[ExI] spike turing test, was RE: ai class at stanford

[Kelly Anderson]

On Thu, Sep 1, 2011 at 10:45 PM, spike <spike66 at att.net> wrote:
>>... On Behalf Of Kelly Anderson
> Subject: Re: [ExI] spike turing test, was RE: ai class at stanford
>
>>...So here's a question.... which would be easier to program? The generic
> Turing test "I'm pretending to be SOME generic human, and you can't tell the
> difference" or the person specific test "I'm Paris Hilton, and you can't
> tell that I'm not"... -Kelly
>
>
> To further cloud an already broad and difficult question, there is a WW2
> fighter game I only played a couple times, with another guy who is an
> excellent fighter pilot gamer.  The mission is to guard a formation of B52s
> with a squadron of fighter planes, when suddenly you are jumped by a
> squadron of Nazi 109s.  A wild air battle ensues, but the part I find
> interesting is that you can play single human or multiple humans, as
> participants along with software allies.  You can review the battle
> afterwards and watch from any plane's point of view, including enemy planes.
> It is impossible to tell the difference between the human pilots and those
> commanded by the computer.  You can't tell from their battle strategy or
> their tactics or their flying skill.
>
> I suppose the computer guided planes use lookup tables derived from how
> humans have played during the development phase, but the point is that with
> a sufficiently large lookup table, a computer is indistinguishable from a
> human in certain settings.  Perhaps much of the time we function as enormous
> lookup tables.  On typical internet chat groups, perhaps the carbon based
> lookup tables aren't even very sophisticated.  That game setting is a form
> of the Turing test perhaps, and the computer passes.

I don't know if you could tell whether chess moves were generated by a
person of a good computer program most of the time either... Perhaps
you could, if the human weren't very good... LOL

Large lookup tables can be done, but are highly inefficient. I'm sure
they have a slightly more elegant solution than that. You could
probably codify a pretty good fighter pilot with less than 100 rules.

1) Try to get behind the other guy (at least for WWI and WWII era fighting)
2) Try to stay out from in front of the other guy.
3) Shoot where the other guy will be when your bullets get there

and so forth. The results of following such rules are generally pretty
good, and human like. Add in a little randomness at the end... and
you're good to go. The more constrained the context, the more amenable
it is to programming. That's why AGI is so hard, because it doesn't
have enough context.

-Kelly