Does it make sense to do a cubeless rollout of a cube position? I see that if I ask for one, gnubg automatically ticks the Cubeful Rollout box and does a cubeful rollout.
But if the bot doesn't understand the position, it may erroneously pass future cubes and give a bad result. For example, 9 out of these 36 games ended in Double/ Pass on the 2-cube rollout, and 14/36 on the 4-cube rollout. This is only a test; I know more than 36 games are required to draw any conclusions. This is a position that Frank Berger has suggested, on DailyGammon, that the bot doesn't understand. "XGID=------a----------BBCCBBAn-:1:1:1:00:0:0:0:0:10 XG believes to be a 60-70% favorite for the straggler player depending on ply/rollout but the positions is only about 6%" . I wanted to see what Gnubg thinks. The 2-ply evaluation suggest 40% wins; rollouts suggest 70%. But can I trust this if the cube actions are wrong? I would expect a cubeless rollout to help eliminate this bias so that games are played to a conclusion (leaving only chequer play errors) /z8AAAEAANvdFg:UYkNAAAAAAAE Note that the Copy window is reversing the players' win-loss breakdown in the results below. The GUI has them the correct way. Cube analysis Rollout cubeless equity -0.1968 Cubeful equities: 1. No double -0.6901 2. Double, pass -1.0000 (-0.3099) 3. Double, take -0.5094 (+0.1808) Proper cube action: No redouble, take (36.8%) Rollout details: ianshaw owns 2-cube: 0.2934 0.2533 0.0028 - 0.7066 0.0398 0.0000 CL -0.1967 CF -0.6901 [0.0418 0.0370 0.0019 - 0.0418 0.0144 0.0002 CL 0.1242 CF 0.1798] gnubg owns 4-cube: 0.2856 0.2425 0.0010 - 0.7144 0.0000 0.0001 CL -0.3652 CF -0.5094 [0.0442 0.0414 0.0011 - 0.0442 0.0235 0.0005 CL 0.2649 CF 0.3471] Full cubeful rollout with variance reduction 36 games, rollout as initial position, Mersenne Twister dice generator with seed 2147483647 Play: world class 2-ply cubeful prune [world class] keep the first 0 0-ply moves and up to 8 more moves within equity 0.16 Skip pruning for 1-ply moves. Regards, Ian
