Shane,

But it seems to me that modeling the physics of sensors and actuators
isn't the main problem -- there are real problems of computational
resources, if one wants to simulate, say, the complexity and diversity
of a forest.  I believe the approach you describe works to simulate the
experience of a robot moving around in a lab environment, but by
eliminating the rich diversity of a natural environment, this approach
may be missing the primary point about why embeddedness in the world is
valuable for a learning or evolving AGI.  In a forest there are so many
different things with so many different behaviors, that can be combined
in so many different ways.  I think that a non-physical-environment-like
simulation with this degree of diversity, would be more useful for a
baby AGI than an accurate simulation of a nondiverse, stale, lab-like
physical-world environment. 

-- Ben G

> -----Original Message-----
> From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On
Behalf
> Of Shane
> Sent: Sunday, March 07, 2004 9:22 AM
> To: [EMAIL PROTECTED]
> Subject: Re: [agi] virtual worlds
> 
> 
> > It's true that making a virtual world with the richness of the real
> > world is damn tough.  I agree that giving an AGI access to
real-world
> > sensors and actuators is important, yet I also believe that a lot of
> > teaching and learning of AGI's can be done in simulated worlds.
IMO,
> > the practical hassles of robotics are larger as the benefits, for
the
> > current stage of AGI work.  However, this is largely an issue of
> > financing -- with bucks to hire a team of robotics engineers to
solve
> 
> Yes, in fact this is a common "issue" in AI.  Some of the guys where I
> work are currently addressing this problem by building simulators
which
> contain full physics models.  Thus you don't just have an agent in the
> environment, but rather you have a fully simulated virtual robot -
even
> the sensors are simulated. The virtual robot has a physical version
too so
> it's possible to set up a virtual and an equivalent real world
experiment
> and then compare the results to check that the two match up.  Of
course
> the
> idea is that when it's finished you will be able to quickly and
cheaply
> set up detailed virtual experiments and be pretty sure that if you
used
> the expensive real robot you would get almost identical results.
> 
> Shane
> 
> -------
> To unsubscribe, change your address, or temporarily deactivate your
> subscription,
> please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

-------
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Reply via email to