Parameter Screening Using Impact Factors and Surrogate-Based ANOVA Techniques
Wenbin Song, Andy J. Keane · 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference · 2006
This paper introduces the concept of parameter impact factors in order to screen important parameters in high dimensional design optimization problems which make use of computationally expensive high fidelity simulation models. Based on a snapshot dataset obtained by evaluating design points produced by Design of Experiments techniques, a simple concept of parameter impact factors is introduced and calculated to obtain preliminary estimates on the importance of parameters in the simulation results. Combined with parallel tuning of hyperparameters used in Gaussian process surrogate models and ANOVA techniques using the progressively built surrogate models, a more accurate estimation on the impact of different parameters can be achieved. Less important parameters can then be fixed in order to reduce the dimensionality of the problem to make the problem more tractable within given computational budget and time constraints.