Reproduce “BALLAST: Bayesian Active Learning with Look-ahead Amendment for Sea-drifter Trajectories under Spatio-Temporal Vector Fields” on OpenResearch

About this paper

Rui-Yang Zhang, Lachlan Astfalck, Edward Cripps, David S. Leslie, Henry B. Moss

A Bayesian active learning framework, BALLAST, was developed to optimize the placement of Lagrangian observers for inferring time-dependent vector fields by simulating future trajectories. This approach, enabled by the efficient VaSE Gaussian Process inference method, demonstrated up to 22% cost savings in drifter deployments and consistently reduced posterior predictive error compared to existing strategies.

Key insights

• Standard EIG active learning is insufficient for Lagrangian observers, as it performs worse than uniform placement by failing to account for the continuous data collection along a drifter's dynamic trajectory.
• Explicitly simulating future trajectories using an ensemble of plausible vector fields (the 'look-ahead' mechanism) is crucial for effective active learning with mobile sensors, providing a more informed utility estimation.
• The computational bottleneck of spatio-temporal GP posterior sampling can be overcome by synergizing standard GP methods (for initial conditions from non-gridded data) with SPDE-based state-space models (for temporal evolution), making complex active learning tractable.

Summary by alphaXiv.