Re: computer-go: minimax and go

[Don Dailey <drd@xxxxxxxxxxxxxxxxxxxxx>]

   From: "Vlad Dumitrescu" <vladdu@xxxxxxxxxxxxxxxxx>

   some more random thoughts...

   > These   discussion  always degrade  to   considering   Go as having  a
   > branching factor of  361 and a calculation of  how much power it takes
   > to do a full width search to 1000 ply.  But that is not the model that
   > makes any sense.
   It makes sense if one is after the Perfect Play -- under the presumtion that
   this quest makes any sense per se. :-)

Yes.  At some point search is more reliable than evaluation, because I
can at  least   imagine  playing perfectly   with  an  infinite  speed
computer, I even know how to write  this program right now.  But given
an infinite speed computer, nobody   knows how to write an  evaluation
function that plays perfect go, without needing a search.

How many programs evaluate  ladders completely statically?  Some of us
seem to believe more in the reliability of local search than in static
evaluation.  Why is it such  a stretch to  believe that some searching
needs to be involved in a scalable algorithm?


   Reading all these posts where the point lies on more philosophical and
   esoteric aspects, I'm struck by the fact that many "disputes" are caused by
   the imperfection of the language. What is "a very sophisticated  pruning
   mechanism"? It means different things to each one of us... :-)

I sensed a  lot of misconceptions too,  based on language.   Mainly, I
think people  imagine that  I am  claiming a full  width, fixed  depth
chess style search with a stupid evaluation is the way to go!

Really, the only point I  want to make is that  if we utilize the full
power  of the computer, our  algorithm will automatically scale.  This
is simple and obvious.

   Selection helps keeping things manageable, but may keep programs from
   discovering new ways of play... A tradeoff as usual! :-) Selection may not
   only remove bad moves, but also moves that lead to a position impossible to
   comprehend for a regular human (like a huge semeai on the whole board).

I believe selectivity  should be  tapered.   You should be  willing to
accept more and more error toward  the leaf nodes.  The characterstics
each program should  have, is that it  reconsiders it's move with more
and   more  calculation, so  that   given enough   time, it  will play
perfectly.  Humans do this, and change their mind, so should computers
in this case.

   As for quantum computers, there isn't needed to have one unit (=quanta) for
   each node in the full tree of the game. But even if it was, it would be a
   huge step forward anyway, to have that amount of computing power
   available...

It's not clear whether quantum  computers will be useful  for Go.   My
understanding is that they might work well  for some specialized tasks
but not for general problem solving.

   Have a nice weekend!
   /Vlad
   --------------------------------------------------------------
   For God's sake, Smithers! It's not rocket science, it's just brain surgery!
   --------------------------------------------------------------

You too!

Don