Death of a theory. Now I am puzzled.

[Dann Corbit]

Geschrieben von:/Posted by: Dann Corbit at 22 June 2004 00:23:58:

I had a theory that piece values would continuously drop with each new depth of search. The notion seems simple enough. The power of a piece can be represented as something akin to potential energy. As we search deeper and deeper, the intrinsic value becomes obvious and is something akin to kinetic energy. (e.g. a 1 ply search shows undervalued benefit of a queen compared to a knight, but a 30 ply search would show massive damage inflicted by the queen).
My theory looked good from plies 1-5. Ply 6 broke it.
bishop_score=375 at 2; stddev=5.260952 : -0.0204613*x^2 + 15.6317*x + 1454.24
(x=291.000000, y=4271.000000), t=62.000000
(x=307.000000, y=4320.000000), t=62.000000
(x=323.000000, y=4375.000000), t=62.000000
(x=339.000000, y=4402.000000), t=64.000000
(x=355.000000, y=4424.000000), t=63.000000
(x=371.000000, y=4433.000000), t=62.000000
(x=387.000000, y=4442.000000), t=63.000000
(xmax=387.000000, ymax=4442.000000), xmax seen verses curve xmax [387 375]
bishop_score=359 at 3; stddev=9.149938 : -0.0604074*x^2 + 43.3334*x + -2645.39
(x=291.000000, y=4850.000000), t=259.000000
(x=307.000000, y=4959.000000), t=262.000000
(x=323.000000, y=5054.000000), t=267.000000
(x=339.000000, y=5112.000000), t=264.000000
(x=355.000000, y=5113.000000), t=265.000000
(x=371.000000, y=5117.000000), t=267.000000
(x=387.000000, y=5080.000000), t=265.000000
(xmax=371.000000, ymax=5117.000000), xmax seen verses curve xmax [371 359]
bishop_score=353 at 4; stddev=13.878643 : -0.0679874*x^2 + 47.9526*x + -2897.13
(x=291.000000, y=5294.000000), t=1176.000000
(x=307.000000, y=5428.000000), t=1221.000000
(x=323.000000, y=5503.000000), t=1220.000000
(x=339.000000, y=5525.000000), t=1223.000000
(x=355.000000, y=5566.000000), t=1225.000000
(x=371.000000, y=5541.000000), t=1224.000000
(x=387.000000, y=5475.000000), t=1222.000000
(xmax=355.000000, ymax=5566.000000), xmax seen verses curve xmax [355 353.158]
bishop_score=341 at 5; stddev=10.606575 : -0.0803571*x^2 + 54.7857*x + -3392.92
(x=275.000000, y=5595.000000), t=3493.000000
(x=291.000000, y=5740.000000), t=3500.000000
(x=307.000000, y=5861.000000), t=3520.000000
(x=323.000000, y=5928.000000), t=3536.000000
(x=339.000000, y=5936.000000), t=3548.000000
(x=355.000000, y=5919.000000), t=3553.000000
(x=371.000000, y=5880.000000), t=3551.000000
(xmax=339.000000, ymax=5936.000000), xmax seen verses curve xmax [339 341]
(This was on a much slower machine):
bishop_score=352 at 6; stddev=13.737753 : -0.0704985*x^2 + 49.6135*x + -2495.49
(x=275.000000, y=5812.000000), t=32773.000000
(x=291.000000, y=5975.000000), t=32941.000000
(x=307.000000, y=6093.000000), t=33238.000000
(x=323.000000, y=6190.000000), t=33534.000000
(x=339.000000, y=6209.000000), t=33548.000000
(x=355.000000, y=6220.000000), t=33691.000000
(x=371.000000, y=6218.000000), t=34385.000000
(xmax=355.000000, ymax=6220.000000), xmax seen verses curve xmax [355 352]



my ftp site {remove http:// unless you like error messages}

[Uri Blass]

Geschrieben von:/Posted by: Uri Blass at 22 June 2004 01:40:59:
Als Antwort auf:/In reply to: Death of a theory. Now I am puzzled. geschrieben von:/posted by: Dann Corbit at 22 June 2004 00:23:58:

I had a theory that piece values would continuously drop with each new depth of search.

I do not understand your theory.
Your theory suggest that evaluation that gives pieces smaller values is better when you search deeper.
Did you test it in games with fixed depth or only by test suites?
The only test that I fully trust is games.
You can use games at fixed depth and if the depth is small you can get thousands of games in a short time(you can play the same set of opponents when they also will use small fixed depth and after the games reset learning).
If you find that at depth 3 bigger values of pieces are better based on games when at depth 5 smaller value of pieces are better than you have evidence.
Otherwise I see no convincing evidence for your theory.
Uri

[Dann Corbit]

Geschrieben von:/Posted by: Dann Corbit at 22 June 2004 01:53:12:
Als Antwort auf:/In reply to: Re: Death of a theory. Now I am puzzled. geschrieben von:/posted by: Uri Blass at 22 June 2004 01:40:59:

I had a theory that piece values would continuously drop with each new depth of search.

I do not understand your theory.
Your theory suggest that evaluation that gives pieces smaller values is better when you search deeper.
Did you test it in games with fixed depth or only by test suites?
The only test that I fully trust is games.
You can use games at fixed depth and if the depth is small you can get thousands of games in a short time(you can play the same set of opponents when they also will use small fixed depth and after the games reset learning).
If you find that at depth 3 bigger values of pieces are better based on games when at depth 5 smaller value of pieces are better than you have evidence.
Otherwise I see no convincing evidence for your theory.
Uri

I arrived at my theory from first principles before I ever tested it in any way. It seems obvious to me. Unfortunately, it also does not seem to hold well.



my ftp site {remove http:// unless you like error messages}

[Slobodan R. Stojanovic]

Geschrieben von:/Posted by: Slobodan R. Stojanovic at 22 June 2004 03:11:58:
Als Antwort auf:/In reply to: Re: Death of a theory. Now I am puzzled. geschrieben von:/posted by: Dann Corbit at 22 June 2004 01:53:12:

I had a theory that piece values would continuously drop with each new depth of search.

I do not understand your theory.
Your theory suggest that evaluation that gives pieces smaller values is better when you search deeper.
Did you test it in games with fixed depth or only by test suites?
The only test that I fully trust is games.
You can use games at fixed depth and if the depth is small you can get thousands of games in a short time(you can play the same set of opponents when they also will use small fixed depth and after the games reset learning).
If you find that at depth 3 bigger values of pieces are better based on games when at depth 5 smaller value of pieces are better than you have evidence.
Otherwise I see no convincing evidence for your theory.
Uri

I arrived at my theory from first principles before I ever tested it in any way. It seems obvious to me. Unfortunately, it also does not seem to hold well.

In chess, best moves are not always the obvious ones.
SL - chess philosopher