EvoNeuro (Evolutionary Robotics and Computational Neuroscience cross-fertilization) is an academic research project (ANR-09-EMER-005) funded by the French research funding agency (ANR). It began the 1st of December 2009, for a duration of 39 months.
Evolutionary robotics (ER) uses algorithms inspired from biological evolution to design robots or software agents on the basis of a high level description of their mission. A lot of research is dedicated to the design of controllers able to exhibit a behavior with desired consequences. In this context, artificial neural networks formalism is the most commonly used for its versatility. Although many interesting behaviors have thus been generated, there is a clear limit in the kind of behaviors to be reached: reactive controllers are easy to get, but more cognitive ones are, up to now, impossible to generate.
Computational models of various brain regions have been proposed for more than thirty years. Computational neuroscience (CN) is the natural complement of experimental brain research, as they help formulating and testing hypotheses about whole mechanisms which are only partially observed using anatomy, physiology or behavior. Moreover, they can be used to make predictions that can in return be tested experimentally. The design of such models is a creative process constrained by experimental data.
The EvoNeuro project aims at exploring the frontiers between ER and CN. On the one hand, the neuromimetic metaphor in ER of neural networks remains quite poor with regard to the wealth of knowledge and know-how researchers in computational neuroscience daily use when designing models of the brain. The only imported concept is the use of artificial neuron models, while the concept- and tool-boxes of CN scientists are much richer than that. On the other hand, there have been very few attempts to use ER for the design of CN models and it was limited to parameter optimization rather than to structural design where the full potential of ER could be used. The process of computational model design and tuning is often a matter of trials and errors, because of the limitations or the lack of experimental data. It could benefit from more systematic explorations using ER techniques.
The project is thus built around two interacting activities: the development of new ER methods by the incorporation of more neuroscience, and the proof of concept of using ER within computational neuroscience, by managing two case-studies in well-established topics (action selection and active exploration in spatial cognition). Finally, experimental neuroscience studies will be necessary to test and validate the resulting models.
- ER: Evolutionary Robotics
NCNR: Computational Neuroscience & Neuro-Robotics