LEAP learns actions from trajectory observations
Original title: Action Learning for Scientific Simulations
LEAP anchors unknown actions to reference processes for reusable objectives in scientific simulations.
Understanding system evolution requires reasoning about complete trajectories in science and engineering. Existing methods miss nonadditive dependencies across time. LEAP recovers stochastic actions from observations by anchoring to a tractable reference process and learning corrections. The method applies to physical systems, orbital dynamics, and molecular simulations for inference, optimization, and control.
- LEAP yields differentiable objectives on path space
- Reusable for path ranking, steering, and constrained optimization
- Captures local and nonadditive temporal dependencies
- Demonstrated on stochastic physical systems and simulations
From the original article
1 min read at arxiv.orgUnderstanding and controlling how systems evolve is central to science and engineering, where many problems require reasoning about complete trajectories.
By Sarah Liaw · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.