About brute force and knowledge...

[Antti Huima <huima@xxxxxx>]

Concerning the question about brute-force search versus knowledge, I have
a hypothesis. Let P be a/the principal variation of optimal play on a
19x19 go board, i.e. a/the canonical game that represents the true
game-theoretical value of the empty board. The hypothesis is that P may
look unreasonable or ununderstandable even to a highest-ranking dan
player.

While this hypothesis might not be very interesting, I state it in order
to present the following note.

The playing style of go has developed under thousands of years. It is
conceivable that people have preferred moves that leave less options and
variations of play for the opponent. It has been stated that "simple moves
win" or something like that. I doubt that the reason were in the internal
structure of the game itself; instead, the reason can be that the more
complex variations you initiate, the more probable it is that either
player will plumber, and thus the game becomes more random for human
players. Playing _very_ complex moves causes a _huge_ amount of variations
to initiate, and a human brain cannot cope with such a chaotic state-space
analysis problem.

Therefore, things like 'building thickness' and 'making good shape' are
not only heuristics, but they are tools human players use to reduce parts
of the state space of the game to a manageable size. 'Removing aji'
means essentially performing the same thing. After 'aji' has been
'removed', opponent plays near the region that was enforced can be ignored
in the hypothetical state-space search.

Thus, it is possible that the reason why go programs should incorprate go
knowledge, understanding of shape for example, is not that it would lead
to perfect or better play, but that it would reduce the state-space to
search (by brute force if you wish) to a space considered by human
players.

The point is, it can be that humans play go after all in a very simplified
way. Horrible positions can perhaps be constructed where best players
don't have a clue what to do, but they never arise in a game, because
humans prefer to simplify things.

Thus, perhaps the complexity of go as a game-three searching problem is
not so overwhelming as it looks, if we just take it for granted that
humans usually play only moves of certain type, usually explainable by
some heuristic concepts instead of an enumeration of 10^10 variations.

I welcome any comments, preferably posted to this list and not directly to
me.

Yours,

-- 
Antti Huima
SSH Communications Security Oy