Assorted tweaks

[Lyudmil Tsvetkov]

Not all backward pawns are due same penalty of course.

Even if you do not make a distinction between files and ranks, still there are at least 2 types of mainstream backward pawns that should be scored separately: the one is bigger penalty for backward f2 and g2 pawns part of the king shelter, and the second bigger penalty for black d6 and e6 central backward pawns.

[d]6k1/8/3p4/4p3/4P3/8/8/6K1 w - - 0 1
this pawn is worth more, 50% more, or even double the standard penalty

[d]6k1/8/4p3/3p4/3P4/8/8/6K1 w - - 0 1
and this pawns is also worth the same bigger/double penalty

Why are those backward pawns due higher penalties than usual?
Well, because such backward pawns in the center, precisely on d6 or e6, largely determine the fight for the center, and the fight for the center largely determines the outcome of the game.

Having such backward pawns is an extremely undesirable situation.

SF does not understand that. And how could it, when the former backward SF code gave same penalties for files d,c and b? Of course, it will not understand that. Now the code is simplified, with all files getting the same values, but I regard this more as making the role of backward pawns more salient, as now even the edge a/h files get the same bigger value, which was much lower before.

One way or another, I do not mess with tuning, especially when you know that SF still gives lower penalty for baclward pawns in the eg. When you do precisely the opposite of what you should and it works, I do not know what to say. It works, because it is tuned to other elements that are also badly tuned, and this is what makes SF generally scale worse than Komodo at long TC.

Looking at the patch Stefan is currently pushing on fishtest, I remember to say that central d6 and e6 backward pawns should have bigger penalties only in the case when they are unopposed, that was my intention for central backward pawns, but, at the same time, concerning backward pawns part of the king shelter, I think it really does not matter if such pawns are opposed or unopposed, the important thing is that they are on their 2nd rank while not being edge pawns. I think 3rd rank and edge pawns will not quite work.

BUT, that said, neither backward pawns part of the shelter, nor central d6 and e6 backward pawns will work in current SF, no matter how reasonable those terms might be and how badly SF needs them.
They will not work, because they require depth of calculations, which can not be reached at 15 sec. for the entire game. There are chances that such patches might pass at 5 min./game testing, but of course there are simply no resources for such testing.

They will of course also pass in 15 sec. games... in 5 years, when SF search and eval will be much more advanced to be able to make use of such deeper terms.

Anyway, if anyone has the resources, I would recommend that specific king-safety patches be tested at possibly longer TC, naturally of course because all lines having to do with king safety are much more prolific and deep.

Interesting, if SF can not introduce some special rules for longer TC testing, only in the case of king safety patches?

[Lyudmil Tsvetkov]

I forgot to say that such penalties should be given only for the mg, as they have more to do with control of center, rather than weak pawns.

In the eg it makes absolutely no sense to give more than the standard penalty.

[Lyudmil Tsvetkov]

Controllers (do not mind the name) are triplets of pawns on the 4th or 5th ranks arranged in a way, so that the most central one of them has one own pawn each to the left and to the right on the same rank one file apart; i.e. not on the adjacent file, but a file apart.

[d]6k1/8/8/8/2P1P1P1/8/8/6K1 w - - 0 1
triplet of controlling pawns on the 4th rank, with e4 being the central controlling pawn

[d]6k1/8/8/2P1P1P1/8/8/8/6K1 w - - 0 1
triplet of controlling pawns on the 5th rank, with e5 being the central controlling pawn

Controllers on the 4th rank deserve some 10-15cps overbonus in all cases, while controllers on the 5th rank some 40-50cps overbonus in all cases.

You might give the bonus in terms of the most central pawn found for the entire triplet.

The bonus is given for the fact that such triplets control almost the entire board, preventing enemy counterplay/breaks, while at the same time helping with own pawn and pieces activity.
4th and 5th rank are the important ranks for this function, as there is where you should try to stop enemy breaks/activity and from where usually own activisation starts.

The bonus should be given only for the mg, as preventing enemy counterplay is much more important in the mg, and besides, although still useful, in the eg this might have some randomness attached to it. As said, in a good engine the number of mg parameters should be couple of times bigger than the number of parameters applied both in the mg and eg.

What do you think of this concept? (well, I must ask a question, although no one will answer, it starts getting too lonesome here)

[Lyudmil Tsvetkov]

I might have mentioned this in the past, but here is a better implementation/values.

This is the second most important imbalance rule after scoring the bishop pair.
If you are able to benefit from the bishop pair, you should also be able to benefit from this rule.

Count the number of pieces and give the following bonus points in increasing order:

1 piece - no bonus
2 pieces - 20cps
3 pieces - 50cps
4 pieces - 90cps
5 pieces - 150cps
6 pieces - 220cps
7 pieces - 300cps

You count all available pieces for each side, apart from the kings.

Why this bonus?

For coordination, for the simple and very true fact that the more pieces you have, the better the coordination between them will be and they will perform better overall.

Coordination will include:

- defending each other, less easy targets
- attacking enemy pawns in harmony, enemy pawns become easier targets
- attacking the enemy king shelter in harmony, shelter becomes easier target, etc.

This is all very important and certainly might decide the outcome of the game.
As you see, a single piece does not get a bonus, as it has nothing to coordinate with, but 2 pieces can defend each other and attack enemy pawns and the king shelter in synchrony.

The more pieces you have, the better the synchronisation will be, therefore the gradually increasing bonus with the increasing number of pieces.

Simple facts:
- 2 rooks get bigger value than the queen. why? because they are able to defend each other, which the queen can not do alone, and as well attack in synchrony enemy pawns and the king shelter
- in many positions minor plus rook hold against the queen with equal pawns, although they are much weaker in terms of material. why? because they are able to defend each oother and the own pawns in synchrony
- 5 pawns are stronger than rook, but rook plus 5 pawns is weaker than 2 rooks. why? because the 2 rooks are able to defend and attack in coordination enemy pawns and the king

etc., etc., examples have no end

This will help you see why 3 minors are stronger in the general case than queen, why it is usually better to have a piece rather than 3 pawns, but also decide better when to change and when to avoid changes.

[d]4k1r1/8/8/8/8/8/8/2B1K2R w K - 0 1
you do not believe in this rule?
this is a draw

[d]1r2k1r1/8/8/8/8/8/8/R1B1K2R w KQ - 0 1
white wins
just one rook added each side, nothing else changed, but white should win in most cases with this imbalance

why?
because 3 pieces coordinate better than 2 pieces, it is preferable to have 3 vs 2 pieces rather than 2 vs 1 enemy piece

[d]3nkb2/4n3/8/8/8/8/8/3QK3 w - - 0 1
another example
white is better here, with equal number of pawns

[d]1r1nkbr1/4n3/8/8/8/8/8/R2QK2R w KQ - 0 1
already black is not only better, but much better, probably winning, with equal number of pawns
only thing changed is 2 rooks added for each side, nothing more

why this change?
very simple, because 5 piecs coordinate better or much better than 3 pieces, although the material count might be fully the same

Is not this a hacky rule?

Very hacky; it is better to have some term for precisely scoring piece coordination, for example piece defence, attack synchronisation/attacking the same squares, etc., but that should be done with a time frame in mind, and, as long as no one has still found the magic formula to score such defence and attacking synchronisation with a time frame, it is better to use a hacky rule in the meantime than none.

Same with horizontal file symmetry for psqt and other purposes.
Is not asymmetric psqt very hacky?
Definitely it is; no one can argue with the fact that psqt should be horizontally symmetric along the center line running between the d and e files.

But, as long as you do not have very refined terms for scoring specific pawn and piece features taking into account the enemy king position, you better implement for a while asymmetric psqt than nothing.

Otherwise agreed, it is better to take into account the king position, and at least some 200-300 parameters could be scored in terms of that.

[Lyudmil Tsvetkov]

Very simple rule: consider any score below 50cps in the eg as a draw, a 0.0 score.

Nothing less, nothing more.
Guess how much you can gain on that, provided that half of the games go into the eg, and of those half another half has scores lower than 50cps?

Meaning, in 1 of every 4 games you will reach eg scores lower than 50cps.

What is the meaning of this rule?

All eg scores below 50cps are certainly draws. If the players are strong and of roughly equal strength, nothing can avoid that.

So, whenever you reach a line in your search tree that has eg score lower than 50cps, you count it as a draw and instead prefer another line, with a lower score, but ending into the mg.

For example, you have one line with 45cps ending in the eg, and another line with 20cps score edning in the mg.
You simply pick the 20cps mg line as best choice.

Or, you have one line with 30cps score ending in the eg, and another line with 10cps score ending in the mg.
You simply pick the 10cps mg line as the best choice.

What is the danger of it?

No dangers, apart from bogus eval scores.

- one way or another, with equal opponents and decent strength, a 50cps edge position in the eg can never be won
- even if 50cps eg advantage is substituted for 10cps mg advantage, even if most games with such low edge will end in a draw, you certainly have bigger chances to win more games while leading in score than lose, so you should improve your results overall

I think this is a very good rule, with very wide implementation - simply discarding all obviously drawn lines early on and choosing instead lines that are still undecided and promise you better chances. One way or another, a 50cps eg score is a 0.0 score.

What do you think of this rule?

[Lyudmil Tsvetkov]

Try this funny, but very logical, king safety rule: give some 10cps bonus, only for the mg, for a king that has an own chain/defended pawn on the same file in front of the king/on a more advanced rank.

Obviously, the king is safer in this way, as it is more difficult to attack the king frontally.

[d]7k/2p1k1p1/3p3p/3Pp3/2P5/1P4P1/2K2P1P/6K1 w - - 0 1
All 4 kings are due the bonus, as they are behind chain pawns:

- Kh8 behind simple h6 chain pawn
- Kg1 behind twice defended chain pawn, g3
- Kc2 behind inner chain pawn on c4
- Ke7 behind spearhead black pawn on e5

As you see, even when in the center, it is much more difficult to attack such kings frontally, as the defended pawns shelters them.

Maybe, if your king safety/shelter is sufficiently well tuned, you could give the bonus only in case the king is in the center, on files c-f. This can help decide when your king is relatively safe even in the center and gain other advantages instead.

[Lyudmil Tsvetkov]

Of course, you will have very hard time implementing this, if you have different imbalance terms that are also well tuned.

But I am certain that imbalance evaluation can always be improved.

[Lyudmil Tsvetkov]

What would be a complex backward pawn?

The same as backward pawn, but that is not scored in engines.

A complex backward pawn would be a pawn, the square in front of which is controlled by 2 enemy pawns, that defends one own pawn on one adjacent file, and has a duo own pawn on another adjacent file on the same rank.

[d]6k1/8/8/1p1p4/1P3p1p/2PP3P/5PP1/6K1 w - - 0 1
c3 is a complex backward pawn - the square in front of it is controlled by 2 enemy pawns, b5 and d5, it has one duo pawn on the same rank on adjacent file, d3, and defends one other own pawn, b4, on another adjacent file

g2 is also a complex backward pawn, this time also part of the king shelter, so very important - 2 enemy pawns control the square in front of it, f4 and h4, it has one duo pawn on the same rank on adjacent file, f2, and defends one other own pawn, h3, on another adjacent file

Do you recognise the pattern?

remove the d3 and d5 pawns, and you have a simple normal backward c3 pawn

remove f2 and f4 pawns, and you have a simple normal g2 backward pawn

So, actually, not very much distinct from normal backward pawns, has absolutely the same meaning, is a relatively frequent feature with quite some weight, but is not scored in engines...

Any such complex backward pawn is worth some 20cps penalty in all cases.

For a start, if it does not quite work, you might try to do just unopposed complex backward pawns, that would be the bigger weakness.

You can certainly gain as much elo with this feature, as you have gained with normal backward pawns, as they are more or less the same.

This is also not very difficult to implement, certainly not more difficult than implementing candidate passers, for example.
If you can gain something from candidate passers, when you specify a lot, it is exactly the same case here, so why not do it?

I wonder who of the SF team will be the first to push such a patch, in a moment when Lucas is not overseeing the framework?

[Lyudmil Tsvetkov]

Well, very similar to the proposed en passant bonus, but this time a clear distinction will be made between normal backward pawns on the 2nd rank, semibackward pawns on the 2nd rank, and unopposed backwardmakers that would care for the cases not covered by the first two features, possibly already available in many engines.

Besides, an unopposed backwardmaker is the much more forceful case.

An unopposed backwardmaker would be any unopposed pawn on the 5th rank that has an enemy pawn on its second rank on adjacent file.

[d]6k1/p4pp1/8/1P2P3/8/8/8/6K1 w - - 0 1
e5 is an unopposed backwardmaker on the 5th rank - f7 enemy pawn is on its 2nd rank on adjacent file

b5 is another unopposed backwardmaker on the 5th rank - a7 enemy pawn is on its 2nd rank on adjacent file

All unopposed backwardmakers on the 5th rank are due some 20cps mg bonus, 10cps eg bonus.

I think this is strictly scientific, it is very difficult to substitute the feature for any other, is not redundant with normal backward and semibackward pawns, and should be fully applicable if such a patch succeeds somewhere.

Who of the SF team will dare push such a patch?

[Lyudmil Tsvetkov]

A dead knight is a knight that has zero mobility, is placed on h1/a1 or g1/b1 squares, and all the squares it attacks are occupied by own pawns blocked by enemy pawns.

[d]1n2k3/3p4/p1pP4/PpP5/1P4p1/5pP1/5P2/4K2N w - - 0 1

Nh1 is such a dead knight above - zero mobility, f2 and g3 squares it attacks are occupied by own pawns blocked by enemy pawns

Obviously, such a knight not only has zero mobility, but has zero mobility permanently or at least for a very long time.

same for Nb8 - zero mobility, all the squares it attacks, a6,c6,d7, are occupied by own pawns, blocked by enemy pawns

This is a permanently disabled knight.

Any dead knight on the h1/a1 or g1/b1 squares deserves a huge penalty, at least 80cps, but might go well over 150-200cps.
The side with such a knight is simply playing without a full minor piece.

Please note, that this is quite different from low or zero mobility pieces on the edge, as in that case the condition might not be permanent, and also from so called trapped knights, where not own, but enemy pawns are attacking all squares the edge knight also attacks, in which case the knight might still be sacrificed for a couple of pawns.

In distinction, with a dead knight, there is simply no escape.

I assure you I have seen, even long TC games, played by top engines with such features.

What I do not know however is if it makes sense to implement it, as the feature might be relatively or quite rare.

In any case, it is ugly to see a top engine playing like this.