opposite castling armageddon

[Ovyron]

You can use them for engine-play. When finished, just set all drawn-games to a win for Black - done!

I don't think this is the same, engines would play very differently if they knew the scoring isn't going to be normal Win=1 Draw=0.5 Lose=0. They're effectively playing a different game.

[Michel]

You can use them for engine-play. When finished, just set all drawn-games to a win for Black - done!

I don't think this is the same, engines would play very differently if they knew the scoring isn't going to be normal Win=1 Draw=0.5 Lose=0. They're effectively playing a different game.

You are right. For example for white a risky move with wdl probabilities 0.5, 0.0, 0.5 now becomes preferable to a "safe" move with probabilities 0.0, 1.0, 0.0.

I am not sure if current A/B engines can be easily changed to accommodate the draw=loss (for white) scoring. It seems they would have to be able to estimate the drawishness of a position, which currently they generally don't do.

Edit: In this respect it is interesting that the latest lc0 net will output wdl scoring.

[lkaufman]

It seems that the results and opinions posted here generally indicate that my proposal does not give White enough advantage to fully offset draw odds, at least with engine play. Therefore I'll propose instead the variant I mentioned in which White has normal castling options but Black has only the queenside castling option, plus draw odds. An opening sequence to produce this would be 1.Na3 Nh6 2.Nb5 Rg8 3.Nc3 Rh8 4.Nb1 Ng8. This is obviously more favorable for White than the initial proposal, but probably not a lot more favorable, as White would usually choose to castle short anyway. Lc0 analysis has White's percentage expectancy rising a couple points, to 70-72% range, which is about right for draw odds. A short shootout I ran this way had White winning 7 wins to 3 draws (= losses). It has the extra merit of allowing for both same and opposite castling, like normal chess, although opposite castling will be much more common. Perhaps it needs a different name though. "Black no short castling Armaggedon" is a bit long!

[pohl4711]

It seems that the results and opinions posted here generally indicate that my proposal does not give White enough advantage to fully offset draw odds, at least with engine play. Therefore I'll propose instead the variant I mentioned in which White has normal castling options but Black has only the queenside castling option, plus draw odds. An opening sequence to produce this would be 1.Na3 Nh6 2.Nb5 Rg8 3.Nc3 Rh8 4.Nb1 Ng8. This is obviously more favorable for White than the initial proposal, but probably not a lot more favorable, as White would usually choose to castle short anyway. Lc0 analysis has White's percentage expectancy rising a couple points, to 70-72% range, which is about right for draw odds. A short shootout I ran this way had White winning 7 wins to 3 draws (= losses). It has the extra merit of allowing for both same and opposite castling, like normal chess, although opposite castling will be much more common. Perhaps it needs a different name though. "Black no short castling Armaggedon" is a bit long!

I did a experiment with normal castling options for white and no castlings allowed for black. That seems quite OK. In my testsings, asmFish scored 58.4% vs. Komodo 10.4 (after 205 games played)

Compared to the results below, this is not bad (but not very good, too.). More experiments are needed. I will continue working on this interesting idea. Another option could be: One pawn material advantage for white (1. Na3 a6 2. Nb1 a5 3. Na3 a4 4. Nb1 a3 5. Nxa3 Na6 6. Nb1 Nb8) - but when doing this, all openings have to be checked by an engine first (when the opening line moves the white queen to a4 for example, the queen can be captured by black rook from a8...).

Test results:
(asmFish 170426 vs. Komodo 10.4, 5'+3'' time-control, singlecore, no ponder, no endgame-bases, LittleBlitzerGUI, 1000 games each testrun(!) except Noomen Gambit-lines (only 246 positions, so 492 games were played) and Noomen TCEC Superfinal (only 100 positions, so 200 games were played))

Stockfish Framework standard 8 move openings: Score 60.3% – 39.7%, draws: 63.4%
FEOBOS v20 contempt 5 top 500 openings: Score 58.7% - 41.3%, draws: 64.1%
HERT 500 set: Score: 60.6% - 39.4%, draws: 60.4%
Noomen Gambit-Lines: Score 59.1% - 40.9%, draws: 59.3%
4 GM-moves short book: Score 60.5% - 39.5%, draws: 57.1%
Noomen TCEC Superfinal (Season 9+10): Score: 62.5% - 37.5%, draws: 50.0%
SALC V5 half-closed: Score 61.6% - 38.4%, draws: 49.2%
SALC V5 full-closed 500 positions: Score 66.5% - 33.5%, draws: 47.7%
Drawkiller (normal set): Score: 65.3% - 34.7%, draws: 33.5%
Drawkiller (tournament set): Score: 65.3% - 34.7%, draws: 33.5%
(no mistake by me: the results of Drawkiller normal and tournament were exactly
the same after 1000 played games!)
Drawkiller (small 500 positions set): Score: 66.4% - 33.6%, draws 30.5%

[Laskos]

Edit: In this respect it is interesting that the latest lc0 net will output wdl scoring.

It's good Lc0 does indeed the wdl scoring, but to me it seems that they use some wrong simple empirical formula for translating (bijecting) score < - > wdl (depending on conditions, it has a complicated approximate empirical translation).

More important, the whole training should be performed again with maximizing w + d instead of w + d/2 (both with w + d + l = 1 constraint). So, basically from scratch.

Lc0 is pathologically more specialized and less prone to shifts in goals than AB engines with hand-made evals. It can take several hours of hand-made changes to the eval to change the goal of an AB engine, but it takes a massive training process for Lc0.

[Michel]

Edit: In this respect it is interesting that the latest lc0 net will output wdl scoring.

It's good Lc0 does indeed the wdl scoring, but to me it seems that they use some wrong simple empirical formula for translating (bijecting) score < - > wdl (depending on conditions, it has a complicated approximate empirical translation).

I had the impression that the value head would now natively be WDL. See https://blog.lczero.org/ .

More important, the whole training should be performed again with maximizing w + d instead of w + d/2 (both with w + d + l = 1 constraint). So, basically from scratch.

Lc0 is pathologically more specialized and less prone to shifts in goals than AB engines with hand-made evals. It can take several hours of hand-made changes to the eval to change the goal of an AB engine, but it takes a massive training process for Lc0.

This may be a valid point. Although I still have the feeling it is sufficient for an engine to be able to estimate the drawishness of a position (for normal chess) to be able to play the altered pay off correctly. That would be for an A/B engine at least. For a LC0 style engine there is also the policy head to consider.

[Nordlandia]

‘Great players never castle until the end of the game, and often never at all, as their king, although often in the middle of the board, nevertheless stands secure.’

[Laskos]

Edit: In this respect it is interesting that the latest lc0 net will output wdl scoring.

It's good Lc0 does indeed the wdl scoring, but to me it seems that they use some wrong simple empirical formula for translating (bijecting) score < - > wdl (depending on conditions, it has a complicated approximate empirical translation).

I had the impression that the value head would now natively be WDL. See https://blog.lczero.org/ .

Ah, I see. It was not so several months ago, it seemed to me they got wdl from the score using some empirical (unsound) formula. Now, it seems to do in training effective computation of it. More flexibility, at least.

More important, the whole training should be performed again with maximizing w + d instead of w + d/2 (both with w + d + l = 1 constraint). So, basically from scratch.

Lc0 is pathologically more specialized and less prone to shifts in goals than AB engines with hand-made evals. It can take several hours of hand-made changes to the eval to change the goal of an AB engine, but it takes a massive training process for Lc0.

This may be a valid point. Although I still have the feeling it is sufficient for an engine to be able to estimate the drawishness of a position (for normal chess) to be able to play the altered pay off correctly. That would be for an A/B engine at least. For a LC0 style engine there is also the policy head to consider.

Is it easy for anything to estimate the drawishness of a position? In all the particular conditions, like strength, opponent, etc.? I tried to generalize the Davidson draw model to accommodate such, with very dirty results.

[Michel]

Edit: In this respect it is interesting that the latest lc0 net will output wdl scoring.

It's good Lc0 does indeed the wdl scoring, but to me it seems that they use some wrong simple empirical formula for translating (bijecting) score < - > wdl (depending on conditions, it has a complicated approximate empirical translation).

I had the impression that the value head would now natively be WDL. See https://blog.lczero.org/ .

Ah, I see. It was not so several months ago, it seemed to me they got wdl from the score using some empirical (unsound) formula. Now, it seems to do in training effective computation of it. More flexibility, at least.

More important, the whole training should be performed again with maximizing w + d instead of w + d/2 (both with w + d + l = 1 constraint). So, basically from scratch.

Lc0 is pathologically more specialized and less prone to shifts in goals than AB engines with hand-made evals. It can take several hours of hand-made changes to the eval to change the goal of an AB engine, but it takes a massive training process for Lc0.

This may be a valid point. Although I still have the feeling it is sufficient for an engine to be able to estimate the drawishness of a position (for normal chess) to be able to play the altered pay off correctly. That would be for an A/B engine at least. For a LC0 style engine there is also the policy head to consider.

Is it easy for anything to estimate the drawishness of a position? In all the particular conditions, like strength, opponent, etc.? I tried to generalize the Davidson draw model to accommodate such, with very dirty results.

Well I was assuming drawishness between identical or at least very similar engines. This can be done with the Texel tuning method. I think it would not be particularly difficult.

[Laskos]

Well I was assuming drawishness between identical or at least very similar engines. This can be done with the Texel tuning method. I think it would not be particularly difficult.

Even in these very particular conditions, how many parameters are there in Texel tuning method, aside eval? Another goal of that method is to make the mapping of the eval to score for the whole game via a logistic. And it fails to do the same mapping for openings and endgames, if my experiments were not that off. And as soon as you deviate from a "Texel engine at a particular time control and hardware", those too few parameters have to be changed. It is good for self-training and self-testing, but one can hardly generalize.