Published: 2010 September
Buchtitel: Proceedings of the Fourth IEEE International Conference on Self-Adaptive and Self-Organizing Systems (SASO 2010)
Verlag: IEEE Computer Society
Artificial evolution has successfully been utilized to train robotic controllers in various types of applications and scenarios. However, for different behaviors to be trained usually different evolutionary setups have to be designed which is a costly procedure. For many current real-world problems in robotics, the effort for finding a good evolutionary setup is still big and programming by hand is a more feasible approach. In this paper, a completely evolvable genotype-phenotype mapping (ceGPM) is studied with respect to its capability of improving the flexibility of artificial evolution. By letting mutation affect not only controller genotypes, but also the mapping from genotype to phenotype, the future effects of mutation can change over time. That makes the training process highly adaptable to the behavior being trained during the training itself. By this means, the need for prior parameter adaptation can be reduced. Experiments with the presented approach in a simulated swarm of mobile robots using controllers based on finite state machines indicate that the ceGPM is capable of robustly adapting to a benchmark behavior. A comparison to a related approach shows significant improvements in evolvability according to a proposed measure.
Download: Media:Final version.pdf
Ein im Beitrag referenziertes Bild ist hier verlinkt: http://www.aifb.kit.edu/images/6/6b/Utrans.pdf