Author: Robert Hyatt
Date: 09:16:19 05/08/01
Go up one level in this thread
On May 08, 2001 at 11:57:07, Larry Proffer wrote: >On May 08, 2001 at 11:43:29, Uri Blass wrote: > >>On May 08, 2001 at 10:58:01, Larry Proffer wrote: >> >>> >>>Thank you for your reply. I concur. >>> >>>What we do know is that Fritz and Junior are, to all intents and purposes >>>'equal', or very nearly 'equal'. >>> >>>If we need to find a winner, it makes *no* difference how many games we play. >>> >>>If we play one game, Fritz has a 50% chance, Junior a 50% chance. >>> >>>If we play one thousand games, Fritz has a 50% chance and Junior has a 50% >>>chance. >>> >>>Number of games is not relevant when they are so closely matched. >> >>I guess that the number of games is relevant and if the number of games is >>bigger the better learner is going to win and it means probably that Fritz is >>going to win after many games because Fritz has a bigger book so it is more easy >>for it to learn to go for lines that the opponent does not understand. >> >>Uri > >1. This, of course, opens up a lot of questions as to why machines were switched >after a few games. This would have killed Fritz's learn files. Why? Everything could have been moved easily. > >It is known that 'book-learning' can have bad effects, particularly after a >string of losses. The effect can be to push the program away from its usual >openings into even worse areas of the book. The desire, of course, is to hope >that it gets pushed away from the losses towards soemthign less bad, but, in >practice, it can be pushed into even worse regions from which there is no >escape. I don't know "where this is known" from.. If you lose a game with a specific opening, and you don't play it again, you won't lose that game again.. > >Books are so large, and the pathways produced by learning so unpredictable, that >this effect is quite common. I have been doing "book learning" for 5 years now. I don't see this at all and I have a big book. > >So another question is why carry out a machine switch whose effect would have >been to kill the learnt data? > > >2. We know about learning in computer chess. Would you tell us if it (comp-comp >learning) has any relevance at all in 'finding the best opponent for Kramnik'. >Doesn't the luck involved with this bi-program learning process (remember, the >learning pathways are almost infinite - we don't know where they lead, and they >may make things worse) just add to the general fact that Fritz still has 50% >chance, and Junior still has 50% chance? Learning is part of the engine. In a match vs a human, a computer had _better_ have good learning skills or it will lose the same game over and over..
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.