Reliability of robot rollouts, part I
ba...@datablast.net
How long does it take Jellyfish to do something like this. That is
how long does it take do do 7776 truncated rollouts on say level 7.
Thanks
On 14 Oct 1997 23:10:51 GMT, bo...@bigbang.astro.indiana.edu (Chuck
Bower) wrote:
snip
>
> The most recent game is from a match between Emil Malek and
>Nick Ballard, with analysis provided by Ron Karr (and JF rollouts
>by Richard McIntosh). In the past I have questioned the use of
>level-5 Jellyfish rollouts. For play decisions, Richard does 7776
>truncated trials with an horizon of 7. ("Truncated rollouts" means
>that JF will play the game for a set number of rolls and then
>evalulate the position and keep a tally of the evaluated results.
>The algorithm is repeated with random dice until 7776 truncated
>games have been played--in Richard's case. "Horizon" is the number
>of rolls played for each side before the evaluation/tally.)
>
snip
Chuck Bower
(NOTE: this is the first of 3 parts.)
Frequenty there are posts asking "how can I best learn to be
a strong BG player?" Standard answers are "read GOOD books, watch
good players, play against good players, study the games of good
players, etc."
Backgammon by the Bay has an excellent WEB page which falls
under the "...study games of good players..." category. Several
annotated games can be found at:
http://www.backgammon.org/bgbb/games.html
From an aesthetics standpoint, I find the page appealing. (But,
hey, I've been defending the Jellyfish GUI, so beware!) Recent
improvements include a new color board layout. Technically, the
choice of games and analysis are top notch. The annotators make
use of Jellyfish rollouts to critique the plays made at the table.
The most recent game is from a match between Emil Malek and
Nick Ballard, with analysis provided by Ron Karr (and JF rollouts
by Richard McIntosh). In the past I have questioned the use of
level-5 Jellyfish rollouts. For play decisions, Richard does 7776
truncated trials with an horizon of 7. ("Truncated rollouts" means
that JF will play the game for a set number of rolls and then
evalulate the position and keep a tally of the evaluated results.
The algorithm is repeated with random dice until 7776 truncated
games have been played--in Richard's case. "Horizon" is the number
of rolls played for each side before the evaluation/tally.)
Four positions in the match caught my eye (meaning I wasn't at
all sure what the right play was, even after reading the rollout
results and Ron's analysis). I decided to perform JF level-6
cubeless rollouts to see if the level-5 truncated results stood up.
I was quite surprised by the outcome!
For those not familiar with JF, level-5 means JF looks at the
postion, uses its neural net (brain) to assign an equity (cubeless
value of the game) for the various legal moves, and chooses the best
of these moves. In "level-6", JF builds a tree. For each legal
move, it looks at the equities of all of the opponent's possible moves
NEXT roll and then makes the choice, assuming the opponent makes
the "best" play. Level-6 has (not surprisingly) been shown to play
a stronger game than level-5.
So why would anyone use truncated level-5 rollouts? I can only
think of one answer: speed! JF level-6 takes a lot of time, partly
because it plays the games all the way through. But even more
time consuming (by an order of magnitude) is the time it takes
looking one roll in the future. Is the speed worth the sacrifice
in accuracy? I've always been skeptical.
There is a "hand-waving" argument which goes like this: "If
a rollout robot makes mistakes, that's OK for comparing two plays,
because it will make similar mistakes for each position (averaged
over the long run) so they cancel, and the true difference between
the plays emerges." In a mathematics class, this kind of argument
would be lucky to receive the grade of D!
Now for some BG history. Prince Joli Konsil wrote a book
in the 70's--The Backgammon Quiz Book (Playboy Press, 1978, sc.).
Some of his answers were controversial, but it was harder back
then to decide. There was no JF! Soon after his book was published
Konsil became BG editor for GAMES magazine. In the July/Aug 1980
issue, he reprinted five of his more controversial positions. BUT,
better yet, he had robot rollouts! A non-commercial(?) program
was used to compare various plays for the five positions. I quote:
David Rothman of Hawthorne, California ...uses a Data General
machine that can take a given backgammon situation and run
1000 trials in half an hour. The computer doesn't play on a
very sophisticated level but this flaw is fairly minor since
it plays "equally poorly" for both players.
Does Konsil's last sentence sound familiar?!?!?!
Years later I ran Expert Backgammon (and recently JFv3.0) rollouts
on Konsil's five controversial positions. See part II of this series
for the positions, surprises, etc.
Chuck
bo...@bigbang.astro.indiana.edu
c_ray on FIBS
Richard McIntosh
It depends primarily on the speed of the processor, and to some extent
on the difficulty of the position. (As a extreme example, if one side
dances, it takes much less time to arrive at the horizon.) JellyFish
3.0 is much faster than 2.x. For the upcoming BGBB annotated game
(#13), my Pentium 150MHz averaged 78 minutes per position, level 5
horizon 7. That is consistent with times for the nearly 3000 positions
that I have rolled out for the BGBB archive in the last year.
Regards,
Richard McIntosh
ba...@datablast.net wrote:
>
> How long does it take Jellyfish to do something like this. That is
> how long does it take do do 7776 truncated rollouts on say level 7.
> Thanks
>
> On 14 Oct 1997 23:10:51 GMT, bo...@bigbang.astro.indiana.edu (Chuck
> Bower) wrote:
> snip
> >
> > The most recent game is from a match between Emil Malek and
> >Nick Ballard, with analysis provided by Ron Karr (and JF rollouts
> >by Richard McIntosh). In the past I have questioned the use of
> >level-5 Jellyfish rollouts. For play decisions, Richard does 7776
> >truncated trials with an horizon of 7. ("Truncated rollouts" means
> >that JF will play the game for a set number of rolls and then
> >evalulate the position and keep a tally of the evaluated results.
> >The algorithm is repeated with random dice until 7776 truncated
> >games have been played--in Richard's case. "Horizon" is the number
> >of rolls played for each side before the evaluation/tally.)
> >
> snip
Morten Wang
I'm eagerly awaiting the next part about the rollouts, but before that
arrives I have to make a slight correction.
* Chuck Bower
| In "level-6", JF builds a tree. For each legal
| move, it looks at the equities of all of the opponent's possible moves
| NEXT roll and then makes the choice, assuming the opponent makes
| the "best" play.
on level 6 JF will pick out the strongest moves on level 5
and analyze them one roll further into the game. then it will pick
the best move. it won't check all the moves (as far as I can
understand by reading the JF documentation). the documentation
doesn't say anything about how many moves it checks though, or if it
only takes the best moves.
Morten!
--
"God does not deduct from our alloted life span
the time spent playing backgammon."
--> Morty on FIBS
--> Backgammon homepage: http://home.sn.no/~warnckew/gammon/