On Thu, 2008-10-23 at 09:38 -0200, Mark Boon wrote: > Don, > > I have figured out the discrepancy in the average game length. As > playout length I count from the start of the game, which gives me > 114-115. I believe you count from the starting position where the > playout starts. Because when I modify my code to do that I also get > 111 moves per game on average. > > But I believe counting from the start is more meaningful if you're > going to use the game-length as a reference number. Because otherwise > you might get a different number depending on how explorative your > search is. If your search tends to explore deeper lines in favour of > wider lines you get a different number. Or at least you'll have to > include the exploration-factor K in your UCT tree as a criteria.
I think you are misunderstanding how this works. I am defining a very simple non UCT bot and you are using terminology that implies something different that I am specifying. My timing is based solely on the opening position so how would it come out any different? We are both counting from the start of the game in this specific case. (I'll make sure my spec makes this unambiguous.) You also talk about how "explorative" your "search" is. The spec doesn't talk about a search or exploration, you just do 1 million uniformly random playouts from some specified position (the starting position.) The score should be the total black wins divided by the total number of games played and for this specific test it should be 0.5 komi. I will build a more sophisticated test soon that you simply run your bot against. The whole idea of this (for me) was to create the simplest reasonable go benchmark that played a "reasonable" game of go (depending on your definition of reasonable) so that hopefully many people might be willing to go to the trouble to implement it. I'm not going for a sophisticated UCT tree based algorithm, because that would require a considerably larger amount of effort for people to write, and it would also be significantly more difficult to define precisely. Having said that, you could easily take a UCT program and with very little effort have a "simple" mode that does what we are asking. The simple bot also serves as a stepping stone for beginners to start implementing their own more capable bot because they can be confident that they got the details correct and bug-free (more or less.) So you could grow a UCT program from this. - Don > > Mark > > > On 22-okt-08, at 20:57, Don Dailey wrote: > > > On Wed, 2008-10-22 at 20:29 -0200, Mark Boon wrote: > >> On Wed, Oct 22, 2008 at 6:07 PM, Don Dailey <[EMAIL PROTECTED]> wrote: > >>> > >>> For one thing, komi is different. I used 0.5 for running this > >>> test. > >>> > >>> I would have use 0.0 but some implementations don't like even > >>> komi's. > >>> > >> > >> But the komi should have no effect on the playout length. I started > >> out with 103 moves, but that was because of a mercy-rule. Without it > >> it's 114. You have 111 rather consistently. And I assume you don't > >> use > >> a mercy-rule. Nor super-ko checking, is that correct? > > > > The score is what I was looking at, but you are right about the > > play-out > > length because the play-outs don't care what komi is. > > > > Do you use the 3X rule? Have you checked the other specs? > > > > My guess is that your playouts are not uniformly random - that is very > > easy to get wrong. Of course that is just a wild guess, it may very > > well be something completely different. > > > > - Don > > > > > > > >> > >> Another thing I have never looked at is AMAF. But that also shouldn't > >> affect playout length I assume. > >> > >> By the way, thanks for all the pointers to 'git from everyone.' It's > >> new to me and at first glance the specs look good, so I'll definitely > >> give it a go. > >> > >> Mark > > _______________________________________________ > 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/
