I use CFG distance only in the tree. The playout uses the concept "adjacent" which is trivial to compute.
The only distance I am concerned about is the distance to the last move, and there are only 4 classes: distance 1,2,3, and >3. So it is cheap. The advantage is in semeais. Moves at CFG distance 3 are the outside liberties of the opponent's string. There was not a big difference compared to the other two heuristics. I found that - CFG is best - max(dx, dy) + (dx + dy)/2 is second best - Manhattan is third. Brian -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Jacques Basaldúa Sent: Monday, January 24, 2011 2:41 PM To: [email protected] Subject: [Computer-go] Computing CFG distance in practice Hi, I got a lot of improvement recently (something you all did long time ago) with proximity heuristics. I am using Manhattan distance: d = max(dx, dy) and d = max(dx, dy) + (dx + dy)/2 where dx = abs(ex - ox) and dy = abs(ey - oy) But many people report CFG distance to be superior. What do you do: a. Compute it in root. Then build a lookup table and use the LUT during playouts and tree search. b. Compute the shortest path from (ox, oy) to (ex, ey) connected by the stones on the board each time you need to evaluate a distance. I don't like a because it looks inefficient as the board changes a lot during the search. I don't like b because it looks computing intense unless there is some smart idea I am missing. Jacques. _______________________________________________ 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
