Low Complexity Robots

Michael Edelman mje at mich.com
Wed Nov 10 13:45:33 EST 1999


....
> >> In article [...], Steve <rstevew at armory.com> wrote:
> >>> This is Mark Tilden's BEAM robotics. It's malarkey.
> >>> Cute malarkey, but still malarkey. The thing keeps
> >>> thrashing is all, it doesn't think any more than a
> >>> pair of dice thinks. It just tries all the numbers
> >>> till something happens.

Not quite true. Tilden's stuff models the way real neural circuits work,
rather than trying to come up with a priori rules that approximate the
outcome of a nueral network.

One interesting thing about neural nets is that they learn, without
needing a seperate memory apart from the nuerons that generate the
movement. If you've even seen a Tilden 'bot fired up for the first time
you'll see it learn to walk *without* the need for an external
validation that says "yes, you're walking, now keep doing that". The
same elements that generate the movement learn the pattern from feedabck
from the environment.

Now, no one's made a "nervous net" that mimics human linguistic behavior
in the way that a lot of well-known top down AI programs do. But I would
argue that Tilden's 'bots teach us far more about how the brain actually
works than any top-down GOFAI program ever did.




More information about the Neur-sci mailing list