Institute of Computer Science III
University of Bonn
Databases * Information Systems * Software Engineering * Pattern Recognition * Image Processing * Artificial Intelligence * Robotics


Adaptive Steuerung von Strategieparametern bei genetischen Algorithmen

Andreas Bergmann

The practical application of systems based on evolutionary concepts is often complicated by the problem of tuning the system parameters. Depending on the parameter setting the evolutionary algorithm converges prematurely or makes a long term random walk through the solution space. A successful problem solving process is only guaranteed on the small edge between these extremes. This report gives an overview of genetic algorithms and evolution strategies and discusses several control techniques for their strategy parameters. A new method is presented applying fuzzy control rules for the dynamic guidance of a genetic algorithm. This approach allows the incorporation of available knowledge about reasonable parameter settings. The fuzzy control rules adjust the inversion, crossover and mutation rate to achieve a desired population diversity. Experimental results of a comparison with an (1+1)-evolution strategy demonstrate that fuzzy control allows better approximation of the target entropy and results in better performance of the underlying genetic algorithm.

Click here to obtain the full paper (PS, gzip, 328413 bytes, 48 pages, german language)


webmaster@www.informatik.uni-bonn.de - 16.12.05