Neural networks of general recurrent type can display complex dynamics including bifurcation scenarios, co-existing oscilatory modes, chaos, etc. Our basic hypothesis states that these dynamical properties are the prerequisite for the existence of cognitive processes in biological and artificial brains. Starting with small artificial networks as multifunctional units, the assumption is that cognitive abilities will arise by mutual interaction of dynamic neuromodules. Cognitive systems adapt and act in a sensori-motor loop which sets the boundary conditions for the self-organized development of the underlying neural structures. Believing in cognition as an emergent process in complex adaptive systems, evolution seems to be the only way out of a situation, where there is no theory - up to now - to guide the construction of behavior oriented neurodynamical control systems. Autonomous robots provide a natural setting for testing these ideas on embodied cognition. Using a special Evolution-Simulation-Environment together with an Artificial Life approach to Evolutionary Robotics we generate modular neural networks for the control of various robot platforms. Some results and examples for behavior relevant dynamical effects will be presented.
Neural Networks, Autonomous robots, Cognitive systems