Well, we do what we can. If you are right, then (for example) simply having a higher penalty for doubled pawns in the shelter should help. If it does, we can then try backward or isolated pawns in the shelter. It's not as good as tuning them all at once, but first we have to prove that the individual terms help.Lyudmil Tsvetkov wrote:I wish you best of luck with any possible experiment.lkaufman wrote:Well, it's worth a test. Doubled pawns also guard key squares, and can also sometimes advance when "hooked", but you might be right. Maybe all weak pawn penalties should be increased by some percentage in the shelter.Lyudmil Tsvetkov wrote:doubled shelter pawn should be a very valid chess knowledge principle, at least according to my understanding.lkaufman wrote: Well, shelter pawns have their own rules, but I don't think we have any extra penalty specifically for doubled pawns in the shelter. I've never thought of that as a valid general principle (I'm not talking about doubled isolated shelter pawns of course). Do you think White would be much better off in this position if we move the g3 pawn to h3? It still provides a hook for an attack. Doubled pawns are generally ok for defense, just not useful for offense.
SF also does not have specific doubled shelter penalty, but it seems to handle such situations for some reason better, maybe more appropriately tuned specific values, or more relevant specific search.
of course, h3 pawn is much better than g3 pawn in the shelter.
- structure is not compromised
- g4 square is defended
- g3 doubled pawn can be fixed by an enemy pawn by blocking it, making both doubled pawns weak
- most importantly, concerning hooks, thanks for teaching me that term, I did not know what it mean until now, there is a major distinction between 'hooking' a normal and a doubled pawn. a normal hooked pawn can advance and pass, preventing opening of attacking lines, while a doubled hooked pawn can only capture or being captured in the vast majority of cases, so in this way it is an excellent opponent target for opening vital lines. for example, with white pawns on f2,g2,g3, after black's h5-h4 break, white can not play g3-g4 to avoid opening lines usually, as the pawn structure is clumsy, while, with white pawns on f2,g2,h3, on black's g5-g4 break, white has the vital option of passing with h3-h4, thus avoiding opening lines near the king. later, h4 pawn can be defended by the g3 pawn, while a doubled g4 pawn can rarely be defended by an f3 pawn. that seems to make the large portion of the distinction.
still, I fail to grasp how are you going to tune this.
you have to tune at the same time, pawn psqt, doubled pawns(possibly also doubled pawns psqt, if you have one), shelter weakness, specific weak pawn shelter weakness, if you try it, etc.
so, you should tune at least 4 evaluation features at the same time, not to mention search, and I fully fail to understand how is it possible to successfully tune those.
that is one of the reasons I am a bit reluctant to become an engine developer.
Some handicap results and conclusions.
Moderators: hgm, Rebel, chrisw
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: Some handicap results and conclusions.
Komodo rules!
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Some handicap results and conclusions.
funny, in some 90% of positions Komodo is losing to SF with king side attacks(of course, the older version I have), Komodo queen would be on the opposite side to where its king is, when SF starts the attack.
that would basically mean attack on the king side, Komodo queen on the queen side.
obviously, Komodo does not have any king-queen proximity bonus, what they call it on cpw, king tropism.
that would basically mean attack on the king side, Komodo queen on the queen side.
obviously, Komodo does not have any king-queen proximity bonus, what they call it on cpw, king tropism.
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: Some handicap results and conclusions.
Plz Larry, let us know about the test result.
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: Some handicap results and conclusions.
The doubled pawn in the shelter idea tested poorly; we'll try the queen-king tropism idea next. I think we tried it long ago and it failed, but the devil is in the details. It is a sound idea.JJJ wrote:Plz Larry, let us know about the test result.
Komodo rules!
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: Some handicap results and conclusions.
Thanks you, I wish you good luck for the next idea. Plz let us know as welllkaufman wrote:The doubled pawn in the shelter idea tested poorly; we'll try the queen-king tropism idea next. I think we tried it long ago and it failed, but the devil is in the details. It is a sound idea.JJJ wrote:Plz Larry, let us know about the test result.
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: Some handicap results and conclusions.
In an other subject. I see on chess.com Komodochess is rated at bullet without playing game. I think a nice idea would be to see the elo rating at bullet or blitz of komodo with some handicap. Player could play against it with knight handicap or rook handicap. I don't think any player would win against knight at bullet, but who knows, and they can play it at blitz.
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Some handicap results and conclusions.
Larry, just to ask you, if you don't mind, if you mind, I will just stop:lkaufman wrote:The doubled pawn in the shelter idea tested poorly; we'll try the queen-king tropism idea next. I think we tried it long ago and it failed, but the devil is in the details. It is a sound idea.JJJ wrote:Plz Larry, let us know about the test result.
how do you define doubled pawns, and how do you define shelter?
certainly, you checked only for doubled pawns on the 2nd and 3rd ranks on the file where the king is as well as the 2 adjacent files?
because SF defines its shelter as going as high as the 6th rank, and considering doubled weakness for any rank above 3 is simply bad.
did you try penalising only g3/b3 doubled shelter pawn?
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Some handicap results and conclusions.
no need to ask, all will fail.JJJ wrote:Thanks you, I wish you good luck for the next idea. Plz let us know as welllkaufman wrote:The doubled pawn in the shelter idea tested poorly; we'll try the queen-king tropism idea next. I think we tried it long ago and it failed, but the devil is in the details. It is a sound idea.JJJ wrote:Plz Larry, let us know about the test result.
if one or 2 ideas succeed, the next 10 will certainly fail, even if reasonable.
that is a lesson from SF framework we should already have learned.
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Some handicap results and conclusions.
another caveat: maybe trying the difference in performance of doubled shelter pawns on and off vs Komodo predecessor and SF would give a wealth of information.
the newer Komodo version would simply not know how to exploit king safety weaknesses, as doubled pawns.
but testing an engine is crazy thing, who knows what other evaluation and search parameters have to play a role?
the newer Komodo version would simply not know how to exploit king safety weaknesses, as doubled pawns.
but testing an engine is crazy thing, who knows what other evaluation and search parameters have to play a role?
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: Some handicap results and conclusions.
I don't want to get into detail, but I'll say that the way we tried was reasonably close to what you would do. As for the king queen tropism idea, it also tested a bit negative, in this case due to the roughly 1% slowdown it entailed; on a fixed depth basis it was slightly plus. This is the case with many ideas; they are helpful if you ignore the cost of computation, but net losers in reality. That's why many of your ideas are not practical. But this is not your problem, it's up to us to judge whether the cost of an idea is low enough to try it.Lyudmil Tsvetkov wrote:Larry, just to ask you, if you don't mind, if you mind, I will just stop:lkaufman wrote:The doubled pawn in the shelter idea tested poorly; we'll try the queen-king tropism idea next. I think we tried it long ago and it failed, but the devil is in the details. It is a sound idea.JJJ wrote:Plz Larry, let us know about the test result.
how do you define doubled pawns, and how do you define shelter?
certainly, you checked only for doubled pawns on the 2nd and 3rd ranks on the file where the king is as well as the 2 adjacent files?
because SF defines its shelter as going as high as the 6th rank, and considering doubled weakness for any rank above 3 is simply bad.
did you try penalising only g3/b3 doubled shelter pawn?
Komodo rules!