Activity Number:
|
293
- SPEED: Computing, Graphics, and Programming Statistics
|
Type:
|
Contributed
|
Date/Time:
|
Tuesday, August 1, 2017 : 8:30 AM to 10:20 AM
|
Sponsor:
|
Quality and Productivity Section
|
Abstract #322865
|
View Presentation
|
Title:
|
Sequential Computer Experiments for Failure Probability Estimation --- a Floor System Example
|
Author(s):
|
Hao Chen* and William James Welch
|
Companies:
|
University of British Columbia and University of British Columbia
|
Keywords:
|
Engineering model ;
Expected Improvement ;
Gaussian Process ;
Rare Probability Estimation ;
Sequential Design
|
Abstract:
|
The application that motivates this research is an engineering model to quantify the relationship between Modulus of Elasticity of joists in a floor system and its corresponding maximum deflection under a static load. Our research objective is to statistically estimate the system's failure probability under the assumption that the engineering model is expensive to run. We first use a Gaussian process (GP) to build a statistical surrogate for the input-output relationship of the engineering model with a modest number of evaluations to obtain an initial estimate of the rare failure probability. We then experiment sequentially, guided by the surrogate, to improve the estimate of the failure probability. We optimize the expected improvement criterion (Ranjan et.al (2008)) in sequentially selecting the new sets of inputs. The process will stop and output the final estimate of the failure probability once all the available resources have been exhausted. In summary, it is a machine learning process that updates the estimate "on the fly" as new information is added and we believe the research is of great application potential in assisting engineers in achieving their practical goals.
|
Authors who are presenting talks have a * after their name.