[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Bug-gnubg] prunning nets commited

From: Joseph Heled
Subject: [Bug-gnubg] prunning nets commited
Date: Tue, 12 Oct 2004 22:02:28 +1300
User-agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.7) Gecko/20040616

I decided to check in my changes. I apologize if I broke anything. guess we can always roll it back.

What you need to do :
1.Open gnubg.weights (0.14) in a text editor, and add 'add-to-weights' AT THE END of it.
  2. change the version number (top of file) to 15.
  3. save as new gnubg.weights

4. Add add-to-gnubgautorc to your .gnubgautorc file (assuming you do want to use the speedup).

I added a command line version to turn it on/off. I did not add a GUI option since I don't know how, but it will be nice to have.

What you gain is speed. You need to check out for yourself how much.
What you lose is that results are different, presumably less accurate. Myself, I think it is virtually identical. The differences are in the order of 0.1%, and if they change any decision it is unlikely to know which is correct without a rollout anyway. If you find cases which contradict that please let me know.

There are two reasons why evaluations are different. When computing the moves in a >0 ply, the pruning (a fast but of lesser quality) net picks up 10 candidates - and the move is the one with the best 0-ply evaluation. So, it is possible that the best move did not make the list. The average error, BTW, is 0.0003 for such an error, but potentially it gets multiplied by 441 in a 2-ply evaluation.

The second reason is that for speed and minimal overhead the best move is chosen using the gammon weights, even when evaluation is cubefull. Again, my experience tells me this is almost a non-issue when picking out moves inside an evaluation, since the two (i.e. the best by g-weights and the cubefull best) are very close to each other, and even if one turns to double/pass and the other to double/take, the equity is still very close.

Please give it a try and see how you feel. I think you will see no difference and just be glad for the speedup. If you used any form of "reduced" evaluations, you should stop using them right away and enjoy much better quality at an even lower price.


Attachment: add-to-weights.gz
Description: GNU Zip compressed data

Attachment: add-to-gnubgautorc.gz
Description: GNU Zip compressed data

reply via email to

[Prev in Thread] Current Thread [Next in Thread]