Abstract

Hamiltonian Monte Carlo (HMC) is arguably the dominant statistical inferencealgorithm used in most popular "first-order differentiable" ProbabilisticProgramming Languages (PPLs). However, the fact that HMC uses derivativeinformation causes complications when the target distribution isnon-differentiable with respect to one or more of the latent variables. In thispaper, we show how to use extensions to HMC to perform inference inprobabilistic programs that contain discontinuities. To do this, we design aSimple first-order Probabilistic Programming Language (SPPL) that contains asufficient set of language restrictions together with a compilation scheme.This enables us to preserve both the statistical and syntactic interpretationof if-else statements in the probabilistic program, within the scope offirst-order PPLs. We also provide a corresponding mathematical formalism thatensures any joint density denoted in such a language has a suitably low measureof discontinuities.

 

Introduction (beta)

None

 

Conclusion (beta)

None