Proceedings of the 1998 Winter Simulation Conference

D.J. Medeiros, E.F. Watson, J.S. Carson and M.S. Manivannan, eds.

BELIEF NETWORKS IN CONSTRUCTION SIMULATION

Brenda McCabe
Department of Civil Engineering
University of Toronto
Toronto, Ontario, M5S 1A4, CANADA

ABSTRACT

A method for automatically improving the performance of

construction operations was developed by the integration
of computer simulation and belief networks. The
simulation model is used to represent the operation and to
determine the effect that changes in resource configuration
have on the model performance. The belief network
provides diagnostic analysis of the performance and
recommendations for changes to the model. Modifications
to the simulation model include selection of alternative
resources and resource quantities.
The method is based upon the assumption that ultimate
objectives, such as lower costs or shorter project duration,
will result from efficient use of resources. Therefore, the
improvement process is focused upon resource
performance instead of cost or duration. The approach is
iterative, and will provide the modeler with results even if
user-defined constraints related to performance limits are
not met.

Several simulation languages are available. CYCLONE

(Halpin 1976) was developed specifically for modeling
construction operations. Many CYCLONE-based systems
have been developed to modify or extend the functionality
of CYCLONE. These include INSIGHT (Paulson 1978),
RESQUE (Chang 1987), UM-CYCLONE (Ioannou 1989),
COOPS (Liu and Ioannou 1994), DISCO (Huang,
Grigoriadis & Halpin 1994), CIPROS (Tommelein & Odeh
1994), STROBOSCOPE (Martinez & Ioannou 1994),
HSM (Sawhney & AbouRizk 1995), and ACPSS (Liu
1996).
More general simulation languages are available, such
as Visual SLAM (Pritsker, O'Reilly & LaVal 1997),
GPSS/H (Crain & Smith 1994), SIMAN/Cinema
(Profozich & Sturrock 1994), and SIMSCRIPT (Russell
1993). These systems are capable of supporting simulation
modeling in any domain including manufacturing,
industrial engineering and construction. The price paid for
increased flexibility, however, is the increased skill level
required by the simulation modeler.
General optimization routines and simulation
environments are often incompatible because the modeling
techniques of simulation and mathematics are so different.
Consequently, many optimization methods have been
developed for simulation models (Azadivar 1992). The
methods that have been developed, and in some cases
automated, have been categorized by Azadivar as 1)
gradient based search methods, 2) stochastic approximation
methods, 3) response surface methods, and, 4) heuristic
search methods.
Gradient-based search methods and stochastic
approximation methods are focused on continuous
movement toward the optimum. These techniques assume
unimodal solution functions and contain algorithms to
identify the direction of the steepest slope. Riggs (1976)
developed an automated sensitivity analysis module for
CYCLONE that required the user to provide the upper and
lower limits of the resource quantities available for the
operation being modeled. Using this method, the user was

INTRODUCTION

Simulation is used to model operations such as

construction because the operation is too complex to model
entirely mathematically, or because there is some
uncertainty in the system. Although simulation has been
documented as an excellent tool for modeling construction
operations (AbouRizk and Halpin 1990, AbouRizk and
Dozzi 1993, Smith and Osborne 1995), it has not
experienced widespread use in industry (Shi & AbouRizk
1994). One obstacle to the acceptance of simulation by the
construction industry is the effort required for
experimentation with the model in order to optimize it
(McCabe 1997).
A method for automatically improving the
performance of a simulated operation was developed
through the integration of computer simulation and
artificial intelligence, specifically belief networks. The
method is not domain specific but requires the interaction
of servers and customers in queue-type structures.

1279

BACKGROUND

McCabe
able to establish the direction toward which the optimal
resource configuration may be found.
The response surface methodology involves fitting
regression models to the results of the simulation run
evaluated at various states of the problem domain.
Azadivar and Talavage (1980) showed that the
effectiveness of this method was greatly reduced if the
regression function contained sharp ridges or flat surfaces.
Heuristic methods may not guarantee that the solution
found is the global optimum because there is often no
assumption that the solution function is unimodal. One
may be confident that the solution found by the method is
very good, but it may not be the optimum. Two formal
heuristic methods have been defined by Azadivar (1992):
complex search and simulated annealing. Methods that rely
upon artificial intelligence, such as genetic algorithms,
rule-based systems, and belief networks also fall into this
category.
Complex search involves using the results of several
simulation runs from different variable parameters to
determine the worst point. The worst point is dropped, a
new point is generated, and the simulation is rerun.
Simulated annealing is a local gradient search method that
evaluates the objective function, say, to minimize the cost,
at an appropriately chosen point. If the new cost is less
than the cost at the previous point, then the new point is
accepted and the old one is dropped. To reduce the
likelihood of being caught in a local minimum, the method
will allow uphill moves based on random variables with
controlled probabilities.
Several heuristic techniques have been developed
specifically for improving construction operations. Wood
and Harris (1980) developed a program that utilized an
iterative technique of simulation and manual cost
evaluation to optimize concrete delivery truck fleets. Their
model was able to analyze various truck and plant
capacities.
AbouRizk and Shi (1994) applied heuristics to a
DELAY statistic to determine whether the number of
resources in a simulation model should be increased or
decreased in order to meet project objectives for
optimizing cost, production, or resource utilization. The
DELAY statistic is equal to the fraction of time a resource
is idle relative to its total working time. The limitation of
the work, as cited by the authors, is that the system
assumed the simulation model itself cannot be modified,
and it could not meet multiple objectives, such as optimal
cost and production.
Shi and AbouRizk (1995) developed a hybrid
simulation and mathematical optimization system for
handling large, complex systems. In this model, the large
system is broken into smaller sections for separate
evaluation of each feasible resource state. The smaller
sections are rejoined by mathematical functions and the
entire project is optimized mathematically. The method

1280

requires significant manipulation by the user to determine

the connection types between the smaller simulation model
sections, development of the mathematical functions that
connect the smaller sections into the entire project, and
fine-tuning.
Tompkins and Azadivar (1995) combined genetic
algorithms with object-oriented programming in ModSim
II to develop a means of optimizing simulation models for
manufacturing systems. The system was intended to
represent corporate policy for minimizing resource
requirements of new operations. Several billion points
could be searched resulting in significantly improved
solutions over random search methods.
Chan and Chua (1996) developed a hybrid
optimization system using genetic algorithms and
computer simulation for use in civil engineering
applications. Because of the constraints imposed by
practical issues of the specific applications, they found that
the genetic algorithms were not allowed to fully optimize
the solutions.
Most of the techniques developed to "optimize"
simulation models are based on modifying resource
quantities, but not resource capacities through the selection
of alternative resources. An alternative resource is a
resource that is able to perform the same function, but has
different parameters that affect its performance. For
example, an alternative for one truck resource may be
another truck of greater capacity but, perhaps, slower
acceleration. Other techniques have also focused on a
single optimization objective, such as cost. The approach
developed in this research focuses on the surrogate
objective of improving performance of all resources based
on five performance indices. The drive to improve
performance instead of cost or project duration leads to the
recommendation of alternative resources.
The issue of finding very good vs. optimal solutions is
not perceived as a problem by the author, especially as it
relates to construction operations. Moreover, if one does
not focus on finding the one optimal point, it follows that
the modeler could be presented with several near-optimal
solutions. Several, equally acceptable solutions may result
when the solution function is rather flat near various
optimal points or when there are several local optimums
that result in similar system performance. Because
construction is vulnerable to innumerable external
influences that may continually affect its performance,
several very good and equally acceptable solutions may be
of more value to the construction planner than a single
optimal solution.
3

BELIEF NETWORKS

This performance improvement method is based upon the

integration of simulation and belief networks. Belief
networks may be described as a form of artificial

Belief Networks in Construction Simulation

intelligence able to incorporate uncertainty and knowledge
into their structures (Pearl 1990). Belief networks are
directed, acyclic graphs (DAG) with nodes representing the
variables in the problem domain, and arcs representing
conditional dependence between the variables (Jensen
1996). Directed means that the arcs have an implicit
direction represented by an arrow. Acyclic refers to the
constraint that when the direction of the arrows are
considered, they may not close upon themselves creating a
closed circle. The node or variable with an arrow pointing
away is the parent node. The node to which the arrow is
pointing is considered the child node.
In diagnostic models including the one developed in
this research, variables may be categorized as causal
variables and effect variables. (Henrion, Breese & Horvitz
1991) Arcs connect the causal variables to the effect
variables to depict the conditional relationships i.e. the
state of the cause variable (true or false) affects the state of
the effect variable. Probabilities are assigned to each
variable based on all possible combination states of the
parent nodes.
Poole, Mackworth and Goebel (1998) provide
guidance for the development of belief networks in four
general steps. First, the variables in the domain must be
determined. This is directly affected by the scope of the
problem definition. Second, the conditional relationships
are defined by connecting the nodes with arc. The resulting
graph must be acyclic. Third, the states of the nodes are
determined. Where possible, the variables are binary to
limit the number of probabilities that must be assigned.
Finally, the probability related to each conditional
relationship is determined. This entails evaluating all
combinations of the states of parent nodes and assigning a
probability value of the child node states for each parent
state combination.
Figure 1 shows the belief network that was developed
for diagnosing performance. Note that the nomenclature
from queuing theory has been used. The variables SQ, SU,
QW, QL, and CD are the performance indices Server
Quantity, Server Utilization, Queue Wait time, Queue
Length, and Customer Delay. The performance indices are
the effect variables in this network and represent the effect
or symptoms of poor performance. The variables
TooFewCustomers, TooManyServers, etc. are possible
causes of poor performance and are the causal variables.
Finally, the variables Cost and Duration have been
included to provide direction for the search, but are not
explicitly incorporated as objectives. Their role will be
discussed shortly.
The arcs linking the variables indicate a conditional
relationship between the variables. Causal variables that
relate to the customer quantity or capacity affect the
customer performance and the queuing performance. The
causal variables that relate to the server quantity or
capacity affect the server performance and the queuing

1281

performance. For example, TooManyCustomers may affect

the performance of the customers as measured by the
indices CD, and the performance of the queues represented
by the variables QW, and QL.
The states of the variables are shown in Table 1 and
Table 2. All Causal nodes are binary. These values
indicate whether the causal variable is affecting the
performance of the resources.

Figure 1: Belief Network for Performance Improvement

The Effect nodes have two or three states, depending
upon the limits. In some cases, such as SU, QL and QW,
the value of the variable is bounded by two limits, an upper
and lower limit. The subscript L indicates the lower limit
of the acceptable range of the value of the variable. The
subscript U indicates the upper limit of the acceptable
range. The variables SQ and CD are bound by only one
limit. The value of SQ is evaluated as equal to zero or
greater than zero. The variable CD has only an upper limit,
as the lower limit is always zero.
Limits are defined by the modeler to provide guidance
to the diagnostic processes. This allows the modeler to test
various resource management strategies and to impose
project or corporate constraints related to the acceptable
performance limits for the particular operation being
simulated.
The Cost and Duration nodes were added to allow the
improvement process to take different approaches to
diagnosing the performance. For example, suppose an
activity has unacceptably long queues at the servers. If the
shortest project duration is the overall objective, then more,
larger servers might be an appropriate strategy. However,
if cost is the objective, then fewer, smaller customers may
better achieve the goal. The iterative process of
improvement considers both perspectives.

McCabe
TooSmall to False for the appropriate resource. During
evaluation of the belief network, the probability of each
Causal variable state is calculated using the concepts of
Bayes' Theorem in algorithms designed to solve the
networks (McCabe, AbouRizk and Goebel 1998).

Table 1: States of Causal Nodes

Causal Node

States

Too Many Servers (TMS)

True

False

Too Few Servers (TFS)

True

False

Too Many Customers (TMC)

True

False

Too Few Customers (TFC)

True

False

Server Too Big (STB)

True

False

Server Too Small (STS)

True

False

Customer Too Big (CTB)

True

False

Customer Too Small (CTS)

True

False

Table 3: Probability Assignment for Variable CD

CTB
F
F
F
F
F
F
F
F
T
T
T
T
T
T
T
T

Table 2 : States of Effect Nodes

Effect
Node

State

Index

QL

QLLQLQLU

QL<QLL

QL>QLU

QW

QWLQWQWU

QW<QWL

QW>QWU

CD

CDCDU

CD>CDU

SQ

SQ=0

SQ>0

SU

SULSUSUU

SU<SUL

Cost

OK

Optimize

Duratin

OK

Optimize

CTS
F
F
F
F
T
T
T
T
F
F
F
F
T
T
T
T

TFC
F
F
T
T
F
F
T
T
F
F
T
T
F
F
T
T

TMC
F
T
F
T
F
T
F
T
F
T
F
T
F
T
F
T

CDCDU
0.99
0.10
0.90
0.99
0.70
0.60
0.95
0.70
0.20
0.00
0.60
0.20
0.99
0.10
0.90
0.99

SU>SUU

The probability related to each relationship was

determined by the author. The probabilities of the states of
CD were evaluated as shown in Table 3. Note that the
probability that CD>CDU is not explicitly shown because
the information is redundant. It may be calculated as
P (CD > CD U ) = 1 P(CD CD U ) .
The parents of CD consist of two sets of conflicting
variables:
CustomerTooBig/CustomerTooSmall,
and
TooFewCustomers/TooManyCustomers. These states
cannot be true at the same time. Therefore, where the
combination of the states of the parents indicate that they
are both true, the probability assigned to that combination
is the same as if they were both false. The strategy results
in no clear decision based upon conflicting states of the
parents.
At the end of the simulation run, statistics are
extracted and the performance indices are calculated for
each queuing location in the simulation model. The value
of each index is compared to its user-defined limits and
then it is entered as evidence to the belief network by
setting the appropriate state of the Effect variables to True.
In addition, resource constraints, such as having only one
of a specific resource or not having any alternative
resources, are entered in the same manner. The lack of
alternatives is modeled by setting the variables TooBig and

1282

If the probability of a Causal variable = True is greater

than 50%, the Causal variable gains a score of one. The
evaluation scores are summed over all of the interaction
locations in the simulation model. As before, there is
concern that conflicting variables will compete for priority.
This may occur where one customer is being served by
several servers.
For example, suppose that one customer has three
servers. The performance of the interaction location of the
first server is acceptable in all respects i.e. all performance
indices at the location are within their specified limits.
Consequently, there is no Causal variable with a
probability of True greater than 50%. At the second
location, the belief network evaluation indicates the
likelihood that CustomerTooBig and ServerTooSmall are
greater than 50%. The third server location results in the
variables CustomerTooSmall and ServerTooBig having a
probability greater than 50%. The customer receives a
score of one for each TooBig and TooSmall. Server2 has
one score in TooSmall, and Server3 has one score in
TooBig. By observation, one may conclude that the
customer should not be changed unless there is a similar
problem at each of the server locations at which the
customer interacts. The best action at this time would be to
adjust the parameters of Server2 and Server3, and to rerun
the simulation to determine the effects of the change.
It might be expected that the belief network should be
capable of avoiding conflict recommendations by
evaluating the simulation performance as a whole instead

Belief Networks in Construction Simulation

of evaluating each server-customer interaction location
individually. However, the separate analysis at each
location allows the method to be very general without
limiting the number of interaction locations or resources
that can be included in the model.
A simple heuristic was developed to handle the
problem of conflicting recommendations. If any resource
accumulates a score greater than zero for both variables of
a conflicting set of variables, then the scores for those two
conflicting states are cancelled. In other words, no change
will be made to a resource if there is a conflicting
evaluation related to that variable. In the above example,
the result would be a recommendation to modify the
servers.
After the simulation model is modified, the simulation
is rerun to determine the effect that the changes have on the
performance. The iterative process continues until all
performance indices at each queuing location are within
their specified limits, or until the process begins to
oscillate. The resource configuration and the resulting
performance may be stored in a database. At the
conclusion of the automated process, the modeler may
review the database. Based on the argument that several
very good solutions is of greater value to the modeler than
one optimal solution, the modeler may more closely
examine the resource assignments from the, say, five
lowest cost simulation runs.
It is possible that the performance from the iteration
that resulted in the lowest cost meet the user-specified
limits. In fact, the user-specified limits may be such that
there is no resource assignment combination that would
meet those limits. Again, this is not of major concern. The
limits are used as a guide for the improvement process and
should not be considered absolute. Although the
performance index limits are used as guidelines for
improvement, they will not restrict the model from
working toward optimal solutions.
4

PROTOTYPE

A prototype was developed to demonstrate the model.

MSBN Microsoft Belief Network Version 1.001 was the
belief network modeling and evaluation environment used
for the prototype. The simulation language used was
AweSim Version 2.0 by Pritsker Corporation (Pritsker,
O'Reilly & LaVal 1997). Microsoft Access relational
database was used to store the input resource assignment
and corresponding output of each simulation run. All of the
software was easily integrated using Microsoft Visual
Basic, which also provided the development environment
for the user interface.
A unique feature that was added to the prototype is
that more than one simulation model may be improved and
compared to determine which results in the most effective
method. Each model that is entered is considered a

1283

scenario. One scenario may differ from another by changes

in user-defined performance limits for the same simulation
model, changes to the availability of resources, or by
development of another simulation model to represent a
different construction method. For example, an
earthmoving operation may be completed using scrapers
and bulldozers or by using trucks and loaders. Each
method would be represented by a simulation model and
constitute a Scenario.
The automated method was validated by testing the
prototype with queuing problems found in literature
(Charmichael 1987). A more complicated model was
developed to demonstrate the method. The model is an
earthmoving operation with loaders, bulldozers, a weigh
scale and unloading spaces as servers, and trucks as
customers. Alternative resources for the loaders, trucks and
bulldozers were identified. The unloading spaces could be
modified only in quantity and the weigh scale resource had
no alternatives and was limited in quantity to one.
The operation works as follows. A truck is loaded by
an available loader and travels to the scale where it is
weighed. The truck proceeds to the unloading area where it
dumps its load, and returns to the loading area. The
material that was left behind holds the unloading space
until a bulldozer is available to move and spread the
material. A schematic of the model is shown in Figure 2.

Loader

Truck
Cycle

Unload
Area
Dozer
Cycle

Weigh
Scale

Figure 2 : Schematic of Earthmoving Operation

Several constraints were put upon the simulation
model itself to ensure the necessary flexibility was
available for the prototype to run. For example, all
information related to the quantity and capacity of the
resources to be used in a simulation run is contained in a
text file. The text file was modified between simulation
runs and read into the model at the start of the run. The
parameters of the resources, including alternative
resources, were a standardized database table. These
constraints were not considered to be limiting to the
simulation modeler, and were specific to the software used
to develop the prototype.
At the start of the program, the modeler provides the
location of the simulation file(s) on the computer. The
program reads the simulation file and extracts information
related to resources, the queue locations, and user-defined
statistics. The identification of the queuing locations in

McCabe
which each resource interacts, and the acceptable limits of
the performance indices are entered or verified by the user,
guided by the information extracted from the simulation
file. If more than one scenario is to be compared, the
modeler is prompted to add the extra information before
the automated improvement process is started.
When the iterative improvement process is complete,
the program scans the database containing the input and
output parameters from each simulation run. The lowest
cost and shortest duration are presented to the modeler
including the resource assignment for that iteration. In
addition, iterations that did not result in lowest cost or
shortest duration but did meet the performance limits are
listed. As mentioned, the modeler may review the database
itself to determine if any other solution is acceptable.
Figure 3 shows the resource assignments for the
demonstration case. Each resource alternative has a
different symbol.

Figure 3 : Resource Allocation per Iteration of

Demonstration Case
Figure 4 shows the corresponding cost and duration
that resulted from the resource assignment. It is evident
that the solution function is somewhat flat around the
optimum points, and that the construction planner may find
more than one of solutions acceptable.
5

CONCLUSIONS

The objective of the research was to develop an automated

method of improving the performance of simulated
operations. Evaluation and diagnosis of the performance
was provided through a belief network. Performance
indices were developed to evaluate and measure the
resource performance within the simulation run. These
indices were input to the belief network as evidence of the

1284

current situation. Output of the network includes diagnosis

of operation performance, and recommendations for
modifications to the model to improve its performance.
Scenario #1 Results
Cost $

Duration

70000

4500
4000

60000

3500
50000
3000
40000

2500

30000

2000
1500

20000
1000
10000

500

0
0

10

20

30

40

50

60

70

80

Iteration Number
Cost

Duration

Figure 4 : Resulting Cost and Duration per Iteration of

Demonstration Case
The model presented here has several limitations.
First, the simulation model itself cannot be automatically
changed. This would require an understanding of the
processes that are being modeled. A method may be
developed if the system is limited to a specific process
domain. However, this was not within the scope of the
research.
The prototype was not optimized for computing time,
and it took several minutes to run the demonstration case.
While this is significantly less effort than what is required
in manual experimentation, increased expectations in
automated systems lessens the effectiveness of the
prototype. It is the opinion of the author that this method
may be effectively applied to process-specific simulation
environments where many of the user-defined limits can be
standardized to reduce input required by the modeler. The
belief network can be integrated with the simulation
environment and computation time related to
communication between software may be saved.
Finally, the method does not allow combinations of
alternatives of a single resource type to be assigned within
the same model, such as three units of alternative #1 and
one unit of alternative #4 to work together.
More research is required to fully automate model
optimization of complex processes. The method discussed
here has demonstrated a feasible method for automated
improvement without limiting the modeler to a single
solution.
6

ACKNOWLEDGEMENTS

The author would like to thank the Decision Theory Group,

Microsoft research for the use of their software MSBN.

Belief Networks in Construction Simulation

REFERENCES
AbouRizk, S.M., Halpin, D.W., (1990) Probabilistic
Simulation Studies for Repetitive Construction
Processes, Journal of Construction Engineering and
Management, Vol. 116, No. 4, 575-594
AbouRizk, S.M., Shi, J., (1994) Automated ConstructionSimulation Optimization, Journal of Construction
Engineering and Management Vol. 120, No. 2, 374-385
AbouRizk, S.M., Dozzi, S.P., (1993) "Application of
Computer Simulation in Resolving Construction
Disputes", Journal of Construction Engineering and
Management, Vol. 119, No. 2, 355-373
Azadivar, F., Talavage, J., (1980) Optimization of
Stochastic Simulation Models, Mathematics and
Computers in Simulation, Vol. 22, 231-241
Azadivar, F., (1992) A Tutorial on Simulation
Optimization, Proceedings of the 1992 Winter
Simulation Conference, Institute of Electrical and
Electronics Engineers, Piscataway, N.J., 198-204
Carmichael, D.G., (1987) Engineering Queues in
Construction and Mining, Ellis Horwood Limited,
Market Cross House, Cooper Street, Chichester, West
Sussex, PO19 1EB, England
Chan, W., Chua, D.K.H., (1996) Civil Engineering
Applications of Genetic Algorithms, Proceedings to
the Third Congress of Computing in Civil Engineering,
June 17-19, 1072-1078
Chang, D., (1987) RESQUE, Ph.D. thesis, University of
Michigan, Ann Arbor, MI
Crain, R.C., Smith, D.S. (1994) Industrial strength
simulation using GPSS/H, 1994 Winter Simulation
Conference Proceedings, Institute of Electrical and
Electronics Engineers, Piscataway, N.J., 502-508
Halpin, D.W., (1976) CYCLONE - A Method For
Modeling Job Site Processes, Journal of the
Construction Division, Vol. 103, No 3, pp. 489-499
Henrion, M., Breese, J.S., Horvitz E.J., (1991) "Decision
Analysis and Expert Systems", AI Magazine, Winter
1991, 64-91
Huang, R., Grigoriadis, A.M., Halpin, D.W., (1994)
Simulation of Cable-Stayed Bridges Using DISCO,
Proceedings of the 1994 Winter Simulation
Conference, Institute of Electrical and Electronics
Engineers, Piscataway, N.J., 1130-1136
Ioannou, P.G., (1989) UM-CYCLONE users manual,
Division
of
Construction
Engineering
and
Management, Purdue University, West Lafayette, IN
Jensen, F.V., (1996) An introduction to Bayesian Networks,
UCL Press, University College London, London, UK,
WC1E 6BT, 12-20
Liu, L.Y., (1996) ACPSS - Animated Construction Process
Simulation System, Proceedings of the Third
Congress on Computing in Civil Engineering, June 1719, 397-403

1285

Liu, L.Y., Ioannou, P.G., (1994) Graphical Object-Oriented

Discrete-Event Simulation System, Proceedings of the
1994 Winter Simulation Conference, Institute of
Electrical and Electronics Engineers, Piscataway, N.J.,
1285-1291
Martinez , J., Ioannou, P.G., (1994) General purpose
simulation with STROBOSCOPE, Proceedings of the
1994 Winter Simulation Conference, Institute of
Electrical and Electronics Engineers, Piscataway, N.J.,
1159-1166
McCabe, B., (1997), "An automated modelling approach for
improving construction performance using simulation
and belief networks", PhD thesis, University of Alberta,
Edmonton, Alta., Canada
McCabe, B., AbouRizk, S.M., Goebel, R., "Belief Networks
for Construction Performance Diagnostics", Journal of
Computing in Civil Engineering, Vol. 12, No. 2, 93100
Paulson, G.C. Jr., (1978) Interactive Graphics For
Simulating Construction Operations, Journal of the
Construction Division, Vol. 104, No. 1, 69-76
Pearl, J., (1990), "Bayesian Decision Methods", Readings in
Uncertain Reasoning, Ed. G. Shafer & J. Pearl, Morgan
Kaufmann Publishers, Inc., San Mateo, California
Poole, David L., Mackworth, Alan, Goebel, Randolph G.,
(1998) Computational Intelligence: A Logical
Introduction, Oxford University Press, New York, N.Y.
Pritsker, A.A.B., OReilly, J.J., LaVal, D.K., (1997),
Simulation with Visual SLAM and Awesim, John Wiley
& Sons, Inc., New York, NY and System Publishing
Corporation PO Box 2161, West Lafayette, IN 47906
Profozich, D.M., Sturrock, D.T., (1994) Introduction to
SIMAN/Cinema Proceedings of the 1994 Winter
Simulation Conference, Institute of Electrical and
Electronics Engineers, Piscataway, N.J., 427-430
Riggs, L, (1976) Sensitivity Analysis of Construction
Operations, Ph.D. thesis, Georgia Institute of
Technology, Atlanta, GA
Russell,
E.C.,
(1993)
SIMSCRIPT
II.5
and
SIMGRAPHICS tutorial, Proceedings of the 1993
Winter Simulation Conference, Institute of Electrical
and Electronics Engineers, Piscataway, N.J., 223-227
Sawhney, A., AbouRizk, S.M., (1995) HSM-SimulationBased Planning Method for Construction Projects,
Journal of Construction Engineering and Management,
Vol. 121 No. 3, 297-303
Shi, J., AbouRizk, S.M., (1994) A Resource-Based
Simulation
Approach
with
Application
in
Earthmoving/Strip Mining, Proceedings of the 1994
Winter Simulation Conference, Institute of Electrical
and Electronics Engineers, Piscataway, N.J., 1124-1129
Shi, J., AbouRizk, S.M., (1995) An Optimization Method
for Simulating Large Complex Systems, Eng. Opt.,
Vol. 25, 213-229

McCabe
Smith, S.D., Osborne, J.R., Forde, M. C.,(1995) Analysis
of Earth-Moving Systems Using Discrete-Event
Simulation, Journal of Construction Engineering and
Management, Vol. 121, No 4, 388-396
Tommelein, I.D., Odeh, A.M., (1994) Knowledge-Based
Assembly of Simulation Networks Using Construction
Deigns, Plans, and Methods, Proceedings of the 1994
Winter Simulation Conference, Institute of Electrical
and Electronics Engineers, Piscataway, N.J., 1145-1158
Tompkins, G., Azadivar, F., (1995) Genetic Algorithms in
Optimizing Simulated Systems, Proceedings of the
1995 Winter Simulation Conference, Institute of
Electrical and Electronics Engineers, Piscataway, N.J.,
757-762
Wood, D.G., Harris, F.C., (1980) Truck Allocation Model
for Concrete Distribution, Journal of the Construction
Division, Vol. 106, No. CO2, 131-139
AUTHOR BIOGRAPHY
BRENDA MCCABE is an Assistant Professor in the
Department of Civil Engineering at University of Toronto,
Canada. Her research interests include probabilistic
reasoning and simulation as they relate to the field of
Construction Engineering and Management. She received a
BASc from University of Toronto in 1994 and PhD from
University of Alberta in 1997. She has eleven years of
industrial experience related to land surveying, oilfield
construction, utility engineering and housing development.

1286