Author: Vasik Rajlich
Date: 05:11:35 12/25/05
Go up one level in this thread
On December 24, 2005 at 21:43:19, Albert Silver wrote: >On December 24, 2005 at 18:47:48, Rolf Tueschen wrote: > >>Perhaps a stupid question, but let me ask it for my own interest. You give a >>couple of types with different results and then after 160 games each you get >>different results (one time some 50 points difference) and then I've read the >>item "ponder=off"; do you know something about the importance of that factor? >>I'm thinking about the importance of that aspect for say FRITZ with its >>tradition or better experience. Please dont go into all details, just give me a >>short estimation for the importance. Without knowing the answer, I see three >>critical points: ponder off, 160 games and difference of 50 points between two >>special types. > >Relax dude. That other part was quite unnecessary. For anyone playing with a >single computer or single cpu/core, "ponder=off" is obligatory. What it means is >that the engines do not think during the time of the opponent. Otherwise one >cannot know if one engine is using the CPU more, if there is a problem in the >equality of the usage of the CPU, etc. Nowadays, there are even dual-core >processors (equivalent to two separate processors in one - the AMD Athlon65 X2 >series is an example), so that people can run engine matches with Ponder=On, but >my processor is not one of them. > >The engines do have separate and equal-sized hash tables (256 MB for me, which >is fine since I have 1 GB Ram total), and when it is their turn to think and >play, they make use of the previously calculated hash tables. > >In theory, there should no change in Elo performance using "ponder=off". The >conditions of the matches I reported are this: > >I tested 3 Rybka settings ('very positional', 'slightly positional', and >'slightly tactical') against identical opponents (Deep Fritz 8, etc.) using >identical openings (the Nunn2 set), in order to give each setting the exact same >conditions, and hopefully better show any performance differences, if any. > >The Nunn2 set, in case you don't know, is a set of 20 opening positions chosen >by GM John Nunn, in which 2 engines play 2 games of each position, once as white >and then as black. So 4 opponents for each setting means 4 matches of 40 games, >or 160 games. > >One can argue that the positions may favor one engine more than another, but the >positions are designed to try and provide a variety of typical types of >opening/middlegame problems to solve, and remember that the engines get to play >both sides of each position. > >I am starting another series to test the Rybka settings against Junior 8 (don't >have J9), and see how it does. > > Albert Of course there was nothing wrong with your tests. I just want to add one thing. Technically, 160 games is not enough - your variance is bigger than the performance difference. In practice, though, I will often make decisions based on even less data than this. If I like a certain change intuitively, and it jumps out to a decent lead, I'll just keep it. Unscientific regards, Vas
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.