Paper presented at Evostar 2018 Parma Italy

Authors: S. Engebråten, J. Moen, O. Yakimenko and K. Glette (sondre.engebraten (a) ffi.no)

Automated design of swarm behaviors with a top-down approach is a challenging research question that has not yet been fully addressed in the robotic swarm literature. This paper seeks to explore the possibility of using an evolutionary algorithm to evolve, rather than hand code, a wide repertoire of behavior primitives enabling more effective control of a large group or swarm of unmanned systems. We use the MAP-elites algorithm to generate a repertoire of controllers with varying abilities and behaviors allowing the swarm to adapt to user-defined preferences by selection of a new appropriate controller. To test the proposed method we examine two example applications: perimeter surveillance and network creation. Perimeter surveillance require agents to explore, while network creation requires them to disperse without losing connectivity. These are distinct application that have drastically different requirements on agent behavior, and are a good benchmark for our swarm controller optimization framework. We show a performance comparison between a simple weighted controller and a parametric controller. Evolving controllers allows for specifying desired behaviors top-down, in terms of objectives to solve, rather than bottom-up.

Paper can be found at https://link.springer.com/chapter/10.1007/978-3-319-77538-8_49

Engebråten, S. A., Moen, J., Yakimenko, O., & Glette, K. (2018, April). Evolving a Repertoire of Controllers for a Multi-function Swarm. In International Conference on the Applications of Evolutionary Computation (pp. 734-749). Springer, Cham.

Source code licensed under GLPv3