| A global optimization algorithm, called the Genetic Algorithm is modified for use in parameter estimation problems for nonlinear systems. The modifications include parent selection schemes, and some new operators such as extinction and immigration, reencoding of zeros, and forced mutation. The Genetic Algorithm is unlikely to be trapped at local minima of error surfaces because of its nature of structured random search and because it uses no gradient information. Straightforward application of the Genetic Algorithm to parameter estimation problems include the parameter estimation of nonlinear Infinite Impulse Response (IIR) filters, training of feedforward and recurrent neural networks and frequency modulation/demodulation in the time domain. With appropriate parameterization, the Genetic Algorithm can also be applied to nonparametric learning of n-dimensional decision regions for pattern classification. The behavior of the algorithm is analyzed and its convergence is demonstrated under suitable hypotheses.; Given a set of input data, if the parameter estimation is restricted to use only a small subset of this data, the Genetic Algorithm can be used to choose the best subset of input data to maximize the accuracy of estimation. The examples investigated for this application are sensor placement for modal identification of an early version of the Space Station and an individual Space Station photovoltaic array.; If a parametric model of the nonlinear system to be identified is unknown, a Volterra series can be used for the modelling. However, the estimation of Volterra kernels usually suffers from high dimensionality and over parametrization because the number of kernels to be estimated increases exponentially with the number of delays and the order of the Volterra series. An algorithm called the Recursive Approximation and Estimation algorithm is introduced to simultaneously choose a parsimonious set of parameters and estimate these parameters. |
