![[au]](/mwf/v12/flags/au.png "Australia")
As you know compmaster najdorf poisoned pawn is an amazing opening. I have won many computer games using that opening however this time i followed the opening till move 22 and at this point according to the book white has a 99.7% chance of winning and black only 0.3% and there is 196 games available to continue down this line, But my opponent deviated from the line at this point and i do not see a way to avoid three-fold repetition and still have a winning position. This is the position:
e2-e4 c7-c5
2 Ng1-f3 d7-d6
3 d2-d4 c5xd4
4 Nf3xd4 Ng8-f6
5 Nb1-c3 a7-a6
6 Bc1-g5 e7-e6
7 f2-f4 Qd8-b6
8 Qd1-d2 Qb6xb2
9 Ra1-b1 Qb2-a3
10 e4-e5 d6xe5
11 f4xe5 Nf6-d7
12 Nc3-e4 h7-h6
13 Bg5-h4 Qa3xa2
14 Rb1-d1 Qa2-d5
15 Qd2-e3 Qd5xe5
16 Bf1-e2 Bf8-c5
17 Bh4-g3 Bc5xd4
18 Rd1xd4 Qe5-a5+
19 Rd4-d2 O-O
20 Bg3-d6 Rf8-d8
21 Qe3-g3 Qa5-a1+
22 Rd2-d1 Qa1-a5+
e2-e4 c7-c5
2 Ng1-f3 d7-d6
3 d2-d4 c5xd4
4 Nf3xd4 Ng8-f6
5 Nb1-c3 a7-a6
6 Bc1-g5 e7-e6
7 f2-f4 Qd8-b6
8 Qd1-d2 Qb6xb2
9 Ra1-b1 Qb2-a3
10 e4-e5 d6xe5
11 f4xe5 Nf6-d7
12 Nc3-e4 h7-h6
13 Bg5-h4 Qa3xa2
14 Rb1-d1 Qa2-d5
15 Qd2-e3 Qd5xe5
16 Bf1-e2 Bf8-c5
17 Bh4-g3 Bc5xd4
18 Rd1xd4 Qe5-a5+
19 Rd4-d2 O-O
20 Bg3-d6 Rf8-d8
21 Qe3-g3 Qa5-a1+
22 Rd2-d1 Qa1-a5+
This is a commonly played line...any ideas...
23.c3 and after 23...Qf5 we have the same(almost) lines with these if black played 21...Qf5.
![[it]](/mwf/v12/flags/it.png "Italy")
Statistics... I will never stop saying this enough: never trust them blindly!
One simple example if you rely on them, you would then see 1.h3 is giving better statistics in engine room than any other first move, even 1.e4! while it's obvious that 1.h3 is a crap move, but many players as black don't know how to counter it.
Even worse is happening in this variation of poisoned pawn, we see as you say statistics of white winning 99.7% of the times... while the truth is that if black is better prepared, black would always draw! This variation you gave is one example, one single variation leading to draw for black completely annihilates that 99.7% wins percentage of white, if only everyone as black would start playing this variation, then the percentage would be 0% wins by white, and not 99.7, lol. This is only one drawish variation black can play on poisoned pawn, there are many others which, according to my analysis, ensure black the draw, despite the over 90% wins percentage I keep seeing. Another point is that in many of these variations, even if black plays the correct drawish variations, going for repetition, then at the second time the same position is reached, the engine goes out of book (it always happens when the same position occurrs twice, the engine will start thinking itself avoiding the book moves) and very often plays a losing move, instead of repeating the repetition move as showed in the book, ensuring the draw. An example of this is given in the variation you just showed, after 21...Qa1+ 22.Rd1 Qa5+, if white plays Rd2 again, the same position will be reached, the engine will start thinking, and there is a high chance (especially at blitz time controls) that it will then avoid the repetition with Qa1+, but will play 23...Qf5?? (even if that move is marked as red in book) entering in a well known losing variation. The engine point of view is that they see the move Qf5 with a little black advantage, while they see Qa1+ as only 000. Then, after they play Qf5, they will enter back in book, and book lose. There are so many such examples... this is only one.
Eros
One simple example if you rely on them, you would then see 1.h3 is giving better statistics in engine room than any other first move, even 1.e4! while it's obvious that 1.h3 is a crap move, but many players as black don't know how to counter it.
Even worse is happening in this variation of poisoned pawn, we see as you say statistics of white winning 99.7% of the times... while the truth is that if black is better prepared, black would always draw! This variation you gave is one example, one single variation leading to draw for black completely annihilates that 99.7% wins percentage of white, if only everyone as black would start playing this variation, then the percentage would be 0% wins by white, and not 99.7, lol. This is only one drawish variation black can play on poisoned pawn, there are many others which, according to my analysis, ensure black the draw, despite the over 90% wins percentage I keep seeing. Another point is that in many of these variations, even if black plays the correct drawish variations, going for repetition, then at the second time the same position is reached, the engine goes out of book (it always happens when the same position occurrs twice, the engine will start thinking itself avoiding the book moves) and very often plays a losing move, instead of repeating the repetition move as showed in the book, ensuring the draw. An example of this is given in the variation you just showed, after 21...Qa1+ 22.Rd1 Qa5+, if white plays Rd2 again, the same position will be reached, the engine will start thinking, and there is a high chance (especially at blitz time controls) that it will then avoid the repetition with Qa1+, but will play 23...Qf5?? (even if that move is marked as red in book) entering in a well known losing variation. The engine point of view is that they see the move Qf5 with a little black advantage, while they see Qa1+ as only 000. Then, after they play Qf5, they will enter back in book, and book lose. There are so many such examples... this is only one.
Eros
![[de]](/mwf/v12/flags/de.png "Germany")
Now, that's interesting. The engine will leave book even if the move is painted green?
yes, always when there is a repetition. I mean the engine will play the green move the first time, but the second time it won't listen to what is in book and will decide itself what to play.
Hm... In case you are playing with R3: Would it help to set a negative contempt?
R3's persistent hash is not affected by this CB book decision to not play repetitions and will happily accept a draw by repetition if that seems to be the best available alternative.
Several years ago, Eros popularized a "refutation" of a Moroczy Bind line I was rather fond of by forcing a repetition that led the engine to pick a different, losing book line. At the time, this kind of pissed me off, but in the long run it turned out to be beneficial because many people have now listed the good line as a loser.
Regards,
Alan
Several years ago, Eros popularized a "refutation" of a Moroczy Bind line I was rather fond of by forcing a repetition that led the engine to pick a different, losing book line. At the time, this kind of pissed me off, but in the long run it turned out to be beneficial because many people have now listed the good line as a loser.
Regards,
Alan
not sure, I think the engine would still probably play a move that it evaluates higher than the repetition move giving 0.00
To avoid confusion, this decision not to allow access to the book in a repeated position, is made by the CB GUI, and has nothing to do with the engine.
Alan
Alan
![[gb]](/mwf/v12/flags/gb.png "United Kingdom")
This does not happen with Polyglot books of course ;)
I've been looking at the CTG spec that Steinar published a while ago, it seems that there are a number of interesting oddities like this with CTG. For instance, if you try to book promotions to anything other than a Queen, moves will start disappearing from the GUI display of your book. The spec that Steinar published isn't quite right though, for instance RED and GREEN positions (not moves) are not at the offset given, they are a few bytes further along. Using the information that Steinar published, it's interesting to look at CTG books and see how many games are imported to make them, I was shocked at how few games there are in Rybka3.ctg for instance, it's tearing up the CB server, having taken out one of your Zappa/b3 lines earlier today :) Nice job Jeroen.
I've been looking at the CTG spec that Steinar published a while ago, it seems that there are a number of interesting oddities like this with CTG. For instance, if you try to book promotions to anything other than a Queen, moves will start disappearing from the GUI display of your book. The spec that Steinar published isn't quite right though, for instance RED and GREEN positions (not moves) are not at the offset given, they are a few bytes further along. Using the information that Steinar published, it's interesting to look at CTG books and see how many games are imported to make them, I was shocked at how few games there are in Rybka3.ctg for instance, it's tearing up the CB server, having taken out one of your Zappa/b3 lines earlier today :) Nice job Jeroen.
I'm not speaking to you until I look at how you beat me this morning... :-)
Regards,
Alan
Regards,
Alan
OK, you've found a position where the move to play was black rather than green, resulting in leaving book on move 7 with NXc6?!. Try out the replacement if you want. :-)
Regards,
Alan
Regards,
Alan
![[il]](/mwf/v12/flags/il.png "Israel")
This can be solved at the engine level.
The engine can be programmed to repeat the move in case that position that happened in the game repeated without calculating.
Uri
The engine can be programmed to repeat the move in case that position that happened in the game repeated without calculating.
Uri
![[us]](/mwf/v12/flags/us.png "United States")
> Statistics... I will never stop saying this enough: never trust them blindly!
> One simple example if you rely on them, you would then see 1.h3 is giving better statistics in engine room than any other first move, even 1.e4! while it's obvious that 1.h3 is a crap move, but many players as black don't know how to counter it.
> Even worse is happening in this variation of poisoned pawn, we see as you say statistics of white winning 99.7% of the times... while the truth is that if black is better prepared, black would always draw! This variation you gave is one example, one single variation leading to draw for black completely annihilates that 99.7% wins percentage of white, if only everyone as black would start playing this variation, then the percentage would be 0% wins by white, and not 99.7, lol. This is only one drawish variation black can play on poisoned pawn, there are many others which, according to my analysis, ensure black the draw, despite the over 90% wins percentage I keep seeing. Another point is that in many of these variations, even if black plays the correct drawish variations, going for repetition, then at the second time the same position is reached, the engine goes out of book (it always happens when the same position occurrs twice, the engine will start thinking itself avoiding the book moves) and very often plays a losing move, instead of repeating the repetition move as showed in the book, ensuring the draw. An example of this is given in the variation you just showed, after 21...Qa1+ 22.Rd1 Qa5+, if white plays Rd2 again, the same position will be reached, the engine will start thinking, and there is a high chance (especially at blitz time controls) that it will then avoid the repetition with Qa1+, but will play 23...Qf5?? (even if that move is marked as red in book) entering in a well known losing variation. The engine point of view is that they see the move Qf5 with a little black advantage, while they see Qa1+ as only 000. Then, after they play Qf5, they will enter back in book, and book lose. There are so many such examples... this is only one.
>
You are a GOD....Forza Azzurri!!
Can anyone tell me where I can get a copy of Compmaster 2
Thanks in advance
Andrew
Thanks in advance
Andrew
![[hr]](/mwf/v12/flags/hr.png "Croatia")
Much thanks
Andrew
Andrew
make your own book and stop being a sheep baaaaa :P
The problem with compmaster book is it's only made of decided games. It simply doesn't contain any draws so it doesn't know them.
Regards,
Lukas
Regards,
Lukas
> The problem with compmaster book is it's only made of decided games.
This doesn't seem to be all that bad an approach for a book created for engine v engine play purely by importing games (ie. no green or red "moves"). The decision you need to make in this case goes something like "do I want to win 50 games and draw 50 games, or do I want to win 80 games, draw 19 games and lose 1 game?"
For a general book, or a Freestyle book, you certainly want to know about the draws too.
![[be]](/mwf/v12/flags/be.png "Belgium")
Hi Nick,
There is a very interesting paradox in this particular topic (keeping or discarding drawn games for book building).
I am sure that many people hope to get a more fighting, less drawish book if they discard drawn games and only keep decided ones for building their book.
This is completely erroneous !
In fact it will happen that the book built without drawn games will favor the most drawish lines in many cases !
Let's see with a little example. Imagine a position with four resonable continuations for which we have 400 games (100 games with move A, 100 with move B and so on)
Imagine that full results of these four continuations are (from the side-to-play point-of-view):
Although the proportion of draws varies very much, all four continuations get a similar overall result : 55%.
If we take the full set of 400 games, we will have equal probabilities for all four moves.
Now let's see what happens if we discard drawn games and build our book from decided games :
In our selected set of decided games we see :
line A : 55% success (55/100; no game has been discarded)
line B : 60% success (30/50; 50 drawn games have been discarded)
line C : 66% success (20/30; 70 drawn games have been discarded)
line D : 100 % success (10/10; 90 drawn games have been discarded)
As most book heuristics strongly favor the more successful lines, move D will be the preferred move in this book, followed by C, B and A.
If we go back to the full set of games we clearly see that having discarded draws we will favor the most drawish line : D which will be the most played move in our book had 90% draws, C had 70% and so on!
Nice, isn't it ?
Marc
PS I fear that quite a few book-makers are not aware of this simple fact.
There is a very interesting paradox in this particular topic (keeping or discarding drawn games for book building).
I am sure that many people hope to get a more fighting, less drawish book if they discard drawn games and only keep decided ones for building their book.
This is completely erroneous !
In fact it will happen that the book built without drawn games will favor the most drawish lines in many cases !
Let's see with a little example. Imagine a position with four resonable continuations for which we have 400 games (100 games with move A, 100 with move B and so on)
Imagine that full results of these four continuations are (from the side-to-play point-of-view):
line A : +55 = 0 -45
line B : +30 =50 -20
line C : +20 =70 -10
line D : +10 =90 -00
Although the proportion of draws varies very much, all four continuations get a similar overall result : 55%.
If we take the full set of 400 games, we will have equal probabilities for all four moves.
Now let's see what happens if we discard drawn games and build our book from decided games :
In our selected set of decided games we see :
line A : 55% success (55/100; no game has been discarded)
line B : 60% success (30/50; 50 drawn games have been discarded)
line C : 66% success (20/30; 70 drawn games have been discarded)
line D : 100 % success (10/10; 90 drawn games have been discarded)
As most book heuristics strongly favor the more successful lines, move D will be the preferred move in this book, followed by C, B and A.
If we go back to the full set of games we clearly see that having discarded draws we will favor the most drawish line : D which will be the most played move in our book had 90% draws, C had 70% and so on!
Nice, isn't it ?
Marc
PS I fear that quite a few book-makers are not aware of this simple fact.
> Nice, isn't it ?
Yes very :) You always come up with good stuff. Does this apply to Polyglot books made from games imported with the "-uniform" option?
The whole idea about not putting draws in the book is to not follow a drawn line when you think that there might be something better. For example, you could have 100 draws resulting from a move, but all those draws were obtained with older hardware, or weaker engines, or at lower depths, or a combination of all of those. This is likely as the arrow of progress goes forwards in computer chess. You wouldn't want to play this move, you'd want your engine to at least attempt to calculate a better move, you might not find a better move and end up playing the drawing move, but that's ok. If you play a new move and win (repeatedly) with it then you've contributed something to theory ;) After this new move has been found, then look at the statistics in a few months time and the story will be very different.
However your analysis leads me to think that the following might be better for CTG books:
Take a "gameset" which includes all games (wins, draws and losses).
1) Build a CTG book from the decisive games
2) Let the CTG book "learn" (ie. adjust weights) from the entire Gameset (inc. draws)
Would this have the advantages of:
1) The book wouldn't play safe draws, so you would have more chance of finding novelties to push theory forwards
2) The book would be less likely to play "move D" in your example, having learned that it's mostly a drawing move
I can see that the compmaster book maker has made extensive use of learning in his book making.
> PS I fear that quite a few book-makers are not aware of this simple fact.
I really don't think it's that simple on this one, otherwise we'd all know the right answer already ;)
Best regards!
Seems to me 21...Nc6 22.Bc7 Qa1+ 23.Rd1 Qb2 24.Nd6 Rf8 leads to a likely draw as well. The line you point out is reminscent of the Marshall, where evaluations and empirical results have a terrific divergence. Whenever you see that there is usually a way out, though it takes work to find it.
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill