Semantics-Aware Program Sampling
Abstract
We present an algorithm for specifying and sampling from distributions over
programs via a perturb-max approximation. Prior work is generally limited to
sampling from priors over program syntax (for example, assigning lower probability
to longer programs). We demonstrate a simple modification to our sampling
algorithm that allows interpolation between such syntactic priors and priors over
semantics (that is, the behavior of the program on its inputs, or the set of class
labels when the program is viewed as a statistical model).