Sequential Monte Carlo with transformations | JoVE Visualize

Area of Science:

Computational Statistics and Bayesian Inference
Sequential Monte Carlo Methods
Bayesian Model Comparison

Background:

Traditional Bayesian inference can be computationally intensive, especially for complex models and large datasets.
Sequential updating of posterior distributions is crucial for real-time data analysis and adaptive modeling.
Challenges exist in performing inference on probability distributions residing in spaces of varying dimensions.

Purpose of the Study:

To develop a novel methodology for sequential Bayesian inference on a sequence of posteriors in spaces of different dimensions.
To introduce an efficient particle transition mechanism using deterministic transformations for inter-dimensional posterior movement.
To create a flexible and general algorithm for Bayesian model comparison applicable in low-acceptance rate scenarios.

Main Methods:

Sequential Monte Carlo (SMC) samplers are employed as the core algorithmic framework.
Deterministic transformations are innovatively used to facilitate particle movement between target distributions of differing dimensions.
Adaptive methods are integrated to enhance the algorithm's efficiency and generality.

Main Results:

A highly flexible and general algorithm for Bayesian model comparison is presented.
The proposed method demonstrates effectiveness in applications where reversible jump Markov chain Monte Carlo (RJMCMC) exhibits low acceptance rates.
Successful application of the methodology to model comparison for mixture models and sequential inference of coalescent trees.

Conclusions:

The developed sequential Monte Carlo approach offers a powerful alternative for Bayesian inference and model comparison, particularly in high-dimensional or complex scenarios.
The use of deterministic transformations provides an effective means for particle propagation across differing dimensional spaces, overcoming a key limitation.
The algorithm's adaptability and demonstrated success in mixture models and phylogenetic inference highlight its broad applicability in data-driven scientific research.