| |
BGonline.org Forums
Two Close Plays with 33 *ROLLOUT* inconclusive
Posted By: Daniel Murphy In Response To: Two Close Plays with 33 *ROLLOUT* inconclusive (leobueno)
Date: Monday, 11 April 2011, at 6:54 p.m.
Thank you for upgrading your rollout(s) to Gnubg 2-ply, which is so much more worthy of time and attention than 0-ply. The problem with 0-ply is not that it's bad, but that it's not as good as 2-ply, and although 0-ply will often come up with very similar numbers, it's difficult to know when it has, and when it has not.
Another change you might make is to the "Stop when best play is enough JSDs ahead: limit 4 (min. 36 games)" parameter. I never use this option at all. I just pick a number of trials that seems prudent for the position and play choices, and then extend as necessary, based on the result so far. But if you want to use the option, I'd set the minimum stopping point much higher, perhaps at 108 trials (but no lower), with (as you have it) JSD = 4.
One other thing you might do -- with your evaluations, not your rollouts -- is use 2-ply (or 4-ply if you don't find it too slow) instead of 3-ply. The more rollouts you do, the more you'll see that even-ply evaluations are much more likely to be similar to your rollout results, the major exception (in the normalish positions that are often rolled out) being holding games with an 18-, 20-, 21- or 22-point anchor. In such positions 3-ply tends to be better. An example is your recently posted problem with a 4-1 to play, for which I think even 3-ply probably overestimates Blue's winning chances after anchoring (in a money game, I think Blue's take after 24/20 6/5 is marginal, not comfortable, as 3-ply says, and 2-ply is way off).
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.