Author: Robert Hyatt
Date: 14:24:40 12/29/97
Go up one level in this thread
On December 29, 1997 at 15:17:29, Don Dailey wrote: > >But I thought most learning algorithms known to us, were speculative >at best? Sometimes a position has to be visited far too often to >be seen in a realistic amount of time. Many programmers have told >me they work great in some positions but more often than not just >don't do the job. The conclusion seems to be that they are a minor >improvement and that's all. > >Am I wrong about this? I haven't been keeping up too well with the >state of the art in chess learning but know the basic prinicples. > >Is the stuff you do similar to the Scherzer article in ICCA a few >years ago? Or are you talking about the book move counter stuff in >the book selection process? > >-- Don > Here I was talking about "cooking" the book. I also do the same sort of learning Tony talked about (permanent hash, called "position learning" in Crafty) but here am only talking about the book. I simply figure out, 10 moves after I leave book, whether I like this position, dislike this position, love this position, or hate this position, and update the book file so I will know "more" the next time I have the opportunity to follow this book line. It works well for me. Matthias reported that he is using the same algorithm in Fritz 5 that I use in Crafty, and that he's happy with how it works also... It's not magic at all, and does work exceedingly well. It is *critical* if you play on a server. Otherwise you *must* have a very good book, or you will get sucked into losing lines over and over.. IE, somehow you find a line that works against Crafty for most any time control, and ends up +3 in your favor. In that case, Crafty will follow that line *exactly* once, and will then vary somewhere earlier (maybe at the last point with a branch, or with a -3, it might even go back a branch or two earlier, believing the entire sub-tree is broken. But it won't play it a second time, *ever*. That part of book learning is not too hard to get working, and it works absolutely perfectly. The only "trick" is to figure out whether a book line is busted or not. IE often the first search after leaving book might show you down a pawn, because you played a gambit, but 8 moves later, you are +2. What do you believe? That's the only real problem. Now if you do as I do, and try to pick out "trends" in addition to missing outright losses, you might learn that *your* program wins more with e4 than with d4, or vice-versa. I do this as well, but the real benefit is to cut lines out that are gross blunders. This way it is perfectly safe to create a book from internet pgn collections, without worrying about whether the players made gross blunders. If you play enough games using that book, you learn which moves are blunders and never play them again. Of course, if your opponent wants to walk down one of those losing paths from the opposite side, you should also be able to detect that following that line is a good thing to do. :) I've been using this for almost 2 years now and really like the "peace of mind" this gives. I don't worry about someone cooking a line and killing me 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.