Author: Robert Hyatt
Date: 09:06:48 06/02/04
Go up one level in this thread
On June 02, 2004 at 07:11:04, Uri Blass wrote: >On June 01, 2004 at 11:50:11, Robert Hyatt wrote: > >>On June 01, 2004 at 04:10:01, Uri Blass wrote: >> >>>On May 31, 2004 at 18:33:30, Robert Hyatt wrote: >>> >>>>On May 31, 2004 at 18:29:09, Matthias Gemuh wrote: >>>> >>>>>On May 31, 2004 at 15:08:51, Robert Hyatt wrote: >>>>> >>>>>>On May 31, 2004 at 13:41:39, Miroslav Nikolic wrote: >>>>>> >>>>>>>Hardware: PIII 800MHz, 256Mb RAM >>>>>>>OS: Windows Xp >>>>>>>GUI: Chessprogram8 (Fritz 8) >>>>>>>Tournament: 10x8x40min, round robin, 8 rounds >>>>>>>Time control: 40'/40, 40'/40 + 40' (rest) >>>>>>>Hash: 32 MB >>>>>>>Ponder: off >>>>>>>Resign: on >>>>>>>Tablebase: Nalimov 4-pieces >>>>>>>TB Cache : 6 MB >>>>>>>Book used: DeepFritz7.ctg for Deep Fritz, H8 for Hiarcs, Select.ctg (by me) >>>>>>>for others >>>>>>>Book learning: off >>>>>> >>>>>> >>>>>>Just a question: Why would you want to turn off a part of a chess program that >>>>>>is not easy to develop in the first place? IE book learning is a part of many >>>>>>engines. Turning it off makes no sense to me... any more than turning off >>>>>>passed pawn evaluation or selective search capabilities... >>> >>>I think that it is better if programmers enable also option to turn off passed >>>pawns evaluation and other parameters. >>> >>>In the last version of movei that I still did not release the user can change >>>the passed pawn evaluation by changing weights including disabling them if you >>>change the relevant weights to 0. >>> >>>I will not be surprised if it is possible to find a better personality by >>>changing weights. >> >>In current crafty, you can change weights, or you can change each specific eval >>term if you so choose. But using the _Default_ book, which depends on learning >>to cull bad lines, and then disabling learning makes no sense whatsoever... >> >> >>> >>>>>> >>>>>> >>>>> >>>>> >>>>> >>>>> >>>>>Strange logic ! >>>>>How can one compare book learning with passed pawn evaluation or selective >>>>>search ? >>>> >>>>Very simple. >>>> >>>>1. I added passed pawn logic to address weaknesses that needed help, in >>>>Crafty's particular playing style. >>>> >>>>2. I _specifically_ designed the opening book around book learning. I have >>>>_no_ hand-tailored lines, my book is made by sucking in reams of PGN data and >>>>then relying on learning to discover which lines are bad or unplayable. >>>> >>>>Is this hard to understand? If I had a hand-customized book, I wouldn't care, >>>>but I don't, and turning off the learning facility simply makes _zero_ sense... >>>> >>>>>To make sure that the outcome of a tournament does not depend on order >>>>>of opponents, book learning has to be switched off. >>>> >>>> >>>>That's nonsense. Does a human have to do a "brain purge" between opponents? >>> >>>We cannot do it with humans but we can do it with engines and it makes better >>>comparison. >> >>How. I have a bad line in my book. Do I _really_ have to play it multiple >>times against each opponent to make a better comparison? Or do I have to take >>the time to hand-edit each "learned result" I get so that the book will be as >>good as possible??? >> >> >>> >>>If you want to compare between different versions when the change is only in the>evaluation then learning add varaible that is not relevant and it is better to >>>compare results when learning is off. >>> >> >>I don't see why. It introduces random noise into an experiment, repeatedly, >>when the program plays an opening it would normally have learned was bad... > >Suppose that I do some change in the evaluation and have no learning and a very >small book. > >Suppose that I want to test if the change was productive by playing games >against Another program. > >Learning is the thing that produce unnecessary noise. > >opponent that learns and is not significantly stronger if you test by only >playing 2 games against different opponents may be lucky to win the first 2 >games and repeat them again and again to get 100% when without learning the >result gives me better information. > >Uri You are doing the _same_ thing Sune did. I did _not_ say _you_ should not disable learning when you are testing _your_ engine. I said "disabling learning in a match or tournament makes absolutely no sense." And it doesn't. I disable learning all the time here so that position learning doesn't screw up test results, for example. But you have _never_ seen those match results posted, nor will you, because they are not matches to compare two programs they are matches to debug something. There anything that helps you is fine. But not in a basement tournament...
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.