Re: DARPA Grand Challenge

Traveler <traveler@xxxxxxxxxx> wrote:
> On 20 Sep 2005 18:29:45 GMT, curt@xxxxxxxx (Curt Welch) wrote:
>
> >DARPA doesn't give a *** if the machine is intelligent.
>
> I doubt this very much. The US defense department would sell its soul
> to the devil for a truly intelligent system.

True.

> It's just that those in
> charge don't think it's possible at this stage of the game but they
> are wrong.

Right, they are trying to buy what they think people can deliver this year.

I think the type of approach you and I are working on will allow a single
lap-top to out perform the racks of computers using the hard-coded number
crunching algorithms that we are seeing being used in the contest. But
until someone gets a general solution working on a real application like
the DARPA problem then it's all just hot air.

The reason I think a general solution will work better, with less
resources, is because the current approaches are over-analyzing the
problem. They are doing things like creating a complete 3D model of the
environment from the sensory data and updating it many times per second (10
or greater?). This model allows the software to know exactly how far away
every object in the immediate environment is to some resolution measured in
inches or less. When I drive a car, I can guarantee you my brain doesn't
have any idea how far away a rock is, or the edge of the road. Instead, my
brain has learned how to react to the changing view of the rock (and
everything else in front of the car), with a corresponding movement of the
steering wheel. It's a dynamic control problem worked out by practice. We
get an intuitive sense how far away the things are by how fast they are
changing shape and we regulate the speed of our reaction with our sense of
that change. No where in there does the brain actually need to figure out
that the rock is 12.23 meters away from the car. It only needs to figure
out when to turn based on the changes it sees happening in the view of the
road.

General AI will map the sensory data, straight to the steering wheel
control data, without any unneeded "middle terms" like a 3D model of the
surrounding environment. And in doing that, it will end up using a
significant less amount of total computation power. Humans just aren't
smart enough to program these optimized control solutions manually. They
have to be self optimizing systems which is what general AI will be. They
will naturally tend to find, and use, the optimum solutions first.

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.