to point out the known -- if you're trying to score positions that are very early in the game, you are trying to solve the whole problem of the game of go at once. unfortunately, that is very hard.
to belabor this for a moment: imagine that you had a perfect score estimator for any given board position. you could run it ( < 19x19) times, with each next possible move, in order to evaluate what the best next move is. i.e. you would have a perfect go player. at the very end of the game, this is a tractable (but difficult, as other people have pointed out) problem for the japanese ruleset. most bots score chinese because you can force the board into a much more easily scorable position without suffering point penalty. even a few (say, 20 or less) moves away from the end of the game, this is hard -- the correct order of yose is nontrivial. s. On Sun, May 27, 2012 at 11:45 AM, Don Dailey <[email protected]> wrote: > > > On Sun, May 27, 2012 at 1:36 PM, Nicolas FRANCOIS <[email protected]> > wrote: >> >> Hi. >> >> I've been studying computer go for a while now, and would like to >> experiment on some ideas. I have one (well, in fact, two) big problem >> though : I can't figure out how to write a correct scoring procedure, >> which, I think, is linked to the problem of life and death. >> >> Could you give me some advices on readings on those subjects >> (especially deciding life and death), or some examples of well written >> codes on the same subject ? > > > Scoring correctly using Japanese style rules is very complicated and > difficult to do well. MCTS programs basically play the game out to the > bitter end using Chinese style scoring with a simple eye rule to prevent the > players from moving directly into their own single point eyes. Then > scoring is trivial. > > Random playouts (subject the eye rules I mentioned) is one way to get a > sense of what lives and dies if you keep statistics. It's far from perfect > but it's reasonable and of course it's more reasonable when combined with a > little knowledge in the playouts. The idea is that if some group > consistently lives after doing a few hundred or thousand random play-outs, > it's probably safe. If it consistently dies, it's probably dead or > certainly unresolved. You can almost be sure of unconditional life if it > never dies. Without a lot of complexity that is a sort of a first order > way to measure life and death - very flawed but also very simple to > implement. I'm not sure any program does this 100% perfectly because > it's a non-trivial problem. > > You can probably test this by getting a large sample of correctly scored > positions and experimenting with various algorithms and observing what goes > wrong and what works. > > Don > > > > > >> >> >> Thank you. >> >> \bye >> >> -- >> >> Nicolas FRANCOIS | /\ >> http://nicolas.francois.free.fr | |__| >> X--/\\ >> We are the Micro$oft. _\_V >> Resistance is futile. >> You will be assimilated. darthvader penguin >> _______________________________________________ >> Computer-go mailing list >> [email protected] >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
