Evolutionary Optimization for Problem Classes with Lamarckian Inheritance
Michael Hüsken and Bernhard Sendhoff
The combination of evolution and learning as two natural principles of optimization, which work on different time scales, has been shown to be very efficient for optimization tasks. We investigate how this combination should be organized to achieve networks that can fast and reliably switch between related problems from one and the same class of problems during operation time. We analyze different methods to evaluate the network's performance during optimization, namely a time averaged evaluation and a more direct averaging over different problems in each generation. In particular, we show that a specific kind of Lamarckian inheritance can be beneficial for the evolutionary optimization of neural networks even for dynamic environments like problem classes.
send email to authors

Last update 29-07-2000, all inquiries to Bernhard Sendhoff