Re: computer-go: A little Arithmetic

["Nicol N. Schraudolph" <nic@xxxxxxxx>]

> and there. If considering 0.5-point difference is a tie (or 2.5-point, what
> ever), otherwise a win or loss, the game only has three output: win, tie,
> and loss, not a 25-point swing case. If this is a-priori, will this hurt?

Doesn't make a difference: we now have 3^(3^361) instead of 50^(3^361)
hypotheses, so you'll now need 50/3 times less game records.  The only
things that helps is to reduce the huge exponent (3^361) by employing
more sophisticated representations.

> As a chemical reaction based computer (human brain) can barely scan couple
> thousands sample data with bunch of rule-of-thumb to figure out a 9K

The brain uses 10^10 - 10^11 processing elements in parallel, in ways
which are still mysterious and baffling to us.  If we really understood
how the brain worked (beyond low-level sensorimotor processing), dan-level
Go programs might well be a trivial consequence.

> evaluation formula, or be a supper player (e.g., 9K+), 361 modern computers
> supposed to check all games human ever played would not have a single
> solution, what kind of model, architecture, or induction we were using?

We learn to play Go well by employing very, very sophisticated inductive
bias.  Humans think about Go in terms of shape, influence, aji, strength,
life&death, sente, urgency, tradition, intuition, etc.  And underneath it
all we have (as Vlad points out) excellent pattern recognition built right
into the hardware.  All this is the byproduct of billions of years of
compute time (evolution) on a massively parallel machine (earth).

Unless one is prepared to invest a comparable amount of CPU cycles,
one cannot expect a learning machine to reach similar sophistication
when starting from nothing.  Machine learning is a useful computational
technique, not a deus ex machina.

Best wishes,

- nic

--
    Dr. Nicol N. Schraudolph 
    IDSIA, Galleria 2       
    CH-6928 Manno, Switzerland
    http://www.idsia.ch/~nic/