Liesel: A Probabilistic Programming Framework for Developing Semi-Parametric Regression Models and Custom Bayesian Inference Algorithms

TL;DR

Liesel is a flexible probabilistic programming framework that facilitates semi-parametric regression modeling and custom Bayesian inference using MCMC, with a modular design and integration of modern machine learning tools.

Contribution

It introduces a novel framework combining R and Python interfaces with graph-based model building and modular MCMC algorithms, enabling advanced Bayesian research.

Findings

Supports complex semi-parametric regression models

Enables customization of inference algorithms

Leverages modern hardware and automatic differentiation

Abstract

Liesel is a new probabilistic programming framework developed with the aim of supporting research on Bayesian inference based on Markov chain Monte Carlo (MCMC) simulations in general and semi-parametric regression specifications in particular. Its three main components are (i) an R interface (RLiesel) for the configuration of an initial semi-parametric regression model, (ii) a graph-based model building library, where the initial model graph can be manipulated to incorporate new research ideas, and (iii) an MCMC library for designing modular inference algorithms combining multiple types of well-tested and possibly customized MCMC kernels. The graph builder as well as the MCMC library are implemented in Python, relying on JAX as a numerical computing library, and can therefore benefit from the latest machine learning technology such as automatic differentiation, just-in-time (JIT) compilation, and the use of high-performance computing devices such as tensor processing units (TPUs). Liesel provides all required tools for efficient and reliable statistical research on complex models and estimation algorithms. Its modular design allows users to expand the model library and inference algorithms, offering the flexibility and customization options to tailor the software to any specific research needs.