Subject:
|
RE: Antenna Bumpers/Insects
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Sat, 5 Dec 1998 13:11:04 GMT
|
Original-From:
|
Carl G. Schaefer, Jr. <cgschaef@futurelinkinc.*NoMoreSpam*com>
|
Viewed:
|
1413 times
|
| |
|
|
<SNIP>
> But on a different tangent - what would be needed to run a neural net on
> an RCX? What kind of processing model is involved when you say "a
> hundred node net"?
Because of the limited on-board computational resources, IMHO you would have
to use offline simulation to train the neural network and then embed the
trained net, provided of course, that the resulting weight and bias matrices
and the multiplication/summing code could fit into the small memory
footprint of the RCX., yet still provide room for all the other overhead
required to make the thing go.
> Last time I checked in on neural nets (now too long ago) I conculded
> that they were basically fraudulent in that they modeled no time domain
> phenomena and could not be used in a control situation (while bioneural
> computation seems to be mostly *about* timing). Evidently that
> situation's changed - what are these things we're discussing?
There are a number of spatiotemporal learning paradigms being investigated,
particularly for just this type of application (neurocontrollers). I've
done some research at George Mason University on what we call amorphous
neural networks, which are highly recurrent neural networks with no regular
structural topology. The amorphous neural network structure (topology) and
the weight/bias space were optimized using two passes of a genetic
algorithm. The resulting networks are not only highly recurrent, but
provide very interesting spatiotemporal characteristics not unlike
deterministic chaotic systems.
> *Would* they fit (not in nqc, but in something else)?
IMHO, not in the RCX, but something maybe larger, like the Jones/Flynn
68HC11-based Rug Warrior mobile robots, might be able to handle it.
Carl
|
|
Message is in Reply To:
 | | Re: Antenna Bumpers/Insects
|
| Hey, I vote that gait discussions are definitely *on* topic - a lot of my thought about LEGO robotics in the past has related to mobility! But on a different tangent - what would be needed to run a neural net on an RCX? What kind of processing model (...) (26 years ago, 5-Dec-98, to lugnet.robotics)
|
3 Messages in This Thread:
 
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
