Piece value testing

[David Dahlem]

I'm experimenting with different piece values in my engine. I really don't want to run thousands of games to test these settings. Is there a good test set that will give a somewhat accurate idea of the best settings?

Regards
Dave

[hgm]

I am afraid there is no substitute for massive testing here.

I have just been on vacation for two weeks, and in my absence I had my engine play (in self play) nearly 8000 games, starting from 58 different starting positions (29 with white and then 29 with the conjugated position for black, to eliminate the white/black bias). All these 29 positions were (supposedly) equal to within a Pawn. Where I expected the difference to be close to half a Pawn, I tried both the version with and without Pawn odds for the side with the stronger piece combination (e.g. I tried both Q vs RR and QP vs RR, and NN vs RP as well as NN vs RPP).

The positions I generated by simply deleting the unequal material from the opening array. Preferably on the Queen side, to interfere as little as possible with castling possibilities. If the difference included Pawns, I tried to delete those Pawns that gave the smallest compensation in terms of development (e.g. delete f2/f7).

I have not analyzed the results yet, but each position is played about 250 times, making the statistical error about 2.5%. As many pieces have been tried several times, and the Pawn-odds score is about 63% at this time control, this should be enough to derive piece values to an accurracy of about 100 cP.

[Jim Ablett]

Hello David,

'AdjustEval', included in Odd Gunner Malin's 'Gradualtest' may be something to try.

http://home.online.no/~malin/sjakk/

rgds,
Jim.

[David Dahlem]

Hello David,

'AdjustEval', included in Odd Gunner Malin's 'Gradualtest' may be something to try.

http://home.online.no/~malin/sjakk/

rgds,
Jim.

Thanks Jim

I'll check it out.

Regards
Dave

[David Dahlem]

Hello David,

'AdjustEval', included in Odd Gunner Malin's 'Gradualtest' may be something to try.

http://home.online.no/~malin/sjakk/

rgds,
Jim.

Hi Jim

Have you used this tool? It doesn't output any results for me. I tried it with Crafty, and still couldn't get it to output any result in the results.txt file.

Regards
Dave

[Jim Ablett]

Have you used this tool?

No, never used it.

Maybe Odd Gunnar will pick up on this thread and help.

Jim.

[David Dahlem]

Have you used this tool?

No, never used it.

Maybe Odd Gunnar will pick up on this thread and help.

Jim.

I sent Odd Gunnar an email, but haven't had a reply yet.

Regards
Dave

[Orlov]

What the problem David?

run cmd, then

GradualTest.exe /i arasan8.epd /o result.txt /t 1 Wcrafty.exe (path must be included)

working fine!

[David Dahlem]

What the problem David?

run cmd, then

GradualTest.exe /i arasan8.epd /o result.txt /t 1 Wcrafty.exe (path must be included)

working fine!

AdjustEval.exe is the tool with the problem, not GradualTest.exe. See the previous messages.

Regards
Dave