Anchors Aweigh
Stephen Hubbard
stu...@ix.netcom.com (Stuart Katz, MD) wrote:
>money game; cube centered; X has one on bar; X to play 61
> +24-23-22-21-20-19-+---+18-17-16-15-14-13-+
> | X O X X O | X | O O O |
> | X O X X O | | O O O |
> | O | | |
> | O | | |
> | | | |
> | | | |[64]
> | | | |
> | | | |
> | | | |
> | | | O |
> | X X X X | | O |
> | X X X X | | O |
> +-1--2--3--4--5--6-+---+-7--8--9-10-11-12-+
>Which anchor should you leave, and why?
>Stuart
Recently I have been reading "New Ideas in Backgammon" by Woolsey and
Heinrich. Lets see if anything has rubbed off. :)
Obviously to win x has to get a hit and do a lot of escaping.
First of all I would keep the anchor on 23 because of being so far
behind in the race--51 pips after this move. Abandoning this anchor
seems gives more of a chance for o to bypass the more advanced anchor
without getting hit.
So the question is whether to leave 21 or 20.
JF recommends leaving 21. Why would this produce more equity?
Leaving 21 seems to put more pressure in the outfield and maintain
more contact .
Keeping 20 seems to help future escape chances.
I would think that this move would help timing also by giving more
escaped checkers to work with. Sound explanation or not?
Thanks for any feedback.
Robert-Jan Veldhuizen
On 17-mei-97 15:49:03, Stuart Katz, MD wrote:
SKM> money game; cube centered; X has one on bar; X to play 61
SKM> +24-23-22-21-20-19-+---+18-17-16-15-14-13-+
SKM> | X O X X O | X | O O O |
SKM> | X O X X O | | O O O |
SKM> | O | | |
SKM> | O | | |
SKM> | | | |
SKM> | | | |[64]
SKM> | | | |
SKM> | | | |
SKM> | | | |
SKM> | | | O |
SKM> | X X X X | | O |
SKM> | X X X X | | O |
SKM> +-1--2--3--4--5--6-+---+-7--8--9-10-11-12-+
SKM> Which anchor should you leave, and why?
I would just leave the 23pt anchor. Advantages are that you will probably not
get hit loose in O's innerboard on the 23 or 24 pt, so you give O no chance of
improving his board.
Also X's chances of hitting a blot seem most likely for the 12, 13 and 16 pts
which O has to clear. The 4 and 5 pt anchor cover those best, the 23 pt doesn't
seem that valuable.
Disadvantage is of course that you loose a deep anchor while being much behind
in the race. But O's distribution in the outfield in not very good and you are
likely to have a few shots there, I think.
But I've been proven wrong several times already... :)
--
Zorba/Robert-Jan
Brian Sheppard
Stephen Hubbard <the...@mindspring.com> wrote in article
<5ll1ib$9...@camel1.mindspring.com>...> > +24-23-22-21-20-19-+---+18-17-16-15-14-13-+
> > | X O X X O | X | O O O |
> > | O | | |
> > | | | |[64]
> > | | | O |
> > | X X X X | | O |
> > | X X X X | | O |
> >money game; cube centered; X has one on bar; X to play 61
>
> behind in the race--51 pips after this move. Abandoning this anchor
> seems gives more of a chance for o to bypass the more advanced anchor
> without getting hit.
>
> So the question is whether to leave 21 or 20.
> JF recommends leaving 21. Why would this produce more equity?
My copy of JF (version 2) plays bar/24, 23/17 at every level,
including rollouts. Perhaps you are using a different version?
> Leaving 21 seems to put more pressure in the outfield and maintain
> more contact .
> Keeping 20 seems to help future escape chances.
> I would think that this move would help timing also by giving more
> escaped checkers to work with. Sound explanation or not?
I don't agree. Breaking the 23 point is definitely best.
The fundamental issue in this type of position is timing. How long
can X preserve his anchors? How long can X preserve his board? When
will X get a shot?
Looking at X's position, we see that it is about to crack
someplace. X has to give up an anchor immediately. And
X's board is compromised by the gap on his 5 point. This is
important to the proper evaluation of this position because it
means that X has a lot of work to do even after hitting O.
On the other side of the table, O is not in danger of cracking.
Sure, there are rolls that force O to give up a point, and even
leave a blot. But those blots are indirect shots, and being hit
is not immediately decisive because X has to bring his army around
and O keeps all his men in play.
So timing is bad for X. And this leads to a clear decision for
which anchor to keep: keep the high anchors.
I have no doubt that there are positions where this rule fails, due
to overriding tactical urgencies, but I don't see any such factors
here. The biggest danger is O scrambling home while X is on the
bar, so X never gets more than a few indirect shots. But X is as
likely to suffer that fate if he breaks another anchor.
This decision is supported by rollouts. I won't give exact numbers,
because I believe JF doesn't play these positions very well. But the
general outline of the game is as follows:
1) No matter what X plays, O should double immediately.
2) X has a clear take after 23/17, and also a take after the other
plays but it is close.
3) O should hit that man on the 17 point, because
4) X should try to cover that man on the 17 point, which will
hinder O's entry and facilitate X's escape. X is quite likely to
be forced to give up another anchor, so it is very important to
keep a passageway for the men as they come out.
5) In general, X should try to play with a minimum of blots, since
staying on the bar will allow O to scamper home.
6) O should, of course, play to minimize blots as well. This is
even more important than minimizing shots, since with X's 5 point
open O can take a redouble with one man on the bar, but would have
to drop if a second blot were exposed.
The problem with 21/15 and 20/14 is that the 15 and 14 points are not
critical for O since they are easy to bypass. But if X establishes the
17 point it will be a pain in the neck for O.
Brian
Stephen Hubbard
"Brian Sheppard" <bri...@mstone.com> wrote:
<snip>
>My copy of JF (version 2) plays bar/24, 23/17 at every level,
>including rollouts. Perhaps you are using a different version?
I am using the same version. I must have either made a mistake setting
up the position or interpreting the equities. Thanks for pointing out
my error.
<snip>
--
Stephen Hubbard
the...@mindspring.com
Stephen Hubbard
"Brian Sheppard" <bri...@mstone.com> wrote:
<snip>
>My copy of JF (version 2) plays bar/24, 23/17 at every level,
>including rollouts. Perhaps you are using a different version?
No,I am using version 2. I either made a mistake setting up the
position or in interpreting the equities produced. Thanks for pointing
Chuck Bower
In article <01bc6462$97109430$3ac032cf@polaris>,
Brian Sheppard <bri...@mstone.com> wrote:
>Stephen Hubbard <the...@mindspring.com> wrote in article
><5ll1ib$9...@camel1.mindspring.com>...
>> stu...@ix.netcom.com (Stuart Katz, MD) wrote:
>> > +24-23-22-21-20-19-+---+18-17-16-15-14-13-+
>> > | X O X X O | X | O O O |
>> > | X O X X O | | O O O |
>> > | O | | |
>> > | O | | |
>> > | | | |[64]
>> > | | | O |
>> > | X X X X | | O |
>> > | X X X X | | O |
>> > +-1--2--3--4--5--6-+---+-7--8--9-10-11-12-+
>> >money game; cube centered; X has one on bar; X to play 61
(snip)
Brian Sheppard opined:
>Breaking the 23 point is definitely best.
>
Good to see you are so sure of this... CRB
(snip a lot of qualitative justification for this last declaration.)
Brian continues:
>The biggest danger
(for X, after playing Brian's suggested play--CRB)
>is O scrambling home while X is on the
>bar, so X never gets more than a few indirect shots. But X is as
>likely to suffer that fate if he breaks another anchor.
I disagree--see my later ramblings... CRB
>
>This decision is supported by rollouts. I won't give exact numbers,
>because I believe JF doesn't play these positions very well.
(snip)
>The problem with 21/15 and 20/14 is that the 15 and 14 points are not
>critical for O since they are easy to bypass. But if X establishes the
>17 point it will be a pain in the neck for O.
Brian's explanation here seems to assume that the primary value of
moving into O's outfield is to make a point there. How about just getting
over to X's side to make the 5-point?
This is a very interesting AND TIMELY problem, the latter because
of Nack Ballard's BG by the Bay annotation of a recent match (Karr-McIntosh)
and his criticisms of Jellyfish rollouts. Also timely (but a paradox?) is
Brian Sheppard's glowing praise for Nack's opinions (of JF rollouts). It
seems to me that if you want credibility, you shouldn't try to jump on both
bandwagons. Either you trust JF rollouts of backgame-ish positions, or
you don't. If you don't trust them, then why quote rollout results of such
positions. I repeat some of Brian's words (from above):
"This decision is supported by rollouts. I won't give exact numbers,
because I believe JF doesn't play these positions very well."
In other words (mine), since JF agrees with me, I'll pass along the info
(qualitatively) to bolster my argument. If JF would have disagreed, I
would quote Ballard.
Readers of this newsgroup know that I am a strong believer in JF
rollouts. There seems to be much expert opinion that JF version 2 misplays
SOME positions, particularly backgame-ish positions, and especially when
rolling a prime. I believe that version 3 (released to some lucky recipients
in beta version) has been reworked to address these deficiencies. Time will
tell how well Fredrik has done in this regard. I have my check made out,
awaiting the release of Jellyfish 3.0 to the public.
Another interesting fact about this position relates to an old question:
"what is the ranking of (two point) backgames in terms of equity." There is
a (smart ass) answer which goes like "whichever two points you have are the
best". The logic here is that you don't usually have a choice. LOW AND
BEHOLD, here you do. You can play 2-4, 2-5, or 4-5 for sure. Take your
pick.
I don't pretend to know which play is best with the 6. There are
arguments for any of the three. My gut feeling is to leave the checkers
on the 23 point alone. The main reason is that this point gives you the
best chances to hit late in the game (i.e. in full backgame mode). A
second reason is that by playing 23/17, you are slotting the point which
O WANT'S TO MAKE MOST! This is usually not a good move (but could be
right here because of other overriding factors). In retrospect, you can
argue that any of the three plays slots "the point O wants to make most".
21/15 slots the 10-point (desirable) and 20/14 "slots" the 5-point (VERY
DESIRABLE).
Brian argues that 23/17 is the best play from a timing standpoint.
I agree that timing is of supreme importance in all backgames. It's not
obvious to me that X has major timing problems here. X seems to have
lots of pips to play, one key is not to get blocked. I certainly could
be way off base here--O doesn't need a full prime in order to give X
headaches = forcing the breaking of X's home board.
Still having my 70's BG upbringing knawing in the back of my mind,
I like 20/14. It leaves you with the best(?) backgame position available
here (between 2-4, 2-5, and 4-5). In "Advanced BG", I recall that
Robertie likes the 2-4 because the 4-point (21-point from X's view) is
hard to prime. O will have to make his/her 5-point, 8-point, and 10-point
for the full prime of the 21-point. It's certainly possible, but as
Brian points out, X's board makes leaving multiple blots dangerous.
How much is O willing to risk in order make these points?
Robertie also talks about the 2-5 backgame as being underrated (at
the time of writing--circa 1990). I have had suprising success with
this two point combination, but must admit that I HAVE NOT compiled
sufficient statistics. One thing for sure will be VERY HARD to prime
the checkers on the 20-point.
Brian makes a very good point (which also is true of all backgames)
that staying on the bar while your opponent clears his high points
(especially points six, five, and four pips away from your lowest
anchor) can cost a lot of equity. To me that argues strongly for
keeping the 23-point. O has a lot of earth-moving to do to get his/her
checkers into the home board and start the bearoff. If the 4-5 backgame
turns out to be X's choice, then staying afloat would be crushing.
One point I haven't seen made by the proponents of the 23/17 play
is that if this checker is subsequently hit, X still has chances to
make either the 24-point or to remake the 23-point and go from there.
My feeling is to be more direct. If you think you want the deep anchor
(23-point) THEN KEEP IT NOW.
The most bothersome feature of X's game (IMHO) is the missing tooth
in X's home board (5-point). My plan is to get two checkers around
and make this ASAP. This feature may also be the weakness of my approach
(keeping the 23-point). Maybe X's home board argues against playing a
pure backgame.
In the light of recent strides in computer game-playing, it is fun
to speculate on the present and future of backgammon. The best neural
net BG robots (TD-Gammon, Jellyfish, Loner, Snowwhite,...) MAY ALREADY
BE the world's best players. I haven't seen any quantitative evidence
to the contrary, but the burden of proof in these situations is always
on the challenger--in this case the robots. Is the evidence out there,
one way or the other? (I know about FIBS. Just because the neural nets
beat up on "average" players doesn't mean they're the best. How do they
perform against humans with ratings above 1800? 1850? 1900?)
In analyzing positions, the world has steamrolled into using Jellyfish
almost exclusively. Even early Jellyfish resistors (for example, Bill
Robertie) have seen the value of JF rollouts. The consensus seems to be
that with the exception of the above mentioned complicated positions, JF
can be trusted. This is the last interesting point of Stu's above position.
Time is running out if you want to be smarter than the robots. Here is
a backgame--JF can't play backgames. Humans, this is your call to come
out of the woodwork and convince the world that the song "In the year 2525"
doesn't apply to backgammon analysis. Take your shots before it's too late.
Ramblingly yours,
Chuck
bo...@bigbang.astro.indiana.edu
c_ray on FIBS
Brian Sheppard
Chuck Bower <bo...@bigbang.astro.indiana.edu> wrote in article
<5m05fr$g3$1...@dismay.ucs.indiana.edu>...
> In article <01bc6462$97109430$3ac032cf@polaris>,
> Brian Sheppard <bri...@mstone.com> wrote:
> >Stephen Hubbard <the...@mindspring.com> wrote in article
> ><5ll1ib$9...@camel1.mindspring.com>...
> >> stu...@ix.netcom.com (Stuart Katz, MD) wrote:
> >> > +24-23-22-21-20-19-+---+18-17-16-15-14-13-+
> >> > | X O X X O | X | O O O |
> >> > | X O X X O | | O O O |
> >> > | O | | |
> >> > | O | | |
> >> > | | | |[64]
> >> > | | | O |
> >> > | X X X X | | O |
> >> > | X X X X | | O |
> >> > +-1--2--3--4--5--6-+---+-7--8--9-10-11-12-+
> >> >money game; cube centered; X has one on bar; X to play 61
>
> of Nack Ballard's BG by the Bay annotation of a recent match
(Karr-McIntosh)
> and his criticisms of Jellyfish rollouts. Also timely (but a paradox?)
is
> Brian Sheppard's glowing praise for Nack's opinions (of JF rollouts). It
> seems to me that if you want credibility, you shouldn't try to jump on
both
> bandwagons. Either you trust JF rollouts of backgame-ish positions, or
> you don't. If you don't trust them, then why quote rollout results of
such
> positions. I repeat some of Brian's words (from above):
>
> "This decision is supported by rollouts. I won't give exact numbers,
> because I believe JF doesn't play these positions very well."
>
> In other words (mine), since JF agrees with me, I'll pass along the info
> (qualitatively) to bolster my argument. If JF would have disagreed, I
> would quote Ballard.
Hang on there, Chuck. Nick and JF are both great players. Each criticizes
the other. Must I always agree with one and disagree with the other?
In this case I have done what I believe to be a reasonable thing. I have
done the rollouts, I have seen that the rollouts support the move that
my intuition says is better. I have reported that in my post.
On the other hand, I did not post any specific numbers, since I have
no reason to believe that JF is playing well, and I do not want to
be in the position of endorsing numbers that may not be reliable.
> Readers of this newsgroup know that I am a strong believer in JF
> rollouts. There seems to be much expert opinion that JF version 2
misplays
> SOME positions, particularly backgame-ish positions, and especially when
> rolling a prime. I believe that version 3 (released to some lucky
recipients
> in beta version) has been reworked to address these deficiencies. Time
will
> tell how well Fredrik has done in this regard. I have my check made out,
> awaiting the release of Jellyfish 3.0 to the public.
I agree 100% with everything you say here.
> Another interesting fact about this position relates to an old
question:
> "what is the ranking of (two point) backgames in terms of equity." There
is
> a (smart ass) answer which goes like "whichever two points you have are
the
> best". The logic here is that you don't usually have a choice. LOW AND
> BEHOLD, here you do. You can play 2-4, 2-5, or 4-5 for sure. Take your
> pick.
My belief is that 4+5 are best here, because X does not have the timing
for the deep backgame.
> I don't pretend to know which play is best with the 6. There are
> arguments for any of the three. My gut feeling is to leave the checkers
> on the 23 point alone. The main reason is that this point gives you the
> best chances to hit late in the game (i.e. in full backgame mode).
In this position I prefer to provoke action immediately. I do not think
X is likely to win by an immediate exchange of hits, because of the open
5 point, but any exchange of hits will help X a lot.
> A second reason is that by playing 23/17, you are slotting the point
which
> O WANT'S TO MAKE MOST! This is usually not a good move (but could be
> right here because of other overriding factors). In retrospect, you can
> argue that any of the three plays slots "the point O wants to make most".
> 21/15 slots the 10-point (desirable) and 20/14 "slots" the 5-point (VERY
> DESIRABLE).
I don't see anything compelling here.
The real risk in breaking the 23-point is that X will be attacked on
the 23 and 24 points and stay on the bar while O romps home, never
giving X a shot.
> Still having my 70's BG upbringing knawing in the back of my mind,
> I like 20/14. It leaves you with the best(?) backgame position available
> here (between 2-4, 2-5, and 4-5). In "Advanced BG", I recall that
> Robertie likes the 2-4 because the 4-point (21-point from X's view) is
> hard to prime. O will have to make his/her 5-point, 8-point, and
10-point
> for the full prime of the 21-point. It's certainly possible, but as
> Brian points out, X's board makes leaving multiple blots dangerous.
> How much is O willing to risk in order make these points?
Breaking the 20 point is reasonable, but it does leave X vulnerable to
a priming risk.
More to the point: look at the timing battle. O needs to move 40 pips
to bring his outfield men home. How will X hold onto his front
and back positions for 5 or more rolls?
> Robertie also talks about the 2-5 backgame as being underrated (at
> the time of writing--circa 1990). I have had suprising success with
> this two point combination, but must admit that I HAVE NOT compiled
> sufficient statistics. One thing for sure will be VERY HARD to prime
> the checkers on the 20-point.
True.
> Brian makes a very good point (which also is true of all backgames)
> that staying on the bar while your opponent clears his high points
> (especially points six, five, and four pips away from your lowest
> anchor) can cost a lot of equity. To me that argues strongly for
> keeping the 23-point. O has a lot of earth-moving to do to get his/her
> checkers into the home board and start the bearoff. If the 4-5 backgame
> turns out to be X's choice, then staying afloat would be crushing.
True.
> One point I haven't seen made by the proponents of the 23/17 play
> is that if this checker is subsequently hit, X still has chances to
> make either the 24-point or to remake the 23-point and go from there.
> My feeling is to be more direct. If you think you want the deep anchor
> (23-point) THEN KEEP IT NOW.
Re-making the deep anchor is a possibility, but it is not the main
equity. Note that if X remakes a deep anchor after being hit it amounts
to an improvement of at least one roll of timing, so there is definite
benefit to remaing it later.
Still, this isn't the point. The point is that the 4+5 anchors offer
the best combined chances of avoiding a prime, getting a shot, and
escaping.
> The most bothersome feature of X's game (IMHO) is the missing tooth
> in X's home board (5-point). My plan is to get two checkers around
> and make this ASAP. This feature may also be the weakness of my approach
> (keeping the 23-point). Maybe X's home board argues against playing a
> pure backgame.
I agree.
> In the light of recent strides in computer game-playing, it is fun
> to speculate on the present and future of backgammon. The best neural
> net BG robots (TD-Gammon, Jellyfish, Loner, Snowwhite,...) MAY ALREADY
> BE the world's best players. I haven't seen any quantitative evidence
> to the contrary, but the burden of proof in these situations is always
> on the challenger--in this case the robots. Is the evidence out there,
> one way or the other? (I know about FIBS. Just because the neural nets
> beat up on "average" players doesn't mean they're the best. How do they
> perform against humans with ratings above 1800? 1850? 1900?)
It bothers me that JF reigns supreme on FIBS. Its rating is a lot higher
that I would guess, given that I play even against JF Level 6 in money
play, and hold a clear edge over Level 5. I don't think my play is world
class. So how does it do so well on FIBS?
> In analyzing positions, the world has steamrolled into using
Jellyfish
> almost exclusively. Even early Jellyfish resistors (for example, Bill
> Robertie) have seen the value of JF rollouts. The consensus seems to be
> that with the exception of the above mentioned complicated positions, JF
> can be trusted.
Bill Robertie invited me to write a feature article for Inside Backgammon
concerning the construction of neural networks. The article highlights
the techniques used to train neural networks, by showing the theoretical
basis of training, and then showing how this leads to characteristic
effects in the program's play. The research for that article involved
literally thousands of positions from over a thousand games against JF
Level 5 (which, if you recall, uses the neural network with no lookahead).
That article shows misplays by JF in very ordinary positions. Often nothing
complicated is going on. Sometimes it is clear why the network is
misplaying,
e.g. the well-known outfield prime defect. But other times I have no idea.
In short, I do not believe that backgames are the only defect in JF's play.
Even if JF played backgames perfectly, it would still have lots of holes
in its understanding.
> Time is running out if you want to be smarter than the robots.
My belief is that neural networks are more-or-less guaranteed to have
problems, simply due to the statistical variation in their training
regimen. Neural nets have only a "weak theory" of how games are won.
(In backgammon, a "strong theory" would be a probability-weighted
search tree of outcomes.) Their ability consists of interpolating the
given position within a large set of similar positions. But network
architecture does not endow neural networks with the ability to
recall specific positions, nor to recall specific evaluations, nor to
extrapolate to positions that have not been seen before.
Slapping a neural network into a backgammon player was revolutionary.
From here on, things will change at an "evolutionary" pace. (I.e., slowly.)
It is very hard to improve a neural network once it "plateaus."
The biggest improvement may be in faster computers and clever
search algorithms that allow us to search deeper, or even do rollouts
or approximate rollouts in real-time.
I agree that programs will develop to the point where they clearly
outplay humans. And even to the point where they play ideally.
But it may take longer than you think.
Brian
Chuck Bower
In article <5m05fr$g3$1...@dismay.ucs.indiana.edu>,
Chuck Bower <bo...@bigbang.astro.indiana.edu> wrote:
>In article <01bc6462$97109430$3ac032cf@polaris>,
>Brian Sheppard <bri...@mstone.com> wrote:
>>Stephen Hubbard <the...@mindspring.com> wrote in article
>><5ll1ib$9...@camel1.mindspring.com>...
>>> stu...@ix.netcom.com (Stuart Katz, MD) wrote:
>>> > +24-23-22-21-20-19-+---+18-17-16-15-14-13-+
>>> > | X O X X O | X | O O O |
>>> > | X O X X O | | O O O |
>>> > | O | | |
>>> > | O | | |
>>> > | | | |[64]
>>> > | | | O |
>>> > | X X X X | | O |
>>> > | X X X X | | O |
>>> > +-1--2--3--4--5--6-+---+-7--8--9-10-11-12-+
>>> >money game; cube centered; X has one on bar; X to play 61
>
> (snip)
Since making this post, I have done some studying of this VERY
INTERESTING position. I will use this (old) post as a benchmark
for further comments.
Since the advent of strong computer players with rollout capability,
a former learning tool--hand rollouts--has been relegated to the "garage
sales". Some have lamented this development. Computers give excellent
statistical reliability (when properly handled and interpreted) and most
of the time use good technique which subsequently produces high quality
results. But the amount learned by the questioner (us humans) is often
limited.
Motivated partially by the outstanding question "does JF play backgames
correctly?" I decided to do some hand rollouts of Stuart's position above.
I did 72 (sequential first roll) trials after playing bar/24, 20/14 and 36
trials after bar/24, 21/15. In all cases I had X starting with possession
of the (2-) cube. Jellyfish level-7 took the O side and I took X.
However, I "cheated" by sometimes asking JF for the "correct" play for my
side as well as between roll equities (to help me decide on cube decisions).
Thus it was really JF level-7 vs. SUPER JF level-7 (modesty prevents me from
calling my side "SUPER c-ray").
Now for some old comments and the (new) rebuttal:
>
> Brian argues that 23/17 is the best play from a timing standpoint.
>I agree that timing is of supreme importance in all backgames. It's not
>obvious to me that X has major timing problems here. X seems to have
>lots of pips to play, one key is not to get blocked. I certainly could
>be way off base here--O doesn't need a full prime in order to give X
>headaches = forcing the breaking of X's home board.
>
My equivocation "It's not obvious...that X has major timing problems..."
was correct, but ultimately misleading. X DOES have major timing concerns.
My later insights: "one key is not to get blocked" and "O certainly doesn't
need a full prime" hit the nail squarely on the head. O has a lot of building
power in this position. Almost every roll offers X a handful of board-
breaking throws. I would guess that in 25% of my hand rollouts X's board
actually worsened (usually breaking the 6-point). This threat was considerably
worse after giving up the 20-point.
> Still having my 70's BG upbringing knawing in the back of my mind,
>I like 20/14. It leaves you with the best(?) backgame position available
>here (between 2-4, 2-5, and 4-5). In "Advanced BG", I recall that
>Robertie likes the 2-4 because the 4-point (21-point from X's view) is
>hard to prime. O will have to make his/her 5-point, 8-point, and 10-point
>for the full prime of the 21-point. It's certainly possible, but as
>Brian points out, X's board makes leaving multiple blots dangerous.
>How much is O willing to risk in order make these points?
>
Downright BRAZEN is my final conclusion! With only a two (and later
three) point home board looking at X's imposing four point board, you'd think
a robot would play safe. JF hit loose on its 4-point at nearly every
opportunity. It appears that JF recognizes the weakness of the missing
tooth (X's 5-point).
> Robertie also talks about the 2-5 backgame as being underrated (at
>the time of writing--circa 1990). I have had suprising success with
>this two point combination, but must admit that I HAVE NOT compiled
>sufficient statistics. One thing for sure will be VERY HARD to prime
>the checkers on the 20-point.
>
This prediction held up well. The 2-5 backgame performed respectably
in the rollouts (-0.67 points per game with full cube in play).
>
> One point I haven't seen made by the proponents of the 23/17 play
>is that if this checker is subsequently hit, X still has chances to
>make either the 24-point or to remake the 23-point and go from there.
This came up in the other candidates as well--the 24 point was
sometimes made.
>
> The most bothersome feature of X's game (IMHO) is the missing tooth
>in X's home board (5-point). My plan is to get two checkers around
>and make this ASAP. This feature may also be the weakness of my approach
>(keeping the 23-point). Maybe X's home board argues against playing a
>pure backgame.
>
True, but it also argues against playing the 4-5 holding game.
In most games, until the 5-point is made, X doesn't present a major
threat. IT'S THAT IMPORTANT!
>
> In analyzing positions, the world has steamrolled into using Jellyfish
>almost exclusively. Even early Jellyfish resistors (for example, Bill
>Robertie) have seen the value of JF rollouts. The consensus seems to be
>that with the exception of the above mentioned complicated positions, JF
>can be trusted. This is the last interesting point of Stu's above position.
>Time is running out if you want to be smarter than the robots. Here is
>a backgame--JF can't play backgames. Humans, this is your call to come
>out of the woodwork and convince the world that the song "In the year 2525"
>doesn't apply to backgammon analysis. Take your shots before it's too late.
>
As mentioned by previous posters, JF much prefers 23/17 at all levels
(version 2.01). Another shortcoming of computer rollouts was emphasized in
my study--YOU CAN ONLY FORCE THE BOT TO FOLLOW YOUR GAME PLAN FOR A SINGLE
ROLL! Even in these games where I gave up one of the two advanced anchors,
JF was invariably recommending that I remake it (at the expense of the anchor
on the 23-point). You can lead a jellyfish to water, but you can't make him
drink....
It's impossible to reach a strong conclusion with hand rollouts becuase
of the stochastic factors--just not enough trials for statistical significance.
But my impressions were that giving up the 20-point was inferior to giving
up the 21-point. JF level-6 rollouts after playing 23/17 indicate X's
equity in the -0.7 to -0.75 range (holding the 2-cube) which is certainly
statistically consistent with my -0.67 after playing 21-15. JF likes holding
games. (Donald Kahn opined in a private communication that this position
should be played as a holding game and not as a backgame. JF agrees.)
I suspect that JF plays the 4-5 holding game correctly (more or less) though
I haven't taken the time to inspect it. In summary, my feeling is that 20/14
is inferior, and that 21/15 and 23/17 are the two survivors. Anyone want to
duke it out (mano a mano)?