Forum Archive :
Jellyfish
I was reading a post from a little while back that talked about a bug in
J/F when it is using its bearoff database. Basically what it came down too
was that jelly voluntarily gave up a gammon because it was blindly
following its bearoff database. In other words the move that was best
overall for a fast bear-off failed to take a chip off when the opponent was
going to win the next move. So first this gave me an assumption which
perhaps someone can confirm or refute. The assumption is that the bearoff
database does not take into account the opponents side of the board. It
contains info such that given a specific position and dice roll should be
played in a specific way to get the pieces off in the quickest fashion long
term,,, without regard for what the opponents position is.
This raised the question to me as to wether there are also other situations
where this might cause jelly to play an incorrect move. Not neccessarily
giving up gammons but reducing its overall chances(equity). I was thinking
that this would occur in situations where only a number of miracle rolls
will win the game. I seem to remember a comment in an match that Kit
Woolsey commented(either Woolsey-Bagai or the demo match in the match quiz
series) that he made a certain move because only miracles could win it so
may as well play for the miracles. This leads me to believe that there must
be positions where jellyfish would play from its database and not notice
that its best or only chances to win were to play for miracle doubles. Or
in other words the best chances lie in making a move where smaller sets of
doubles(miracles) will win it.
Granted this cannot be a loss of much equity but hey I was thrilled when I
found out that jelly made a blunder in a situation that even I would easily
see what to do so thought I would see what everyone else thinks or if
someone could come up with a concrete example. I tried to dream one up
using a 4 chip end game but there didnt seem to be any situations there
that would meet these requirements. After I post this I think I am gonna
try to find the specific position that I was thinking of and check it out.
In the meantime would love to hear if anyone else can think of such a
situation. Its a small victory but against jelly I'll take any I can get.
...
I found it and it didnt take long. Below is the position I was speaking of
in my previous post.
X to play 6-3
13 14 15 16 17 18 19 20 21 22 23 24
+-----------------------------------------+
| . . . . . . | | O . O O O . |
| . . . . . . | | O . O O O . |
| . . . . . . | | . . O . O . |
| . . . . . . | | . . . . . . |
| . . . . . . | | . . . . . . |
| | | |
| . . . . . . | | . . . . . . |
| . . . . . . | | . . . . . . |
| . . . . . . | | . X X X . . |
| . . . . . . | | X X X X X . | +-+
| . . . . X . | | X X X X X X | |2|
+-----------------------------------------+ +-+
12 11 10 9 8 7 6 5 4 3 2 1
Jellyfish with database turned on plays 8/2 3/off. So presumably in a close
race this is the fastest way to get all men off and is based on the most
likely sequence of rolls to occur. In the match 8/5 6/off was played and it
was suggestd by Kit that he played this because only miracle doubles would
win so may as well play for miracles. This play obviously makes 5-5 more
productive next roll taking 4 men off instead of only 3. 6-6 obviously does
this with either move but this move makes another set of doubles better. In
a close playing for big doubles is not best but when trailing it is.
So wooo wooo,,, yet another Jelly flaw :)... though I should mention that
if you turn the bear-off database off Jelly appears to recognize this fact
and makes Kit's play. Which raises the question. When is jelly stronger,,,
with or without this database. Does it gain enuogh equity in close races to
justify the equity it gives up when it needs miracles? My guess would be
yes but I dont know for sure. What is the minimum for Jelly to be behind in
a race before it starts giving up equity? 4 chips behind? 5?
|
|
Michael J. Zehr writes:
For reference:
8/2 3/off lets X bear off in an average of 7.55 rolls
8/5 6/off lets X bear off in an average of 7.57 rolls
Against O's current position, the first play wins about 3.6% and the
second play wins about 4%.
The difference is that O almost always bears off after 5, 6, or 7 rolls,
and the right play for X bears off .5% more often after on roll 5 and
.4% more often on roll 6 at the expense of bearing off less often on
rolls 7 and 8.
If the only thing in the JF's bearoff database is the average number of
turns to bearoff, then that would explain a mistake such as this. If
the bearoff database also contained the chances of bearing off in 1,2,
3, ... rolls, then JF would make the right move here, but the database
would be quite a bit larger (by a factor of roughly 10, depending on
accuracy and storage methods).
When does JF make mistakes using the bearoff database? To answer that
we need to answer two questions: how much equity does JF give up
normally using the neural net instead of the bearoff database? How far
behind does one have to be before it's right to not make the moves that
let you bearoff in the lowest average number of moves?
I imagine the biggest difference this can ever make is 5/216, or about
2.3% in wins, or .046 in equity. This occurs when your opponent is off
in one roll on a double and two rolls on a non-double, and you can make
your play now to make one extra double work for you on the next turn.
So you gain when your opponent doesn't roll doubles (5/6) and you roll
that one specific double (1/36). We can see that in this particular
position, 8 rolls from bearing off (before moving) and 2-3 rolls (4-6
checkers) down in the race, it only makes a difference of about .4% or
.008 in equity.
Thanks for bringing up an interesting topic for study.
-Michael J. Zehr
|
|
Fredrik Dahl writes:
JF has another database for short bearoffs which kicks in when both sides
have 6 or fewer checkers (read the manual ;-)).
This base gives probability distribution for the number of rolls to bear
off, and these are used to calculate equities by convolution.
I didn't really do any research into this, but I figured that for bearoffs
longer than this, playing for minimizing the expected number of rolls to
get off couldn't cost much.
Fredrik Dahl.
|
|
|
|
Jellyfish
- Backgame play (Brian Sheppard, Feb 1997)
- Bearoff database bug (Vince Mounts+, May 1998)
- Cube strategy (Fredrik Dahl, July 1995)
- Doubling at 2-away, 2-away (Michael Bo Hansen, May 1998)
- JF tackles New Ideas in Backgammon (Nigel Gibbions+, Mar 1998)
- Review of JF Player 3.0 (Geoff Oliver, Apr 1997)
- Review of JF Tutor 1.0 (John Bazigos, Feb 1995)
- Showdown in Texas (Chuck Bower, July 1997)
- Strengths and weaknesses (Daniel Murphy, Jan 1998)
- What is a Jellyfish? (John S Mamoun, Dec 1996)
- Why the name? (Fredrik Dahl, Dec 1996)
From GammOnLine
Long message
Recommended reading
Recent addition
|
| |
|