Statically bounded-memory delayed sampling for probabilistic streams

Probabilistic programming languages aid developers performing Bayesian inference. These provide constructs and tools for probabilistic modeling automated Prior work introduced a language, ProbZelus, to extend functionality unbounded streams of data. This demonstrated that the delayed sampling inference algorithm could be extended in streaming context. ProbZelus showed while effectively deployed on some programs, depending model under consideration, is not guaranteed use bounded amount memory over course execution program. In this paper, we present conditions program’s which will execute memory. The two are dataflow properties core operations sampling: m -consumed property unseparated paths . A program executes if, only it satisfies properties. We propose static analysis abstracts these soundly ensure any passes properties, thus sampling.