Use of Discretization and Solution History in Stochastic Optimization

Abstract

A practical improvement for stochastic algorithms used with numerical models is proposed. Combined use of solution space discretization and solution history is implemented. Discretization gives the possibility to use a solution history where all the solutions evaluated are stored. Use of the solution history guarantees that a single solution candidate is calculated only once. This approach is useful for the stochastic algorithms that typically evaluate many solution candidates The approach is most useful in the cases where a stochastic algorithm decreases the search space during the optimization process. The improvement proposed was tested in optimization of high-speed induction motors modeled with 2D finite element analysis software. With a genetic optimization algorithm, the average time saving for seven separate optimization runs was 39%.