Analog Neural Networks in Autonomous Systems

iii Summary Signal processing in animals is characterized by high parallelism, robustness, adaptivity and low power consumption, which results in a superior performance in sensory motor control tasks compared to artiicial systems. The analog neural network approach tries to come closer to the original than other artiicial systems by exploiting the underlying hardware's physical properties rather than avoiding most of them. However, learning analog neural networks used so far did not yet result in too many practical applications, because the artiicial neural networks mainly in use rely on a numerical accuracy that is typically not provided by analog hardware implementations. Natural nervous systems demonstrate high robustness, as required for an analog hardware implementation. Hence, a deeper understanding of`hardware-friendly' learning principles may enable the construction of artiicial systems that share more of a natural sys-tem's properties than they do today. This thesis provides new results on how diierent principles of animal learning can be performed employing a simple Hebbian learning rule and short-term dynamics. A combination of classical and operant conditioning was implemented in neural networks comprising neuromimetic analog hardware circuits controlling a physical mobile robot. These networks enabled the robot to learn local maneuvers, in which forward and backward movements were combined. A variety of further networks is introduced performing visual preprocessing, adaptive feedback, adaptive timed responses, parallel learning at diierent rates and other tasks. The robustness of this approach is demonstrated by the innuence of hardware failures and mismatches of, e.g., sensor preprocessing. Analytical considerations reveal similarities to other learning methods and clarify network properties. A key issue is the stability of the resulting behavior. A physical system controlled by learning hardware with several plastic elements was not demonstrated before. Moreover, a new neural network approach to adaptive self-localization is introduced, which enables a mobile robot to enhance its position estimation in an adaptive way. An analytical convergence result, performance in a simulated environment and, to a minor extent, tests in physical robots are reported. Consequences of the work presented here and future aspects of analog neural networks are outlined in a concluding chapter. Netze werden in einem abschlieeenden Kapitel zusammengefaat.