Environment-driven Open-ended Evolution with a Population of Autonomous Robots

Nicolas Bredeche (1, 2) , Jean-Marc Montanier (1, 2)

1 LRI - Laboratoire de Recherche en Informatique
2 TAO - Machine Learning and Optimisation

This paper summarizes work done since 2009 on running swarm of autonomous robots with Environment-driven Dis- tributed Evolutionary Adaptation algorithms (EDEA). The motivation behind this work is to design algorithms that can cope with unknown environmental pressure, that is to learn (through evolution) efficient survival strategies in order to ad- dress a priori unknown constraints existing in the environ- ment. We are concerned with a fixed-size population of au- tonomous robotic agents facing unknown, possibly changing, environments, which may require the robots to evolve partic- ular behaviors such as cooperative behaviors, specialized be- haviors, etc. We describe a particular flavor of EDEA, termed minimal EDEA (mEDEA), and summarize results obtained so far, both with real robots and in simulation. This algo- rithm is shown to be both efficient in unknown environment and robust with respect to abrupt, unpredicted, and possibly lethal changes in the environment. Moreover, the ability of mEDEA to evolve various types of behavior, including coop- erative and specialized behaviors, is summarized.

Domaines: Intelligence artificielle [cs.AI]

Nicolas Bredeche, Jean-Marc Montanier. Environment-driven Open-ended Evolution with a Population of Autonomous Robots. Evolving Physical Systems Workshop, 2012, East Lansing, United States. ⟨hal-00731422⟩