Doğrusal olmayan regresyon model parametrelerinin nokta ve aralık tahmini için bir yaklaşım
Özet
A common used approach for obtaining point estimates of nonlinear regression model parameters is applying Least Squares (LS) approach by using derivative-based iterative algorithms. The Gauss-Newton, the Steepest Descent and Levenberg-Marquardt (L-M) algorithms are widely used derivative-based algorithms which are defined in the literature. However, if the initial values for these algorithms are not well-defined, some of the problems ocur during the search, e.g. trapping to local solutions and non-convergence to global solution. In this study, derivative-free optimization algorithms, Nelder-Mead Simplex (NMS) algorithm and Genetic Algorithm (GA), are used as alternative approaches to derivative-based algorithms to obtain point estimations of nonlinear regression model parameters. The detailed informations are given about NMS algorithm and GA with defining the advantages and disadvantages. It is known that defining the proper values of GA tuning parameters effect global solution. Therefore, optimal GA tuning parameters are defined by using Taguchi experimental design. In addition, in the study, a hybrid algorithm, called GANMS and obtained by combining the advantageous properties of NMS and GA, is used for point estimation of parameters. The Bootstrap approach is used for interval estimation of model parameters. The point and interval estimation approaches, defined in the study, are applied on a data set from the literature and estimation of the negative-exponential regression model parameters are obtained. The obtained solutions are compared with the previous results. It is seen that the GA with proper tuning parameters or GANMS hybrid algorithm can be used as an optimization tool for point estimation of nonlinear regression models.