Evolutionary computation of a deterministic switching regressions estimator (Q1775952)

This article reports the results of simulation investigations of the performance of an evolutionary method (based on genetic algorithm) for computation of a deterministic switching regression estimator proposed by \textit{S. M. Goldfeld} and \textit{R. E. Quandt} [see S.M. Goldfeld and R.E. Quandt (eds.), Studies in nonlinear estimation. (1976; Zbl 0418.62090)]. Distinctive aspects of the method include (1) a combination of simple and random chromosomal crossover, (2) extension of the principle of natural selection to the parametric set-up of a genetic algorithm. This extension provided a set of internal parameters for the evolutionary approach that was shown to be robust for all experiments it was applied to. Results of experiments also indicate that the evolutionary method duplicates, significantly faster, estimates obtained by an existing enumerative method in samples small enough to permit enumeration. It also provides the ability to calculate the estimator in much larger sample sizes than it is possible with the enumerative approach. A real data example from the USA gasoline market is presented.