Publications

Detailed Information

Matrix-based Bayesian network for reliability assessment and decision making of complex large-scale systems : 대규모 복잡 시스템의 신뢰성 해석 및 의사결정을 위한 행렬기반 베이지안 네트워크

DC Field Value Language
dc.contributor.advisor송준호-
dc.contributor.author변지은-
dc.date.accessioned2020-05-19T07:50:57Z-
dc.date.available2020-05-19T07:50:57Z-
dc.date.issued2020-
dc.identifier.other000000159406-
dc.identifier.urihttps://hdl.handle.net/10371/167658-
dc.identifier.urihttp://dcollection.snu.ac.kr/common/orgView/000000159406ko_KR
dc.description학위논문(박사)--서울대학교 대학원 :공과대학 건설환경공학부,2020. 2. 송준호.-
dc.description.abstractTo secure the resilience of the modern society, the reliability of core engi-neering systems should be evaluated accurately. However, this remains chal-lenging due to various factors affecting the system performance in addition to the component and system events as well as the complex definition of the sys-tem event. The issue with regards to the probabilistic modeling of multiple fac-tors can be addressed by Bayesian network (BN), whose graph-based repre-sentation facilitates mathematical formulation of the causal dependence be-tween multiple variables. However, as the number of components increases, the conventional BN becomes inapplicable due to the consequent exponential increase in both the memory demand for quantifying probability distributions and computational cost for optimization.
In this thesis, the memory issue is addressed by developing matrix-based BN (MBN) in which as an alternative data structure to store the distributions, the conventional tables are replaced by conditional probability matrices (CPMs). This strategy facilitates (1) exploiting the regularity in the definition of system events during quantification, and (2) if necessary, enabling approx-imate inference within the framework of BN. The efficiency and applicability of MBN are demonstrated by the numerical examples. In addition, these ap-plications illustrate how MBN can be quantified using the existing system reli-ability methods.
Another issue addressed by the thesis is the optimization of large-scale systems. To this end, the objective function of the corresponding optimization problem, i.e. the sum of the expectations of utilities, is replaced by a proxy measure so as to reduce the computational cost from exponential to polyno-mial order. In order to promote the applications of the proposed method to a wide class of problems, the mathematical condition is derived by which the optimization problems employing proxy objective functions remains equiva-lent to the exact ones. Moreover, the proposed proxy measure allows the ana-lytical evaluation of a set of non-dominated solutions in which the weighted sum of multiple objective values is optimized. By using the strategies devel-oped to compensate the errors by the proposed approximation, the proposed method can not only show an improved accuracy, but also identify even a larger set of non-dominated solutions than the exact objective function of weighted sum. The numerical examples demonstrate the accuracy and effi-ciency of the proposed method.
Finally, to enhance the applicability to a wider class of problems, the MBN is extended from binary-state systems to multi-state systems. To this end, the definitions and the BN operations for MBN are modified. The extended MBN is demonstrated by three types of multi-state systems, i.e. multi-state se-ries-parallel systems, multi-state k-out-of-N:G systems, and flow capacity of networks: The applications are realized by for each system, developing the strategy for quantification and the formulations for inference.
-
dc.description.abstract현대사회의 재난 복원력(disaster resilience)을 확보하기 위해서는 사회중추시스템에 대한 정확한 신뢰성분석이 필수적이다. 그러나 이러한 시스템들은 주로 상호 의존적인 다수의 구성요소에 의하여 상태가 결정되는 데다가 구성요소와 시스템 사건 외에도 다양한 요인이 시스템 성능에 영향을 주어 이들을 대상으로 신뢰성분석을 수행하는 것은 어려운 과제이다. 이를 해결하기 위한 효과적 방법론으로 베이지안 네트워크(Bayesian network, BN)를 들 수 있다. BN은 다수의 변수 간 복잡한 인과관계를 도식화하여 나타냄으로써 고차원 확률변수의 정식화를 용이하게 하기 때문이다. 그러나 기존 BN 방법론은 시스템사건의 확률질량함수를 저장하는데 필요한 메모리와 최적화를 수행하는 데 요구되는 연산비용이 구성요소의 개수에 기하급수적으로 늘어나는 한계가 있어, 대규모 시스템에 적용될 수 없다.
먼저, 메모리의 한계를 해결하기 위하여, 본 논문은 행렬기반 베이지안 네트워크(Matrix-based Bayesian network, MBN)를 확률질량함수를 저장하기 위한 데이터 구조의 대안으로써 제안한다. MBN은 조건확률행렬(Conditional probability matrix, CPM)을 도입하여 기존 테이블 기반 저장방식의 한계를 극복한다. 즉, MBN을 활용하여 (1) 확률분포 저장의 효율성을 높이기 위하여 시스템사건의 정의에 내포된 규칙성을 활용할 수 있으며, (2) 근사추론이 필요한 경우 이를 BN 방법론과 결합할 수 있다. 이에 더하여, MBN이 두 개 이상의 상태를 갖는 시스템에 적용될 수 있도록 제안된 기본 정의를 확장하였다. 적용 예제들을 통하여 MBN의 활용성과 효율성을 증명하였으며, 각 예제에서 주어진 시스템사건을 MBN으로 모델링하기 위한 방법 또한 제시함으로써 MBN의 적용을 구체적으로 보였다.
한편, BN을 거대 복잡 시스템의 최적화에 적용하기 위하여, 본 논문은 목적함수에 해당하는 효용변수(utility variable)의 기댓값에 대한 근사식을 제시한다. 이를 통하여, 최적화 연산비용의 증가추세를 구성요소 개수에 대한 다항적(polynomial) 증가로 완화할 수 있다. 제안한 방법론의 실용성을 확보하기 위하여, 근사식을 활용한 최적화 문제가 정확식을 활용한 경우와 동일할 수 있는 조건을 제시하였다. 한편, 다중 목적 최적화를 수행하는 경우, 제안된 근사함수를 활용하여 효용변수 간 가중합을 최소화하는 방식으로 단일해가 아닌 비지배해 집합을 연산할 수 있다. 근사오차를 보완하기 위하여 제안된 경험적 방법은 근사오차를 보완할 뿐 아니라, 다중 목적 최적화를 수행하는 경우 정확한 가중합 식으로는 구할 수 없는 해까지 찾을 수 있다. 수식 유도와 적용 예제들을 통하여 제안된 방법론의 성능을 분석하고 정확성과 효율성을 증명하였다.
마지막으로, MBN의 적용성을 향상시키기 위하여 MBN을 두 가지 상태를 갖는 시스템을 넘어 여러 상태를 갖는 다중 상태 시스템에도 적용될 수 있도록 확장하였다. 이를 위하여 기존에 개발된 MBN 정의와 연산을 수정하였다. 확장된 MBN 방법론의 효용을 증명하기 위하여 세 가지 시스템, 즉, 다중 상태 직렬-병렬 시스템, 다중 상태 k-out-of-N:G 시스템, 네트워크 흐름 용량을 위한 MBN 정량화 기법을 개발하고 신뢰성해석을 수행하였다.
-
dc.description.tableofcontentsChapter 1. Introduction 1
1.1 Motivation 1
1.2 Objectives and scopes 4
1.3 Organization and nomenclature 5
Chapter 2. Matrix-based Bayesian network (MBN) for efficient memory storage and flexible inference 8
2.1 Introduction 8
2.2 Background and related work 9
2.2.1 Bayesian network (BN) 9
2.2.2 Conditional probability tables (CPTs) 10
2.2.3 Converging structure in BN and related research efforts 11
2.3 Proposed data structure of probability mass function (PMF): conditional probability matrix (CPM) 14
2.3.1 Definitions for CPM 14
2.3.2 Strategy to identify disjoint rules for CPM 18
2.3.3 Construction of CPM for deterministic functions 21
2.3.4 Application of other methodologies to the reliability block diagram (RBD) example 23
2.4 Probabilistic inference using MBN 27
2.4.1 Basic operations for inference using MBN 27
2.4.1.1 Example of the inference using MBN 30
2.4.2 Approximate BN inference using non-exhaustive CPMs 31
2.4.2.1 Non-exhaustive CPMs by deterministic selection 32
2.4.2.2 Non-exhaustive CPMs by stochastic selection 33
2.4.2.3 Approximate inference of the reliability block diagram example by MBN 37
2.5 Numerical examples 40
2.5.1 MBN application to connectivity of random graphs 40
2.5.2 MBN application to connectivity of Sioux Falls benchmark network 44
2.5.2.1 Conditional probability distribution of each random variable 46
2.5.2.2 Approximate inferences of probabilistic queries 48
Chapter 3. Proxy objective functions for efficient optimization of large-scale systems 51
3.1 Introduction 51
3.2 Quantification of Influence Diagram (ID) for complex systems using MBN 55
3.2.1 Background: ID for large-scale systems 55
3.2.1.1 Issue in optimization caused by converging structure 56
3.2.2 Extension of MBN to ID 57
3.3 Proxy objective function for optimizing ID that has multiple strategically relevant decision variables 58
3.3.1 Performance analysis 61
3.4 Multi-objective optimization using proxy objective function 65
3.4.1 Optimization using weighted sum as objective function 65
3.4.2 Selecting basis decision rules 68
3.4.3 Attainable solutions by the proposed scheme 69
3.5 Numerical examples 71
3.5.1 Illustrative example: a system with three components 71
3.5.2 Example RBD 76
3.5.3 Truss bridge structure 78
3.5.4 Sioux Falls benchmark network 82
Chapter 4. Generalized MBN for multi-state systems 87
4.1 Introduction 87
4.2 Generalized MBN for multi-state events 88
4.2.1 Composite state 88
4.2.2 Extended definitions and inference operations 91
4.3 Application I: multi-state series-parallel (MS-SP) system 95
4.3.1 Definition of random variables (r.v.s) for MS-SP system 96
4.3.2 MBN quantification of system event 99
4.3.3 System structure optimization of MS-SP system 100
4.3.4 Numerical examples 102
4.3.4.1 MS-SP system with two subsystems 102
4.3.4.2 MS-SP system with four subsystems 103
4.4 Application II: multi-state k-out-of-N:G system 108
4.4.1 Definition of r.v.s for k-out-of-N system 110
4.4.2 MBN quantification of system event 114
4.4.3 Value of information (VoI) 119
4.4.4 Numerical examples 119
4.5 Application III: flow capacity of network 121
4.5.1 Definition of r.v.s for network flows 123
4.5.2 MBN quantification of system event 126
4.5.2.1 Decomposition of system event based on flow 128
4.5.2.2 Proposed decomposition based on cut 131
4.5.2.3 BN inference by combining decomposition and sampling 135
4.5.3 Component probability based importance measure (CIM) 136
4.5.3.1 Deterministic bounds from incomplete decomposition 137
4.5.3.2 Stochastic estimation from decomposition and sampling 138
4.5.4 Numerical examples 139
4.5.4.1 Sioux Falls transportation benchmark network 139
4.5.4.2 Eastern Massachusetts (EMA) highway benchmark network 143
Chapter 5. Conclusions 147
5.1 Summary 147
5.2 Recommendations for future studies 151
References 154
Appendix. Illustrative examples of quantifying conditional probability matrix (CPM) of system event 158
Appendix A. Multi-state series-parallel system 158
Appendix B. Multi-state k-out-of-N:G system 159
Appendix C. Flow capacity of network 161
Abstract in Korean 164
-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subject.ddc624-
dc.titleMatrix-based Bayesian network for reliability assessment and decision making of complex large-scale systems-
dc.title.alternative대규모 복잡 시스템의 신뢰성 해석 및 의사결정을 위한 행렬기반 베이지안 네트워크-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.department공과대학 건설환경공학부-
dc.description.degreeDoctor-
dc.date.awarded2020-02-
dc.contributor.major구조신뢰성-
dc.identifier.uciI804:11032-000000159406-
dc.identifier.holdings000000000042▲000000000044▲000000159406▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share