I finally found time to try fill_board, and it made the program a little weaker.
I tested against gnugo level 10, 19x19 board, vs Many Faces with 10K playouts per move. Before fill_board it won 1183 of 1411, 84%, 1.9% 95% confidence interval. After it won 1375 of 1676, 82%, 1.8% 95% confidence interval This is not enough games to prove that fillboard is weaker, but it is likely. My playout move generator implementation: First, make one of the local tactical suggested moves If none, make one of the 3x3 local pattern moves If none, sometimes make a capture (more likely for bigger groups) If none, fill_board: pick a random empty point. If it is not on the 1st or 2nd line and there are no stones in 8 adjacent points, play it Otherwise, play a random move (constrained to not fill eyes, etc) I remember that fill_board in Mogo is important to add variety when there are many playouts, so perhaps the result would be different with longer time limits. I still use vanilla UCT with a pretty large exploration constant, so perhaps UCT is giving me more exploration already than Mogo so I don't need this. David _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
