Transfer optimization in complex engineering design

• Main Author: Tan, Alan Wei Min
• Other Authors: Ong Yew Soon
• Format: Theses and Dissertations
• Language: English
• Published: 2019
• Subjects: DRNTU::Science::Mathematics::Applied mathematics::Optimization; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/105865; http://hdl.handle.net/10220/47873
• Institution: Nanyang Technological University

The optimization task is an important element in most real-world design processes. It facilitates the construction of robust, high performance machines that are essential in today's dynamic, safety-focused world. However, in the domain of complex engineering design, the objective functions to be optimized generally do not have closed-form solutions, are expensive to evaluate, and is non-convex. Consequently, traditional optimization algorithms that typically require up to several thousand exact function evaluations to obtain near-optimal solutions, are problematic to cost-effectively utilize. The solution to this issue has traditionally been in the form of surrogate-assisted optimization, where approximations of the expensive functions are constructed and, subsequently, searched. However, despite the improved results afforded by this methodology, a fairly substantial amount of initial evaluations is still required in order to build sufficiently accurate representations of the problem under investigation. The improvements to designs in most practical settings are often incremental, with progress made gradually over time. Consequently, there is an abundance of knowledge from previous (or in-progress) design efforts that are related (but distinct), that can be utilized to augment the optimization task in order to achieve accelerated performance. Additionally, modern design cycles are often broken down into manageable pieces that are addressed by numerous teams working together in tandem. As such, relevant information can become available at various stages of an ongoing optimization task. Careful exploitation of this knowledge can be extremely beneficial in real-world settings, particularly for computationally expensive optimization problems. In other words, in such cases, approaching the optimization task in a tabula rasa manner increases the duration (and thereby, cost) of the design stage...