|
In response to a critical mass of questions that I have witnessed
generally, and on this newsgroup particularly, about the top neural-
network backgammon-software JellyFish (i.e., "JF") before, during,
and especially after the recent commercial release of the derivative
program JF Tutor 1.0, and in anticipation of the imminent commercial
release of JF Tutor 1.0's co- and sub-derivative JF Analyzer 1.0,
scheduled for commercial release later this month,
I have moderately annotated, below, the draft of a review of JF Tutor 1.0
and JF Analyzer 1.0 that I sent to backgammon-promoter Carol Joy Cole at
her request for inclusion in the December, 1994 issue of her worthwhile
monthly newsletter, "Flint Area BackgammoNews."
Unfortunately for at least me, she had a space-limit of one page for the
review and didn't (successfully) send me an edited version of it before
she put it to press, but she had specified the former constraint to me
as a desideratum and had numerous adverse circumstances, including
delayed submission of other material and a broken print-camera, which at
least contributed to the latter.
However, and I hope fortunately, you, the readers of the newsgroup
rec.games.backgammon and/or of its FAQ, here and now have access to much
more of the "inside story" behind the review; and since the version that
appeared in Carol's newsletter not only necessarily edited much out but
also did not have the benefit of subsequent questions that have
materialized from readers and non-readers of the review alike, I think and
hope that even readers of that version will find reading the annotated,
pre-edited version below worthwhile.
Thanks to Carol for inviting me to submit this review, to JF's
creator Fredrik Dahl for trusting me to be one of the first and
continuing members of his select group of beta-testers of JF, and to
those of you who have provided me on FIBS with prompt, positive feedback
on the version that appeared in Carol's newsletter.
Happy reading, backgammon-friends.
-- John Bazigos
e-mail: jbazigos@Kate.ibmPCUG.CO.UK
FIBS-handle: doc
> JELLYFISH: COMMERCIALLY AVAILABLE, STATE-OF-THE-ART, NEURAL-BASED,
> GENERAL-PURPOSE BACKGAMMON-SOFTWARE
>
> The recent commercial release of Fredrik Dahl's neural-based
> backgammon-program JellyFish (i.e., "JF") constitutes a milestone in the
> history of backgammon. For a modest price, it offers the public
> user-friendly access to the best general-purpose backgammon-software to
> date, which will almost certainly result in order-of-magnitude
> improvements in understanding and playing-strength for novices,
> intermediates, experts, and masters alike. Using players need only a
> version of Microsoft Windows at least as recent as 3.1 running on an IBM
> PC-compatible personal computer with an Intel 80x86 CPU chip at least as
> recent as the 386 to run the program and thus avail themselves of the
> master-level training-sessions JF Tutor provides, the rollout-facilities
> JF Analyzer provides, and the insights the Tutor and Analyzer together
> provide for both arbitrary-length matches and non-match (i.e., "money")
> games.
My copy of JF Tutor 1.0 accepts match-lengths between and including 1 and
32,767 (= 2^15 - 1) points.
>
> Both I and another top player who has had greater experience with the
> program independently concluded that JF already plays world-class
> backgammon.
Carol's version referred to him as "another more experienced player",
which was (unintentionally) misleading and/or relatively irrelevant:
misleading if an attempted parahprase, and less relevant than his greater
experience with JF, not backgammon generally, if not. That other top
player, by the way, is Kit Woolsey, but I was unsure at the time whether
he wanted to be identified in the newsletter as a co-author of all the
various claims of his that I cited therein, so I felt unauthorized in
disclosing his identity; because I have since seen him make on
rec.games.backgammon claims that imply all of his cited herein, I feel
implicitly authorized in doing so here and now.
> In fact, when I checked on 27 December 1994, it was ranked 38th of about
> 1700 and rated 1820.43 on the worldwide First Internet Backgammon Server
> (i.e., "FIBS"), where it analyzes at a level, 5, one-ply shallower than
> at its top level, 6 -- at which I normally use it in real time (on my
> 33.2 MHz 486)!
Incidentally, my anecdotal evidence so far indicates that
-- on a 386, level 5 is the highest one playable in real time;
-- on a 486, level 6 is playable in real time only if
-- the chip-speed is greater than 33.2 MHz
and/or
-- the chip-speed is at least 33.2 MHz and includes a math co-processor;
-- on a "586" (i.e., Pentium), level 6 is playable even at 33.2 MHz;
and
-- a math co-processor, like that on my 33.2 MHz 486 DX, increases visible
response-time by a factor of between 2 and 3.
I am one of the many who will appreciate inclusion of further such data in
subsequent postings to this newsgroup, so that we can all profitably factor
them into our decisions concerning future purchases of hardware.
> By comparison, Gerry Tesauro's seminal backgammon-playing
> neural network TD-Gammon (i.e., "TD") was ranked 12th and rated only
> slightly more than 50 points higher than JF's level 5 at that time.
That difference had narrowed to less than 50 points when I checked today,
about one month after I submitted the presently annotated draft.
> The aforesaid other top player and I independently concluded, further,
> that rollouts by JF (Analyzer) not only are the most trustworthy of any
> general-purpose backgammon-software, but also, given the large number of
> trials that the relatively great speed at which they can be generated
> enables practically, the most trustworthy results generally for all but a
> few types of (technical endgame) positions. Indeed, his and my sometimes
> joint but mostly separate use of JF as a primary research-tool for most
> of 1994 has resulted in our routinely considering statistically
> significant enough rollout-results from JF to have oracle status for all
> but the minority of positions for which we have access to a better,
> special-purpose oracle -- e.g., Hugh Sconyers' CD-ROM system Bearoff
> Equities & Backgame Probabilities for bearoffs, Johannes Schmitt's
> BackgammonBase for longer races, and in-house software for some other
> computationally tractable position-types.
>
> Further, since JF's rollout-results constitute probably the strongest
> evidence yet for each side's outcome-distribution (i.e., single-,
> double-, and triple-win percentages) and associated match-equity or
> non-matchgame equity, for all but a few types of (technical endgame)
> positions, these results can be used to test hypotheses concerning
> long-interesting theoretical issues, like match-equities and/or
> opening-plays. Rollouts by JF are statistically efficient because of its
> use of arbitrary-ply stratified sampling of dice-rolls, and a
> particularly good empirical data-source for match-equities because of the
> high and dead-equal level of play on both sides that even the best
> databases of human-versus-human matches suffer from lacking.
This, of course, includes Hal Heinrich's huge such database -- which, when
it contained apparently "only" slightly more than 1,000 matches,
constituted the primary source of data for Kit Woolsey's most recent
published revision ("Inside Backgammon", 1992; "How to Play Tournament
Backgammon", 1993) of his seminal table of match-winning probabilities
("Backgammon Times", 1981) as a function of scores in matches where
neither side is more than 15 points from winning the match.
> As examples, my massive JF-rollout of backgammon's initial position has
> contributed to respective increases in my estimates of the
> opening-roller's equity and of the game's cubeless
> (back)gammon-frequency;
My working-estimates for several years consecutively until last year were
based on the traditional 11:10 advantage of the opening roller being
uniformly distributed over single-, 20% double-, and 4% triple-games, for
figures of +0.059 and 20%, respectively; Jerry Godsey's "Expert Backgammon
Opening Rollout Results Sheet" ("Chicago Point", 70; July, 1994), which
reflected 27,216 trials on EXBG 2.0 and/or 2.1 for each (plausible)
opening-play, implied cubeless estimates of +0.045 and 24.9%, respectively;
and now, my aforesaid JF-rollouts imply cubeless and cube-adjusted values
of +0.065 and +0.083, respectively, for the opening roll, and about 26.7%
for (back)gammon frequency. Though I formerly regarded the 24.9% figure as
suspiciously higher than Kit's long-accepted 20%, I now regard it as a
lower bound (for match-equity derivations).
> my massive
> JF-rollouts of all plausible plays for all opening rolls has confirmed
> most of my suspicions thereabout but also provided me with more reliable
> data on their outcome-distributions, further contributed to those two
> respective estimate-increases, and led me to one revolutionary discovery
> so far;
Would you believe that the routinely overlooked, neglected, ridiculed, or
usually at best hastily dismissed play making the deuce-point with an
opening roll of 64 not only seems best for non-match play but also seems
to increase the opener's gammon-advantage and to be best for most
match-scores?! I never used to until late last year, and still wouldn't
stake my soul on it, but JF-rollouts, my subsequent analysis of them, and
the results of my experimenting with it have led me to the
working-conclusion that it is the best play theoretically, and even better
practically for the serious student of the play. (Incidentally, as I
recall, TD used to make this play before Gerry Tesauro coerced it into
choosing between more "acceptable" opening-plays generally; though I am
sure Gerry, being the fine researcher that he is/was, had his reasons for
doing this --whatever they were-- I hereby humbly express my opinion that
it was inelegant for the implementation of TD, and disturbingly limiting
to the discovery of what neural-nets can teach us (about backgammon) by
processing their own experience.)
> and my massive JF-rollouts of all plausible plays in response to at
> least JF's highest-equity play for each opening roll has provided me
> with unprecedented understanding not only of how to respond to my
> opponent's first play, but also of strategic trade-offs, and of
> significant errors in all published match-equity tables.
I will provide details of all these, and more, things in my forthcoming
magazine "Backgammon Quarterly", for which I have been collecting and
working through ideas for the last four years (If you are interested in
receiving it, please send me e-mail stating so and including your postal
address).
> As, among other things, a professionally trained and experienced software
> engineer, I take pleasure in reporting that Fred Dahl and his small team
> of programmers performed the obviously professionally desirable but
> nonetheless unusual feat of providing childishly simple and enjoyable
> access to software with leading-edge functionality in their December,
> 1994 release of JF Tutor 1.0, and seem firmly on-track to repeat the feat
> for their scheduled January, 1995 release of JF Analyzer 1.0.
Carol ommitted the adjective "childishly" in the immediately preceding
sentence, but I like its connotation: One of my tenets for engineering
user-interfaces is that they should be suitable and fun for the
psychological child within each of us to use; this holds even more
strongly than usual for interfaces to game-software designed for a
user-population that is not entirely comfortable with (personal)
computers.
> JF's clear, terse, and informative 15-page user's manual provides
> general information on the program in a modest tone, and then gently
> steps the user through the mechanics of (un)installation and
> confirmation, starting a new game or (arbitrary-length) match,
> saving/loading a position to a DOS-file, its six levels of play,
> evaluating a position on level(s) 5 and/or 6, editing a position, JF's
> interactive comments, rolling the dice automatically and/or manually,
> cube-action, checker-movement, user-modification of the colors for
> any subset of five colored entities in the graphical interface,
> miscellaneous bells and whistles (such as [a] human-versus-human
> option), and useful shorcuts. A normally intelligent backgammon-player
> with a
> rudimentary knowledge of Windows should be able to start using JellyFish
> almost immediately, and could easily be almost perfectly adept in its
> use in his/her first day if (s)he so desires; and I think that
> JF-adeptness problems should be scarce and minor, since
> backgammon-players are generally more than normally intelligent and
> Windows is so user-friendly for beginners.
>
> Again, the recent commercial availability of JF constitutes a monumental
> event in the history of backgammon by making available the strongest
> general-purpose backgammon learning-tool to date for all players through
> a highly user-friendly interface at a bargain-price. Since our
> privileged introduction to it in the first half of 1994, I and a couple
> of other top players have used it continually to verify, refute, and
> confidently increment our beliefs about backgammon, from those as
> specific as estimates of outcome-distributions for particular positions
> to those as general as what constitute the most plausible values for a
> match-equity table. This has already resulted in an order-of-magnitude
> increase in our understanding of the game, and should result in another
> as we continue this process in 1995. I will share my thus-derived
> insights (and others) in my forthcoming magazine "Backgammon Quarterly";
(Again, if you are interested in receiving it, please send me e-mail
stating that and including your postal address.)
> incidentally, now that JF has become
> commercially avaiable, backgammon-magazinists will face an
> unprecedentedly difficult challenge in validly solving problems that one
> cannot easily, validly, and confidently solve with just one or more
> pieces of commercially available software.
Carol, perhaps inadvertently, transformed this prediction into a statement
of possibility by replacing "will" with "may" in the sentence cited
immediately above, but I maintain the prediction.
> In any case, JF is by far the best generally available investment to
> improve one's game, and, like (almost) all worthwhile products and
> services to improve one's backgammon but extremely so, should easily pay
> for itself rather quickly for almost all who play the game for stakes,
> since it costs only $220 ($110 for the Tutor, and $110 extra for the
> Analyzer); in fact, I think that the overwhelming majority of this
> review's readers should consider being able to run JF sufficient
> justification for buying and/or using Microsoft Windows.
>
> As for its weaknesses, notably in technical endgames, they do not alter
> the facts that its play is master-level and its rollouts almost all
> trustworthy. Besides, given the nature of neural-network
> program-architecture, its demonstrated success in learning backgammon so
> far, and Fred Dahl's care and competence in improving JF as much as
> possible at each point, this defect should be remedied in subsequent
> releases -- beginning with that of JF Analyzer 1.0 in late January, 1995.
|
![[Long message]](pics/Long.gif)
![[GammOnLine forum]](pics/GammOnLine.gif)
From GammOnLine
![[Recommended reading]](pics/Star2.gif)
Recommended reading
![[Recent addition]](pics/New.gif)
Recent addition