ALABI: Active Learning for Accelerated Bayesian Inference

TL;DR

ALABI is an open-source Python package that accelerates Bayesian inference for expensive models by using active learning with Gaussian Process surrogates, significantly reducing the number of costly model evaluations needed.

Contribution

The paper introduces ALABI, a novel active learning framework that integrates Gaussian Process surrogates with various MCMC methods to efficiently perform Bayesian inference on high-dimensional, computationally expensive models.

Findings

Speeds up MCMC by 10-1000x for models with evaluation times >1s.

Effectively handles complex and high-dimensional posterior structures.

Demonstrates substantial efficiency gains across diverse test cases.

Abstract

We present Active Learning for Accelerated Bayesian Inference (\texttt{alabi}): an open-source Python package for performing Bayesian inference with computationally expensive models. Given a forward model and observational data to construct a likelihood and priors, \texttt{alabi}\ uses a Gaussian Process (GP) surrogate model trained to predict posterior probability as a function of input parameters, and employs active learning to iteratively improve GP predictive performance in high-likelihood regions where the GP is most uncertain. \texttt{alabi}\ provides a uniform interface for using Markov chain Monte Carlo (MCMC) with different packages, including the affine-invariant sampler \texttt{emcee}, and nested samplers \texttt{dynesty}, \texttt{multinest}, and \texttt{ultranest}. This approach facilitates accurate estimation of the desired posterior distribution, while reducing the number of computationally expensive model evaluations required by factors of thousands. We demonstrate the performance of \texttt{alabi}\ on a variety of test cases, including where inference is challenging due to complex posterior structure or high dimensionality. We show that \texttt{alabi}\ offers a substantial improvement for likelihood functions with evaluation times $\gtrsim 1$\,s, speeding up MCMC computations by a factor of $10-1000\times$ when tested on problems with up to 64 dimensions.