Distributed intelligent systems like self-organizing wireless sensor and actuator networks are supposed to work mostly autonomous even under changing environmental conditions. This requires robust and efficient self-l...Distributed intelligent systems like self-organizing wireless sensor and actuator networks are supposed to work mostly autonomous even under changing environmental conditions. This requires robust and efficient self-learning capabilities implementable on embedded systems with limited memory and computational power. We present a new solution called Spiral Recurrent Neural Networks (SpiralRNN) with an online learning based on an extended Kalman filter and gradients as in Real Time Recurrent Learning. We illustrate its stability and performance using artificial and real-life time series and compare its prediction performance to other approaches. SpiralRNNs perform very stable and show an ac-curacy which is superior or similar to other state-of-the-art approaches. In a memory capacity evaluation the number of simultaneously memorized and accurately retrievable trajectories of fixed length was counted. This capacity turned out to be a linear function of the size of the recurrent hidden layer, with a memory-to-size ratio of 0.64 for shorter trajectories and 0.31 for longer trajectories. Finally, we describe two potential applications in building automation and logistics and report on an implementation of online learning SpiralRNN on a wireless sensor platform under the TinyOS embedded operating system.展开更多
文摘Distributed intelligent systems like self-organizing wireless sensor and actuator networks are supposed to work mostly autonomous even under changing environmental conditions. This requires robust and efficient self-learning capabilities implementable on embedded systems with limited memory and computational power. We present a new solution called Spiral Recurrent Neural Networks (SpiralRNN) with an online learning based on an extended Kalman filter and gradients as in Real Time Recurrent Learning. We illustrate its stability and performance using artificial and real-life time series and compare its prediction performance to other approaches. SpiralRNNs perform very stable and show an ac-curacy which is superior or similar to other state-of-the-art approaches. In a memory capacity evaluation the number of simultaneously memorized and accurately retrievable trajectories of fixed length was counted. This capacity turned out to be a linear function of the size of the recurrent hidden layer, with a memory-to-size ratio of 0.64 for shorter trajectories and 0.31 for longer trajectories. Finally, we describe two potential applications in building automation and logistics and report on an implementation of online learning SpiralRNN on a wireless sensor platform under the TinyOS embedded operating system.
