Efficient Global Optimization Under Conditions of Noise and Uncertainty – A Multi-Model Multi-Grid Windowing Approach

1. Abstract Incomplete convergence in numerical simulation such as computational physics simulations and/or Monte Carlo simulations can enter into the calculation of the objective function in an optimization problem, producing noise, bias, and topographical inaccuracy in the objective function. These affect accuracy and convergence rate in the optimization problem. This paper is concerned with global searching of a diverse parameter space, graduating to accelerated local convergence to a (hopefully) global optimum, in a framework that acknowledges convergence uncertainty and manages model resolution to efficiently reduce uncertainty in the final optimum. In its own right, the global-to-local optimization engine employed here (devised for noise tolerance) performs better than other classical and contemporary optimization approaches tried individually and in combination on the “industrial” test problem to be presented.