I was just watching a match from the 2005 WC, Rumker vs. Carlston, and Magriel, who was commenting, said at a certain moment that you shouldn't split after your opponent roled a double 5 as his opening roll. He also said it was a very advanced play very few players knew about. Does anyone here know more about this specific situation? I mean, why is it better not to split?

I don't understand the question. You can't get double five as your opening roll, so I'm assuming something else has to have happened first, and it's pretty hard to answer when you don't know what that is.

One thing that comes to mind is opponent making the 2 or 3 point with 53 or 64. In that case it's usually (always?) correct to bring two builder down with rolls like 32 and 43.

Magriel is right. After an early 55, say as the reply to a non-splitting opening roll, your opponent is poorly placed to play a priming or blocking game. His best game plan is to attack after you split, hoping to play a blitzing game (for which he's well-positioned) or a mutual holding game where both sides get a high anchor. You counter his desire by building instead of splitting.

This is an application of the same principle. When your opponent makes low points in his board to start the game, the value of splitting drops and the value of building and priming rises. In the early game, you should be constantly thinking this way to evaluate all the murky 'split or build?' decisions that come up.

I think you're referring to the game here. The sequence of rolls is 53 (8/3 6/3), 55 (13/3(2)), 63. Carlston splits with 24/18 13/10, which Magriel says is wrong, but doesn't say what is better--24/15 maybe? Honestly I would've played 24/18 13/10 too, so I don't think I can be much help. My best guess is something like this: because your opponent now has 10 checkers in the attacking zone and a point made, he is more inclined to attack given the chance. Leaving a checker on the bar point is more likely to lead to a blitz, because now White's 6s and 1s will both hit and put a very threatening attacker on the 7-point. If red is in the outfield after 24/15, and white hits with a 3, the hitter is at least somewhat further from the action. This is all guesswork though since the split looked normal to me.

Of course I could be completely off the mark and the right play is 13/4 or something crazy like 13/10 13/7 (this is nuts, I think). I don't have any bots installed right now so I can't see what they think, maybe someone else can check it out?

Edit: in after Robertie it seems. Seriously 13/10 13/7 is right here? That still looks nuts.

according to the snowie rollouts i did:

1. 13/7 13/10 -0.138
2. 13/4 -0.161
3. 24/15 -0.165
4. 24/18 13/10 -0.180

Cool, thanks djk. Wow, I gotta recalibrate myself I guess...those plays are not close.

Does this count as well when your opponent makes a 3 point home board with 33 or 44 then?

No, because in those positions your opponent has used up all his builders and has little left to attack you.

i decided to extend the rollouts ( i only did 324 games at first) and the results are changing if im reading them correctly. how many games does one need to rollout to get accurate results? im using the full cubeful setting on snowie4

Depends on the ply level you're using. 2-ply and 3-ply have variance reduction, which is very helpful. I recommend using 2-ply, 648 trials for the best balance of time and accuracy.

thank you. i have one more thing i wanted to clear up. i just did 3-ply 324 game rollouts for the 4 moves i mentioned earlier, and i'm a little confused.
i get 2 equity #s for each roll, one on the same line as the move, and the other after the words 'live cube rollout:' which numbers are accurate? i used the full cubeful setting if that matters.

i would think the ones next to live cube rollout are right, but they have 24/15 and 24/18 13/10 both being better than 13/7 13/10 and 13/4, which you've already said is wrong. the other numbers have 13/4 and 13/7 13/10 being the top 2.

edit: here's a screenshot

"Cubeful" is what indicates the live cube. I never use cubeful rollouts for checker plays. I used to hear, that besides taking a lot longer, that I couldn't really trust them anyway. Maybe the software has developed since, though.

I guess since Snowie still puts the plays in the order of the cubeless results, when you're doing a cubeful rollout means, that the programmers don't really trust the cubeful results either?

Cubeless rollouts are able to use Snowie's variance reduction routine, which (I believe) makes a given number of cubeless rollouts as accurate as a rollout 10-15 times longer without variance reduction. (Don't hold me to these numbers, as I'm remembering them from a long time ago, when the Snowie people first implemented the algorithm.) So 324 cubeless rollouts with variance reduction ~= 3000-4000 raw cubeless rollouts.

The cubeful rollouts don't use variance reduction, and they take a little longer as well, so 324 cubeful rollouts are just a very small sample.

Most people rely on the cubeless rollouts for higher ply levels just because of the time factor. But if you do several thousand cubeful rollouts of this position and those results held up, that would be a theoretically startling result. Magriel and I would certainly take notice, I assure you.

Thanks for the explanation. I assume you havn't tried several thousand cubeful rollouts yourself then? I'm curious so I'll have the results in like 2 days i guess

Correct. I use cubeless rollouts myself, usually 648 2-play for most normal positions. In addition to the variance reduction, there is supposedly some sort of correction algorithm to handle the effect of cube position, although I don't know that anyone has every tried to test this.

well, my full cubeful, 3ply on both checker moves and cube decisions, 3888 games rollouts for the same 4 moves finished overnight, but apparently snowie just decided to delete it i can say though that 13/7 13/10 was definitely the best play with -.127 equity. i think 13/4 was next with ~-.160, and then 24/15 was a little worse with something like ~-.167. i didn't get to see the results for 24/18 13/10 at all, but i dont think it really matters.

Thanks much for doing the work. That's a good data point to have -- the cubeful rollouts backed up the cubeless rollouts given a sufficient number of trials. I'd have been surprised if it were otherwise, but it's nice to have some hard confirmation.