SPUX: Scalable Particle Markov Chain Monte Carlo for uncertainty quantification in stochastic ecological models

TL;DR

SPUX introduces a parallelized Particle MCMC framework that significantly accelerates uncertainty quantification in complex ecological models by efficiently distributing computational load across multiple units.

Contribution

It presents SPUX, a Python-based scalable Particle MCMC method with adaptive load balancing for efficient Bayesian inference in stochastic ecological models.

Findings

Achieves significant speed-ups in model calibration.

Effectively manages computational load during re-sampling.

Demonstrates applicability to predator-prey models.

Abstract

Calibration of individual based models (IBMs), successful in modeling complex ecological dynamical systems, is often performed only ad-hoc. Bayesian inference can be used for both parameter estimation and uncertainty quantification, but its successful application to realistic scenarios has been hindered by the complex stochastic nature of IBMs. Computationally expensive techniques such as Particle Filter (PF) provide marginal likelihood estimates, where multiple model simulations (particles) are required to get a sample from the state distribution conditional on the observed data. Particle ensembles are re-sampled at each data observation time, requiring particle destruction and replication, which lead to an increase in algorithmic complexity. We present SPUX, a Python implementation of parallel Particle Markov Chain Monte Carlo (PMCMC) algorithm, which mitigates high computational costs by distributing particles over multiple computational units. Adaptive load re-balancing techniques are used to mitigate computational work imbalances introduced by re-sampling. Framework performance is investigated and significant speed-ups are observed for a simple predator-prey IBM model.