I was astonished to see GNU Backgammon assess its own cube decision as
"very bad" (> 0.16 equity loss). I am using different settings for
evaluation (during the game) and analysis (after the game), but cannot
remember a discrepancy this large. There is also a pronounced odd/even
effect with respect to the plies.

The score (after 5 games) is: gnubg 8, axel 6 (match to 11 points)
Move number 52: gnubg on roll, cube decision?

GNU Backgammon Position ID: 0DmAnUFtM8YhAA
Match ID : MAFgAYAAMAAE
+-1--2--3--4--5--6-------7--8--9-10-11-12-+ O: gnubg
| O O O O X O | | X X O O | 8 points
| O O O X O | | X X O O | On roll
| | | X O |
| | | |
| | | |
| |BAR| |^ 11 point match (Cube: 1)
| | | |
| | | |
| X | | X |
| X | | X |
| X X | X | X O | 6 points
+24-23-22-21-20-19------18-17-16-15-14-13-+ X: axel
Pip counts: O 101, X 200

I took ("When no clue, take, you might learn something"). What are your
thoughts (about the position and suitable rollout settings)?

Best regards

Axel

On Sunday, July 3, 2022 at 9:03:59 PM UTC+1, Axel Reichert wrote:
> I was astonished to see GNU Backgammon assess its own cube decision as
> "very bad" (> 0.16 equity loss). I am using different settings for
> evaluation (during the game) and analysis (after the game), but cannot
> remember a discrepancy this large. There is also a pronounced odd/even
> effect with respect to the plies.
>
> The score (after 5 games) is: gnubg 8, axel 6 (match to 11 points)
> Move number 52: gnubg on roll, cube decision?
>
> GNU Backgammon Position ID: 0DmAnUFtM8YhAA
> Match ID : MAFgAYAAMAAE
> +-1--2--3--4--5--6-------7--8--9-10-11-12-+ O: gnubg
> | O O O O X O | | X X O O | 8 points
> | O O O X O | | X X O O | On roll
> | | | X O |
> | | | |
> | | | |
> | |BAR| |^ 11 point match (Cube: 1)
> | | | |
> | | | |
> | X | | X |
> | X | | X |
> | X X | X | X O | 6 points
> +24-23-22-21-20-19------18-17-16-15-14-13-+ X: axel
> Pip counts: O 101, X 200
>
> I took ("When no clue, take, you might learn something"). What are your
> thoughts (about the position and suitable rollout settings)?
>
> Best regards
>
> Axel

Looks like a clear ND/T to me. However, I note several points.
1) We are in the middle of Wimbledon.
2) A new edition of the New Yorker just came online today.
3) There are some great articles in the London Review of Books that I haven't read yet.
4) Same as 3) for the New York Review of Books.

How are points 1 to 4 relevant?

They're relevant because they indicate that, with so much else to watch and read, I'm not
exactly super-motivated to spend time defending my verdict.
Obviously with a match lead, gnubg should be conservative with the cube.

Paul

On July 3, 2022 at 2:03:59 PM UTC-6, Axel Reichert wrote:

> I was astonished to see GNU Backgammon assess its
> own cube decision as "very bad" (> 0.16 equity loss)....

> GNU Backgammon Position ID: 0DmAnUFtM8YhAA
> Match ID : MAFgAYAAMAAE

> I took ..... What are your thoughts (about the position and
> suitable rollout settings)?

It's hard to understand which is which but apparently Gnubg
doubled (and you took) when it shouldn't double. If that's the
case, 4-ply hint says ND (I rarely do rollouts) and says D/T is
-0.281 which is almost twice the > -0.16 treshold.

Since I don't understand (and don't want to underdtand bogus)
detailed cube analysis numbers, I don't pay much attention to
how Gnubg may be making "PR sacrificing" cube decisions but
in the past I had caught and posted many examples of Gnubg's
making "PR sacrificing" checker decisions during the game by
looking ahead to upcoming dice rolls.

How did your game end? Can you copy/paste it here as text?

From now on, I may try to catch such cube "errors" even with
my very limited understanding of the "cube skill bullshit"...

MK

On 7/3/2022 4:03 PM, Axel Reichert wrote:
> I was astonished to see GNU Backgammon assess its own cube decision as
> "very bad" (> 0.16 equity loss). I am using different settings for
> evaluation (during the game) and analysis (after the game), but cannot
> remember a discrepancy this large. There is also a pronounced odd/even
> effect with respect to the plies.

This is the sort of position that GNU and XG still have some trouble
with. The match score also complicates matters. At 3-away/5-away,
when most of the 3-away player's wins are gammons, but the 5-away
player still has considerable winning chances, doubling is typically
a mistake; for a start, it decreases the 3-away player's gammon value
(possibly to zero, if the 5-away player redoubles). The take is easy
and will probably be easy next turn as well.

I suspect that the "contradictory" behavior you observed will be much
reduced at a normal match score such as 7-away/7-away. Whatever
heuristics GNU is using to adjust for the match score are probably
not working too well in this position.

Below is an XG rollout with stronger-than-usual settings. The
standard rollout settings were yielding high-variance results,
which is a sign that the bot doesn't understand the position
very well. Even these settings were yielding pretty high
variance, but at least the final result is in line with what
I would expect from general principles (i.e., ND/T).

XGID=aABBBbBcbB--C---Ac-ca-----:0:0:1:00:8:6:0:11:10

X:Player 2 O:Player 1
Score is X:8 O:6 11 pt.(s) match.
+13-14-15-16-17-18------19-20-21-22-23-24-+
| X O | | O O |
| O | | O |
| O | | O |
| | | |
| | | |
| |BAR| |
| | O | |
| | | |
| X O | | |
| X X O O | | X O X X X |
| X X O O | | X O X X X X |
+12-11-10--9--8--7-------6--5--4--3--2--1-+
Pip count X: 101 O: 200 X-O: 8-6/11
Cube: 1
X on roll, cube action

Analyzed in Rollout
No double
Player Winning Chances: 62.75% (G:47.44% B:0.84%)
Opponent Winning Chances: 37.25% (G:5.60% B:0.26%)
Double/Take
Player Winning Chances: 61.44% (G:48.89% B:1.13%)
Opponent Winning Chances: 38.56% (G:5.92% B:0.32%)

Cubeful Equities:
No double: +0.609
Double/Take: +0.411 (-0.198)
Double/Pass: +1.000 (+0.391)

Best Cube action: No double / Take
Percentage of wrong pass needed to make the double decision right: 25.2%

Rollout:
2014 Games rolled with Variance Reduction.
Dice Seed: 271828
Moves: 4-ply, cube decisions: XG Roller+
Search interval: Large
Confidence No Double: ± 0.018 (+0.591..+0.627)
Confidence Double: ± 0.031 (+0.380..+0.442)

eXtreme Gammon Version: 2.19.207.pre-release, MET: Kazaross XG2

---
Tim Chow

On 7/4/2022 5:53 PM, MK wrote:
> in the past I had caught and posted many examples of Gnubg's
> making "PR sacrificing" checker decisions during the game by
> looking ahead to upcoming dice rolls.

You still believe this nonsense? Even by your standards, this
is nuts. How do you explain the fact that even when you change
the dice seed and play from this position, GNU makes exactly
the same decisions every time?

---
Tim Chow

The take is easy
and will probably be easy next turn as well.

> Tim Chow

Doubtful.

Stick

On July 7, 2022 at 7:17:27 PM UTC-6, Tim Chow wrote:

> On 7/4/2022 5:53 PM, MK wrote:
>> in the past I had caught and posted many examples of
>> Gnubg's making "PR sacrificing" checker decisions
>> during the game by looking ahead to upcoming dice rolls.

> You still believe this nonsense?

Don't you yet?

> Even by your standards, this is nuts. How do you explain
> the fact that even when you change the dice seed and
> play from this position, GNU makes exactly the same
> decisions every time?

This can become a multi-faceted interesting discussion.

Let's start.

First, I assume you mean if you change the dice seed not
from this position but from previous position, i.e. before
the bot makes its decision. Otherwise, it's already done
and it's too late.

Second, I assume you understand that when you say it
will make the same decision, you mean "the same wrong,
inferior decision", or else you would be missing the point
by a mile...

My comment was meant to be a semi-serious, semi-snide
explanation, expecting that nobody else would ridicule
themselves offering explanations.

Do you realize and accept that your blabbering bullshit
explantion is no better than mine that you sneer at...?

Also, I was referring to past examples of checker plays
which may have changed since then, while here we are
talking a cube decision. Very different things...

Starting with Jellyshit, I have argued for and provided
statistical evidence for that games played with manual
dice against the bots "enfold", "develop" in such easily
observable different ways, that if anyone denies to or is
incapable to see it, I won't take them seriously enough to
debate with them.

In the past, I was the one begging to be given a chance to
prove my points. Now I don't give a shit. Except, of course,
if you guys make it easy for me to demonstrate to you all,
in a way open to the public and if you all are willing to bet
money on it against me.

I have given you step-by-step instructions to prove it to
yourselves in the past. I don't think that any of you ever
made a genuine effort. I won't waste my time again with
repeating them but I can perhaps add this one new way.

Recently, I posted here an external dice dll for XG reading
numbers from text files, including the source code and
instructions to compile it as a dll or as an exe.

Gnubg already has the feature of reading dice from files.
So, let it read from a text file and use my dll to make XG
get rolls from the same file. Play money games against
both bots thus with the same dice rolls.

Per you guys' own argument, (and mine:)) both bots play
almost exactly the same way, except for very well known
differences like their playing the opening 64's and a few
others. You will see that depending on who rolls the dice,
the bots will diverge on best moves. But, why am I even
wasting my time with you half brained "believer" morons
when I should know better that you neither won't or can't
consider what I say even as mere hypothetical possibilities...

MK

On 7/19/2022 7:25 AM, MK wrote:
> On July 7, 2022 at 7:17:27 PM UTC-6, Tim Chow wrote:
>
>> On 7/4/2022 5:53 PM, MK wrote:
>
>>> in the past I had caught and posted many examples of
>>> Gnubg's making "PR sacrificing" checker decisions
>>> during the game by looking ahead to upcoming dice rolls.
>
>> You still believe this nonsense?
[...]
> My comment was meant to be a semi-serious, semi-snide
> explanation, expecting that nobody else would ridicule
> themselves offering explanations.

Ah, so your answer is no. Some vestige of sanity still remains
in that brain of yours, fighting to get out.

---
Tim Chow

On July 19, 2022 at 7:30:23 AM UTC-6, Tim Chow wrote:

> On 7/19/2022 7:25 AM, MK wrote:

>> My comment was meant to be a semi-serious, semi-
>> snide explanation, expecting that nobody else would
>> ridicule themselves offering explanations.

> Ah, so your answer is no. Some vestige of sanity still
> remains in that brain of yours, fighting to get out.

As typical, you pick one sentence from my post and
repond to it with onle sentence also. Perhaps your
brain (however much you have) filters out the rest
in self-defense. I hope that others may have gotten
a little more from it than you.

Why don't we go back to your reply to Axel and do a
retake. You started your bullshit by saying:

> This is the sort of position that GNU and XG still
> have some trouble with.

Since you are generalizing from it, can you give some
other examples of "this sort of position"..?

And you finished with a rollout by saying:

> Even these settings were yielding pretty high variance,
> but at least the final result is in line with what I would
> expect from general principles (i.e., ND/T).
> .....
> Moves: 4-ply, cube decisions: XG Roller+

If you had kept going up the ladder, you would have seen
that unfortuately for you, the D/T comes back at 6-ply.

So, let's take a more close look at what bots say about it.

In Gnubg, which Axel was playing against, we see:

Beginner = ND/T
Casual Play = ND/T
Intermediate = ND/T
Advanced = ND/T
Expert = D/T **
World Class = ND/T
Supremo = ND/T
Grand Master = D/T **
4-ply = ND/T

He didn't say which level he was playing against but it
is very odd that contradicting decisions are interleaved.

In XG, which you prefer for some reason, we see:

1-ply = D/T
2-ply = D/P !!
3-ply red = D/T
3-ply = D/T
XG Roller = D/T
4-ply = D/T
5-ply = D/T
XG Roller+ = ND/T **
6-ply = D/T
7-ply = ND/T **
XG Roller++ = ND/T **

Not only contradicting decisions are also interleaved but
there is also the bonus oddity of D/P at 2-ply.

I asked for it but Axel didn't post the game record here in
text format which would have taken only a few seconds
of copy/paste. I wonder if he was afraid that I could find
something more to point at in there..?

You guys' denial of the obvious will only delay it for you
(unfortunately) but eventually you all will what pieces of
shit are the gamblegammon bots that you worship...

MK