A strange Beowulf Beta

[Dann Corbit]

Geschrieben von:/Posted by: Dann Corbit at 03 June 2004 00:08:44:

Try this one out:
ftp://cap.connx.com/pub/chess-engines/n ... eowulf.exe
You may find it amusing.



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

[David Dahlem]

Geschrieben von:/Posted by: David Dahlem at 03 June 2004 16:35:15:
Als Antwort auf:/In reply to: A strange Beowulf Beta geschrieben von:/posted by: Dann Corbit at 03 June 2004 00:08:44:

Try this one out:
ftp://cap.connx.com/pub/chess-engines/n ... eowulf.exe
You may find it amusing.

What is amusing about it? The only thing i notice is a huge difference in file size, 272 kb vs 1.18 mb.
Regards
Dave

[Dann Corbit]

Geschrieben von:/Posted by: Dann Corbit at 03 June 2004 19:41:49:
Als Antwort auf:/In reply to: Re: A strange Beowulf Beta geschrieben von:/posted by: David Dahlem at 03 June 2004 16:35:15:

Try this one out:
ftp://cap.connx.com/pub/chess-engines/n ... eowulf.exe
You may find it amusing.

What is amusing about it? The only thing i notice is a huge difference in file size, 272 kb vs 1.18 mb.

For every evaluation weight (117 different terms), a computer decided what it;s value should be, not a human. It will be interesting to see if it does better than formerly.



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

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 04 June 2004 03:23:17:
Als Antwort auf:/In reply to: Re: A strange Beowulf Beta geschrieben von:/posted by: Dann Corbit at 03 June 2004 19:41:49:

Try this one out:
ftp://cap.connx.com/pub/chess-engines/n ... eowulf.exe
You may find it amusing.

What is amusing about it? The only thing i notice is a huge difference in file size, 272 kb vs 1.18 mb.

For every evaluation weight (117 different terms), a computer decided what it;s value should be, not a human. It will be interesting to see if it does better than formerly.

Interesting.
I believe there is an unexploited (by most) potential in such techniques.
I would not even be surprised if this was one of the big secrets of commercial
programs.
Perhaps you could give some examples of how the weights changed, e.g. material
values, passed pawn values?
Can you tell if the method is stable or if some of the parameters diverge?
Perhaps if there are divergencies it's a symptom the solver is trying to do an impossible job of fitting.
One might also try strange initial values to test if the method works at all, ie. rook less worth that a pawn and similar nonsense.
-S.

[Dann Corbit]

Geschrieben von:/Posted by: Dann Corbit at 04 June 2004 04:22:32:
Als Antwort auf:/In reply to: Re: Gothmog 0.4.8 en-passant bug geschrieben von:/posted by: Sune Fischer at 04 June 2004 03:23:17:

Try this one out:
ftp://cap.connx.com/pub/chess-engines/n ... eowulf.exe
You may find it amusing.

What is amusing about it? The only thing i notice is a huge difference in file size, 272 kb vs 1.18 mb.

For every evaluation weight (117 different terms), a computer decided what it;s value should be, not a human. It will be interesting to see if it does better than formerly.

Interesting.
I believe there is an unexploited (by most) potential in such techniques.
I would not even be surprised if this was one of the big secrets of commercial
programs.
Perhaps you could give some examples of how the weights changed, e.g. material
values, passed pawn values?
Can you tell if the method is stable or if some of the parameters diverge?
Perhaps if there are divergencies it's a symptom the solver is trying to do an impossible job of fitting.
One might also try strange initial values to test if the method works at all, ie. rook less worth that a pawn and similar nonsense.
-S.

I iterate over 12000 chess positions with a single known solution.
I start with a depth of 2 and solve for every parameter. I do it again a few times and the parameters generally settle down to some value.
Then I set the depth to 3 and repeat the experiment.
Then I set the depth to 4 and repeat the experiment.
I take each parameter from a value that must be too low and iterate with a fixed increment until it must be too high. I get (for each point) the number of positions actually solved and also the time it took to solve them. Sometimes, there is a clear, parabolic curve with a maximum for the points. Here is an example:
rook_score=525 at 4; stddev=2.561553 : -0.0026725*x^2 + 2.80831*x + 4809.58
(x=471.000000, y=5540.000000), t=1147.000000
(x=492.000000, y=5542.000000), t=1147.000000
(x=513.000000, y=5550.000000), t=1146.000000
(x=534.000000, y=5546.000000), t=1145.000000
(x=555.000000, y=5545.000000), t=1144.000000
(x=576.000000, y=5540.000000), t=1144.000000
(x=597.000000, y=5534.000000), t=1143.000000
(xmax=513.000000, ymax=5550.000000), xmax seen verses curve xmax [513 525.909]
With a value of 471 centipawns for a rook, 5540 of the problems were solved with a search depth of 4 plies. It took 1147 seconds.
As you can see, the most positions were solved when 513 was chosen for the centipawn value of a rook. If we fit a parabola using least squares fit through the data points, we arrive at 525 as the best score (forget about the .909 because I have already added 0.5 to the score so the integer rounding will work).
Other times, the solution count does not change or jumps around randomly, but the time is affected. An example would be null move pruning.


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