University of Illinois Urbana-Champaign

A Survey of Parameterization Techniques for Bayesian Network Models for Human Reliability Analysis

O’Leary, Joseph; Zhao, Yunfei; Groth, Katrina

Content Files
A67.docx
F67_Final.docx

Permalink

https://hdl.handle.net/2142/121855

Description

Title
A Survey of Parameterization Techniques for Bayesian Network Models for Human Reliability Analysis
Author(s)
  • O’Leary, Joseph
  • Zhao, Yunfei
  • Groth, Katrina
Issue Date
2023
Keyword(s)
  • Bayesian networks
  • Human reliability analysis
  • Machine learning
  • Parameter Estimation
Abstract
Bayesian networks are a type of probabilistic machine learning model which consist of a directed acyclic graph and a set of prior distributions or conditional probability tables. Parameterization and quantification of the prior distributions and tables is a challenge in domains with sparse data such as Human Reliability Analysis (HRA), and often relies on expert elicitation or highly noninformative priors. While there are several methods for computing human failure probabilities, they are often limited by assumptions about independence of performance influencing factors and the independence of tasks completed by the human operator(s) under study. Bayesian networks allow the analysis to explicitly and rigorously incorporate information about dependencies between performance influencing factors and tasks. This paper presents a survey of techniques for the parameterization and quantification of Bayesian network priors and conditional probability tables for HRA to improve the quality of resulting Bayesian network models. The paper considers cases where numerical data is available but limited, where numerical data is absent and the analysis must rely upon either direct expert elicitation or elicitation of probabilistic information from published literature, and where both numerical data and information from experts are jointly available. Techniques include approaches for both parameter learning of root node probability distribution parameters and conditional probabilities of child nodes from both numerical sources and expert sources. Fully quantified Bayesian networks have twofold value to the HRA analyst. The structure of the network allows the analyst to make logical statements about what factors are significant to the completion of a task and how those factors are related to one another. The fully quantified network allows the analyst to make numerical predictions about the probability of a human failure event, to assess the sensitivity of such a result to various possible changes in the network, and to then make practical, well-founded design recommendations to improve the reliability of the system under study.
Type of Resource
text
Genre of Resource
Conference Paper/ Presentation
Language
eng
Handle URL
https://hdl.handle.net/2142/121855
Sponsor(s)/Grant Number(s)
  • U.S. Nuclear Regulatory Commission Grant No. 31310020M0002
  • University of Maryland, College Park

Owning Collections

PSAM 2023 Conference Proceedings PRIMARY