Author: Laszlo Gaspar
Date: 07:01:33 12/07/05
Go up one level in this thread
On December 07, 2005 at 07:25:26, Tord Romstad wrote: >On December 07, 2005 at 05:31:51, Barry Culp wrote: > >>On December 06, 2005 at 23:11:38, Uri Blass wrote: >> >>>results of more matches >>> >>>Rybka(2 plies)-Movei(4 plies) 12.5-7.5 >>>Rybka(3 plies)-Movei(5 plies) 7.5-12.5 >>>Rybka(4 plies)-Movei(6 plies) 11-9 >>>Rybka(5 plies)-Movei(7 plies) 10.5-9.5 >> >>What's the point of running these handicap matches ? Please explain > >It is commonly believed that the difference between a search to >depth N and a search to depth (N+d) is less significant when N is >big. This seems intuitively plausible, but as far as I know we >don't have much experimental support for this hypothesis. > >Experiments like Uri's could help us to learn more about this. > >Tord Dear Tord and others, This is my first post on this forum but I would like to share my experience in this topic, since I also have an own engine and researched a bit in this direction. Interesting but I found that every ply increase in search gives an almost same ELO increase. This means that my weak engine which is let's say 450 ELO weaker than Fruit at depth=4 is almost equal at depth=6 (when Fruit remained at 4 ply) and better with depth=7 and so on. This increase seems to be different but constant in relation of two specific engine and it's around 200ELO, which is quite big. I think this is very important to know and it has quite a few consequenses: 1. Any dumb engine can be champion on a fast hardware. 2. You can improve your engine quite a lot if you improve the evaluation function only or tune your extensions. 3. You can neglect the evaluation function and concentrate on search only to reach higher depth. 4. Not important to search everything to high depth just the important lines(extensions, reductions). 5. Time management is important. 6. Testing development version of an engine can be more simple, since a fixed, low depth test match can be done quickly. 7. In my opinion the ELO rating can be estimated for any time control by simple maths though still I didn't figure out the exact formulation (we have to calculate the average search depth for a specific TC, which is not hard if we know the average branching factor). In the view of these I consider Uri's data very interesting and useful although the number of games is low to justify the aboves. Best regards, László
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.