

Mon Apr 30 15:36:05 CDT 2007

On Mon, Apr 30, 2007 at 01:30:55PM -0500, dpa wrote:
> "AI is a wonderful study. It has the grand goal of making philosophy an
> empirical science. Long may it wave. However, it has not been
> particularly useful in designing robots."
> 
> I agree.
> 
> Discuss.
> 
> And then the last paragraph:
> 
> "I believe the design model for the robot is not a synthetic human
> with independent thinking, but a sheep dog that acts as an agent
> for the shepherd. The sheep dog has autonomy (like a thermostat)
> without independence."
> 
> Those familiar with my own vapid ramblings will recognize the same
> model.  I usually say to think of robots not as "replacement humans"
> but rather as "intelligent tools."  But it is the same basic concept.

I have two responses.

First, I think "making philosophy an empirical science" is a dishonest
summary of AI which seems to have been constructed for the purpose of
sneering at it, which I find distasteful. That the followup statement,
  > However, it has not been
  > particularly useful in designing robots.
happens to be true doesn't justify the preceding dishonest summary.
If someone were to write
  > The robot community aims to build useless worms. They are not there 
  > yet, though they believe it may happen. Long may they squirm...
  > but their ideas have not been particularly effective in converting 
  > Moore's law into economic productivity.
then even if that claim about economic impact was correct, the
paragraph would be a flaky and pointlessly nasty thing to write. Robot
builders not only aim to build more than worms, they've in various ways
already built much more than worms.

(I'm not actually sure that robots in the thermostat style are less
economically important than AI-ish ideas, so perhaps my true fact
after snide slam isn't actually a true fact. However, my impression is
that currently AI-ish work has a bigger impact. It's hard to gauge the
economic effects of some AI-ish ideas like data mining, but they seem
to be significant. I also don't know how much of economically
important niches like IC verification can be honestly credited to
techniques developed for related AI problems like theorem proving, but
it might be significant. And I'm pretty sure one particular AI-ish
field, planning, is very significant, enough that it alone may well
exceed the impact of robots. My impression (mostly from the business
press) is that complicated logistics stuff is essentially always
arranged with planning software, and that that is because the
efficiencies are significant (not just 1%, but perhaps 10% or more).
Do robot gains rival a 10%-or-so gain in efficiency in a big sector of
the economy yet?)

AI might be more fairly characterized (and criticized) as emphasizing
the characteristics that we think of as human to the exclusion of
things that we disdain to worry about because lizards and insects get
them right. But even if there's been some arrogantly blindered
nonsense from that perspective, there's also been an impressive amount
of useful progress, not just in the economic impact I mentioned but in
things like face recognition, speech recognition, and game playing.
And there's no particular "philosophical" about it (faces? chess?
planning?) and if anything the robot builders might be justified in
criticizing it not for being empirical but for insufficiently
empirical to be relevant. (Screwing around with theorem proving, e.g.,
what's up with that?)

Second, in robots as in many other fields, you see people solving
problems by first reducing messy reality to some idealization which
preserves important features of the grotty real-world problem problem,
then trying to get a useful answer from the idealized problem. This
approach doesn't always work, but it has been so successful on so many
problems that I think it's worth considering it as a candidate
approach even when it hasn't worked terribly well so far. I think you
can legitimately criticize many AI people for horribly underestimating
the difficulty of the first step in the robot problem, an
underestimate which is difficult to excuse when you look at how much
work nature puts into the first step. But even if AI has had more than
its share of arrogant idiots, I don't think it's safe to wave away its
work on the second step for your intelligent tools. For example, even
with the amount of computer power we have, problems like learning and
visual navigation are probably practical: look at bees. Dismissing the
existing AI work on that, or hoping to surpass it effortlessly with
one's judiciously non-AI design innovations, might not be wise.

I think an only-usefully-provocative position might be something like
"far too little of the AI stuff is useful until you have predigested
the world into a relatively clean model, and in robotics that
predigestion is perhaps impossible, and certainly so hard (and so much
harder than the mainline AI tradition acknowledges) that their work is
largely irrelevant to robots." (And if you want to be really careful,
you could say "today's robots.") But given the successes of AI-ish
approaches in areas where the predigestion has proved practical (like
trucking manifests), in other areas where the predigestion was there
from the get-go (like playing Chess, or verifying the proof of the
Four-Color theorem), and in some who-asked-for-that surprise areas
(like face recognition), and given the existence proof of humans
dominating a world full of cleverly designed less-AI-ish autonomous
tools, I think it's a bit early to dismiss high-level thinking. Does
anyone have a good track record in predicting even a decade in advance
which AI-ish things will finally become practical as computer power
and algorithmic ideas advance? The optimists are regularly embarrassed
--- but the "not in my lifetime" answer hasn't been embarrassment-free
either.

-- 
William Harold Newman <william.newman at airmail.net>
PGP key fingerprint 85 CE 1C BA 79 8D 51 8C  B9 25 FB EE E0 C3 E5 7C
Ubi saeva indignatio ulterius cor lacerare nequit. -- Jonathan Swift's epitaph