M. Šimandl and P. Hering (Czech Republic)
System identification, nonlinear parameters estimation, structure adaptation, probability density function, multi layer perceptron network.
Application of neural networks in identification of nonlin ear stochastic systems is treated. The stress is laid on a parameters estimation and structure adaptation of the net works. They are trained by a global filtering method allow ing to determine conditional probability density functions of network parameters. The Gaussian sum approach used for parameters estimation of network gives better results than the commonly used prediction error methods, and it is an interesting alternative to sequential Monte Carlo meth ods. The approach also enables structure adaptation which is given by pruning of insignificant connections from an a priori chosen large network. The designed structure adap tation method utilizes conditional probability density func tions of the parameters obtained from the estimation algo rithm to measure saliency of the network connections and it represents a generalization of the extended Kalman filter based pruning method. The proposed training and pruning approach is demonstrated in a numerical example.
Important Links:
Go Back
