Modern Backgammon
cubist
can't seem get hold of a _functional_ email adress and I'm seriously
contemplating sending a snailmail to Gammon Press.
Michael Strato
For the book, contact Bill Robertie at Gammon...@msn.com
Regards,
Michael
"cubist" <rind...@hotmail.com> wrote in message
news:f3d0b253.02050...@posting.google.com...
Frank Mazza
Gammon Press
PO Box 294
Arlington
MA 02476
U.S.A.
Tel.: +1 781 641 2091
Fax: +1 301 299 7409
E-mail: robe...@world.std.com
Timely service.
And I highly recommend the book. Bill does an excellent job of
recognizing, naming and explaining those features of bot play that
have so influenced the modern game.
Plus there are lots of diagrams, a whole set of illustrative problems
with solutions, and a full annotation of a 25 point match between Nack
Ballard and Jerry Grandell. What more could you want!
Kudos to Robertie!
Frank (TransientResponse) Mazza
Hank Youngerman
Bill's permission) at www.redtopbg.com.
On Tue, 07 May 2002 17:45:39 GMT, fxmaz...@yahoo.com (Frank Mazza)
wrote:
cubist
I finally managed to reach Robertie and order the book, thanks.
Frank Mazza
I'll check that out. In Bill's book the one problem I had was too few
board diagrams in the match annotation section. Now I can look at
every board position, read Bill's comments and see Snowie's verdict
all at the same time.
Thanks for doing that.
Frank Mazza
Chase
wrote:
>And I highly recommend the book [_Modern Backgammon_]. Bill does an excellent job of
>recognizing, naming and explaining those features of bot play that
>have so influenced the modern game.
>
>Plus there are lots of diagrams, a whole set of illustrative problems
>with solutions, and a full annotation of a 25 point match between Nack
>Ballard and Jerry Grandell. What more could you want!
High on my wish list would be that he included his roll outs for the
positions. A lot can be learned from comparing win and gammon
percentages for the various candidates.
Dean (Chase)
__________________________________________________________
To repsond via email, remove the "REMOVE" from my address.
Craig Campbell
news:3cd9bbd...@news.metrocast.net
> High on my wish list would be that he included his roll outs for the
> positions. A lot can be learned from comparing win and gammon
> percentages for the various candidates.
>
>
>
> Dean (Chase)
> __________________________________________________________
> To repsond via email, remove the "REMOVE" from my address.
Chase:
I don't think BR did rollouts for this great book. If he did
rollouts huge blunders like diagram 2-10 wouldn't have found there way
into print.
BR states that the position after the safe 13/11, 12/6 is only a
minimal cube so take a small risk this turn to get to a clear cube.
In actuality the position after 13/11, 12/6 is about as optimal as a
cube can get so the extra risk taken by BR's play is unjustified.
CTCampbell
--
Posted via Mailgate.ORG Server - http://www.Mailgate.ORG
Craig Campbell
news:e5340e01968c97ac372...@mygate.mailgate.org
Chase:
I am retracting the 'huge blunder' from my previous note. The
position after 13/11, 12/6 being roughly the same as a bar point
holding game and familiar to me was clearly a marginal take making BR's
reasoning weak. This caused me to not look at the other position as
closely as I should have. The world doesn't end after playing 13/9,
11/9(2) and geting hit, so the extra risk taken isn't all that great.
My bad. You always risk looking like a fool when you claim someone
as good as BR made a 'huge blunder'.
CTCampbell
PS
I still think this is a mistake that would have been uncovered by
a rollout.
Chase
<ccamp...@aol.com> wrote:
>"Chase" <dem...@REMOVEhotmail.com> wrote in message
>news:3cd9bbd...@news.metrocast.net
>
>> High on my wish list would be that he included his roll outs for the
>> positions. A lot can be learned from comparing win and gammon
>> percentages for the various candidates.
>>
>> Dean (Chase)
>
> I don't think BR did rollouts for this great book.
I'd be extremely surprised if Robertie didn't do extensive roll outs
on all these positions. One doesn't give examples of how the bots have
changed our thinking without consulting the bots on the positions.
>If he did
>rollouts huge blunders like diagram 2-10 wouldn't have found there way
>into print.
A Snowie 3-ply evaluation lists the recommended play as a blunder. But
a full roll out confirms what Robertie says, moving the suggested play
into the top spot by a comfortable margin (~0.040).
> BR states that the position after the safe 13/11, 12/6 is only a
>minimal cube so take a small risk this turn to get to a clear cube.
>In actuality the position after 13/11, 12/6 is about as optimal as a
>cube can get so the extra risk taken by BR's play is unjustified.
My roll outs indicate that Robertie's analysis is entirely correct.
After 13/11 12/6, followed by an average non-hitting roll (e.g. 6-2),
Black has a small double (~0.050), and White a trivial take (~0.260).
If Black instead plays 13/9 11/9(2), Black's double is huge (he is
borderline too good), and White has a very huge pass (~0.310). This is
all as Robertie reports.
Chase
<ccamp...@aol.com> wrote:
>"Craig Campbell" <ccamp...@aol.com> wrote in message
>news:e5340e01968c97ac372...@mygate.mailgate.org
>
>Chase:
> I am retracting the 'huge blunder' from my previous note. The
>position after 13/11, 12/6 being roughly the same as a bar point
>holding game and familiar to me was clearly a marginal take making BR's
>reasoning weak.
According to my roll outs, the take is very huge. Dropping would be a
double whopper with cheese (0.257). Seems like Robertie's reasoning is
on target here.
> I still think this is a mistake that would have been uncovered by
>a rollout.
Actually, what looks like a very large error on Robertie's part at
3-ply, is very much confirmed by the roll outs. As I said in an
earlier post, I'd be very surprised if Robertie hasn't done extensive
roll outs on all of these positions. I just wish he had included them
in the book. It would save guys like me a lot of time. :)
Craig Campbell
That fact that your rollout results differed from mine caused me to
check into something. The seed used makes a significant difference! I
assumed that variance reduction handled this. I am performing some
tests to try and determine how many iterations need to be ran to get
the rollout results to converge when different seeds are used. I think
this would provide a minimum number of iterations for a rollout to be
accurate. Any thoughts?
Chase
<ccamp...@aol.com> wrote:
>Chase:
>
> That fact that your rollout results differed from mine caused me to
>check into something. The seed used makes a significant difference! I
>assumed that variance reduction handled this. I am performing some
>tests to try and determine how many iterations need to be ran to get
>the rollout results to converge when different seeds are used. I think
>this would provide a minimum number of iterations for a rollout to be
>accurate. Any thoughts?
>
>CTCampbell
I don't know a lot about statistics, so I don't know how many trials
would be necessary for our results to converge. Maybe some of the math
people can answer this question.
What settings did you use for your rollouts? I used full, 324 games,
2-ply (medium), seed 1, with racing database. I hope to eventually
roll all the positions out at 3-ply, 100%, huge, but that will take
some time.
Craig Campbell
> would be necessary for our results to converge. Maybe some of the math
> people can answer this question.
>
> What settings did you use for your rollouts? I used full, 324 games,
> 2-ply (medium), seed 1, with racing database. I hope to eventually
> roll all the positions out at 3-ply, 100%, huge, but that will take
> some time.
>
>
> Dean (Chase)
> __________________________________________________________
> To repsond via email, remove the "REMOVE" from my address.
Dean:
I am using the max settings (3-ply, huge, 100%, cubeful) and I
originally did 441 iterations truncated at 14 ply. I am currently
running 2 rollouts in parallel using different seeds to see if and when
they converge. After 150 iterations the average cubeless equity for the
safe play is .492 and both runs seem to be very stable (not changing
much per iteration). With the limited cube equity that White gets in
this position that is a pretty good cube but a little less than an
absolutely optimal cube.
I will post how many iterations it takes to get them to sync up.
Chase
<ccamp...@aol.com> wrote:
> I am using the max settings (3-ply, huge, 100%, cubeful) and I
>originally did 441 iterations truncated at 14 ply. I am currently
>running 2 rollouts in parallel using different seeds to see if and when
>they converge. After 150 iterations the average cubeless equity for the
>safe play is .492 and both runs seem to be very stable (not changing
>much per iteration). With the limited cube equity that White gets in
>this position that is a pretty good cube but a little less than an
>absolutely optimal cube.
> I will post how many iterations it takes to get them to sync up.
>
>CTCampbell
I am not so trusting of cubeful rollouts in positions where cubeless
rollouts suggest that Snowie may be misevaluating things at 3-ply. I'm
not all that knowledgeable about Snowie (or rollouts in general), but
I read in Jeremy Bagai's book that Oasya says doing cubeless rollouts,
where Snowie applies its magic formula to adjust for the cube, is
generally more accurate than doing cubeful rollouts.
Craig Campbell
> rollouts suggest that Snowie may be misevaluating things at 3-ply. I'm
> not all that knowledgeable about Snowie (or rollouts in general), but
> I read in Jeremy Bagai's book that Oasya says doing cubeless rollouts,
> where Snowie applies its magic formula to adjust for the cube, is
> generally more accurate than doing cubeful rollouts.
>
> Dean (Chase)
> __________________________________________________________
> To repsond via email, remove the "REMOVE" from my address.
Dean:
I hope that is not the case. I have absolutely 0% confidence in
Snowie's magic formula. I have seen it make rediculous cube adjustments
way to often. I much prefer doing deep ply GNUbg analysis to Snowie
rollouts. This works much better for a position that tends to resolve
itself within 6 ply. I wouldn't use Snowie at all, but I am not
confident in GNUbg's rollouts at this time.
Douglas Zare
Craig Campbell wrote:
> > I am not so trusting of cubeful rollouts in positions where cubeless
> > rollouts suggest that Snowie may be misevaluating things at 3-ply. I'm
> > not all that knowledgeable about Snowie (or rollouts in general), but
> > I read in Jeremy Bagai's book that Oasya says doing cubeless rollouts,
> > where Snowie applies its magic formula to adjust for the cube, is
> > generally more accurate than doing cubeful rollouts.
>
> I hope that is not the case. I have absolutely 0% confidence in
> Snowie's magic formula. I have seen it make rediculous cube adjustments
> way to often.
Could you give examples? I have to concur that the cube adjustment is much
better than a cubeful rollout when the accuracy of the absolute evaluations
is in doubt.
> I much prefer doing deep ply GNUbg analysis to Snowie
> rollouts. This works much better for a position that tends to resolve
> itself within 6 ply. I wouldn't use Snowie at all, but I am not
> confident in GNUbg's rollouts at this time.
Well, aren't deep-ply evaluations just truncated rollouts with good variance
reduction? If you don't trust the static evaluations in a few turns, why
trust the 6-ply evaluations?
Douglas Zare
Craig Campbell
> reduction? If you don't trust the static evaluations in a few turns, why
> trust the 6-ply evaluations?
>
> Douglas Zare
Douglas:
No, deep-ply evaluations are not just truncated rollouts with good
variance reduction. IMHO the neural nets have two primary areas of
deficiency in their chequer play.
The first is the problem of 'all' AI methods that aren't
supplemented by brute force and that is they add 2 + 2 and get about
4. What I mean by this is that they always estimate the answer, even
when it is calculable. This is not usually a big problem equity wise.
The second area concerns timing issues. The nets are apt to misplay
positions where timing is a substantial factor, and hence botch the
evaluation. A misevaluation in this type of position can be quite
substantial.
A 6-ply search goes a long way to aleviating these problems. The
calculable problems just get counted. The timing problems become
apparent after 3 rolls for each side.
Craig Campbell
The rollout results that I received seem reasonable to me.
Position after 13/11, 11/9(3) .865 cubeful
Position after 13/11, 12/6 .836 cubeful (actually .418 with the
opponent holding a 2 cube)
After 13/11, 12/6 I would estimate that you get hit about 10/36
losing virtually all of these games. The results suggest 11/36. I can
live with that. I guess what bothered me the most was the statement
that this isn't much of a cube. With the cube still centered all you
lose is your ability to double your opponent out once you reach a
position where you would rather have doubled them in. Kind of like a
last roll double.
PS
I ended the parallel rollouts at 700. They differed at that point
by .002 cubeful equity. I would guess that a 1000 iterations should
do the trick. But that is based on a single test, very limited value.
Chase
<ccamp...@aol.com> wrote:
> The rollout results that I received seem reasonable to me.
>Position after 13/11, 11/9(3) .865 cubeful
>Position after 13/11, 12/6 .836 cubeful (actually .418 with the
>opponent holding a 2 cube)
> After 13/11, 12/6 I would estimate that you get hit about 10/36
>losing virtually all of these games. The results suggest 11/36. I can
>live with that. I guess what bothered me the most was the statement
>that this isn't much of a cube. With the cube still centered all you
>lose is your ability to double your opponent out once you reach a
>position where you would rather have doubled them in. Kind of like a
>last roll double.
I'm in the middle of doing some full, 3-ply, 100%, huge rollouts on
this position. So far I'm getting 0.839 for 13/11 11/9(3) and 0.803
for 13/11 12/6. This looks fairly close to what you are getting. After
13/11 12/6, Snowie has White winning a little over 27%, which looks
like a very easy take. After 13/11 11/9(3), White wins a little over
22%, a much closer decision. I'll post the full results when I have
them.
Douglas Zare
Craig Campbell wrote:
> > Well, aren't deep-ply evaluations just truncated rollouts with good variance
> > reduction? If you don't trust the static evaluations in a few turns, why
> > trust the 6-ply evaluations?
>
> No, deep-ply evaluations are not just truncated rollouts with good
> variance reduction.
Do tell, what are they then? Specify exactly. Can you find positions (let's say
away from the database) in which Snowie's 3-ply evaluation is not what is
suggested by the preset "4-ply," "5-ply," and "6-ply" truncated rollouts?
There are other theoretical possibilities than truncated rollouts. However, the
natural ones take incredibly long amounts of time to implement in backgammon. Do
you claim that this is what gnu uses?
> IMHO the neural nets have two primary areas of
> deficiency in their chequer play.
> The first is the problem of 'all' AI methods that aren't
> supplemented by brute force and that is they add 2 + 2 and get about
> 4. What I mean by this is that they always estimate the answer, even
> when it is calculable. This is not usually a big problem equity wise.
Can you give any position in which this is a real problem? If not, why call it a
primary area of deficiency? I don't see this as an issue in backgammon.
> The second area concerns timing issues. The nets are apt to misplay
> positions where timing is a substantial factor, and hence botch the
> evaluation. A misevaluation in this type of position can be quite
> substantial.
> A 6-ply search goes a long way to aleviating these problems. The
> calculable problems just get counted. The timing problems become
> apparent after 3 rolls for each side.
I think you have greatly overestimated the value of looking ahead 3 rolls for each
side. You don't know that the bot is playing correctly in the middle. You don't
know that the errors are unbiased at the end. Three exchanges is not long enough.
(Three exchanges is the default for Snowie minirollouts, by the way.)
It's easy to call almost everything that the bots do wrong a timing problem.
However, I think that is an oversimplification of the problem. There are plenty of
fundamental defects in the bots' evaluations, some of which are corrected by
rollouts, and some of which are not.
Douglas Zare
Chase
cubeless, in both win and gammon percentages, and looks like to yields
a more efficient cube.
Money session. Score X-O: 0-0
X to play (2 2)
+24-23-22-21-20-19-------18-17-16-15-14-13-+
| O O O O | | O O X |
| O O O | | O |
| O | | | S
| | | | n
| | | | o
| |BAR| | w
| | | | i
| O X | | | e
| X O X | | |
| X X X O X | | X |
| X X X O X | | X X |
+-1--2--3--4--5--6--------7--8--9-10-11-12-+
Pipcount X: 93 O: 141 X-O: 0-0/Money (1)
CubeValue: 1
1. R 13/11 11/9(3) Eq.: 0.839
0.1% 8.4% 77.8% 22.2% 2.0% 0.0%
95% confidence interval:
- money cubeless eq.: 0.620 ą0.015.
Rollout settings:
Full rollout,
504 games (equiv. 18976 games),
played 3-ply (huge, 100%),
random seed, with race database.
2. R 13/11 12/6 Eq.: 0.795 (-0.044)
0.0% 6.6% 72.7% 27.3% 2.0% 0.0%
95% confidence interval:
- money cubeless eq.: 0.501 ą0.013.
Rollout settings:
Full rollout,
504 games (equiv. 25228 games),
played 3-ply (huge, 100%),
random seed, with race database.