I think all uct/mc programs will have this problem unless they add something extra to avoid it. During playouts using random moves the playouts will think that the ladder usually works. Local 3x3 patterns aren't enough since they can't tell the difference between ladders that work and ladders that don't work. Some programs read ladders during playouts so they know the local result and can play correctly. Many Faces doesn't read ladders during playouts, since I think it's too slow, but I found a different solution to the problem.
Mogo and Crazystone must some solution to this problem, but I don't know what it is. This should only show up for long ladders, since the UCT tree part of the search will read a ladder correctly. David From: computer-go-boun...@computer-go.org [mailto:computer-go-boun...@computer-go.org] On Behalf Of terry mcintyre Sent: Saturday, May 02, 2009 9:49 AM To: computer go Subject: [computer-go] a ladder example I promised an example of a monte carlo program mistakenly starting a ladder; here it is. I played white; Leela had a 2 stone handicap and 45 minutes on the clock. Leela's move 32 initiates a ladder. Unfortunately for Leela, I have a ladder breaker at D16. Leela's analysis was optimistic when it initiated the ladder, but a few moves later, it became aware of the looming shadow of a great doom. There are a few large captures later in the game. The corner fight at the end had me biting my nails. Terry McIntyre <terrymcint...@yahoo.com> "Here you can see how slightly a people needs to be governed." - Carl Schurz, immigrant, 1848. What happened since?
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
