|
From: | SP LEE |
Subject: | Re: [gnugo-devel] Re: Full board strategy component |
Date: | Thu, 31 Jul 2003 17:06:15 -0700 |
As a very rusty Shodan, if that, my ability to judge the reliability of the data being generated is limited, so I really appreciate feedback! If indeed J7 is sente, and not an overplay, is this because it is splitting the various black groups? The FBS algorithm's first guess is that the lone White stone peaking out at J5 (K5 being totally blocked off) is ratherout-numbered, at present, and that the split groups are both fairly healthy. Some prepatory moves could raise the viability of a J7 play. In fact, the highestrecommended move, L13, is deemed to improve the J7 rating slightly. One thing I'd like to see is if known sente points could be fed to the Full Board Strategy component. And, if known boundaries of groups could be integrated in a meaningful way. How much improvement is necessary (and what would it look like) before therecommendations provided by this sort of component becomes helpful? I assume "usually reasonable" full-board suggestions are better than random choices, if they can be derived at a minimum cost in cpu's, and they can be overridden bythe more precise local reading algorithms. - Phil Freihofner
Maybe you are stronger than me. Playing J7 directly seems to be indeed overplay. Of course when there are more white stones around the center the situation will be different and the numbers will also change.
Unfortunately I'm almost a new comer and can't say when the code will be useful. As far as I know, there is a influence module in gnugo and it does the similar thing. There is also a cosmic patch which generates a different version of the influence module.
However, I do think gnugo need some improvement in this area. Take an example from the game posted by Dan in http://mail.gnu.org/archive/html/gnugo-devel/2003-07/msg00125.html. At move 49, gnugo finds the best move F9 which is only valued 11 points. At this early stage of the game, there are many big moves almost everywhere. With the cosmic patch, gnugo finds a good move at N7 valued 40 points. When I replayed the game with the cosmic patch, there are still some cases the best moves are less than 10 points in the early stage.
Apparently, I think, when the best move is valued not so high, it means gnugo doesn't know very well what to do in the specific situation, especially possible followup moves may not be seen very well. I hope tuning the influence module can improve gnugo in this aspect.
SP Lee _________________________________________________________________MSN 8 with e-mail virus protection service: 2 months FREE* http://join.msn.com/?page=features/virus
[Prev in Thread] | Current Thread | [Next in Thread] |