Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

BP: Fachverband Biologische Physik

BP 20: Poster IIIa

BP 20.16: Poster

Wednesday, March 20, 2024, 11:00–14:30, Poster B

Efficient Bayesian Inference of Active Brownian Motion using Reinforcement-learned Brownian Bridges — •Sascha Lambert and Stefan Klumpp — Georg-August-Universität Göttingen, Institut für Dynamik Komplexer Systeme

Active Brownian Particles (ABPs) describe microswimmers by a Langevin equation that combines stochastic fluctuations and deterministic forces, including the swimmer's active propulsion and the interactions with obstacles.

Full Bayesian inference of the parameters of the Langevin equation can be achieved within the framework of Particle Pseudo-Marginal Metropolis-Hastings algorithms (PMMH). These algorithms sample from the full posterior of the model parameters, conditioned on experimental observations. A key strategy in these algorithms is using a particle filter for resampling the trajectories between observations.

The sampling efficiency is largely determined by uncertainties of the experimental measurements, with lower uncertainties resulting in reduced acceptance rates of newly generated samples. The rotational degrees of freedom and the resulting memory of ABPs produce strong temporal correlations that are not easily resolved using particle filters, as they only filter based on individual observations. We construct an approximative Brownian Bridge to increase the particle swarm's coherence and to guide the swarm between the observations. This approach increases the resampling efficiency by several orders of magnitude.

The exact guidance parameters depend on the full non-linear system and can be learned through reinforcement learning.

Keywords: Active Brownian Motion; Bayesian Statistics; Machine Learning