L. Markolf, O. Stursberg: "Tailored Transformations for Constraint Satisfaction in Learning of Parametric Controllers for Nonlinear Systems". 10th Int. Conf. on Control, Decision, and Information Technologies, 2024.
Abstract
The paper on hand addresses the problem of guaranteed satisfaction of ellipsoidal state and input constraints when learning parametric controllers for discrete-time nonlinear but control-affine systems with additive bounded disturbances. The parametric controllers are proposed to be compositions of parametric functions (like neural networks), for which state-dependent transformations lead to ranges that are equal to sets of admissible control inputs derived from offline computed reachable state sets. In this way, constraint satisfaction is guaranteed for all control laws that can be learned through the choice of the parameter vector.
BibTex
@article{markolf2024tailored,
abstract = {The paper on hand addresses the problem of guaranteed satisfaction of ellipsoidal state and input constraints when learning parametric controllers for discrete-time nonlinear but control-affine systems with additive bounded disturbances. The parametric controllers are proposed to be compositions of parametric functions (like neural networks), for which state-dependent transformations lead to ranges that are equal to sets of admissible control inputs derived from offline computed reachable state sets. In this way, constraint satisfaction is guaranteed for all control laws that can be learned through the choice of the parameter vector.},
author = {Markolf, L. and Stursberg, O.},
journal = {10th Int. Conf. on Control, Decision, and Information Technologies},
title = {Tailored Transformations for Constraint Satisfaction in Learning of Parametric Controllers for Nonlinear Systems},
year = 2024
}
URL