An Extended Neuron Model for Efficient Time-Series Generation and Prediction

Modelling of spatio-temporal patterns with neural networks is an important task for a large number of applications which require adaptive control. In this paper, the use of an extended neuron model in neural networks is proposed to achieve a given dynamic network behaviour. The new neuron model is based on the implementation of exponential excitation decay and the introduction of temporal refractoriness of the neuron output as observed in biological nerve cells. A learning algorithm based on error-backpropagation for the resulting network is derived. A benchmark test on prediction of the chaotic Macky-Glass differential equation and real-live experiments with controlling the movement of a walking machine leg are performed. The results suggest superior time-series modelling ability of the presented approach in terms of network trainability and computation complexity.