Date

2014

Author

Bozağaç, Cumhur Doruk

Metadata

Show full item record

Item Usage Stats

173
views

73
downloads

Cite This

Computer simulations are widely used for design optimization purposes. The problem becomes challenging when design variables are high dimensional and when the simulation is computationally expensive. In this work we propose a methodology for metamodeling of dynamic simulation models via Multivariate Adaptive Regres sion Splines (MARS). To handle incomplete output processes, where the simulation model does not produce an output in some steps due to missing inputs, we have devised a two-level metamodeling scheme. The methodology is demonstrated on a dynamic radar simulation model. The prediction performance of the resulting meta model is tested with four different sampling techniques (i.e., experimental designs) and 16 sample sizes. We also investigate the effect of alternative coordinate system representations on the metamodeling performance. The results suggest that MARS is an effective method for metamodeling dynamic simulations, particularly, when expert judgment is not readily available. Results also show that there are interactions between the coordinate system representations and sampling techniques, and some sampling-representation-size combinations are very promising in the solution to this type of problem. The technique is then applied to develop proxy models (PMs) of atomic models in Discrete Event System Specification (DEVS) simulations to replace an atomic model with a PM that uses the fitted metamodel. The mechanisms required for replacing an atomic model and integrating the PM into a simulation are described in detail. The methodology is tested by replacing a radar atomic model in a military engagement simulation with the PM and the method’s benefits and chal- lenges are discussed thoroughly. As a final step of this research, the response surfaces of the original simulation and the PM integrated simulation is analyzed. Both simulations are used in a Particle Swarm Optimization based optimization procedure and the results are compared. Results obtained from this particular case suggest that metamodeling of computationally intensive atomic models is feasible, and even with relatively small number of observations, we can apply metamodeling to sub-components of DEVS simulation models successfully.

Subject Keywords

Engineering design, Engineering design, Regression analysis, Experimental design

URI

http://etd.lib.metu.edu.tr/upload/12616735/index.pdf
https://hdl.handle.net/11511/23233

Collections

Graduate School of Natural and Applied Sciences, Thesis