Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> General Discussion >> Next challenge for AI?
        
(Message started by: ddyer on Nov 16^th, 2006, 6:27pm)

Title: Next challenge for AI?
Post by ddyer on Nov 16^th, 2006, 6:27pm

I'm going to throw just a little bit of a bomb, but I'm
completely serious.

Hundreds of man-years (at least) have already been
expended trying to produce a Go program with little
to show. Go programmers have had the advantage
of a well-developed theory of the game, and hundreds
of genuine experts to draw techniques from; which
ought to help. Go already is the next challenge,
and not one that is likely to be met soon. Furthermore,
there was a bona-fide million dollar prize for a shodan
Go program - which I think has expired because the
person offering it - umm - expired.

So maybe Arimaa is an interesting challenge, maybe the
next to be taken up and met, but it's not the challenge.

Title: Re: Next challenge for AI?
Post by Fritzlein on Nov 16^th, 2006, 8:22pm

Sure, Go is a fantastic game. It has stood the test of time as a human activity, and it has proven to be very computer-resistant. Go has (or had) a huge challenge prize, and despite widespread efforts, that prize was unclaimed. Moreover, the Go challenge was not for beating the World Champion, but only for beating a shodan. The top Go players may be 14 ranks ahead of the top Go computers. In comparison Arimaa has a small prize, a small pool of developers, and the top human at Arimaa is merely 2.5 ranks ahead of the top Arimaa computer.

That said, do you think Go is inherently more computer-resistant than Arimaa? Consider the billions (yes, billions) of man-hours that have gone into human study of Go. Suppose you learned the rules of Go, and got together with a few dozen other folks who were also just learning, and played against each other for four years, with no outside assistance. Would even one of you achieve shodan level in Go? When I say no outside assistance, I mean no teachers, no books, no master games to study. Just by knowing the rules and teaching each other, how deep into the knowledge of Go would your hardy band penetrate? If the best of you reached 3 kyu, you would have a smaller lead over the top Go computer than humans now have over the top Arimaa computer.

At the moment, we human players know very little about Arimaa. How could we know a lot? There are no teachers, no books to read, no theory but what we create ourselves. Among the things we don't know is how deep the water is. Our ratings span about 6.5 ranks: are there inherently ten ranks of depth within Arimaa? Twenty? As many levels as Go?

I submit that, while we know that Go is very deep and very computer resistant, we also don't know if Go is the strategy game AI challenge. We know so little about Arimaa, we can't even know that Go will elude computers for longer than Arimaa does.

Title: Re: Next challenge for AI?
Post by ddyer on Nov 16^th, 2006, 8:38pm

I won't quibble - I think we're pretty much in agreement.

My intuition is that Arimaa will prove to be a much easier
AI problem than Go, at least because the computer side
can be advanced much faster than the human side.

I think it is widely agreed that the successful computer
programs for Chess are not AI programs at all, they're
merely very fast and clever at counting wood. Arimaa
looks like it's well designed to thwart that approach.