By the way Mark, If you find you are not doing one of the things below exactly the same as I am, I would be willing to change my bot temporarily to see if it makes the result match yours (assuming it's trivial to make the change.)
- Don On Thu, 2008-10-23 at 13:53 -0400, Don Dailey wrote: > On Thu, 2008-10-23 at 15:00 -0200, Mark Boon wrote: > > Thanks again for more explanations. I think the AMAF is clear to me now. > > > > > > >> When you say you count all the playouts starting from an empty board, > > >> then > > >> I have no idea how our outcome can be different by 3-4 moves, > > >> which is > > >> coincidentally the average depth of a uniform tree of 1,000,000 moves > > >> on a 9x9 board. > > > > > > Well we are doing 2 completely different things - so it is > > > surprising to > > > me that we actually came that close. > > > > > >> > > > > > > This I don't understand. Both do a playout starting from an empty > > board, trying every move but ko-captures and eye-filling moves (using > > the same definition of eye). These are not completely different > > things. These should be EXACLTY the same things. The only thing > > different is that in my program the first few moves are selected by > > my UCT search instead of random selection as in your program. But I > > believe that does not affect the average game length one bit. > > I believe this might have a small effect on the game length, but it's > probably not the explanation. > > > > > When I just run playouts from an empty board I get the same average > > length, confirming that UCT selection does not influence the game- > > length. I didn't assume you were doing a UCT search. I didn't > > understand what kind of search you did exactly but in any case failed > > to see how it would change the average game-length. > > You are getting about 3.5 extra moves. I count pass moves including the > 2 at the end, and I assume you do the same but if you were not that > would bring your count down - so that doesn't explain the problem. > > This is exactly why I wanted to do this - to get verification between > bots. So we must explore these 2 possibilities: > > 1. I am doing it wrong. > > 2. You are doing it wrong. > > > The things to check for both of us are: > > 1. proper 1 point eye definition? > 2. simple ko testing (are you testing for simple ko?) > 3. proper node count accounting. > 4. properly random (in a perfectly uniform way.) > 5. N*3 max game length (this is probably not the issue even if it's > wrong) > 6. Do we stop after 2 passes? > 7. Never include a pass move in playout UNLESS no other move is > possible. > 8. suicide NOT allowed. > 9. PSK never tested in the playouts. > > > > > > This is still something I don't understand. Are there others who > > implemented the same thing and got 111 moves per game on average? I > > tried to look through some posts on this list but didn't see any > > other numbers published. > > So far I don't think anyone else has published. Please go down my > checklist above and I'll do the same. If I implemented it wrong, I did > it in all 3 bots I have completed so far, but that means little, they > were all almost direct ports of each other in C-like languages. > > - Don > > > > > > > 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/
