Novel Stopping Criterion for Optimization

Abstract

A novel method for identification of steady state is demonstrated as the termination criterion for the optimization stage of modeling empirical data. The method was tested on a variety of applications. It is described, and its utility is demonstrated on modeling simulated data and is also validated using two laboratory scale experiments. The novel stopping criterion for optimization, based on identifying steady state of a random subset of the sum of squared deviations with respect to iteration number, was formerly explored for neural network training. The novel stop-optimization criterion was tested on a different variety of applications involving various kinds of objective functions. On all the cases, the novel stop-optimization criterion gives equivalent results (as measured by model residuals) to the best possible results, with a sufficient (not excessive) number of iterations and without a priori knowledge of the optimization problem (scale, end-point values, and other classic stopping criteria).