The ambition of this page is
to
define and discuss the concept of shape in the context of
reinforcement
learning.
A useful definition of shape is the following:
Shape is the set of pattern features that help to predict
features
of the Go board.
In other words, shape can be thought of as a set of auxiliary features
that
can help us to answer questions.
Local shape is the set of local
pattern features that help to predict local features of the Go board,
relative to a particular intersection. For example this could be a set
of template patterns of the form { Black | White | Empty | Don't Care }
for a fixed scope about a point. This corresponds roughly to the
concept of a pattern database in
traditional Computer Go development.
However, shape is not restricted to local shape. We could also consider
more abstract concepts of shape, corresponding to patterns over higher
order features. For example, we may be interested in the shape of a set
of groups, which could be defined as patterns over features of those
groups.