Re: computer-go: perfect play

[Matt Gokey <mgokey@xxxxxxxxxx>]

Whew, my question about the difficulty of chess and go and distance of play from a
perfect go-omniscient entity aptly coined "GoGod" (which I did not know at the time)
has sparked a very interesting debate.

"GoDevil" in my opinion is an amusing concept but only slightly interesting because as
it was pointed out here GoGod vs GoDevil or GoDevil vs GoDevil games are essentially
the same as GoGod vs GoGod since they both have perfect knowledge of the game tree and
GoDevil could gain no advantage from GoGod or GoDevil himself based on his weaknesses
or playing style.  And as for playing fallible humans neither GoGod nor GoDevil would
be happy to sit down and continue to play an already won position probably gained
after the human placed his first stone, likely a game theoretical mistake.   Perhaps
GoGod, being such a benevolent entity, would let us win occasionally so we don't feel
so bad . . . ;-)

>From a game theoretical perspective isn't the empty board one of three positions for
black: 1) Won 2) Lost 3) Drawn?  Isn't this a fact, a truth that we know absolutely?
If this is the case, once one has knowledge of the full game tree, the game is as
pointless as Tic Tac Toe. ;-)  Also, komi was brought up as a point of discussion.
Isn't komi (my go terms are a little rusty) a score adjustment for white with two
objectives - to make the game fairer for white and to eliminate drawn positions?  By
introducing komi then you make the game even less interesting for GoGod because the
empty board is now either a won position for black or a won position for white, and as
komi is increased, more and more full perfect games are won for white, until at some
value all perfect games are won for white meaning the empty board is a won position
for white (assuming perfect play, of course).

I still think the concept of mythical GoGod can be useful in understanding the game
and learning some insights into the game.  The recent posts going down the path of
trying to derive a game theoretical handicap limit and thus a theoretical limit on
GoGod's rating is interesting and I think developing a reasonable limit (not the
actual value) like this is good practical research that could lead to some insights
into the game.  However, my guess is the true value of this handicap is 1.  Let me
explain: with perfect play a handicap of 1 would be enough to tip the tables and make
the empty board a won position for black.

I would have to say that I agree with Don in many of his posts and responses.  My own
thoughts as I wrote the questions were that humans really don't play chess or go very
well, in a game theoretical sense, that is.  From a practical consideration, given the
sheer complexity of the games, and what we do know about how we play, the masters are
truly amazing though.  In "The Inmates are Running the Asylum", Cooper calls computer
programmers Homo-logicus, a different breed from Homo-sapien;  I think the best
go/chess/etc. players are different breed too:  Homo-strategum.

Now about go complexity vs chess complexity.  If you look at the two games completely
on technical grounds only, go scoring complications and end of game recognition
coupled with no clear workable evaluation and the size of the game tree and number of
positions - its a hands down win for go.  On these terms alone go is vastly more
complex than chess - by probably many orders of magnitude!  And this is evident in how
difficult it has been for computer go programming.  For human players, however, the
picture looks very different.  For humans, it may be that Go is little more difficult
than chess to master, but it doesn't seem greatly so and it is actually easier for a
child to learn and easier to become novice-proficient.  Based on all the discussion
here over the last few weeks its decidedly undecided precisely how far away master go
and chess players are from so-called perfect game theoretical play, however, it seems
both have at least significant room for improvement.  Perhaps chess has a little less
room for improvement at the top.  So what is going on here? IMHO, I think humans are
taking advantage of aspects of go which allow them to reduce a game orders of
magnitude more complex than chess on technical grounds, to approximately the same
difficulty (perhaps a bit more difficult) on practical grounds.   Go has fuzzy logic
written all over it - patterns (local and whole board), assessment of probabilities,
mutual subgame degree of independence, fuzzy borders between groups, estimating
scoring, reading on higher levels than individual stones, in addition to more concrete
parts like the endgame, reading ladders, life and death analysis (partly
patterns/partly reading) all of which are mixed together to make decisions.  It also
is relatively static with regard to placement of pieces making it much easier to read
(for humans) and much more visual (and we have millions of years of evolution on our
side for very robust visual pattern recognition).

That said, I don't believe great computer go engines will play like humans or that we
should try to emulate human play, but I do believe we must learn how to take advantage
of some of the same characteristics of go that humans take advantage of.  A crucial
question is how can we take advantage these characteristics and using what software
techniques?  What are some promising areas of research or study that we can apply to
go?  Does anyone have some ideas on this?

Matt