David Fotland suggested another rule, tell me what you think. His rule is to stop the game at or beyond move N*N*2 where N is the board size and score the board no matter what. But in the play-outs he first plays 1 move before making this test.
If a game actually lasts that long, the play-outs are not worth much but the game is probably over anyway and the board gets scored as is. My own implementation depends on there being no eyes bigger than 1 point, because it's fast and easy to score a board that has only 1 point eyes. But a simple modification makes it work without any appreciable speed sacrifice: If a game ends beyond move N*N*2 then use the slow scoring routine that doesn't assume the board is complete. It should be very rare that this routine has to be used. - Don On Sat, 2008-10-11 at 16:15 -0400, Don Dailey wrote: > On Sat, 2008-10-11 at 21:46 +0200, Denis fidaali wrote: > > > > On point 13. > > 1 stone captures, may eventually be "hard" for some implementation. > > I think using game length as a criterion gives more freedom. > > I definitely considered that. My own program returns the number of > stones captured (and a list of those stones if I ask for it) and so it's > easy for me, but wonder if there are programs where this would be > difficult to detect? > > My arguments in favor of the rule is that it's not quite as artificial > as setting some arbitrary limit and I think this would not unduly delay > the terminations of games. With a limit of hundreds of moves, some > games would go hundreds of moves beyond what is necessary. > > Are there any MC programs out that cannot detect whether they made a one > stone capture and it would be unduly difficult to fix this? > > > > > > > > Then you have to specify what to do with those pathological games anyway. > > Do you score them "as they are", or do you drop them. I do count them. > > You score them as is. You may not get the "correct" score, but after > all these ARE random games and we expect this to be extremely rare. > > > > The rule for counting is probably quite standard too. I give a point for > > each stone present. Then for each empty stone, i give a point to the side > > who has all orthogonal control of it, if any. > > Yes. Tromp/Taylor which are Chinese rules. > > > > > > I usually implements super-ko checking very late, if i do at all. > > > > I often start by selecting either the first or the last "best" move, > > rather than picking one at random. Although picking at random > > makes me more comfortable somehow. It could be interesting > > to see how much difference it makes. > > Of course everyone is free to implement any variation they want, > controlled by switches and such. The idea is to have an extremely > reliable benchmark and this will also be very helpful for new developers > as this also would serve as a good "first program" for new developers > and they could have confidence that they got all the details right. > > > > > > Finally, about the size. Is it supposed to be 9x9 only ? > > It can be anything you want, but I suggest that you develop a program > capable of playing on any board size. > > Having said that, you can get a little more speed if you code to just 1 > specific board size. Lazarus compiles to any ODD board size but it's > fixed at compile time and I have to recompile to get other board sizes. > I doubt this optimization is worth very much and it's annoying that I > have to fetch a different binary in order to play other board sizes. > > Also, I don't really keep up with how new processor families change the > rules. I know that I USED to get a nice speedup several years ago by > using as many constants as possible (as opposed to variables) but I have > no idea whether (or how much) this is true today. > > > 9x9 only gives more freedom to fine tune for this size. > > It may make the implementations less useful, if someone > > knows another sizes the light-policy could succeed on. > > (like maybe studying 7x7 and such ?) > > How do you define "succeed?" It works on all board sizes, it just > works better on smaller boards. You could equally say it doesn't > work on any board size since there are ways to build stronger programs. > > > - Don > > > > > > > _________________________________________________________________ > > Email envoyé avec Windows Live Hotmail. Dites adieux aux spam et virus, > > passez à Hotmail ! C'est gratuit ! > > http://www.windowslive.fr/hotmail/default.asp_______________________________________________ > > computer-go mailing list > > computer-go@computer-go.org > > http://www.computer-go.org/mailman/listinfo/computer-go/ > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
