| |
BGonline.org Forums
15552-game rollout of 41Z-65R-43
Posted By: Daniel Murphy In Response To: 15552-game rollout of 41Z-65R-43 (Nack Ballard)
Date: Tuesday, 29 December 2009, at 12:58 a.m.
Nack writes: Your 15k rollout shows the margin to be .002. It's worth pointing out, I think, that the 0.002xx margin comes with 95% confidence intervals of 0.00588 (for 24/20 23/20) and 0.00392 (for 9/5 8/5) -- still quite large, relative to the difference. In Tim's earlier, shorter rollout where the margin was 0.011, the 95% confidence intervals for both plays were also large, of course -- about 0.020 (rounded).
Tim does rightly note that the longer rollout is still "TCTC," but it cannot be overemphasized how much attention must be paid to the reported uncertainties, when trying to figure out what a rollout has told us.
Tim writes: I've argued—at least on rec.games.backgammon, if not here—that truncated rollouts might sometimes be preferable to full rollouts, especially given a fixed amount of computation time.
Well, I would agree with that, if the word "especially" were left out -- I can't think of any reason other than time limitations that truncation should be preferred to full. Truncated rollouts will have lower uncertainties, but ... so what? If you valued low uncertainties with few trials, you could do your rollouts trunctated to depth two. There's no escaping the truncated result's dependency on the evaluations the bot makes at the end of the truncation period. More precise does not mean more accurate.
assuming your goal is to get at the truth of the position as opposed to what the bot thinks. -- I'm not sure what to make of that; it does not sound like a meaningful opposition of terms. You may do a rollout trying to "get at the truth of the position," but the rollout result can't be anything but what the bot "thinks" with the particular parameters that governed the bot's play of both sides of the position.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.