[img]http://s6.************/d9ha2g701/Screenshot_from_2013_07_21_18_21_31.png[/img]
green shot screen capture

The last possibility can be ruled out, as this also holds for the expert vs expert

As far as me concerns the only proof that is delivered is that linux gnub has some flaws. Great! For one time let's move from the laboratory to the real world of backgammon, and get a feel of the seventies.

There is very little difference between expert and advanced. Both are 0-ply rollouts. See my explanation below.

Here is what I recommend you try:

Player 0 is the top player. Set it to use expert or lower.

The only difference between expert and the settings below it is noise. They all are 0-ply rollouts, with no look-ahead. As skill level is decreased, increasing amounts of random noise are added to (and subtracted from) the equities returned by the neural net. At settings below expert, therefore, it is possible that the play selected will not be the one preferred by the neural net, even at 0-ply.

Player 1 is the bottom player. Set it for world class. That is the 2-ply look-ahead that most of us use in our work with GnuBg. Make sure Player 1 is the player given the opening 64 to play. His lesser-skilled opponent will make the reply.

Rollout 5184 trials at a minimum.

Sometimes you can get a statistically significant result with fewer trials, but the result only tells you which play is best. In order for the margin between plays to be accurate, you need more trials. For example, suppose the top play beats the second by by 0.03 with 95% confidence. The 95% applies only to the estimate that the top play is best. The margin 0.03 does not share the same 95% confidence. To reduce variance in the margin, you need many more trials.

In my work, I often use the scale adopted by Nack Ballard and David Rockwell. Its rigor is required only for early game rollouts, typically during the first three rolls.

For plays between 0 and 0.01 from best I use 62k = 62208 trials.
Between 0.01 and 0.02, 32k = 31104
Between 0.02 and 0.03, 15k = 15552
Between 0.03 and 0.04, 7k = 7776
Above 0.04, 5k = 5184

Hope this helps.

Mike

White - Pips 158

Black - Pips 167
Black to Play 6-4

We are having so much fun slotting the bar point in this thread! I offered up this position as a counterexample to see whether anyone would fall into the trap. You are quite right that without the distraction of a hit, it is quite a large mistake to slot the bar point. With 54-split, 63, the hit gives you the tempo you need to protect the slotted bar point.

Perhaps it is surprising, however, that rollouts favor the pointing play, 8/2 6/2, by a significant amount. Running is second best.

Mike

I don't take it personal. In this case my choice definitely would be 24/18 13/9. I absolutely detest the opponent making his 5 point.

For my pc this is going to take days. I hope it gets not overheated, as there is a heatwave going on here at the moment.
Why don't you use ratio standard deviation/value?

I wonder why the noise of player(0) will drop, if i set player(1) to worldclass.

Forget that last remark.

The primary reason is because I don't know what it is!

Mike

This is my 2 cents.
wikipedia:
¨standard deviation is commonly used to measure confidence in statistical conclusions. For example, the margin of error in polling data is determined by calculating the expected standard deviation in the results if the same poll were to be conducted multiple times. The reported margin of error is typically about twice the standard deviation *– the radius of a 95 percent confidence interval.¨

Now suppose we allow a standard deviation of .1 for an outcome of about .2. Then there is 95% certainty that the value is between .1 and .3. If for an outcome of .9, the value is between .8 and 1.0. So we are interested in standard deviation/value. This ratio is set default .1. So for an estimated equity of 1.00, the standard deviation is ratio*equity=.1. For an estimated equity of .2, the standard deviation is ratio*equity= .02.

[img]http://s6.************/bxr5b0hv5/Screenshot_from_2013_07_22_19_41_49.png[/img]
screen capture windows 7

Dealing myself with a heatstroke and fatigue, and betraying my lack of tournament experience, i would like post-crawford to be replaced by crawford.

FYI: It is pretty common for GnuBg and Snowie rollouts to run overnight. If you want, you can stop, and later restart, a GnuBg rollout without affecting the result.

Sounds you got a result that satisfies you.

How many trials did you end up with?

Mike

more than 1500. Smells like you or accomplice did your own rollouts.
I stopped the rollouts many times along the way, and saw that the outcome was stable.

Did you do handrollouts? And did you like the play?

Here is an expert advice on playing two down from the mid point with opening 6x:
DON'T DO THAT!

To me you are but a stranger. My pc is busy at the moment doing very extensive rollouts, and results and conclusions will be presented within a week.

Yo! Yogiman,

Dmitriy Obukhov is the real thing.

Since 2005, Iancho Hristov has been keeping a list backgammon players worldwide who have the lowest average PR (Performance Rating as determined in XG) over a span of 50 matches. Not all the world’s best players are represented, because Iancho does not always have access to their match files. Many of best players in the world, however, will be found on his list. Here is the link: Iancho’s PR Ranking List. Check out the name at slot 34.

Opening 64 is something Dmitriy has studied extensively. With backgammon giant Paul Weaver, he is the author of an article on the subject that will be published in Primetime Backgammon this week. It will analyze the best play for opening 64 at every possible score in a 7-point match. If anyone is an expert on the subject, he is the guy.

Glad to hear you are continuing the rollouts. As I said before, I have not attempted anything (or seen anything except the XGR++ evals I cited above). I bet you are having fun and learning a ton. I look forward to seeing your results.

Hope it has cooled down enough in your part of Europe to run the CPU all day and night long!

Mike

That was a hellawa intro! My post was a joke. I am not trying to convince Yogiman to stop playing two down. Seems to me he has made his choice, and no one can convince him to do otherwise.

The reliability of this list is based on the precision of the XG. A lot of very tiny equity mistakes of the bot can add up to something that could change the ordering of the list. And again, some ¨minor¨ moves can work out better for the human brain. To me it seems there must be an element of chance to it.

The tiny random mistakes will add up to pretty much 0, unless you're suggesting that you discovered systemic mistakes that bots make during evaluation.

in any case, whether he's #34 on a list or #43 he's still an expert

I used the word can, and mentioned the ordering of the list.

The bot has created a huge set of rules upon which to act and upon which to evaluate, and I have not the ability to judge about the randomness of mistakes, so systematic mistakes I can not rule out, with due respect to the developers. However, you do make a good point.