Author: Robert Hyatt
Date: 15:03:45 08/05/02
Go up one level in this thread
On August 05, 2002 at 17:39:46, Vincent Diepeveen wrote: >On August 04, 2002 at 17:00:29, Jay Scott wrote: > >>On August 04, 2002 at 13:26:05, Vincent Diepeveen wrote: >> >>automatic tuning with temporal differences: >>>it's a rude and primitif algorithm which can be considered lucky if the plus >>>and minus sign are already guessed right. >> >>Gerry Tesauro might have something to say about that. :-) >> >>manual tuning: >>>The pro's are not far off perfect tuning nowadays. Of course the parameters >>>can get improved, but definitely not their tuning. >> >> You're making strong claims, but I don't see any reasoning to back them up. Can >>you prove it, or are you blowing smoke? > >If i throw a few important values of diep in the autotuner i get >weird values back. > >No one ever reported success. The only program which is so called >a success i have seen with my own eyes on icc. in fact i've played against >it with DIEP hundreds of games. I've in total seen about 500 games or so >of the thing called knightcap. > >It wasn't true simply. > >tuning a bunch of parameters at a time isn't working. It's hard proof. >a program that's outsearching opponents by 1 to 2 ply was getting >smoked completely. a few years ago 2 ply was a lot. > >nowadays it says less obviously when you get above 10 ply. > >So you are having the move. Show me that it works! > >Get crafty source code. modify the tuning to crafty. There you go. >Perhaps a week of work at most. > >I only see a lot of people chat about technical terms. 'back propagation'. >Hehe. Lucky i forgot all those things. > >I know what it did for me and i know what it did for others. And i know >it wasn't good. > >And it's very easy to realize why. > >It has no domain dependant knowledge, despite that patterns it is supposed >to tune ARE domain dependant written. > >In short that's a contradiction which no tuner can make up for! > >Then we didn't hit the subject of granularity, making a method to draw the >right conclusions out of a result and another zillion of things which are >not able to do either :) > >The best example is book learning. Despite years of work the only thing >book learning can do for you is repeat won lines and avoid lost lines. > >Smarter than that you can't get I disagree there. "result learning" is but one type of book learning. Mchess learned to add book lines after playing a game. Crafty learns that the positions right out of book are favorable or not so favorable (not won or lost, just good or bad) and uses that information as well the next time it comes this way in the book... > >Have you ever thought of that, how much *effort* has been put in learning >in chess the last years, and how little good results it has brought? > >Best regards, >Vincent
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.