Author: Robert Hyatt
Date: 18:16:29 06/25/01
Go up one level in this thread
On June 25, 2001 at 15:31:52, Chris Carson wrote: >Bob, > >I have not seen a "this type of position" at 40/2 that anyone can get more than >one win with. Each so called "this type of position" only leads to a single win >that a program with learning avoids later, so I do not see the "this type of >position" argument holidng up. If you can provide me the "this type of >position" and some lost games, I would appreciate it. First, remember that Fritz learns using the same algorithm that Crafty uses. If fritz drops out of book a pawn (or more) ahead, it will like that opening. If things turn sour 30 moves after book, I don't see how book learning is going to help at all. The GM will discover something like "hey, it has no idea about distant majorities and will trade into a lost ending to win a pawn." At that time, the game ends, because the computer and the GM are searching toward different goals, and the computer doesn't understand the GM's goal at all. Second, suppose the GM discovers that one of the "short openings" is a winner. ie c3 and Qc2 and then the program falls into a hole and blows up. There is nothing it can learn since c3 took it out of book from the get-go. This is like the famous fritz vs tiger opening that hit on ICC. Except that was not a GM pulling the plug. It was a low-rated human that found it. > >I am looking for 40/2. At blitz or ICC time controls that are not long, then >bets are off since the computer can be tricked with little time to think. >Tacticts that work at short time controls do not seem to work at long time >controls for more than a game or two. If you can provide me 40/2 games and >results, I would appreciate it. As you have said many times, blitz has no >relationship to 40/2 games. > I think you might be asking for "too much". It takes many games to find these "key positions". With a single game taking 4-6 hours, it is unlikely anyone is going to try this, except for Kramnik (since he stands to win a million dollars or so, he might be willing to "do the time" but I doubt he will tell anyone until it is over. :) >If the "this type of position" exists, it will be a breakthrough. Right now the >programs are leading by about a 2 to 1 victory margin and noone can find a >consistant winning "this type of position". GM Kramnik is working on it, but >even he needs 3 months playing the program and I doubt a single "this type of >position" will emerge, he will win, but it will be with diversity. > >Please only respond with the "this type of position" and some 40/2 results to >back it up. Thanks! :) > >Best Regards, >Chris Carson One such known position is the "trojan horse" attack. It occurs naturally in several openings and when the computer doesn't have a "book" it generally eats the knight and gets blasted. Another is the classic stonewall with white having pawns at c3, d4, e3 and f4, and black having pawns at d5 and e6. I think GCP posted such a position although I have no idea how he reached it vs Crafty (assuming it was crafty) since crafty doesn't like such positions and won't castle into the firing line. The mistake is that black often complacently castles kingside right into a ready-made pawn-storm and mating attack. GM players will be much more subtle. I watched "cptnbluebear" notice that I had a bug in my pawn evaluation and didn't realize two different things on two different cases. One was that in a king-only ending, split pawns are better than connected passers. The other had to do with a majority and not understanding it. He created endgame after endgame where this beat crafty... until I fixed it at Roman's encouragement.
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.