Average SE over many drawn game positions

[opraus]

There must be 'something' to this ...

Evaluating drawn game positions [from Ed Schroders' million.pgn],

Where:

Opening is defined as missing no more than 1 minor plus 2 pawns.
Ending is defined as having pawns, but no more than 2 Rooks plus 2 minors.
Middle is all the rest.
Material balance is equal.

I extracted only 3000 positions for each game phase. There were certainly many more

Code: Select all

Opening - Average SE :  +10
Middle-game - Average SE :  +21
Endgame - Average SE : +8

I am wondering if the average opening should be closer to +30 [a tempo] ?
I could certainly have ran the tests longer with 100x as many positions, but this only took about an hour, and it help me find a bug in my king activation code.

-David

[opraus]

I ran through 300k 'opening' positions to get an average score of +11.87 cp in favor of white.

This takes considerably less time than playing 1000 5 minute games, but I am not so sure of the significance, or what the numbers should be ....

My 'suspicion' is, that the +11.87 is still low, but who knows.

Any ideas? Should it be 2-3 x higher? Or, 0 ?

Then I can start looking for what might be low. Or high.

Has anyone done this? I think Crafty will run through an epd file, but it may not keep a running total for the static evals ...

Does this make any sense to anyone ?

-David

[Aleks Peshkov]

My 'suspicion' is, that the +11.87 is still low, but who knows.

You are free to scale up 3 times your personal evaluation coefficients (mobility, castling bonus and so on) to get some 19th century chess book value of 30 cp, if you like.

[opraus]

Hi Aleks,

Yes, I do not know what it ought to be, but even the transition through the phases does not seem correct.

White certainly has *some* advantage in the opening, and perhaps 1/2 that in middle, until completely gone in the end [for drawn games].

Seems sensible ...

Of course, all this is from a lousy chess player ):

This caught my curiosity, because at first, I was getting all negative numbers .... so that could not have been right.

-David

[Aleks Peshkov]

What exactly is your "Average SE" value? I suspect that the nature of this value have no strong correlation to chess advantage.

[opraus]

Average SE is average static evaluation over all 3k positions. I would pass each position to my static evaluation routine, and accumulate the scores. Finally, divide by number of positions.

-David

[Aleks Peshkov]

IMHO the value of tempo better to be calculated as difference in evaluation before and after each non-capture game move. I think the difference between white-to-black and black-to-white positions should be comparable.

Absolute SE are random values of source games, it is greatly influenced by popularity of various openings among humans and their main lines.