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

Re: [computer-go] Modern brute force search



Vincent,

You are basically  saying that you believe only  a single opening move
wins for  white and the rest  of white's initial move  choices must be
either forced draws or losses for white.

I would  find it remarkable  if this was  the case even with  just the
first move.

I  do believe  there are  many positions,  maybe even  most positions,
where only a single move is best in the game theoretic sense.

But  you cannot be  correct about  there being  only a  single winning
line, here's the proof by contradiction based on your premise:

  You  say there is  only 1  winning line  and that  chess is  won for
  white.  Therefore  only ONE white move  wins (the rest  must lose or
  draw or you are already wrong.)

  Black has 20 possible responses to any first white move.  But all 20
  responses MUST  lose against the single winning  white move (because
  of your premise that white wins by force.)

  Even though  blacks 20 responses all  lose, white must  still have a
  response for  each of  those 20 moves.   That means we  have already
  branched to 20 different lines of play, all of them based on perfect
  play by both sides. 
 
  But  you  say that  there  is  only 1  line  of  play.   There is  a
  contradiction here.

The only possible way the 1 line argument can hold is if the game is a
theoretic  draw, in which  case the  only possibility  is that  1 move
draws for white, everything else loses.   Then it must be the same for
blacks first response to white, etc until the game ends in a draw.

So there are 4 possible cases here and only 1 of them can be true:

   1. Game is won by white,  many lines of "best" play.   
   2. Game is won by black,  many lines of "best" play.
   3. Game is a draw,  only 1 single "best" line of play.
   4. Game is draw, many lines of "best" play.

Option 3 is virtually impossible, though I can't furnish a proof.

Option  2 is  extremely unlikely,  although  I have  heard "tongue  in
cheek" arguments for  this based on the idea that  black might be able
to  take advantage  of the  fact that  white has  to be  the  first to
commit!

   
- Don




   X-Sender: diep@xxxxxxxxxxxxxxxxx
   Date: Thu, 11 Nov 2004 18:43:46 +0100
   From: Vincent Diepeveen <diep@xxxxxxxxxxxxxxxxx>
   Cc: computer-go@xxxxxxxxxxxxxxxxx
   X-Virus-Scanned: by XS4ALL Virus Scanner
   X-Scanned-By: MIMEDefang 2.42

   At 10:05 11-11-2004 -0500, Don Dailey wrote:
   >
   >Richard,
   >
   >Vincent is NOT saying that there  is a player who knows the best line.
   >He  is saying that  if a  hypothetical computer  was built  that could
   >calculate the best line, the top players would commit it to memory.
   >
   >Of course I think Vincent underestimates the difficulty of doing this.
   >I doubt there is a single "best line", there must be many ways to play
   >a perfect game of GO.

   I am sure there is just 1 line.

   in chess egtb's we can clearly see that the optimal play involves usually a
   single line.

   Only when it is real simple then there is sometimes more choices.

   >Just as a silly  example, let's say it turns out that  Chess is a draw
   >from the  opening position.  Both  white and black players  must learn

   Please assume white begins and wins. At the top level a top GM tries to
   draw with black and hopes for a win with white. 

   If white begins and wins things are more difficult in chess.

   idemdito in go. black begins and wins. no question about it.

   >every possible way to achieve this result.  Let's say that it's a draw
   >against all  20 of  the first possible  moves.  The black  player must
   >learn the correct  response to every one of these  20 moves because he
   >doesn't  know  which  white  will  choose.   The  white  player  could
   >specialize in just one of those first moves.  However it will probably
   >be  the  case that  black  has more  than  one  "drawing" response  to
   >whatever  move white chooses.   White will  be forced  to know  how to
   >respond to each  of these.  Each player can  take turns "specializing"
   >in one  subtree at each  point, but it's  easy to see this  will still
   >quickly mushroom out of control.

   A major problem for software, and this problem is not easy to solve, is
   that the longest line from a human viewpoint is very trivial to see usual.
   It's the medium term 'computerish tactics' that is hard.

   >A really  strong human would  benefit enormously from  this knowledge,
   >perhaps the  equivalent of several  hundred ELO rating points,  but it
   >still wouldn't guarantee  a draw because he simply  could not remember
   >the whole tree of possibilities.    What a really great player could 

   You seem to underestimate what professional players, go or chess usually know.

   Usually they know more books than even the biggest hand written openings
   book of a chessprogram. And that's several million big.

   >attempt to do is memorize enough of the tree to get him to positions
   >he feels very comfortable with, but this wouldn't be fool-proof.

   memorizing is very simple. most players never forget any game they played.

   it would be wishful thinking that in go the things aren't the same. 

   >In GO, it's  even more unlikely that even a  top player could memorize
   >enough sequences  to play perfectly, and  I even doubt  it would bring
   >the  level of  play up  very much  since it  would probably  be common
   >practice to quickly get away  from the most "obvious" continuations in
   >order to nullify the memorization.

   In Go from human viewpoint many sequences are near to forced. The play
   happens at a very tiny area of the board. That's why reading deep in go is
   not a single problem.

   Even a real poor go player has no problems to calculate a deep forced win
   there.

   In such forced wins usually there is many moves which from computer
   viewpoint are 'unforced' by any definition. 

   A human just knows: "this move is needed to garantuee X Y Z".

   No computer knows that. Go or chess.

   Just imagine a pro player. 

   A good example is perhaps 10x10 draughts. Right from the start software
   found tactical tricks which no human could ever find. Any trick from a
   world champion was found instantly.

   Yet for example my friend Marcel Monteba who plays at a master level
   strength (doesn't have a title yet), he easily can beat any software.

   >From a mathematical viewpoint it is very difficult to explain to scientists
   why chess is so much simpler.

   >From human viewpoint it is very simple.

   In chess, just centralize your pieces and you win.

   In draughts the center is not so important like in chess.

   So building up is very important. 

   It's very difficult to explain to a program why a certain move doesn't
   build up the position.

   Now imagine this, draughts is 10x10 board from which 50 squares get used.

   Go has exactly the same symmetry like chess, but also the center initially
   is not so important.

   That means that the move selection from human viewpoint is very simple and
   from a computer viewpoint real terrible.

   So a perfect playing database when using opponent modeling (see v/d Herik
   and others for this publication www.icga.org), it is very well possible
   that the computer plays a move after which the very deep win is just very
   logical for a professional player as he just sees 1 move at a time as the
   rest sucks. Especially in Go this scenario is very likely.

   So if professional players, who already will go up a lot in strength by
   being able to research what perfect play looks like, if you put them
   against a 361 stone database, then obviously it's real difficult to fool them.

   Just don't underestimate how good human learning is. 

   So far human learning has always outclassed anything else in the universe.

   >- Don
   >
   >   Date: Thu, 11 Nov 2004 08:21:20 -0600
   >   From: Richard Brown <rbrown@xxxxxxxxxxxxxxxxx>
   >
   >   Vincent Diepeveen wrote:
   >   > In whatever game we play.
   >   > 
   >   > The 361 stone go database will beat of course every player very easily,
   >   > just like a 32 stone database in chess will beat the strongest
   players very
   >   > easily.
   >   > 
   >   > The reality is of course that by the time we have a 32 stone database
   and
   >   > that there will be a 361 stone database for go, that players have looked
   >   > into it and already know the best line by head.
   >
   >   Really?  Do you know a go player who knows the "best line"?  Wow!
   >
   >   > So they can win with white in chess and with black in go very easily
   when
   >   > using the optimum line.
   >
   >   Well, sure, theoretically.  That is to say, _if_ they know the optimum
   line.
   >
   >   Again, I must ask, who is this go player, who knows the optimum line?
   >
   >   I think that the rest of the go world would like to meet this great
   master!
   >
   >   -- 
   >   Richard L. Brown             Office of Information Services
   >   Senior Unix Sysadmin         University of Wisconsin System
   >				780 Regent St., Rm. 246
   >   rbrown@xxxxxxxxxxxxxxxxx              Madison, WI  53715
   >
   >   _______________________________________________
   >   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/