Surprised by XG

[peps...@gmail.com]

I'm very surprised by XG's play here -- 18/15 7/6 because it crunches on 33.
Why not 18/14 which plays well on all next rolls?

Thank You.

Paul

XGID=-BBB-CCa-----B----bbbbbbbA:1:-1:-1:11:2:8:0:11:10
X:XG Roller+ O:Daniel

Score is X:8 O:2 11 pt.(s) match.
+13-14-15-16-17-18------19-20-21-22-23-24-+
| X | | O O O O O |
| | | O O O O O |
| | | O O |
| | | |
| | | |
| |BAR| |
| | O | |
| | | |
| | | | +---+
| O X | | X X X X X X | | 2 |
| O X | | X X X X X X | +---+
+12-11-10--9--8--7-------6--5--4--3--2--1-+
Pip count X: 74 O: 96 X-O: 8-2/11
Cube: 2, X own cube
X to play 11

1. XG Roller+ 18/15 7/6 eq:+1.076
Player: 96.68% (G:22.96% B:0.06%)
Opponent: 3.32% (G:0.01% B:0.00%)

2. XG Roller+ 7/5(2) eq:+1.076
Player: 96.85% (G:22.62% B:0.06%)
Opponent: 3.15% (G:0.01% B:0.00%)

3. XG Roller+ 18/14 eq:+1.076
Player: 96.97% (G:22.19% B:0.06%)
Opponent: 3.03% (G:0.00% B:0.00%)

4. XG Roller+ 7/6 7/4 eq:+1.075 (-0.001)
Player: 96.72% (G:22.79% B:0.06%)
Opponent: 3.28% (G:0.01% B:0.00%)

5. XG Roller+ 18/16 7/6(2) eq:+1.067 (-0.009)
Player: 96.24% (G:23.03% B:0.06%)
Opponent: 3.76% (G:0.01% B:0.00%)


eXtreme Gammon Version: 2.10, MET: Kazaross XG2

[peps...@gmail.com]

On Sunday, December 10, 2023 at 12:44:11 PM UTC, peps...@gmail.com wrote:
> I'm very surprised by XG's play here -- 18/15 7/6 because it crunches on 33.
> Why not 18/14 which plays well on all next rolls?

Actually, the stats kind of answer my own question.
My own play of 18/14 does actually win more games, and probably for the reason I said.
But the extra games is only 0.3%. Does this look right? So the prob of 33 followed by me
entering is 1%. So XG is right if I win around a third of the games when that happens.
This looks totally reasonable.
I can see how XG's play is more gammonish because it plays to 6 to avoid an inefficient bear-in.
And apparently the extra gammons compensate. Instructive example of the difference between
bots and strong players. I think Stick would play 18/14, as would any good human player.

Paul

[MK]

On December 10, 2023 at 6:09:53 AM UTC-7, peps...@gmail.com wrote:

>> I'm very surprised by XG's play here --
>> 18/15 7/6 because it crunches on 33.

I agree that 18/15 7/6 is not the best play.

>> Why not 18/14 which plays well on all
>> next rolls?

> I think Stick would play 18/14, as would
> any good human player.

Not being one of them/you sick world-class
or even merely good gamblegammon players,
I can't speak for them but I wouldn't play that.

I wouldn't argue that 18/14 is not better than
18/15 7/6 but I don't see why the latter is the
better play either.

Stacking on high points is better for bearing
off and there is no reason to hurry the back
checker. Stacking on 5-point here is actually
better than on 6-point. So, this is the perfect
roll for 7/5(2) especially because 66 won't
hurt as it will be unplayable and every other
roll will play either better or not any worse
after 7/5(2).

MK

[Timothy Chow]

For what it's worth, below is an XGR+ rollout of the position. Murat's
play is on top, albeit by a very narrow margin. Note that the numbers
are displayed to 4 decimal places instead of the usual 3 decimal places.

XGID=-BBB-CCa-----B----bbbbbbbA:1:-1:-1:11:2:8:0:11:10

Score is X:8 O:2 11 pt.(s) match.
+13-14-15-16-17-18------19-20-21-22-23-24-+
| X | | O O O O O |
| | | O O O O O |
| | | O O |
| | | |
| | | |
| |BAR| |
| | O | |
| | | |
| | | | +---+
| O X | | X X X X X X | | 2 |
| O X | | X X X X X X | +---+
+12-11-10--9--8--7-------6--5--4--3--2--1-+
Pip count X: 74 O: 96 X-O: 8-2/11
Cube: 2, X own cube
X to play 11

1. Rollout¹ 7/5(2) eq:+1.0841
Player: 97.30% (G:21.85% B:0.21%)
Opponent: 2.70% (G:0.02% B:0.00%)
Confidence: ±0.0010 (+1.0831..+1.0851) - [77.9%]

2. Rollout¹ 18/15 7/6 eq:+1.0835 (-0.0006)
Player: 97.10% (G:22.33% B:0.21%)
Opponent: 2.90% (G:0.01% B:0.00%)
Confidence: ±0.0011 (+1.0824..+1.0845) - [18.6%]

3. Rollout¹ 18/14 eq:+1.0829 (-0.0011)
Player: 97.35% (G:21.62% B:0.22%)
Opponent: 2.65% (G:0.00% B:0.00%)
Confidence: ±0.0010 (+1.0820..+1.0839) - [3.3%]

4. Rollout¹ 7/6 7/4 eq:+1.0823 (-0.0018)
Player: 97.11% (G:22.16% B:0.21%)
Opponent: 2.89% (G:0.02% B:0.00%)
Confidence: ±0.0010 (+1.0812..+1.0833) - [0.2%]

¹ 2592 Games rolled with Variance Reduction.
Dice Seed: 271828
Moves and cube decisions: XG Roller+
Search interval: Gigantic

eXtreme Gammon Version: 2.19.211.pre-release, MET: Kazaross XG2

---
Tim Chow

[MK]

On December 12, 2023 at 9:27:48 PM UTC-7, Timothy Chow wrote:

> For what it's worth, below is an XGR+ rollout
> of the position. Murat's play is on top, albeit
> by a very narrow margin.

> X to play 11
> 1. Rollout¹ 7/5(2) eq:+1.0841

> 2. Rollout¹ 18/15 7/6 eq:+1.0835 (-0.0006)

> 3. Rollout¹ 18/14 eq:+1.0829 (-0.0011)

> 4. Rollout¹ 7/6 7/4 eq:+1.0823 (-0.0018)

As plain analysis,

2-ply plays really bad
3-ply picks 1=18/14, 2=7/5(2), 3=18/15 7/6
xg-roller picks the same as 3-ply
4-ply picks 1=7/5(2), 2=18/14, 3=18/15 7/6
xg-roller+ picks 1=18/15 7/6, 2=7/5(2), 3=18/14
xg-roller++ picks 1=18/14, 2=18/15 7/6, 3=7/5(2)

I find this pretty fucking funny. Except for 3-ply
and xg-roller all other plies give every different
conbination possible. :)

Granted they are all very close but shouldn't we
expect a little more consistency?

So then, do you guys take xg-roller+ rollout more
accurate than plain xg-roller++?

I suppose an xg-roller++ rollout would take some
hours but will it agree with the xg-roller+ rollout?

I personally wouldn't bet on it. So, what gives..?

Interestingly Noo-BG's various ply analyses are
just as inconsistent but then even the 0-ply with
maximum noise rollout picks 18/15 7/6 as the
best play...

I'm surprising myself that I'm taking interest in
a position discussion but it was valuable to see
that you can't make rocket science out of BG and
that, not always but often, mine or anyone else's
judgment is as good as any bg botts.

MK

[Timothy Chow]

On 12/13/2023 2:55 AM, MK wrote:

> Granted they are all very close but shouldn't we
> expect a little more consistency?

No. Why would anyone expect that when the plays are so close?

---
Tim Chow

[MK]

I'm not complaining about "some inconsistency"
but I would expect at least "some consistency".

They are not the same thing. I had said:

"Except for 3-ply and xg-roller all other plies
give every different conbination possible."

Obviously you missed my point that the results
were "anti-consistent", (as though on purpose).

Here they are again in more detail:

1. 2-ply 18/14 eq:+1.0888
2. 2-ply 18/15 7/6 eq:+1.0865 (-0.0023)
3. 2-ply 18/16 7/6(2) eq:+1.0834 (-0.0054)
4. 2-ply 18/17 7/6 7/5 eq:+1.0787 (-0.0101)
5. 2-ply 18/16 7/5 eq:+1.0754 (-0.0135)

1. 3-ply 18/14 eq:+1.0841
2. 3-ply 7/5(2) eq:+1.0839 (-0.0002)
3. 3-ply 7/6 7/4 eq:+1.0838 (-0.0003)
4. 3-ply 18/15 7/6 eq:+1.0795 (-0.0045)
5. 3-ply 18/16 7/6(2) eq:+1.0746 (-0.0095)

1. XG Roller 18/14 eq:+1.0839
2. XG Roller 7/6 7/4 eq:+1.0820 (-0.0019)
3. XG Roller 7/5(2) eq:+1.0808 (-0.0031)
4. XG Roller 18/15 7/6 eq:+1.0791 (-0.0048)
5. XG Roller 18/16 7/6(2) eq:+1.0721 (-0.0118)

1. 4-ply 7/5(2) eq:+1.0803
2. 4-ply 18/14 eq:+1.0787 (-0.0016)
3. 4-ply 7/6 7/4 eq:+1.0757 (-0.0045)
4. 4-ply 18/15 7/6 eq:+1.0742 (-0.0061)
5. 4-ply 18/16 7/6(2) eq:+1.0682 (-0.0120)

1. XG Roller+ 18/15 7/6 eq:+1.0766
2. XG Roller+ 7/5(2) eq:+1.0764 (-0.0003)
3. XG Roller+ 18/14 eq:+1.0762 (-0.0004)
4. XG Roller+ 7/6 7/4 eq:+1.0757 (-0.0010)
5. XG Roller+ 18/16 7/6(2) eq:+1.0676 (-0.0091)

1. XG Roller++ 18/14 eq:+1.0821
2. XG Roller++ 18/15 7/6 eq:+1.0818 (-0.0003)
3. XG Roller++ 7/5(2) eq:+1.0817 (-0.0004)
4. XG Roller++ 7/6 7/4 eq:+1.0803 (-0.0019)
5. XG Roller++ 18/16 7/6(2) eq:+1.0754 (-0.0067)

Forget any consistency among all 5 choices
but even just the first 2 choices coincide only
once and quite interestingly betwwn the lowest
2-ply and the highest XG Roller++ analyses..!

Contrary to your comments in the other thread,
they do indeed get more consistent with rollouts
but again with unexpected exceptions. Let's look:

First with moves and cube decisions: XG Roller

1. Rollout¹ 18/15 7/6 eq:+1.0844
2. Rollout¹ 7/6 7/4 eq:+1.0841 (-0.0004)
3. Rollout¹ 7/5(2) eq:+1.0827 (-0.0017)
4. Rollout¹ 18/14 eq:+1.0812 (-0.0033)
5. Rollout¹ 18/16 7/6(2) eq:+1.0731 (-0.0114)
324 Games rolled with Variance Reduction.

1. Rollout¹ 18/15 7/6 eq:+1.0855
2. Rollout¹ 7/6 7/4 eq:+1.0851 (-0.0004)
3. Rollout¹ 7/5(2) eq:+1.0836 (-0.0019)
4. Rollout¹ 18/14 eq:+1.0829 (-0.0026)
5. Rollout¹ 18/16 7/6(2) eq:+1.0745 (-0.0110)
648 Games rolled with Variance Reduction.

1. Rollout¹ 18/15 7/6 eq:+1.0854
2. Rollout¹ 7/5(2) eq:+1.0849 (-0.0005)
3. Rollout¹ 7/6 7/4 eq:+1.0846 (-0.0007)
4. Rollout¹ 18/14 eq:+1.0837 (-0.0017)
5. Rollout¹ 18/16 7/6(2) eq:+1.0756 (-0.0098)
1296 Games rolled with Variance Reduction.

Then with moves and cube decisions: XG Roller+

1. Rollout¹ 7/5(2) eq:+1.0854
2. Rollout¹ 18/15 7/6 eq:+1.0830 (-0.0023)
3. Rollout¹ 18/14 eq:+1.0824 (-0.0030)
4. Rollout¹ 7/6 7/4 eq:+1.0812 (-0.0042)
5. Rollout¹ 18/16 7/6(2) eq:+1.0762 (-0.0092)
324 Games rolled with Variance Reduction.

1. Rollout¹ 7/5(2) eq:+1.0852
2. Rollout¹ 18/15 7/6 eq:+1.0831 (-0.0021)
3. Rollout¹ 18/14 eq:+1.0814 (-0.0039)
4. Rollout¹ 7/6 7/4 eq:+1.0810 (-0.0042)
5. Rollout¹ 18/16 7/6(2) eq:+1.0764 (-0.0088)
648 Games rolled with Variance Reduction.

1. Rollout¹ 7/5(2) eq:+1.0846
2. Rollout¹ 18/15 7/6 eq:+1.0834 (-0.0012)
3. Rollout¹ 18/14 eq:+1.0825 (-0.0021)
4. Rollout¹ 7/6 7/4 eq:+1.0817 (-0.0029)
5. Rollout¹ 18/16 7/6(2) eq:+1.0766 (-0.0080)
1296 Games rolled with Variance Reduction.

Finally with moves and cube decisions: XG Roller++

1. Rollout¹ 18/15 7/6 eq:+1.0836
2. Rollout¹ 7/5(2) eq:+1.0832 (-0.0004)
3. Rollout¹ 18/14 eq:+1.0826 (-0.0010)
4. Rollout¹ 7/6 7/4 eq:+1.0820 (-0.0016)
5. Rollout¹ 18/16 7/6(2) eq:+1.0763 (-0.0073)
324 Games rolled with Variance Reduction.

1. Rollout¹ 18/15 7/6 eq:+1.0838
2. Rollout¹ 7/5(2) eq:+1.0837 (-0.0001)
3. Rollout¹ 18/14 eq:+1.0823 (-0.0015)
4. Rollout¹ 7/6 7/4 eq:+1.0819 (-0.0019)
5. Rollout¹ 18/16 7/6(2) eq:+1.0774 (-0.0064)
648 Games rolled with Variance Reduction.

1. Rollout¹ 7/5(2) eq:+1.0835
2. Rollout¹ 18/15 7/6 eq:+1.0834 (-0.0001)
3. Rollout¹ 18/14 eq:+1.0830 (-0.0004)
4. Rollout¹ 7/6 7/4 eq:+1.0821 (-0.0014)
5. Rollout¹ 18/16 7/6(2) eq:+1.0765 (-0.0070)
1296 Games rolled with Variance Reduction.

As would be expected, the results get more precise
with more trials, (i.e. the equity differences shrink).

And the different strength bots don't change their
minds with longer rollouts, except in the last one
which is the longest rollout for the strongest bot..!
(It took over 12 hours on low priority setting but it
was well worth it as we can see :)

It all looks like I am vindicated that ER/PR would
be horse muffins even if calculated by doing long
enough strong rollouts but they are worthless cow
pies even more based on simpler analyses done
during the games even by XG Roller++ :( So, just
go with 2-ply and be happy :))

MK

[Timothy Chow]

On 12/22/2023 10:48 AM, MK wrote:

> Contrary to your comments in the other thread,
> they do indeed get more consistent with rollouts
> but again with unexpected exceptions. Let's look:

There is nothing unexpected about your results. You have
omitted the most important quantity from your rollout results,
which is the confidence interval. It is to be expected that
there will be random fluctuations within, say, two times the
width of the confidence interval.

---
Tim Chow

[MK]

Confidence interval is a result, not a parameter of
the rollouts I am doing to compare their results. If
I use it as a paramater, it will defeat my purpose.

But there may be something I don't understand in
what you said and would like to hear you explain.

After achieving my own challenge of consistently
beating the bots behind me, I gradually took off my
"brain mask" for the past few years, being no longer
worried about keeping my mind uninfected by what
I call "fancyful bullshits" like equities, cube points,
etc. I feel fairly safe by now that I won't become one
of you all who try to play like the bots, (thus not be
able to beat them at their own game), and I'm more
willing to understand those "fancyful bullshits" in
order to better argue against them.

I have always argued that cube skill formulas to
extrapolate cubeful equities and match tables to
extrapolate "matchful" equities injected systematic
errors of unknown magnitute, each by itself or both
in combination.

I have also argued that they would accumulate in
proportion to the lengths of rollouts.

In my above rollouts, I was trying to isolate random
and systematic errors by comparing many rollouts
of different lengths and/or different plies. I realize
that a few rollouts aren't significant but it's a start.

Perhaps you can help me out with it by suggesting
a better way doing and comparing, cubeless/cubeful
and matchless/matchful (single games/multi-game
matches).

MK