Subject:
|
Re: AI in RCX
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Mon, 6 Nov 2000 07:03:26 GMT
|
Viewed:
|
669 times
|
| |
|
|
Hi! Sorry I missed out on the discussion... hope this is not too late to be of
some use.
I have to agree with Andy here. This type of thing is doable- not easy, but
doable. I've implemented some fairly substantial reinforcement learning
algorithms on the RCX, which are of similar mathematical complexity
to generic neural network programming. This approach does require patience- a
full run over the large matrices required for the RL problem takes ~5-10
seconds per step in my particular application, which means that the robot
moves, and stops and thinks for quite some time, and only then does it move
again.
Yiu, from what you describe, this means that (most likely, unless you are
very, very careful in avoiding dependency on floating point math) you will have
to train the robot first, and then just do pure reaction, if you want to do
real time. Learning and reacting together most likely won't be possible in real
time.
Of course, the math involved here means that legOS is probably the best tool
for this task. Yiu, you'll want to check out legOS.sourceforge.net for more
details on that. Anyway, good luck with the project!
Luis
P.S. If anyone wants more details on the RL problem that I tackled, you'll have
to buy our book:
http://www.amazon.com/exec/obidos/ASIN/1893115844/tieguyorg/
I'd love to publish it on the web but I'm afraid the contract won't let me ;)
Also, for those of you who have been waiting, I finally got my copy from the
printer, so Amazon should have their copies in hand shortly.
In lugnet.robotics, Andy Gombos <gombos_2000@yahoo.com> writes:
> I think that you should not immediately discount the idea of nueral networks
> in the RCX. Bert van Dam has done amazing stuff with networks on the RCX,
> and using no PC code.
> The point of the RCX is to try, learn, and invernt, not to say that it is
> impossible.
>
> Andy
>
> Rama Hoetzlein wrote:
>
> > Steve Baker wrote:
> >
> > > I think this is "do-able" as a project. [refering to Neural Networks
> > AI on RCX]
> >
> > I will believe it when I actually see it. Until then, it remains to me a
> > conceptual possibility (like all projects), but a practical
> > impossibiltiy.
> >
> > My primary reason for suggesting that the goal of the project be reduced
> > was to suggest how one could reconcile one's goals with practical
> > problems. By focusing on small, simple projects first (such as FSMs), it
> > is possible to gradually build up to larger projects, such as autonomous
> > neural-network lego robots, while still keeping difficult goals in the
> > distance. This is, of course, how to gain experience - which is much
> > more important to me that the final product.
> >
> > I do not doubt that neural networks on the RCX are possible in theory.
> > The question is, will the final result be successful - acceptably fast,
> > acceptably robust, acceptably working!. Without having the actually
> > results in hand, the only thing we have to go on is our imagination and
> > experience - which cannot serve as a practical measures of what is
> > possible.
> >
> > Therefore, although I will agree that neural networks on the RCX are
> > possible in theory... I will not agree that they are practical -- until
> > a see a successful working example.
> >
> > Rama
> >
|
|
Message is in Reply To:
 | | Re: AI in RCX
|
| I think that you should not immediately discount the idea of nueral networks in the RCX. Bert van Dam has done amazing stuff with networks on the RCX, and using no PC code. The point of the RCX is to try, learn, and invernt, not to say that it is (...) (24 years ago, 5-Nov-00, to lugnet.robotics)
|
3 Messages in This Thread:
 
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
