Acceleration of Direct Model Optimization Methods by Function Approximation
Michael Syrjakow, Helena Szczerbicka, Michael R. Berthold, Klaus–Peter Huber · KOPS (University of Konstanz) · 1996
In recent years optimization of simulation models has become a very important application field of direct optimization strategies. The search process of these iterative strategies is only based on cost function values and does not require any additional analytical information like gradients etc. Today the most common direct methods for global optimization are Genetic Algorithms, Evolution Strategies, and Simulated Annealing. All these methods apply sophisticated probabilistic search operators which Imitate principles of nature. Although these operators have been proven to be well-suited for global search the required computational effort (number of required cost function evaluations) still remains a big problem. In this paper we focus on acceleration methods for direct global optimization strategies. Our approach is based on cost function approximation. As approximation techniques we use a simple grid-based method and RecBFNs (Rectangular Basis Function Networks), a special kind of neural networks. The methods we have developed have been applied successfully to model optimization as well as to a selection of mathematical test problems. The encouraging results presented in this paper show that it is possible to optimize simulation models both successfully and with tolerable computational effort.