Arimaa Forum (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
Arimaa >> General Discussion >> Would "DeepBlue/Arimaa" beat the best hu
(Message started by: Cobra on Apr 7th, 2009, 1:22am)

Title: Would "DeepBlue/Arimaa" beat the best hu
Post by Cobra on Apr 7th, 2009, 1:22am
Would a DeepBlue-style computer running the best Arimaa bots with parallel search and hardware optimizations beat the best human players?

Perhaps another way to explore this question: If you increased the speed of the best bots by 1, 2, 3, ... orders of magnitude, would they beat the best human players? How many orders would it take?

Just curious what thoughts people have in this area.

Title: Re: Would "DeepBlue/Arimaa" beat the bes
Post by tize on Apr 7th, 2009, 6:03am
To be able to beat the best humans I think the bots of today need to search 5 ply, at least. And I think they search about 3 ply on a normal midgame position (based on that marwin searches 2-3 ply midgame).

This means that the hardware need to be about 20k times faster.

But this is just my guess...

Title: Re: Would "DeepBlue/Arimaa" beat the bes
Post by Fritzlein on Apr 7th, 2009, 9:12am
That's an interesting question.  Unfortunately, we have very little data, so any answer is mostly speculation.

My personal expectation is that every time you double the speed of hardware, the software player running on it gains 50 to 100 Elo points.  Taking hardware that is 1000 times faster than the Challenge hardware gives you ten doublings, for a gain 500 to 1000 Elo points.  Given that clueless is performing around 1950 and chessandgo around 2500, our advantage is only 550 Elo points, so a supercomputer could very well surpass our best.

However, everything in the above sentence is suspect.  The rating gain per doubling might not be linear; the gain might be linear but not be in the 50 to 100 point range; the supercomputer might be more than 1000 times faster than challenge hardware, and chessandgo might be more than 550 points ahead of clueless.

Therefore you might as well choose whichever answer you would like to be true, because either way you can defend it.  :)

Title: Re: Would "DeepBlue/Arimaa" beat the bes
Post by jdb on Apr 7th, 2009, 2:25pm

on 04/07/09 at 09:12:11, Fritzlein wrote:
That's an interesting question.  Unfortunately, we have very little data, so any answer is mostly speculation.

My personal expectation is that every time you double the speed of hardware, the software player running on it gains 50 to 100 Elo points.  Taking hardware that is 1000 times faster than the Challenge hardware gives you ten doublings, for a gain 500 to 1000 Elo points.  Given that clueless is performing around 1950 and chessandgo around 2500, our advantage is only 550 Elo points, so a supercomputer could very well surpass our best.


To add my speculation to your speculation,

I would guess a doubling in speed is worth much closer to 50 elo than 100. Having said that, even 50 elo points is a major improvement.



Arimaa Forum » Powered by YaBB 1 Gold - SP 1.3.1!
YaBB © 2000-2003. All Rights Reserved.