Author: Robert Hyatt
Date: 07:23:51 02/01/01
Go up one level in this thread
On February 01, 2001 at 06:25:59, Enrique Irazoqui wrote: >On January 31, 2001 at 22:27:39, Robert Hyatt wrote: > >>On January 31, 2001 at 15:50:08, Enrique Irazoqui wrote: >> >>>On January 31, 2001 at 14:40:01, Robert Hyatt wrote: >>> >>>>On January 31, 2001 at 14:03:09, Enrique Irazoqui wrote: >>>> >>>>>On January 31, 2001 at 13:47:43, Vincent Diepeveen wrote: >>>>> >>>>>>On January 31, 2001 at 11:04:05, Enrique Irazoqui wrote: >>>>>> >>>>>>>On January 31, 2001 at 09:29:49, Vincent Diepeveen wrote: >>>>>>> >>>>>>>>On January 31, 2001 at 08:35:34, Sandro Necchi wrote: >>>>>>>> >>>>>>>>>On January 31, 2001 at 01:00:18, Jouni Uski wrote: >>>>>>>>> >>>>>>>>>>http://www.computerschach.de/tourn/cad2001/cad2001.htm >>>>>>>>>> >>>>>>>>>>In Gadeques tournament Deep Fritz - Shredder 5 ended 10-10. But 14 games were >>>>>>>>>>won by white! And I thought, that whites advantage is minimal in computer chess. >>>>>>>>>>Have programs killer books or what? >>>>>>>>> >>>>>>>>>As far as Shredder 5.0 is concerned there are no killer variations. The book is >>>>>>>>>made mainly for the human players and with a lot of alternatives to make it play >>>>>>>>>different lines. There are very few very long variations. Of course there are >>>>>>>>>good move against weak ones, but not deep variations. >>>>>>>>>So, it is a sort of compromise to make the program fun to play with. >>>>>>>>>Since we drew 6 games and lost 7, there is still a a lot of room for >>>>>>>>>improvements... >>>>>>>>> >>>>>>>>>Sandro Necchi >>>>>>>>>> >>>>>>>>>>Jouni >>>>>>>> >>>>>>>>I'm saving this message to quote later in my life. >>>>>>> >>>>>>>You can dump it. In my games, Shredder 5 didn't play one single killer line. >>>>>>> >>>>>>>Enrique >>>>>> >>>>>> >>>>>>What is your definition of a killer line? >>>>> >>>>>A line that doesn't exist in opening theory and gives decisive advantage to the >>>>>program that plays it. >>>> >>>>I don't want to join the argument, particularly. But my definition is a bit >>>>different. I would call a line a "killer line" if it is chosen specifically >>>>because it leads to a win against a specific opponent. >>> >>>The problem is that we don't know what was in the mind of the book maker, so we >>>must decide whether or not it is a killer line based on other criteria. For >>>instance, A leaves book at move 10, B stays in book until move 28 and then >>>starts computing with a mate evaluation. I have seen this sort of thing. If the >>>line in question doesn't exist in human games you can be sure it's a cooked >>>line. >> >> >>That I would agree with. That is why I generally avoid the "cooked book" >>threads. Part of the definition of "cooking" is based on the word "intent". >>And without a good 1-800-psychic connection to help, I can't figure out the >>"intent" of someone that made a book. Of course, I _told_ you I chose a >>well-known guico variant to play against Belle. It was a perfectly well-known >>variation right out of MCO10 in fact. But my intent was to win a pawn and >>keep it. >> >> >> >>> >>>> IE the line everybody >>>>is smashing tiger with, 1. h4 and 2. h5 is a non-theory line that is known to >>>>lead to a win in nearly every game. By your definition that is _also_ a killer >>>>line. >>> >>>Sure. Funny kind of, though. :) >>> >>>> But If I play some games vs some other program, and I discover that if >>>>I play some variation of the Guioco Piano, I will win most of the games against >>>>that program, then I would call _that_ line a killer line as well. >>> >>>I wouldn't. I understand your point, but there is also a "moral" and a >>>"competent" issue here, I think. In this case, the program that loses to a known >>>line of the Giuoco Piano has a lousy book or at least a lousy line it shouldn't >>>play. It's fault, then, because the author of the book should have known better. >>>Crafty may kill it, but it is not a killer line. To me, I mean. >>> >>>>IE I did this very thing against Belle for several years, as I hav mentioned >>>>before. Belle did this against other programs (myself included) for the same >>>>reason. >>>> >>>>I consider either type of opening as a "cooked book"... >>> >>>But in those times books and learners were much more primitive than today. Now >>>there is little justification for a program falling into a line it dislikes, and >>>even less to keep playing it time and again. >> >>THe issue isn't avoiding the killer the _second_ time. I believe I can >>do that with 100% reliability today. > >Well done, but I still find programs that have lousy learners and fall time and >again for the same line they lose to. That's quite unforgivable nowadays, don't >you think? I wonder how many Elo points a good and aggressive learner is worth >in long matches, but must be quite a few. > I wouldn't be surprised. When I do C vs C matches, I _always_ turn learning off and widen up the book a bit, as otherwise they play almost the same openings as when one side wins, the other side loses, and learning tends to cancel out since they would be sharing the same book. If I use two books, and white wins the first game, then black will try to win the second game the same way, etc... I have done some tweaking recently on the learning after seeing a couple of SSDF debacles. IE it is dangerous to learn too "quickly" on a server as you just might learn everything is bad, based on losing games here and there. But in a single engine vs engine match, learning can be critical. >> The intent is to avoid it the _first_ >>time because the other programmer found the hole but has never played it against >>me in public prior to "the game"... >> >> >> >> >>> >>>So how do you identify a killer line. >>> >>>1 - It's not theory. >>>2 - Quits book with a winning position. >>> >>>I guess it's debatable, but anyway we know one when we see one. :) >> >> >>Think about the tiger game. The position is _not_ won out of book by >>any stretch. White is very probably lost after playing h4 and h5, in >>fact, as he has totally deserted the center. And if black castles >>queen-side, white's advanced (and weakened) pawns are just targets and >>nothing else. The line is based on the fact that tiger will always castle >>into that pawn storm. IE the famous "stonewall attack" is well-known theory, >>but it could be a killer line against the right program... > >It is an anti-Tiger killer line, you are right. But it is an exceptional case. >Usually cooks follow theory until a given moment and then they become "opponent >specific". > >Enrique > >>> >>>Enrique >>> >>> >>>>> >>>>>>A definition of mine could be: "knowing in advance that you make a full >>>>>>point with it against a certain program X1 which is having book X2". >>>>>> >>>>>>I remember 60 moves killer lines in mchess >>>>> >>>>>I don't know how many, but Mchess was full of them. Evals of +2 or more >>>>>immediately after book were not so rare. Sometimes Mchess left book with a mate >>>>>evaluation. :) >>>>> >>>>>In my games, the new books of Deep Fritz, Nimzo 8 and Gandalf are too recent to >>>>>be cooked, but the books of Junior 6 and Gambit are old enough, and still I >>>>>didn't see any killer lines played by Shredder 5. As far as I can tell, we are >>>>>not facing a new "Mchess case". You can download the games and take a look at >>>>>the lines. >>>>> >>>>>Enrique
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.