Weighted linear associative memory approach to nonlinear parameter estimation (Q2565004)

A new method for nonlinear parameter estimation (identification) is introduced. It is based on a weighted linear associative memory, a methodological approach inherent to the artificial neural networks. The idea of linear associative memory is presented in the introduction. The weighting itself improves the linear associative memory performance by taking into account how close a response vector is to the observed one. By getting closer the weight of the parameters in the cost function is increased. The complete theory of parameter estimation by weighted linear association is developed. The results are illustrated in two examples: the van der Pol equation and the somatic shunt cable model of neuron, which is considered also in the stochastic regime. In conclusion, the author claims that the introduced method significantly improves the accuracy and efficiency of the estimation.