First move advantage

[mwyoung]

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

Can this be given as a Elo estimate? Example if GM Carlsen could only play black in all his games, how much lower would GM Carlsen Elo be?

[lkaufman]

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

Can this be given as a Elo estimate? Example if GM Carlsen could only play black in all his games, how much lower would GM Carlsen Elo be?

That's pretty easy to determine from database stats. The elo value of the White pieces grows with higher ratings. At low GM level it's around 35, at high GM level around 40, and at World Champ (2800) level near 45. So 45 would be a good estimate for your question.

[Ajedrecista]

Hi again:

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

Can this be given as a Elo estimate? Example if GM Carlsen could only play black in all his games, how much lower would GM Carlsen Elo be?

That's pretty easy to determine from database stats. The elo value of the White pieces grows with higher ratings. At low GM level it's around 35, at high GM level around 40, and at World Champ (2800) level near 45. So 45 would be a good estimate for your question.

Thanks for your answer to my previous question.

Regarding your last answer of 45 Elo of white advantage at WC level: Carlsen (as everyone) play white around 50% of times (not 100%), so the rating drop should be 45/2 Elo more less, is not it? It is a simple comment but I do not like to speculate with these scenarios.

Regards from Spain.

Ajedrecista.

[Lyudmil Tsvetkov]

Looking at the games of the match between Houdini and Komodo, run by Clemens Keck (http://www.clemens-keck.de/ReplayZone/1010.htm ), I notice that the engines have somewhat different evaluation of the general starting position (or openings very close to the starting position, probably deviating only slightly from perfect play): barring some exceptions, like Houdini estimating very well white's chances in the Caro-Kann and the Petrov (some +35 centipawns), Houdini sticks to an evaluation of some 9-12 centipawns in favour of white, while Komodo thinks white's advantage is somewhere around 22-25 centipawns.
I remember also that Rybka exhibited generally some 25 centipawns in white's favour.

Me, personally, I think that objectively white's advantage should be estimated at some 14-15 centipawns (do not ask me how I am trying to guess this, it is purely intuitive, although a while ago I experimented some erratic calculations). So, if this is anywhere near about the truth (most probably very very far from it), Houdini in general should be a bit closer to such an estimate.

What is your opinion on the matter? Maybe you have some relevant data, estimates, etc. I think this is very important for a top program to have a consistent evaluation and might even have a bearing on other stages of the game.

Best regards, Lyudmil

PS. I hope this posting is not to wild to answer

I can answer your question rather precisely. I have two Aquarium IDeA projects going, one using Houdini 3, the other Komodo. Both have very deeply analyzed all the openings (to around moves 15 to 20 usually) of all games played at 2700 Elo average and above in recent times, plus many others. The resultant minimaxed tree gives the opinion of each program as to the true advantage of the White pieces as well as of every opening. Both Komodo and Houdini are showing +.15 for the initial position (there is some rounding done by Aquarium, this isn't an exact figure), so basically you are right on the money. Currently Komodo gives both 1.d4 and 1.e4 as +.15 and 1.Nf3 as +.11, while Houdini gives 1.d4 and 1.Nf3 as +.15 and 1.e4 as +.11. Against 1.e4 Komodo gives French as best and the other "big three" (1...e5, Sic., Caro) as equal second. Houdini gives French and Sicilian as equal best, 1...e5 as third, but doesn't like the Caro. Against 1.d4 they both rate Slav and Gruenfeld as the two best defenses (K prefers Slav slightly, H prefers Gruenfeld slightly.
So basically, there is remarkably little difference between their evaluations when done this way.

Hi Larry.
Thanks for answering, and congrats on Komdo btw.
I still have problems with making compatible the results of your Aquarium tests and the evals of both top engines I see on Clemens Keck's site.

For Komodo specifically I see:
+0.23 after 1.e4 c5 2.Nf3 Nc6 3.d4 cxd4
+0.23 after 1.e4 c6 2.d4 d5 3.e5
+0.19 after 1.d4 d5
+0.22 after 1.d4 Nf6 2.c4
+0.26 after 1.e4 e5 2.Nf3 Nf6 3.d4
+0.21 after 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6
+0.27 after 1.d4 Nf6 2.c4,
etc.

And the engine basically sticks to those values for quite some time (meaning fairly good consistency).

For me, this seems quite different from the 0.15 values you have stated.

I think you understand very well the implications of couple of centipawns' difference that could gradually transpose to a more significant one.

Best, Lyudmil

Well, there is no mystery here. The scores you report are probably from a 25 to 30 ply search. My scores are from a similar (or deeper) search starting on average around ply 30! The deeper you search, the closer the score for a sound defense gets to the theoretical zero score for a draw.

Best,
Larry

Larry,
I do not wish to be insolent in the least, it is just that the topic seems very interesting to me, so I have a right to cross-investigate you
If Houdini's average is 10 centipawns and Komodo's average is 23 centipawns at ply 25, and you get 15 centipawns for both engines at ply 30 with Houdini starting from a lower evaluation mark and Komodo from a higher eval mark, then your theory about deeper search leading to an approximation to the theoritical zero score for a draw should not be quite right.
I am of the exactly opposite opinion: that a deeper search should lead to an increase in score for the attacking engine (meaning the engine currently leading in score).

Besides, I do not quite see, that if we know that 15 centipawns might be a reasonable value for white's advantage in the starting position, an engine could return somewhat different scores at ply 25. I think this is deep enough, and the absolute value of the white advantage at such depth should not be too distinct from those hypothetical '15' centipawns.

The reason why I raised this matter is that it could be useful for engine developers in the process of tuning of weights. Now imagine that you know what the approximate value of the starting position is, then you can use this knowledge to fine-tune your engine's parameters, add new ones, etc., bearing in mind that you are targeting an eval of some +15 centipawns, and not +10 or +30. I think this definitely might be of some help.

[lkaufman]

Hi again:

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

Can this be given as a Elo estimate? Example if GM Carlsen could only play black in all his games, how much lower would GM Carlsen Elo be?

That's pretty easy to determine from database stats. The elo value of the White pieces grows with higher ratings. At low GM level it's around 35, at high GM level around 40, and at World Champ (2800) level near 45. So 45 would be a good estimate for your question.

Thanks for your answer to my previous question.

Regarding your last answer of 45 Elo of white advantage at WC level: Carlsen (as everyone) play white around 50% of times (not 100%), so the rating drop should be 45/2 Elo more less, is not it? It is a simple comment but I do not like to speculate with these scenarios.

Regards from Spain.

Ajedrecista.

No, the difference between having White and having Black is 2 x 45 = 90, so the difference between half Black and all Black is back to 45.

Larry

[Lyudmil Tsvetkov]

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

Hi Mike.
We were trying to tie the single ply white's advantage to evaluation.
If you know to evaluate the opening position correctly, you actually know quite a bit about chess.
If you evaluate the opening position imperfectly, then your chess knowledge is somewhat limited.
The opening position is the most important of all positions.
I think the discussion makes sense, because modern top engines start going after perfection, and it really matters if you evaluate the opening position with +10, +15 or +30 centipawns.

[Lyudmil Tsvetkov]

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

But how do we know that 3 tempi equal a pawn?
For me, after 3 consecutive white moves, black is dead lost.
And is it possible to relate the advantage of a single ply to the cumulative advantage of so many plies, bearing in mind chessboard realities? (limited piece material, relatively easy access to the enemy king, etc.)

I am sorry for prodding on, but this seems very interesting to me indeed.

[lkaufman]

Looking at the games of the match between Houdini and Komodo, run by Clemens Keck (http://www.clemens-keck.de/ReplayZone/1010.htm ), I notice that the engines have somewhat different evaluation of the general starting position (or openings very close to the starting position, probably deviating only slightly from perfect play): barring some exceptions, like Houdini estimating very well white's chances in the Caro-Kann and the Petrov (some +35 centipawns), Houdini sticks to an evaluation of some 9-12 centipawns in favour of white, while Komodo thinks white's advantage is somewhere around 22-25 centipawns.
I remember also that Rybka exhibited generally some 25 centipawns in white's favour.

Me, personally, I think that objectively white's advantage should be estimated at some 14-15 centipawns (do not ask me how I am trying to guess this, it is purely intuitive, although a while ago I experimented some erratic calculations). So, if this is anywhere near about the truth (most probably very very far from it), Houdini in general should be a bit closer to such an estimate.

What is your opinion on the matter? Maybe you have some relevant data, estimates, etc. I think this is very important for a top program to have a consistent evaluation and might even have a bearing on other stages of the game.

Best regards, Lyudmil

PS. I hope this posting is not to wild to answer

I can answer your question rather precisely. I have two Aquarium IDeA projects going, one using Houdini 3, the other Komodo. Both have very deeply analyzed all the openings (to around moves 15 to 20 usually) of all games played at 2700 Elo average and above in recent times, plus many others. The resultant minimaxed tree gives the opinion of each program as to the true advantage of the White pieces as well as of every opening. Both Komodo and Houdini are showing +.15 for the initial position (there is some rounding done by Aquarium, this isn't an exact figure), so basically you are right on the money. Currently Komodo gives both 1.d4 and 1.e4 as +.15 and 1.Nf3 as +.11, while Houdini gives 1.d4 and 1.Nf3 as +.15 and 1.e4 as +.11. Against 1.e4 Komodo gives French as best and the other "big three" (1...e5, Sic., Caro) as equal second. Houdini gives French and Sicilian as equal best, 1...e5 as third, but doesn't like the Caro. Against 1.d4 they both rate Slav and Gruenfeld as the two best defenses (K prefers Slav slightly, H prefers Gruenfeld slightly.
So basically, there is remarkably little difference between their evaluations when done this way.

Hi Larry.
Thanks for answering, and congrats on Komdo btw.
I still have problems with making compatible the results of your Aquarium tests and the evals of both top engines I see on Clemens Keck's site.

For Komodo specifically I see:
+0.23 after 1.e4 c5 2.Nf3 Nc6 3.d4 cxd4
+0.23 after 1.e4 c6 2.d4 d5 3.e5
+0.19 after 1.d4 d5
+0.22 after 1.d4 Nf6 2.c4
+0.26 after 1.e4 e5 2.Nf3 Nf6 3.d4
+0.21 after 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6
+0.27 after 1.d4 Nf6 2.c4,
etc.

And the engine basically sticks to those values for quite some time (meaning fairly good consistency).

For me, this seems quite different from the 0.15 values you have stated.

I think you understand very well the implications of couple of centipawns' difference that could gradually transpose to a more significant one.

Best, Lyudmil

Well, there is no mystery here. The scores you report are probably from a 25 to 30 ply search. My scores are from a similar (or deeper) search starting on average around ply 30! The deeper you search, the closer the score for a sound defense gets to the theoretical zero score for a draw.

Best,
Larry

Larry,
I do not wish to be insolent in the least, it is just that the topic seems very interesting to me, so I have a right to cross-investigate you
If Houdini's average is 10 centipawns and Komodo's average is 23 centipawns at ply 25, and you get 15 centipawns for both engines at ply 30 with Houdini starting from a lower evaluation mark and Komodo from a higher eval mark, then your theory about deeper search leading to an approximation to the theoritical zero score for a draw should not be quite right.
I am of the exactly opposite opinion: that a deeper search should lead to an increase in score for the attacking engine (meaning the engine currently leading in score).

Besides, I do not quite see, that if we know that 15 centipawns might be a reasonable value for white's advantage in the starting position, an engine could return somewhat different scores at ply 25. I think this is deep enough, and the absolute value of the white advantage at such depth should not be too distinct from those hypothetical '15' centipawns.

The reason why I raised this matter is that it could be useful for engine developers in the process of tuning of weights. Now imagine that you know what the approximate value of the starting position is, then you can use this knowledge to fine-tune your engine's parameters, add new ones, etc., bearing in mind that you are targeting an eval of some +15 centipawns, and not +10 or +30. I think this definitely might be of some help.

I guess my explanation is not valid for Houdini. I think that Houdini does not value space as highly as Komodo does, so White's early space advantage in most openings looks better on Komodo. Based only on the opening position, maybe Houdini is a bit closer to the "ideal" value of 15 than Komodo is, but that is just one of many common opening positions. I often see Houdini giving a zero (or very near zero) score to opening positions in some King's Indian lines where Komodo gives White a normal (near 15) plus, and where White's winning percentage agrees with Komodo. I'm not sure that Komodo is right more often than Houdini, but that is my belief.
I have done some very deep opening analysis together with other theoreticians using Houdini, and very often we end with a zero score, usually meaning a draw. Some day all major openings will be pretty much analyzed to a draw, though no one will be able to remember all the analysis. I think it's normal that evals get closer to zero as we search deep into the endgame.

[mwyoung]

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

Can this be given as a Elo estimate? Example if GM Carlsen could only play black in all his games, how much lower would GM Carlsen Elo be?

That's pretty easy to determine from database stats. The elo value of the White pieces grows with higher ratings. At low GM level it's around 35, at high GM level around 40, and at World Champ (2800) level near 45. So 45 would be a good estimate for your question.

Here is my result of a small 1 year sample of Magnus Carlsen games. Games from chessgames.com. Run with Fritz elo calculator.

2012 rating as white 2820
2012 rating as black 2766

First move advantage seems very important at the top levels.

[Lyudmil Tsvetkov]

The first move advantage as defined by Hans Berliner, is - very obviosly - 0.5 plies.

That's pretty clear and I do not understand any requirement of discussion about this particular fact.

That is obviously true but not helpful by itself. If three tempi equal a pawn (as many books say), then first move advantage is 1/6 of a pawn. Since in the opening a pawn value is usually defined as less than one average pawn, if we say 0.9 pawns then 1/6 of that is .15, the exact figure we agreed on in this thread!

Can this be given as a Elo estimate? Example if GM Carlsen could only play black in all his games, how much lower would GM Carlsen Elo be?

That's pretty easy to determine from database stats. The elo value of the White pieces grows with higher ratings. At low GM level it's around 35, at high GM level around 40, and at World Champ (2800) level near 45. So 45 would be a good estimate for your question.

Here is my result of a small 1 year sample of Magnus Carlsen games. Games from chessgames.com. Run with Fritz elo calculator.

2012 rating as white 2820
2012 rating as black 2766

First move advantage seems very important at the top levels.

Hi Mark.
The sample might not be very representative, but it obviously points at something.
And this would also suggest that 15 centipawns (or whatever the real value of the first move advantage is) is quite a noticeable advantage in chess.