|
From: | Joseph Heled |
Subject: | Re: [Bug-gnubg] Problems with latest Windows Build |
Date: | Wed, 03 Nov 2004 20:26:26 +1300 |
User-agent: | Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.7) Gecko/20040616 |
Ned Cross wrote: else is having the trouble.
5) Of course the biggest problem of all is the sometimes poor results of the pruning net in rollouts combined with the inability to use the old evaluations at reduced speed. 2-ply rollouts performed at 50% checker and 33% cube speed have proven quite accurate and considerably faster than 2-ply 100%. 2-ply prune rollouts, while faster per number of trials, are not proving as accurate, and have very high standard error rates, making them possibly unusable.
Can you provide a convincing proof of that? Of course you are free to use whatever setting you like, but the reduced evaluations are worse in every test I did. For example it has about 10 times the error rate when tested over the contact benchmark. (I just did several thousand positions. will do the full test when I get more computing power).
Of course those are done using my reduced code (only 33% for moves), and GNUbg might be different implementation, I have not looked at this for a long time.
But I simply can't believe large differences. And you want to compare full 2ply with pruning and full2 ply vs. reduced, not against one another.
-Joseph
[Prev in Thread] | Current Thread | [Next in Thread] |