Author: Ingo Bauer
Date: 12:45:56 07/27/04
Go up one level in this thread
On July 27, 2004 at 15:26:36, Bryan Hofmann wrote: >On July 27, 2004 at 14:57:02, Robert Hyatt wrote: > >>On July 27, 2004 at 13:32:54, Ingo Bauer wrote: >> >>>Hi >>> >>>>>Second round with Crafty having black just started and of course I deleted all >>>>>learning values from the previous round! >>>> >>>>Not sure for the reasoning for doing this. Humans are not allowed to learn >>>>anything when they play white and lose, so that they can use that information >>>>when playing black??? >>> >>>Of course I delete the learning for both engines!!! All I wanted to do is having >>>the same clean start for both engines. The engine that is having white would >>>have the possibility to learn something for its "black" game that the other >>>engine could not do when having black first. >> >>That logic is broken. If an engine plays black, it learns for "both sides". If >>an engine is white, the same thing holds. Disabling learning seems wrong, since >>it is a part of each engine, depending on how well it is implemented. Tuning >>bits of a program on or off on a whim seems somehow wrong unless the goal is >>_not_ to measure the strength of the entire "entity" but rather to measure the >>strength of a subset... >> >> >> (And, yes the black-first engine >>>could learn something for its white game, but who knows if that is identical?) >>> >>>>Why not just disable learning completely? >>> >>>Yep, you are right here. I could (and should) have done this. Have not thought >>>about it but do you think that deleting it after each round is doing any harm? I >>>am pretty sure that does not matter. I will do it for future games. >> >>For "Nunn matches" I don't think it matters since in theory, the same position >>will not be reached twice since each opening is different. >> >>However, the idea of keeping learning active makes sense since each program >>plays the same opening from both sides. What it learns from one side ought to >>influence it when it plays the other side, like a human... >> > >I'll add it does exactly that! Here is a small test I did with Crafty dealing >with the Learn off vs on. I ran a 5/2 match Crafty 19.15 vs Ruffian 1.0.5 using >the Nunn I & II openings and mirroring them so both engines played each opening >as White & Black. As you can see there is a 65 point difference in strength. >Since there were no books used in this match this represents the strength >difference of position learning only. > > >Crafty 19.15 Learn Off - Ruffian 1.0.5 : 17.5/60 9-34-17 >(==1=0000==000==00==110000000==010=1001010=01=00=0=0000=00100) >Elo : -154 >Margins : >68 % : (+ 47,- 33) -> [-187,-107] >95 % : (+ 90,- 70) -> [-224, -64] >99.7 % : (+131,-112) -> [-266, -23] > > >Crafty 19.15 Learn On - Ruffian 1.0.5 : 22.5/60 13-28-19 >(=0==0100==01==01==010=000000===101=011110=01==1=0=0000010000) >Elo : -89 >Margins : >68 % : (+ 39,- 37) -> [-126, -50] >95 % : (+ 76,- 78) -> [-166, -13] >99.7 % : (+113,-123) -> [-211, 24] Nice, What about a match with disabled learning for both engines now? That result in comparison with learning on for both would be intersting! Of course it is cheking how good the quality of lerning is but you are doing allready 60 games in comparison to my 40 the difference in learning capability is bigger. If there is a big difference I will appologize myself to Bob and do my testing anyhow if he agrees or not to my proposal! Bye Ingo
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.