Mario Negrello (Fraunhofer IAIS, Germany), Martin Huelse (University of Wales, UK) and Frank Pasemann (Fraunhofer IAIS, Germany)
Copyright: © 2008
Neurodynamics is the application of Dynamical system’s theory (DST) to the analysis of the structure and function of recurrent neural networks (RNNs). In this chapter, we present recurrent neural networks artificially evolved for the control of autonomous robots (Evolutionary Robotics), and further analysed within dynamical system’s tenets (Neurodynamics). We search for the characteristic dynamical entities (e.g. attractor landscapes) that arise from being-environment interactions that underpin the adaptation of animat’s (biologically inspired robots). In that way, when an efficient controller is evolved, we are able to pinpoint the reasons for its success in terms of the dynamical characteristics of the evolved networks. The approach is exemplified with the dynamical analysis of an evolved network controller for a small robot that maximizes exploration, while controlling its energy reserves, by resorting to different periodic attractors. Contrasted to other approaches to the study of neural function, neurodynamics’ edge results from causally traceable explanations of behavior, contraposed to just correlations. We conclude with a short discussion about other approaches for artificial brain design, challenges, and future perspectives for Neurodynamics.