| |
BGonline.org Forums
Ultimate XG Settings – help please
Posted By: Timothy Chow In Response To: Ultimate XG Settings – help please (Taper_Mike)
Date: Friday, 6 February 2015, at 11:32 p.m.
I agree with almost everything Taper_Mike wrote. Just one comment about this paragraph:
By convention, the number of trials is usually set to be a mulitple of 1296. Popular choices are 5k = 5184, 7k = 7776, 10k = 10368, 15k = 15552, 20k = 20736, 31k = 31104, 46k = 46656, and 62k = 62208. Anything less than 5k can be problematic. The top play may be identified with high confidence, but the exact margins by which it beats other plays will still have a lot of variance.
It's been years since I paid any attention to the confidence percentages. I look only at the confidence interval. A small confidence interval means that it's unlikely that the equity will change by more than that if you roll it out further. The only thing I'd quibble with is Mike's comment that "anything less than 5k can be problematic." There are plenty of positions for which even 324 trials will produce a confidence interval of ±0.001.
Having said that, I will say that it is a good habit, whenever statistical computations are involved, to explicitly say ahead of time what you're planning to do. For example, if you perform a rollout and then stop it as soon as it declares "99% confidence" then this won't actually give you 99% confidence because the 99% figure is calculated assuming that you're not going to watch the rollout as it progresses and stop it when that threshold is reached. By standardizing the number of trials to being one of a small set of predetermined values, you avoid this statistical trap. For this reason, I would recommend that you pick some rollout setting to be your "standard" setting. The setting should be long and strong enough that you will rarely feel the need to do anything longer and/or stronger, but fast enough that you don't get impatient. 5184 trials at 4-ply/4-ply would seem to be a reasonable default given your system, but it doesn't matter too much as long as you stick to it.
The other thing that you might consider is to pick a fixed random seed, different from what anyone else might pick but that you keep constant for all your rollouts. This is useful because then if someone else performs a rollout of the same position, they can combine your results with theirs to get a more reliable estimate. If you use the default seed then your rollout might just replicate someone else's computation without adding any new information.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.