Asymmetrical evaluation

[lauriet]

Hi All,

My engine displays an asymmetrical evaluation with search depth:

1 6 0 B4A3 -16cp
2 84 0 A6D6 H7H8 -57cp
3 220 0 A6XF6 G7XF6 B4B3 -18cp
4 2908 0 A6A2 H7G8 A2XE2 G8F7 -59cp
5 10030 0 A6A2 E2F4 B4B3 G7G5 A2A7 -21cp
6 20814 0 A6A2E2F4 A2D2 H7G8 D2H2 G8F7 -65cp
7 188578 4 A6E6 E2F4 E6E4 F4G2 B4C5 G7G5 E4A4 -23cp
8 1146981 19 A6A2 E2F4 B4XB5 H7H6 B5C4 H4H3 A2A6 G7G5 -91cp
9 2475004 71 B4XB5 H4H3 A6A1 E2D4 B5A4 D4F3 A1H1 F3G1 H1XG1 -52cp

Even plys evaluate worse than odd plys.

Is this just another thing I dont understand about search or do you think I have a bug ?



Regards
Laurie.

[Stan Arts]

After odd plies, you are up a move and could improve your position. (Usually.)
After even plies, the opponent got a move in improving his position.

This effect becomes less noticeable as the length of variations start to vary more with things like extentions/reductions.

[PK]

Stan's explanation is spot-on. It can be added that You can try to alleviate this effect by giving a small bonus for the side to move (something like 10 centipawns in the middlegame, 5 centipawns in the endgame).

[lauriet]

But doesn't the Qsearch continue until no more captures, so it could finish on an even or an odd ply.
A 3 ply search may finish after 8 more capture....so 11 ply and a 4 ply search may finish after 7 more captures...so an 11 ply search.

??????

Laurie

[PK]

You may check for yourself how often the first ply of quiescence search cannot improve on evaluation score.

[mar]

my bet: qsearch

[bob]

Hi All,

My engine displays an asymmetrical evaluation with search depth:

1 6 0 B4A3 -16cp
2 84 0 A6D6 H7H8 -57cp
3 220 0 A6XF6 G7XF6 B4B3 -18cp
4 2908 0 A6A2 H7G8 A2XE2 G8F7 -59cp
5 10030 0 A6A2 E2F4 B4B3 G7G5 A2A7 -21cp
6 20814 0 A6A2E2F4 A2D2 H7G8 D2H2 G8F7 -65cp
7 188578 4 A6E6 E2F4 E6E4 F4G2 B4C5 G7G5 E4A4 -23cp
8 1146981 19 A6A2 E2F4 B4XB5 H7H6 B5C4 H4H3 A2A6 G7G5 -91cp
9 2475004 71 B4XB5 H4H3 A6A1 E2D4 B5A4 D4F3 A1H1 F3G1 H1XG1 -52cp

Even plys evaluate worse than odd plys.

Is this just another thing I dont understand about search or do you think I have a bug ?



Regards
Laurie.

On even plies, the opponent gets one additional move, which should make the score drop a little. On odd plies, your program gets one extra move, which should make the score jump a bit. Also do you have a bonus in your evaluation for side to move (something like if wtm score += bonus else score -= bonus;)??

[Dann Corbit]

Rotate and mirror the positions and analyze those also. That may give you a clue, especially if you print out details in a log.

Something that will cause such a problem is giving the color <x> bonus to color <y> for some term.
Typical cut and paste problem.