Opening Rolls Rollouts JellyFish Ver 3.0
Unknown
For those of you interested I've rolled out all the opening moves top
contenders. Trials 1296 Standard Deviations 0.008. I'm not sure of the
statistical relevance of the data with this S.D. but they,re here for
your perusal.A few points I noticed though were slotting ones own 5
point with 41 is definitely an inferior choice, for 21 its definitely
second place but for 51 seems like a good choice given the general
accepted opinion is that JellyFish doesnt "steer" its games towards
backgames.
Spliting to the opponents 5 point with 43 and 54 produced better
results in the rollouts.
The inner board point making numbers 31 42 53 64 all produced equities
higher than the evaluation.
Top moves marked with *
Dice Move Move L7 EV L6 Roll Dice Move Move
L7 EV L6 Roll
21* 13/11 24/23 -.005 +.017 62 24/18 13/11
+.010 +.006
21 13/11 6/5 -.007 -.008 62* 24/16
-.003 +.008
21 24/21 -.018 -.008 62 13/5
-.020 -.015
21 13/10 -.026 -.027 62 24/18 24/22
-.041 -.042
21 24/22 24/23 -.035 -.056
43 13/9 24/21
+.010 -.003
31* 8/5 6/5 +.140 +.155 43 13/9 13/10
+.005 -.004
43* 24/20 13/10
+.001 +.009
41* 13/9 24/23 -.010 -.006 43 24/20 24/21
-.007 +.003
41 13/8 -.022 -.037
41 24/20 24/23 -.026 -.023 53* 8/3 6/3
+.052 +.077
41 13/9 6/5 -.032 -.031 53 13/8 24/21
+.029 +.024
41 24/20 6/5 -.048 -.066 53 13/8 13/10
+.008 -.003
51* 13/8 24/23 +.008 +.004 63* 24/18 13/10
+.002 .000
51 13/8 6/5 -.014 -.004 63* 24/15
+.002 .000
51 24/18 -.025 -031 63 24/18 24/21
-.031 -.047
63 13/4
-.039 -.037
61* 13/7 8/7 +.098 +.107
54* 13/8 24/20
+.025 +.034
32* 24/21 13/11 +.002 +.011 54 13/8 13/9
+.022 +.015
32 13/10 13/11 -.005 +.004 54 24/15
+.002 .000
32 13/10 24/22 -.007 -.008
32 13/8 -.022 -.037 64 24/14
+.014 +.007
32 24/21 24/22 -.027 -.027 64 24/18 13/9
+.011 +.008
32 13/10 6/4 -.049 -.038 64* 8/2 6/2
+.005 +.011
42* 8/4 6/4 +.095 +.122 65* 24/13
+.079 +.066
52 13/8 24/22 +.010 +.001
52* 13/8 13/11 +.008 +.007
52 13/8 6/4 -.030 -.041
52 13/6 -.040 -.054
achim_mueller
In article <3427e1f1...@news.demon.co.uk>, says...
>
>For those of you interested I've rolled out all the opening moves top
>contenders. Trials 1296 Standard Deviations 0.008. I'm not sure of the
I've done the same on level 5 (truncated rollouts, 7 moves), 12960 trials each.
Similar results. For those who are interested (there are a few more rollouts
there and this database shall grow...) go to
home.deltacity.net/~acepoint/index.htm. Please note, that it's still under
construction and that my english is far from perfect :-)
always good dice
amue
Chuck Bower
In article <3427e1f1...@news.demon.co.uk>, Midas <> wrote:
>For those of you interested I've rolled out all the opening moves top
>contenders. Trials 1296 Standard Deviations 0.008. I'm not sure of the
>statistical relevance of the data...
I can shed some light on this. Take the difference between two
plays (of the same roll!) and compare to the following table:
difference confidence level
0.014 90%
0.017 95%
0.026 99%
This is based on the 0.008 standard deviation quoted. What "confidence
level" really means is:
If you could do a VERY LARGE NUMBER of rollouts ("VERY LARGE NUMBER"
means the biggest number you can think of), what are the chances that
the order of the two plays will remain the same?
If the difference between two plays is less than 0.015, then statistically
you can call them equal. (0.015 is not a HARD number. Some might say
0.02 and others 0.01. But at some point you have to draw the line
statistically, and the range 0.01 to 0.02 for Midas's table should satisfy
most statisticians, I suspect.)
The other question is: "How well is JF playing these positions?"
Given that JFv3.0 level-6 is playing its twin, it is fair to say that
the two players are equal! But for any given position, the "defender"
may have an easier position than the "offender", or vice versa, so the
rollout results could have a "systematic error". These are difficult
to estimate. As a wild guess, I'd say 0.01 for these opening roll
positions. All the more reason to consider two plays "equal" if they
are within a certain distance of each other, that distance being around
0.02, give or take a bit, IMHO.
Chuck
bo...@bigbang.astro.indiana.edu
c_ray on FIBS
Gary Wong
(Midas) writes:
> For those of you interested I've rolled out all the opening moves top
> contenders. Trials 1296 Standard Deviations 0.008. I'm not sure of the
> your perusal.
Thanks for that article. One thing surprises me a bit though, in the
results for 62:
> Top moves marked with *
>
> Dice Move Move L7 EV L6 Roll
> 62 24/18 13/11 +.010 +.006
> 62* 24/16 -.003 +.008
> 62 13/5 -.020 -.015
> 62 24/18 24/22 -.041 -.042
Why on earth is JF so eager to play 24/16 (and similar results running to
the outfield on 63 and 64)? It even ranks it better than 24/18 13/11
(though admittedly the distinction is lost in the statistical noise).
The way I see it:
Advantages of 24/18 13/11:
* 24/18 leaves the back men in communication (16 return shots on the
bar point; only 4 on the 9 point).
* Even if the blot on the 9 point is missed after 24/16, you'll be scrambling
to safety it next roll, reducing your mobility.
* 24/18 leaves good chances of anchoring on the opponent's bar point;
24/16 gives up anchoring for now.
* 24/18 13/11 gives a spare builder for the 5 point; 24/16 doesn't put a
man anywhere useful (what's the use of being on your opponent's 9 point?)
* 24/18 gives better outfield coverage; it protects against the opponent
bringing builders to his 9, 10 or 11 points (24/16 only protects the
10 and 11 points)
Advantage of 24/16:
* Only gives your opponent 14 shots at your blot on his 9 point (2 of his
rolls make the 9 point), instead of 26 shots on the bar point (6 make it)
Are the reduced shots on the runner really worth giving up all the advantages
of 24/18 13/11? Let's see what the outcomes of the next two rolls are:
24/16 24/18
Blot hit, return hit (best) 48 320
Blot missed (good) 792 360
Blot hit, no return hit (bad) 384 400
Blot pointed on (worst) 72 216
---- ----
1296 1296
Hard to say whether the chance of hitting back with 24/18 is worth the risk
of him making his bar point on your head. There's no doubt that 24/18
increases the complexity of the game, which must make it a better choice
against a weaker player. I'd tend to play 24/18 anyway, and it puzzles me
that Jellyfish favours 24/16 in the rollouts. Does it see something there I'm
missing? What do other humans think?
Cheers,
Gary (GaryW on FIBS).
--
Gary Wong, Computer Science Department, University of Auckland, New Zealand
ga...@cs.auckland.ac.nz http://www.cs.auckland.ac.nz/~gary/
Chuck Bower
Gary Wong wrote:
It seems that Gary has chosen to analyze the merits of the two plays in
the "old fashioned" method of PRE-robot days. First off, I emphasize that
Midas's JF rollouts say that THESE TWO PLAYS ARE EQUAL. But
Gary's question is still pertinent. "Why doesn't JF feel that 24/18, 13/11 is
BETTER?" (BTW, "old fashioned" should NOT be interpreted as "worse",
just "different", at least for now.)
Even if we agree with Gary's breakdown of advantages vs. disadvantages
(and I don't completely), you also must weigh "how good" or "how bad". If
the bad are rare, but when they happen are REALLY bad, then they could
outweigh the good, even if the good occur more frequently. That is in effect
what JF rollouts do--they not only include the qualitative features (good vs.
bad) but also the quantitative (HOW good vs. HOW bad).
Take Gary's summary table of the next 1296 outcomes. 24/18 has more
chances for "best", but also more chances for "worst". When JF rolls these out,
the "best" and "worst" seem to offset (I'm obviously simplifying), resulting in
EQUAL results regardless of the play chosen.
Another similar (though clearly not identical) roll is 65. You could play
24/18,
13/8 (and certainly some players have experimented with this). It has most (but
not all) of the advantages Gary listed for the 63 24/18 play. You could certainly
make a case for 24/18 with 65, but very few players (including JF) see its merits
over the conventional, unimaginative, beginner play of 24/13.
Are there any human experts who play 24/16 with an opening 62 at $ play? I
don't know of any. There may be some who play it at DMP, though. It does
seem that Gary's argument that 24/18, 13/11 "complicates" things, and thus may
be better for a strong player vs. a weak player. However, consider the following:
Suppose your opp opens 24/16 with 62 and you then roll 31. How do you play it?
JFv3.0 evaluations say the following: level-5 and level-6 prefer 13/9*, but
level-7
makes the 5-point. Which is correct? Heck if I know! But JFv3.0 level-6
rollouts
(864 trials) say that 13/9* is better than 5(2) by 0.026 in cubeless $ equity, and
this
is statitically signifcant at better than 95% confidence. If you make plays that
your
opp. isn't used to (even an expert opp), then even one misplay because of
unfamiliarity could swing the play in your favor.
During the 1994 World Cup Lecture, Joe Sylvester related a story about
playing against an opponent who made unconventional openings, and specifically
mentioned 43 played 24/20, 24/21. Joe then said (paraphrasing) "I was
perplexed. I wasn't sure how to respond to this!" Midas's rollouts show this way
of opening 43 as about equal to the other, more common plays. Now, how do
you respond with 21 to this play? JFv2.0 level-6 rollouts (432 trials) say that
if
you'd BETTER HIT BOTH or you're giving away a chunk of equity (0.07).
Maybe you thought that was "obvious", but often "obvious" proves out to be
"obviously wrong"!
Bottom line is that there is still a lot we (including the bots) don't know
about
this game.
Leo Bueno
On 23 Sep 1997 13:33:17 -0700, Achim Mueller wrote:
>I've done the same on level 5 (truncated rollouts, 7 moves), 12960 trials each.
>Similar results. For those who are interested (there are a few more rollouts
>there and this database shall grow...) go to
>home.deltacity.net/~acepoint/index.htm.
Tried going to the site; it did not load.
==================================================================
E-mail: leob...@accesspro.net
Snail-mail: P.O. Box 440545, Miami, FL 33144-0545 (USA)
Test De Cubanidad: http://www.accesspro.net/leobueno/cubanidad.htm
Castro Fall Poll: http://www.accesspro.net/leobueno/adiosfidel.htm
==================================================================
achim_mueller
In article <342cee75...@snews.zippo.com>, leob...@bridge.net says...
...
>Tried going to the site; it did not load.
>
I'm sorry to hear that, full url is
http://home.deltacity.net/~acepoint/index.htm.
The server is awfully slow sometimes, but they promissed to work on it. Better
luck next time.
Always good dice
amue
Gary Wong
Chuck Bower <bo...@bigbang.astro.indiana.edu> writes:
> Gary Wong wrote:
> > Why on earth is JF so eager to play 24/16 (and similar results running to
> > the outfield on 63 and 64)? It even ranks it better than 24/18 13/11
> > (though admittedly the distinction is lost in the statistical noise).
[snip]
> > Hard to say whether the chance of hitting back with 24/18 is worth the risk
> > of him making his bar point on your head. There's no doubt that 24/18
> > increases the complexity of the game, which must make it a better choice
> > against a weaker player. I'd tend to play 24/18 anyway, and it puzzles me
> > that Jellyfish favours 24/16 in the rollouts. Does it see something there I'm
> > missing? What do other humans think?
[snip]
> It seems that Gary has chosen to analyze the merits of the two plays in
> the "old fashioned" method of PRE-robot days. First off, I emphasize that
> Midas's JF rollouts say that THESE TWO PLAYS ARE EQUAL. But
> Gary's question is still pertinent. "Why doesn't JF feel that 24/18, 13/11 is
> BETTER?" (BTW, "old fashioned" should NOT be interpreted as "worse",
> just "different", at least for now.)
OK, I guess I have to plead guilty to that one (weighing up the positions
under my own opinion, "old fashioned" or whatever else). I certainly agree
the evidence demonstrates that Jellyfish finds statistically no difference
between the two plays, when playing against itself. But I'm not quite ready
to make the leap of faith to infer from that evidence that 24/18, 13/11 isn't
a significantly better move for _me_ to make against another human. One of
several factors could swing the scales in either direction:
* Jellyfish might be misplaying the resultant positions that one play
is more likely to lead to slightly, reducing its estimation of the
equity a tiny amount;
* I might make significantly more errors after one play than the other, in
which case I should either learn a more appropriate play for the subsequent
positions, or select the play leading to a position I am more likely to be
able to play correctly;
* I should also (all other things being equal) seek to complicate or
simplify the position, depending on how well I'm likely to be able to
play the position relative to my opponent. Jellyfish on the other hand,
in training against itself, will never artificially seek or avoid
complication (deciding from the estimated equity only).
* As you point out in your message: I might be able to turn a position to
my advantage by avoiding a `traditional' play if this will lead to a
position my opponent is less familiar with than the ordinary one.
Again, this will (or should) affect the play I make slightly, but not the
one Jellyfish selects.
Essentially I suppose the problem is that the equity for any backgammon
position does NOT just depend on the position of the chequers, the value of
the cube, and the match score -- it ALSO depends on _who_ is playing each
side. Jellyfish rollouts give us very accurate estimates of the equity of a
position being played by Jellyfish on each side; it doesn't necessarily
follow that Jellyfish's equity for a position needs to be anything like _my_
equity for the same position against a particular opponent, or even that the
ordering of equities for various moves are equivalent. It is quite possible
that Jellyfish wins more (or the same) number of games playing 24/16 against
itself than 24/18, but I win more against an average human playing 24/18 than
24/16. And unfortunately it still doesn't answer the question: given
`perfect' subsequent play by both players, is 24/18 13/11 a better move?
> Bottom line is that there is still a lot we (including the bots) don't
> know about this game.
Yep, agreed hands down on that one :-)
Leo Bueno
Went back; here's what I get.
Sorry
The resource requested /~acepoint/index.htm. cannot be
found.
If you feel that this is a configuration error, please contact the
administrators or the author of the referring page.
Roxen·Challenger/1.1.1alpha1, at http://home.deltacity.net/.
Chuck Bower
Kevin Bastian wrote:
> Am I correct in concluding that, based on Midas' rollouts (below), the only
> time I should exercise a post-Crawford free-drop in match play after
> winning the opening roll is if that opening roll is a 4-1? (assuming I play
> the other opening rolls correctly, of course!) The rollout showed the
> equity after its preferred move of 13/9 24/23 as -.006. I understand (sort
> of) the "margin of error" in the results, i.e., if Jellyfish says an
> opening roll is -.001 that might just as well be +.001; they're so close as
> to be statistically equivalent. Or something like that. On the other hand,
> if it evaluates it as -.006, and I have nothing else to go on, why not take
> that as a "sell" signal and drop the double?
>
> OK, all that babble aside, is the only free-drop opener a 4-1?
(snip, including Midas's rollout results)
Be careful. There are a few things which must be considered:
1) Can trailer use gammons? I have yet to see free-drop decisions at
other than 2-away, 1-away p.C. covered rigorously.
2) Keep in mind the uncertainties, both statistical and systematic. For
example,
my JFv3.0 level-6 rollouts (864 trials) came up with 41 played 13/9, 24/23
giving opening roller a slight advantage at $ play (+0.006, s.d. = 0.010).
Thus
if you combine and weight the results by number of rollouts, you get -0.002
with s.d. around 0.006. So the true JFv3.0 level-6 rollout results could
easily
be positive.
3) Relative strength of players could be a factor as well.
Kevin Bastian
Am I correct in concluding that, based on Midas' rollouts (below), the only
time I should exercise a post-Crawford free-drop in match play after
winning the opening roll is if that opening roll is a 4-1? (assuming I play
the other opening rolls correctly, of course!) The rollout showed the
equity after its preferred move of 13/9 24/23 as -.006. I understand (sort
of) the "margin of error" in the results, i.e., if Jellyfish says an
opening roll is -.001 that might just as well be +.001; they're so close as
to be statistically equivalent. Or something like that. On the other hand,
if it evaluates it as -.006, and I have nothing else to go on, why not take
that as a "sell" signal and drop the double?
OK, all that babble aside, is the only free-drop opener a 4-1?
I'll eagerly await the experts' views! :-)
Thanks,
Kevin Bastian
(KevinB on FIBS)
Midas wrote in article <3427e1f1...@news.demon.co.uk>...
> For those of you interested I've rolled out all the opening moves top
> contenders. Trials 1296 Standard Deviations 0.008. I'm not sure of the
> statistical relevance of the data with this S.D. but they,re here for
> your perusal.A few points I noticed though were slotting ones own 5
> point with 41 is definitely an inferior choice, for 21 its definitely
> second place but for 51 seems like a good choice given the general
> accepted opinion is that JellyFish doesnt "steer" its games towards
> backgames.
> Spliting to the opponents 5 point with 43 and 54 produced better
> results in the rollouts.
> The inner board point making numbers 31 42 53 64 all produced equities
> higher than the evaluation.
>
> Top moves marked with *
>
> Dice Move Move L7 EV L6 Roll Dice Move Move
> L7 EV L6 Roll
> 21* 13/11 24/23 -.005 +.017 62 24/18 13/11
> +.010 +.006
> 21 13/11 6/5 -.007 -.008 62* 24/16
> -.003 +.008
> 21 24/21 -.018 -.008 62 13/5
> -.020 -.015
> 21 13/10 -.026 -.027 62 24/18 24/22
> -.041 -.042