Stationary Fokker - Planck Learning for the Optimization of Parameters in Nonlinear Models

摘要

A new stochastic procedure is applied to optimization problems that arise in the nonlinear modeling of data. The proposed technique is an implementation of a recently introduced algorithm for the construction of probability densities that are consistent with the asymptotic statistical properties of general stochastic search processes. The obtained densities can be used, for instance, to draw suitable starting points in nonlinear optimization algorithms. The proposed setup is tested on a benchmark global optimization example and in the weight optimization of an artificial neural network model. Two additional examples that illustrate aspects that are specific to data modeling are outlined.