Subject:
|
Re: Design
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Tue, 6 Dec 2005 14:03:09 GMT
|
Original-From:
|
PeterBalch <PETERBALCH@COMPUSERVE.COMstopspammers>
|
Viewed:
|
1168 times
|
| |
|
|
> dan miller wrote:
> Why the negativity about constrained or simulated
> worlds? We are not Nature; we are not constrained to work in a design • space
> that never changes.
I think that the whole AI community was chastened by SHRDLU (or, at least,
99.9% of them were). It seemed like a huge success: "Why did you clear off
the blue cube?" "To put the red pyramid on top of it." It looked like a
real human-language interface. All it needed was to be scaled up. A new,
faster computer would do the trick.
Nowadays, you won't find a single AI worker who thinks that. The received
wisdom is that SHRDLU was dealing with a toy world in a way that didn't
scale well.
New algorithms have replayed the same story over and over again in the last
40 years. Yes, everyone starts by trying their new algorithm in a toy world
and it looks great. But then they try to apply it to the real world and it
doesn't work well. So they tweak it and tune it hoping that just one more
adjustment will let it live up to its promise. And six months later they
give up and admit that maybe a different approach would work better. I've
been down that road time and time again and so has every AI researcher.
Only 1 in a thousand algorithms moves successfully from a toy world to the
real world.
It's time we admitted that it's a research paradigm that doesn't work.
Hence the negativity.
(So whay am I spending time trying to write PlaniSim to simulate a "real"
2D world? 'Cos I don't learn from my mistakes :-).)
> Animals very clearly live in 'constrained environments'. Obviously fish
> can't fly, ants can't live underwater, etc. So although they live in the
> real world, the domain they are expected to function in may be very
> specialized.
Yes there's no clear dividing line between what I'm calling "toy worlds"
and the "real world".
The difference is that the real world is dirty while toy worlds are clean.
The real world is totally bl**dy minded. It throws counter-examples and
pathalogical-cases and things you hadn't thought of at you so fast that
they far outnumber the cases you'd prepared for.
(Last night, I was trying to simulate a dynamically-balanced hopping-robot
in PlaniSim. I gave it some very slightly rough terrain so it wouln't be
too easy a problem. What I hadn't forseen was that when it hopped, the
sideways force on the foot was often such that the foot slipped and then,
with luck, would catch on a bump. After a hundred such slips, the "electric
motor" I was using to rotate the counter-balance was going so fast it
couldn't provide the torque needed to move the leg quickly. A good
toy-world can include enough real-world dirt to catch me out like that but
if I built a real hopper I'd hit a hundred times more problems.)
> fish can't fly,
> the domain they are expected to function in may be very
> specialized.
That's something I worry about.
I believe strongly that building mobile robots to operate in toy worlds is
a waste of time. (I do it for fun, not because I think it has a future.) I
believe we should emulate animal evolution by attempting to always situate
them in the real world. We know that approach worked in the past. And yet
...
Life evolved in water. Have we chosen a domain that's too hard? Animals
didn't attempt the land for a long time. The standard biological excuse is
to do with dessication and UV but maybe land is very much harder for
locomotion. Maybe we're trying to walk before we can swim.
Peter
|
|
Message has 2 Replies:
 | | Re: Design
|
| (...) I had a question from a student the other day (I teach Biophysics at a university) who asked why engineering doesn't follow the same mechanism as evolution (if evoution works so well, why not emulate it?). The brief answer is becuse, for (...) (19 years ago, 6-Dec-05, to lugnet.robotics)
| | | Re: Design
|
| (...) I don't have the dialog verbatim - but something like: Human : Are there any steeples in the box? SHRDLU: I don't know what a "steeple" is. Human : Define a steeple as a cube with a pyramid resting on top of it. SHRDLU: OK Human : Are there (...) (19 years ago, 7-Dec-05, to lugnet.robotics)
|
5 Messages in This Thread:
  
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
