Can someone explain to me in English why 10/6* 5/off is better in this position than 10/6/1? Or has GNU backgammon gone loopy?

0-ply is loopy. Use at least 2-ply.

Settings, Analysis: change predefined settings for both cube and checkerplay to either world class or supremo then save settings.

I think your play is fine, but 0-ply evaluations are not to be trusted.

Thanks... tried it on Supremo and the weird move is now very bad!

So I just ran it through Extremegammon and surprisingly it didn't pick up 10/6*/1 on it's first pass. On 3-ply it has it as a 0,155 blunder, but it came to it senses when I forced XGroller+. It's still closer than I would expect between 10/5 4/off and hitting, 10/1 without the hit is a blunder though.

GNU will never get all plays right even with these settings, but it will suck a lot less than on 0-ply.

If you highlight the top 3 plays there should be a button below called "Rollout". It might take a bit of time for it to roll out this position enough times to get a trustworthy result with GNU (and you probably need to change the rollout settings just like you did the analysis settings).

Anyway I did a quick rollout in Extreme:


1. Rollout¹ 10/6* 6/1 eq:+2.077
Player : 88.17% (G:79.45% B:23.35%)
Opponent: 11.83% (G:0.00% B:0.00%)
Confidence: &plusmn; 0.023 (+2.054<E<+2.100)
Duration: 8 minutes 47 seconds

2. Rollout¹ 10/5 4/0ff eq:+2.008 (-0.069)
Player : 88.50% (G:78.71% B:17.76%)
Opponent: 11.50% (G:0.00% B:0.00%)
Confidence: &plusmn; 0.020 (+1.988<E<+2.028)
Duration: 9 minutes 39 seconds

3. Rollout² 10/1 eq:+1.908 (-0.169)
Player : 87.41% (G:79.06% B:13.77%)
Opponent: 12.59% (G:0.00% B:0.00%)
Confidence: &plusmn; 0.031 (+1.877<E<+1.939)
Duration: 5 minutes 25 seconds


¹ 635 Games rolled with Variance Reduction.
Moves and cube decisions: 3 ply

² 299 Games rolled with Variance Reduction.
Moves and cube decisions: 3 ply


eXtreme Gammon Version: 1.21, MET: Rockwell-Kazaross



EDIT: I hadn't even noticed what play 0-ply liked best originally, kinda assumed it was 10/1, but yeah 10/6* 5/off is really bad

Cool, interesting analysis.

I can understand conceptually why it's close between hitting and not hitting, but I couldn't understand why it would suggest leaving 2 blots... anyway, it just looks like it was an anomaly so order is restored to the Universe!

I was playing GNU on the lower levels and had the same weird stuff happen to me. I switched to playing it on the grandmaster level, which is 3-ply, and it's been much better for learning. I get crushed over and over, which is frustrating, but I've also improved a lot and find myself making better plays against humans. Even though it's very difficult to win without a little luck at that level, I've been working on getting my error rate down which has been very beneficial.

Yeah, well I've been playing GNU at the default Expert level and really it doesn't make many mistakes at all, if you go back and analyse the game on Grandmaster or something. Like maybe 1 or rarely 2 "doubtful" plays per game, so in practice it doesn't play too differently between Expert and the higher-ply levels.

I did find when I left it on Expert but introduced a small amout of "noise" to make it easier, it would play really well but then make a complete mess of bearing off, playing obviously stupid moves. I think that's because it uses a separate bearing-off move engine.

It looks like a "feature" of the evaluation function. It values putting a 4th man on the bar. It values getting the first man in. It apparently values each of these more than leaving a blot in the home board (or it would play 10/5 4/off or 10/6/1). Once it gives you another roll after 10/6/1, it realizes that you always get a man in, and it makes the obviously right play. LOL bots on 0-ply.

Heh, you might LOL at 0-ply, but it makes the same play as (or a move within a tiny fraction of a percent of the same equity of) Grandmaster mode 99% of the time (by my estimate). I just played a match on Expert (default, 0-ply) and analysed it on Supremo (2-ply), and it didn't make a single "doubtful" or worse move the whole match (5 or 6 games), except for one doubtful double in the last one. Which is why I'm surprised the analysis came up with such a weird move recommendation on this occasion.

In that case, I think it's trying to increase your chances of getting a backgammon. You're pretty much locked up for the gammon, and getting one more white checker on the bar will increase you BG chances. If you get hit, you'll be right back around in no time, even if you roll 66 and can't enter. Meanwhile, adding one more checker to the bar makes it that much more difficult for white to enter and get out of there. So in this case, gnu knows that you're not going to lose this game, so it wants you to maximize your win, which means you should be shooting for the backgammon.

Yeah, but 10/6/1 is obviously way the best for winning backgammons. The position is actually more interesting at 4-away/2-away, or at DMP, where the backgammon (and gammon at DMP) are dead. All three reasonable plays (10/6/1, 10/5/1, 10/5 4/off) should run pretty close with backgammons out of the picture.

I did a few rollouts, and with backgammons active, 10/6*/1 wins by a mile (duh). When you can use a gammon, but not a backgammon (4-away/4-away, 4-away/2-away, etc.), 10/5/1 comes out on top. At DMP, it looks like 10/5 4-off is coming out on top, although that rollout will take for freaking ever to reach 3 JSD, and all three choices run within .004 MWC of each other, so who knows.