[go: up one dir, main page]
Published July 6, 2024 | Version v1
Software Open

Artifact for Programmable MCMC with Soundly Composed Guide Programs

Creators

  • 1. ROR icon Carnegie Mellon University
  • 2. ROR icon Peking University

Description

In probabilistic programming with a newly proposed coroutine-based programmable inference framework, the user provides (i) a model coroutine and (ii) a
sequential composition of guide coroutines. The model coroutine specifies a probabilistic model for Bayesian inference. Meanwhile, the sequential
composition of guide coroutines customizes the Block Metropolis-Hastings (BMH) algorithm, where we successively run the guide coroutines, each of which is
followed by an MH acceptance routine. Each guide coroutine only updates a subset (i.e., block) of random variables. The model and guide coroutines communicate
with one another by message passing, and their communication protocols are described by guide types.

This artifact is a program analysis tool for statically checking the soundness of a probabilistic program in this coroutine-based framework. The artifact
offers three functionalities:
  • Type-equality checking: check structural type equality of guide types.
  • Type inference: infer guide types of model and guide coroutines (using the first functionality for structural-type-equality checking)
  • Coverage checking: check whether the support of sequentially composed guide coroutines coincides with the support of a model coroutine.

Files

README.pdf

Files (1.0 GB)

Name Size Download all
md5:68d6618eb662c766033d26da442648b1
1.0 GB Download
md5:489275fdb653510b2149fb5180fd3ff8
273.8 kB Preview Download

Additional details

Software

Repository URL
https://github.com/LongPham7/GuideTypes/tree/subguide_types
Programming language
OCaml, Python