Automating Incremental Inference for Probabilistic Programs

Abstract

This thesis proposes a method that leverages incremental inference to improveinference efficiency in complex probabilistic programs, providing an algorithm-agnostic approach that does not rely on any single sampling method. The methodbuilds on an existing incremental inference framework, which samples from oneprobabilistic program and uses the results to perform inference on a related pro-gram. Our approach uses this framework to sample from simplified programs,thereby bypassing direct inference on complex programs. These simplificationsare constructed by automatically detecting certain program patterns, both struc-tural and parametric, that increase program complexity, and applying correspond-ing changes to simplify them. On a set of input programs, we empirically showwhich patterns increase complexity and demonstrate how the constructed changesachieve more efficient inference than baseline sampling methods.