Multi-Task Feature Learning-Based Response Surface Method for Performance-Based Seismic Design Optimization

Abstract

In the context of performance-based earthquake engineering, it is crucial to quantify uncertainties in engineering demands, damages, and performance of a building by conducting dynamic analyses with multiple ground motions that reflect the uncertainty of seismic hazards. However, performing a large number of dynamic analyses that include the elastoplastic range requires significant computational costs. To address this issue, response surface methods that approximate dynamic analyses with cost-effective statistical models are expected to be beneficial. This paper proposes a response surface method for multiple ground motions based on multi-task feature learning, demonstrates its efficiency compared to treating ground motions independently, and investigates its applicability to the performance-based seismic design optimization.