Design and Optimization under Uncertainties : A Simulation and Surrogate Model Based Approach

This thesis deals with development of complex products via modeling and simulation, and especially the use of surrogate models to decrease the computational efforts when probabilistic optimizations are performed. Many methods that can be used to perform probabilistic optimizations exist and this thesis strives to present and demonstrate the capabilities of a few of them. Hopefully, this information can be helpful for someone who wants to choose a method. Knowledge about several different topics is required to perform a probabilistic optimization. First, it is necessary to incorporate the probabilistic behavior into the analysis by estimating how the uncertainties and variations in the model and its parameters are affecting the performance of the system. The focus in this thesis is on sampling based methods to estimate these probabilities. Secondly, an optimization algorithm should be chosen so that the computer can search for and present an optimal solution automatically. The probabilistic optimization process can be computationally demanding since numerous simulations of the model are performed each time the value of the objective function is estimated. It is therefore desirable to speed up the process by incorporating computationally effective surrogate models. This is especially important if the simulated model is computationally demanding on its own, e.g. a finite element model with many nodes. Each of these topics is presented in its own chapter of this thesis. A few methods are presented and their performances demonstrated for each topic. Surrogate models can also be used to improve the performances of optimization algorithms when the desire is to optimize computationally expensive objective functions. With this in mind, efforts have been made to improve the Complex-RF optimization algorithm. A modified algorithm is presented in this thesis and the main difference is that it creates and utilizes surrogate models iteratively during the optimization process. The modified algorithm is compared with Complex-RF and is demonstrated to be superior for computationally expensive models