Parameter Optimization of Complex Simulation Models

Abstract

Today, a great shortcoming of most of the available simulation tools is, that optimization is not sufficiently supported by appropriate optimization techniques. This situation is mainly caused by the lack of qualified strategies for optimization of simulation models. Traditional indirect optimization techniques based on exploitation of analytical information like gradients etc. cannot be applied, because only goal function values are available which have to be calculated by an often very expensive simulation process. In the first part of this paper, some powerful direct optimization algorithms are presented which work iteratively, only requiring goal function values. These strategies are combinations of direct global and local optimization methods (Genetic Algorithms, Simulated Annealing, Hill-Climbing, etc.) trying to merge the advantages of global and local search. Our developed strategies have been implemented and integrated into REMO (REsearch Model Optimization package) representing a software tool for experimentation with simulation models. At the moment REMO is extended and completely reimplemented in Java to make it available on the World Wide Web. The ongoing development process of REMO is described in the second part of this paper.