[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] Perfect sub-tree move path (was Brute Force...)...



>   And even given awesome data encoding techniques, also vastly surpasses
>   the  any storage  mechanism available  now or  for centuries  into the
>   future,  presuming electron spin  as the  memory mechanism  and entire
>   planets   as   the  atomic   databanks.    This   leads   me  to   the
>   conclusion...the perfect "sub-tree" play paths is a theory with little
>   or no value.

Maybe, but  not so  sure.  You  are thinking only  in terms  of todays
technology and imagining only a few decades out.  The physical laws of
nature seem to preclude the 32 man chess database, but I remember that
they said silicon had it's limits  many years ago and they have pushed
way way  past them.   The quantum world  is hardly understood  and who
knows  what is  possible?  Perhaps  not in  the 2000's  but  maybe the
2100's?

There  are  other  ways  to  skin  a cat  too.   Maybe  some  kind  of
meta-database  approach can give  us the  equivalent of  the omnicient
database  with several  orders of  magnitude less  data  combined with
sophisticated rules that we can't yet imagine.

But our discussion is more  theoretical than practical, you are trying
to apply it directly practical.

Nevertheless,  if such  a  database existed,  it  would be  incredibly
valuable and the quality of play would mushroom considerably in chess,
perhaps not so in GO.

Even  though we  know it's  not practical  to keep  very much  of this
database in your head, the human brain can fill in a lot of holes with
meta-data.  You  could still learn thousands of  opening sequences and
find pathways  to positions you  were quite comfortable  with.  Simply
training   against  such   a   database  would   improve  your   skill
significantly and be worth billions of these database positions.

- Don


   Date: Fri, 12 Nov 2004 09:16:21 -0800 (PST)
   From: "Jim O'Flaherty, Jr." <jim_oflaherty_jr@xxxxxxxxxxxxxxxxx>
   Reply-To: computer-go <computer-go@xxxxxxxxxxxxxxxxx>
   Sender: computer-go-bounces@xxxxxxxxxxxxxxxxx
   X-Scanned-By: MIMEDefang 2.42

   Don (and others),

   Consider Vincent's  original assertion  - there is  a perfect  line of
   play.  As  you (and  others) have pointed  out, the perfect  "line" of
   play is  actually a tree  subset of all  possible moves at  each node.
   This tree notion results in  an exponential explosion, one that vastly
   exceeds the memorization capabilities of humans.

   And even given awesome data encoding techniques, also vastly surpasses
   the  any storage  mechanism available  now or  for centuries  into the
   future,  presuming electron spin  as the  memory mechanism  and entire
   planets   as   the  atomic   databanks.    This   leads   me  to   the
   conclusion...the perfect "sub-tree" play paths is a theory with little
   or no value.

   Now  consider that  we are  talking  about two  human (and  presumably
   finite  and  fallible)  players.   Given  player  A  cannot  perfectly
   memorize the huge trees of  decisions occurring just a small number of
   moves out  (Go or Chess),  the pressure on  player A is to  choose the
   move with  the tree subset with  the shortest "long  branch".  This is
   the move  leading to a Player A  win in the least  number of remaining
   moves.  This  means that  there are that  many fewer options  Player A
   must have  committed to memory to  ensure he stays  with the "perfect"
   sub-trees.  It increases his odds of completing the game "perfectly".

   However, this same  pressure acts on player B  in exactly the opposite
   way.  Player  B’s goal will be  to ensure he chooses  the move that
   requires player A have the  maximum number of future choices committed
   to  player A’s  limited memory.   Player B  is playing  to maximize
   Player A’s fallibility.   Player B is not playing  to win, lose, or
   even to tie while player  A is successfully remaining on the "perfect"
   play track.  Player  B is merely trying to push player  A into a board
   state  for which  player A  (being fallible  which  implies forgetful)
   cannot recall the "perfect" move.

   If  player  B eventually  succeeds  in getting  player  A  to make  an
   "imperfect" move, the game moves into an state where the outcome is no
   longer guaranteed, and returns to probabilistic.

   >From  all of  this,  I make  the  following assertions:  A) There  is
   currently  no  means  by   which  to  produce  a  “provable”  or
   “certain” perfect sub-tree move path data set.  There is neither
   the computation  power nor  the storage capacity  to deal with  it for
   Chess or Go.   Checkers is now converging on  producing that solution.
   Chinook’s endgame DBs are now finally approaching the first half of
   the game.  They  started from the end game  and are working backwards.
   This is not  reasonably possible with Go.  I speculate  it is not even
   reasonably  possible with  Chess,  but have  not  given it  sufficient
   thought  to be  sure.  B)  If ever  a single  “perfect  path” is
   discovered through either Chess or  Go having the appearance all moves
   from game start  to end are forced, the motivation on  player B to not
   cooperate in following that path  will cause it to NEVER BE COMPLETELY
   PLAYED, thereby taking away the  path’s claim to value.  C) For the
   foreseeable future  (centuries), it is going to  be creativity (likely
   human, unlikely computer) that finds lines off of know sequences which
   alter  play.  This  can  be seen  as  Go Joseki  DBs are  occasionally
   updated when some 9 dan, after spending years researching and playing,
   discovers a nuance  of weakness in a particular  response, exploits it
   and then the world finds out when he utilizes it in a game.

   For me, this leads me to  be quite optimistic about Go’s future.  I
   am  not so  interested  in Chess  anymore  thanks to  the brute  force
   techniques finally playing at GM levels.


   Jim O’Flaherty


   --- Don Dailey <drd@xxxxxxxxxxxxxxxxx> wrote:

   > 
   > Joss,
   > 
   > >   1) White wins and so black has no real "best move", (s)he therefore
   > >   plays as well as possible, but always knowing that a loss was inevitable
   > >   assuming that white plays perfect responses.
   > >   ...
   > >   ... If not, black may as well resign after the first move 
   > >   and so the single best line is simply: White moves, black resigns.
   > 
   > Personally, I don't think there is much basis for defining "resign" as
   > a best move (even in a  losing position) which is essentially what you
   > are doing.   "Best" means better than  the rest and it  makes sense to
   > consider it valid for  the "the rest" to be a null  set.  (At least it
   > makes sense in set theory to consider null sets valid sets.)
   > 
   > As you  point out,  there are  other ways to  define "best."   You can
   > define "best" based  on opponent modeling, trying to  make things more
   > complicated for the  opponent.  But then this is  not really chess, it
   > is a different  game which is no long deterministic.   I admit that it
   > is  probably more like  the game  as humans  play it!)   It's probably
   > likely that based on this definition the "best" move may actually turn
   > a draw into  a loss in some  cases if it can increase  your chances of
   > winning against a fallible opponent.
   > 
   > This is  just for fun,  but if you  want to extend your  definition of
   > "best" and make  it a little more "human", I  would try something like
   > this:
   > 
   >   1.  If position  is a win, the  best move leads to  checkmate in the
   >       least  number of  moves.  If  you are  on the  losing  side, the
   >       "best" move delays the inevitable for as long as possible.
   > 
   >   2.  Draws  are little  trickier  due to  50  move rule,  insufficent
   >       material, etc.  So I would go with this:
   > 
   >       Both  sides  want  to  delay   the  game  as  long  as  possible
   >       (presumably to  give the opponent  a chance to  blunder.)  There
   >       are 4 ways  to draw (not including by  agreement which we define
   >       to always be "bad"):
   > 
   >           1. Three fold repetition.
   >           2. stalemate
   >           3. Insufficient material.
   >           4. 50 move rule.
   > 
   >       OBSERVATION:   All games in chess must eventually terminate, if not by 
   >                      an outright win or stalemate,  by an eventual draw by repetiton.
   > 
   >       Throw  out  the 50  move  rule  for  this definition,  it's  not
   >       technically relevant and is  a human concieved device to shorten
   >       the game.
   > 
   >       Throw out insufficient material.  This is another device to make
   >       the game  much nicer for human  players and is  not intrinsic to
   >       the game.  All  of these kind of games  terminate eventually due
   >       to repetition.
   > 
   >       So the best  move in a drawn position, is  the move which delays
   >       the  "bitter end"  the  longest,  assuming no  50  move rule  or
   >       insufficient material!
   > 
   > Please note, even 3 fold repetition  is not a draw unless claimed, but
   > for calculating  this new hypothetical  definition of "best"  move, we
   > assume that  it MUST  be claimed when  it first occurs,  otherwise the
   > game would NEVER terminated  in many positions.  (We could technically
   > even throw out repetition, but  then our omnicient computer would hang
   > up in an infinite loop trying to find the end of the game and it would
   > always play for repetition when a win wasn't possible.)
   > 
   > - Don
   > 
   > _______________________________________________
   > computer-go mailing list
   > computer-go@xxxxxxxxxxxxxxxxx
   > http://www.computer-go.org/mailman/listinfo/computer-go/
   > 

   _______________________________________________
   computer-go mailing list
   computer-go@xxxxxxxxxxxxxxxxx
   http://www.computer-go.org/mailman/listinfo/computer-go/

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/