Two-Phase Interaction-Kernel Learning for Stochastic Particle Systems
A data-driven method for recovering the unknown interaction kernel of a stochastic interacting-particle system from observed trajectories. Kernel density estimation converts trajectories into an empirical density obeying the mean-field equation; a first phase combines importance sampling with an adaptive threshold to identify the dominant kernel terms, and a second phase refines their coefficients on the full dataset, turning kernel identification into a tractable regression problem.
2501.00714
A stochastic interacting-particle system evolves many agents through pairwise drift and diffusion, and its collective behaviour is governed by an interaction kernel that is usually unknown. This pape…