Action learning scores the plausibility of entire trajectories from observed data

The team says action learning scores whole trajectories without known forces, a potential energy function or simulator likelihood. They report detecting a hidden 0.01% perturbation in the Sun’s mass from 31-body orbital trajectories and reusing one learned molecular action for trajectory completion, steering, constrained optimization and temperature transfer. A coauthor says the learned objective can incorporate new goals and constraints without retraining.