| Title | A Specification of a Control System for Adaptive Time Stepping |
| Authors | Gustaf Söderlind |
| Year | 2005 |
| Document type | Report |
| Status | Submitted |
| Language | eng |
| Abstract English | To increase their computational efficiency, time-stepping methods for the solution of ODEs and DAEs are usually adaptive. A control system for the adaptivity includes the automatic control of step size as well as method order, and can be constructed using proven design principles from control theory. It is a ''superstructure,'' surrounding the basic method, that manages the computational process.<br> <br> As the control system is a dynamical system in its own right, its properties need to be properly analyzed and tested, motivating a separate study of control system dynamics. Here we propose a complete specification of a control system for adaptive time-stepping, based on controlling local error estimates. The control system is specified in Matlab to facilitate ease of implementation and is executable on its own. The entire control system is independent of the basic discretization method, and can be tested stand--alone, without any specific ODE or DAE method.<br> <br> With respect to step size control, a collection consisting of PI controllers and digital filters is included. For<br> controlling the order, error estimates for two alternative methods of different orders are monitored. The estimates are fed to the same step size controller, which will produce different step size suggestions to estimate relative efficiency. The order change is based on simple time integration of the efficiency measure. Anti-windup is used to counteract excessive step size increases for method orders currently not in use. We further discuss the choice and handling of the norm used for measuring local errors. |
| Keywords | adaptivity control theory digital filters error control stepsize control, |
Questions: webmaster
Last update: 2013-04-11
Centre for Mathematical Sciences, Box 118, SE-22100, Lund. Telefon: +46 46-222 00 00 (vx)