The Role of Non-linearity for Evolved Multifunctional Robot Behavior Martin H¨ ulse, Steffen Wischmann, and Frank Pasemann Fraunhofer Institute, Autonomous Intelligent Systems, 53754 Sankt Augustin, Germany {martin.huelse, steffen.wischmann, frank.pasemann}@ais.fraunhofer.de http://www.ais.fraunhofer.de/~INDY Abstract. In this paper the role of non-linear control structures for the development of multifunctional robot behavior in a self-organized way is discussed. This discussion is based on experiments where combina- tions of two behavioral tasks are incrementally evolved. The evolution- ary experiments develop recurrent neural networks of general type in a systematically way. The resulting networks are investigated according to the underlying structure-function relations. These investigations point to necessary properties providing multifunctionality, scalability, and open- ended evolutionary strategies in Evolutionary Robotics. 1 Introduction Evolutionary robotics (ER) as the study and development of behavioral con- trol for autonomous robots through self-organizing processes based on artificial evolution is a widely accepted approach [10,14]. With respect to natural evo- lution and simplest forms of natural life there are many researches criticizing the dissatisfying outcomes of current work in ER [4,5]. In [5] it is claimed that open-ended evolutionary processes are necessary to overcome crucial limitations of current ER models and to generate more complex and interesting results. However ER models providing open-ended evolutionary processes are imple- mented, the agents must be incrementally evolved. With respect to behavioral control this means control structures must facilitate incremental evolution. Such an approach should also cope with the scalability problem of ER models in general [1,3,4]. The crux of incremental control structure evolution is the integration of new behavioral functionality without loosing existing capabilities. In this paper we propose an approach to make this problem more tractable. We present incremen- tally evolved control structures which are systematically investigate to study the underlying dynamical properties and control principles providing (1) coordina- tion of different behavioral tasks, and (2) the development of multifunctionality. In [2] it is claimed that a serious and systematical analysis of concrete examples of evolved agents are the prerequisite for dynamical explanation and ”abstract- ing” general principles” of situated autonomous agents. Therefore we present robotic tasks which at first might seem rather simple, but this simplicity allows J.M. Moreno, J. Madrenas, and J. Cosp (Eds.): ICES 2005, LNCS 3637, pp. 108–118, 2005. c  Springer-Verlag Berlin Heidelberg 2005