Forum Archive :
||12 April 2010
What settings should I use to have an almost perfect rollout with snowie?
Bill Robertie writes:
The real question is accuracy per unit of time. If time is no problem, just
use Snowie 3-ply cubeful rollouts for 2592 trials and you'll get the best
answer available, although you might wait a couple of days on even today's
A few years ago I tried to solve this problem using the earliest version of
Snowie 4.0. I created a suite of 50 cube problems covering all of the major
position types, and rolled them out using Snowie 3-ply cubeless, 1296
trials. The result of this I accepted as the "correct" answer for each
problem. I then duplicated the rollouts for various combinations of trial
length and ply level, as well as Snowie's raw evaluations on each ply
The results of this experiment showed a sweet spot at 2-ply cubeless, 648
trials, so that's what I used as my rollout settings for a long time. The
exception were various sorts of non-standard or backgame positions, where I
didn't trust anything less than Snowie 3-ply.
Lately, as a result of some positions published in this forum, I've
upgraded my rollouts to 2-ply cubeful, 1296 trials. There was more of a
discrepancy between the cubeful and cubeless rollouts than I originally
expected, and although these new settings take longer, especially in early
game positions, I think the time is worth it.
Peter Hallberg writes:
Here are some crude guidelines:
* In a attacking position it's ok to use truncated rollouts.
* In priming games use full rollout.
* As a really good estimate of whether the equity is going to change on
rollout (compared to 3-ply eval) do a 1-ply full 1296 games. If the
equity is pretty much the same as 3-ply I usually don't want to put much
more effort into it.
* You often need to make more precise rollouts for cubes because the right
plays is determined by the equity. Selecting a move is less sensitive to
error margin because you just want to know which move is best rather than
how good it actually is.
As a note: With Gnu and XG you have excellent ways to stop a rollout based
on confidence interval and they are way faster than Snowie.
- Advice (David Montgomery, Apr 1996)
- Cautionary tale (Kit Woolsey, Sept 1995)
- Combining rollouts (Gregg Cattanach+, Dec 2003)
- Confidence intervals (Bob Koca, Nov 2010)
- Confidence intervals (Timothy Chow, May 2010)
- Confidence intervals (Gerry Tesauro, Feb 1994)
- Cubeless vs centered-cube rollouts (Ron Karr, Dec 1997)
- Duplicate dice (David Montgomery, June 1998)
- How reliable are rollouts? (David Montgomery, Aug 1999)
- Level-5 versus level-6 rollouts (Michael J. Zehr, June 1998)
- Level-5 versus level-6 rollouts (Chuck Bower, Aug 1997)
- Positions with inaccurate rollouts (Douglas Zare, Oct 2002)
- Reporting results of rollouts (David Montgomery, June 1995)
- Rollout settings (Lokicol+, Apr 2010)
- Settlement limit (Michael J. Zehr, Apr 1998)
- Settlement limit (Kit Woolsey, Dec 1997)
- Settlement limit in races (Alexander Nitschke, Dec 1997)
- Some guidelines (Kit Woolsey, Apr 1996)
- Standard error and JSD (rambiz+, Feb 2011)
- Standard error and JSD (Stick+, Oct 2007)
- Systematic error (Chuck Bower, Oct 1996)
- Tips for doing rollouts (Douglas Zare, June 2002)
- Truncated rollouts (Gregg Cattanach, Oct 2002)
- Truncated rollouts: pros and cons (Jason Lee+, Jan 2006)
- What is a rollout? (Gregg Cattanach, Dec 1999)