| Autor(en): |
Alexandru-Ciprian Zavoianu, Edwin Lughofer, Gerd Bramerdorfer, Wolfgang Amrhein und Erich Peter Klement |
| Journal: |
SYNASC 2013, 15th International Symposium on Symbolic and Numeric Algorithms for Scientific Computing |
| Jahr: |
2013 |
| Zusammenfassung: |
The task of designing electrical drives is a multi-objective optimization problem (MOOP) that remains very slow even when using state-of-the-art approaches like particle swarm optimization and evolutionary algorithms because the fitness function used to assess the quality of a proposed design is based on time-intensive finite element (FE) simulations. One straightforward solution is to replace the original FE-based fitness function with a much faster-to-evaluate surrogate. In our particular case each optimization scenario poses rather unique challenges (i.e., goals and constraints) and the surrogate models need to be constructed on-the-fly, automatically, during the run of the evolutionary algorithm. In the present research, using three industrial MOOPs, we investigated several approaches for creating such surrogate models and discovered that a strategy that uses ensembles of multi-layer perceptron neural networks and Pareto-trimmed training sets is able to produce very high-quality surrogate models in a relatively short time interval. |
