rybka randomizer versus gm joe ghallager analyses

[ozziejoe]

I thought I would look at a rybka randomizer (http://www.rybkachess.com/index.php?aus ... 2.3+readme) generated database (and success rates of different moves) and compare it to what gm Joe gallagher says is good in his book "starting out: the caro-kann

Here is the starting position

I ran about 1200 games with ply =7 and rybka randomizer playing both sides.


[d]rnbqkbnr/pp2pppp/8/2ppP3/3P4/2P5/PP3PPP/RNBQKBNR b KQkq -[d]


Joe says in this position white pretty much has to play dxc5, or black has an easy game, basically a french with the white bishop outside the pawn chain. Lets see if this is true.

The most commen move by rybka was actually nf3! Already a deviation and white wins 56% of the time (n=338). However, that is only because of some crappy lines black plays. Black has simply to play bg4 and he has no worse than equality (black wins 57%; n = 57; i need more samples!)

[d] rnbqkbnr/pp2pppp/8/2ppP3/3P4/5N2/PPP2PPP/RNBQKB1R b KQkq -[d]

So this common position seems to have no teeth. Joe gallagher also says C3 is not much of a move. However, white won 64% of the time with this move (n =170). Almost everthing black tried, sucked donkeys (e.g., nc6 yields 36%, cxd4 yields 23%, n=11, too small). the only glimmer of hope is h5 which scored 55% (n=11). I may put randomizer on this position for awhile and see what is going on. (Humans, in contrast, only won 42% when playing c3 (n =489), though i suspect that is because the weaker player tended to play that move)

The main move is dxc5 (though i can tell you people under 1800 rarely play it, because they are so concerned about leaving the e5 pawn stranded). One recommended line is E6 (black wins 51%, n =69)! This leads to an "innocous version of the french with white having an extra move.


[d]rnbqkbnr/pp3ppp/4p3/2PpP3/8/8/PPP2PPP/RNBQKBNR w KQkq -[d]

I have to run many more games to get deaper into the tree, but I find this method of analyzing openings to be intriguing. I need to find a better line against c3. This does indeed give me some troubles in practical games. I might also revisit the whole 3 ..c5 variation. I wonder if 3...bf5 is better?

[tano-urayoan]

But in your first diagram is black to move so why white play dxc5, two consecutive moves?
In your second diagram well black has the better of a french, caro-kan advance variation after Bg4 because he pressures d4 immediately (attacking one of his defenders).
In the third position you suggest the main move in the c5 variation.1.e4 c6 2.d4 d5 3.e5 c5 4.dxc5 Nc6 5.Bb5 e6 6.Be3 Nge7 7.c3 Bd7 8.Bxc6 Bxc6 9.Nf3 Nf5 10.Bd4. This is all theory here the Informator monograph recommends 10. Bd4 Nxd4 11. Qxd4 Qa5 12. b4 Qa6 13. a4 b6 with equal game for the pawn. But Gm Movsesian played 14. Qf4 and commenting this move Movsesian said: 'This move was played by young Croatian IM Brkic and brings the whole idea of black's pawn sacrifice under question.'

My point concerning your experiment is a lot of moves has been tested already in opening theory. Must productive will be check the assesment of opening analysis to check its value.
Opening theory is in par with strategic principles and I dont know how well acquainted are engines in this respect. This is a more radical example being this a close position with a lot of maneuvering required.

I recommend you to try the randomizer in a more advance position (for advance i mean closer to the middlegame, in this example maybe move 10 of the main variation). Also Mr Kling ideas of looking at human games where multiple moves have been tried could get a better grip of the position.

P.d. If you dont like what you get in your games you could switch to Bf5 (for something is the main move)

[Rolf]

I thought I would look at a rybka randomizer (http://www.rybkachess.com/index.php?aus ... 2.3+readme) generated database (and success rates of different moves) and compare it to what gm Joe gallagher says is good in his book "starting out: the caro-kann

Here is the starting position

I ran about 1200 games with ply =7 and rybka randomizer playing both sides.


[d]rnbqkbnr/pp2pppp/8/2ppP3/3P4/2P5/PP3PPP/RNBQKBNR b KQkq -[d]

Please let me give a short comment. It's an old saying in statistics that you get what you have predefined with your setting. You have found out the crucial point when you mentioned that perhaps the computer vs computer results alone couldnt show the realistic answer for a chess position if then human players should profit from these results. Unless you have a complete survey like in the famous endgame computer results. I dont have the correct advice for you how to improve it but you must find a safer way so that your numbers have a valid meaning for human players. As a form of exploration of a position you can well analyse with your method by changing the depth of the thought process. To get certain hints for novelties perhaps. But without human made content input you wont find no guarantee in such computer results. Because in the end a somewhat worse looking line is the best because the endings looks promissing but what the machine cant find out with the actual programs. Sorry if I have just described the trivially known but the topic interested me a bit.

[Marek Soszynski]

Joseph,

The weakest part of engines' play are openings (which is why they rely on opening books based on grandmaster games). The Randomizer function is most interesting but its precise worth is yet to be determined. It will work better the further away the position is from the start of a game. Using engines to analyse the very early part of the game is not the best use of them.

[Marc Lacrosse]

Joseph,

The weakest part of engines' play are openings (which is why they rely on opening books based on grandmaster games)

Top class competition opening books are not based on grandmaster games any more.
For more than three years by now...

For a simple reason :
In the present state of affairs, the number of available public games per year :
- is less than 200,000 per year for master class human players
- is more than 600,000 per year for 2600+ engines

Marc

[Marek Soszynski]

Top class competition opening books are not based on grandmaster games any more.

Perhaps I should have said originally based on a core of grandmaster games. In any case, engines use opening books because among other things the book moves are better than what engines will come up with in analysis.

[ozziejoe]

These are many good points in this thread. thank you for your responses.

My hope is that computers can help with the analysis, even early in the opening.


My problem is that i don't trust theory, even early on. I mean it is usually right and the people making the theory are more expert than I, but how many times has something been considered unsound, and later it becomes the main move,or at least one of the popular moves. Indeed, i think 3..c5 for black was once not considered very good. Then the experts (e.g., ghalager) said it was good, and now i've learned a recent analysis says it might not be good. This is move 3 and people don't agree!

Here is another example. Ghallagher says that what is now called the short system was virtually unheard of prior to 1990

[d]rn1qkbnr/pp2pppp/2p5/3pPb2/3P4/5N2/PPP2PPP/RNBQKB1R b KQkq -[d]

Here's what ghallager says "Short started to play it in top class tournaments. At first people thought it was just a feeble variation that would soon disappear, but when vishy anand repeatdly employed it to make the great anatoly karpov suffer the system really took off"

We humans are a bit too prone to fads.

I agree we have to look at the human analysis , and we players have to think very much in terms of practical play. Yet, i think computers are getting to the point where they can impact even early opening theory.


I am still hoping that the "macho grob" is a good defense, no matter what the experts say:)

[Dr.Ex]

[d]rnbqkbnr/pp3ppp/4p3/2PpP3/8/8/PPP2PPP/RNBQKBNR w KQkq -[d]

I have to run many more games to get deaper into the tree, but I find this method of analyzing openings to be intriguing. I need to find a better line against c3. This does indeed give me some troubles in practical games. I might also revisit the whole 3 ..c5 variation. I wonder if 3...bf5 is better?[/quote]

4.c3?! Nc6 is just fine for black. 4.dxc5 and 4.c4 or 4.Nf3 Nc6 5.c4 are critical.

[ozziejoe]

C4 in particular looks violent, Mischa, and is likely to give black many "sweating moments" over the board.

Ok, I am going to assume you are right Mischa. So my main question will now focus on the program, rather than the position. I want to know if rybka randomizer can pick up that c3 is weaksih ( ?!). And i want to know the conditions it picks this up. Then, in future, when a position is not so clear, I hope i can use randomizer in the most effective manner.

Of course my time would probably be better spent figuring out how to deal with c4, or getting better at the endgame:)

best
J

[tano-urayoan]

4.dxc5 is the main line. 4. Nf3 could be answered by Nc6 or Bg4 4. c4 is interesting, a bunch of central tension, needs more analysis.