BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Gradient-augmented supervised learning of feedback control laws fo
 r high-dimensional nonlinear dynamics - Dante Kalise (Imperial College Lon
 don)
DTSTART:20211117T120000Z
DTEND:20211117T123000Z
UID:TALK165427@talks.cam.ac.uk
DESCRIPTION:High-dimensional Hamilton-Jacobi-Bellman PDEs naturally arise 
 in feedback control synthesis for high-dimensional dynamics\, and their nu
 merical solution must be sought outside the framework provided by grid-bas
 ed discretizations. In this talk\, we discuss the construction of optimal 
 feedback laws for high-dimensional nonlinear dynamics circumventing the di
 rect numerical approximation of the HJB PDE. Our feedback law recovery is 
 cast in a supervised learning framework\, through the generation of a synt
 hetic dataset from samples of the HJB solution and its gradient. This grad
 ient-augmented formulation scales efficiently with respect to the dimensio
 n of the control system\, and is complemented with sparse optimization to 
 recover a feedback law of reduced complexity. We present different archite
 ctures for feedback recovery\, including polynomial approximation\, tensor
  decompositions\, and deep neural networks.\nThis talk is based on joint w
 orks with G. Albi\, B. Azmi\, S. Bicego\, S. Dolgov\, K. Kunisch and L. Sa
 luzzi.
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR