Fri Dec 23 12:53:47 EST 2011
Is inference always generate-and-test?
More specifically, I'm thinking of the task "design somthing that has
these properties.". In general, there is no algorithm that goes from
a description of the properties to the description of something that
implements those properties (a logic and a model?).
I'm inclined to think (in a vague an unrooted way) that any form of
"algorithmic inference" is really just fixing parameters an already
existing / well-defined family of models, not *ever* the construction
of a model out of the blue.
Why is this important? Compilation of specifications. Is it so that
a model can be seen as a "constructive proof" of a set of properties?
Is this kind of thinking formalized in some way?