Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Diminishing returns

Author: Anthony Cozzie

Date: 07:24:59 04/29/04

Go up one level in this thread


On April 29, 2004 at 09:26:21, Tony Werten wrote:

>On April 29, 2004 at 08:40:03, Anthony Cozzie wrote:
>
>>On April 29, 2004 at 03:13:07, Tony Werten wrote:
>>
>>>Hi all,
>>>
>>>a while ago we had some discussions about diminishing returns in search for
>>>chess.
>>>
>>>My opinion was that you can't prove that with programs searching d vs d+1 ply
>>>depth because the advantage of the d+1 program gets smaller. ie at d=1 it has a
>>>100% depth advantage, at d=2 it's 50% etc.
>>>
>>>Some people claimed that you can't compare it that way because bla bla
>>>exponential something bla :)
>>>
>>>Well, I found an easier way to explain it.
>>>
>>>A few assumption:
>>>
>>>The easiest way to win is when you see a trick, your opponent doesn't see.
>>>
>>>The depth that needs to be searched to see a trick is equally divided. ie there
>>>are as many tricks hidden 1 ply away as there are tricks at 2 ply ( it doesn't
>>>really matter but it's easier to visualize )
>>>
>>>w is player d+1
>>>b is player d
>>>
>>>d=1: b sees tricks 1 ply away, w sees ply 1 and 2 => w sees 2.0x as many tricks
>>>d=2: b:1,2 w:1,2,3 => w: 1.5x
>>>d=3: b:1,2,3 w:1,2,3,4 => w: 1.3x
>>>...
>>>d=10: b: 1..10 w: 1..11 => w:1.1 x
>>>
>>>
>>>
>>>Conclusion: There may or may not be diminishing returns in chess, but d vs d+1
>>>are not going to prove it, because those matches by itself are a clear example
>>>of diminishing returns regardless what game is played.
>>>
>>>disclamer: I know chess isn't only about tricks, but it is an advantage to see
>>>more of them then your opponent. Clearly the win percentage is depending on
>>>other (random) stuff as well. BUT When you see less more, the advantage becomes
>>>less.
>>>
>>>Tony
>>
>>I'm not even sure I agree with "the tricks are equally divided".
>
>Me neither, but for the thinking it is easier.
>
>>It would be
>>possible to get some sort of statistics for this, but my guess is trick %
>>declines with depth :(
>>
>>Even so, just the extra positional help makes more depth worth it IMO.
>
>It's the same. Replace "depth to see tricks" with "depth needed to find the
>correct positional move" and the same story holds.
>
>Don't misunderstand, I know an extra ply of depth gives more strength. But every
>extra ply will give less extra strength. Not because of chess ( or checkers or
>whatever ), it's just a logical consequence that comes from search.
>
>Maybe not even search alone.
>
>If you drive 10 km/h faster than me, it makes a big diffence if I'm driving 10
>km/h. You will get there twice as fast. When driving 50, the difference is a lot
>smaller. Certain researchers should call that "diminishing returns in car
>driving" if they are consequent.
>
>Tony
>
>>
>>anthony

I think you are right, I was just pulling a Uri and nitpicking :) Also, I think
the diminishing effects of search are greatly tied to eval.  If you have a
perfect eval, any search beyond depth 1 is wasted. It seems logical that the
better your eval, the less you need to search.  If you have a pst program, you
will probably get a big boost from going from 18 -> 19 ply.  This is why I think
it would be very interesting to redo the move-change-with-depth experiment.

IMHO, search is becoming less interesting _if_ you have a good parallel search
and a good eval.  With a good implementation of DTS, it seems like it is
possible to get a 6X speedup on an 8-way opteron.  This means that even a slow
program will be able to get 13-14 ply without forward pruning.  A faster program
will be getting 16+. And those are middlegame depths - it would be more like 20
and 25 in the ending.  That finds almost anything :)

anthony



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.