GradualTest and AdjustEval

[Peter Fendrich Peter Fend]

Geschrieben von: / Posted by: Peter Fendrich Peter Fendrich at 07 February 2002 23:08:52:
I downloaded the new version of GradualTest and found the (for me) new AdjustEval and got kind of interested without grasping the whole picture.
Odd Gunnar!
Please post some words about the intention with this program.
Testing personalities? Example?
Peter

Terra pages

[Odd Gunnar Malin]

Geschrieben von: / Posted by: Odd Gunnar Malin at 07 February 2002 23:35:00:
Als Antwort auf: / As an answer to: GradualTest and AdjustEval geschrieben von: / posted by: Peter Fendrich Peter Fendrich at 07 February 2002 23:08:52:

I downloaded the new version of GradualTest and found the (for me) new AdjustEval and got kind of interested without grasping the whole picture.
Odd Gunnar!
Please post some words about the intention with this program.
Testing personalities? Example?
Peter

Instead I would tell you why I made it and when I used it so you get a picture.
I downloaded the big qtest.epd from Dann Corbit (more than 60000 positions)
Run this through GradualTest with crafty as an engine. First with depth=1 and taking all right solution as input to a new test with depth=2 and so on to depth=8.
Command for GradualTest was:
GradualTest /i qtest.epd /g s1.epd /c "new\nsd 1" wcrafty.exe
GradualTest /i s1.epd /g s2.epd /c "new\nsd 2" wcrafty.exe
....
GradualTest /i s7.epd /g res.epd /c "new\nsd 8" wcrafty.exe
Then I took the resultfile as startfile and did the same with Yace.
The result was a file with ca. 1000 position.
This was used as an input for AdjustEval to tune my engine for the CCT4
I have ca. 10 parameters and it took me about 2 hour's to get the 'optimal' tuning.
Instead of tuning all parameters on one run, I did a 'binary search' with big step first and smaller and smaler until max was given.
Ex.:
set bishoppair|0|100|10
This would let it run through all testpositions with first:
set bishoppair 0
then
set bishoppair 10
and so...
The resultfile would give you the optimal setting, so I used this to smaller the step.
set bishoppair|78|88|1
When this was set I went to next parameter.
Some parameters was tuned together but most wasn'tr because of the timepresure.
I did go through this loop until no parameters had to be changed.
I know that you could get into top that isn't the max because the curve don't need to form nice with only one top, so the best is to let it adjust all parameters at once but this take much time so maybe if you adjust the parameters that deal with the same thing together or sort out what position are openings, midlegame and endgame culd minimize the parameters who need to be tuned on each set.
Odd Gunnar

[Dann Corbit]

Geschrieben von: / Posted by: Dann Corbit at 08 February 2002 01:00:15:
Als Antwort auf: / As an answer to: Re: GradualTest and AdjustEval geschrieben von: / posted by: Odd Gunnar Malin at 07 February 2002 23:35:00:
[snip]

Instead I would tell you why I made it and when I used it so you get a picture.
I downloaded the big qtest.epd from Dann Corbit (more than 60000 positions)
Run this through GradualTest with crafty as an engine. First with depth=1 and taking all right solution as input to a new test with depth=2 and so on to depth=8.
Command for GradualTest was:
GradualTest /i qtest.epd /g s1.epd /c "new\nsd 1" wcrafty.exe
GradualTest /i s1.epd /g s2.epd /c "new\nsd 2" wcrafty.exe
....
GradualTest /i s7.epd /g res.epd /c "new\nsd 8" wcrafty.exe
Then I took the resultfile as startfile and did the same with Yace.
The result was a file with ca. 1000 position.

Can I get a copy of this file?


My Ftp Site

[Odd Gunnar Malin]

Geschrieben von: / Posted by: Odd Gunnar Malin at 08 February 2002 01:54:56:
Als Antwort auf: / As an answer to: Re: GradualTest and AdjustEval geschrieben von: / posted by: Dann Corbit at 08 February 2002 01:00:15:

[snip]

Instead I would tell you why I made it and when I used it so you get a picture.
I downloaded the big qtest.epd from Dann Corbit (more than 60000 positions)
Run this through GradualTest with crafty as an engine. First with depth=1 and taking all right solution as input to a new test with depth=2 and so on to depth=8.
Command for GradualTest was:
GradualTest /i qtest.epd /g s1.epd /c "new\nsd 1" wcrafty.exe
GradualTest /i s1.epd /g s2.epd /c "new\nsd 2" wcrafty.exe
....
GradualTest /i s7.epd /g res.epd /c "new\nsd 8" wcrafty.exe
Then I took the resultfile as startfile and did the same with Yace.
The result was a file with ca. 1000 position.

Can I get a copy of this file?

Hi.
I sent you the file.
Take a look here, third line from bottom of the message.
Odd Gunnar