Johannes P. Schlöder, Interdisciplinary Center for Scientific Computing (IWR), Heidelberg University
Monday, November 26, 2018, 11:00
"Room 01-012, Georges-Köhler-Allee 102, Freiburg 79110, Germany"
Models of dynamic processes (e.g., initial value, boundary value or optimal control problems with ODE, DAE or PDE) typically contain some parameters that have to be determined from measurements of the process behavior. Good parameter values are indispensable for predictive models.
The talk discusses properties of parameter estimation problems and presents numerical formulations for several model classes, types of measurements, and measurement errors.
Derivative-based numerical optimization methods in the framework of direct multiple shooting methods combined with generalized Gauss-Newton methods are described. The main properties and the performance of these methods are discussed, including convergence behavior, regularization aspects, real-time adaptation, assessment of the statistical quality of the solution and implications for experimental design.
Numerical results including applications from biology, environmental science, aerospace and gait analysis demonstrate the wide range of application of the methods.