| |
BGonline.org Forums
Rollout Settings (for Alex/Steve)
Posted By: Stick
Date: Sunday, 24 August 2008, at 10:01 p.m.
Since I didn't have the time before to properly respond to Alex's post or Steve's post I thought I'd create an entirely new thread to address them.
When deciding whether to use Snowie or GNU for a particular rollout, I'd say easily over 95% of the time use GNU. The main reason is it's significantly faster. The bots are arguably the same strength for money, with the nod going to GNU for match play hands down. If they're the same playing strength there's no reason to use Snowie. Also, equally important to me is that GNU rolls out all candidate plays simultaneously whereas Snowie rolls out 1 play, then the other play, then if there's a 3rd play. This means that with GNU you get to see the results as it happens so if there's a clear difference after X number of trials, you can abort the rollout and move on to the next. With Snowie this isn't possible, you have to make your wild guess from the beginning how many trials each play will need and go from there.
GNU's bearoff database goes out to either the 12 point or midpoint. GNU also, to my knowledge, toss around the erratic Snowie bearoff cubes.
GNU's biggest exploitable weakness is massive backgames and its cube handling in these backgames. This shouldn't matter much to you as it never comes up in real play. You also won't be rolling out backgames anyway, the time it takes to roll out a backgame is astronomical in most cases and the results you get are untrustworthy with either bot. I'd take Snowie over GNU in a super backgame, it's such a minor point it's almost not worth mentioning.
I always use the highest settings with Snowie, 3 ply checker play, 3 ply cube, I put the cube cap at 4 to cut down on any larger than average cubes affecting the results, also make sure the seed is set to 0 (random seed). The absolute minimum number of trials I'll do with either bot is 108. I never truncate until reaching the bearoff database. If you do a live cube rollout with Snowie you'll need at least twice as many trials to get the same statistical significance you would if you were doing a cubeless rollout. Depending on the position you can decide for yourself which you need.
My person default settings for GNU can be found here:
http://www.bgonline.org/gnu.html
Of course you'll want to change the number of trials to something reasonable, but other than that everything should be good. Any questions let me know.
If your rollout results for GNU & Snowie disagree, which is semi rare but quite plausible, you'll want to first make sure you did both rollouts on high settings. If you did you may want to post the positions/rollouts on the forums or ask someone like myself about them and see if we notice anything you may have overlooked. Lastly, could just be a position they disagree on, it can happen.
0 ply rollouts, worth the time or a waste of time? I agree with everything Chuck said noting that GNU is fairly quick even on 2 ply so why not just do the 2 ply rollout and not have to second guess the results because you gimped it with 0 ply to begin with? If you have a newish computer this shouldn't be a problem, even Chuck's rusty old 1.6 GHz still does ok w/GNU. If you have anything over 2.0 (or multiple cores) you should be happy.
0 ply will get you decent results quicker, so if you're really pressed, it's certainly better than nothing. If you want to make sure you're understanding a position and learning the right things though, I'd put in the 2 ply time.
Stick
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.