| |
Reinforcement Learning and Computer
Go (RLGO) |
| Features |
Complexity:
A concrete feature is a
feature that can be computed directly in terms of observations that we make about
the state or about state transitions. Vague notions cannot be
considered concrete features unless we have a specific way to compute
and verify the value for that notion in terms of observations about the
state.
An abstract feature is the
opposite of a concrete feature. It cannot be computed directly in terms
of observations, and represents a higher order notion that can only be
computed in terms of other features.
The order of a feature is
the level of abstractness. Concrete features are order 0. The order of
an abstract feature is 1 + the highest order of any features used to
compute it.
Both current and predictive features can be described by questions and
answers.
Scale:
A point feature is a property of a particular stone or intersection.
A compound feature is a
property of a particular region (for example a string, group, or
eyespace)
A global feature is a
property of the whole board (for example whether a ko-fight is
happening, or the estimated overall score)
Many features exist at multiple scales. See segmentation
for some discussion of different regions.
