I am on holiday from now (actually before now but I can't change the past)
to Wednesday, with very few time commitments.
So I'm deciding to read Axel's stuff on Isight.
Part of the point of this posting is to motivate myself to do it.
My technique for getting the material will be to google
"Isight backgammon Axel". Let's see what turns up.

It turns up this URL:
https://bkgm.com/articles/Reichert/insights-with-isight.pdf

If this is the wrong reference, please let me know.
I will add future posts to this thread if I find I have any constructive
observations. However, my definition of "constructive" differs
from that of many people. For example, I like the part of
the first hundred or so places in the decimal expansion of pi where
it goes 062862089986280. But many people might think memorizing
pi is a waste of time.

Paul

On Monday, April 18, 2022 at 10:53:56 AM UTC+1, peps...@gmail.com wrote:
> I am on holiday from now (actually before now but I can't change the past)
> to Wednesday, with very few time commitments.
> So I'm deciding to read Axel's stuff on Isight.
> Part of the point of this posting is to motivate myself to do it.
> My technique for getting the material will be to google
> "Isight backgammon Axel". Let's see what turns up.
>
> It turns up this URL:
> https://bkgm.com/articles/Reichert/insights-with-isight.pdf
>
> If this is the wrong reference, please let me know.
> I will add future posts to this thread if I find I have any constructive
> observations. However, my definition of "constructive" differs
> from that of many people. For example, I like the part of
> the first hundred or so places in the decimal expansion of pi where
> it goes 062862089986280. But many people might think memorizing
> pi is a waste of time.
>
> Paul

I can calculate square-roots to two decimal places or so pretty effortlessly, so the
fact that Isight avoids square-root computation is not a strong selling point for me.
Also, not sure why the Keith count gets mentioned so much in the abstract, whereas
other standard methods get mentioned only once.
Overall, I'm not particularly impressed by the abstract, but of course I will read further.

Paul

On Monday, April 18, 2022 at 10:53:56 AM UTC+1, peps...@gmail.com wrote:
> I am on holiday from now (actually before now but I can't change the past)
> to Wednesday, with very few time commitments.
> So I'm deciding to read Axel's stuff on Isight.
> Part of the point of this posting is to motivate myself to do it.
> My technique for getting the material will be to google
> "Isight backgammon Axel". Let's see what turns up.
>
> It turns up this URL:
> https://bkgm.com/articles/Reichert/insights-with-isight.pdf
>
> If this is the wrong reference, please let me know.
> I will add future posts to this thread if I find I have any constructive
> observations. However, my definition of "constructive" differs
> from that of many people. For example, I like the part of
> the first hundred or so places in the decimal expansion of pi where
> it goes 062862089986280. But many people might think memorizing
> pi is a waste of time.
>
> Paul

As a somewhat experienced player (PR between 5 and 6, I would guess),
I am deciding not to read the three references on races. But I'm sure much in those
articles would be new to me. Another point is that references have references which have references
and so on. It's not quite an infinite regression, but it's a regression that inhibits progress,
and may be part of the reason I don't know as much maths as Tim does.
When I try to learn maths, I consult a reference if I think it will help me understand a point.
But the reference has a reference. And that reference has a reference. And that reference has a reference.
Etc. etc. etc. Finally, in desperation, I might ask Tim. Then he says (rightly) that I should ask elsewhere.
Then people writing in the more standard forums like stackexchange refer me back to Tim.
(I made the last bit up, but I think it sounds good, anyway. The other bits are not made up).

I know the simple rules like 8/9/12, 10%+2. And I know what EPC means.
If anyone feels, that the external references in Axel's article need to be consulted as prerequisites,
please let me know.

Paul

On Monday, April 18, 2022 at 10:53:56 AM UTC+1, peps...@gmail.com wrote:
> I am on holiday from now (actually before now but I can't change the past)
> to Wednesday, with very few time commitments.
> So I'm deciding to read Axel's stuff on Isight.
> Part of the point of this posting is to motivate myself to do it.
> My technique for getting the material will be to google
> "Isight backgammon Axel". Let's see what turns up.
>
> It turns up this URL:
> https://bkgm.com/articles/Reichert/insights-with-isight.pdf
>
> If this is the wrong reference, please let me know.
> I will add future posts to this thread if I find I have any constructive
> observations. However, my definition of "constructive" differs
> from that of many people. For example, I like the part of
> the first hundred or so places in the decimal expansion of pi where
> it goes 062862089986280. But many people might think memorizing
> pi is a waste of time.

Huge blunder(s) in the discussion of the pips-even race with 5 checkers on the ace against one checker on the 5.
with the stacked player to roll. After the roll, it is double/pass. But the pass is not optional, as incorrectly stated
People might simply stop reading if they see that pass described as "optional". I will continue though.
Axel might be getting confused with the optional pass positions where the on-roll player has a single checker
on the 6 (not the 5). It's doubtful that a strong player would make such an error.
But this is a specialised and very mathematical phase of the game. And it's absolutely plausible that a
weak player could make a strong contribution here.
So it doesn't stop me reading (besides, I've already made a commitment to do so), but I think it would turn off many players.
Urgently needs correction. (Apologies if this has already been done and my link is outdated.)

Paul

On Monday, April 18, 2022 at 10:53:56 AM UTC+1, peps...@gmail.com wrote:
> I am on holiday from now (actually before now but I can't change the past)
> to Wednesday, with very few time commitments.
> So I'm deciding to read Axel's stuff on Isight.
> Part of the point of this posting is to motivate myself to do it.
> My technique for getting the material will be to google
> "Isight backgammon Axel". Let's see what turns up.
>
> It turns up this URL:
> https://bkgm.com/articles/Reichert/insights-with-isight.pdf
>
> If this is the wrong reference, please let me know.
> I will add future posts to this thread if I find I have any constructive
> observations. However, my definition of "constructive" differs
> from that of many people. For example, I like the part of
> the first hundred or so places in the decimal expansion of pi where
> it goes 062862089986280. But many people might think memorizing
> pi is a waste of time.
>
> Paul

I don't know Axel personally (beyond reading some of his posts here) and
know nothing about his career.
But I believe that, if he wants to (or does) work in areas related to non-linear parameter
optimization or similar areas of maths and data science, the Isight paper could make
a great marketing tool for career development.
For this purpose, I would totally change the title to focus on the career-relevant techniques.
I would also initially hide the fact that the application to the optimization is to play
a board game better, since board games are often seen as frivolous.
So there could be another "career" version where the title is (something like): "A surprising application of
non-linear parameter optimization" and the fact that the application is backgammon races is announced in the
first sentence of the abstract, rather than the title.

Paul

On April 18, 2022 at 5:10:07 AM UTC-6, peps...@gmail.com wrote:

Jesus Christ, man! I felt so bad about following up
to my own post once. Don't you feel any shame in
being the only one to follow up to your own post
four times in a row?! Stop fucking yourself in public!

MK

On Monday, April 18, 2022 at 10:53:56 AM UTC+1, peps...@gmail.com wrote:
> I am on holiday from now (actually before now but I can't change the past)
> to Wednesday, with very few time commitments.
> So I'm deciding to read Axel's stuff on Isight.
> Part of the point of this posting is to motivate myself to do it.
> My technique for getting the material will be to google
> "Isight backgammon Axel". Let's see what turns up.
>
> It turns up this URL:
> https://bkgm.com/articles/Reichert/insights-with-isight.pdf
>
> If this is the wrong reference, please let me know.
> I will add future posts to this thread if I find I have any constructive
> observations. However, my definition of "constructive" differs
> from that of many people. For example, I like the part of
> the first hundred or so places in the decimal expansion of pi where
> it goes 062862089986280. But many people might think memorizing
> pi is a waste of time.

Despite the focus on ease of application, the section on adjusted pip count
says nothing about mnemonics or other tools to aid the reader's memory for applying the
rules. My mnemonic is "Strictly low 2 2 3 2 1 1 relative all 1 high crossover extra."
That's my personal way of summarizing the rules so that I can remember them.

The order in which the rules are written seems arbitrary, but it matters crucially, for ease of memory.
I think the right order is:
1. Rule for acepoint stack
2. Rule for two point
3. Rule for three point.
4. Rule for 4/5/6 points.
5. Rule for extra checker.
6. Rule for crossover.

What are the advantages of my ordering? Essentially, the separation between the relative and absolute
rules is much clearer and cleaner: the first three rules are all absolute and the final three are all relative.
Absolute rules are easier than relative rules (as the paper says) hence these are given first.
The rules concerning numbered points, and the other rules are also presented in two separate blocks and
not interspersed.
The numbered-points rules are in the obvious numerical order.
Extra checker counting is simpler than crossover counting: if all other considerations are equal, the simpler the rule,
the nearer it should be to the beginning.

Paul

On Monday, April 18, 2022 at 12:23:51 PM UTC+1, MK wrote:
> On April 18, 2022 at 5:10:07 AM UTC-6, peps...@gmail.com wrote:
>
> Jesus Christ, man! I felt so bad about following up
> to my own post once. Don't you feel any shame in
> being the only one to follow up to your own post
> four times in a row?! Stop fucking yourself in public!
>
> MK

It's not ideal, but if I don't do it this way, my comments simply won't get done,
bearing in mind all my other commitments after this short holiday.
I'm sure Axel is interested, and probably Tim, too.
Realistically, the choice is probably between this tedious series of solipsistic posts,
and not posting on this topic at all.
You would probably prefer the no-posting option, but I'm sure Axel would probably
prefer the post-obsessively option.
Tim probably prefers the post-obsessively option too.

Paul

On Monday, April 18, 2022 at 10:53:56 AM UTC+1, peps...@gmail.com wrote:
> I am on holiday from now (actually before now but I can't change the past)
> to Wednesday, with very few time commitments.
> So I'm deciding to read Axel's stuff on Isight.
> Part of the point of this posting is to motivate myself to do it.
> My technique for getting the material will be to google
> "Isight backgammon Axel". Let's see what turns up.
>
> It turns up this URL:
> https://bkgm.com/articles/Reichert/insights-with-isight.pdf
>
> If this is the wrong reference, please let me know.
> I will add future posts to this thread if I find I have any constructive
> observations. However, my definition of "constructive" differs
> from that of many people. For example, I like the part of
> the first hundred or so places in the decimal expansion of pi where
> it goes 062862089986280. But many people might think memorizing
> pi is a waste of time.
>
> Paul

I think I've said this before, but the claim that the Isight method works better,
in the sense of error minimization, than other comparable methods, is utterly
without justification. Note that an unjustified claim is not necessarily a false
claim.
The problem with the claim is that there is no evidence of separation of the data
into in-sample and out-of-sample portions. If you use the same data to calibrate
the method, as you do to test the accuracy of your method, then this is clear (if inadvertent)
cheating.
I think I brought up this point before, and Axel clarified that the method is also validated
with out-of-sample tests, too. That's good (if true) but it's needed in the actual paper.

Paul

On 4/18/2022 7:23 AM, MK wrote:
> On April 18, 2022 at 5:10:07 AM UTC-6, peps...@gmail.com wrote:
>
> Jesus Christ, man! I felt so bad about following up
> to my own post once.

Ah, so that's why you use sock puppets!

---
Tim Chow

"peps...@gmail.com" <pepstein5@gmail.com> writes:

> Also, not sure why the Keith count gets mentioned so much in the
> abstract, whereas other standard methods get mentioned only once.

Because the Keith count was the best so far (and by far, see table
9). Also, because I built heavily on Tom's outstanding work and really
felt obliged to mention his outstanding work.

Axel

"peps...@gmail.com" <pepstein5@gmail.com> writes:

> I don't know Axel personally (beyond reading some of his posts here)
> and know nothing about his career. But I believe that, if he wants to
> (or does) work in areas related to non-linear parameter optimization
> or similar areas of maths and data science, the Isight paper could
> make a great marketing tool for career development.

I did several talks on this over the years. Normally, the software
Isight is used for parameter optimization of, say, automotive or
aerospace parts, classical engineering problems. But most engineers like
to play with their favourite toys and my presentations of applying the
tool to something completely different were well received. It just
showed how generally these tools can be applied and always made for a
fun discussion with people who know optimization, but do not know
backgammon.

> board games are often seen as frivolous

Yes. That was one of the more frequent remarks from the audience.

> So there could be another "career" version where the title is
> (something like): "A surprising application of
> non-linear parameter optimization"

Almost! "Process Integration and Design Optimization: A Playful
Approach"

Best regards

Axel

On Tuesday, April 19, 2022 at 2:46:17 AM UTC-4, Axel Reichert wrote:
> "peps...@gmail.com" <peps...@gmail.com> writes:
>
> > Also, not sure why the Keith count gets mentioned so much in the
> > abstract, whereas other standard methods get mentioned only once.
> Because the Keith count was the best so far (and by far, see table
> 9). Also, because I built heavily on Tom's outstanding work and really
> felt obliged to mention his outstanding work.
>
> Axel

Especially if you use the Keith Count on the regular and apply known fixes to some of its common problems.

Stick

Stick Rice <bananaboater315@gmail.com> writes:

> if you use the Keith Count on the regular and apply known fixes to
> some of its common problems.

If you care to let me know the common problems and its fixes, it is
likely a simple task to run my slightly augmented script over both
databases (the one I used for calibrating and the one I later used for
verifying).

Best regards

Axel

On April 18, 2022 at 8:16:01 AM UTC-6, Tim Chow wrote:

> On 4/18/2022 7:23 AM, MK wrote:

>> Jesus Christ, man! I felt so bad about following up
>> to my own post once.

> Ah, so that's why you use sock puppets!

I quoted this to refer to it in my reply to Paul below.

-------------------------------------------------------

On April 18, 2022 at 5:37:53 AM UTC-6, peps...@gmail.com wrote:

> It's not ideal, but if I don't do it this way, my comments
> simply won't get done.....
> Realistically, the choice is probably between this tedious
> series of solipsistic posts, and not posting on this topic at all.

Okay, sorry. :( I misjudged where you were going with it.

> I'm sure Axel is interested, and probably Tim, too.

I think you are overesteeming Tim. As you can see from
his comment above, when he doesn't have or runs out of
anything of substance to say, he starts with his stupid
one-liners (as he started doing in the other thread also).

> You would probably prefer the no-posting option,

No, I do, in fact, prefer to read and participate in lengthy
discussions with detailed posts also.

> but I'm sure Axel would probably prefer the post-obsessively
> option. Tim probably prefers the post-obsessively option too.

We all often argue obsessively, in trying to be right and
can't just let go of it but that's okay with me as long as
we make efforts to offer arguments backed up data and
logic, even if they gradually become scattered/diluted
and any progress made becomes incrementally small.

It's very visible how Tim and you are two peas in a pod
but I won't put Axel in the same basket with you two yet.
Especially Tim, for me, acts like an inferiority complexed
little a piece of greasy turd who always needs to float to
the top, not realizing how pathetic he sounds, still talking
about my sock puppets, etc... :(

MK

On Thursday, April 21, 2022 at 2:12:48 AM UTC+1, MK wrote:
> On April 18, 2022 at 8:16:01 AM UTC-6, Tim Chow wrote:
> > On 4/18/2022 7:23 AM, MK wrote:
>
> >> Jesus Christ, man! I felt so bad about following up
> >> to my own post once.
> > Ah, so that's why you use sock puppets!
> I quoted this to refer to it in my reply to Paul below.
>
> -------------------------------------------------------
> On April 18, 2022 at 5:37:53 AM UTC-6, peps...@gmail.com wrote:
>
> > It's not ideal, but if I don't do it this way, my comments
> > simply won't get done.....
> > Realistically, the choice is probably between this tedious
> > series of solipsistic posts, and not posting on this topic at all.
> Okay, sorry. :( I misjudged where you were going with it.
> > I'm sure Axel is interested, and probably Tim, too.
> I think you are overesteeming Tim. As you can see from
> his comment above, when he doesn't have or runs out of
> anything of substance to say, he starts with his stupid
> one-liners (as he started doing in the other thread also).
I don't see it that way. I'm somewhat negatively surprised at myself,
that I'm not coming up with much (if anything) of value that
I didn't already say in Nov 2020. But I didn't know that the series
would end up so weak when I started it.
Tim hasn't responded much but that's presumably because my content
has been poor.

> > You would probably prefer the no-posting option,
> No, I do, in fact, prefer to read and participate in lengthy
> discussions with detailed posts also.
> > but I'm sure Axel would probably prefer the post-obsessively
> > option. Tim probably prefers the post-obsessively option too.
> We all often argue obsessively, in trying to be right and
> can't just let go of it but that's okay with me as long as
> we make efforts to offer arguments backed up data and
> logic, even if they gradually become scattered/diluted
> and any progress made becomes incrementally small.
>
> It's very visible how Tim and you are two peas in a pod
Yes, as far as I can tell, Tim and I are strikingly similar in many ways.
But we've never met IRL.

> but I won't put Axel in the same basket with you two yet.
> Especially Tim, for me, acts like an inferiority complexed
> little a piece of greasy turd who always needs to float to
> the top, not realizing how pathetic he sounds, still talking
> about my sock puppets, etc... :(
>
> MK

On Wednesday, April 20, 2022 at 5:26:29 PM UTC-4, Axel Reichert wrote:
> Stick Rice <bananab...@gmail.com> writes:
>
> > if you use the Keith Count on the regular and apply known fixes to
> > some of its common problems.
> If you care to let me know the common problems and its fixes, it is
> likely a simple task to run my slightly augmented script over both
> databases (the one I used for calibrating and the one I later used for
> verifying).
>
> Best regards
>
> Axel

Keith Count ~80 to ~100 pips can double when leading by 5
Keith Count ~100 pips to 120pips can double when leading by 6
Keith Count 120 pips + can double when leading by 7

The most common problem and its fix can't be applied. It's recognizing that the position is not suited to the Keith Count and simply using another option, like EPC.

Stick

On 4/20/2022 9:12 PM, MK wrote:
> I think you are overesteeming Tim. As you can see from
> his comment above, when he doesn't have or runs out of
> anything of substance to say, he starts with his stupid
> one-liners (as he started doing in the other thread also).

I see a typo: where you wrote "he doesn't have or runs out"
you meant to write, "I don't have or run out."

---
Tim Chow

Stick Rice <bananaboater315@gmail.com> writes:

> Keith Count ~80 to ~100 pips can double when leading by 5
> Keith Count ~100 pips to 120pips can double when leading by 6
> Keith Count 120 pips + can double when leading by 7

Thanks.

I take this to mean said number of pips *before* increasing by 1/7. Also
you meant probably (corresponding to Tom's original wording):

A player should double if his count exceeds the opponent's count by no
more than 5 (for his count >80 and <=100)

Similarly for 6 and 7 with your brackets given above. I also assumed
that redoubling occurs 1 pip later:

A player should redouble if his count exceeds the opponent's count by no
more than 4 (for his count >80 and <=100)

Similarly for 5 and 6 for the longer races.

This helps a bit, shaving off 5 % from the total error. Playing around
with the numbers manually, I could improve still a little bit, but my
method is still considerably better:

| Count | Tom's Database | Axel's Database |
|------------------------+----------------+-----------------|
| Original Keith | 1262 | 1165 |
| Stick's Modified Keith | 1204 | 1202 |
| Axel's Modified Keith | 1182 | 1176 |
| Axel's Isight Method | 1064 | 965 |

Best regards

Axel

On Thursday, April 21, 2022 at 7:22:20 PM UTC+1, Axel Reichert wrote:
> Stick Rice <bananab...@gmail.com> writes:
>
> > Keith Count ~80 to ~100 pips can double when leading by 5
> > Keith Count ~100 pips to 120pips can double when leading by 6
> > Keith Count 120 pips + can double when leading by 7
> Thanks.
>
> I take this to mean said number of pips *before* increasing by 1/7. Also
> you meant probably (corresponding to Tom's original wording):
>
> A player should double if his count exceeds the opponent's count by no
> more than 5 (for his count >80 and <=100)
>
> Similarly for 6 and 7 with your brackets given above. I also assumed
> that redoubling occurs 1 pip later:
>
> A player should redouble if his count exceeds the opponent's count by no
> more than 4 (for his count >80 and <=100)
>
> Similarly for 5 and 6 for the longer races.
>
> This helps a bit, shaving off 5 % from the total error. Playing around
> with the numbers manually, I could improve still a little bit, but my
> method is still considerably better:
>
> | Count | Tom's Database | Axel's Database |
> |------------------------+----------------+-----------------|
> | Original Keith | 1262 | 1165 |
> | Stick's Modified Keith | 1204 | 1202 |
> | Axel's Modified Keith | 1182 | 1176 |
> | Axel's Isight Method | 1064 | 965 |
>
> Best regards

Axel,

I'm sure Stick knows this.
I think his claim is that he can improve on Isight by using the Keith count
as his main algo but then using some type of adjustment in some cases.
But he's totally vague on what the adjustments are and what the cases are.
It's possible that, as a pro, he can't reveal all his secrets.
Certainly chess pros keep plenty of opening secrets.

Paul

"peps...@gmail.com" <pepstein5@gmail.com> writes:

> On Thursday, April 21, 2022 at 7:22:20 PM UTC+1, Axel Reichert wrote:
>>
>> A player should double if his count exceeds the opponent's count by
>> no more than 5 (for his count >80 and <=100)
>>
>> Similarly for 6 and 7 with your brackets given above. I also assumed
>> that redoubling occurs 1 pip later:
>>
>> A player should redouble if his count exceeds the opponent's count by
>> no more than 4 (for his count >80 and <=100)
>>
>> Similarly for 5 and 6 for the longer races.

[...]

>> I could improve still a little bit, but my method is still
>> considerably better

[...]

> I think his claim is that he can improve on Isight by using the Keith
> count as his main algo but then using some type of adjustment in some
> cases.

After all my work, I am quite confident that *within my parameterized
framework* the Isight method is the optimum one (without getting too
technical here). This of course leaves plenty of room for even better
method *outside* this framework.

Stick essentially proposed to distinguish between 4 different race
lengths, while my framework caters for only 2. Accompagning these 2
additional thresholds (say, 100 and 120 pips) with their "bandwidths"
for the size of the doubling window introduces 4 new parameters in
total.

Adding this complexity and mental overhead barely halves the performance
gap to my method, which does not even consider different race lengths.
Since I am currently reading "Life is simple" about Occam's Razor, this
reminds me on Ptolemy's epicycles. (-;

My gut feeling says that any major improvement on race cube decisions
will not be based on adding this feature or that (at considerable cost!)
to my (quite general) framework, but rather on something fundamentally
different, namely EPCs along with a doubling criterion matched to them.

But this will be tricky, as mentioned already years back in my
article. Last year I played around with Jean-Luc Seret's "pipples",
which are supposed to give a very accurate estimate of the EPC, see
https://www.bkgm.com/articles/GOL/Dec00/pipples.htm:

"for 85% of the positions the error is smaller than 10 pipples"

Since 1 pipple is 1/100 of a roll, 10 pipples are worth 0.816 pips. From
Table 6 in my article I expected a total error on Tom Keith's database
roughly in the range of the Keith count, which was confirmed by my test.

And to my surprise, even when I eliminated all positions from Tom's
database that did not stick (no pun intended) to Seret's many (and quite
limiting!) restrictions (

- No checkers outside
- At least 7 checkers
- At most 7 checkers on any single point
- Pipcount from 30 to 70

), on this much smaller database (it was designed for these positions!)
Seret's method still had a total error about 25 % higher than my Isight
method. I combined Trice's EPC doubling criterion with Seret's pipple
calculation. Trice's criterion is not the culprit, see the last row of
table 6 in my article. It is the lack of accuracy of the EPC
approximations.

So currently I am running out of ideas for races. (-;

Best regards

Axel

On Thursday, April 21, 2022 at 10:01:24 PM UTC+1, Axel Reichert wrote:
> "peps...@gmail.com" <peps...@gmail.com> writes:
>
> > On Thursday, April 21, 2022 at 7:22:20 PM UTC+1, Axel Reichert wrote:
> >>
> >> A player should double if his count exceeds the opponent's count by
> >> no more than 5 (for his count >80 and <=100)
> >>
> >> Similarly for 6 and 7 with your brackets given above. I also assumed
> >> that redoubling occurs 1 pip later:
> >>
> >> A player should redouble if his count exceeds the opponent's count by
> >> no more than 4 (for his count >80 and <=100)
> >>
> >> Similarly for 5 and 6 for the longer races.
> [...]
> >> I could improve still a little bit, but my method is still
> >> considerably better
> [...]
> > I think his claim is that he can improve on Isight by using the Keith
> > count as his main algo but then using some type of adjustment in some
> > cases.
> After all my work, I am quite confident that *within my parameterized
> framework* the Isight method is the optimum one (without getting too
> technical here). This of course leaves plenty of room for even better
> method *outside* this framework.
>
> Stick essentially proposed to distinguish between 4 different race
> lengths, while my framework caters for only 2. Accompagning these 2
> additional thresholds (say, 100 and 120 pips) with their "bandwidths"
> for the size of the doubling window introduces 4 new parameters in
> total.
>
> Adding this complexity and mental overhead barely halves the performance
> gap to my method, which does not even consider different race lengths.
> Since I am currently reading "Life is simple" about Occam's Razor, this
> reminds me on Ptolemy's epicycles. (-;
>
> My gut feeling says that any major improvement on race cube decisions
> will not be based on adding this feature or that (at considerable cost!)
> to my (quite general) framework, but rather on something fundamentally
> different, namely EPCs along with a doubling criterion matched to them.
>
> But this will be tricky, as mentioned already years back in my
> article. Last year I played around with Jean-Luc Seret's "pipples",
> which are supposed to give a very accurate estimate of the EPC, see
> https://www.bkgm.com/articles/GOL/Dec00/pipples.htm:
>
> "for 85% of the positions the error is smaller than 10 pipples"
>
> Since 1 pipple is 1/100 of a roll, 10 pipples are worth 0.816 pips. From
> Table 6 in my article I expected a total error on Tom Keith's database
> roughly in the range of the Keith count, which was confirmed by my test.
>
> And to my surprise, even when I eliminated all positions from Tom's
> database that did not stick (no pun intended) to Seret's many (and quite
> limiting!) restrictions (
>
> - No checkers outside
> - At least 7 checkers
> - At most 7 checkers on any single point
> - Pipcount from 30 to 70
>
> ), on this much smaller database (it was designed for these positions!)
> Seret's method still had a total error about 25 % higher than my Isight
> method. I combined Trice's EPC doubling criterion with Seret's pipple
> calculation. Trice's criterion is not the culprit, see the last row of
> table 6 in my article. It is the lack of accuracy of the EPC
> approximations.
>
> So currently I am running out of ideas for races. (-;
>
> Best regards
>
> Axel

I think you and Stick are both right.
If we stick to an algo with clear params, your method wins.
Stick is pointing out that, as one of the very greatest players in the world,
he can outperform your algo.
How does he do this?
He has heuristics that say things like (and these are just examples):
"Here we have a strange position which is flat and smooth from points 1 to 4 but with large
stacks on the low points. For this type of position, [some count] does best."
(I found a position with that characteristic where Isight didn't seem to work well -- admittedly just one position).
Or "Here, we have a strange race where the pip count is even but X has many fewer checkers. In this situation, I do..."

With such an intuitive way of thinking, combined with experience, I'm sure that Stick (or any of the top ten players in the world) can outperform a
simple parameterised algo.

Now, if you agree with me on the above, then it's also likely that the best approach for top players might be "Use X algo as a base but vary heuristically
and intuitively according to experience."
If you accept the above (which seems totally plausible), there is no reason to assume that if you replace X by Isight, you do better than replacing X by the Keith
count. The way to test such hypotheses is to see how popular Isight becomes among the top players.

Paul

On Friday, April 22, 2022 at 2:22:09 PM UTC+1, peps...@gmail.com wrote:
> On Thursday, April 21, 2022 at 10:01:24 PM UTC+1, Axel Reichert wrote:
> > "peps...@gmail.com" <peps...@gmail.com> writes:
> >
> > > On Thursday, April 21, 2022 at 7:22:20 PM UTC+1, Axel Reichert wrote:
> > >>
> > >> A player should double if his count exceeds the opponent's count by
> > >> no more than 5 (for his count >80 and <=100)
> > >>
> > >> Similarly for 6 and 7 with your brackets given above. I also assumed
> > >> that redoubling occurs 1 pip later:
> > >>
> > >> A player should redouble if his count exceeds the opponent's count by
> > >> no more than 4 (for his count >80 and <=100)
> > >>
> > >> Similarly for 5 and 6 for the longer races.
> > [...]
> > >> I could improve still a little bit, but my method is still
> > >> considerably better
> > [...]
> > > I think his claim is that he can improve on Isight by using the Keith
> > > count as his main algo but then using some type of adjustment in some
> > > cases.
> > After all my work, I am quite confident that *within my parameterized
> > framework* the Isight method is the optimum one (without getting too
> > technical here). This of course leaves plenty of room for even better
> > method *outside* this framework.
> >
> > Stick essentially proposed to distinguish between 4 different race
> > lengths, while my framework caters for only 2. Accompagning these 2
> > additional thresholds (say, 100 and 120 pips) with their "bandwidths"
> > for the size of the doubling window introduces 4 new parameters in
> > total.
> >
> > Adding this complexity and mental overhead barely halves the performance
> > gap to my method, which does not even consider different race lengths.
> > Since I am currently reading "Life is simple" about Occam's Razor, this
> > reminds me on Ptolemy's epicycles. (-;
> >
> > My gut feeling says that any major improvement on race cube decisions
> > will not be based on adding this feature or that (at considerable cost!)
> > to my (quite general) framework, but rather on something fundamentally
> > different, namely EPCs along with a doubling criterion matched to them.
> >
> > But this will be tricky, as mentioned already years back in my
> > article. Last year I played around with Jean-Luc Seret's "pipples",
> > which are supposed to give a very accurate estimate of the EPC, see
> > https://www.bkgm.com/articles/GOL/Dec00/pipples.htm:
> >
> > "for 85% of the positions the error is smaller than 10 pipples"
> >
> > Since 1 pipple is 1/100 of a roll, 10 pipples are worth 0.816 pips. From
> > Table 6 in my article I expected a total error on Tom Keith's database
> > roughly in the range of the Keith count, which was confirmed by my test.
> >
> > And to my surprise, even when I eliminated all positions from Tom's
> > database that did not stick (no pun intended) to Seret's many (and quite
> > limiting!) restrictions (
> >
> > - No checkers outside
> > - At least 7 checkers
> > - At most 7 checkers on any single point
> > - Pipcount from 30 to 70
> >
> > ), on this much smaller database (it was designed for these positions!)
> > Seret's method still had a total error about 25 % higher than my Isight
> > method. I combined Trice's EPC doubling criterion with Seret's pipple
> > calculation. Trice's criterion is not the culprit, see the last row of
> > table 6 in my article. It is the lack of accuracy of the EPC
> > approximations.
> >
> > So currently I am running out of ideas for races. (-;
> >
> > Best regards
> >
> > Axel
> I think you and Stick are both right.
> If we stick to an algo with clear params, your method wins.
> Stick is pointing out that, as one of the very greatest players in the world,
> he can outperform your algo.
> How does he do this?
> He has heuristics that say things like (and these are just examples):
> "Here we have a strange position which is flat and smooth from points 1 to 4 but with large
> stacks on the low points. For this type of position, [some count] does best."
> (I found a position with that characteristic where Isight didn't seem to work well -- admittedly just one position).
> Or "Here, we have a strange race where the pip count is even but X has many fewer checkers. In this situation, I do..."
>
> With such an intuitive way of thinking, combined with experience, I'm sure that Stick (or any of the top ten players in the world) can outperform a
> simple parameterised algo.
>
> Now, if you agree with me on the above, then it's also likely that the best approach for top players might be "Use X algo as a base but vary heuristically
> and intuitively according to experience."
> If you accept the above (which seems totally plausible), there is no reason to assume that if you replace X by Isight, you do better than replacing X by the Keith
> count. The way to test such hypotheses is to see how popular Isight becomes among the top players.

More succinctly, take the informal approach: "Use [algo X] as an initial approx and adjust using experience and intuition."
Even if algo X is the best when used in an automated botlike fashion (and there's strong evidence that X = Isight), this by no means indicates that
algo X is the best for the informal approach.
The problem is that it's hard to compare algos with the informal approach.
I think Stick's method is [Keith Count + informal intuitive adjustments]
There's no evidence that Isight beats this (or matches this) and there's also no evidence that [Isight + informal intuitive adjustments] beats Stick's method (or matches it).

Paul

On 4/22/2022 9:32 AM, peps...@gmail.com wrote:
> The problem is that it's hard to compare algos with the informal approach.
> I think Stick's method is [Keith Count + informal intuitive adjustments]
> There's no evidence that Isight beats this (or matches this) and there's also no evidence that [Isight + informal intuitive adjustments] beats Stick's method (or matches it).

With a player who is sufficiently active, and whose matches are
recorded, one could accumulate some evidence. But I gather that
Stick doesn't play a lot of tournament backgammon nowadays.

---
Tim Chow

On Saturday, April 23, 2022 at 9:18:32 PM UTC+1, Tim Chow wrote:
> On 4/22/2022 9:32 AM, peps...@gmail.com wrote:
> > The problem is that it's hard to compare algos with the informal approach.
> > I think Stick's method is [Keith Count + informal intuitive adjustments]
> > There's no evidence that Isight beats this (or matches this) and there's also no evidence that [Isight + informal intuitive adjustments] beats Stick's method (or matches it).
> With a player who is sufficiently active, and whose matches are
> recorded, one could accumulate some evidence. But I gather that
> Stick doesn't play a lot of tournament backgammon nowadays.
>

No, I think that, once you accept that strong players are using intuitive adjustments of known algos, hypotheses about
which algos form the base are untestable in principle.
For example, suppose Isight recommends D/P in a position where Keith Count recommends D/T, and a player passes.
We can't know whether the player is adjusting the Keith Count to a pass or using Isight.
We can make statements like "If you make the simplifying assumption that strong players use a specific algo consistently,
then Isight is the most usual choice."
But here the simplifying assumption is false. And it's perfectly consistent with the above that Keith Count with intuitive adjustment is
the method of choice.

Paul