Big error of JellyFish
Gilles BAUDRILLARD
I play JellyFish at level 7.I am leading 1-away,4-away and we are in
post-Crawford games.
JF win the opening roll with 2-1 and plays ... 13-11,6-5 !!!
Of course,it can't work :
-If I am unable to hit the blot on JF's 5-pt,I will just use my free
drop and refuse the incoming cube.
-If I hit,I accept the cude and JF will start the game in a very
inferior position : one more man back ( two if I roll 6-4 ) and
my back men having already started to move ...
In some of Kit's annotated matches,I have seen this error
made by top level players,but JF does not have the excuse
to be tired at the end of a long match :-)
Gilles ( 'gillesb' on FIBS )
James Eibisch
>I play JellyFish at level 7.I am leading 1-away,4-away and we are in
>post-Crawford games.
>
>JF win the opening roll with 2-1 and plays ... 13-11,6-5 !!!
>
>Of course,it can't work :
>-If I am unable to hit the blot on JF's 5-pt,I will just use my free
>drop and refuse the incoming cube.
>-If I hit,I accept the cude and JF will start the game in a very
>inferior position : one more man back ( two if I roll 6-4 ) and
>my back men having already started to move ...
It's quite a gammonish move (if one can be gammonish after the first
move) and this is perfect strategy at 4-away. If you miss, would you
always drop? What if you reply with 31, 42, or 66? You're more likely
not to hit, of course.
Perhaps you are, in fact, more likely to do something constructive and
relatively safe if JF opens with 13-11 24-23. I like JF's play at this
score.
--
_ N : E : T : A : D : E : L : I : C : A
James Eibisch ('v') -- http://www.revolver.demon.co.uk --
Reading, U.K. (,_,) -- Now showing: Invaders 1978: --
======= a faithful version of Taito's original
Patti Beadles
In article <57vca3$s...@belzebul.imaginet.fr>,
Gilles BAUDRILLARD <gba...@imaginet.fr> wrote:
>In some of Kit's annotated matches,I have seen this error
>made by top level players,but JF does not have the excuse
>to be tired at the end of a long match :-)
No, but jellyfish also doesn't have a network that understands match
scores and free drops. As far as I know, jellyfish looks at the wins,
gammons, and backgammons for each play, factors in match equities, and
then chooses the best play based on which one will give it the highest
match equity. The play equities are cubeless, though, so there's no
concept of, "If I don't get hit, he automatically drops."
-Patti
--
Patti Beadles | Knowing how to play well doesn't
pat...@netcom.com/pat...@gammon.com | win-- you must actually do it.
http://www.gammon.com/ |
or just yell, "Hey, Patti!" | No longer pa...@velo.com
Brian Sheppard
Patti Beadles <pat...@netcom.com> wrote in article
<pattibE1...@netcom.com>...
> No, but jellyfish also doesn't have a network that understands match
> scores and free drops. As far as I know, jellyfish looks at the wins,
> gammons, and backgammons for each play, factors in match equities, and
> then chooses the best play based on which one will give it the highest
> match equity. The play equities are cubeless, though, so there's no
> concept of, "If I don't get hit, he automatically drops."
>
This is a good start at understanding the problem, but it doesn't
go far enough to clarify my understanding.
Since JellyFish factors in match equities, then knows that at 1-4
its chances are not much different than at 1-3. I also understand
that JellyFish level 7 includes some lookahead capability. If it
sees just a couple of plies it should follow the logic behind the
free-drop calculation.
Patti: can you clarify?
thanks
Brian
Ron Karr
An attempt to clarify:
The question is, what is meant by "factors in match equities"? I
believe that what happens is this:
When Jellyfish looks ahead (on Level 7), it goes through all possible
combinations of the dice for both players, figures out what the best
move is in each case, and comes up with a weighted average of equities.
But the question is: how does it come up with the "best move" for each
of the plays it examines in the lookahead? My understanding is: it
considers the match score only in the sense that it adjusts the equities
based on the relative value of gammons at the CURRENT match score. It
does NOT (as Patti pointed out) take into account that the cube may be
turned after a given play, and that the player doubled may either take
or drop, thereby changing the equity.
So, for example, if Jellyfish rolls a 2-1 at 4 away-1 away, it evaluates
the equity of the slotting play vs the splitting play, assuming that the
match score is (and will REMAIN) 4 away-1 away. It DOES take into
account that gammons count at double their usual value for the trailer,
while they don't count extra at all for the leader. That's probably why
it prefers slotting to its usual play of splitting.
Ron
John Quinnelly
Gilles BAUDRILLARD wrote:
>
> I play JellyFish at level 7.I am leading 1-away,4-away and we are in
> post-Crawford games.
>
> JF win the opening roll with 2-1 and plays ... 13-11,6-5 !!!
>
Why is this play wrong? Is it the opening play itself or is it wrong at this match
point?
John
Kit Woolsey
John Quinnelly (jo...@ntr.net) wrote:
It is wrong at this match situation. The reason is not particularly
obvious, and several top-level players make the same error. Here is what
is going on:
First of all, it is clear that the person who is trailing in the match
will double as soon as he legally can, since he has everything to gain
and nothing to lose. Therefore, this game (and the one after it, if
there is one) will be played with the cube on 2.
Given that the cube will be on 2, there is very little difference between
being ahead 1 away - 4 away and being ahead 1 away - 3 away. In both
cases, the trailer simply will have to win the next two games (or win a
gammon in the first game).
Since there is little difference between these two scores, it should be
clear what the leader's strategy is when he is doubled at the 1 away - 4
away score. If he is the favorite in the game he should take the double,
while if he is the underdog he should pass and play another game at
essentially the same score. The only difference is that he can no longer
afford to pass the double regardless of the position, since this would
put his opponent only 2 away, thus making the following game the last
game of the match (since once again the cube will be turned
immediately). This is called the free drop.
Now, we can analyze the play of the opening 2-1 roll under these
circumstances. If you slot the five point, you are hoping not to be hit
and to make the five point on the next turn. However, you will never get
a chance to make that five point! The reason is that if your opponent
misses the shot (and doesn't roll something super like 3-3 or 6-6
himself) he will be the underdog and will pass when you turn the cube.
If he hits the shot you will be a clear underdog and he will happily take
the double. Consequently, slotting the five point can only lose, never gain.
The above is a very difficult concept to fully understand. If you can
follow the logic involved, you are taking a giant step to becoming a
winning match player.
Kit
Brian Sheppard
Ron Karr <ka...@apple.com> wrote in article <32A3FB...@apple.com>...
> But the question is: how does it come up with the "best move" for each
> of the plays it examines in the lookahead? My understanding is: it
> considers the match score only in the sense that it adjusts the equities
> based on the relative value of gammons at the CURRENT match score. It
> does NOT (as Patti pointed out) take into account that the cube may be
> turned after a given play, and that the player doubled may either take
> or drop, thereby changing the equity.
You believe that the "gammon prices" are frozen at the start of a search?
> So, for example, if Jellyfish rolls a 2-1 at 4 away-1 away, it evaluates
> the equity of the slotting play vs the splitting play, assuming that the
> match score is (and will REMAIN) 4 away-1 away. It DOES take into
> account that gammons count at double their usual value for the trailer,
> while they don't count extra at all for the leader. That's probably why
> it prefers slotting to its usual play of splitting.
I can understand why JF would initially rank the slotting play higher,
since a linear extrapolation shows that gammons are better for us than
for the opponent. But tree search is supposed to overcome such
shortsightedness, not reinforce it!
A normal variation in a three-ply tree search would look like the
following:
Ply 1: 2-1 played by slotting
opponent has the decision about whether to double
Ply 2: opponent plays some number
We decide whether to double (and we do double)
Opponent takes or drops according to his match equities
Ply 3: We roll
You can verify that if JF performs an actual 3-ply search, then it is
guaranteed to get the right answer on this problem, since all the relevant
decisions occur within a 3-ply horizon.
My conclusion is that JellyFish's tree search does not consider the
possibility that one player might double. Do you agree?
Brian
Ron Karr
Brian Sheppard wrote:
>
>
> My conclusion is that JellyFish's tree search does not consider the
> possibility that one player might double. Do you agree?
>
Yes.
This is one area where we humans (theoretically) still retain some edge
over JF. We can consider things like "if I make this play, I will have
an efficient double next time", while JF only takes into consideration
cubeless winning probability (although as you point out, it wouldn't be
hard for a program to include this).
Ron
James Eibisch
On Wed, 4 Dec 1996 08:37:15 GMT, kwoo...@netcom.com (Kit Woolsey)
wrote:
[snip]
>Consequently, slotting the five point can only lose, never gain.
Excellent explanation, thanks Kit. Wish I'd thought of it by myself
though :)
William C. Bitting
Kit Woolsey (kwoo...@netcom.com) wrote:[cuts and more cuts]: : >
: : > I am leading 1-away,4-away and we are in post-Crawford games.
: : > JF wins the opening roll with 2-1 and plays ... 13-11,6-5 !!!
: : > Of course,it can't work :(clipped)
: : Why is this play wrong? Is it the opening play ..or play at this score?
: It is wrong at this match situation. The reason is not particularly
: the double. Consequently, slotting the five point can only lose, never gain.
: The above is a very difficult concept to fully understand. If you can
: follow the logic involved, you are taking a giant step to becoming a
: winning match player.
: Kit
Yes, very difficult for me anyhow! :) I can see that JF will never make
the 5 point. What I don't "quite" understand is why this makes a
difference? While I presume that the correct play, then, is 24-23 13-11
(13-10?), it is not clear to me why this is better at this match score?
An aside: TD_one seemly always plays 6-5 13-11, so I gather at double
match point games this is not an error(?). Would the same then be true
for 2-away 2-away games; i.e. 6-5 13-11 is not incorrect?
Where this reasoning leads me is to the proposition that as trailer at -4
-1 that I don't want to succeed in gaining an immediate advantage
because the leader will drop thereby robbing me of an opportunity to win
the match through achieving a gammon with the cube at 2. I.E., play with
the goal of first having your double accepted so that in the event you
win a gammon, you also win the match.
Obviously, that cannot be very good reasoning, so I give myself an
"unsatisfactory" rating for this effort. Such logic would seem to lead to
the conclusion that as trailer with the opening roll of 3-1, one sould
not play 6-5 8-5 because the leader will drop a double before your
next roll.
Or, is it the risk that the blot on 5 is hit, making you the underdog
that one is avoiding? That would at least accomplish the mission of
having your double accepted. ::)) And if one wants to avoid that at -4
-1 why doesn't TD_one want to avoid it at -1 -1?
So, I see that the 5 point will never be made in the -4 -1 (while it
certainly will be in the -1 -1) situation, I just can't grasp the
relevance of that outcome. I come back to the proposition that at -4 -1
that the trailers first objective is to play the game with the cube at 2.
Well, it will make for a good riddle to go back to from time to time with
the hopes of piercing the vail! william, wcb on FIBS
Ron Karr
William C. Bitting wrote:
>
> Kit Woolsey (kwoo...@netcom.com) wrote:
> : John Quinnelly (jo...@ntr.net) wrote:
> : : Gilles BAUDRILLARD wrote:
> [cuts and more cuts]: : >
>
> : : > I am leading 1-away,4-away and we are in post-Crawford games.
> : : > JF wins the opening roll with 2-1 and plays ... 13-11,6-5 !!!
> : : > Of course,it can't work :(clipped)
>
>
> the 5 point. What I don't "quite" understand is why this makes a
> difference? While I presume that the correct play, then, is 24-23 13-11
> (13-10?), it is not clear to me why this is better at this match score?
> An aside: TD_one seemly always plays 6-5 13-11, so I gather at double
> match point games this is not an error(?). Would the same then be true
> for 2-away 2-away games; i.e. 6-5 13-11 is not incorrect?
>
Having just won a match from an opponent who made the same error, I'll
try another explanation. (BTW, I hit 2 checkers with 4-6, she doubled,
I took as a big favorite. If I had missed, I would have dropped.)
This situation occurs ONLY when the match leader has a free drop. It
does not apply at -1, -1 or -2, -2 or -1, -3, etc. Only at
post-Crawford where the trailer needs an even number.
When my opponent is at Crawford, and I have an odd # of points, I want
to double at my first opportunity, since I have nothing to lose, and I
stand to win 2 points instead of 1 (or 4 points instead of 2, if gammons
are relevant). Obvious, right?
Suppose I roll 2-1 first. I'm not going to get a chance to double until
after my opponent plays. We know that the splitting and slotting plays
with 2-1 are very close. In fact, let's assume for the sake of argument
that they are exactly equal, so that normally it's a tossup which play
to make.
But what IS different about the two plays is the volatility after the
opening move. When I split, nothing really dramatic happens most of the
time. He has a few great numbers (doubles) and some good numbers that
give him a slight edge, after which he takes. A lot of the time he ends
up with a slight disadvantage, so he drops.
When I slot, on the other hand, the volatility is larger. Numbers that
hit the blot on the 5 point give him a big edge. Numbers that don't hit
give me a big edge.
Under normal circumstances, the volatility doesn't matter; only the
equity. But at this match score, the only games that we actually end up
playing are the ones where HE has the edge (because those are the only
ones he takes). If I slot, his average edge is going to be greater in
the games we do play. So it's in my interest to split and try to reduce
the volatility.
And in practice, it's even better than this, because when I split and my
opponent rolls some average number, it's often not clear who has the
edge (for example, would you take after rolling 5-3?). But it's easy to
tell who has the edge after the slotting play.
Ron
Kit Woolsey
: Where this reasoning leads me is to the proposition that as trailer at -4
: -1 that I don't want to succeed in gaining an immediate advantage
: because the leader will drop thereby robbing me of an opportunity to win
: the match through achieving a gammon with the cube at 2. I.E., play with
: the goal of first having your double accepted so that in the event you
: win a gammon, you also win the match.
: Obviously, that cannot be very good reasoning, so I give myself an
: "unsatisfactory" rating for this effort. Such logic would seem to lead to
: the conclusion that as trailer with the opening roll of 3-1, one sould
: not play 6-5 8-5 because the leader will drop a double before your
: next roll.
: Or, is it the risk that the blot on 5 is hit, making you the underdog
: that one is avoiding? That would at least accomplish the mission of
: having your double accepted. ::)) And if one wants to avoid that at -4
: -1 why doesn't TD_one want to avoid it at -1 -1?
: So, I see that the 5 point will never be made in the -4 -1 (while it
: certainly will be in the -1 -1) situation, I just can't grasp the
: relevance of that outcome. I come back to the proposition that at -4 -1
: that the trailers first objective is to play the game with the cube at 2.
: Well, it will make for a good riddle to go back to from time to time with
: the hopes of piercing the vail! william, wcb on FIBS
:
Ok, I'll try again. Let's suppose we are considering two plays, both of
which give you 50% chances of winning. However:
Play A is the gambling play. A lot swings on your opponent's response.
Half the time (when he rolls badly) your winning chances will be 70%, but
the other half the time (when he rolls well) your winning chances will be
30%.
Play B is the conservative play. Half the time your winning chances will
be 51%, while the other half of the time your winning chances will be 49%.
Since your opponent has a free drop, he will always drop if he is the
underdog. Therefore:
If you make play A, half the time he will drop and the other half the
time your winning chances will be 30%.
If you make play B, half the time he will drop and the other half the
time your winning chances will be 49%.
Obviously you should make play B.
This is what is going on (although to a lesser degree, of course) with
slotting the five point. If he hits you have a big disadvantage, while
if he misses you have a big advantage (which you won't be able to
utilize). On the other hand if you play 24/23, 13/11, most of the time
you will be pretty close to equal. That shows why slotting is incorrect.
Kit
Dave McNair
Kit Woolsey wrote:
> Ok, I'll try again. Let's suppose we are considering two plays, both of
> which give you 50% chances of winning. However:
>
> Play A is the gambling play. A lot swings on your opponent's response.
> Half the time (when he rolls badly) your winning chances will be 70%, but
> the other half the time (when he rolls well) your winning chances will be
> 30%.
>
> Play B is the conservative play. Half the time your winning chances will
> be 51%, while the other half of the time your winning chances will be 49%.
>
> Since your opponent has a free drop, he will always drop if he is the
> underdog. Therefore:
>
> If you make play A, half the time he will drop and the other half the
> time your winning chances will be 30%.
>
> If you make play B, half the time he will drop and the other half the
> time your winning chances will be 49%.
>
> Obviously you should make play B.
>
> This is what is going on (although to a lesser degree, of course) with
> slotting the five point. If he hits you have a big disadvantage, while
> if he misses you have a big advantage (which you won't be able to
> utilize). On the other hand if you play 24/23, 13/11, most of the time
> you will be pretty close to equal. That shows why slotting is incorrect.
>
> Kit
Is everything clear now? Just to confuse matters further...
I've recently been playing about with the skewed equity tables I found
on
Roberto Gobbo's pages.
Let's say you reach this scenario having deigned to play someone ranked
700 pts below you on FIBS (unlikely for most of us, but maybe not for
the
likes of Kit!).
The doubling window at this score (if you assume a third of your wins
are
gammons) is 0-70%. Thus if you make play A, half the time you will reach
your
optimum doubling pt!
Technically, the weaker player could take, but of course the weaker
player is
unlikely to realize it's a take, so we're back to square one
(unfortunate
metaphor in rec.games.bg).
Obviously, this is an extreme case, but in a match between unevenly
ranked
players, it seems the "free" drop may be far from free to the weaker
player.
So I'll stick my neck out and say you CAN take these post-Crawford
doubles
from a better player EVEN when you're an underdog in the current game.
Convinced?
\Dave McNair
Ron Karr
Dave McNair wrote:
>
(snip)
>
> Obviously, this is an extreme case, but in a match between unevenly
> ranked
> players, it seems the "free" drop may be far from free to the weaker
> player.
> So I'll stick my neck out and say you CAN take these post-Crawford
> doubles
> from a better player EVEN when you're an underdog in the current game.
>
> Convinced?
>
> \Dave McNair
I haven't seen these tables, but you need to be careful how you
interpret them. The reason the weaker player is supposed to take with
only 30% winning chances (or whatever, depending on ratings difference)
is because he's considered to have only 30% winning chances in the
starting position! Technically, the "free drop" means: drop if your
chances in the current game are worse than your chances in the next
game. Whether you're far stronger or far weaker, your chances after
slotting the 5 point and being hit are going to be worse than in the
starting position.
Ron
Dave McNair
Ron Karr wrote:
>
> Dave McNair wrote:
> >
> (snip)
> >
> > So I'll stick my neck out and say you CAN take these post-Crawford
> > doubles
> >
> > Convinced?
> >
> > \Dave McNair
>
> I haven't seen these tables, but you need to be careful how you
> interpret them. The reason the weaker player is supposed to take
> starting position! Technically, the "free drop" means: drop if your
> chances in the current game are worse than your chances in the next
> game. Whether you're far stronger or far weaker, your chances after
> slotting the 5 point and being hit are going to be worse than in the
> starting position.
>
> Ron
You're right that the table gives the weaker player about 30% in a
1-pointer. I mistakenly used a skewed equity table together with an
orthodox position evaluation (Kit's 70-30 split was clearly based on
each player having equal ability).
Here the stronger player starts with 70% winning chance. So using skewed
position evaluation, half the outcomes of the gambling play A would give
say 90% winning chances, and the other half 50% winning chances. Whereas
conservative play B would give around 71% half the time and 69% the
other.
After this adjustment, the reasons for making the conservative play
still hold good.
\Dave Mc