Basic ace deuce backgame cube/ RO
Posted By: Rick Janowski In Response To: Basic ace deuce backgame cube/ RO (Bob Koca)
Date: Tuesday, 16 July 2013, at 11:21 a.m.
In Response To: Basic ace deuce backgame cube/ RO (Bob Koca)
A number of factors make me wary of accepting XGís cube action conclusions in this position without more detailed consideration:
(a) Large disparity between XGR++ evaluations and rollouts for both probability percentages and equities. This indicates that the bot doesnít understand the position (and resulting positions) as well as it might and may be mishandling the cube in rollouts to a significant extent.
(b) The disparity between predicted cubeful equities from the different rollout settings also indicates that even on the higher settings it may still be mishandling the cube to a significant extent.
(c) The high cubeless equities from the rollouts (0.65 and above) are a very strong indication of pass positions rather than take positions. These cubeless equities are derived from the percentages from the double/take (D/T) sides of the rollouts. Ideally, these are best derived from a separate cubeless rollout, but using the D/T figures should provide fairly good estimates.
(d) Cube efficiency values for the taker, back-calculating from the rollout data, are excessively high to the point where they are unrealistic. The rollout results indicate cube efficiency values ranging from 0.94 to 0.97, whereas 0.8 would be excessive. Moreover, a takerís cube efficiency is normally less than the doublerís. The analogy here is the nearer to the target you are the easier it is to hit the bullís eye. This is very strong indication of poor take action on future redoubles, either by taking passes or passing takes or a combination of both.
Curiously, although there is significant disparity between the probabilities and equities predicted from XGR++ evaluations and the various rollouts, they are all very consistent in predicting relative proportions of single games, gammons and backgammons. This suggests that the cash-point window (dead to live) can be predicted with high accuracy:
Dead-cube cash-point = 0.521 +/- 0.015, and Live-cube cash point = 0.592 +/- 0.015.
Considering the more realistic cube-efficiency values of 0.7 and 0.8, the cash-points would be 0.572 and 0.578 respectively.
I donít dispute that technically this might well be a double/take but if so I think the margin is far smaller than the rollouts indicate. Practically though there are good reasons to pass as the takerís side has much tougher containment-related checker play than that of the doubler. Although this might be true for bots too, their errors are likely to be significantly smaller and less numerous than most humans. A bot might win say 43% of games in such positions whereas a human might struggle to win 40% against an equally skilled human, in my view.
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.