| In this thesis we study a general class of optimization problems where the objective function involves the evaluation of random factors. The real-world origins of this type of problems usually are the design and optimization of large complex systems. From an engineering point of view, efficiency is crucial when approaching this type of problems, due to the scale and complexity of it as well as the constraint of time and computational resources. We present in this thesis a new framework, building on the foundation of the goal softening idea in ordinal optimization, that decomposes the optimization process into two components, the breadth process, which concerns the search in the design space, and the depth process, which concerns the evaluation of design performance. By comparing the contribution of the two processes to the objective, we can dynamically and optimally control the progress of an optimization procedure. Algorithms designed under this framework are shown to be highly efficient.; Our approach is substantiated by two case studies. One is the pricing of American-style options. The other is the scheduling problem for apparel manufacturing systems. In both problems, significant improvements are shown comparing our approach to existing methods. |
