Speaker
Mr
Sören Geffken
(Universität Bremen)
Description
The goal of this talk is to present current research activities aiming at improved efficiency and stability within the ESA NLP-Solver WORHP. It is designed to solve high dimensional sparse non-linear optimisation problems.
The underlying SQP method is inherently sequential, therefore parallelism cannot be exploited
straightforwardly.
In order to obtain the best solver performance the parameter configuration should be adapted accordingly for every problem itself. An approach running several solver instances using different parameter sets in parallel has been developed and proven highly beneficial on a given set of problems. The First-Across-The-Line approach stops all instances when the first local optimum has been found, thus improving the solver's speed and stability as well. Furthermore, the approach allows the user to experiment with specialised algorithms within the optimisation as the threads using basic parameter settings serve as safeguards guaranteeing the solver to converge as usual.
In order to improve the solver's efficiency for one special problem, the new operational mode can be used to automatically attune the solver's parameters accordingly. Again the solver is started with several instances at once, but this time the Best-Of-All mode is used in order to obtain the best local optimum and the corresponding parameter settings. Additionally, the mode enables users to perform parameter sweeps to further improve the solver's configuration.
The results presented show the improvement of the solver's performance on the state-of-the-art CUTEst test set that has been solved faster and with more optimal solutions found compared to the traditional single-core approach. The parameter attunement is applied to specific single problems from the test set as well.
| Applicant type | First author |
|---|
Primary author
Mr
Sören Geffken
(Universität Bremen)
Co-author
Prof.
Christof Bueskens
(Universität Bremen)
