Transfer prior knowledge from surrogate modelling: A meta-learning approach. (February 2022)

• Record Type: Journal Article
• Authors: Cheng, Minghui; Dang, Chao; Frangopol, Dan M.; Beer, Michael; Yuan, Xian-Xun
• Is Part Of: Computers & structures. Volume 260(2022)
• Publication Date: 2022-02
• Subjects: Meta-learning-based surrogate modelling -- Model-agnostic meta-learning -- Knowledge transfer -- Surrogate modelling; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2021.106719
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: Propose a meta-learning-based surrogate modelling (MLSM) framework for knowledge transfer. Provide the definition of similar tasks and identify the scope of the framework. Outline the applications to global sensitivity analysis, reliability analysis, and optimization. Demonstrate the ability of the framework to transfer prior knowledge and show the computational efficiency. Abstract: Surrogate modelling has emerged as a useful technique to study complex physical and engineering systems in various disciplines, especially for engineering analysis. Previous studies mostly focused on developing new surrogate models and/or applying existing surrogate models to practical problems. Despite the computational efficiency, the surrogate for a new task is often built from scratch and the knowledge gained from previous surrogate modelling for similar tasks is neglected. As the need for quickly modifying simulation models to reflect design changes has significantly increased, one potential solution is to utilize prior knowledge from surrogate modelling. In this study, a novel meta-learning-based surrogate modelling framework is presented. The framework includes two phases: a meta-training and a few-shot learning phase. A meta-model that represents a family of tasks and the adaptation of this model to a new task with few data points are the results of the first and second phase, respectively. The study specifies the scope of the framework by classifying similar tasks. Applications of the framework to global sensitivity analysis, optimization, and reliability analysis are also addressed. Four numerical experiments are performed to demonstrate the feasibility and applicability of the framework.