ETH - Intelligent Control Systems Group
Tuesday, June 21, 2022, 11:00
Hybrid: Seminar room building 102, first floor + Online with ZOOM
By learning the model error from data, as well as providing a quantification of the residual uncertainty, Gaussian process-based model predictive control (GP-MPC) has exhibited impressive performances in real-world applications, receiving considerable attention in the learning-based control community. GP-MPC employs a linearization-based covariance matrix propagation to approximate the probabilistic chance constraints in the stochastic optimal control problem (OCP). However, the number of optimization variables corresponding to the covariance propagation constitutes a major factor prohibiting its when facing strong limitations on the computational budget. In this talk, we tackle this problem by making use of an inexact Newton-type method, decoupling the covariance propagation from the solution of the optimal control problem (OCP). By eliminating the corresponding variables from the OCP, as well as saving on computationally expensive GP evaluations, the computational requirements are greatly reduced. We illustrate properties of the resulting algorithm at the hand of a simple numerical example and conclude with a discussion of the algorithm's asymptotic convergence properties.
Room 102 SR 01-012(16)
Zoom login for online participants:
Meeting-ID: 627 9173 7415