[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: A little Arithmetic
> If board configurations can be generalized, there should exit a learning
> system that can generalize these configurations. The real odd balls can be
> dealt by 9K programs. If board configuration can not be generalized, no
> learning scheme will work.
Generalizability is not a property of board positions per se, it is
a property of *representations* of board positions. Learning can only
formulate *short* statements about Go from a million games; to get
anywhere your hypothesis space must contain short statements that
generalize well, i.e., capture useful regularities about Go. The
primitive representation of board positions you proposed does not
contain such statements.
Example: "don't peep where you can cut" is a short statement that
generalizes well; given good definitions of "peep" and "cut" as
primitives (wherein much trickery of course lies buried), learning
such a rule is easy. Try to express this rule in terms of raw board
positions, and the universe is not big enough to hold its description,
let alone learn it.
- nic
--
Dr. Nicol N. Schraudolph
IDSIA, Galleria 2
CH-6928 Manno, Switzerland
http://www.idsia.ch/~nic/