Monte-Carlo Tree Reductions for Stochastic Games - Technologies and Applications of Artificial Intelligence - page 235

Tabl e 2. Using minimax at the end

Policy Against rand-mm

win lost draw

with 0.01[sec] per move
move-group-random-mm 169

1 330

move-group-cycle-R-mm 186

2 312

move-group-cycle-M-mm 166 0 334

group-nodes-mm 24 5 471

chance-nodes-mm 240 0 260

with 1[sec] per move
move-group-random-mm 265

0 235

move-group-cycle-R-mm 290 0 210

move-group-cycle-M-mm 293 0 207

group-nodes-mm

121 0 379

chance-nodes-mm

310 0 190

players apply minimax to find the best move. Policies are used when pieces
are unrevealed and otherwise a minimax search is done. These modified players
move-groups-random-mm , move-groups-cycle-R-mm , move-groups-cycle-M-mm ,
group-nodes-mm and chance-nodes-mm are compared to rand-mm player.
Results of table 2 show that adding minimax search in the perfect information
part of games improves all the players. This enhancement makes chance-nodes-
mm the best player with 0 . 01[sec] and 1[sec] per move. In case of 1[sec] per move,
move-groups-cycle-R-mm and move-groups-cycle-M-mm are closed to chance-
nodes-mm . In all these experiments, rand-mm obtains quasi null scores.

Fig. 2. chance-nodes-mm with X random plies against chance-nodes-mm