Author: Joachim Rang
Date: 04:03:47 12/13/05
Go up one level in this thread
On December 13, 2005 at 06:01:45, Chrilly Donninger wrote: >> >> >>Hi Chrilly, >> >>I find that hard to believe. One setting scores 40% against Rybka and another >>75-80% but at the same time the latter scores less against Shredder? You did not >>wrote by how much the more selective search scores better against Shredder but >>nevertheless in my tests I never experienced something like that. >> >>I often experience the same like Vas, that a tuning shows significant >>improvement in selfplay but no significance against a broad range of opponents. >>But I never observed that something which helped in selfplay hurts against other >>engines. It may be without effect against others but not worse. >> >>So far I was never able to tune Fruit against a specific opponent which resulted >>in a signficant difference not to say in doubling the score (40% to 80%). So >>either I simply lack the understanding how to do this or you played not enough >>games to get a statistical significant result and observed just noise. Given the >>little time between your first post about Rybka and this one I give both >>possibilities the same probability. ;-) >> >>regards >> >>Joachim >> >One explanation could be: As a tester you can only change some settings. But I >changed the search on the software-side almost completly. It was additionally >tuned against Shredder to play very aggressive. This worked quite good and gave >>80% against Shredder. The standard-version is - on a single - 70-75%. >Against Rybka this over-aggressive style did not work properly. But the >standard-version which is more tuned had no problems with Rybka. > >Chrilly While I find such a phenomenon not too unlikely I'm a bit surprised by the figures - 40% vs. 75% is rather huge. Furthermore I would suspect that changes in eval would have bigger effect on different opponents than changes in search. However if you mimic one pruning scheme it is clear that the opponent with the same pruning scheme can't find the holes in it - another opponent can. I'm also surprised that history pruning did not work in Hydra. In Fruit it works against all different opponents (no easy to exploit holes). But I'm beginning to sound like Uri so I stop here... ;-) Joachim
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.