Scalable simulation-based inference using multifidelity methods

Overview

Bayesian inference is a popular statistical framework for estimating the parameters of statistical models based on data.  However, Bayesian methods are well known to be computationally intensive. This fact inhibits the scalability of Bayesian analysis for real-world applications involving complex stochastic models. Such models are common in the fields of cell biology, ecology and epidemiology.

Multifidelity estimators are a promising class of asymptotically unbiased estimators for dealing with the scalability challenges (Warne et al., 2022; Prescott et al., 2024). These approaches use approximate simulators and randomised bias corrections tuned to optimise the trade-off between computational cost, variance of the estimator.  The resulting estimators can be orders of magnitude more statistically efficient than alternatives.

Research activities

This project will involve:

• Implementing Bayesian methods for a challenging stochastic model from, biology, ecology or epidemiology.
• Implementing multifidelity samplers for Bayesian inference
• Evaluating the performance of multifidelity samplers numerically

Outcomes

The outcomes of the project include:

• multifidelity inference algorithms and software
• performance assessments of multifidelity methods.
• rules-of-thumb for configuration and tuning of multifidelity methods

Skills and experience

The following skills will be necessary:

• programming abilities (preferred languages are MATLAB, R, Python, or Julia).
• Introductory stochastic modelling and probability theory
• Understanding of statistical inference (classical or Bayesian) is also desirable.

Start date

1 November, 2024

End date

28 February, 2025

Keywords

• simulation-base inference
• likelihood-free inference
• stochastic simulation algorithms
• stochastic processes
• Monte Carlo methods

Contact

Contact the supervisor for more information.