Machine Learning, Dynamical Systems and Control



Much of the power of modern AI derives from our ability to fit highly expressive models to large sets of high-dimensional data, by numerical optimization. Physics and dynamics-informed AI poses additional challenges beyond the standard issues of scale and dimensionality: the goal is to identify accurate physical models that support downstream control and decision making with guaranteed performance. This goes well beyond the standard demands of ML/AI, where the optimization goal is often just to accurately label a given training set. Physical data are structured: individual signals are often sparse in an appropriate basis, batches of signals are often low-rank, while data generated by systems with a few important underlying degrees of freedom typically concentrate near low-dimensional manifolds. From a computational perspective, correctly inferring these low-dimensional models in noisy, incomplete or otherwise unreliable data typically requires us to solve a high-dimensional nonconvex optimization problem, to global optimality! Clearly this requires new perspectives, since efficient nonconvex optimization is impossible in the worst case. Our team will innovative new algorithms and computational schemes to address the emerging challenges posed for optimization in the context of complex dynamical systems.




Stacks Image 5
Krithika Manohar, Thrust Lead: Optimization and Sensors
Assistant Professor

Website: [ VIEW ]

Department of Mechanical Engineering


Research: Data-driven modeling, sensors & sensor placement, dynamical systems & machine learning



BELOW: Manohar et al, Optimized sampling for multiscale dynamics SIAM MMS 17, 117 (2019).