Content uploaded by Stefan Conrady
Author content
All content in this area was uploaded by Stefan Conrady on Feb 13, 2019
Content may be subject to copyright.
INTERFACES
Vol. 47, No. 1, January–February 2017, pp. 1–21
http://pubsonline.informs.org/journal/inte/ ISSN 0092-2102 (print), ISSN 1526-551X (online)
THE FRANZ EDELMAN AWARD
Achievement in Operations Research
Bayesian Networks for Combat Equipment Diagnostics
David Aebischer,aJohn Vatterott, Jr.,aMichael Grimes,aAndrew Vatterott,aRoderick Jordan,aCarlo Reinoso,aBradford
Alex Baker,aWilliam D. Aldrich,aLuis Reinoso,aRodolfo Villalba,aMichael Johnson,aChristopher Myers,aStefan
Conrady,aJoseph A. Tatman,aSuzanne M. Mahoney,aDarrin L. Whaley,aAmanda B. Heplera
aU.S. Army Communications Electronics Command, Aberdeen, Maryland 21001
Contact:
david.a.aebischer.civ@mail.mil (DA), johnjr@stltrades.com (JV), mgrimes@vettechgrp.com (MG), andrewv@stltrades.com (AV),
roderickj@stltrades.com (RJ), carlor@stltrades.com (CR), abbaker@vettechgrp.com (BAB), billa@stltrades.com (WDA),
luisr@stltrades.com (LR), rudyv@stltrades.com (RV), michaelj@stltrades.com (MJ), chrism@stltrades.com (CM),
stefan.conrady@bayesia.us (SC), jatatman@innovativedecisions.com (JAT), smmahoney@innovativedecisions.com (SMM),
dlwhaley@innovativedecisions.com (DLW), abhepler@innovativedecisions.com (ABH)
https://doi.org/10.1287/inte.2016.0883
Copyright: ©2017 INFORMS
Abstract. The lives of U.S. soldiers in combat depend on complex weapon systems and
advanced technologies. In combat conditions, the resources available to support the oper-
ation and maintenance of these systems are minimal. Following the failure of a critical
system, technical support personnel may take days to arrive via helicopter or ground
convoy—leaving soldiers and civilian experts exposed to battlefield risks. To address this
problem, the U.S. Army Communications Electronics Command (CECOM) developed a
suite of systems, Virtual Logistics Assistance Representative (VLAR), with a single pur-
pose: to enable a combat soldier to maintain critical equipment. The CECOM VLAR team
uses an operations research (OR) approach to codifying expert knowledge about Army
equipment and applying that knowledge to troubleshooting equipment diagnostics in
combat situations. VLAR infuses a classic knowledge-management spiral with OR tech-
niques: from socializing advanced technical concepts and eliciting tacit knowledge, to
integrating expert knowledge, to creating an intuitive and instructive interface, and finally,
to making VLAR a part of a soldier’s daily life. VLAR is changing the Army’s sustainment
paradigm by creating an artificial intelligence capability and applying it to equipment
diagnostics. In the process, it has generated a sustainable cost-savings model and a means
to mitigate combat risk. Through 2015, VLAR saved the Army $27 million in direct labor
costs from an investment of $8 million by reducing the requirement for technical support
personnel. We project additional direct costs savings of $222 million from an investment of
$60 million by the end of 2020. Most importantly, VLAR has prevented an estimated 4,500
casualties by reducing requirements for helicopter and ground-convoy movements. This
translates to short- and long-term medical cost savings of over $9 billion. In this paper, we
discuss the OR methods that underpin VLAR, at the heart of which lie causal Bayesian
networks, and we detail the process we use to translate scientific theory and experiential
knowledge into accessible applications for equipment diagnostics.
Keywords:
expert systems
•
military
•
Bayesian networks
•
diagnostics
•
knowledge elicitation
U.S. Army soldiers in combat zones face numerous
challenges each day. One such challenge is operating
and maintaining complex electronic weapon systems.
Soldiers depend on operational command, control,
communications, computers, intelligence, surveillance,
and reconnaissance (C4ISR) equipment to conduct
operations against the enemy and for life-support
functions. The U.S. Army’s Communications Electron-
ics Command (CECOM) is responsible for life cycle
management of these C4ISR systems and for con-
necting combat soldiers, wherever they are, to the
C4ISR knowledge base. When equipment fails, lives are
in danger; therefore, fast and accurate diagnosis and
repair of equipment may mean the difference between
1
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
2Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
life and death. For that reason, we developed a suite
of systems, the Virtual Logistics Assistance Represen-
tative (VLAR), for use by soldiers in extreme com-
bat conditions. For example, consider a typical com-
bat situation where VLAR might be used: a soldier is
engaged with the enemy, is in the dark and in freezing
rain, has limited knowledge of a piece of equipment
that has failed, and has no tools to fix the problem,
but has fellow soldiers who are depending on him or
her to take the necessary steps to bring their life-saving
equipment back into operation. With such a situation
in mind, the VLAR knowledge-engineering team (here-
after referred to as the team) tried to think counter-
factually as it developed VLAR. Team members would
constantly ask each other what would have happened
differently had a soldier been able to repair a piece of
critical equipment. Could the ability to enable a crit-
ical diagnosis and repair have prevented an injury or
death? With VLAR, a soldier can diagnose and repair
the majority of problems without outside assistance;
without VLAR, external assistance is nearly always
required.
The conflicts in Iraq and Afghanistan brought sig-
nificant changes in enemy tactics and caused a prolif-
eration of new systems designed to counter those tac-
tics. Most technologies associated with these systems
were geared toward nonlinear battlefield scenarios that
fell outside the realm of existing Army equipment and
training. The complexity of these systems necessitated
making significant investments in civilian contractor
personnel for on-site technical support. This contractor
logistics support (CLS) strategy was effective in pro-
viding an immediate knowledge base, but ineffective
and inefficient as an on-site support strategy, primar-
ily due to the prevalence of remote combat outposts
(COPs) in Afghanistan and their associated logistical
challenges, as we discuss next in Combat Outposts. The
CLS strategy is also costly, averaging about $500,000
per person-year. As a result of the Afghanistan conflict,
CECOM was presented with a new and multivariate
challenge to providing effective and efficient support.
Combat Outposts
Enemy tactics in Afghanistan compel Army comman-
ders to place small troop formations of 15–20 person-
nel on COPs in remote areas, which are characterized
by both high concentrations of indigenous personnel
and high enemy activity. COPs are frequently in moun-
tainous areas, inaccessible by ground vehicles and
separated from the sustaining base of technical person-
nel because of a combination of enemy threat, geogra-
phy, and environmental conditions. A three-to-five-day
resupply cycle by helicopter or ground convoy is typi-
cal for these locations. The knowledge gap and physical
distance between personnel at the sustaining base and
personnel at the COPs can result in a catastrophic situ-
ation if critical equipment fails between cycles. VLAR’s
primary purpose is to close that knowledge gap.
COPs are small in area, usually about the size of a
football field, with a small, flat cleared area for heli-
copters to take off and land. They are generally staffed
by infantry personnel who are highly trained in com-
bat tactics but have limited experience in the operation
and maintenance of electronics systems. To be effec-
tive, COP personnel must be autonomously capable of
attacking and defending against the enemy. In addi-
tion to a robust mixture of light weapons, the COP
equipment suite includes a satellite communications
(SATCOM) system, a diesel engine-driven electrical
power generator, a long-range surveillance and target-
ing system, a force-protection camera system, a vari-
ety of tactical radios, and a variety of soldier-operated
counter-improvised explosive device systems. If any
of these systems fail, troops are at risk and forced to
rely on secondary or tertiary systems that may reduce
their combat effectiveness. Failure of the main electric
power generator, for example, would affect all electri-
cally powered systems on the COP, causing troops to
use battery-powered backup systems, such as tactical
radios, for command and control. When the batteries
discharge on these backup systems, the combat unit is
in an emergency situation. Any misdiagnosis of equip-
ment faults exacerbates the problem—initiating a chain
of events that lengthen system downtime and put the
lives of COP personnel at risk. To avoid such risks, COP
equipment problems must be diagnosed quickly and
accurately.
Mitigating the Risk with VLAR
Recent U.S. Department of Defense budget reductions
forced CECOM to realize that complete dependence on
the previous CLS strategy would no longer be afford-
able or sustainable. In addition, external support visits
were inefficient. The locus of equipment knowledge
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 3
had to shift to a better balance between on-site and
external support. The Army undertook rebalancing
efforts to rightsize its support personnel mix; however,
the rightsizing effort needed an enabling mechanism,
ensuring that the correct mix of off-site experts was still
available, but establishing the primacy of on-site sol-
dier maintenance tasks. VLAR, which uses an innova-
tive OR process to codify CECOM’s expert-knowledge
base and make it usable for soldiers, is that enabling
mechanism. The primary benefit of using VLAR is
that it enables fast and accurate equipment diagnoses
at the point of need. From this foundation of accu-
rate field-level information, we examine a number of
other positive effects, including lessened risk for sol-
diers and civilians in combat environments, increased
operational availability, cost savings, and more effec-
tive training. To understand these positive effects, we
must first understand why VLAR is successful in trans-
ferring usable knowledge to soldiers at the point and
time of need. To do so, we examine the process of mod-
eling the problem domain and engineering knowledge.
Process and Methods
Three tools were necessary to build VLAR: causal
Bayesian networks (CBNs) to represent the domain
knowledge in a qualitative and quantitative frame-
work, a knowledge-engineering process to extract and
codify tacit knowledge from experts, and a graphi-
cal user interface (GUI) to make the CBN knowledge
layer usable and instructive for the soldier. These tools
enable a VLAR system that puts the collective knowl-
edge and experience of CECOM personnel at a sol-
dier’s fingertips and allows that soldier to identify
faults and underlying causes quickly and accurately.
Causal Bayesian Networks
CECOM evaluated a number of methodologies that
would be mathematically precise, robust enough to
handle complex diagnostics, and flexible enough to
handle different types of input data. CBNs have an
inherent ability to represent complex, multivariate
relationships, and to use machine-learning techniques
for updating models. The graphical representations
within them facilitate an understanding of the how and
why of the causal path structures to enable soldiers
to learn as they perform diagnostics. Therefore, CBNs
meet all the criteria for building a system that can
effectively encode tacit expert knowledge, efficiently
complement this knowledge with case data, and make
these data compatible with an intuitive user interface
(Pearl 2015).
CBNs consist of a qualitative (i.e., structural) com-
ponent and a quantitative component. The qualitative
component is a directed acyclic graph (DAG) in which
the nodes represent variables in the problem domain
and the arcs represent cause-and-effect relationships
between those variables. The quantitative component
sets joint probability distributions for each state in each
variable using conditional probability tables (CPTs)
and conditional probability statements, for example,
the probability of A being true, given that B is true.
The quantitative component is mathematically rooted
in Bayes’ rule and the chain rule (Pearl 1988). In the
appendix, we provide details on these mathematical
underpinnings.
The basic building block of a CBN is the parent-
child relationship. A node’s calculated probability dis-
tribution is based on its distribution CPT and that of
its parent(s). Figure 1depicts a portion of a typical
VLAR CBN, which models the complex thermody-
namic relationships in play during the operation of a
diesel engine.
This network shows that the parent node (Injector
Pump Timing) has a causal effect on its child node
(Black Smoke), which is denoted by the directed arc
going from Injector Pump Timing to Black Smoke.
Figures 2(a) and 2(b) show the CPTs for each node.
The parent node is a root node (i.e., it has no par-
ents). The CPT for this node is, therefore, the prior
probability of each variable state. For example, the
prior probability that Injector Pump Timing is Normal
is 90 percent. This is shown in the CPT and in the
node probability distribution. (Note that in Figures 2(a)
and 2(b), the nodes names, Injector_Pump_Timing1
and Black_Smoke1, respectively, are assigned by the
Bayesian network software we used. In the network
and throughout this paper, the generic descriptors,
Injector Pump Timing and Black Smoke, are used to
refer to these nodes.)
The child node is a leaf node (i.e., it has no children).
Therefore, its CPT is the probability of each state given
the state of its parent(s). CPTs read like a conditional
statement. For example, in Figure 2(b), we can read the
first line of the CPT for the Black Smoke node as the
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
4Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
Figure 1. (Color online) This subset of a typical VLAR CBN for diesel engine diagnosis depicts the basic components and
terminology. The probability distributions for the variable states are marginal; that is, the probabilities (e.g., Black Smoke is
Yes is 1.08 percent) are for each state with no evidence entered.
Variable, parent node
Probability distribution
Directed arc
Probability distribution
Variable, child node
Variable states
Variable states
Figure 2(a). (Color online) The CPT for this parent node
shows the prior probabilities of all states of Injector Pump
Timing.
probability that Black Smoke is Yes, given that Injector
Pump Timing is Normal, is 0.2 percent, and the proba-
bility that Black Smoke is No, given that Injector Pump
Timing is Normal, is 99.8 percent.
Referring back to Figure 1, the probability distri-
bution in the Black Smoke node shows marginal dis-
tribution as Yes being 1.08 percent and as No being
98.9 percent. This marginal distribution is computed
by summing the probabilities of each state of the child
node, given its parent’s state. For example, the proba-
bility of Black Smoke is Yes, given Injector Pump Tim-
ing is Normal (0.2 percent), times the probability that
Injector Pump Timing is Normal (90 percent), is 0.18
percent. This process is repeated for each state of the
parent node and then summed. See Figure A.2 in the
appendix for details on these calculations.
Figure 2(b). (Color online) This CPT shows the probability
of a Yes or No state for Black Smoke, conditioned on the state
of Injector Pump Timing.
Although the arc direction is from cause to effect,
information flows in both directions. The marginal
probability distributions of all nodes in the causal
Bayesian network are updated in light of new evidence
using a propagation algorithm derived from Bayes’
theorem (Pearl 1982). This is shown when setting evi-
dence in either the parent or child node. The user
enters evidence by clicking the appropriate value of a
state variable. Figure 3shows how evidence set in the
parent node changes the probability distribution in the
child. Note how the probability distributions change in
the child node exactly according to the applicable row
in the CPT. The child node now shows the conditional
probability of Black Smoke, given that Injector Pump
Timing is Advanced.
Conversely, evidence entered in the child node
changes the probability distribution of the parent node
(Figure 4).
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 5
Figure 3. (Color online) When the user enters evidence for the parent node (Injector Pump Timing is Advanced), the
probabilities of the child node states (Black Smoke) are updated.
Figure 4. (Color online) If the user instead enters evidence for the child node (Black Smoke, Yes), the probabilities of the
parent node states (Injector Pump Timing) are updated. This update is computed by finding the joint probability that Black
Smoke is Yes and Injector Pump Timing is Normal and dividing that value by the marginal probability of Black Smoke. This
calculation is found using the chain rule; see Figure A.2 in the appendix.
Injector Pump Timing Injector Pump TimingBlack SmokeBlack Smoke
P(IJP)
Normal
Ye s
Sum
Ye s
Ye s
Advanced
No
No
90%
0.2%
99.8% 89.82% 1.08%
No
98.92%
0.18%
0.66%
11%
6%
Retarded
4%
89%
Ye s
No
6%
94%
5.34%
0.24%
3.76%
P(BS | IJP) P(BS, IJP)
= P(IJP)*P(BS | IJP) Over IJP
P(BS) = ∑P(BS, IJP)
Normal
16.7% = 0.18/1.08
Advanced
Retarded
61.1%
22.2%
Normal
90.8%
Advanced
Retarded
5.4%
3.8%
P(BP, IJP)
P(BS)
P(IJP | BS) =
This omnidirectional functionality makes CBNs a
very natural fit for diagnostics when there is utility in
reasoning forward from contextual test evidence and
in reasoning backward from five-senses observation.
DAGs establish the structure and the causal relation-
ships between variables; the CPTs parameterize those
relationships. Conditional probability statements are
at the heart of Bayesian networks; by extension, they
are at the heart of VLAR. Reasoning with Bayesian
networks is possible because the structures approxi-
mate human reasoning (Pearl 1988).
We will first closely examine the qualitative aspect of
CBNs (i.e., DAGs). DAGs use three basic node and arc
structures: indirect effect, common effect with multiple
causes (collider), and common cause of multiple effects
(confounding). These structures, used individually or
in combination, can be used to model any relationship
between variables in any domain. This is because each
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
6Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
node structure approximates a unique aspect of human
reasoning. The unique properties of these structures
are best described in terms of their marginal and
conditional dependencies and independencies (Pearl
2015). In Figures 5(a)–5(c), we show how the parent
and child nodes used in Figure 1are connected into
a more comprehensive CBN for diesel engine fuel-
system diagnostics. We use that expanded network to
demonstrate each structure, explain how it relates to
human reasoning, explain how it is used to describe
specific relationships, and show each structure’s utility
for VLAR. This expanded view of the CBN for diag-
nosing diesel engines depicts the three basic structures:
indirect effect (Figure 5(a)), common cause (with mul-
tiple effects) (confounder) (Figure 5(b)), and common
cause (of multiple effects) (collider) (Figure 5(c)).
The indirect-effect structure is useful for compact
representation of problem domains. Figure 6shows a
generic indirect-effect structure and an indirect-effect
structure abstracted from our expanded network (Fig-
ure 5(a)). The key to understanding this structure is to
relate it to a chain of events. Humans reason that if
events are causally connected, or temporally ordered,
we can assume evidence for one event means all events
preceding that event are true, or satisfied. For example,
consider a car that does not start. We reason that if we
turn the key in the ignition switch and the engine turns
over, all events necessary to make the engine turn over
are satisfied. We would not, then, go back and check if
the ignition switch was operating. Instead, we would
investigate other paths, such as no fuel, that would
cause the engine not to start. The key to implement-
ing this structure is to understand its marginal and
conditional dependencies and independencies. A rep-
resents a cause, B represents an intermediate node,
and C represents an effect. A and C are marginally
dependent and conditionally independent. With no
evidence entered in B, information can flow from A
to C and from C to A through an open path through the
intermediate node. We can reason forward, for exam-
ple, knowing an electrical signal at A has an effect
on an electrical signal at B, and B has an effect on C.
Knowing A tells us something about C; conversely,
knowing C tells us something about A. If we know B,
however, and set evidence on it, we no longer need to
know A. Evidence on the intermediate node (B) blocks
the path between cause and effect; the result is that A
and C are independent.
For our expanded network example, knowing the
state of Injector Pump Timing makes any knowledge
about Fuel Injection Pump Alignment of no further
value and makes further alignment testing unneces-
sary. The knowledge about alignment is, in effect,
stored in the joint probability distribution of Injector
Pump Timing. This structure has many ramifications
for troubleshooting, especially of electrical and optical
paths, where we want to eliminate as much wire or
cable as a possible problem in the shortest amount of
time and with the fewest number of tests.
Next, we will examine the common-cause-of-
multiple-effects structure (i.e., the confounding struc-
ture) from a human-reasoning perspective and from
the perspective of its unique dependence and indepen-
dence properties. In human reasoning, given evidence
for a certain effect, we seek out other evidence to sub-
stantiate or refute possible causes for that effect. For
example, if we observe that the grass is wet in our
front yard and we want to determine if the wetness is
the result of overnight rain or a sprinkler left on, we
look across the street at our neighbor’s grass to support
our reasoning process. The properties of a confounder
structure support that type of reasoning. Effects in a
confounder are marginally dependent and condition-
ally independent, given the cause. Figure 7shows a
generic version of a confounding structure on top and
a confounding structure abstracted from our expanded
network (Figure 5(b)) on the bottom.
With no evidence entered in the Cause node and evi-
dence set in Effect 1, information flows from Effect 1
to Effect 2 through the Cause node. Identifying a spe-
cific effect would increase the probability of other com-
mon effects and give us information about the probable
cause. In this case, Effect 1 and Effect 2 are marginally
dependent. If we enter evidence for the cause, however,
this closes the information path from Effect 1 to Effect
2 through the Cause node. Given the Cause, Effect 1
and Effect 2 have become independent. The sequence
in Figures 8(a)–8(d) demonstrates the marginal depen-
dence and conditional independence properties using
the abstraction from our expanded CBN. If the cause is
unknown, information about an effect provides infor-
mation about other effects and increases certainty
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 7
Figure 5. (Color online) (a) Indirect-effect structures imply marginal dependence between a cause (Fuel Injection Pump
Alignment) and an effect (Black Smoke) and independence between a cause and an effect conditioned on the intermediate
variable (Injector Pump Timing); (b) Common-cause (confounder) structures imply marginal dependence between effects
(Black Smoke and Engine Runs Erratically) and independence between effects conditioned on the cause (Fuel Injector Nozzles);
(c) Common-effect (collider) structures imply marginal independence between causes (Injector Pump Timing and Fuel Injector
Nozzles) and dependence between causes conditioned on the effect.
(a)
(b)
(c)
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
8Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
Figure 6. (Color online) In indirect-effect structures, cause and effect are marginally dependent and conditionally independent
if no evidence is entered in the Intermediate node (B); Cause (A) and Effect (B) are independent if evidence is set for B.
about the cause. If the cause is known, information
about other effects is of no further value.
The sequence shown in Figures 8(a) to 8(d) shows
how we can use these unique properties to model
complex relationships in VLAR and then use the
model to impact diagnostic processes. One of our
objectives in VLAR is to minimize the number of
troubleshooting steps. Expert knowledge tells us that
Figure 7. (Color online) The top part of the figure shows
a generic confounding structure; the bottom part shows a
confounding structure abstracted from the working network.
The confounding structure encodes marginal dependence
between Effect 1 (Black Smoke) and Effect 2 (Engine Runs
Erratically) and conditional independence between Effect 1
and Effect 2.
Black Smoke and Engine Runs Erratically are both
robust, five-sense indicators of possible Fuel Injector
Nozzle issues. We build a structure to reflect that
possibility; therefore, low-cost tests, such as simple
observations and cylinder drop tests (e.g., shutting
off the fuel supply to each cylinder one at a time),
always precede more expensive tests (e.g., injector
nozzle testing). Specifically, in this case, observing
Figure 8(a). (Color online) The diagram illustrates marginal
distributions with no evidence entered.
Figure 8(b). (Color online) Entering evidence for Black
Smoke changes the probability distribution for Engine Runs
Erratically. This demonstrates marginal dependence between
causes. Information is propagated from child node to child
node via an open path through the parent node.
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 9
Figure 8(c). (Color online) Evidence entered for the common
cause (Fuel Injector Nozzles) closes the path between the
child nodes.
Figure 8(d). (Color online) Entering evidence for Black
Smoke does NOT change the probability distribution for En-
gine Runs Erratically. The information path through the Fuel
Injector Nozzles node is blocked.
Black Smoke transmits information about the Engine
Runs Erratically condition, which is useful in reducing
uncertainty about the Fuel Injector Nozzles. Knowing
the cause obviates the need to do any further testing
for indicators. The CBN network and the user are both
focused on determining the cause.
We build on that notion by explaining the final basic
structure: the common effect with multiple causes,
or collider structure. This structure also approximates
human reasoning with unique properties. As humans,
we naturally process causal information by eliminat-
ing or reducing the probability of noncauses based on
the strength of evidence for a known cause. Return-
ing to the car-starting example, if we know that a car
does not start and that the battery fails a capacity test,
then our belief in other possible causes (e.g., related
to the starter or ignition switch) will decrease signif-
icantly. Colliders are designed with properties that
enable us to represent such relationships. Colliders are
marginally independent and conditionally dependent,
given the effect. Figure 9shows a generic form of the
collider structure on the top and a collider structure
Figure 9. (Color online) A collider structure encodes mar-
ginal independence and conditional dependence between
causes with no evidence set for the effect and dependence
between causes when there is evidence set for the effect.
abstracted from our expanded network (Figure 5(c)) on
the bottom.
In this collider structure, either Cause 1 or Cause 2
can trigger the Effect. With no evidence entered on the
Effect node, there is a closed path between Cause 1
and Cause 2. Knowing Cause 1 tells us nothing about
Cause 2 and vice versa. These causes are marginally
independent. Conversely, when evidence is entered
for the Effect node, evidence set on either cause will
update the probabilities of the other cause(s). Given the
Effect, Cause 1 and Cause 2 are then dependent. The
sequence in Figures 10(a)–10(c) illustrates the inherent
explaining-away feature of the collider structure; that
is, it shows how colliders transmit information with
evidence set and (or) not set.
Colliders enable the explaining-away feature of
Bayesian networks. The explaining-away phenomenon
occurs when confirming evidence for one cause of an
effect explains that effect, causing probabilities for the
other possible causes of that effect to decrease. This fea-
ture is unique to Bayesian networks and is not available
through other classical statistical methodologies. The
collider effect that provides the explaining-away fea-
ture is enabled only when an evidence finding has been
entered into the common-effect node or a descendant
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
10 Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
Figure 10(a). (Color online) The marginal distribution is
shown without any evidence entered for the nodes.
Figure 10(b). (Color online) Entering evidence for Injector
Pump Timing changes the distribution of the common effect,
but not of the other cause. This demonstrates that the com-
mon causes are marginally independent. The path from
Injector Pump Timing to Fuel Injector Nozzles is blocked by
Black Smoke.
Figure 10(c). (Color online) Conditioning on the collider by
setting evidence about Black Smoke opens the path between
causes through the effect. Information flows from Injector
Pump Timing through Black Smoke to update the probabil-
ity of Fuel Injector Nozzles.
of the common-effect node. This is known as condi-
tioning on a collider, and it tracks the natural effect of
human reasoning and the logic of processing evidence,
not just an anomaly of using Bayesian networks. This
collider effect is part of the concept of directional sepa-
ration (i.e., d-separation), which Pearl (1982) describes
in detail.
In these three basic structures, we have tools with
which to represent the most complicated relationships,
illuminate the nexus of theoretical and experiential
knowledge, and elucidate the finer points of trou-
bleshooting. These structures are the enabling mech-
anism that makes the artificial intelligence in VLAR
less artificial. To specify those relationships, however,
we need a framework and robust process for using
these tools to build a CBN—from its structural under-
pinnings to the CPT. Building the structure and the
CPT requires omnidirectional and multidimensional
thinking. This is a daunting task considering the sheer
number of ways in which machinery can fail and
the physical principles of electrical theory, wave the-
ory, radio frequency, and thermodynamics that could
be applicable. To take full advantage of the omni-
directional reasoning power of CBNs, however, we
must first determine how to quantify the knowledge
from experts so it subsumes that theoretical and expe-
riential knowledge. Codifying the knowledge from
experts into CBNs requires a structured and efficient
knowledge-engineering process.
Knowledge Engineering with CBNs
CECOM has little or no hard data on field usage and
troubleshooting events—only artifacts of supply trans-
actions that cannot be traced back to actual events
and whose trouble-ticket records include limited con-
text. However, CECOM has a robust cadre of experts
who are stationed worldwide and have years of experi-
ence working with soldiers and equipment in combat
environments. The team had to determine an expert-
knowledge framework and then develop a process
from this framework with which to capture this expert
knowledge in specific equipment domains and on spe-
cific CECOM equipment. It did so using the Define,
Structure, Elicit, Verify (DSEV) process, which we dis-
cuss next.
DSEV is a rigorous process for eliciting knowledge
from subject matter experts (SMEs) to build a CBN.
As Figure 11 shows, DSEV is a cyclical process pre-
ceded by two critical steps. The first step, framing
the problem, ensures that the VLAR diagnostic model
is neither too broad nor too specific. A well-framed
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 11
Figure 11. (Color online) The DSEV process for building CBNs from expert knowledge is named for its key components, a
cycle of defining, structuring, eliciting probabilities for nodes (variables), and verifying a few variables at a time.
problem translates well to logical troubleshooting cate-
gories (e.g., will not start, will not transmit) and serves
to keep the team and the SMEs focused on a work-
able subset of the problem. The second step, readying
the experts, prepares the team and the SMEs for the
elicitation process. This involves orienting the experts
to the qualitative and quantitative aspects of the CBN,
explaining the process of translating expert knowledge
into probabilities, and making the experts aware of
the biases and calibration issues that can affect those
translations. In the DSEV cycle, the team first leads
the experts through a process of defining a small set
of nodes to add to the CBN, adding this set into the
existing CBN structure, eliciting the CPTs for these
nodes, and immediately verifying the functionality of
the expanded model with the experts. This process
continues incrementally—adding sets of nodes to the
model until the CBN is ready for a validation process
that prepares it for use. Finally, the DSEV process con-
tinues for the model in the field by providing a mecha-
nism for collecting case data and then using these data
to refine the model through machine learning (Hepler
et al. 2015).
The team customized the DSEV process in several
ways. First, as part of a ready-the-experts preprocess,
team members followed a robust process for educating
themselves on the operating principles of the system
being modeled. This included studying and analyzing
technical documentation for the system, formal train-
ing on the target system (which the system manufac-
turer furnished), and observation of experts at work.
This process provided credibility for the team and
accelerated knowledge transfer. Second, the team seg-
mented the diagnostic models into two categories of
knowledge. The first category describes what the vari-
ous on-equipment indicators and sensors tell the user
about why the system is not functioning correctly. It
also describes how this information can be used to
reduce uncertainty through a process of asking first-
best questions. These questions refer to a process by
which we compute the amount of entropy (i.e., uncer-
tainty), which is reduced by asking a specific ques-
tion and entering the answer to that question as evi-
dence. The intent is to remove the maximum amount of
uncertainty using the fewest number of questions. The
standard DSEV process involves brainstorming a list
of nodes that provide information about a target node.
The team refined this process by brainstorming lists of
causal relationships between failure modes and indica-
tors, taking into account all theoretical and operating
principles. This is the first step in using artificial intel-
ligence processes to build a network that approaches
a diagnostic event as an expert does—by eliminating
unnecessary tests and isolating the problem from a
theoretical and experiential perspective. The second
knowledge category is the process of troubleshooting,
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
12 Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
or making the art of troubleshooting accessible in the
presentation.
VLAR Graphical User Interface
Given the relationships specified in the CBN, the VLAR
GUI describes how the team translates these relation-
ships into simple text. This provides a means for sol-
diers to directly access the CBN knowledge in an
intuitive and instructive manner. This GUI produced
a dramatic training effect, in addition to expected
improvements in troubleshooting speed and accuracy.
The VLAR GUI is the presentation layer of this
advanced OR system, translating and displaying the
knowledge and analytic skills of human experts from
the underlying CBNs. Users who do not have prior
experience with built-in theoretical principles must be
able to benefit from using them. They must also under-
stand how and why the theoretical underpinnings of
the CBN alter the manner in which specific diagnoses
are approached. The requirements of the GUI, and the
assumptions we made in its development, are the same
as those we made for the overall VLAR tool: it must
assume austere combat conditions, high enemy threat,
and limited tools, experience, and time. We assume
that CBNs, in their basic form, are extremely complex
and not intuitive for an untrained user. The VLAR
GUI removes the challenge of users directly interfacing
with the CBN by presenting information in an intuitive
and logical manner, leveraging the inherent structural
properties—the marginal and conditional dependen-
cies and independencies described above—of CBNs
as if a seasoned domain expert was walking a sol-
dier through an advanced diagnostic procedure. This
process is achieved through the careful crafting and
synthesizing of displayed text and graphics. However,
the user display is only part of the GUI. When a user
answers the binary-choice or multiple-choice questions
displayed on the GUI, these answers are recorded as
evidence in the CBN. To accomplish this function, the
GUI uses a middle layer, a system net, to act as the
interface between the GUI and all CBNs in the network.
VLAR CBNs provide an exhaustive model of faults
for a given piece of equipment. This model requires
many CBNs to make up a complete and comprehen-
sive diagnostic network. For example, a VLAR suite for
a diesel generator would, at a minimum, have sepa-
rate CBNs for faults relating to engine starts, fuel sys-
tems, output power, and safety circuits. The system
net allows the VLAR GUI to move seamlessly in and
between those CBNs. This function is critical because
multiple root-cause problems are commonplace with
CECOM equipment. It is often necessary to fix one
problem, or at least ascertain there is no problem,
before moving to the next. A typical engine-start prob-
lem may have both electrical and mechanical issues;
thus, accurately diagnosing the problem requires tran-
sitioning between CBNs. The system net enables this
transition, but makes it transparent to the user. Based
on user input to the GUI, the system net determines
the most appropriate CBN to start the diagnostic pro-
cess and continues to monitor and analyze changes in
CPTs across all CBNs in the network, ensuring that the
CBN and GUI are synchronized. The system net also
controls images and text files displayed in the GUI and
the order and sequence in which they are displayed to
the end user. The order and sequence are dictated by
most-probable states, and the most-probable states are
a function of evidence propagated through the CBN in
the form of joint probability distributions.
Thus far in this paper, we have described how the
team uses the powerful CBN tool and the associated
knowledge-engineering processes to address the prob-
lem statement. Soldiers can learn VLAR without any
previous training and start troubleshooting immedi-
ately, making it invaluable in combat situations in
which speed and precision are critical. We now exam-
ine specific cases for the VLAR methodology.
VLAR Methodology: Solutions for Combat
Soldiers
Over the past four years, the team has undertaken
diagnostics development projects for three equipment
types: tactical generators, High Antennas for Radio
Communication (HARCs), and Satellite Transportable
Terminals (STTs). In each of the initial development
projects undertaken, the team had to adapt the DSEV
and knowledge-engineering processes to represent one
or more unique physical properties in such a manner as
to preserve the integrity of the theory and to maintain
the soldier usability and accessibility imperatives.
The team’s methodology is to apply the combination
of CBN tools and knowledge-engineering tools (as we
describe in the previous sections) to Army equipment
problems that deal with uncertainty, and specifically
to those problems that necessitate codifying complex
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 13
principles of physics and making them accessible to
soldiers. The team made three key assumptions that
were common to all projects. First, all tools would be
designed to operate with standalone diagnostic capa-
bility, with no requirement for network or Web ser-
vices. All knowledge and diagnostic functionality had
to reside on the soldier’s mobile device. Standalone
functionality was critical because any failure that effec-
tively disconnected the unit’s communications net-
work, such as a generator or satellite communications
(SATCOM) failure, would make online or network-
connected diagnostic solutions unusable. Second, from
a design perspective, the team assumed combat sce-
narios at a COP: limited electrical power, no network
connectivity, limited tools and test equipment, lim-
ited experience, and imminent threat. By designing to
this standard, the team could refine the interface to a
life-saving mode (i.e., a soldier might have only min-
utes to restore the system to operational status) while
ensuring functionality in all scenarios. Third, the team
assumed limited hard data and the primacy of expert
knowledge for modeling equipment diagnostics. For
all projects, the development effort focused on identi-
fying and understanding the theories, rules, and laws
governing the operation of the systems and on inte-
grating these into the CBNs. This was a significant
challenge, because a deep understanding of theoreti-
cal principles would routinely create what seemed to
be counterintuitive temporal relationships. In develop-
ing VLAR, the team overcame this challenge by apply-
ing the knowledge-engineering process to a group of
experts with a unique combination of theoretical and
experiential knowledge. Next, we explain the VLAR
knowledge-engineering processes, demonstrate spe-
cific VLAR tools, and compare VLAR use versus non-
VLAR use in diagnostic scenarios.
The CECOM equipment knowledge base resides in
a diverse group of technical personnel (e.g., engi-
neers, trainers, and field support personnel) assigned
to support Army units worldwide and tasked with
providing direct and indirect technical assistance to
soldiers on CECOM systems. These technical experts
gained this rich set of knowledge from formal train-
ing and, collectively, hundreds of years of experience
in maintaining equipment in combat situations. The
team applied DSEV processes to elicit from person-
nel in this group some of the tacit knowledge of
troubleshooting procedures—specifically those proce-
dures that were confounded by theoretical and oper-
ating principles that soldiers did not understand or
know well. These principles of electricity, electronics,
fiber optics, and radio frequency (RF) can be used to
good effect to analyze any aspect of any CECOM sys-
tem. However, failure to understand how to do such
analysis can lead an inexperienced soldier to incor-
rect conclusions, thereby increasing troubleshooting
time and reducing diagnostics accuracy. The output
of the knowledge-engineering and extraction process
for these projects was a CBN-based artificial intelli-
gence process for diagnostic procedures that closely
matched the nuanced approach taken by technical per-
sonnel and field experts who have substantial experi-
ence working with and applying those laws to system
diagnosis.
VLAR Knowledge Extraction
Expert-knowledge extraction for VLAR is a process by
which we develop a compact representation of the joint
probability distribution from a diagnostic cause-and-
effect relationship. Experts help us to model systems
based on how they fail. They first help us to graph these
failure modes into one of the three common node and
arc structures. We use combinations of the basic struc-
tures to create CBNs for each subsystem—exhaustive
of all possible failure modes for that subsystem. The
experts then assist in filling in the CPTs by estimat-
ing the prior probabilities for each node state and esti-
mating the conditional probabilities for each state in
the parent-child relationship. This process is consistent
across all VLAR efforts. The difficulty lies in accurately
representing the mix of experiential and theoretical
knowledge. A good example is found in the analy-
sis of causal paths within the CBN for the excitation
circuit on the Army’s 30-kilowatt (30 kW) generators.
Ohm’s Law and Kirchhoff’s Law govern the operation
of this circuit (Figure 12). Examining this circuit in
those terms is useful to illustrate how they would affect
decisions made while troubleshooting and how encod-
ing those laws in CBNs can be both helpful and instruc-
tive to soldiers. Consider a case in which the excitation
circuit windings are shorted. To inexperienced techni-
cians without VLAR, this condition would appear as
a zero voltage signal at a linear test point, and they
would incorrectly interpret it as a strong indication of
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
14 Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
Figure 12. (Color online) This example shows an excitation circuit as modeled in the CBN by codifying Ohm’s Law and
Kirchoff’s Law. For brevity, we omit a discussion of the input to these processes, as indicated by the arrows.
a problem with the control circuit or the electrical path
leading to that test point. This interpretation would
delay isolation of the correct problem and could trig-
ger a supply requisition for the incorrect part. Through
VLAR’s knowledge-extraction process, this condition
appears to the soldier as it would to an experienced
technician: as an anomalous voltage drop across the
control circuit. Moreover, the soldier would correctly
interpret it as a loss of the normal voltage drop across
the excitation winding, and would be able to correctly
isolate the component in minimum time and correct
the initialization of the supply system.
The Field_Flash node in Figure 12 is set for evi-
dence of fail. This evidence would come from a user
observation of an under-voltage condition and an ini-
tial test of the voltage regulator. In this scenario,
the VLAR GUI would then direct the user to test
the K23_Output_Voltage (following the directions we
show in Figure 13). Note that the Field_Flash collider
has four parent nodes. We focus on two parents—
Field_Flash_Resistance and K23_Output_Voltage—
and the collider. In Figure 12, we show the parents
within circles. The voltage at K23_Output is mini-
mal in this scenario, and the test criteria given in the
GUI would direct the user to record the test as a fail.
Setting evidence for fail in the K23_Output_Voltage
node drives the fail state in the Field_Flash_Resistance
node to be very probable (84.4 percent). This is
consistent with how we previously described con-
ditioning on a collider: K23_Output_Voltage and
Field_Flash_Resistance are conditionally dependent
given evidence for Field_Flash. This probability distri-
bution for Field_Flash_Resistance is counter to a linear,
non-VLAR approach that would interpret zero voltage
at K23 as an indication of a failure in the circuit supply-
ing power to K23 and would trigger an attempt to first
test and analyze that portion of the circuit. This capa-
bility to encode both theory and experience represents
the essential value of VLAR’s OR-based approach to
diagnostics. We find similar examples in other VLAR
products.
The VLAR HARC project demonstrated the team’s
ability to use expert-knowledge extraction to address
specific systemic operational problems. HARC is a
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 15
Figure 13. (Color online) The GUI provides directions for testing the voltage of the K23 circuit.
radio that uses fiber optics to extend the antenna
to high elevations to increase range. The HARC
system mounts on an aerostat platform—with the
ground-based and airborne components of the sys-
tem connected, electrically and optically, by a tether.
Army units in Afghanistan depend on the HARC, and
on other sensors on the aerostat, for critical surveil-
lance and command and control functions. However,
when first deployed, the HARC system’s poor reliabil-
ity caused excessive and unnecessary aerostat recover-
ies, which resulted in low availability of all the aerostat
payloads. The team used its knowledge-engineering
process to build a causal structure for the HARC
system and subsystems. In that process, team mem-
bers discovered that the HARC operation was particu-
larly sensitive to optical loss, or excessively low light-
wave amplitude, and that multiple causes for excessive
optical loss were present. Experts assisted in quantify-
ing lab-setting optical-loss levels and then in estimat-
ing probabilities for optical loss in the dynamic envi-
ronment of the airborne components of the HARC.
The resulting CBN included some previously unob-
served variables as predispositional nodes to account
for each environmental or operational factor affect-
ing the uncertainty of light amplitude. The conditional
probability tables for the HARC VLAR incorporate
the knowledge that external forces (e.g., bending and
stretching) could have nonlinear effects on different
wavelengths in a single strand of fiber. The resultant
CBN was then able to detect subtle levels of optical
degradation and support knowledge-driven aerostat
recovery decisions. This entire knowledge-engineering
process equated to a rigorous process of separating and
isolating ground and airborne faults, which ultimately
made troubleshooting faster and more accurate and led
to increased payload availability.
To illustrate the effectiveness of the HARC know-
ledge-extraction process, consider a common scenario
in which an optical fiber is bent or stretched on the
ground side of the aerostat platform such that it causes
stress on one fiber strand inside a fiber optic cable and
high optical loss on one of the wavelengths active on
that strand. To the radio operator, this appears as a
failure-to-transmit condition for one of the four radios
connected to the system. Without VLAR, the operat-
ing parameters of the system mask the true state of the
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
16 Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
optical fiber. An inexperienced technician would see
evidence of green fiber uplink readings (i.e., accept-
able light levels and wavelength, one of seven active
on the strand, provides the light-level threshold indica-
tion for the entire fiber path from the ground through
the tether to the airborne components. The technician
would also see evidence of a red laser-detector state
(i.e., light level below the acceptable threshold) for one
wavelength on the airborne side and a red fiber down-
link indication. Without VLAR, the technician would
reason that with a fiber uplink above that threshold,
light levels are OK up to the airborne components,
and would incorrectly interpret this condition as a
single-wavelength failure of the laser detector at the
airborne module. This would initiate an unnecessary
aerostat recovery. VLAR builds in expert knowledge on
fiber optics and builds testing processes based on that
knowledge. In this case, experts guided the team by
showing that the command-link wavelength alone is a
poor proxy for fiber optical path health because it is the
lowest wavelength in the strand and that higher wave-
lengths are more susceptible to light loss from bending
and stretching. The team built this knowledge into the
CBN and GUI so that the technician, given this same
evidence, would interpret it as a probable fiber prob-
lem and initiate a series of optical fiber tests to deter-
mine the exact point of fiber failure. This real-world
scenario demonstrates how the VLAR-based approach
saves time and enables the technician to restore the
HARC to full mission-capable status in minimum time
without an unnecessary aerostat recovery. As with
the VLAR generator effort, the solution for HARCs is
rooted in the knowledge-engineering process and in
making complex principles of optical theory accessible
to soldiers who are responsible for equipment mainte-
nance and operations.
Impact and Calculation of Benefits
VLAR is the first and only Army effort to incor-
porate operations research in diagnostic processes
and empower soldiers with new and robust capa-
bilities. For each of its combat equipment applica-
tions, VLAR has helped achieve quantifiable improve-
ments in availability and troubleshooting accuracy by
combat soldiers, reductions in troubleshooting time,
cost savings from personnel reductions and logistics
efficiency, and significant improvements in helping
soldiers understand complex systems. In its initial beta
test in 2012, Army mechanics used VLAR to diag-
nose a 30 kW generator that would not start. Without
VLAR, multiple technicians had attempted to diagnose
this generator problem with no success for over three
months. The team provided the technicians with an
early version of the VLAR 30–60-kilowatt (30–60 kW)
generator application. Without any training on the
application, a technician was able to use VLAR to iso-
late two separate nonstandard and complex faults (one
mechanical and one electrical) and return the system
to fully repaired status within 30 minutes. This suc-
cess was repeated in numerous tests by soldiers using
the completed 30–60 kW generator VLAR application,
culminating in the final user test at the Letterkenny
Army Depot. The Letterkenny test required untrained
soldiers to troubleshoot a population of 180 unservice-
able 30 kW and 60 kW generators. None of these gen-
erators had been evaluated since being returned from
combat operations in Afghanistan and all were des-
ignated for complete overhaul. Soldiers using VLAR
evaluated all 180 generators and determined that only
20 needed to be overhauled. The remaining 160 gener-
ators needed only minor repairs and testing to bring
to them full mission-capable status. This equates to a
30 percent increase in generator availability, which is
attributable to VLAR use. VLAR-enabled troubleshoot-
ing was 100 percent accurate during testing, and the
average troubleshooting time dropped from over one
hour per generator at the beginning of the test to less
than 30 minutes per generator by the end of the test.
The team saw a similar effect with the VLAR HARC
fielding. In Afghanistan, use of VLAR increased HARC
availability rates by 30 percent. VLAR is the primary
reason that no unnecessary aerostat recovery (i.e.,
returning the aerostat to the ground for maintenance
or diagnostic function) has occurred because of HARC
maintenance since the VLAR HARC fielding. In addi-
tion, VLAR users documented a 75 percent decrease
in troubleshooting time, a 60 percent increase in trou-
bleshooting accuracy, and a 100 percent reduction in
part returns, where the part was determined to be ser-
viceable. HARC systems were, at peak, supported by
a team of more than 25 field service representatives
in Afghanistan at a cost of over $12 million per year.
Due to the complex nature of most HARC faults, on-
site visits by these field representatives to remote sites
were usually required. Within six months of fielding
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 17
VLAR, CECOM reduced its technical support person-
nel requirement by 50 percent. By the end of 2016,
HARC was fully supported by aerostat operators and
crews using VLAR. Because of the success achieved
to date, CECOM has designated VLAR as an inte-
gral part of its fleet-management strategy. Specifically,
CECOM mandated VLAR development for 60 per-
cent of the most complex CECOM-managed combat
systems—approximately 60 systems. To support this
strategy, CECOM has budgeted $10 million per year for
VLAR development.
The most significant benefit of VLAR is more diffi-
cult to quantify. VLAR saves lives by reducing the risk
for soldiers and civilians in combat situations. Research
on combat operations in Iraq and Afghanistan shows
that one of every 50 vehicle-convoy operations resulted
in a combat injury or death. Helicopter missions
accounted for one combat fatality for every 8,000 fly-
ing hours. Based on generator and HARC fleet den-
sity and repair history, VLAR use has avoided over
180,000 ground and air movements and has avoided
4,500 casualties. The short-term and long-term costs
associated with that number of casualties are estimated
at over $2 million per soldier, resulting in direct and
indirect actual and projected U.S. cost savings of $9 bil-
lion (Bilmes 2013). Clearly, each minute of equipment
availability saved and each helicopter flight or vehicle
convoy avoided translates directly into reduced risk in
combat. The benefits are summarized in Table 1.
Extending VLAR’s Influence
The key challenge in developing VLAR was to demon-
strate the quantifiable and qualifiable improvements
to the diagnostic process available through advanced
Bayesian causal methods.
Table 1. The direct financial benefits of VLAR provide a favorable return on the project investment; however, these benefits
are dwarfed by the value of reducing casualties.
Direct financial benefits •Achieved cost savings of $27 million through 2015 from reduced expert personnel requirements from an
investment of $8 million.
•Savings of $222 million are projected through 2020 from an investment of $60 million.
•Investment has paid for itself.
Benefits of casualty avoidance •Of every 50 combat convoy operations, 1 results in an injury or death.
•Reducing requirements for helicopter and convoy operations reduces risks to soldiers and civilians.
•Based on generator and HARC fleet-size and repair rates, the use of VLAR in Afghanistan has avoided
over 180,000 helicopter and ground-vehicle movements to date, reduced the number of casualties
(killed and wounded) by 4,430, and provided direct and indirect actual and projected U.S. cost savings
of $9 billion.
The Army has little visibility on maintenance actions
performed on systems in the field. Therefore, it has
little information to support supply and maintenance
decisions (e.g., which repair parts it should stock and
which systems it should overhaul). Accurate informa-
tion at the equipment level can inform such decisions.
Because VLAR has been proven to help soldiers make
accurate diagnoses on equipment in the field, it can
also be useful for higher-echelon tasks. It represents
the first viable Army solution for providing fully con-
textualized on-platform maintenance data. By using
VLAR for maintenance and taking advantage of its
embedded reporting capability, logisticians can now
have a granular and fully contextualized set of data
on maintenance operations. VLAR data are now being
tested in a failure reporting, analysis, and corrective-
action system, which will allow supply personnel to
project spares requirements based on real equipment
data instead of supply-transaction history. VLAR is
also being tested in maintenance prioritization mod-
ules that determine the specific pieces of equipment
that require overhaul. It enables a tuning of this pro-
cess to ensure maximum fleet availability at mini-
mum cost. VLAR is currently using case files generated
from actual maintenance events to assist logisticians
in making sparing and provisioning decisions. The
team’s knowledge-engineering methodology has been
proven across multiple system domains (e.g., electri-
cal, electromechanical, electronic, fiber optic, and radio
frequency), and VLAR is now considered to be a
viable strategy for improving sustainment and shift-
ing the Army maintenance paradigm to an OR-based
approach.
Through a rigorous demonstration, evaluation, and
vetting process, VLAR has proven to be an instru-
ment of change at high levels of the Army Materiel
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
18 Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
Command because of its ability to provide a balance
between analytical complexity and usability. The
Army supports its combat forces by a number of sus-
tainment processes broadly categorized under sup-
ply and maintenance operations. All those operations
share a common requirement: to know, with a high
degree of certainty, what is needed and where it is
needed. The foundation of maintenance operations is
an accurate diagnosis of the problem. OR has changed
Figure A.1. The chain rule describes a probability distribution in terms of conditional probabilities. In this example, we use
probability notation to show the mathematics behind displayed probabilities for basic parent-child node pairs.
XY
Node X Node Y
Each node’s CPT P(X) P(Y|X)
Calculation to generate displayed P(X) P(X) P(Y) PP(X, Y) over
probabilities (since X has no parents) X PP(Y|X) ∗ P(X)
Where P(Y|X)and P(X)
are from the CPTs
Each node’s displayed probabilities P(X) P(Y)
Figure A.2. The marginal probabilities of Black Smoke are calculated by multiplying (1) the CPTs of Black Smoke conditioned
on the various states of Injector Pump Timing states by (2) the probabilities of those states, and totaling the result. The
conditional probabilities are computed from the equations in Bayes’ rule. This probability tree diagram illustrates the
computations of the marginal and conditional probability distributions.
Injector Pump Timing Black Smoke Injector Pump TimingBlack Smoke
P(IJP)
Normal
Yes
Sum
Yes
Yes
Advanced
No
No
90%
0.2%
99.8% 89.82% 1.08%
No
98.92%
0.18%
0.66%
11%
6%
Retarded
4%
89%
Yes
No
6%
94%
5.34%
0.24%
3.76%
P(BS ⎪IJP) P(BS, IJP) = P(IJP)*P(BS⎪IJP) P(IJP ⎪BS) = P(BP, IJP)
P(BS)
P(BS) = ∑P(BS, IJP)
Over IJP
Normal
16.7% = 0.18/1.08
Advanced
Retarded
61.1%
22.2%
Normal
90.8%
Advanced
Retarded
5.4%
3.8%
the process by which the Army does maintenance. OR
infuses the VLAR product with a rich set of capabili-
ties that advance the art and science of diagnosis, and
it does so at the edge of tactical operations—where the
lives of soldiers depend on their equipment.
Appendix
This appendix illustrates the application of Bayes’ rule for the
generic case (Figure A.1) and a specific example (Figure A.2).
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 19
Figure A.1 is based on the chain rule of probability (Pearl
1988):
P(X,Y)P(Y/X) ∗ P(X),
and the knowledge that when we sum X out of the joint
distribution P(X,Y), we arrive at P(Y).
References
Bilmes LJ (2013) The financial legacy of Iraq and Afghanistan: How
wartime spending decisions will constrain future national
security budgets faculty research. Accessed September 20, 2016,
http://watson.brown.edu/costsofwar/files/cow/imce/papers/
2013/The%20Financial%20Legacy%20of%20Iraq%20and%20
Afghanistan.pdf.
Hepler AB, Tatman JA, Smith GR, Buede DM, Mahoney SM, Tatman
SM, Marvin FF (2015) Bayesian network elicitation facilitator’s
guide. Report, Innovative Decisions, Vienna, VA.
Pearl J (1982) Reverend Bayes on inference engines: A distributed
hierarchical approach. Accessed September 19, 2016, https://
www.aaai.org/Papers/AAAI/1982/AAAI82-032.pdf.
Pearl J (1988) Probabilistic Reasoning in Intelligent Systems: Networks
of Plausible Inference (Morgan Kauffmann, San Francisco).
Pearl J (2015) An Introduction to Causal Inference (CreateSpace Inde-
pendent Publishing Platform, Seattle).
David Aebischer is Chief of the Special Operations
Branch of the U.S. Army CECOM Training Support Divi-
sion (TSD) and is the founder of the CEDAT VLAR project.
David is an honorably discharged veteran of the U.S. Air
Force with over 35 years of government service—including
over four years of service supporting soldiers in Iraq and
Afghanistan. David holds a Master of Business Administra-
tion and Technology Management from Monmouth Univer-
sity and is pursuing a doctoral degree in decision science
from Walden University. David has received extensive tech-
nical training in electrical, electromechanical, and electronics
systems and also continues to train on advanced Bayesian
Network techniques and applications.
John Vatterott, Jr. is the president and founder of Amer-
ican Trade School (ATS), which operates two distinct busi-
nesses: a brick-and-mortar school that specializes in training
electricians, heating and air conditioning technicians, and
IT technicians; and the Tactical Power division that pro-
vides military personnel training, diagnostic software appli-
cations, and course curriculum development for the U.S.
Department of Defense (DoD). Throughout his 25+years of
experience in education, John has worked extensively with
the DoD in educational development. John created an R&D
and Customized Training Division in 1995 where, for the
next eight years, he provided extensive training and curricu-
lum development to the U.S. Army’s Communications and
Electronics Command (CECOM) Power and Environmental
Logistics Assistance Representatives (P&E LARS). John is
a current member of the National Fire Protection Agency,
International Code Council, Air Conditioning Contractors
Association, and Refrigeration Service Engineering Society.
Michael Grimes is an entrepreneur with 27 years of expe-
rience building complex programmatic systems that are at
the cutting edge of technology. Michael is the founder of Vet-
eran Technology Group—a veteran-owned consulting and
design company he started with his brother in 2009, and
had been the project manager/technical creator for the U.S.
Army’s CEDAT VLAR program working with American
Trade School’s Tactical Power Team. Michael enlisted in the
United States Air Force at age 17 and served with distinc-
tion during Desert Shield and Desert Storm, earning the
National Defense Medal, Kuwait Liberation Medal, Merito-
rious Service Medal, and other honors until being honorably
discharged in 1995 to pursue his education at the Univer-
sity of Southern Illinois studying administration of justice.
Michael’s experience in Airframe maintenance and back-
ground in computer science, skills he acquired while serving
in the military and enhanced throughout his career, gives
him unique insight into diagnostics and analytics to design
programs that maximize system performance while reduc-
ing risk and driving down costs.
Andrew Vatterott is the Lead Knowledge Engineer at
American Trade School for the U.S. Army’s CECOM Equip-
ment Diagnostic Analysis Tool, Virtual Logistics Assistance
Representative (CEDAT VLAR) Project. Andrew attended
Dominican University for studies in business management
and statistics, where he served as president of the school’s
business club before accepting his position with American
Trade School’s Tactical Power team in 2011. He has since
continued with extensive education and training in Bayesian
statistics, expert elicitation, technical writing, fiber optics,
radio frequency, satellite communications, and other elec-
tromechanical systems, which uniquely qualify him for each
VLAR project. As the Lead Knowledge Engineer, Andrew
oversees the Tactical Power team’s development efforts
across nine VLAR projects.
Roderick Jordan is the lead instructor for the Tactical
Power Integration Team (TPIT) at American Trade School.
He joined the U.S. Army in 1987 and trained as a Turbine
Engine Driven Generator Repair Specialist. Roderick served
in combat where his competence of electrical mechanics
allowed his Patriot Missile unit to record a 98% high level of
readiness without power interruptions. With his unique cre-
dentials, Roderick assists with the VLAR program by com-
municating his extensive electromechanical knowledge and
troubleshooting experience for equipment selected for the
VLAR program.
Carlo Reinoso is a Knowledge Engineer at American
Trade School for the U.S. Army’s CEDAT VLAR program.
Carlo attended both Saint Louis University and the Uni-
versity of Missouri-St. Louis where he focused his studies
on computer science and communications. In 2012, Carlo
accepted a position with ATS and the tactical power team
tasked with expert elicitation and knowledge capture, and
integrating them into a graphical user interface that can
diagnose and troubleshoot tactical systems. As a Knowledge
Engineer, Carlo has become an essential asset in shaping,
expediting, and streamlining the development process of
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
20 Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS
the VLAR program, and he takes pride in enhancing and
improving the interworking of GUI to allow the user to con-
veniently and effortlessly interact with the VLAR program.
Bradford Alex Baker is a Senior Software Engineer for
the U.S. Army’s CEDAT VLAR project. He joined Veterans
Technology Group/American Trade School in 2014 and is
based in Chicago. Alex has worked on the VLAR interface
bringing many enhancements to its structure and improv-
ing and refining the underlying code to make a cleaner,
more functional interface. Before joining Veterans Technol-
ogy Group, he worked on a variety of applications for one of
the largest proprietary trading firms in Chicago, including
a click-trading application for commodity markets, an auto-
mated trading platform that responded to commodity and
volatility index market signals in under 10 microseconds,
and a proprietary market data research grid. Alex received
his BS from Purdue University.
William D. Aldrich is a veteran of the U.S. Army, serving
as a Cavalry Scout from 1991–1999. William earned an Asso-
ciates of Arts from Arapahoe Community College in 1994.
William was recruited in 2012 by American Trade School as
a Knowledge Engineer, initially as a technical lead for the
HARC (High Antenna for Radio Communication), due to his
experience with fiber optic and radio frequency broadcast-
ing equipment. He soon became certified as a Bayesian Belief
Knowledge Engineer by Innovative Decision Inc. and started
developing Bayesian Belief Networks for the CEDAT VLAR.
Luis Reinoso is a Bayesian Knowledge Engineer for the
CEDAT VLAR project. Luis attended Saint Louis Univer-
sity and the University of Missouri–St. Louis for studies in
nursing and English. Luis was recruited by American Trade
School (ATS) in 2013 as a Knowledge Engineer to join the
newly formed Tactical Power team and is tasked with proce-
dural technical writing, internal review and editing, image
creation, and expert elicitation. Luis has been involved in the
development of seven U.S. Army projects that include train-
ing and troubleshooting systems for power generation and
communication.
Rodolfo Villalba is a Knowledge Engineer at American
Trade School with the Tactical Power team. He graduated
from Dominican University with a BA in psychology and a
minor in criminology. Since joining American Trade School,
Rodolfo has completed multiple trainings in professional
elicitation techniques and Bayesian Belief Networks. He is
responsible for building and integrating probability models
based on expert elicitation, technical procedure writing, and
accompanying procedural graphics across multiple tactical
power and communication systems into a GUI.
Michael Johnson is a Bayesian Knowledge Engineer for
the CEDAT VLAR Project. Michael attended Saint Louis Uni-
versity and studied aviation science. He joined ATS Tactical
Power in 2014 as a Knowledge Engineer tasked with captur-
ing expert knowledge through elicitation. As a member of
the Tactical Power team, Michael gained an extensive educa-
tion, valuable experience and training in Bayesian statistics,
satellite communications, and electromechanical systems.
Christopher Myers is the Lead Generator Technician and
Trainer for the United States Army’s CEDAT VLAR and the
Tactical Power Integration Team (TPIT) at American Trade
School. Christopher joined the United States Marine Corps in
1990 and trained as an Electrical Equipment Repair Special-
ist with emphasis on generator troubleshooting and repair.
As Lead Generator Technician at American Trade School,
Christopher developed troubleshooting techniques for the
CEDAT VLAR on the TQG “Bravo” 30/60 kW generator sets.
Christopher has created a clearer schematic for the 15 kW
TQG generator set, which simplifies the troubleshooting pro-
cedures for the operator and maintainer, and is also working
with Knowledge Engineers on the 7.5 kW generator project.
Stefan Conrady studied electrical engineering at the Uni-
versity of Ulm, Germany, and has more than 15 years of
experience in decision analysis, analytics, market research,
marketing, and product strategy with Fortune 100 com-
panies in North America, Europe, and Asia. In his role
as Managing Partner of Bayesia USA and Bayesia Sin-
gapore, he specializes in applying Bayesian networks for
research, analytics, and reasoning. In this context, Stefan has
recently co-authored a new book, it Bayesian Networks and
BayesiaLab—A Practical Introduction for Researchers.
Joseph A. Tatman has more than 30 years’ experience in
the application, research, and teaching of decision and risk
analysis. He specializes in Bayesian networks in a broad
array of applications including the analysis of complex
political-economic-military problems, diagnostic and effec-
tiveness analysis of complex systems, analysis of risk in
system development and operations, and the application of
large data sets in the health and medical sciences. Dr. Tatman
has taught Bayesian networks to a broad spectrum of stu-
dents, companies, and organizations. He graduated from the
University of Notre Dame with a BS in electrical engineering,
from the U.S. Air Force Institute of Technology with an MS
in electrical engineering, and from Stanford University with
a PhD in engineering-economic systems (now management
science and engineering).
Suzanne M. Mahoney has more than 40 years of experi-
ence in the information technology and modeling and sim-
ulation communities. She earned BS and MA degrees in
mathematics at the University of Michigan, and a PhD in
information technology from George Mason University. Her
research focuses on the knowledge representation required
to automatically construct large Bayesian networks from
reusable Bayesian network fragments. Working first for Infor-
mation Extraction and Transport, Inc. and then Innovative
Decisions, Inc., Dr. Mahoney has applied her expertise to
leading edge programs for the military and intelligence
communities.
Darrin L. Whaley has 27 years of experience as an oper-
ations research analyst and Marine Corps artillery officer.
As an analyst, Mr. Whaley has applied a wide array of
analytical techniques to a diverse set of challenges. He
has applied optimization techniques to military recruiting,
Aebischer et al.: Bayesian Networks for Combat Equipment Diagnostics
Interfaces, 2017, vol. 47, no. 1, pp. 1–21, ©2017 INFORMS 21
training, manpower, equipment fielding, shipboard voyage
planning, and transportation networks. He has applied mul-
tiobjective decision analysis to theater security cooperation,
Analysis of Alternatives for equipment procurement, and
logistics process comparison. He graduated from the United
States Naval Academy with a Bachelor of Science in Ocean
Engineering and from Naval Postgraduate School with a
Master of Science in Operations Research.
Amanda B. Hepler is an experienced consultant pro-
viding expertise in statistics, Bayesian networks, eviden-
tial analysis, and computer programming. Dr. Hepler has
a strong research background in probabilistic approaches
to evidential reasoning. During her tenure as an Intelli-
gence Community Postdoctoral Fellow, her research team
conducted cutting edge research in the field of forensic
statistics. She previously served as a Research Fellow at
University College London exploring the use of Bayesian
networks in forensics. In 2001, Dr. Hepler graduated summa
cum laude from Towson University, earning a bachelor’s
degree in mathematics, and obtained a PhD in statistics from
North Carolina State University in 2005. Since joining Inno-
vative Decision, Inc., Dr. Hepler has gained practical experi-
ence applying statistical and operations research methods to
improve decision making for various government clients.