Re: New AI Prize

Fred Flinstone <[EMAIL PROTECTED]> wrote in message
news:[EMAIL PROTECTED]
> Why do we need a computer program that can understand a complex artificial
> world? Wouldn't it be better to have a machine that understood the real
> world?

A machine understanding of the real world would be useful,
but tricky to attain.

> Creating an AI and a complex artificial world is just making the
> problem even harder and the result would be useless.

Not at all. Complex artificial worlds have a lot of advantages
for training up of an AI, when compared to the real world.

First, the link between the AI and the world it occupies is just
software. Once it has been developed you can replicate it endlessly,
as long as you have enough computer memory and processing
power. For real world interaction you would need real physical
robots, and if you wanted e.g. to make them evolve to a high level
by natural selection you would need thousands or millions of them.
Not cheap.

Secondly, an artificial world can be sped up by piling on the
processing power. With enough fast processors you could in
theory get through thousands of generations every second.
This is significant because life took billions of years to get to
the level it has attained naturally, and scientists probably
wouldn't want to wait billions of years for AI entities to reach
the same level.

Thirdly there's replicability of situations. You don't know if
robot A is better than robot B in a certain situation. No problem,
you face them both with that situation. With software, you can
guarantee that both robots face exactly the same circumstances,
so any difference in outcome must be due to differences in the
way the robots behave. And you can perform extensive "what-if"
analysis. What if the robot had approached the situation more
cautiously? What if it had been able to move slightly faster?

Artificial worlds have been used in AI, but these worlds are
generally very simple, and they map to a very restricted subset
of the real world domain, such as rectangular blocks which can
support other blocks. Or topologically  toroidal (i.e. the edges
wrap around) 2-d "ponds" in which predators and prey interact,
and there are no features apart from the interacting organisms.
But I would argue that the simplicity and limited size of these
environments limits the possibilities for the development of
intelligence.


>
> I don't think a prize is necessarily a good thing, the smart people who
have
> a chance of creating AI will realize that it would be a gamble to spend
> their lives trying to create AI in the hope of winning the prize at the
end.

If there was a second prize, one for developing a suitable environment
in which AI entities could thrive, I think I'd be tempted to have a go.
But yes, it would be a gamble. And the problem with AI is how do you
decide when the goal has been reached?  That would be a serious
problem if there was a big money prize involved, one which various people
thought they had earned but the controllers of the purse strings disagreed.

> People have families to support and need a decent income. Money isn't a
> really a good incentive anyway, the peope who create AI will create it
> because they're genuinely fascinated by it. With the way things are going
I
> don't think AI is going to just pop up from out of a laboratory one day
with
> a fanfare blasting out, it's looking like its going to slowly evolve into
> existence.

Money can be a good incentive, after all it is what allows the people
to support their families. But there needs to be a balance, enough small
sums floating around to ensure that nobody starves while trying to
attain their heady goals, plus a few large incentives to spur them on.

I think true AI will evolve out of games, because of smart opponents
making for an improved gaming experience. Multiplayer games reduce
the need for emergent game AI's, but they'll still be needed e.g. for
slack periods when not many real people are logged on to a game, but
the few that are want to control large armies.

Martin Gradwell.

[ comp.ai is moderated.  To submit, just post and be patient, or if ]
[ that fails mail your article to <[EMAIL PROTECTED]>, and ]
[ ask your news administrator to fix the problems with your system. ]