Author: Vincent Diepeveen
Date: 17:14:13 06/17/98
Go up one level in this thread
On June 17, 1998 at 16:08:57, Don Dailey wrote: >>>sure, but suppose you know that the typical SSDF tester plays maybe 20 games >>>between two programs before he goes on to a new opponent for one of them? And >>>suppose you know there are *many* testers that will run such short matches. If >>>you provide 20 cooks in your book, you get 20 wins. And by the time the other >>>program has learned to avoid the cooks, the match on that machine is over and >>>the learned results don't prevent a 20-0 match result. >>> >>>As I said, if you know the testing methodology, you can figure out ways to >>>turn it into an advantage. It's happened more than once... >> >>Sure, but in the pre-learner and narrow book times. Nowadays it's practically >>impossible. If you look at the games played in the last year, you will find very >>few cooks and all played by the same program. Overall wide books and learners >>take care of killer lines so efficiently that they are not a real problem >>anymore. >> >>Enrique > >I think you both have valid points. Bob is right that the match size >could have an impact. I'm not sure how good the learners really are >and I suspect cooks still have an impact. I used to detest this but >I have come to believe it is just a fact of life. Even with a human >match, each player has advantages depending on various factors, like >how fast the games are played, which country it is in, how long the >match is and who knows what other factors. It's like life in general, >you work with the advantages and disadvanges you are presented with >and try to cope. Even in chess I believe there is a lot of random >factors and just plain luck. For the size of the current tournament books used in SSDF i think the match size is very important for the game result when learning. Most programs play against for example 1.c4 or 1.b3 only 1 or 2 lines. So if the program wins first game with such a line and the learner has the luck to learn this the first game, then it wins further all white games if it's an aggressive learner. if i remember well look to the Fritz5 versus Mcpro6 games. The option is to have a very wide book. It's not hard to have a very wide book. In fact it can be made easy and automatically. The main problem of a very wide book is that you are not sure that you win, because a very wide tournament book with behind it an even wider book covering it means that book is hard to test even with 1000 games. Also a result might be that the allround enginestrength becomes more important. The wider your book the more different game positions you get. For example: instead of being able to just play slav or just playing well in aggressive variations with open files and pawn formations that cannot change anymore, you then will need to have an engine that is an allround player who plays well in every position. This is gonna get tough considering that most grandmasters from which Karpov is an extreme example also play a very limited number of openings, especially with black against 1.e4. So this would mean that computer programs need to play as well or better as most GM's but even more allround in fact. The human solution to the problem is that he simply does not repeat the game (human still not playing allround chess, as he remains in the same wood of variations), but this is gonna get hard for a program, if we assume that you cannot escape the narrow book, and the fact that compared to playing strength the learning strength does not even approach human D class level. >- Don 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.