On-line evolution of robot controllers by an island evolution approach

From Master Projects
Jump to: navigation, search

has title::On-line evolution of robot controllers by an island evolution approach
status: finished
Master: project within::Computational Intelligence and Selforganisation
Student name: student name::Rachid Siallioui
Start start date:=2012/01/10
End end date:=2012/07/31
Supervisor: Evert Haasdijk
Second reader: has second reader::Gusz Eiben
Thesis: has thesis::Media:Thesis.pdf
Poster: has poster::Media:Posternaam.pdf

Signature supervisor



The European research project Symbrion focuses on the development of novel adaptation and evolution principles for multi-robot organisms. These robot organisms have to evolve autonomously their controllers on-line and on-board to adapt to different environments and tasks. Besides the traditional approach - with off-line, off-board, and centralised evolution - three classifications of embodied evolution are offered in recent research: the encapsulated evolution approach, where robots run the evolution algorithms locally and perform fitness evaluations autonomously; the distributed evolution approach, where each robot has a single genotype and is controlled by the corresponding phenotype and by exchanging genetic information the evolution process emerges; and finally the combination of the early two approach, resulting in parallel evolution algorithms running on different robots, where each robot is an island and genetic information is exchanged through intra-island variation and inter-island migration. In this paper we study the latter approach and experiment with this method on a multi-robot organism in robot simulator Webots. We study with who robots should communicate (e.g. neighbour, best performing one) and what individual they should communicate (e.g. average, best, multi-kulti). By measuring the number of (successful) migrations, the distance traveled by the multi-robot organism, and the population diversity and by comparing these results with similar results on the encapsulated approach we try to get an indication of the best of the two evolution approaches.