computer-go: perfect play

["Vlad Dumitrescu" <vladdu@xxxxxxxxxxx>]

Hi all!

these threads have been (and hopefully still will be) very interesting to
follow. So I feel compelled to submit my piece of wisdom (all the two pence
worth of it!)

*** First, I think it is obvious that perfect play is not possible without
knowledge of the full game tree. This knowledge can be explicit or implicit
(as in "my dad analyzed the position from here and told me this is the
winning sequence"). Even if some shortcut may exist, for it to indubitably
be "perfect" it has to be based on someone's analyze of the full tree.

There is actually an alternative to that. Someone might discover a smart
analogy with some mathematical model where things can be proven using
mathematical methods (just about the way it is done for the simple endgame
in "Mathematical Go Endgames"). In my opinion, at the present this
alternative is also out of reach for us, non-omniscient people. But
developing such a theory is at least possible for a human being, so all hope
is not gone :-)

Thus it is beyond anyone's power to prove anything about the strength of the
perfect player.

*** Second, even if someone had the "cheat sheet" for playing the perfect Go
game, the question remains whether there is a winning strategy for black
and/or white. Maybe there isn't any - and thus the game (at least in the
beginning) would still be open to play. This way, a game between two perfect
players would not just involve at most one move! :-)

There is a side-issue here: how do we know what the right komi should be
between two perfect players?

*** Third, I find it difficult to believe that the perfect play would
involve a single or a few lines of play. One of the reasons Go is a
wonderful game is that it allows for a wide variety of strategies and styles
of play. So I don't really find a difference between GoGod and GoDevil, they
just play differently from each other. The only difference is only for the
human facing them :-)

Wouldn't be interesting to watch "perfect Go Seigen" face "perfect
Fujisawa", or "perfect Takemiya" vs "perfect Shusaku"?

*** Now all this is just speculation and a bit of philosophy. And it applies
as well to chess or any game. Since full tree search is not a practical
thing to do, we are facing the problem of finding shortcuts, guessing our
way ahead and letting actual play validate our assumptions.

It is really a kind of genetic evolution, the game is slowly evolving and
the computer programs do too, at their own pace. Mutations arrive when a
top-player decides to try some unusual variation; if it works more will
imitate it, analyze it and develop it - the cross-over mechanism.

*** I don't want to get into a dispute, but I really wonder about one of
Vincent's latest affirmations:
    >In chess you can go beyond that. It has happened that Kasparov had
prepared games up to 30 moves deep.
    >He's not doing that on his own. He has a whole team to do it with him.

Does that say anything about Kasparov's strength? With a good enough team,
one would need only to have good memory in order to play well... (of course,
the opponent should not know that and not play strange moves to create
confusion)

---------
I hoped that some kind of conclusion would magically make itself obvious by
writing these thoughts. Sorry, I got no better insight, but maybe someone
else will!

Happy playing and coding!
/Vlad