Re: computer-go: Learning from existing games

[Heikki Levanto <heikki@xxxxxx>]

On Mon, Jan 13, 2003 at 03:56:32PM -0500, Don Dailey wrote:
> [...] I  think that  even  moves themselves could be "judged" (very weakly
> and inaccurately of course) by the final results of the game.  I would
> never recommend this as a way to proceed but it's certainly the case that
> the moves of the winning side are better than the moves of the loser.

First of all, I agree with your general observations that judging moves is
not the way to go, but that (at some level) the winner must have played
better moves.

That said, I must say that that matter is more complex than what we (or me
at least) seem to be assuming. After all, some moves do have definitive
values. This is taught to every go player, at least when studying yose
(endgame). There we have schoolbook examples that a hane-and-block is worth
three points in sente, or some other move is worth five points in gote,
provided that white has enough ko-threats.  The valuation tends to be not
only numerical, but require other considerations as well - sente/gote being
the most obvious, but various ko's complicate the matter, as do the
questions like who may choose between a large seki and a small uncoditional
life... Still, moves do seem to have values. 

Even in the beginning. I have seen a pro claiming somewhere that ifyou have
a 3-4 stone, and instead of the normal 5-3 shimari, you play closer at 4-3,
you loose two points (not sure of the exact value). So, at some level people
do value individual moves. 

Of course such values beg the question "compared to what" - obvious
references are to the best available move (like when estimating the cost of
joseki mistakes), or to not playing at all (like in some yose), or ...

Conversely, the idea of evaluating positions has its problems as well. One
can estimate the final score, but in the end it does not matter if I win by
a point or by a hundred. Another way to estimate positions would be the
probability of winning from a given position, which makes more sense when
making decisions that involve risks (should I invade or try to kill). Even
these estimates are seldom one-dimensional numbers, but hinge on
probabilities, strategies, and possibilities...


[...] 
> I doubt  thinking in terms of  moves instead of positions  can get you
> anywhere, but temporal difference learning is the same idea applied to
> positions and has been successful in many games.

True. Just to confuse matters, we can always define the value of a move to
be the difference of the values of the position before and after the move,
so we can use all position-evaluation tricks to evaluate moves...

Anyway, in any given go position, there are a small number of reasonable
moves that can be considered, and a great number of totally stupid moves
that no human would even consider. Any go program must be able to discard
the 90% of rubbish moves quickly. Here a simple move evaluation could make
sense. (especially since the same move may be totally irrelevant or the best
possible, depending if there is anything more valuable around...)


Best regards

	Heikki

P.S. Difficult game, isn't it? That's why we find it so fascinating!

-- 
Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>