| |
BGonline.org Forums
Oops -- fixed that
Posted By: Daniel Murphy In Response To: How comes Blue wins more games after redoubling?! (Daniel Murphy)
Date: Saturday, 11 June 2011, at 4:51 p.m.
Corrected!
Here, however, I think there may be a meaningful explanation for why the rollout erroneously report more wins after "double/take" than after "no double." Remember (as noted by Tim in the linked-to April '11 post) that in a cubeful rollout, whenever a cube decision is reached for which the bot (re)doubles and passes, that trial stops, and the bot estimates the probability of that trial resulting in a win, gammon or backgammon if the trial had continued. The estimate will be wrong if the bot misevaluates the cube position. The final rollout result will be correspondingly affected.
Noticing that this particular Gnubg rollout was done with 3-ply cube, I suspect that a 3-ply evaluation of the original position overestimates Blue's win percentage, in comparison to what the rollout found. One might suspect that if Blue does not double now, many favorable sequences will lead to a double/pass later, at which point the bot will evaluate the position and end the trial. One might also suspect that if the bot overestimates Blue's equity for cube action in the original position, it's likely that the bot will all also overestimate Blue's equity in subsequent decisions after "no double" in the original position.
And that is, I think, the most important factor explaining why the rollout wrongly reports that Blue's "double/take" win percentage is higher than Blue's win percentage after "no double."
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.