An evolutionary nested sampling algorithm for Bayesian model updating and model selection using modal measurement. (1st June 2017)

• Record Type: Journal Article
• Authors: Qian, Feng; Zheng, Wei
• Is Part Of: Engineering structures. Volume 140(2017:Jun. 01)
• Page Start: 298
• Page End: 307
• Publication Date: 2017-06-01
• Subjects: Nested sampling -- Bayesian inference -- Model updating -- Model selection -- Evolutionary algorithm -- Modal measurement; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2017.02.048
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: An evolutionary nested sampling algorithm for Bayesian model updating and model selection. The proposed algorithm is more efficient than standard nested sampling. The proposed algorithm is able to tackle multiple-solution problems. The algorithm termination condition is proposed. The proposed algorithm is robust to noise and can estimate model evidence. Abstract: Nested sampling (NS) is a highly efficient and easily implemented sampling algorithm that has been successfully incorporated into Bayesian inference for model updating and model selection. The key step of this algorithm lies in proposing a new sample in each step that has a higher likelihood to replace the sample that has the lowest likelihood evaluated in the previous iteration. This process, also regarded as a constrained sampling step, has significant impact on the algorithm efficiency. This paper presents an evolutionary nested sampling (ENS) algorithm to promote the proposal of effective samples for Bayesian model updating and model selection by introducing evolutionary operators into standard NS. Instead of randomly drawing new samples from prior space, ENS algorithm proposes new samples from previously evaluated samples in light of their likelihood values without any evaluation of gradient. The main contribution of the presented algorithm is to greatly improve the sampling speed in the constrained sampling step by use of previous samples. The performances of the proposed ENS algorithm for model updating and model selection are examined through two numerical examples.