|
> Marty, may I ask when your book will be available and can you give a
> brief synopsis of it's content/style?
When will it be available? Good question. The best I can estimate now is:
in a few months. I'm extending and upgrading rollouts, incorporating some
feedback from the cognoscenti. But it's 99% done. The most annoying thing
left is to wrangle over format. I think I have a good layout now, but
these things must be negotiated, and once negotiated, actually made real.
Fear not, I do layout very quickly.
The book is match analysis. The style is my usual, only perhaps better. In
about August 2002 I analyzed a match for GammonVillage: Senkiewicz vs.
Tardieu. Unfortunately, the editors put it in the "Bot" section; I used
Snowie 3 to help analyze, but unlike some other specimens of analysis in
GammonVillage "Bot," my analysis went way beyond creative rubber-stamping
of evaluations and rollouts. The "bot" aspect was secondary, in my of-
course-quite-unbiased opinion.
My book has three matches in similar vein, but better analysis than I did
for Senkiewicz-Tardieu: very detailed, but more thorough, and with more
emphasis on generalization. I include a "Lessons" section after every
game. But no Celestial Pronouncements. Lots of Snowie 3.2 rollout data is
included (I favor transparency over opaqueness!), with good settings (at
least 2-ply lookahead, score-based play, live 3-ply cube, 100% huge--but
often 3-ply lookahead instead of 2-ply, all other parameters the same:
also, 95% confidence intervals included). There are lots of diagrams, in
Monte Carlo font, including diagrams of alternative plays and related
positions.
I say a lot about bots versus humans--hard to avoid when you use bot
data--but that's not the book's main theme.
My next effort will use GNUBG, but since I started this project almost two
years ago, I used what I thought was the strongest bot at the time. I
rolled out hundreds of positions, usually several alternatives per
position. Rollouts took @#$#@)(* forever on my 1GHz Athlon Windows box and
it's too late to redo them; by the time I'd finish redoing, a better bot
would be available and I'd still be behind.
The three matches are all by Malcolm Davis: against Ed O'Laughlin, against
me, and against Frank Talbot. They're good matches. All the players kindly
consented to give commentary. Their commentary was illuminating enough to
make me rework some of the analysis, add and extend rollouts, etc. Their
comments are included, often verbatim as transcribed from audio.
I don't think the book is dry at all. It's not for the faint-hearted, but
I've worked hard to make it engaging and readable. I realize match
analysis is often a hard sell because it's a lot of work to read, so I did
my best to make the reader's task easier. Initial feedback is quite
positive. One very strong player said he felt he didn't need a backgammon
board to go through the games. I think almost everyone can get something
out of it.
I'm still waiting for more feedback from Mary Hickey! Hint hint!! Her
feedback will make the book even better. She's witty and knowledgeable and
she writes well.
The book is now a bit less than 600 8.5 x 11" pages. It might get all the
way to 600. I have no idea what the price will be. I should think at least
40 USD. There should be an advertisement for it in Walter Trice's new
book. See http://www.fortuitouspress.com for info and updates.
> Marty, I wonder what procedure you followed: starting with writing down
> your own opinions about a play, and then consulting the bot and
> rewriting and adding to the analysis? Reason I am asking is that
> otherwise it is difficult not to rubberstamp the bot's opinions.
I started with a Snowie analysis file: 3-ply evaluations. Initially I took
the attitude that if I disagreed with Snowie, except in certain types of
positions where I know it has weaknesses, then Snowie was probably right.
For all positions I found interesting and had some doubt about my play
versus Snowie's, I did rollouts: usually initially 2-ply lookahead, 3-ply
live cube, score-based play, 100% speed, huge search space. If I still had
doubts, or if outcomes were very close, or if positions were complicated
(backgame, tactical, lopsided score or special situation like DMP), I'd
use 3-ply rollouts. Somewhere in the middle of all that I'd write down my
thoughts about a position.
Finally I'd either have the position elucidated to my satisfaction, or I'd
have doubts about how important various factors were. In the latter case
I'd give my opinion with the best justification I could.
I think I succeeded overall in my aims. Of course there will be some
mistakes, and some of those will be bias in favor of the bot. But
hopefully there won't be many. I try to be as objective as I can.
Another thing: Maybe 1/5 of all rollouts resulted in reversals of the
Snowie 3.2 3-ply evaluation. Of course, I didn't roll out all positions,
only the ones I thought were interesting or instructive. But the fact that
there were so many reversals means there's still very much room for
improvements to bot technology. It also means that analysts have to stay
on their toes!
|
![[GammOnLine forum]](pics/GammOnLine.gif)
![[Long message]](pics/Long.gif)
![[Recommended reading]](pics/Star2.gif)
Recommended reading
![[Recent addition]](pics/New.gif)
Recent addition