|
From: | Harri Salakoski |
Subject: | Re: [gnugo-devel] Semeai node limits |
Date: | Fri, 17 Mar 2006 18:01:33 +0200 |
Narugo project has gnugo java porting sub-project and I am planning to do similar what you descriped. Goal is use computing power outside game context to tune patterns and create intelligence inside game engine instead so much focusing search speed. http://sourceforge.net/projects/narugoHave anybody done anything like genetic algorithms or autotuning on pattern values??
t. harri----- Original Message ----- From: "Jens Yllman" <address@hidden>
To: "GNU Go development" <address@hidden> Sent: Friday, March 17, 2006 9:10 AM Subject: Re: [gnugo-devel] Semeai node limits
On Mar 13, 2006, at 12:22 PM, Gunnar Farnebäck wrote:Dan wrote:Perhaps the semeai nodes should be increased but 2000 is too much. The current default is 500.<snippage>With 1000 nodes I measure a regression slowdown by 9% and with 1500 a slowdown by 17%. /GunnarRegressions are not run often ... it would be more pertinent to ask a) whether gnugo picks better moves with a higher node limit, and b) whether the cost time-wise is acceptable for ordinary usage. If the moves are not better, that would be a topic for investigation; it suggests a brittle algorithm which has been tweaked to work at certain settings, but not under deeper study.
For me it often feels that many changes in pattern values and patterns are to 'highlight' something soon enough with the current depth/nodes/cache settings. Not to realy make GNU Go stronger. Because changing depth/nodes/cache makes GNU Go as clueless as before. That not to say all work that is done is stupid. But maybe more work should go into more 'intelligence'. It feels little bit stupid of me to say this who never have time to contribute. Maybe a study of regressions that change when changing depth/nodes/cache would show that pattern values should not be static. Maybe some pattern values should maybe slowly decrease when more analyzing is done. Or maybe there should be patterns that have diffrent versions depending on how much analyzing is done. With that I mean that maybe some patterns need more restrictions than today. But at lower settings not have time for that. Just some thoughts. Have anybody done anything like genetic algorithms or autotuning on pattern values?? Jens _______________________________________________ gnugo-devel mailing list address@hiddenhttp://lists.gnu.org/mailman/listinfo/gnugo-devel
[Prev in Thread] | Current Thread | [Next in Thread] |