Kingside fianchetto

[Lyudmil Tsvetkov]

[d]r1b2rk1/p4pb1/3p1n1p/qB1Pn1p1/2p1P3/2N3N1/PP3PPP/R1BQ1RK1 b - - 0 16

Of course, current engines will easily see the above position favours black, but that would not be true of all positions with kingside fianchettoed bishops, especially when they are closer to the opening stage and of a more closed nature.

[carldaman]

Struggling to scroll through the miles-long Rybka-dedicated thread, I came here again.

I made the effort to count the number of times Fischer used a kingside fianchetto. For a reference year I took 1970, so when Fischer was already in a top from, but his openings still were not dictated by tactical considerations, related to the forthcoming Candidate matches and World Championship Final.

For the year 1970, I counted 80 serious tournament games played by Fischer (I excluded blitz and simul games), in 40 of which he employed a kingside fianchetto, either with white or black. In another 4 games, he used a queenside fianchetto, and in 36 games no fianchetto appeared. Games with lacking fianchetto are due mainly to Fischer quite often playing the Ruy Lopez and the Sicilian with white, but he also frequently fianchettoed his king bishop in the Sicilian.

So, roughly in 55% of all games Fischer used kingside fianchetto. I noticed at that, that when he played weaker opponents, he often avoided to fianchetto his king bishop, obviously thinking he could win the easy way, even with weaker lines.

The ration of kingside to queensde fianchettos is 10 to 1, indicating that queenside fianchettos are really far less promising than kingside ones.
How do you interpret those numbers?

For me, Fischer of 1970-1972 is the tsrongest player in history, and also theoretically best prepared.

How many engines would play without book kingside fianchettos in more than half of all games?

So if you ask me again, yes, I think a kingside fianchetto is an optimal structural approach, and engines should do everything possible to include in their reperoires this mightly weapon as often as possible.

So I would give not 30cps, but even half a pawn for a bishop on g2/g7 in the opening. At first it might be difficult, but then engines will learn how to play those lines.

But are there any engines out there that actually listen when strictly positional terms are treated?
Engines want tactics, tactics and more tactics. And therefore, they would not like to change also their search so that it benefits also positional, instead of only tactical terms.

So again, how do you interpret the above statistics?

The bolded statement is a very critical observation. This is why I always insist that there has to be willingness to take a step back, accepting no Elo gain or even a moderate loss for some important positional enhancements, in order to set up the foundation from which the engine can actually begin to learn to play real high-level positional chess from the earliest stages of the game.

I also don't care if they release it as a separate 'analysis version' if there is no Elo gain, as long as they don't neglect these positional considerations and their consequences towards the resulting quality of analysis. Otherwise, the engine will primarily be useful for beating up on *other engines* in endless testing cycles, which is ultimately not the most fulfilling prospect.

Regards,
CL

(...and yes, the Fischer of 1970-72 is the best and most dominant player ever)

[Ralph Stoesser]

I'm not at my computer right now and I believe that top engines will find fast that black is better in that last position you gave, but only by search. Static eval should favour white here, not only because of the material advantage. Do you know how to get the static eval printout for a position in SF? (Hint: Start SF from a command line window, type in "position fen YOUR-FEN", hit the ENTER key and then type "eval" and submit again with the ENTER key. Then you get the eval printout and a sub-section about white's and black's bishop values). Since you often have a very good sense for static eval improvements, you often should use this tool to be able to see the status quo.

[Ralph Stoesser]

Carl, everybody is free to fork SF and to do exactly what you think should be done. But for sure it will not happen with the official version. Especially not until the Elo rat race isn't finally decided. Taking steps back only because of so called human knowledge is very risky. You would never know if you ever could make enough steps forward again. Might be interesting for some people to have a good Fisher personality in an engine, tactically 600 ELO weaker but with much human-like knowledge, but I doubt you would find enough programmers who would contribute for your fork to make enough progress in that direction. But dreaming is always allowed.

[carldaman]

Carl, everybody is free to fork SF and to do exactly what you think should be done. But for sure it will not happen with the official version. Especially not until the Elo rat race isn't finally decided. Taking steps back only because of so called human knowledge is very risky. You would never know if you ever could make enough steps forward again. Might be interesting for some people to have a good Fisher personality in an engine, tactically 600 ELO weaker but with much human-like knowledge, but I doubt you would find enough programmers who would contribute for your fork to make enough progress in that direction. But dreaming is always allowed.

Hi Ralph,

A fork would indeed be interesting, but there's nothing wrong with an analysis version, either, that would leave the normal version alone and free to "rat race".

There are plenty of talented programmers working on engines in the 2500-2800 (ccrl) range; it's not that far-fetched that one would be interested in a SF fork, or perhaps incorporating ideas into their own engine.

I'm just throwing philosophical ideas out there for consideration. We have some nice specific ideas from folks like Lyudmil that could be still included into other engines as well, not necessarily just SF.

I'm still eagerly hoping to see a decent patch that would reward extending the pawn chain towards the enemy Kingside (if he has castled that way), while penalizing the extension of a chain towards the Queenside, or the side where the opposing King ain't.

Such a patch is long overdue, probably because it is not so simple to achieve successfully. However, even a small Elo loss would be justifiable, since the engine [any engine, that is] would begin to approach the game in a much more sound manner. It is this latter consideration that I think has been neglected by the programming community. Elo gains act as blinders on a horse. I'd expect more independent, daring, and outside-the-box thinking from at least some corner of the community in respect to this subject matter.

Best regards,
CL

[Lyudmil Tsvetkov]

I'm not at my computer right now and I believe that top engines will find fast that black is better in that last position you gave, but only by search. Static eval should favour white here, not only because of the material advantage. Do you know how to get the static eval printout for a position in SF? (Hint: Start SF from a command line window, type in "position fen YOUR-FEN", hit the ENTER key and then type "eval" and submit again with the ENTER key. Then you get the eval printout and a sub-section about white's and black's bishop values). Since you often have a very good sense for static eval improvements, you often should use this tool to be able to see the status quo.

Thanks Ralph.

I am at great pains trying to learn a bit of programming, but for the time being have not advanced much more than memorising a couple of basic operators and learning what a loop is...

And me, who thought you could probably submit a fianchetto patch.

Also, as Carl rightly points out, the pointy chain fix for SF is long overdue.

[Lyudmil Tsvetkov]

Struggling to scroll through the miles-long Rybka-dedicated thread, I came here again.

I made the effort to count the number of times Fischer used a kingside fianchetto. For a reference year I took 1970, so when Fischer was already in a top from, but his openings still were not dictated by tactical considerations, related to the forthcoming Candidate matches and World Championship Final.

For the year 1970, I counted 80 serious tournament games played by Fischer (I excluded blitz and simul games), in 40 of which he employed a kingside fianchetto, either with white or black. In another 4 games, he used a queenside fianchetto, and in 36 games no fianchetto appeared. Games with lacking fianchetto are due mainly to Fischer quite often playing the Ruy Lopez and the Sicilian with white, but he also frequently fianchettoed his king bishop in the Sicilian.

So, roughly in 55% of all games Fischer used kingside fianchetto. I noticed at that, that when he played weaker opponents, he often avoided to fianchetto his king bishop, obviously thinking he could win the easy way, even with weaker lines.

The ration of kingside to queensde fianchettos is 10 to 1, indicating that queenside fianchettos are really far less promising than kingside ones.
How do you interpret those numbers?

For me, Fischer of 1970-1972 is the tsrongest player in history, and also theoretically best prepared.

How many engines would play without book kingside fianchettos in more than half of all games?

So if you ask me again, yes, I think a kingside fianchetto is an optimal structural approach, and engines should do everything possible to include in their reperoires this mightly weapon as often as possible.

So I would give not 30cps, but even half a pawn for a bishop on g2/g7 in the opening. At first it might be difficult, but then engines will learn how to play those lines.

But are there any engines out there that actually listen when strictly positional terms are treated?
Engines want tactics, tactics and more tactics. And therefore, they would not like to change also their search so that it benefits also positional, instead of only tactical terms.

So again, how do you interpret the above statistics?

The bolded statement is a very critical observation. This is why I always insist that there has to be willingness to take a step back, accepting no Elo gain or even a moderate loss for some important positional enhancements, in order to set up the foundation from which the engine can actually begin to learn to play real high-level positional chess from the earliest stages of the game.

I also don't care if they release it as a separate 'analysis version' if there is no Elo gain, as long as they don't neglect these positional considerations and their consequences towards the resulting quality of analysis. Otherwise, the engine will primarily be useful for beating up on *other engines* in endless testing cycles, which is ultimately not the most fulfilling prospect.

Regards,
CL

(...and yes, the Fischer of 1970-72 is the best and most dominant player ever)

I always enjoy your presence in a thread, Carl.

If you allow me one little remark: better positional understanding and stronger play are not mutually exclusive, on the contrary, they reinforce each other.

As far as I understand it, the problem with most engines not being able to implement positional terms in their structure stems basically from the way their search is tuned to all other non-positional eval parameters.

They have tuned for years the search function to work with specific eval parameters of a non-positional nature, but when you introduce more positional parameters, you have to also tune all your search.

The point is that, with a pointy chain for example, you would like to look at different portions of the tree than in the standard case of positions dominated mainly by factors like mobility and attacks.

[Ralph Stoesser]

The directed pawn chain rule is very specific and complex, rather difficult to implement and will probably slow down the eval considerably. It is a rule about closed positions where engines don't understand much in general so far. As a chess playing entity myself I meanwhile have the feeling there must be more simple principles behind that complex rule, something easier and more basic, with a better chance to succeed ELO wise. The rule is good and right, essy to understand for humans, but not in right language for a recent chess engine if may use that analogy. Again, I believe the task is to find more basic principles behind this specific rule. I have noticed that Lyudmil novadays proposes many simple and basic ideas that are easy to try out. In my opinion this is exactly the right direction to proceed.

[carldaman]

Struggling to scroll through the miles-long Rybka-dedicated thread, I came here again.

I made the effort to count the number of times Fischer used a kingside fianchetto. For a reference year I took 1970, so when Fischer was already in a top from, but his openings still were not dictated by tactical considerations, related to the forthcoming Candidate matches and World Championship Final.

For the year 1970, I counted 80 serious tournament games played by Fischer (I excluded blitz and simul games), in 40 of which he employed a kingside fianchetto, either with white or black. In another 4 games, he used a queenside fianchetto, and in 36 games no fianchetto appeared. Games with lacking fianchetto are due mainly to Fischer quite often playing the Ruy Lopez and the Sicilian with white, but he also frequently fianchettoed his king bishop in the Sicilian.

So, roughly in 55% of all games Fischer used kingside fianchetto. I noticed at that, that when he played weaker opponents, he often avoided to fianchetto his king bishop, obviously thinking he could win the easy way, even with weaker lines.

The ration of kingside to queensde fianchettos is 10 to 1, indicating that queenside fianchettos are really far less promising than kingside ones.
How do you interpret those numbers?

For me, Fischer of 1970-1972 is the tsrongest player in history, and also theoretically best prepared.

How many engines would play without book kingside fianchettos in more than half of all games?

So if you ask me again, yes, I think a kingside fianchetto is an optimal structural approach, and engines should do everything possible to include in their reperoires this mightly weapon as often as possible.

So I would give not 30cps, but even half a pawn for a bishop on g2/g7 in the opening. At first it might be difficult, but then engines will learn how to play those lines.

But are there any engines out there that actually listen when strictly positional terms are treated?
Engines want tactics, tactics and more tactics. And therefore, they would not like to change also their search so that it benefits also positional, instead of only tactical terms.

So again, how do you interpret the above statistics?

The bolded statement is a very critical observation. This is why I always insist that there has to be willingness to take a step back, accepting no Elo gain or even a moderate loss for some important positional enhancements, in order to set up the foundation from which the engine can actually begin to learn to play real high-level positional chess from the earliest stages of the game.

I also don't care if they release it as a separate 'analysis version' if there is no Elo gain, as long as they don't neglect these positional considerations and their consequences towards the resulting quality of analysis. Otherwise, the engine will primarily be useful for beating up on *other engines* in endless testing cycles, which is ultimately not the most fulfilling prospect.

Regards,
CL

(...and yes, the Fischer of 1970-72 is the best and most dominant player ever)

I always enjoy your presence in a thread, Carl.

If you allow me one little remark: better positional understanding and stronger play are not mutually exclusive, on the contrary, they reinforce each other.

As far as I understand it, the problem with most engines not being able to implement positional terms in their structure stems basically from the way their search is tuned to all other non-positional eval parameters.

They have tuned for years the search function to work with specific eval parameters of a non-positional nature, but when you introduce more positional parameters, you have to also tune all your search.

The point is that, with a pointy chain for example, you would like to look at different portions of the tree than in the standard case of positions dominated mainly by factors like mobility and attacks.

Thanks, Lyudmil, for the kind words. I'm glad you're back posting again regularly, too

Yes, better positional understanding should always result in stronger play, if it wasn't for coding redundancies, overlaps and other unforeseen conflicting factors. With engines, that will often rear its ugly head, it's just the nature of the beast.

I just think that something as fundamental as pointy chains deserves more than just one stab at implementing properly, and may even require (temporary) sacrifices in other areas to get things going. It is better to implement something credible, than to do nothing at all.

It may look like a step back, but we're looking for a sounder foundation, or rather, a better springboard for future gains. Sometimes, this sort of strategic backstepping is the only sure eventual way forward. This happens a lot in other areas; I know I have done this many times in life, and often it was the only successful way to maneuver out of a bad situation.

The trouble here is that an engine sitting at the top of the charts is not viewed as a bad situation, so why change if nothing positive shows up on the ratings bottom line?!

[Ralph Stoesser]

Opening a command line window and starting an executable is not exactly programming. It was just a proposal, you don't have to do anything. Regarding me, I have no time to code something for SF at the moment. I would like to because I see many nice things to try atm! Sigh. Best wishes!