Domain knowledge-guided Bayesian evolutionary trees for estimating the compression modulus of soils containing missing values

Created by W.Langdon from gp-bibliography.bib Revision:1.8414

@Article{Zhang:2024:engappai,

• author = "Wenchao Zhang and Peixin Shi and Huajing Zhao and Zhansheng Wang and Pengjiao Jia",
• title = "Domain knowledge-guided Bayesian evolutionary trees for estimating the compression modulus of soils containing missing values",
• journal = "Engineering Applications of Artificial Intelligence",
• year = "2024",
• volume = "133",
• pages = "108356",
• keywords = "genetic algorithms, genetic programming, Compressive modulus, Missing data analysis, Bayesian additive regression trees",
• ISSN = "0952-1976",
• URL = "https://www.sciencedirect.com/science/article/pii/S0952197624005141",
• DOI = "doi:10.1016/j.engappai.2024.108356",
• abstract = "The soil compressive modulus (Es) is a critical parameter in geotechnical design, but its accurate estimation remains challenging because of incomplete geotechnical investigation data and inherent uncertainties in geomaterial properties. This study introduces a novel framework, called Bayesian evolutionary trees enhanced with missingness incorporated in attributes (BETm), for predicting Es with missing data. BETm encompasses three essential components: missingness incorporated in attributes (MIA), genetic programming (GP), and Bayesian additive regression trees (BART). The MIA module is integrated to impute missing data, specifically accounting for geotechnical patterns. The GP component constructs high-order variables to effectively capture the underlying parameter relationships. BART conducts a comprehensive uncertainty analysis on the importance of input variables, providing probabilistic predictions of Es with confidence intervals. BETm was applied to a dataset comprising 2955 geotechnical samples from 101 boreholes in Suzhou, China. Comparative experiments demonstrated its effectiveness compared to several state-of-the-art methods. The proposed model achieved superior accuracy, with a coefficient of determination (R2) of 0.977, mean absolute error of 0.044 MPa, and mean absolute percentage error of 33.8percent. Compared to a random forest model using average values for imputation, this model showed a 6percent improvement in R2. Furthermore, BETm performed competitively for extreme values (outside the normal range of 4-14 MPa), underscoring its robustness in predicting Es with missing data",

}

Genetic Programming entries for Wenchao Zhang Peixin Shi Huajing Zhao Zhansheng Wang Pengjiao Jia

Citations