budd@arizona.UUCP (12/30/83)
Something I find curious in seismo!flinn's article from AGJ is that it doesn't mention Joseki. It would seem that one fairly easy criterion one could use to measure the ability of computer GO systems would be their ability to respond correctly to common Joseki situations, say those from Kosugi and Davies book ``38 Basic Joseki''. Does anybody know if this question has been addressed in any published systems?
jim@randvax.ARPA (Jim Gillogly) (12/31/83)
-------- I don't think performance on Joseki situations is a good overall measure of the skill of a program. I can imagine a program that does a local search for tactical problems and life-or-death situations (you didn't say Fuseki, so I don't have to look at interactions with other corners) and has stored all the lines in Ishida's Joseki dictionary (or some more massive work). That's all reasonable technology from the chess programming world. The Zobrist and Ryder programs each played a credible opening. However, none of the hacks we chess programmers found to "cook" chess will work in the middle game. In chess you can make up for bad strategic thinking with superior tactics, but that won't hack it in Go. I'll go further: I bet I could write a program that would play Joseki at a 2 kyu (amateur) level (i.e. as well as I could), but we need all new hacks to achieve even 10 kyu in the middle game because the strategic concepts aren't well understood. (Ever try to tell a program what "sabaki" means?) Jim Gillogly I/ / randvax!jim I_/ jim@rand-unix I
ucbesvax.turner@ucbcad.UUCP (01/05/84)
#R:arizona:-702500:ucbesvax:38500004:000:697 ucbesvax!turner Jan 4 21:13:00 1984 Yeah, how *does* one define sabaki to a computer, anyway? I've taken it to mean that one's attacks and defenses can take many forms, and that both the attacks and defenses can be deferred for several moves, in a pinch. This seems very combinatorial. To define a "light shape" as a local collection of "light shape" moves begs the question. One must measure the cluster's ability to withstand attacks from all angles, with a possibly large (2-4) number of approach moves from the opponent before any response is necessary to avoid total loss. I would want a heavily-optimized tesuji-reader before even embarking on the design of a sabaki evaluator. --- Michael Turner (ucbvax!ucbesvax.turner)