Lehrstuhl Mathematik & InformatikNeural Control within the BMFT-Project Neres
Abstract. Whereas the identification and control of linear systems is well understood, this does not apply in general to nonlinear systems. Here, neural nets open up new paths for the treatment of multidimensional nonlinear systems as well as the possibility of adaptive readjustments to changes of the environment and of the system parameters. The advantages of neural control are of particular value for robotics. On the subsymbolic level, the goal is a symbiosis between sensorics and neural signal processing and control. However, we do intend to use traditional AI-techniques in cases where a robust knowledge representation is required which goes beyond the subsymbolic level, e.g. for space representation. In many applications, the problem is to extract significant control parameters from visual sensor data in a robust and efficient manner. For this task, neural nets are suited particularly well. Mathematical models for machine learning as well as unifying dynanical concepts will be utilized to achieve quantitative, generalizable results with respect to the efficiency of neural nets, by taking into account the real world requirements for control tasks with respect to performance, reliability and fault tolerance. Speech is of special significance for the dialogue with autonomous systems. Since neural nets have lead to encouraging results in speech processing, corresponding techniques will also be applied in robotics.