Re: Рейтинг программ для игры в Го
|
|
Кто-нибудь сравнивал MoGo c Many Faces of Go при условиях, близких к приведённым мной (см. от 29, August, 2008)?
А то последняя на 32 ядрах обыграла MoGo на неизвестном (мне) количестве ядер (http://www.grappa.univ-lille3.fr/icga/tournament.php?id=181).
Хотя комментарий к другому выигрышу (где MoGo играл только на 2 ядрах): "CzechBot and ManyFaces1 played the sequence shown to the left. With Black 15 ManyFaces1 started a ladder which clearly does not work. CzechBot presumably knew that it did not work, and felt that it could safely tenuki with move 20. But it was wrong: after the 23-24 exchange, the ladder does work, and ManyFaces1 captured the laddered stones. This sequence looked to me like one that might occur between a human dan-player and a human 3-kyu (CzechBot is in fact rated at 3-kyu on KGS). Judging from this sequence, and from others later in the game, CzechBot was being thoroughly outplayed. Of course, though CzechBot is a version of MoGo, it was only playing on a dual processor, while ManyFaces1 was running on a fast 32-core system, with 24 times the total power; so we should expect ManyFaces1 to play better. However, it appears that MoGo, a fairly pure MC program, scales better than ManyFaces, which uses MC but incorporates the knowledge of the old classical ManyFaces. MoGo's creators have found that it plays better the more processors it can use, without showing any signs of plateauing; and its best results have been achieved on an 800-node system. But Many Faces not only fails to improve with more than 32 cores, it plays worse. Even 32 may be too many for it: its author David Fotland said in the course of this game "32 core is too fast ... it does so many playouts the Many Faces knowledge is less effective".", - свидетельствует об ограниченной масштабируемости Many Faces, а значит о более низком абсолютном рейтинге, о котором идёт речь в данной теме. Тогда как MoGo на 800х4 ядрах обыграла 8p игрока на доске 19х19 с форой лишь в 9 камней, но проиграла с форой в 7 и "15 minutes each sudden death" (http://www.computer-go.info/h-c/index.html).