Technical Fault Diagnosis with

Bayesian Networks & BayesiaLab

1
 Overview
• Introduction: Our Company and Technology
• Motivation
• We have no data
• Example: my car won’t start!
• Logic is limited
• Key Concepts
• Bayesian Networks
• Directed Acyclic Graphs
• Constructing and Evaluating a Bayesian Network Model
• Diagnosis model
• Resource optimization for diagnosis and repair.

BayesiaLab.com 2
 Your BayesiaLab Team Today

stefan.conrady@bayesia.us stacey.blodgett@bayesia.us clare.gora@bayesia.us

BayesiaLab.com 3
 Co-founded in 2001
by Dr. Lionel Jouffe &
Dr. Paul Munteanu

stefan.conrady@bayesia.us 4
BayesiaLab.com 5
 Teaching Edition
Academic Edition

Desktop BayesiaLab 6
Bayesia Market
Simulator
Code Export Module
Professional
Software

BayesiaLab
WebSimulator

Web
Application Bayesia Expert
Knowledge Elicitation
Environment
(BEKEE)

Bayesia Engine API for

Bayesia Engine API for
API Modeling and
Inference
Network Learning

BayesiaLab.com 6
7
 Bayesian Networks & BayesiaLab
A Practical Introduction for Researchers
• Free download:
www.bayesia.com/book
• Hardcopy available on Amazon:
http://amzn.com/0996533303

20,000
Downloads

BayesiaLab.com 8
 Webinar Slides, Data, and Recording Available

stefan.conrady@bayesia.us 9
Motivation & Objective
Diagnostic Reasoning & Resource Optimization
 Medical Diagnostics

Sample Uniformity Single

Clump Uniformity Marginal Bland Normal
Code of Cell Epithelial Bare Nuclei Mitoses Class
Thickness of Cell Size Adhesion Chromatin Nucleoli
number Shape Cell Size
1000025 5 1 1 1 2 1 3 1 1 2
1002945 5 4 4 5 7 10 3 2 1 2

Wisconsin Breast Cancer Database

1015425
1016277
1017023
3
6
4
1
8
1
1
8
1
1
1
3
2
3
2
2
4
1
3
3
3
1
7
1
1
1
1
2
2
2
1017122 8 10 10 8 7 10 9 7 1 4
1018099 1 1 1 1 2 10 3 1 1 2
1018561 2 1 2 1 2 1 3 1 1 2

See Chapter 6 in Book 1033078 2 1 1 1 2 1 1 1 5 2

BayesiaLab.com 11
 Medical Diagnostics
Augmented
Markov
Blanket
SC=0.4

See webinar recorded on February 9, 2018

stefan.conrady@bayesia.us 12
www.BayesiaLab.com 13
 Physics Laboratory

BayesiaLab.com 14
 DEAD
END

NO DATA

www.BayesiaLab.com 15
 “Mean Time Between Failures”
Mean Hours Between Failure

BayesiaLab.com 16
 Motivation
Reliability

BayesiaLab.com 17
Example from Marketing Mix Optimization Webinar

Generic
2000

18
 The New Paradigm: Bayesian Networks
Under the Hood

BayesiaLab.com 19
 Is it supposed to
look like that?
Workshop

BayesiaLab.com 20
 Deductive Logic
Limitations of Logic
• “Classical logic has no explicit mechanism for representing the degree of
certainty of premises in an argument, nor the degree of certainty in a
conclusion, given those premises.”
Source: J. Williamson, Handbook of the Logic of Argument and Inference: The Turn Toward the
Practical

LOGIC IS NOT ENOUGH!

BayesiaLab.com 21
 2000 Years After Aristotle 1763
PHILOSOPHICAL
TRANSACTIONS
Bayes’ Theorem for Conditional Probabilities

H: Hypothesis
E: Evidence
= P(E | H) P(H)
P(H | E)
P(E)
“Probability of
H given E”

stefan.conrady@bayesia.us 22
 Probabilistic Reasoning
Mathematical Formulation of Inductive Reasoning
• “Bayesian inference is important because it provides a normative
and general-purpose procedure for reasoning under uncertainty.”
Source: Inductive Reasoning: Experimental, Developmental, and Computational
Approaches, edited by Aidan Feeney and Evan Heit
• “People do not appear to be natural Bayesians, i.e. they do not
seem to follow Bayesian reasoning when making a decision.”
Source: Wheaton, Kristan J.; Lee, Jennifer; and Deshmukh, Hemangini. “Teaching
Bayesian Statistics To Intelligence Analysts: Lessons Learned.” Journal of
Strategic Security 2, no. 1 (2010): 39-58.

BayesiaLab.com 23
 Why is this so important?
Human Cognitive Limitations and Biases Under Uncertainty

Human Reasoning ≈ Normative Reasoning

Disease Symptom

FALLACIES Human Reasoning ≠ Normative Reasoning

BayesiaLab.com 24
 I have just updated
my belief given this
Workshop new evidence.

BayesiaLab.com 25
Conceptual Advantages of Bayesian
Networks for Diagnostic Reasoning
 Bayesian Networks

BayesiaLab.com 27
 Bayesian Networks for Diagnostic Reasoning
Key Properties

• Representation of the Joint Probability Distribution

• No distinction between dependent and
independent variables
• Omni-directional Bayesian inference
• Nonparametric
• Probabilistic
• Causal
• Intuitive

BayesiaLab.com 28
 Bayesian Networks for Diagnostic Reasoning
Omni-Directional
Inference

BayesiaLab.com 29
 Bayesian Networks for Diagnostic Reasoning
Omni-Directional
Inference

BayesiaLab.com 30
 Directed
Acyclic Graphs

Bayesian
Networks
Direct Acyclic Graphs
 Directed Acyclic Graphs
What is a DAG? Qualitative!
• DAG=Directed Acyclic Graph
• DAGs are nonparametric
• A Node represents a variable in a domain, regardless of whether it is observable or
unobservable.
• A Directed Arc has represents a potential causal effect. The arc direction indicates
the assumed causal direction, i.e. “A→B” means “A causes B.”
• A Missing Arc encodes the definitive absence of a direct causal effect, i.e. no arc
between A and B means that there exists no direct causal relationship between A
and B and vice versa.

BayesiaLab.com 32
 Directed Acyclic Graphs
Why DAGs?
• Easy, natural encoding of causal
knowledge or assumptions (compare to
Knowledge Encoding in Chapter 4).
• Clear implications for association.
• Clear criteria for identification.

BayesiaLab.com 33
 Directed Acyclic Graphs
There are only three types of structures in a DAG:

BayesiaLab.com 34
 Directed Acyclic Graphs
Structures in a DAG
• Indirect Connection

The type of causal

structure has major
implications regarding
• Common Cause the association of nodes,
i.e. what we can observe!

• Common Effect (Collider)

www.BayesiaLab.com 35
 Structures in DAGs
Indirect Connection
• Implication for Causality
• A causes B via node C.
• Implication for Association
• Marginally (or unconditionally), A and B are dependent. This means that without
knowing the value of C, learning about A informs us about B and vice versa, i.e. the
path between the nodes is unblocked and information can flow in both directions.
• Conditionally on C, i.e. by setting Hard Evidence on (or observing) C, A and B become
independent. In other words, by “hard”-conditioning on C, we block the path from A
to B and from B to A. Thus, A and B are rendered independent, given C.

BayesiaLab.com 36
 Structures in DAGs
Indirect Connection

Marginally dependent/open
Conditionally independent/blocked

BayesiaLab.com 37
 Structures in DAGs
Common Cause
• Implication for Causality
• C causes both A and B
• Implication for Association
• Marginally (or unconditionally), A and B are dependent, i.e. the path
between A and B is unblocked.
• “Hard”-conditioning on C renders A and B independent. In other words,
if we condition on the common cause C, A and B can no longer provide
information about each other.

www.BayesiaLab.com 38
 Structures in DAGs
Common Cause

Marginally dependent/open
Conditionally independent/blocked

BayesiaLab.com 39
 Structures in DAGs
Common Effect (Collider)
• Implication for Causality
• C is the common effect of A and B.
• Implication for Association
• Marginally (i.e., unconditionally), A and B are independent, i.e. the
information flow between A and B is blocked.
• Conditionally on C — even with Virtual or Soft Evidence — A and B become
dependent.
• If we condition on the collider C, information can flow between A and B, i.e.
conditioning on C opens the information flow between A and B.

www.BayesiaLab.com 40
 Structures in DAGs
Common Effect

“Explaining Away”
Intercausal
Reasoning

Marginally independent/blocked
Conditionally dependent/open

BayesiaLab.com 41
 Mathematical Formalism
The BayesiaLab Workflow Research Software

BayesiaLab.com 42
 A desktop software for:
• encoding
• learning
• editing
• performing inference
• analyzing
• simulating
• optimizing
with Bayesian networks.

BayesiaLab.com 43
Constructing a Bayesian Network Model
Without Data!
 Map of Analytic Modeling

Reasoning with Theory

BayesiaLab.com 45
 Model Construction
Overview
1. Diagnostic Model
• Search for what can go wrong.
• Create a check-list to minimize risk.
• Identify “most probable” explanation.
2. Optimization Model
• Maintenance and repair under constraints.

BayesiaLab.com 46
 Model Construction
Key Considerations
• We need to think about
how the “world works,”
not how we reason, e.g.,
we are not building a
decision tree.

BayesiaLab.com 47
 Model Construction
Considerations for Diagnostic Model
• Components, conditions
• Function, performance
• Probabilistic, deterministic relationships
• Cost to observe, test
• “Leaks”, i.e, unspecified causes
Considerations for Optimization
• Cost to repair, replace

BayesiaLab.com 48
 Model Construction
Simplified Gasoline Engine Model

Ignition

Compression Fuel

Combustion

BayesiaLab.com 49
 Model Construction
Simplified Gasoline Engine Model

Components

Functions

BayesiaLab.com 50
 Model Construction
Simplified Gas Engine Model

Icon on node indicates

that the node has a cost
to observe.

BayesiaLab.com 51
 Model Construction
Cost of Diagnostic Observation

$ $$ $$$

BayesiaLab.com 52
 Model Evaluation
Pre-Flight Check

BayesiaLab.com 53
 Model Evaluation
100.0%

• Creating a check list using

99.5%
BayesiaLab’s Adaptive
99.0%
Questionnaire.

Expected Reliability
98.5%

98.0%

97.5%

97.0%

96.5%

96.0%

95.5%
0 1 2 3 4 5 6 7 8

No. of Inspection Steps

BayesiaLab.com 54
 Model Construction
Adaptive
Simplified Engine Model
Questionnaire

BayesiaLab.com 56
 Adaptive Questionnaire
on the WebSimulator
BayesiaLab.com 57
 Model Evaluation
Influence Path
Example:
Influence Paths
IP Lighting to
Combustion

IP Lighting

Combustion

BayesiaLab.com 58
 Model Evaluation
Influence Paths

BayesiaLab.com 59
 Model Evaluation
Influence Paths

BayesiaLab.com 60
 Model Evaluation
Influence Paths

BayesiaLab.com 61
 Model Evaluation
Influence Paths

BayesiaLab.com 62
 The Real World

All I have is $50. What can you

do to get me back on the road?

63
This tank needs battle-ready in 12 hours!

64
 Reasoning Under Constraints
Objective
• Achieve maximum reliability within given constraints.
• Establish optimum order of actions for maximum reliability gain.
Optimization Scenarios
• Maintenance and repair
• Within a fixed budget
• Within a limited timeframe
• Managing repair escalation
• By cost
• By difficulty level

BayesiaLab.com 65
 Reasoning Under Constraints
Additional Considerations

• We are now also performing interventions, i.e., repair actions, instead &
of merely observing conditions.
• We have an overall resource constraint that applies to diagnosis and
repair.
• For each component, we need to decide among the following options:
• Neither test nor repair/replace: $0
• Test and then repair/replace if needed $x or $x+$y
• Always repair/replace $y

BayesiaLab.com 66
 Health Policy
Optimization

See webinar recorded on March 16, 2018

BayesiaLab.com 67
 Reasoning Under Constraints
Implementation as a Bayesian Network

BayesiaLab.com 68
 Reasoning Under Constraints
Implementation as a Bayesian Network

Original Condition

BayesiaLab.com 69
 Reasoning Under Constraints
Implementation as a Bayesian Network

Decision

BayesiaLab.com 70
 Reasoning Under Constraints
Implementation as a Bayesian Network

Cost

BayesiaLab.com 71
 Reasoning Under Constraints
Implementation as a Bayesian Network

Outcome

BayesiaLab.com 72
 Reasoning Under Constraints
Implementation as a Bayesian Network

Resource=Budget

BayesiaLab.com 73
 Reasoning Under Constraints
Implementation as a Bayesian Network

Power Supply Combustion

BayesiaLab.com 74
 Reasoning Under Constraints
To be optimized
Implementation as a Bayesian Network

BayesiaLab.com 75
BayesiaLab.com 76
BayesiaLab.com 77
BayesiaLab.com 78
BayesiaLab.com 79
 Reasoning Under Constraints
Further Applications
• Warranty fraud detection
• Exploring improbable but catastrophic scenarios
• Planning of sabotage
• Minimum effort for saboteur
• Maximum effort to repair
• Easy to deny

BayesiaLab.com 80
 Implementation Example

BayesiaLab.com 81
 BEST
Bayesian Expert System for
Troubleshooting
• bayesia.com/best-overview

BayesiaLab.com 82
 BayesiaLab Trial
Try BayesiaLab Today!
• Download Demo Version:
www.bayesialab.com/trial-download
• Apply for Unrestricted Evaluation Version:
www.bayesialab.com/evaluation

BayesiaLab.com 83
Seminar on Intelligence Analysis
Virginia Tech Applied Research Center in Arlington on September 11

84
 BayesiaLab Courses Around the World in 2018
• August 29–31
London, UK
• October 29–31
Chicago, IL
• December 4–6
New York, NY

Learn More & Register: bayesia.com/events

stefan.conrady@bayesia.us 85
 6th Annual BayesiaLab Conference in Chicago
November 1–2, 2018
Chicago

BayesiaLab.com 86
 Thank You!
stefan.conrady@bayesia.us BayesianNetwork

linkedin.com/in/stefanconrady facebook.com/bayesia

BayesiaLab.com 87