|
From: | Alain Baeckeroot |
Subject: | Re: [gnugo-devel] 3x3 point never played |
Date: | Mon, 24 Jan 2005 12:40:01 +0000 |
User-agent: | KMail/1.7.1 |
Le Jeudi 13 Janvier 2005 03:45, Gunnar Farnebäck a écrit : > Alain wrote: > > On a 9x9, gnugo 3.6 against himself > > The 3x3 point is not played until the end game, > > That is certainly too late. It seems more appropriate to do it at move > 10, or maybe move 8. Relevant code is around line 355 of fuseki.c: I tried differents things on that file: - play the sansan at move 10 and let gnugo 3.6 finish the game.... with various params (size cache, --large-scale, owl depth semeai depth) Here are my first conclusions (i can give you more details , and sgf files) - I (6kyu) thinks black make a global mistake by choosing that joseki: W can destroy B startegy by playing sans san in 8 or 10, and win the game. - With default params, gnugo 3.6 live in the corner and win, so W should have played that for win. - default + large-scale, gnugo 3.6 is dead in the corner, so W is right not having played. - with compile options: CFLAGS="-march=athlon-xp" ./configure --enable-cache-size=512 --enable-experimental-semeai --enable-semeai-node-limit=8000 --enable-owl-threats --enable-owl-node-limit=16000 --enable-large-scale --enable-experimental-reading then gnugo W is alive, and win should have win the game. > > /* On 9x9, only play open corners after the first move if nothing > * else useful is found. > */ > if (board_size == 9 && stones_on_board(color) > 0) > empty_corner_value = 5; > > I remember that it was necessary to be careful not to just blindly > occupy corners on 9x9, but this seems way conservative. Maybe an empty > corner value of say 12 or 15 would be more appropriate. I like the lot of idea gnugo has, and i m afraid that rising the value will kill this "imagination". I thinks if the defensive engine get stronger, gnugo will be able to read that it can live there, and so will play it for victory :) > > > and then is considered as dead, whereas it can live !!! > > Well, lets say that I have no difficulty seeing how GNU Go might go > wrong in reading this. With such a strong position black has all sorts > of nasty moves available which could normally quickly be ruled out. > Try "gnugo -l SanSan-is-Alive.sgf -L 26 --decide-owl G3 -o vars.sgf" > to see what moves were actually considered. ;) Well , humm, sansan might be alive if played early, and is probably dead in the end game. There are still hallucination from gnugo in the end game, when white play negative move (-2 or -3). This is due to difference of strengh of the attack module and the defense. The attacker thinks it can do something, but once the move is played and B has answered, the life of the group is in the hand of the defense module. As they are not identical, the defense module has to defend a W group and doens't know how to do it so the group is considered dead, so it dies of oldness, and will not go to Valhalla ;) (i have other sgf file to explain that clearly, i prepare an other mail with all that clearly exposed) Alain
sansan-retry-3.6.sgf
Description: Text document
sansan-retry-large-scale.sgf
Description: Text document
sansan-retry-large-scale-3.6.sgf
Description: Text document
sansan-retry.sgf
Description: Text document
[Prev in Thread] | Current Thread | [Next in Thread] |