From: Reijer Grimbergen ETL GO JP> Date: 16 may 1997 Subject: Re: Real chess computing Larry Kaufman WIZARD NET> writes: > I don't think the branching factor is the main issue. The key point is > the difficulty of evaluating positions. In checkers, winning material is > everything. In chess, it is dominant, and most other positional factors can > be evaluated by simple rules and summed. In shogi, material is still very > important, but the relative importance of position vs. material varies > greatly thruout the game in ways that are very hard to express in clear, > programmable rules. In Go, even to count who is ahead in "material" (i.e. > points) is very difficult, as it's often very unclear which groups are alive > and which are not. This is why, I believe, that the best pc checker program > is near Championship level, the best chess program at Grandmaster level, the > best Shogi program at 1 Dan level, and the best Go program at perhaps 7-10 > kyu level. The branching factor is a minor consideration, since a larger > branching factor makes the game more difficult for both human and computer, > and also because the computer partially compensates by using the Alpha-Beta > algorithm. As I wrote in my previous post, evaluation function and branching factor are in my opinion the key differences between the two games. However, I disagree with you about the difficulty of programming both problems. I think it is not too difficult making a good evaluation function using rules for good and bad shape piece configurations. I think such an evaluation function only needs a couple of hundred rules to evaluate a position correctly. I do think there should be a different evaluation function for the middle game and for the endgame. The evaluation of a shogi endgame is very different from the evaluation of the middle game, with the king being the center of attention and material playing a minor role. This week I have been struggling with the moment at which an endgame evaluation should be invoked instead of a middle game evaluation. I hope to be able to build a good evaluation function in the near future to prove my point. As for the branching factor, a larger branching factor does not make the game more difficult for human players, but does make it more difficult for computers. Old cognitive evidence shows that strong game players only look at a couple of candidate moves in each position. This means that in chess (branching factor 35), shogi (branching factor 80) and Go (branching facto 250) the computer needs to do increasingly more work to compensate for the fact that the human branching factor for all of these games is less than 3 (of course only for expert players). I agree that the Alpha-Beta algorithm partly compensates, but it is the "partly" that makes the games more difficult. > I do agree that the success of computers (and Deep Blue's victory in > particular) in chess should tend to increase the popularity of shogi vs. > chess, especially since shogi programs are now good enough to be sparring > partners for the average player without being a threat to professionals for > the foreseeable future. I think that whoever wishes to convert chess > players to shogi should point to the superiority of humans over computers in > shogi as evidence that shogi is more of a creative, imaginitive game than > chess, which is more of a logical exercise. I do not agree. Even though grandmaster chess players sometimes like to see themselves as artists, chess in Holland can be found on the sportspages and that is in my opinion where it belongs. Chess is a competition with other human minds and the goal of beating others has not changed now there is a machine that can beat us all. I am not sure about the developments in chess research (this might decrease now the ultimate peak seems to have been climbed), but I doubt very much that shogi will win popularity over chess because of Deep Blue's victory. Reijer -- Reijer Grimbergen Electrotechnical Laboratory Palcious Tsukuba 302 1-1-4 Umezono 1-24-8 Ninomiya Tsukuba-shi, Ibaraki-ken Tsukuba-shi, Ibaraki-ken 305 JAPAN 305 JAPAN E-mail: grimberg etl go jp Tel: 0298-59-1606 WWW: http://www.etl.go.jp:8080/etl/suiron/~grimberg Tel: +81-(0)298-54-5080 extension: 67431 Fax: +81-(0)298-58-5918