Training multilayer perceptrons with the extended Kalman algorithm
Advances in neural information processing systems 1
Information-based objective functions for active data selection
Neural Computation
Neural network exploration using optimal experiment design
Neural Networks
Neural Networks for Modelling and Control of Dynamic Systems: A Practitioner's Handbook
Neural Networks for Modelling and Control of Dynamic Systems: A Practitioner's Handbook
Functional Adaptive Control: An Intelligent Systems Approach
Functional Adaptive Control: An Intelligent Systems Approach
Sequential Monte Carlo Methods to Train Neural Network Models
Neural Computation
From experiment design to closed-loop control
Automatica (Journal of IFAC)
Toward the training of feed-forward neural networks with the D-optimum input sequence
IEEE Transactions on Neural Networks
A priori and a posteriori machine learning and nonlinear artificial neural networks
TSD'10 Proceedings of the 13th international conference on Text, speech and dialogue
Posterior estimates and transforms for speech recognition
TSD'10 Proceedings of the 13th international conference on Text, speech and dialogue
Hi-index | 0.00 |
System identification is a discipline for construction of mathematical models of stochastic systems based on measured experimental data. Significant role in the system identification plays a selection of input signal which influences quality of obtained model. Design of optimal input signal for system modeled by multi-layer perceptron network is treated. Because the true system is unknown, the design can be constructed only from the actually obtained model. However, neural networks with the same structure differing only in parameters values are able to approximate various nonlinear mappings therefore it is crucial maximally to use available informations to select suitable input data. Hence a global estimation method allowing to determine conditional probability density functions of network parameters will be used. The Gaussian sum approach based on approximation of arbitrary probability density function by a sum of normal distributions seems to be suitable to use. This approach is a less computationally demanding alternative to the sequential Monte Carlo methods and gives better results than the commonly used prediction error methods. The properties of the proposed experimental design are demonstrated in a numerical example.