JOURNAL OF

COMPUTING IN CIVIL ENGINEERING

Volume 14 Number 1 January 2000

TECHNICAL PAPERS
River Stage Forecasting in Bangladesh: Neural Network Approach
Shie-Yui Liong, Wee-Han Lim, and Guna N. Paudyal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Artificial Neural Network for Measuring Organizational Effectiveness
Sunil K. Sinha and Robert A. McKim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Augmented IFN Learning Model
Shih-Lin Hung and J. C. Jan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Comparison of Neural and Conventional Approaches to Mode Choice Analysis
Tarek Sayed and Abdolmehdi Razavi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Genetic Programming-Based Approach for Structural Optimization
Chee Kiong Soh and Yaowen Yang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Finite State Verification of Intelligent Transportation Systems
John W. Baugh Jr. and GopalaKrishna Reddy Kakivaya . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
IntelliGIS: Tool for Representing and Reasoning Spatial Knowledge
Xudong Jia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Automatic Road Detection in Grayscale Aerial Images
Katherine Treash and Kevin Amaratunga . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Integrating Document Management with Project and Company Data
Dany Hajjar and Simaan M. AbouRizk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
 RIVER STAGE FORECASTING IN BANGLADESH:
NEURAL NETWORK APPROACH
By Shie-Yui Liong,1 Wee-Han Lim,2 and Guna N. Paudyal3

ABSTRACT: A relatively new approach, artificial neural network, was demonstrated in this study to be a highly
suitable flow prediction tool yielding a very high degree of water-level prediction accuracy at Dhaka, Bangladesh,
even for up to 7 lead days. The goodness-of-fit R2 value, root-mean-square error, and mean absolute error,
ranging from 0.9164 to 0.9958, 0.0788 to 0.2756 m, and 0.0570 to 0.2050 m, respectively, were obtained from
the training and verification simulation. In addition, the high degree of accuracy is accompanied with very small
computational time. Both results make the artificial neural network a desirable advanced warning forecasting
tool. Sensitivity analysis was also performed to investigate the importance of each of the input neurons. The
sensitivity study suggested a reduction of three from eight initially chosen input neurons. The reduction has
insignificantly affected the prediction accuracy level. The finding enables the policymakers to reduce the un-
necessary data collection at some gauging stations and, thus, results in lower costs.

INTRODUCTION Therefore, a sound flood forecasting tool is absolutely called

for.
Bangladesh is known as the land of six seasons, the most Historically, rainfall and river stage data at selected loca-
dominant being the rainy season, which lasts from May to tions are available for the past 32 years. Since 1990, water
September. The country is located at the Tropic of Cancer at stage data from about 40 river gauging stations and rainfall
longitude 90⬚ E and has a land area of 145,000 km2. About data from 46 rainfall stations within Bangladesh have been
60% of the land in Bangladesh is located on the world’s largest recorded. At present, the General Model (Mike 11 1992) of
delta comprising three of the world’s most unstable rivers, the the Surface Water Modelling Centre (SWMC) has been used
Ganges, the Brahmaputra, and the Meghna, which annually during the 6-month monsoon period to provide real-time flood
flood during the monsoon period. The rivers flowing into forecasts for 1, 2, and 3 lead days. The boundary data received
Bangladesh drain some of the wettest catchment areas on earth through a radio network are the input to the model. Fairly
with average annual rainfalls as high as 1,100 cm. The major accurate flood forecasts [e.g. SWMC (1996)] have been made
rivers and their tributaries have their origins outside Bangla- at key locations throughout the country.
desh, and only about 7.5% of their total catchment area of Conventional models such as the General Model require a
about 1,500,000 km2 lie within Bangladesh. Ninety percent of great deal of detailed data (e.g., topographical map, river net-
their annual flows originate outside the country. works and characteristics, soil characteristics, rainfall, and run-
The country experienced regularly severe floods such as off data). Often, these data are not available, which pose a
those in 1954, 1955, 1974, 1987, 1988, and 1996. The problem great difficulty in model calibration. In addition, a sufficiently
of floods in Bangladesh came to the forefront after two con- long lead time for forecasting is required to take the necessary
secutive severe floods of 1954 and 1955. The damages due flood evacuation measures. In the case of dissemination of
to floods became colossal when floods of severe magnitude flood information and other flood evacuation measures, com-
occurred again in the two consecutive years 1987 and putational speed of the models used are of absolute impor-
1988. Floods in 1987 inundated about 40% of the land area tance. Conventional models are not ideal for real-time fore-
(57,000 km2), affected about 30,000,000 people, and caused casting due to the large amount of detailed information they
about 1,800 deaths and estimated flood damages of require and the associated long computation time. A relatively
U.S.$500,000,000. In August and September 1988, the flood new approach, artificial neural network (ANN), is proposed in
inundated 60% of the land area (82,000 km2) including Dhaka, this study to overcome the above-mentioned problem.
the capital of Bangladesh, affected 45,000,000 people, and In this paper ANN is applied to forecast, for different lead
caused about 2,330 deaths and an estimated damage of days, the river stage at Dhaka based on river stages measured
U.S.$1.2 billion. In 1996, the flood was less severe. Never- at various upstream gauging stations and at Dhaka inclusively.
theless the flood-affected area was 36,000 km2. In this study forecasting models for up to 7 lead days are
Following the two successive catastrophic flood events in studied. Sensitivity analyses on the gauging stations’ data are
1987 and 1988 it has been recognized (Hoque 1994) that it is also conducted. The sensitivity analysis serves to eliminate
necessary to (1) examine the effectiveness of flood control gauging stations that insignificantly affect changes in the river
options; (2) improve the accuracy of current flood forecasting stage at Dhaka.
practices; (3) extend prediction beyond the major rivers in
Bangladesh; and (4) evaluate a sound flood action policy. NEURAL NETWORKS (NNs)

1
ANN is a computing paradigm designed to mimic the hu-
Assoc. Prof., Dept. of Civ. Engrg., Nat. Univ. of Singapore, 10 Kent man brain and nervous system. ANN has a big role to play in
Ridge Crescent, Singapore 119260.
2
Res. Engr., Dept. of Civ. Engrg., Nat. Univ. of Singapore, 10 Kent
the field of hydrology where complex natural processes dom-
Ridge Crescent, Singapore 119260. inate. The high degree of empiricism and approximation in the
3
Chf. Tech. Adviser, Surface Water Modelling Centre, House No. 476, analysis of hydrologic systems may find the use of ANN
Rd. 32 New DOHS, Dhaka, Bangladesh. highly suitable. In other words, when the possibility of rep-
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. resenting the complex relationships between various aspects of
To extend the closing date one month, a written request must be filed the processes in terms of physical or conceptual modeling is
with the ASCE Manager of Journals. The manuscript for this paper was
submitted for review and possible publication on December 16, 1998.
very remote, the ANN may play an important role.
This paper is part of the Journal of Computing in Civil Engineering, ANN problems can be classified into several groups such
Vol. 14, No. 1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0001– as pattern association, pattern classification, pattern clustering,
0008/$8.00 ⫹ $.50 per page. Paper No. 19885. and pattern mapping. In this study, the problem of capturing
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 1
 planning and design purposes, but a series of detailed infor-
mation is needed. The physical model also has been applied
quite successfully (Samuels and Reeve 1991; Syme and Pau-
dyal 1994; and SWMC 1996). The model makes use of the
capability of geographic information system technology to re-
trieve useful information such as topography, soil character-
istics, etc. Modeling a large river network such as that in
Bangladesh can pose a potential problem to modelers as large
volumes of data are needed. Model calibration is relatively
difficult because either the required data are not available or
data are expensive and time consuming to collect.
The application of ANN for advanced flood forecasting is a
promising alternative approach because it requires minimal
data. Feedforward NNs are closely related to statistical models
that are a data-driven approach and more suited to forecasting
applications. One of the major strengths of ANNs is their abil-
ity to recognize or correctly classify patterns that have never
been presented to the network before. ANNs appear unique in
their ability to extract the essential features from a training set
FIG. 1. Typical NN Configuration
and use them to identify new inputs or generalization.
The ANNs’ application to simulation and forecasting prob-
lems in water resources is relatively recent, particularly to hy-
the complex relationships between the input and the output drologic problems. French et al. (1992) used the ANN to fore-
patterns falls under the category of pattern mapping. Although cast rainfall patterns in space and time. A series of applications
there are many types of ANNs, by far the most widely used of ANN to catchment runoff forecasting followed. They in-
and well suited to this system is the feedforward NN or the clude works of Karunanithi et al. (1994), Bonafe et al. (1994),
multilayer perceptron, which is organized as layers of com- Hsu et al. (1995), Cheng and Noguchi (1996), and Liong et
puting elements (called neurons) connected by weighted con- al. (1996) to predict flow at the catchment outlet with inputs
nections between layers [Fig. 1(a)]. Typically, there is an input such as rainfall, upstream flow, and/or temperature only. Re-
layer, which receives inputs from the environment, an output cently, Thirumalaiah and Deo (1998) applied ANN to forecast
layer, which outputs the network’s response to its input, and river stage at a downstream station for different lead days us-
one or more intermediate layers called the hidden layers. ing only the most recent river stage at various stations as in-
The action of each neuron is to compute a response from puts. They determine the number of input stations for ANN
the weighted sum net of its inputs Xi and bias ␪ from neurons using a trial-and-error method. In the present study, however,
connected to it, using it predetermined activation function a more systematic way to determine the number of input sta-
F(net) [Fig. 1(b)]. The output of a neuron becomes the inputs tions using sensitivity analysis will be presented.
of other neurons in a higher layer. Many types of activation
function are in use; however, the most widely used is the sig-
moid function. Mathematically the above two steps may be
expressed as

冘
N

net = wi Xi ⫹ ␪ (1)
i=1

1
F(net) = (2)
1 ⫹ e⫺net

The feedforward network is trained using the back-propa-

gation method. Training is a process of adjusting the connec-
tion weights in the network so that the network’s response best
matches the desired response. Although this can be addressed
as an optimization method, the back-propagation methods
avoid this costly exercise by using an approximation to a gra-
dient descent method. Specifically, at each setting of the con-
nection weights, it is possible to calculate the error committed
by the network simply by taking the difference between the
desired and the actual response determined by the network.
Details on NNs can be found, for instance, in Rumelhart et al.
(1986), Simpson (1990), Fausett (1994), and Haykin (1994).

NNs AS FORECASTING TOOLS

Generally, forecasting models can be divided into statistical
and physically based approaches. Statistical approaches deter-
mine relationships between historical data sets, whereas phys-
ically based approaches model the underlying physical pro-
cesses directly. Hoque (1994) and Karim and Chowdhury
(1995) used different types of statistical distributions for flood FIG. 2. Bangladesh River System and Gauging Stations Used
frequency analysis in Bangladesh. This analysis is useful for in Study

2 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

DHAKA RIVER STAGE FORECASTING WITH NN inate from outside the country, the water level at Dhaka is
assumed to be mainly affected by flow coming from these
Advanced and accurate forecasting of river stages in Bang- major rivers. Thus, rainfall data within the country is not con-
ladesh is important for water resources planning and flood sidered in the study. Daily river stage data from a total of eight
warning. During a monsoon season an accurate river stage gauging stations are selected and used in the study (Fig. 2).
forecast is tremendously useful because it would allow the The eight stations are Dhaka (ST 1); Rajshahi (ST 2); Chilmari
authority to issue a warning of the impending flood and to (ST 3); Durgapur (ST 4); Kanairghat (ST5); Amalshid (ST6);
take evacuation measures, if necessary. At the same time, it Habiganj (ST 7); and Gorai RB (ST 8). It should be noted that
would help in regulating outflows from dams and/or reservoirs these stations are the most upstream gauging stations on the
during this period. major rivers in Bangladesh, and the distances of the upstream
In this paper river stage forecasting at Dhaka is the focus boundary stations from Dhaka range from about 160 to 355 km.
of the study. The city of Dhaka is located in the flat region Data records for 3 years (1992, 1993, and 1995) are selected
near the confluence of three major magnificent rivers, the Gan- for NN training because they represent the variety of river
ges, the Brahmaputra, and the Meghna. It should be noted that stage data of the considered 6-year record. The remaining rec-
to prevent flooding of the city, embankments or dikes have ords of data from 1991, 1994, and 1996 are reserved for ver-
been constructed surrounding the city. Thus, it is very impor- ification of the trained forecaster. The following relationship
tant to have an accurate forecasting tool with long lead days between the input and output neurons is proposed:
for the authority to undertake proper flood prevention and mit-
igation related measures. X1(t) = f [X1(t ⫺ ␣), X2(t ⫺ ␣), . . . , X8(t ⫺ ␣)] (3)
The application of ANN is first demonstrated on the forecast
of daily river stage at Dhaka for the rainy season period from where Xi (t) represents the daily river stage at time t of the
May 1 until September 30, each year from 1991 to 1996 (in various stations i = 1, 2, . . . , 8, and ␣ = number of lead days
1996, the period of record ends on July 17). The interest in from 1 to 7 days. The NN’s architecture and parameters are
the study is to develop a good forecasting model to predict first optimized. In the study, the following NNs characteristics
the river stage data at Dhaka up to 7 lead days. Long lead (Fig. 3) are adopted: (1) The input layer has 8 neurons rep-
days are essential so that the policymakers may take necessary resenting the daily water levels of the eight stations; (2) only
flood mitigation and/or evacuation measures. one hidden layer with 30 neurons is used; (3) 1 neuron in the
Because 90% of the annual flows of the major rivers orig- output layer represents the water level at Dhaka for a certain
number of the lead days; (4) 2,000 training epochs as the stop-
ping criteria; (5) sigmoid function is used as the activation
function; (6) a learning rate and a momentum constant of 0.1;
and (7) an initial weight distribution of ⫾0.3 is used. The input
and output river stages were normalized to range from 0 to 1
with 0 and 1 being equivalent to the minimum and the max-
imum river stages, respectively, of the respective station re-
corded. A cross-validation method (Bishop 1994; Haykin
1994) was used to overcome the overfitting problem. Rather
than letting the neural network learn the training set until a
minimum root-mean-square error (RMSE) is achieved, the net-
work is cross-validated with the testing set at the end of each
learning epoch. The testing set is extracted from a small por-
tion (about 5–10%) of the whole training data set. Should
overfitting occur, the testing set’s error learning curve will start
increasing instead of decreasing although the training set still
continues learning. A commercially developed NN software,
NeuroShell 2 (NeuroShell 1993), is used in the study.
FIG. 3. ANN Structure in Forecasting Dhaka River Stage The performance of the predictions resulting from both the

TABLE 1. Water-Level Prediction Errors and Goodness-of-Fit: Eight-Input Neuron Model

Goodness-of-Fit Water Level
Error’s standard
Lead time RMSE MAE Maximum error deviation
Status (days) R 2 value (m) (m) (m) (m)
(1) (2) (3) (4) (5) (6) (7)
Training 1 0.9960 0.0765 0.055 ⫹0.345 0.0765
2 0.9904 0.1171 0.088 ⫹0.554 0.1152
3 0.9827 0.1504 0.118 ⫹0.806 0.1505
4 0.9762 0.1819 0.138 ⫹0.974 0.1804
5 0.9638 0.2219 0.167 ⫹0.986 0.2065
6 0.9583 0.2355 0.177 ⫹1.032 0.2354
7 0.9451 0.2672 0.198 ⫹1.169 0.2640
Verification 1 0.9938 0.0781 0.057 ⫹0.471 0.0749
2 0.9837 0.1219 0.097 ⫹0.472 0.1132
3 0.9697 0.1537 0.132 ⫹0.672 0.1515
4 0.9610 0.1851 0.149 ⫹0.626 0.1762
5 0.9380 0.2307 0.183 ⫹0.706 0.2001
6 0.9385 0.2272 0.180 ⫹0.644 0.2223
7 0.9195 0.2569 0.202 ⫹0.891 0.2450
Note: ⫹ implies that measured value is underestimated.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 3

training and the verification is evaluated by the following mea-
sure for goodness-of-fit, coefficient of efficiency R2 (Nash and
Sutcliffe 1970; ASCE 1993), RMSE, and mean absolute error
(MAE) expressed as:

冘
p

[(Xm)i ⫺ (Xs)i]2

冘
R =1⫺
2 i=1
p (4)
[(Xm)i ⫺ (X
¯ m)i]2
i=1

冑冘
p
1
RMSE = [(Xm)i ⫺ (Xs)i]2 (5)
p i=1

冘
p
1
MAE = 兩 (Xm)i ⫺ (Xs)i 兩 (6)
p i=1

where the subscripts m and s represent the measured and the

simulated water levels, respectively; p = total number of events
considered; and X̄m = mean value of the measured data. The
R2 can be used to indicate the relative assessment of model
performance in dimensionless measures, and RMSE and MAE
can give a quantitative indication of the model error in units
of the variable. It should be noted that the R2 value of 1, in
which case the value of RMSE and MAE is 0, implies a per-
fect match between the measured and the simulated data.
Table 1 lists the R2, RMSE, and MAE values resulting from
training and verifications for various lead days ranging from
1 to 7 days. The R2 values, RMSE, and MAE for training and
verifications range from 0.9195 to 0.9960, 0.0765 to 0.2672
m, and 0.0550 to 0.2020 m, respectively, meaning that very
good predictions with relatively small errors have been
achieved. Figs. 4(a–d) show some samples of comparison be-
tween the measured and the simulated water levels for 1, 3,
5, and 7 lead days, respectively.
The impact of the water stage changes at each of the eight
gauging stations on the Dhaka water stage is then conducted.
Two sensitivity tests—a step-by-step sensitivity analysis and
a multiple regression based sensitivity analysis—are applied
in this study.

SENSITIVITY ANALYSIS
The two sensitivity tests are carried out to determine the
relative significance of each of the input neurons. Both anal-
yses aim to reduce the number of input neurons should they
prove to be insignificant to model performance. Reduction of
input neurons would result in a decrease in unnecessary data
collection that in turn leads to cost reduction.
The first sensitivity test, a step-by-step method, is carried
out on the trained forecaster by varying each of the input neu-
rons, one at a time, at a constant rate. Various constant rates
(⫺5, 5, 10, and 20%) are selected in the study. For every input
neuron, the percentage change in the output, as a result of the
change in the input neuron, is observed. The sensitivity of each
input neuron is calculated by the following:

冘冉 冊
N
1 % change in output
Sensitivity Level of Xi (%) = ⫻ 100
N j=1 % change in input j
(7)
where N (=841) = number of data sets used in the study. FIG. 4. Comparison between Simulated and Measured Water
Results show that the change in output neuron is unaffected Levels with Eight Input Neurons for Verification Set: (a) 1 Lead
by the rate an input neuron (⫺5, 5, 10, or 20%) is varied. Fig. Day; (b) 3 Lead Day; (c) 5 Lead Day; (d) 7 Lead Day
5(a) shows how sensitive the water-level changes at each of
the gauging stations are and how they will affect the changes water level at Dhaka); this unexpected response may imply
in water level at Dhaka. ST 1 (Dhaka itself) and ST 3 are the that water at ST 6 and ST 8 do not flow toward Dhaka at all.
two most sensitive gauging stations. ST 6 and ST 8 have a ST 6 and ST 8 are then eliminated from the input neurons.
reverse impact on Dhaka water level (i.e., increasing water Thus, the total number of input neurons reduces to six. A sim-
levels at ST 6 and/or ST 8 will result in a decrease of the ilar sensitivity test is again carried out on the model with six
4 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
FIG. 5. Sensitivity Levels of Input Stations for 1 to 7 Lead Days Using Step-by-Step Sensitivity Test: (a) Eight Input Stations; (b) Six
Input Stations; (c) Five Input Stations

input neurons. Results are presented in Fig. 5(b). ST 1 and ST and the output neuron. The input and output neurons are first
3 still are the most sensitive stations, whereas ST 5 displays related through a power function [(8)], that is then transformed
an insignificant reverse impact. The sensitivity test is, once into a logarithmic equation [(9)] as shown below
again, conducted on the model with only five input neurons
(without ST 5). The final sensitivity of each of the five input X1(t) = b0[X1(t ⫺ ␣)]b1[X2(t ⫺ ␣)]b2 ⭈ ⭈ ⭈ [X8(t ⫺ ␣)]b8 (8)
neurons is shown in Fig. 5(c).
The second sensitivity test, a multiple regression based sen- ln[X1(t)] = ln b0 ⫹ b1 ln[X1(t ⫺ ␣)] ⫹ b2 ln[X2(t ⫺ ␣)]
sitivity analysis, is carried out by observing the coefficients of
the log linear regression equation relating the input neurons ⫹ ⭈ ⭈ ⭈ ⫹ b8 ln[X8(t ⫺ ␣)] (9)

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 5

FIG. 6. Sensitivity Levels of Eight Input Stations for 1 to 7 Lead Days Using Multiple Linear Regression Based Sensitivity Test

where bi = model coefficients estimated using multiple linear both models. This implies that the sensitivity analysis con-
regression. Each of the model coefficients can be interpreted ducted has indeed removed the insignificant gauging stations.
as the sensitivity of the input neuron relative to the output As can be seen in Figs. 7(a–c), the training sets for different
neuron. The larger the coefficient is, the more sensitive is the lead days, as expected, yield higher R2 values, and thus less
corresponding input neuron to the output neuron. RMSE and MAE, than the verification sets. However, a similar
Fig. 6 shows the sensitivity level of each of the eight input conclusion cannot be drawn for maximum error and error’s
neurons. Results from this regression based sensitivity test standard deviation [Figs. 7(d and e)]. As the number of lead
show a similar trend as that in the step-by-step sensitivity test days increases, again as expected, the R2 values decrease in
[i.e. only five gauging stations (ST 1, ST 2, ST 3, ST 4, and both models. Both maximum error and error’s standard devi-
ST 7) are sensitive to Dhaka water level]. ation in the verification sets are smaller than that of the train-
ing sets. This implies that the training data sets have been well
selected to represent the available data sets. Comparison be-
DHAKA RIVER STAGE FORECASTING WITH FIVE tween measured and simulated water levels resulting from the
INPUT NEURONS five-input neuron model yields an almost identical plot as that
of the eight-input neuron model (Fig. 4).
With the five significant input neurons a new NN forecaster A sample of advanced forecasting on the verification set
is then trained. Table 2 lists the R2 values, RMSE, MAE, and using the five-input neuron model is shown in Fig. 8. Fig. 8(a)
other measures of prediction accuracy resulting from the five- shows a comparison between the measured and simulated wa-
input neuron model. The R2 values, RMSE, and MAE for ter levels on the verification set for 7 lead days. Results are
training and verifications range from 0.9164 to 0.9958, 0.0788 also presented in a scatter plot [Fig. 8(b)]. Figs. 8(a and b)
to 0.2756 m, and 0.0570 to 0.2050 m, respectively, meaning show that the simulated river stages at Dhaka mimic their mea-
that very good predictions with relatively small errors have sured counterparts very closely. Most of the prediction errors
been achieved. Figs. 7(a–e) graphically display the differences are within the ⫾0.5 m error range. It should be noted that the
in values (R2 values, RMSE, MAE, maximum error, and er- maximum error occurs at the beginning of the monsoon sea-
ror’s standard deviation, respectively) listed in Tables 1 and 2 son, of the year 1991, at which the flow was very low. The
resulting from the eight- and five-input neuron models. No relatively large error prediction may be due to the small num-
significant difference is observed between the results from ber of training patterns available at this low flow range.

TABLE 2. Water-Level Prediction Errors and Goodness-of-Fit: Five-Input Neuron Model

Goodness-of-Fit Water Level
Error’s standard
Lead time RMSE MAE Maximum error deviation
Status (days) R 2 value (m) (m) (m) (m)
(1) (2) (3) (4) (5) (6) (7)
Training 1 0.9958 0.0788 0.057 ⫹0.360 0.0780
2 0.9903 0.1187 0.089 ⫹0.535 0.1187
3 0.9834 0.1535 0.116 ⫹0.747 0.1536
4 0.9748 0.1871 0.140 ⫹0.985 0.1843
5 0.9614 0.2292 0.170 ⫹0.989 0.2115
6 0.9557 0.2427 0.179 ⫹1.017 0.2417
7 0.9417 0.2756 0.201 ⫹1.126 0.2727
Verification 1 0.9934 0.0791 0.058 ⫹0.475 0.0761
2 0.9853 0.1164 0.093 ⫹0.441 0.1163
3 0.9740 0.1529 0.123 ⫹0.614 0.1525
4 0.9602 0.1868 0.150 ⫹0.631 0.1791
5 0.9376 0.2315 0.183 ⫹0.705 0.2049
6 0.9347 0.2340 0.184 ⫹0.678 0.2301
7 0.9164 0.2617 0.205 ⫹0.857 0.2557
Note: ⫹ implies that measured value is underestimated.

6 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

FIG. 7. Comparison of Performance between Eight- and Five-Input Neuron Models: (a) R 2 Values; (b) RMSE; (c) MAE; (d) Maximum
Error; (e) Error’s Standard Deviation

FIG. 8. Simulated and Measured Daily Water Levels for Five-Input Neuron Model at 7 Lead Days—Verification Set: (a) Water-Level
Plot; (b) Scatter Plot

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 7

CONCLUSIONS Karim, M. A., and Chowdhury, J. U. (1995). ‘‘A comparison of four
distributions used in flood frequency analysis in Bangladesh.’’ Hydro.
The application of an NN has been successfully demon- Sci. J., 40(1), 55–66.
strated on flow forecasting in Dhaka, Bangladesh. The flow Karunanithi, N., Grenney, W. J., Whitley, D., and Bovee, K. (1994).
‘‘Neural networks for river flow prediction.’’ J. Comp. in Civ. Engrg.,
prediction obtained has a very high degree of accuracy even
ASCE, 8(2), 201–220.
for a 7-lead-day model. Although ANNs belong to a class of Liong, S. Y., Cheu, R. L., and Chan, W. T. (1996). ‘‘Near-perfect hydro-
data-driven approaches, it is important to determine the dom- graph forecasting with neural network.’’ J. Instn. of Engrs. Singapore,
inant model inputs, as this reduces the size of the network and Singapore, 36(2), 54–57.
consequently reduces the unnecessary data collection. Sensi- Mike 11—A micro-computer based modelling system for rivers and chan-
tivity analysis has been introduced to verify the importance of nels. (1992). Danish Hydraulic Institute, Horsholm, Denmark.
Nash, J. E., and Sutcliffe, J. V. (1970). ‘‘River flow forecasting through
each of the input neurons. The analysis leads to the elimination
conceptual models, Part 1—A discussion of principles.’’ J. Hydrol.,
of three (out of a total of eight) less sensitive input neurons. Amsterdam, 10(3), 282–290.
The removal of the less sensitive input neurons insignificantly NeuroShell 2. (1993). Ward Systems Group, Inc., Md.
reduces the accuracy degree of the eight-input neuron model. Rumelhart, D., Hinton, G., and Williams, R. (1986). ‘‘Learning internal
Thus, the study demonstrates two results: (1) The suitability representations by error propagation.’’ Parallel distributed processing:
of an NN for flow prediction with high accuracy at a fraction Exploration in the microstructure of cognition, D. Rumelhart and J.
McClelland, eds., Vol. 1, MIT Press, Cambridge, Mass., 318–362.
of the computational time required by the conventional rain-
Samuels, P. G., and Reeve, C. E. (1991). ‘‘River flooding and morphology
fall-runoff models; and (2) a technique for detecting less sen- in Bangladesh.’’ Proc., Int. Symp. on Envir. Hydr., J. H. W. Lee and
sitive input neurons in an effort to reduce unnecessary data Y. K. Cheung, eds., Balkema, Rotterdam, The Netherlands, 1377–1382.
collection and operational cost. Simpson, P. K. (1990). Artificial neural systems: Foundations, paradigms,
applications and implementations. Pergamon, New York.
Surface Water Modelling Centre (SWMC). (1996). ‘‘Surface water sim-
ACKNOWLEDGMENTS ulation modelling programme phase III.’’ Final Report, Volume II: Sur-
The first two writers wish to acknowledge the financial support given face water models, River Research Institute, Ministry of Water Re-
by the National University of Singapore under research project sources, Government of the People’s Republic of Bangladesh in
RP950685. Water-level data have been provided by Danish Hydraulic cooperation with Danida, Surface Water Modeling Centre, Dhaka,
Institute, Horsholm, Denmark, and are hereby acknowledged with grati- Bangladesh.
tude. Syme, W. J., and Paudyal, G. N. (1994). ‘‘Bangladesh flood management
model.’’ Proc., 2nd Int. Conf. on River Flood Hydr., W. R. White and
J. Watts, eds., Wiley, Chichester, England and New York, 167–176.
APPENDIX I. REFERENCES Thirumalaiah, K., and Deo, M. C. (1998). ‘‘River stage forecasting using
artificial neural networks.’’ J. Hydrologic Engrg., ASCE, 3(1), 26–32.
ASCE Task Committee on Definition of Criteria for Evaluation of Wa-
tershed Models. (1993). ‘‘Criteria for evaluation of watershed models.’’
J. Irrig. and Drain. Engrg., ASCE, 119(3), 429–442. APPENDIX II. NOTATIONS
Bishop, C. M. (1994). ‘‘Neural networks and their applications.’’ Rev. of
Scientific Instruments, 65, 1803–1832. The following symbols are used in this paper:
Bonafe, A., Galeati, G., and Sforna, M. (1994). ‘‘Neural networks for
daily mean flow forecasting.’’ Proc., 5th Int. Conf. on Hydr. Engrg. bi = model coefficients;
Software (HYDROSOFT ’94), W. R. Blain and K. L. Katsifarakis, eds., e = 2.718281828;
Vol. 1, Computational Mechanics, Southampton, U.K. and Boston, F(net) = activation function;
131–138. f[ ] = nonlinear function between input and output neurons;
Cheng, X., and Noguchi, M. (1996). ‘‘Rainfall-runoff modelling by neural
network approach.’’ Proc., Int. Conf. on Water Resour. and Envir. Res.,
ln( ) = natural log;
University of Kyoto, Kyoto, Japan, Vol. 2, 143–150. N = total number of data sets;
Fausett, L. (1994). Fundamentals of neural networks: Architectures, al- net = weighted sum of inputs to neuron;
gorithms, and applications. Prentice-Hall, Englewood Cliffs, N.J. p = total number of events;
French, M. N., Krajewski, W. F., and Cuykendall, R. R. (1992). ‘‘Rainfall R2 = goodness-of-fit;
forecasting in space and time using a neural network.’’ J. Hydrol., wi = weight for i inputs;
Amsterdam, 137, 1–31. Xi = water level for input station i;
Haykin, S. (1994). Neural networks: A comprehensive foundation. Mac- Xi (t) = water level for input station i at time t;
millan, New York. (Xm)i = measured water level for input station i;
Hoque, M. M. (1994). ‘‘Evaluating design characteristics for floods in
Bangladesh.’’ Proc., 2nd Int. Conf. on River Flood Hydr., W. R. White
(X̄m)i = mean of measured water level for input station i;
and J. Watts, eds., Wiley, Chichester, England and New York, 15–26. (Xs)i = simulated water level for input station i;
Hsu, K. L., Gupta, H. V., and Sorooshian, S. (1995). ‘‘Artificial neural Y = water level for output neuron;
network modeling of the rainfall-runoff process.’’ Water Resour. Res., ␣ = number of advanced warning days; and
31(10), 2517–2530. ␪ = threshold value.

8 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 AUGMENTED IFN LEARNING MODEL
By Shih-Lin Hung,1 Member, ASCE, and J. C. Jan2

ABSTRACT: Solving engineering problems is a creative, experiential process. An experienced engineer gen-
erally solves a new problem by recalling and reusing some similar instances examined before. According to
such a method, the integrated fuzzy neural network (IFN) learning model was developed and implemented as
a computational model for problem solving. This model has been applied to design problems involving a
complicated steel structure. Computational results indicate that, because of its simplicity, the IFN model can
learn the complicated problems within a reasonable computational time. The learning performance of IFN,
however, relies heavily on the values of some working parameters, selected on a trial-and-error basis. In this
work, we present an augmented IFN learning model by integrating a conventional IFN learning model with two
novel approaches—a correlation analysis in statistics and a self-adjustment in mathematical optimization. This
is done to facilitate the search for appropriate working parameters in the conventional IFN. The augmented IFN
is compared with the conventional IFN using two steel structure design examples. This comparison reveals a
superior learning performance for the augmented IFN learning model. Also, the problem of arbitrary trial-and-
error selection of the working parameters is avoided in the augmented IFN learning model.

INTRODUCTION tion (BP, Rumelhart et al. 1986) learning algorithm. Vanluch-

ene and Sun (1990) applied the back-propagation learning
Solving engineering problems—such as those of analysis algorithm to structural engineering. Gunaratnam and Gero
and design—is a creative, experiential process in which the (1994) discussed the effect of the representation of input/out-
experiences and combined knowledge of engineers serve as put pairs on the learning performance of a BP neural network.
resources. An experienced engineer generally solves a new Also, several other researchers have applied neural networks
problem in the following stages. First, he or she recalls in- to engineering design and related problems (Hajela and Berke
stances that were similar but produced resolution, while prop- 1991; Ghaboussi et al. 1991; Kang and Yoon 1994; Stephen
erly considering the functional requirements of those instances. and Vanluchene 1994; Elkordy et al. 1994).
Then, the engineer attempts to derive the solution from these The learning processes of back-propagation (BP) supervised
similar instances through adaptation or synthesis. After the neural network learning models, however, always take a long
problem is solved, the new instance is then stored in her/his time. Therefore, several different approaches have been de-
memory as an additional knowledge resource for solving fu- veloped to enhance the learning performance of the BP learn-
ture problems. ing algorithm. Such an approach is to develop parallel learning
The described stages can be implemented as a computa- algorithms on multiprocessor computers to reduce the overall
tional model for problem solving, one that utilizes a case base computational time (Hung and Adeli 1991b, 1992, 1994b;
of previously solved problems when solving a new one. In Adeli and Hung 1993a). Another approach is to develop more
symbolic artificial intelligence (AI), case-based reasoning effective neural network learning algorithms to reduce learning
(Carbonell 1981) is an effective means of facilitating computer cycles (Adeli and Hung, 1993a, 1994a; Hung and Lin, 1994).
program development. It attempts to solve problems by di- A third approach involves the development of a hybrid learn-
rectly accessing the case base. This approach relies on the ing algorithm—for instance, by integrating a genetic algo-
explicit symbolic representation of a case base established rithm with neural network algorithms to improve the overall
from experience. With a given case base, case-based reasoning learning performance (Hung and Adeli 1991b, 1994b).
uses a representation involving specific episodes of problem Another category of learning in ANN includes unsupervised
solving, not only to solve a new problem, but also to learn neural network learning models that are generally used in clas-
how to solve the new problem. Based on the approach of case- sification problems (Carpenter and Grossberg 1988; Adeli and
based reasoning, the way to solve engineering problems has Hung 1993b). In structural engineering, Adeli and Park (1995)
received considerable attention (Maher et al. 1995). employed a counterpropagation neural network (CPN), which
Artificial neural networks (ANNs), on the other hand, con- combines supervised and unsupervised neural networks, to
stitute a different AI approach, one that has made rapid ad- solve complicated engineering design problems. That investi-
vances in recent years. Such networks have the ability to de- gation concluded that a CPN learning model can learn how to
velop, from training instances, their own solutions to a class solve complicated structural design problems within a reason-
of problems. The method of representation used by ANNs, able computation time. Recently, authors Hung and Jan (1997,
essentially a continuous function, is conducive to generaliza- 1999) presented an integrated fuzzy neural network (IFN)
tion beyond the original set of training instances used in their
learning model in structural engineering. The IFN learning
development. The feasibility of applying ANN computing to
model combined a novel unsupervised fuzzy neural network
engineering problems has received increasing interest, partic-
(UFN) reasoning model with a supervised neural network
ularly on supervised neural networks with the back-propaga-
learning model using the adaptive L-BFGS learning algorithm
1 (Hung and Lin 1994). The IFN learning model was applied to
Assoc. Prof., Dept. of Civ. Engrg., National Chiao Tung Univ., 1001
Ta Hsueh Rd., Hsinchu, Taiwan 300, R.O.C. steel beam design problems. That work contended that the IFN
2
PhD Candidate, Dept. of Civ. Engrg., National Chiao Tung Univ., learning model is a robust and effective ANN learning model.
1001 Ta Hsueh Rd., Hsinchu, Taiwan 300, R.O.C. In addition, the IFN model can interpret a large number of
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. instances for a complicated engineering problem within a rea-
To extend the closing date one month, a written request must be filed sonable computational time, owing to its simplicity in com-
with the ASCE Manager of Journals. The manuscript for this paper was
putation. However, the performance of the IFN learning model
submitted for review and possible publication on June 1, 1999. This paper
is part of the Journal of Computing in Civil Engineering, Vol. 14, No. is heavily affected by some working parameters that are prob-
1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0015–0022/$8.00 lem dependent and obtained via trial and error.
⫹ $.50 per page. Paper No. 21139. In this work, we present a more effective neural network
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 15
learning model, called the augmented IFN learning model, by instances U1, U2, . . . , UN and X is a new instance. Instance
integrating a conventional IFN learning model with two newly Uj is defined as a pair, including input Uj,i and its correspond-
developed approaches. The first approach, correlation analysis ing output Uj,o. If there are M decision variables in the input
in statistics, is employed to assist users in determining the and K items of data in the output, the input Uj,i and output Uj,o
appropriate working parameter to be used in the fuzzy mem- of instance Uj are represented as vectors of the decision vari-
bership function. The second approach, self-adjustment in ables and data, and are denoted as Uj,i = {u1j , u 2j , . . . , u M
j } and
mathematical optimization, is used to obtain appropriate Uj,o = {o1j , o2j , . . . , ojK}. Similarly, the new instance X can also
weights, systematically, for each decision variable required in be defined as a pair including input Xi and unsolved output
the input of training instances. The novel model is imple- Xo, respectively. The input Xi is a set of decision variables as
mented in C language on a DEC3000/600 workstation. The Xi = {x1, x 2, . . . , x M }. The output Xo is currently a null vector.
augmented IFN learning model proposed herein is applied to The learning stage in the IFN model is performed in two
two structural engineering problems to verify its learning per- sub-ANN models concurrently. First, the offline assistant su-
formance. The first example is a steel beam design problem. pervised neural network model is trained, based on the adap-
The second example involves the preliminary design of steel tive L-BFGS supervised learning algorithm using these N
structure buildings. The two examples are also used to train a given instances. In the UFN reasoning model, however, the
conventional IFN learning model, for the sake of comparison. learning process simply involves selecting appropriate work-
ing parameters for the fuzzy membership function and weights
REVIEW OF IFN LEARNING MODEL for each decision variable in the input. The process is imple-
mented in the following steps. The first step attempts to de-
This section briefly reviews the integrated fuzzy neural net- termine the degree of difference between any two distinct in-
work (IFN) learning model (Hung and Jan 1997, 1999). The stances in base V, which contains P instances randomly
IFN learning model combines two sub-ANN learning models. selected from instance base U. Therefore, a total of T = C P2 =
One is a novel, unsupervised fuzzy neural network (UFN) rea- P(P ⫺ 1)/2 degrees of difference must be computed. The func-
soning model: a single-layered laterally connected network tion of degree of difference, diff (Ui,i, Uj,i), is employed to mea-
with an unsupervised competing learning algorithm. The other sure the difference, dij , of two inputs Ui,i and Uj,i for instances
is an offline assistant model: a supervised neural network Ui and Uj in V. The function is defined as the modified square
learning model with the adaptive L-BFGS learning algorithm of Euclidean distance and represented as
(Hung and Lin 1994). The IFN learning model is schemati-

冘
M
cally depicted in Fig. 1.
Assume that U is an associated instance base, with N solved dij = diff (Ui,i, Uj,i) = ␣m(wi u mi ⫺ wj u mj )2 (1)
m=1

where wi and ␣m denote predefined weights and are used to

represent the degree of importance for the ith instance in the
instance base for the mth decision variable in the input. The
weights wi are generally set as constant one. The weights ␣m,
however, are set by trial and error. After the values of dij for
all instances in base V are computed, the average of the sum
of dij , denoted as d̄ij = avg(dij) = 兺t=1
T
(dij)/T, can be computed.
The second step entails determining the fuzzy membership
function. The relationship of ‘‘similarity’’ between any two
instances is represented using a fuzzy membership function.
In the UFN reasoning model, a quasi-Z-type membership func-
tion is used and defined as
␮(dij) = f (Rmax, Rmin, dij)

=
再0
Rmax Rmin ⫺ Rmindij

1
(Rmax ⫺ Rmin)dij
if dij ⱖ Rmax
if Rmin < dij < Rmax

if dij ⱕ Rmin (2)

The terms Rmax and Rmin are two working parameters that define
the upper and lower bounds of ‘‘degree of difference.’’ The
lower bound Rmin is set as a constant 10⫺5. The upper bound,
however, is set as Rmax = ␩d¯ ij . The term ␩ is a real number
between zero and one and it was set by trial and error. Con-
sequently, the degree of similarity for instances Ui and Uj can
be determined from the fuzzy membership function.
After learning in the UFN and in the assistant supervised
learning model is completed, any new instance X can be solved
via the IFN learning method. The reasoning in UFN is per-
formed through a single-layered, laterally connected network
with an unsupervised competing algorithm, and it is imple-
mented in three steps. The first step involves searching for
some instances from the instance base U that resemble the new
instance X according to their inputs; that is, the degree of dif-
ference, dXj , between the inputs, Xi and Uj,i, for instance X and
instance Uj in base U is calculated. The input Xi of instant X
FIG. 1. Conventional IFN Learning Model is presented to the first node and the input Uj,i of instance Uj
16 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
is presented to the ( j ⫹ 1)th node simultaneously. The degree stances to the new instance can be found and, ordinarily, no
of difference dXj is then computed using (1) as diff (Xi, Uj,i) = solution can be generated via the UFN reasoning model. On
兺Mm=1 ␣m(wx x
m
⫺ wj u mj )2. The second step entails representing the other hand, a loose similar relationship is adopted under
the relationships among the new instance and its similar in- the case of a larger Rmax. Accordingly, a large number of in-
stances, with dXj less than Rmax. The fuzzy set, Ssup,X , of ‘‘similar stances are taken to be ‘‘similiar instances’’ and the solution
to X ’’ was then expressed as follows: generated via these similar instances is inferior. Here, the lin-
ear correlation analysis in statistics is employed to facilitate
Ssup,X = (␮1 /S1) ⫹ (␮2 /S2) ⫹ ⭈ ⭈ ⭈ ⫹ (␮p /Sp) ⫹ ⭈ ⭈ ⭈ (3) the determination of appropriate value of Rmax in the fuzzy
where Sp = pth similar instance to instance X; the term ‘‘⫹’’ membership function. The analysis is a process that aims to
denotes a union operator; and ␮p = corresponding fuzzy mem- measure the strength of the association between two sets of
bership value. variables that are assumed to be linearly related.
The final step involves generating the output Xo vector of For the above instance base U with N instances, the corre-
instance X by synthesizing the outputs of its similar instances lation analysis in the fuzzy membership function is imple-
according to their associated fuzzy membership values using mented in the following steps. The first step is to determine
center of gravity (COG) or mean of maximum (MOM) meth- the degree of difference between any two instances in the base
ods (Hung and Jan 1999). For the given unsolved instance X, U using the aforementioned function of degree difference in
assume that P similar instances are in fuzzy set Ssup,x ; they are (1). Hence, a total of (C N2 ⫹ N ) resembling samples Sij(Ui,o,
classified into C distinct clusters according to their outputs. Uj,o, dij) can be compiled. A resembling sample contains two
Then, the output Xo of instance X yielded via these two meth- instances’ outputs (Ui,o and Uj,o) and the corresponding degree
ods are defined, respectively, as follows: of difference (dij). Thereafter, two arrays, At and Bt, can be

冘
p
assorted from resembling samples in the case of dij less than
or equal to a prescribed value, say t. The elements in At and
␮k Sk,o Bt are the first and second items, respectively, of these resem-

冘
k=1
Xo = p (COG) (4) bling samples. Next, the accumulative correlation coefficient,
␮k Ac㛭CORREL(At, Bt, t), is calculated for arrays At and Bt with
k=1 the degree of difference less than or equal to t. Assume that

冘
C for a total P resembling samples with dij less than a prescribed
␮ kmaxS max,o
k t, the arrays At and Bt can be denoted as

冘
k=1
Xo = C (MOM) (5) At = {ak 兩ak = Ui,o 僆 Sij , for dij ⱕ t} = {a1, a2, . . . , ap}
␮ max
k Bt = {bk 兩bk = Uj,o 僆 Sij , for dij ⱕ t} = {b1, b2, . . . , bp}
k=1

The value of the accumulative correlation coefficient equals

where ␮k denotes the membership value for the kth similar
instance in fuzzy set Ssup,x . Correspondingly, ␮ max
k denotes the Cov(At, Bt, t)
membership value for the most similar instance S kmax in the kth Ac㛭CORREL(At, Bt, t) = (6)
␴At␴Bt
cluster.

冘
p
The reasoning process of the UFN depends on determining 1
the degree of similarity among X and Uj . Consequently, no Cov(At, Bt, t) = (ak ⫺ ␮At)(bk ⫺ ␮Bt) s.t. dij ⱕ t (7)
p k=1
solution can be generated by the UFN reasoning model if the
new instance entirely differs from all instances in the instance where ␴At and ␴Bt are standard errors of arrays At and Bt; ␮At
base, e.g., all dXi are greater than Rmax. In addition, using in- and ␮Bt are the means of At and Bt. The formulas expressed in
appropriate working parameters would allow for the possibility (6) and (7) represent the relationship between the accumulative
that no similar instances can be derived. For the above issues, correlation coefficient to any value of t. An accumulative cor-
the undertrained adaptive L-BFGS supervised neural network relation curve can be plotted as a function of t and Ac㛭COR-
is used as an assistant system to generate an approximate out- REL(At, Bt, t). Fig. 2 demonstrates an accumulative correlation
put for the new instance. curve for two arrays At = sin (x) and Bt = 0.95(x ⫺ x 3/3! ⫹
x 5/5!) such that x = ⫺␲ ⫹ (␲/15)i, i = 0 to 30. Shown in Fig.
AUGMENTED IFN LEARNING MODEL 2, the curve falls from one to zero as the value of t increases,
and it resembles the quasi-Z-type fuzzy membership function
In the conventional IFN learning model, working param- defined in (2) for the case of Rmin equal to zero. Note that the
eters (such as wj , ␣m, Rmax, and Rmin) are selected subjectively
by users, and generally, on a trial-and-error basis. Conse-
quently, the learning performance is highly affected by these
parameters, especially Rmax and ␣m. In this work, two novel
approaches are employed for assisting the users to determine
these parameters and weights systematically. One approach,
correlation analysis in statistics, is used to determine the ade-
quate Rmax value in the fuzzy membership function. The other
approach, self-adjustment based on mathematical optimization
theory, is employed to find proper values for weights ␣m.

Correlation Analysis for R max in Fuzzy

Membership Function
In conventional IFU, the similarity measurement between
two instances heavily depends on the value of parameter Rmax.
A small value of Rmax implies that a strict similar relationship
between instances is utilized. Consequently, most of the in-
stances are sorted as dissimilar. As a result, few similar in- FIG. 2. Example of Accumulative Correlation Curve

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 17

appropriate Rmax equals a certain value of t, such that instances The iteration is terminated as the value of 储g (s⫹1)储 or
in the instance base U have a certain degree of correlation. 兩E(␣(s⫹1)
m ) ⫺ E(␣m(s))兩 is sufficiently small. The term g(s) is the
Obviously, the smaller the t implies a larger accumulative negative gradient vector of function E(␣m). For simplicity, the
correlation coefficient, indicating a strong relationship between superscript (s), denoted as the sth iteration, is ignored. Here-
the two arrays, e.g., the strongest correlation, t = 0, between inafter, vector g is derived using chain rule and denoted as

冘
the two arrays refers to the case in which the instances in the N

two sets are identical and the value of Ac㛭CORREL(At, Bt, t) ⭸E(␣m) 1 ⭸Ei
g= = (11)
equals one. In such a case, no solution to a new instance can ⭸␣m 2N i=1 ⭸␣m
be generated via the UFN reasoning model except for when
冘 冉冘 冊
⫺2
⭸Ei
identical instances exist in the instance base. In order to avoid =2 ( yk ⫺ oik ) f (dip)
this issue here, we set Ac㛭CORREL(At, Bt, t) equal to 0.8 as ⭸␣m k p

冉冘 冘
the lower bound for similarity measurement. The value of t
corresponding to this lower bound is adopted as the appropri- ⫻ opk f ⬘(dip)(u im ⫺ u pm)2 f (dip)
ate value of R max. p p

Self-Adjustment Approach for Selecting Weights ␣m

Except for R max, the selected weights ␣m also significantly
⫺ 冘 冘
p
冊
f (dip)opk
p
f ⬘(dip)(u im ⫺ u pm)2
(12)
affect the learning performance for the conventional IFN. This
occurrence has been investigated in the earlier work (Hung where index p denotes the pth similar instance in fuzzy set
and Jan 1999). The learning results indicated that significant Ssup,Ui , and indices m and k denote the mth decision variable
improvements were achieved as the weights were gradually in the input and the kth data item in the output vectors, re-
updated via a basis of heuristic knowledge associated with spectively.
learning problems. In this study, a more systematical approach
—self-adjustment based on mathematical optimization—is Augmented IFN Learning Model
adopted to facilitate the search for appropriate weights. In this section, we present an augmented IFN learning
For the above instance base U with N instances, the self- model. Fig. 3 schematically depicts the procedures of the aug-
adjustment approach can be briefly stated as consisting of the mented IFN learning model. Instead of using a constant work-
following steps. First, set up the corresponding working pa- ing parameter Rmax, as in conventional IFN, the appropriate
rameters, Rmax, Rmin, and wj , where Rmax is determined using parameter is determined using correlation analysis. Mean-
the aforementioned correlation analysis approach and where while, the appropriate weights ␣m are adapted during the self-
parameter Rmin and weights wj are set as constants in this work. adjustment process through a mathematical approach. These
Meanwhile, weights ␣m for each decision variable in the input two approaches, called self-organized learning, are used to en-
are directly initialized as one. Then, based on these working hance the learning capability of the conventional IFN learning
parameters, the outputs for training instances are found via the model. The procedure of the augmented IFN learning model
UFN reasoning model. Then the error, Ei, between the com- can be summarized as follows:
puted and desired outputs, Y and Ui,o, for training instance Ui,
is calculated. The system error, E, for a total N instances is • Self-organized learning phase
then defined as half of the average sum of errors and denoted Step 0. Train the adaptive supervised L-BFGS learning
as neural network model offline.

冘 Step 1. Initialize parameter Rmin as constant 10⫺5 and

N
1
E= Ei (8)
2N i=1

冘
K

Ei = ( yk ⫺ oki )2 (9)
k=1

where oki and yk = kth items of data in desired and computed

outputs, Ui and Y, respectively. Note that the system error is
an implicit function of the weights ␣m as E(␣m).
Weights ␣m in the UFN reasoning model are adjusted to
reduce the system error as much as possible. This goal can be
achieved if a set of appropriate weights, ␣m, are used. The
problem, then, can be considered as an unconstrained opti-
mization problem, that is, searching a set of optimum weights
by iteration to minimize the system error. In the mathematical
optimization approaches, the conjugate gradient (CG) method
has been proved an efficient means of solving the problem.
The weights ␣m are updated in each iteration, say the (s ⫹
1)th iteration, as ␣(s⫹1)
m = ␣m(s) ⫹ ␭d (s). The term ␭ is step length
and is set as a constant in this work. The search direction
d (s) is defined as

再 d (s) = g (s) = ⫺ⵜE(␣m(s))

d (s)
=g (s)
⫹␤ d (s) (s⫺1)
if s = 0
if s > 0
(10a,b)

where
(g (s))Tg (s)
␤(s) = , and g (s) = ⫺ⵜE(␣m(s))
(g (s⫺1))Tg (s⫺1)
FIG. 3. Augmented IFN Learning Model

18 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 weight ␣m for each decision variable, on a heuristic a. If Lb < Lp, then Mn = Mp
basis or by trial and error.
Step 2. Calculate the degree of difference dij (i ≠ j) b. If Lp < Lb < Lr, then Mn
among all instances in base V.
= Cb[Mp ⫺ (Mp ⫺ Mr)((Lb ⫺ Lr /Lr ⫺ Lp))] ⱕ Mp
Step 3. Determine the parameter Rmax using correlation
analysis and set Rmax = t such that Ac㛭CORREL(At, Bt, c. If Lb > Lr, then Mn
t) = 0.8.
Step 4. Set the fuzzy membership function, ␮(dij) =
f (Rmin, Rmax, dij), defined in (2).
Step 5. Adjust the weight ␣m for each decision variable
= Mcr = Cb
␲
Lb 冑冉 冊
␲E
Lb
2

Cw Iy ⫹ EIy GJ

in input, using the self-adjustment approach. where Mp = plastic moment strength; Lp and Lr = limited
• Analysis phase (after learning phase is completed) laterally unbraced length for full plastic bending capacity
Step 6. Present the new (unsolved) instance X to the and for inelastic lateral-torsional buckling, respectively;
UFN reasoning model, and perform a similarity mea- Cb = moment gradient coefficient; Mr = limited buckling
surement between X and instance Ui in the base U using moment; E = elasticity modulus of steel; Cw = warping
a single-layered, lateral-connected competing network. constant; Iy = moment of inertia about the minor axis;
Step 7. If more than one similar instance is found in G = shear modulus of elasticity of steel; and J = torsional
Step 6, generate the solution Xo for the new instance constant.
using fuzzy synthesis approaches in (4) or (5), and go 4. Finally, confirm whether or not the section of selected
to Step 9. Otherwise go to next step. member satisfies the requirements of flexural design
Step 8. Compute the solutions via the undertrained as- strength, 0.9Mn ⱖ Mu, and shear design strength, 0.9Vn
sistant adaptive L-BFGS supervised learning model and ⱖ Vu, where Vn is the nominal shear strength.
go to next step.
Step 9. Feedback the new instance into the base U. The above steps are repeated until a satisfactory lightest mem-
Meanwhile, further learning in the assistant supervised ber is obtained. Eight hundred instances, created according to
learning model is launched offline. the aforementioned design process, are used in the present ex-
ample to train and verify the conventional and augmented IFN
learning models. Of these, 600 are used as training instances
APPLICATIONS and the remaining two hundred (200) are used as verification
instances. Seven decision variables are used as inputs in order
to determine the plastic modulus Z of a lightest W-section steel
Here the novel augmented IFN learning model is applied to beam with compact section. The seven decision variables are
problems of engineering design. Two structural design exam- the yielding strength of steel fy, factored maximum bending
ples are presented to verify the learning performance of the moment Mu, the live load wl, the factored maximum shear
augmented IFN learning model. The first example is a steel force Vu, the span of the beam L, the moment gradient coef-
beam design problem. The second is the preliminary design ficient Cb, and the unbraced length Lb. Notably, the first de-
of steel structure buildings. The two examples are also used cision variable, fy , was not used in the earlier work (Hung and
to train a conventional IFN learning model for comparison. Jan 1999), as the yielding strength of steel was identical for
each instance.
The parameters Rmin and wi are set as constant 10⫺5 and one,
Steel Beam Design Problem respectively. The weights ␣m, however, are initialized as [1, 9,
1, 1, 1, 1, 1] for the seven decision variables. Of these, the
The problem is the design of the lightest W-section simply second weight, ␣2, for decision variable, factored maximum
supported steel beam with compact section, under LRFD Spec- bending moment Mu, is set based on the finding investigated
ification (Manual 1994). The problem was studied in our ear- in the earlier work (Hung and Jan 1999). Using these param-
lier work (Hung and Jan 1999), with those results demonstrat- eters and weights, the augmented IFN learning model is
ing that the conventional IFN learning model is superior to a trained in four different cases with different numbers of train-
stand-alone supervised neural network with L-BFGS learning ing instances in base V. The number of training instances is
algorithm when the number of training instances was large increased from 100 to 400 with an increment of 100. After
enough. This work investigates only the learning performances performing the correlation analysis in four different bases of
of the conventional IFN and the novel augmented IFN learning V, the accumulative correlation coefficients can be computed
models. with respect to any specified value t. With the value of Ac㛭
The LRFD design method, applied to the effect of lateral CORREL(At, Bt, t) equal to 0.8, the values of t(=Rmax) for four
torsional buckling for beam with compact section, is summa- different cases are obtained. Those values are 0.044, 0.034,
rized in the following steps: 0.033, and 0.030, respectively, for 100, 200, 300, and 400
training instances in base V, and are displayed in Fig. 4. In-
terestingly, according to this figure, the greater the number of
1. Determine the factored moment Mu at midspan of a given instances in bases V implies a higher correlation among these
beam. instances. Restated, the value of Rmax decreases with an in-
2. Select one potential member with plastic section modu- crease of the number of training instances. After the value of
lus Z so that the required nominal moment Mn = Zfy sat- Rmax is determined, the self-adjustment approach is launched
isfies 0.9Mn ⱖ Mu, where fy is the specified yield strength to search the appropriate weights ␣m for each decision variable
of the beam. in the input.
3. Estimate whether the lateral supports are close enough After self-adjustment is achieved, a set of weights ␣m is
to design the beam using plastic analysis or whether we obtained, e.g., the weights ␣m are adjusted to values [1.42,
should use the fully plastic moment strength Mp without 6.93, 1.02, 1.01, 1.01, 1.51, 0.73] for the case of V with 400
plastic analysis, for a given unbraced length Lb. The training instances. Note that the first, second, and sixth weights
nominal moment Mn is then obtained according to the (␣1, ␣2, and ␣6) are changed by more than the other weights.
following formulas with different conditions: These adjustments illustrate that the first and sixth decision
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 19
 for the 200 verification instances. According to this table, the
learning performance is improved as the number of training
instances increases for the augmented and conventional IFN
learning models. This same table reveals that the average er-
rors for verification instances are 3.47% and 4.47% for the
augmented and conventional IFN learning models, respec-
tively. In sum, the learning performance of the augmented IFN
learning model is much better than that of the conventional
IFN learning model in this example.
Based on the newly adjusted weights ␣m, correlation anal-
ysis is next performed for the case of 400 instance bases in V
again. According to those results, the accumulative correlation
coefficients are higher for any specified value t than those ob-
tained from the initial set of weights. For example, the accu-
mulative correlation coefficient is from 0.8 to 0.9 at the value
t equal to 0.03 (Fig. 4).
FIG. 4. Accumulative Correlation Curve for Steel Beam De-
sign Example Preliminary Design of Steel Structural Buildings
In the complete design of a structure, the preliminary design
variables, fy and Cb, are more important than we would be led stage is mainly a creative, experiential process that involves
to believe by a process of trial and error. However, the weight the choice of structure type, selection of the material, and de-
␣2 for the second decision variable, Mu, is adjusted from 9 to termination of the sections of beams and columns in the struc-
6.93. The occurrence reveals that the factored moment, Mu, is ture. An experienced engineer is likely to carry out this stage
the most important decision variable in the input, as compared more quickly than does an inexperienced one. The basic con-
with other decision variables. Then, the 200 verification in- figuration of the structure at this stage should satisfy the spec-
stances are used to verify the learning performance of the con- ified design code, such as LRFD for steel structures. To satisfy
ventional and augmented IFN learning models. The verifica- the prescribed constrains and achieve minimum expenditure
tion in the augmented IFN is based on the newly adjusted for materials and construction, this stage becomes a looped
weights ␣m with the appropriate Rmax. The working parameters optimization decision-making process. Hence, a good initial
and weights used in the conventional IFN, however, are se- development of the basic form, with sections of beams and
lected by trial and error. Fig. 5 compares the computed and columns satisfying the aforementioned constraints, will reduce
desired outputs of the 200 verification instances for the aug- the number of redesign cycles. After a basic structure is de-
mented and conventional IFN learning models. The correlation termined, the structural analysis stage involves analyzing the
coefficients for the computed and desired outputs are 0.997 initial guessed structure and computing the service loads on
and 0.992 for the augmented and conventional IFN learning the members. Also, the maximum lateral drift of the structure
model, respectively. Table 1 summarizes the computing results and the drifts between floors are computed if lateral loads are
considered. The present example involves a complete design
structure that satisfies the conditions that the service loads
should not exceed the strength of the members; the drifts
should be within the prescribed limits, and the structure should
be economical in material (e.g., minimum weight), construc-
tion, and overall cost. In this example, the augmented IFN
learning model is trained to learn how to implement the pre-
liminary design of buildings satisfying the conditions of utility,
safety, and economy in only one design cycle. For simplicity,
only regular buildings with a rectangular layout—such as
most factory buildings—are considered herein. Also, the
beams in every floor have the same sectional dimensions, as
do the columns.
In this example, 416 instances are used. They are randomly
divided into 380 training instances and 36 verification in-
stances. Seven decision variables are used as inputs to deter-
mine the sections of beams and columns of a building that
FIG. 5. Comparison of Computed and Desired Outputs for satisfies the given specifications. The seven decision variables
Verification Instances: Steel Beam Design Example and their limits are described as follows:

TABLE 1. Results of Steel Beam Design Problem 1. Number of stories = [9, 10, 11, 12, 13, 14, 15]
2. Bay length in long-span direction (X direction) = 9 to 12
Average Error (Rmax) for 200
m
Verification Instances
3. Bay length in short-span direction (Y direction) = 6 to 9
Number of training Conventional IFN Augmented IFN m
instances in U(V) learning model learning model 4. Number of bays in both directions = [3, 4, 5]
(1) (2) (3) 5. Seismic zone coefficient = [0.18, 0.23, 0.28, 0.33]
100 (100) 9.33% (0.040) 8.31% (0.044) 6. Live load (kgw/m2) = 200 to 350
200 (200) 8.31% (0.040) 5.50% (0.034) 7. Wall load (kgw/m) = [100, 200]
300 (300) 5.57% (0.040) 4.47% (0.033)
400 (400) 4.92% (0.040) 3.82% (0.030)
600 (400) 4.47% (0.040) 3.47% (0.030)
Other corresponding decision variables used in the stage of
preliminary design are assumed to be constant. In practice, the
20 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
sections of beams and columns of a building are classified into
certain groups for convenience in construction, instead of sep-
arate consideration being given to each element. Here, a build-
ing with three groups of steel elements in both beams and
columns is considered. The three groups are upper, medium,
and lower. For a building with N stories, these three groups
are defined as upper group: floors from 2N/3 to N; medium
group: floors from N/3 to 2N/3; and lower group: floors from
one to N/3. An instance contains seven decision variable inputs
and six data items as outputs.
The parameters Rmin and wi are set as constant 10⫺5 and one,
respectively. The weights ␣m, however, are initialized as [1, 1,
1, 1, 1, 1, 1] for the seven decision variables. Using these
parameters and weights, correlation analysis in the augmented
IFN learning model is performed first to determine the work-
ing parameter Rmax. Fig. 6 displays the result of correlation
analysis. With the value of Ac㛭CORREL(At, Bt, t) equal to 0.8, FIG. 7. Comparison of Computed and Desired Outputs
the values of t(=Rmax) are obtained. They are 0.12 for beams (Beams) for Verification Instances: Preliminary Design of Steel
and 0.178 for columns, respectively. Structural Buildings
After the fuzzy membership function is defined, the self-
adjustment approach is launched to obtain the adequate
weights ␣m for each decision variable in the input. The weights
␣m for each decision variable are updated from [1, 1, 1, 1, 1,
1, 1] to [1.471, 1.369, 0.416, 0.008, 1.104, 0.825, 0.513] for
beams and to [1.106, 0.997, 0.542, 0.009, 1.232, 0.999, 0.997]
for columns. Interestingly, the weights of the fourth decision
variable in beams and columns are both self-adjusted close to
zero. This observation indicates that this decision variable
(number of bays in both directions) is insignificant in the input.
Consequently, this decision variable can be neglected. The 36
verification instances are used to verify the learning perfor-
mance of the augmented and conventional IFN learning mod-
els, respectively. Notably, the augmented IFN is verified on
the basis of the newly adjusted weights ␣m with the adequate
Rmax. The working parameters and weights used in the con-
ventional IFN, however, are selected on a trial-and-error basis.
Fig. 7 depicts the correlation between the computed and de- FIG. 8. Comparison of Computed and Desired Outputs (Col-
sired outputs of beams for the 36 verification instances for the umns) for Verification Instances: Preliminary Design of Steel
augmented as well as conventional IFN learning models. Sim- Structural Buildings
ilarly, Fig. 8 displays the correlation between the computed
and desired outputs for columns for the thirty-six verification TABLE 2. Results of Preliminary Structural Design Problem
instances using the two IFN learning models. Table 2 sum-
marizes the learning results for thirty-six verification instances. NUMBER OF ITEMS IN OUTPUT FOR
According to this table, the average percentage errors for VERIFICATION INSTANCES
beams and columns are 13.81 and 9.36 for the conventional Conventional IFN Augmented IFN
IFN learning model. However, these errors are reduced to 6.17 Error Learning Model Learning Model
and 6.1 for beams and columns, respectively, for the aug- (%) Beam Column Beam Column
mented IFN learning model. The augmented IFN learning (1) (2) (3) (4) (5)
model significantly improves in terms of learning. This ex-
0⬃5 12 43 52 51
5 ⬃ 10 29 29 39 31
10 ⬃ 15 21 22 13 20
15 ⬃ 20 29 10 3 6
>20 17 4 1 0
Average error for 36 ver-
ification instances 13.81% 9.36% 6.10% 6.17%
Note: Each instance has six items, three beams, and columns, in out-
put.

ample also illustrates that the augmented IFN learning model

yields a substantially better learning performance than that of
the conventional IFN learning model.

CONCLUDING REMARKS
This work presents an augmented IFN learning model by
integrating two newly developed approaches into a conven-
FIG. 6. Accumulative Correlation Curves for Example of Pre- tional IFN learning model. These approached are a correlation
liminary Design of Steel Structural Buildings analysis in statistics and self-adjustment in mathematical op-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 21
timization, which collaboratively enhance the learning capa- model for image recognition.’’ Integrated Computer-Aided Engrg.,
bility of the conventional IFN. The first approach, correlation 1(1), 43–55.
Adeli, H., and Hung, S. L. (1995). Machine learning—neural networks,
analysis in statistics, assists users in determining the appro- genetic algorithms, and fuzzy systems. Wiley, New York.
priate working parameter used in the fuzzy membership func- Adeli, H., and Park, H. S. (1995). ‘‘Counterpropagation neural network
tion. The second approach, self-adjustment in mathematical in structural engineering.’’ J. Struct. Engrg., ASCE, 121(8), 1205–
optimization, obtains appropriate weights, systematically, for 1212.
each decision variable in the input of training instance. The Carbonell, J. G. (1981). ‘‘A computational model of analogical problem
augmented IFN learning model proposed herein is applied to solving.’’ Proc., 7th Int. Joint Conf. on Artificial Intelligence, William
Kaufmann, Los Altos, Calif., 1, 147–152.
engineering design problems. Two structural design problems Carpenter, G. A., and Grossberg, S. (1988). ‘‘The ART of adaptive pattern
were addressed to assess the learning performance of the aug- recognition by a self-organizing neural network.’’ IEEE Comp., 21,
mented IFN learning model. Based on the results of this work, 77–88.
we can conclude the following: Elkordy, M. F., Cheng, K. C., and Lee, G. C. (1994). ‘‘A structural dam-
age neural network monitoring system.’’ Microcomp. in Civ. Engrg.,
9(2), 83–96.
1. The problem of arbitrary trial-and-error selection of the Ghaboussi, J., Garrett, J. H., and Wu, X. (1991). ‘‘Knowledge-based mod-
working parameter (Rmax) in fuzzy membership function, eling of material behavior and neural networks.’’ J. Engrg. Mech.,
encountered in the conventional IFN learning model, is ASCE, 117(1), 132–153.
avoided in the newly developed augmented IFN learning Gunaratnam, D. J., and Gero, J. S. (1994). ‘‘Effect of representation on
model. Instead of arbitrary Rmax, the appropriate value of the performance of neural networks in structural engineering applica-
Rmax is determined using correlation analysis in statistics. tions.’’ Microcomp. in Civ. Engrg., 9(2), 97–108.
Hajela, P., and Berke, L. (1991). ‘‘Neurobiological computational modes
Thus, the new learning model provides a more solid sys- in structural analysis and design.’’ Comp. and Struct., 41(4), 657–667.
temic foundation for IFN learning than the conventional Hung, S. L., and Adeli, H. (1991a). ‘‘A model of perceptron learning with
IFN learning model. a hidden layer for engineering design.’’ Neurocomputing, 3(1), 3–14.
2. In the conventional IFN learning model, the weights ␣m, Hung, S. L., and Adeli, H. (1991b). ‘‘A hybrid learning algorithm for
denoting the importance of the mth decision variable in distributed memory multicomputers.’’ Heuristics—J. of Knowledge
Engrg., 4(4), 58–68.
the input, are set on a trial-and-error basis. For compli- Hung, S. L., and Adeli, H. (1992). ‘‘Parallel backpropagation learning
cated problems, the appropriate weights are difficult to algorithms on Cray Y-MP8/864 supercomputer.’’ Neurocomputing,
obtain because of a lack of the relevant heuristic knowl- 5(6), 287–307.
edge. In due course, the value is commonly initialized Hung, S. L., and Adeli, H. (1994a). ‘‘An adaptive conjugate gradient
as one for most of the examples. This problem is avoided learning algorithm for effective training of multilayer neural net-
by the newly developed learning model. Instead of an works.’’ Appl. Math. and Computat., 62(1).
Hung, S. L., and Adeli, H. (1994b). ‘‘A parallel genetic/neural network
assumed constant, the appropriate weights are deter- learning algorithm for MIMD shared memory machines.’’ IEEE Trans.
mined through the self-adjustment approach in mathe- on Neural Networks, 5(6), 900–909.
matical optimization. Therefore, the augmented IFN Hung, S. L., and Lin, Y. L. (1994). ‘‘Application of an L-BFGS neural
learning model provides a more solid mathematical foun- network learning algorithm in engineering analysis and design.’’ Proc.,
dation for neural network learning. 2nd Nat. Conf. on Struct. Engrg., Vol. III, Chinese Society of Structural
Engineering, Taipei, Taiwan, 221–230 (in Chinese).
3. For each training instance, decision variables in the input
Hung, S. L., and Jan, J. C. (1997). ‘‘Machine learning in engineering
are generally selected subjectively by users. As a result, design—an unsupervised fuzzy neural network learning model.’’ Pro-
some trivial decision variables may be adopted in the ceedings of Intelligent Information Systems (IIS’97), IEEE Computer
input for some complicated examples. Based on the self- Society, California, 156–160.
adjustment approach, the importance of a decision vari- Hung, S. L., and Jan, J. C. (1999). ‘‘Machine learning in engineering
able in an input can be derived systematically. Therefore, analysis and design—an integrated fuzzy neural network learning
model.’’ Computer-Aided Civ. and Infrastruct. Engrg., 14(3), 207–219.
insignificant or redundant decision variables in the input Kang, H.-T., and Yoon, C. J. (1994). ‘‘Neural network approaches to aid
can be neglected. simple truss design problems.’’ Microcomp. in Civ. Engrg., 9(3), 211–
4. The results illustrate that the value of the appropriate Rmax 218.
gradually falls with an increase in the number of in- Kishi, N., Chen, W. F., and Goto, Y. (1997). ‘‘Effective length factor of
stances. Consequently, not only is the learning perfor- columns in semirigid and unbraced frames.’’ J. Struct. Engrg., ASCE,
mance for training instances enhanced, but also the per- 123(3), 313–320.
Kitipornchai, S., Wang, C. M., and Trahair, N. S. (1986). ‘‘Buckling of
formance for verification instances is improved. Notably, monosymmetric I-beams under moment gradient.’’ J. Struct. Engrg.,
a small value of Rmax indicates that a strict similarity ASCE, 112(4), 781–799.
measurement is utilized in the UFN reasoning model and Kruse, R., Gebhardt, J., and Klawonn, F. (1994). Foundations of fuzzy
the possibility that no similar instances can be derived systems. Wiley, New York.
for any new instance also increases. Maher, M. L., Balachandran, M. B., and Zhang, D. M. (1995). Case-
based reasoning in design. Lawrence Erlbaum Assoc., Mahwah, N.J.
Manual of steel construction—load and resistance factor design. (1994).
ACKNOWLEDGMENTS American Institute of Steel Construction, Inc., Chicago, Ill.
Rumelhart, D., Hinton, G., and Williams, R. (1986). ‘‘Learning represen-
The writers would like to thank the National Science Council of the tations by back-propagation errors.’’ Parallel distributed processing,
Republic of China for financially supporting this research under Contract Vol. 1, D. Rumelhart et al., eds., MIT Press, Cambridge, Mass., 318–
No. NSC 86-2221-E-009-070. 362.
Stephen, J. E., and Vanluchene, R. D. (1994). ‘‘Integrated assessment of
APPENDIX. REFERENCES seismic damage in structures.’’ Microcomp. in Civ. Engrg., 9(2), 119–
128.
Adeli, H., and Hung, S. L. (1993a). ‘‘A concurrent adaptive conjugate Vanluchene, R. D., and Sun, R. (1990). ‘‘Neural networks in structural
gradient learning on MIMD machines.’’ J. Supercomputer Appl., 7(2), engineering.’’ Microcomp. in Civ. Engrg., 5(3), 207–215.
155–166. Zadeh, L. A. (1978). ‘‘Fuzzy set as a basics for a theory of possibility.’’
Adeli, H., and Hung, S. L. (1993b). ‘‘A fuzzy neural network learning Fuzzy Sets and Sys., 1(1), 3–28.

22 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 COMPARISON OF NEURAL AND CONVENTIONAL APPROACHES TO
MODE CHOICE ANALYSIS
By Tarek Sayed1 and Abdolmehdi Razavi2

ABSTRACT: This paper describes a new approach to behavioral mode choice modeling using neurofuzzy mod-
els. The new approach combines the learning ability of artificial neural networks and the transparent nature of
fuzzy logic. The approach is found to be highly adaptive and efficient in investigating nonlinear relationships
among different variables. In addition, the approach only selects the variables that significantly influence the
mode choice and displays the stored knowledge in terms of fuzzy linguistic rules. This allows the modal decision-
making process to be examined and understood in great detail. The neurofuzzy model is tested on the U.S.
freight transport market using information on individual shipper and individual shipments. Shipments are dis-
aggregated at the five-digit Standard Transportation Commodity Code level. Results obtained from this exercise
are compared with similar results obtained from the conventional logit mode choice model and the standard
back-propagation artificial neural network. The advantages of using the neurofuzzy approach are described.

INTRODUCTION to identify and exclude those input variables that do not have
Mode choice modeling is probably the most important ele- a significant contribution. This would lead to a more efficient
ment of transportation planning. It affects the general effi- model. This is particularly important in cases where a large
ciency of travel and the allocation of resources (Ortuzar and number of potential input variables exist but only a subset of
Willumsen 1994). The development of mode choice models them would actually affect the output. Conventional tech-
has recently witnessed significant advances in many fields, in- niques of variable selection such as stepwise modeling (Mas-
cluding passenger and freight transport. A large number of ters 1993) may not produce optimal results.
mathematical models have been used to model the traveler’s It is therefore desired to develop efficient models that are
choice of mode and destination and the shipper’s choice of able to correctly simulate the input/output relationship using a
mode, shipment size, and supply market, among others. Such minimal set of input variables while providing insight into the
models are not only becoming almost intractable but also data decision-making process. The models should also adopt ap-
intensive, difficult to calibrate and update, and intransferable proaches that use a few constraining assumptions as possible
(Abdelwahab and Sayed 1998). These models cover a wide so that they can be successfully used for a wide variety of
range of mathematical complexity and accuracy. As is often applications. It is also beneficial to have the ability to insert
the case, the models that are mathematically manageable, due any available knowledge or expertise into the model if nec-
to the inherent constraints, compromise on their performance. essary.
Linearity and the use of certain distributions are among lim- In this paper, the merits of adopting the neurofuzzy ap-
iting assumptions that handicap some of these conventional proach will be discussed. It will be shown that the proposed
approaches. approach has the potential to provide similar or better accuracy
In a recent paper, Abdelwahab and Sayed (1999) showed compared with ANNs while overcoming their shortcomings.
that artificial neural networks (ANNs) could be successfully In the present study, a bimodal example is used to illustrate
used for mode choice modeling. The results obtained from a the proposed approach, but the extension to multimodal cases
standard back-propagation neural network were compared or similar applications is straightforward.
with similar results from the two most widely used conven-
CONVENTIONAL MODE CHOICE MODELS
tional models, namely, logit and probit functions. ANNs pro-
duced equal or higher predictive accuracy of mode choice As mentioned previously, the two most widely used con-
compared with conventional models. In addition, ANNs were ventional mode choice models are the logit and probit models.
believed to be less data intensive and less time consuming to They both belong to the family of behavioral discrete choice
develop and apply than currently used conventional models. models [see Ben-Akiva and Lerman (1987)]. A full discussion
However, despite the flexibility and the good performance of these models is not within the scope of this study. There-
of ANNs in modeling nonlinear relationships, they have often fore, only relevant aspects of the models will be highlighted.
been criticized for acting as ‘‘black boxes.’’ The knowledge The logit model is perhaps the most widely used mode
contained in the ANN model is kept in the form of a weight choice model, particularly in multimodal contexts (i.e., when
matrix that is hard to interpret and can be misleading at times. choice is made among more than two competing modes). The
The efficiency of ANN models is also another point of con- probability of choosing a mode is usually based on a linear
cern. Because it is not always possible to determine the sig- combination function known as the utility expression ex-
nificance of the input variables in advance, any potential can- pressed as
didate may be included in the model. It is therefore important
Uik = ␣k ⫹ ␤1Xik1 ⫹ ␤2Xik2 ⫹ ⭈ ⭈ ⭈ ⫹ ␤j Xik j ⫹ εi = Vik ⫹ εi (1)
1
Asst. Prof., Dept. of Civ. Engrg., Univ. of British Columbia, 2324
Main Mall, Vancouver, BC, Canada V6T 1Z4. E-mail: tsayed@civil.
where Uik = utility of individual i choosing mode k; Vik =
ubc.ca deterministic utility of individual i choosing mode k; ␣k = con-
2
Res. Assoc., Dept. of Civ. Engrg., Univ. of British Columbia, 2324 stant for mode k; X1, X2, . . . , Xj = explanatory variables (can
Main Mall, Vancouver, BC, Canada V6T 1Z4. be generic or mode-specific); ␤1, ␤2, . . . , ␤j = corresponding
Note. Associate Editor: Hassan A. Karimi. Discussion open until June calibration coefficients; and εi = random error term. It is as-
1, 2000. To extend the closing date one month, a written request must sumed that each individual chooses the alternative mode that
be filed with the ASCE Manager of Journals. The manuscript for this
paper was submitted for review and possible publication on January 19,
‘‘appears’’ to have the highest utility value. To predict an in-
1999. This paper is part of the Journal of Computing in Civil Engi- dividual’s choice of mode, the utility function should be trans-
neering, Vol. 14, No. 1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/ formed into a probability. The logit model transforms the util-
0001-0023–0030/$8.00 ⫹ $.50 per page. Paper No. 20068. ity function into probability in the following manner:
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 23
 Pk(i) = (1 ⫹ eVi j⫺Vik)⫺1 (2) is converted to an output according to a transfer function (typ-
ically a sigmoid function).
where Pk(i) = probability that individual i chooses mode k. An There is a wide range of architectures for ANNs among
implicit assumption in the logit transformation function is that which the most widely used architecture, three-layer feedfor-
all error terms are assumed to be independent and identically ward was adopted by Abdelwahab and Sayed (1998) for mode
distributed with a Weibull distribution. This assumption leads choice modeling. In this type of network, there are two dis-
to one of the most fundamental weaknesses of the logit model, tinctive modes, namely, training and query modes. In the train-
the independence of irrelevant alternative property (McFadden ing mode, a training set consisting of input/output patterns is
1978). However, this weakness is more common in multino- presented to the networks, and the weights are found through
mial logit formulations, where similar modes share attributes. an iterative process. The back-propagation learning algorithm
The coefficients of the utility expression are estimated by the (Rumelhart and McClelland 1986) is often used to find the
maximum likelihood method that obtains these coefficients weights such that the difference between the expected outputs
through an iterative search process throughout the data set. and the ones calculated by the network is sufficiently small.
Another technique of transforming the linear utility function After the network is trained, it is presented with the query data
into a bounded probabilistic relationship is described by Fin- (input variables only) to determine how accurately the network
ney (1971) and Daganzo (1979). The probability unit (probit, has simulated the input/output relationship,
in short) is based upon the relationship between the standard
normal distribution and its associated cumulative distribution, NEUROFUZZY APPROACH
which displays a sigmoid curve representing the relationship Fuzzy Logic and Fuzzy Sets
between the utility of choosing a mode and its attributes. In
this case, however, the assumption that the error terms are Fuzzy logic is another area of artificial intelligence that has
independent and identically distributed is relaxed. The probit been successfully applied to an increasing number of appli-
model is considered to be the most versatile discrete choice cations. The concept of fuzzy logic was introduced by Zadeh
model available (Hausman and Wise 1978; Daganzo 1979). Its (1965). Basically, it is a superset of conventional (Boolean)
relatively limited application, compared with the logit model, logic that has been extended to handle imprecise data and the
has been attributed to the computation difficulty involved in concept of partial truth. In fuzzy logic, variables are ‘‘fuzzi-
the estimation of its multivariate normal function. However, fied’’ through use of so-called membership functions that de-
in the binary case, the computation difficulty is significantly fine the membership degree to fuzzy sets, or as they are often
reduced to a level that makes it comparable to the logit for- called ‘‘linguistic variables.’’ A fuzzy subset A of a universe
mulation. A binary probit model can be stated as of discourse U is characterized by a membership function
uA(x), which associates each element x 僆 U a membership
Pk(i) = ⌽((Vik ⫺ Vi j)/␴) (3) uA(x) in the interval [0, 1] that represents the grade of mem-
bership in A.
where ⌽(⭈) = standardized cumulative normal distribution To each variable, a small number of fuzzy sets (linguistic
function; and ␴ = standard deviation of (εk ⫺ εj), which is the variables) are assigned whose membership functions overlap
scale of the utility function that can be set to an arbitrary and cover the necessary range of variation for that variable.
positive value, such as ␴ = 1. Consequently, statements that contain linguistic variables (e.g.,
transit fare is inexpensive) may take ‘‘truth values’’ of other
ANNs than ‘‘true’’ or ‘‘false,’’ represented by a number in the range
ANNs are a subfield of the artificial intelligence technology [0,1]. This provides an opportunity to define logical operators
such as ‘‘AND’’ or ‘‘OR’’.
that has gained strong popularity in a rather large array of
One way of representing fuzzy sets is by using B-spline
engineering applications where conventional analytical meth-
functions. B-spline functions (Fig. 1) are piecewise polyno-
ods are difficult to pursue or show inferior performance. Spe-
mials of order k that have been widely used in surface fitting
cifically, ANNs have shown a good potential to successfully
applications. They can be used to represent fuzzy membership
model complex input/output relationships where the presence functions because they have several important properties
of nonlinearity and inconsistent/noisy data adversely affects (Brown and Harris 1994):
other approaches.
An ANN consists of a large number of interconnected com- • The degree of membership can be evaluated using a sim-
putational elements called ‘‘neurons’’ that are organized in a ple, fast, and stable recurrence relationship.
number of layers. The connection between each pair of neu- • The output of the function is nonzero in only a small part
rons is called a link and is associated with a ‘‘weight’’ that is of the input space, which means that knowledge is stored
a numerical estimate of the connection strength. Each neuron locally across only a small number of basis functions.
in a layer receives and processes weighted inputs from neurons • The basis functions form a partition of unity as 兺 pi=1
in the previous layer and transmits its output to neurons in the uAi(x) ⬅ 1, thus producing accurate smooth approxima-
following layer. The weighted summation of inputs to a neuron tion.

FIG. 1. B-Spline Fuzzy Membership Functions

24 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 The order of the B-spline functions determines their Therefore, it can be trained to perform an input/output map-
smoothness. They can be used to implement crisp fuzzy sets ping, just as with an ANN, but with the additional benefit of
(k = 1), the standard triangular fuzzy membership functions (k being able to provide the set of rules on which the model is
= 2), or other smoother representations. A univariate B-spline based. This gives further insight into the process being mod-
function of order k is nonzero only over k intervals that are eled.
generated by a (k ⫹ 1) knots. A multivariate B-spline function Several merger types of ANNs and fuzzy systems have been
can be formed by taking the tensor product of n univariate reported in the literature. They include various representations
functions (Brown and Harris 1994). and architectures and therefore are suitable for different ap-
plications. Methods proposed by Kosko (1992) and Jang
Fuzzy Systems (1993) are among many variations that combine neural net-
works and fuzzy system. Although the former uses competitive
Fuzzy systems use fuzzy implications or ‘‘IF THEN rules’’ networks to generate rules for fuzzy systems, the latter pro-
to process information. One such rule may look like this: poses a hybrid back-propagation/least-square learning method
to tune the parameters of a so-called adaptive network-based
IF ‘‘Transit Fare’’ is INEXPENSIVE and ‘‘Transit Waiting fuzzy inference system.
Period’’ is Short However, many of these methods suffer from what is often
THEN ‘‘Preferred Mode of Transportation’’ is Transit referred to as the ‘‘curse of dimensionality’’ (Brown and Har-
ris 1994), which makes them limited to problems involving a
The part of the rule between the ‘‘if’’ and ‘‘then’’ is the small number of variables. To illustrate this ‘‘curse,’’ one may
rule’s premise or antecedent. This is a fuzzy logic expression consider a fuzzy system with N input variables, each having
that describes to what degree the rule is applicable. The part M membership functions. In such a system, as many as MN
of the rule following the ‘‘then’’ is the rule’s conclusion or combinations (or potential fuzzy rules) would exist. This ex-
consequent. The consequent part of the rules may result in a ponential growth of fuzzy rules with number of inputs makes
fuzzy or crisp variable. In cases where the consequent is in it impractical to use most existing architectures for problems
fuzzy form but a crisp number is desired, an inverse operation of high dimensionality (such as the problem in this study).
called ‘‘defuzzification’’ needs to be done. Fuzzy systems are The neurofuzzy approach adopted in this study is based on
defined by a number of fuzzy rules, a number of membership the well-developed adaptive B-spline networks, which exhibit
functions, and mechanisms to apply logical operators. There many desirable properties for nonlinear function approxima-
are numerous successful applications of fuzzy systems in con- tion. In addition, it does not suffer from the curse of dimen-
trol and modeling. They are most suitable for situations where sionality (as will be described later).
an exact model of a process is either impractical or very costly
to build but an imprecise model based on the existing human Adaptive B-Spline Networks
expertise can do the job. In such cases, fuzzy systems are
The input space of a B-spline network (Fig. 2) is covered
considered the best alternative, though they do not perform
by a set of multidimensional overlapping B-splines. The
optimally.
weighted sum of these basis functions produces a nonlinear
adjustable input/output mapping. The network output can be
Neurofuzzy Models represented by the following equation:
The knowledge contained in fuzzy systems are transparent
冘
p

to the user but cannot be acquired directly from data. ANNs, y= aiwi = a(x)Tw (4)
on the other hand, have the ability to learn the knowledge from i=1

a set of data, but the knowledge gained is hidden from the where y = output; a(x) = (a1(x), . . . , ap(x)) = vector of basis
user. The concept of neurofuzzy systems has emerged in recent function outputs when presented by input x = (x1, . . . , xn); and
years as researchers have tried to combine the transparent, lin- w = (w1, . . . , wp) = vector of network weights. The power
guistic representation of a fuzzy system with the learning abil- behind this type of network comes from the direct equivalence
ity of an ANN (Brown and Harris 1994). A neurofuzzy system between the union of a set of fuzzy rules in the form
uses an ANN learning algorithm to determine its parameters
(i.e., fuzzy sets and fuzzy rules) by processing data samples. IF (x is Ai) THEN ( y is Bj)cij (5)

FIG. 2. Typical B-Spline Neurofuzzy Network (Brown and Harris 1994)

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 25

and the weighted sum of the multidimensional fuzzy input mem- tion to calculate parameters for a spline representation of an
bership functions given in (1), where (x is Ai) and ( y is Bj) rep- n-dimensional function. In its most general form, the B-splines
resent linguistic expressions for the input and output respectively; could be of any dimension. For the above example, the first
and ci j is the rule confidence that relates the ith fuzzy input set five splines will be 1D, and the rest of them will be 2D. The
to the jth fuzzy output set. This means that a weight can be fuz- membership functions for the first variable are equivalent to
zified to produce a rule confidence vector ci which can then be the 1D splines, but the membership functions for the second
defuzzified to produce the original weight. This relationship al- and third variables are equivalent to the 1D splines whose
lows the network to be trained in a weight space leading to con- tensor product has produced the aforementioned 2D splines.
siderable reduction in computational cost (the network’s output is In the ASMOD algorithm, for any model structure (i.e., spe-
cific combination of subnetworks, the number and location of
linearly dependent on the weight set), while explaining the output
splines), one can use the training data to calculate the mean
with linguistic rules and the associated rule confidences.
square error of the output. The algorithm starts from the sim-
plest structure (e.g., only first variable in one subnetwork with
Curse of Dimensionality two triangular splines) and iteratively refines its structure until
As described earlier, the number of potential fuzzy rules in some stop criteria is satisfied. In each step among a number
a neurofuzzy system is an exponential function of the dimen- of potential (single) changes to the structure, the one with the
sion of the input space. In reality, many of these rules would best performance is selected and the process continues. Ad-
be redundant for modeling purposes, and therefore a suitable dition of a new input, combining an existing input to a sub-
technique should start from a simple architecture and build on network, splitting a subnetwork, and deleting an input are all
it as necessary. possible changes to the structure. A more detailed explanation
One such approach uses analysis of variance (ANOVA) of the algorithm can be found in (Kavli 1993; Brown and
(Brown and Harris 1995) for decomposition of the output Harris 1994).
function of dimension n as; Adding splines in the middle of existing ones, deleting
splines, and changing the order of splines are also some other

冘 冘冘
n n⫺1 n
changes that are considered by the algorithm. Some measure
f (x) = f0 ⫹ fi (xi) ⫹ fi, j (xi, xj) ⫹ ⭈ ⭈ ⭈ ⫹ f1,2,...,n(x) (6) of statistical significance is used as stopping criteria. Among
i=1 i j=i⫹1
many such measures, Bossley et al. (1995) stated that, for
in which f0 represents a constant (the function bias); and other noisy data, the Bayesian statistical significance measure ap-
terms represent univariate, bivariate, and other subfunctions, pears to perform well and therefore is used in this study. The
respectively. In many cases, a majority of the terms, particu- Bayesian statistical significance measure is given by (Brown
larly those of higher dimensions, are zero or negligible. For and Harris 1994)
such cases, the ANOVA representation allows the input/output K = L ln(J ) ⫹ p ln(L) (7)
mapping to be approximated using a limited number of, as
they are often called, ‘‘subnetworks’’ of much lower dimen- where K = performance measure; p = size of current model;
sions. An example of this decomposition is shown in Fig. 3, J = mean square error; and L = number of data pairs used to
where a 5D function is decomposed into a 1D and two 2D train the network.
subnetworks. The ASMOD algorithm has been applied to a wide variety
A fuzzy rule within each subnetwork may have the form: of modeling problems and has been compared with other neu-
rofuzzy methods. The algorithm appears to perform at least as
IF (x2 is large AND x3 is small) THEN (y is small) well as the best of the other methods (Brown and Harris 1994).
Specifically, Brown and Harris (1994) reported the following
with confidence c
desirable properties:
where c = rule confidence. A rule confidence of 0 indicates
that the rule is not contributing to the output, whereas a rule • It has a smaller number of parameters and a reduced im-
confidence of 1 indicates that the rule is completely true. Val- plementation cost.
ues between 0 and 1 allow the rules to partially fire. The num- • A higher ability to generalize exists.
ber of fuzzy rules used in each subnetwork depends on the • The rule base is more transparent because it is smaller.
number of membership functions that are used to fuzzify the
inputs of that subnetwork. In the above example, assuming DATA
that five membership functions are used for each variable, the To study the feasibility and advantages of using the neu-
first, second, and third subnetworks consist of 5, 25, and 25 rofuzzy approach, a bimodal example for shipment of com-
fuzzy rules, respectively. The consequent part of all the rules modities was selected. Data have been extracted from the 1977
will be ‘‘ORed’’ (i.e., summed in this case) together. Commodity Transportation Survey prepared by the U.S. Bu-
ASMOD, or adaptive spline modeling of observation data reau of Census in 1981. The survey has the most disaggregate
(Kavli 1993), is an algorithm that uses the above decomposi- data publicly available on intercity freight movement, in which
each record represents an individual shipment. Data have been
selected randomly in a multistage sampling process. Among
the possible modes of shipment for which data were available,
only rail and Interstate Commerce Commission-regulated mo-
tor carriers for shipments over 500 lb were selected.
The database used for this study consisted of 7,500 records.
Each record holds information on the chosen mode along with
27 attributes. These attributes and their definitions are listed
in Table 1. They represent shipment attributes (e.g., size), mo-
dal attributes (e.g., transit time for each mode), or market at-
tributes (e.g., geographic location). These attributes are used
FIG. 3. Additive ANOVA Decomposition of Neurofuzzy Rule as input or explanatory variables to model the shippers’ de-
Base (Brown and Harris 1995) cision-making process.
26 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
 TABLE 1. Variable Codes and Definitions Conventional Mode Choice Models
Variable The coefficients for the logit model were first examined us-
code Definition ing all 27 variables. Second, a logit model was estimated using
(1) (2) the same training data set but after limiting the number of
mode Mode of transport (1 = rail, 2 = truck) independent variables to as few as possible without compro-
ss Shipment size (lb) mising the predictive ability of the model. The process elim-
numrec Number of records (shipments) on given OD pair inates variables that are not contributing to the explanatory
ton Total tonnage on OD pair (tons)
vl Value (1977 U.S. $/lb) power of the model (as evident from their low t-statistic) and
dn Density (lb/cu ft) variables that were estimated with signs contrary to a priori
sld Shipment is in solid state (1 = solid; 0 otherwise) expectations given the competitive nature of the shipping in-
prt Shipment is in particulate state (1 = particulate; 0 other- dustry. The resulting model, which is called the ‘‘most effi-
wise) cient’’ conventional model, included 12 variables and is re-
lq Shipment is in liquid state (1 = liquid; 0 otherwise) ported in Table 2. It should be noted that the magnitude of the
gas Shipment is in gaseous state (1 = gaseous; 0 otherwise)
frz Shipment requires freezing (1 = yes; 0 = no) coefficients depends on the scale of the respective variable and
shk Shipment requires stock protection (1 = yes; 0 = no) therefore is not useful in analyzing the model.
tmp Shipment requires temperature control (1 = yes; 0 = no)
oth Shipment requires other protection (1 = yes; 0 = no) ANN Mode Choice Models
lf Shelf life (days); 99 means nonperishable
d1 Regional dummy variable (1 = region 1; 0 otherwise) The ANN models were developed using three-layer feed-
d2 Regional dummy variable (1 = region 2; 0 otherwise) forward networks. The number of nodes in the first layer is
d3 Regional dummy variable (1 = region 3; 0 otherwise) equal to the number of input variables (i.e., 27 for the first
d4 Regional dummy variable (1 = region 4; 0 otherwise)
dist Shipping distance in miles
network). The input variables were scaled to be between 0 and
tcost Truck freight charges (cents/cwta) 1. The number of output nodes should be selected equal to the
rrel Reliability of transit time by rail (% on time) number of modes present. However, an exception is the bi-
rcost Rail freight charges (cents/cwt) modal case where the use of one output node is also an option.
rtime Rail transit time (days) The output variable was coded 1 for truck shipments and 0
rld Loss and damage by rail (%) for rail shipments. This reduces the number of weights in-
trel Reliability of transit time by truck (% on time)
tld Loss and damage by truck (%)
volved. After a number of trials, the number of neurons in the
ttime Truck transit time (days) hidden layer was selected as five. A midrange threshold (i.e.,
a 0.5) was used to predict the mode choice for the testing data
cwt = hundredweight.
set. The results of this model are presented in Table 3.
To build the most efficient ANN model, a method proposed
by Abdelwahab and Sayed (1998) was used. They proposed
METHODOLOGY
to examine the importance of each input variable by looking
To compare the three approaches discussed earlier (conven- at the weights between the input nodes and hidden layer. An
tional, ANN, and neurofuzzy), the database was divided into
two data sets. The first data set, called ‘‘training data set,’’ TABLE 2. Results of Most Efficient Logit Mode Choice Model
was used to build a model by each of the three methods. This Logit
data set had 5,000 records that were selected randomly from
Independent Estimated
the original set. The remaining 2,500 records were then used
variable coefficient t-statistic
as ‘‘testing data set’’ to compare the predictive ability of the (1) (2) (3)
developed models. To avoid any empirical bias, the rail-truck
modal shares (i.e., ⬃37 and ⬃63%) of both data sets were const 0.78034 1.85399a
ss ⫺6.91E-05 ⫺21.2102b
kept similar to the modal shares of the original data. Because dn 2.25E-03 4.34128b
the statistical performance of the logit and probit models was vl 7.78E-03 1.6482a
found to be similar (Ghareib 1996; Abdelwahab and Sayed gas 1.14693 2.60439b
1998), only the logit model will be used in this study to rep- frz 1.35583 4.21887b
resent the conventional approach. The following models were Shk ⫺0.98164 ⫺6.52961b
developed and tested: dl 0.3356 6.05226b
lf ⫺3.74E-03 ⫺2.9702b
rcost ⫺7.36E-03 ⫺8.44882b
• The most efficient logit model using a selected set of var- tcost ⫺2.28E-03 ⫺2.89204b
iables rtime 1.07652 8.80011b
ttime ⫺1.20075 ⫺5.48263b
• An ANN model using all 27 variables
• The most efficient ANN model using a selected set of Percent correctly predicted 86.3 —
variables a
Significant at 90% level.
b
• The neurofuzzy model Significant at 95% level.

The Software Neuframe (Neuframe 1997) was used to build TABLE 3. Prediction Results of ANN Model (Using All 27 Input
the neurofuzzy model described in this study. Neuframe is an Variables)
integrated group of intelligence technology tools that include Predicted by ANN
neural networks and neurofuzzy models.
Percent
Observed Rail Truck Total correct
RESULTS (1) (2) (3) (4) (5)
In this section, the results obtained by each of the four mod- Rail 650 255 905 71.8
Truck 23 1,572 1,595 98.6
els are presented. The discussion of the results will then follow Total 673 1,827 2,500 88.9
in the next section.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 27
input variable whose connection weights are all low is prob-
ably not contributing to the network output. A new data set
excluding this variable can then be constructed and tested. If
the network performance is adequate, then this variable can be
excluded. The one by one reduction of the number of outputs
is continued until any further reduction would result in a sig-
nificant drop in the performance.
Several runs were made on the training data set, at each run
one or more variables whose weights are all low were ex-
cluded and the performance compared. After several runs, a
network consisting of 11 input variables was found to give an
overall success rate of 88.6%, which is similar to the network
constructed in the previous run (with all 27 variables). Any
attempts to further reduce the number of variables resulted in
a relatively significant drop in the success rate. The variables
used and the results of the network are given in Tables 4 and
FIG. 4. Neurofuzzy Model
5. A comparison of Tables 3 and 5 reveals that the exclusion
of 16 explanatory variables had almost no effect on the overall
TABLE 6. Fuzzy Rules Used by Neurofuzzy Model
success rate of the model. It is also interesting to note that
only 8 variables are common between Tables 2 and 4. Subnetwork Rule
number number Rule
Neurofuzzy Mode Choice Model (1) (2) (3)
1 1 IF ‘‘Shipment Size’’ is Small AND ‘‘Other Protec-
The training data set was used in a similar way to build and tion’’ is Not Needed
train the neurofuzzy model. A schematic view of the model is THEN ‘‘Selected Mode’’ is Truck (1.00)
2 IF ‘‘Shipment Size’’ is Medium AND ‘‘Other Pro-
shown in Fig. 4. As shown in the figure, only 4 out of the 27 tection’’ is Not Needed
variables were used by the neurofuzzy model. The Bayesian THEN ‘‘Selected Mode’’ is Rail (0.87) OR ‘‘Se-
statistical significance measure was used as the stopping cri- lected Mode’’ is Truck (0.13)
3 IF ‘‘Shipment Size’’ is Large AND ‘‘Other Pro-
teria. Consequently, a very compact model was obtained that tection’’ is Not Needed
only requires 10 fuzzy rules to reach its decision. The first THEN ‘‘Selected Mode’’ is Rail (0.81) OR ‘‘Se-
subnetwork included 6 rules while the other two subnetworks lected Mode’’ is Truck (0.19)
4 IF ‘‘Shipment Size’’ is Small AND ‘‘Other Protec-
included 2 rules each. The 10 rules are listed in Table 6. It is tion’’ is Needed
also interesting to note that 1 of the 4 variables, reliability of THEN ‘‘Selected Mode’’ is Rail (0.63) OR ‘‘Se-
transit time by truck (trel), was not included in any of the lected Mode’’ is Truck (0.37)
previous models. 5 IF ‘‘Shipment Size’’ is Medium AND ‘‘Other Pro-
tection’’ is Needed
The figures in the brackets following each mode in the con- THEN ‘‘Selected Mode’’ is Rail (0.86) OR ‘‘Se-
sequent part represent the corresponding degree of belief. Tri- lected Mode’’ is Truck (0.14)
angular membership functions (k = 2) are used by the model 6 IF ‘‘Shipment Size’’ is Large AND ‘‘Other Pro-
tection’’ is Needed
for all of the linguistic variables. The membership functions THEN ‘‘Selected Mode’’ is Rail (0.79) OR ‘‘Se-
are shown in Fig. 5. lected Mode’’ is Truck (0.21)
7 IF ‘‘Truck Freight Charges’’ is Low
THEN ‘‘Selected Mode’’ is Rail (0.17) OR ‘‘Se-
TABLE 4. Variables Used in Most Efficient ANN Model lected Mode’’ is Truck (0.83)
8 IF ‘‘Truck Freight Charges’’ is High
Variable THEN ‘‘Selected Mode’’ is Rail (1.00)
code Definition 9 IF ‘‘Truck Reliability’’ is Low
(1) (2) THEN ‘‘Selected Mode’’ is Rail (0.74) OR ‘‘Se-
lected Mode’’ is Truck (0.26)
ss Shipment size (lb) 10 IF ‘‘Truck Reliability is High
vl Value (1977 U.S.$/lb) THEN ‘‘Selected Mode’’ is Rail (0.34) OR ‘‘Se-
frz Shipment requires freezing (1 = yes; 0 = no) lected Mode’’ is Truck (0.66)
shk Shipment requires shock protection (1 = yes; 0 = no)
prt Shipment is in particulate state (1 = particulate; 0 other-
wise) The model’s ability to simulate the decision-makers behav-
oth Shipment requires other protection (1 = yes; 0 = no) ior was then assessed using the testing data set. A midrange
dist Shipping distance in miles
rcost Rail freight charges (cents/cwta)
threshold similar to the ANN models was used to define the
tcost Truck freight charges (cents/cwta) output as one of two modes. The result of the assessment is
rtime Rail transit time (days) presented in Table 7.
ttime Truck transit time (days)
a
cwt = hundredweight. Classification Threshold
An important issue in the application of ANN and neuro-
TABLE 5. Prediction Results of Most Efficient ANN Model (Us- fuzzy models to simulate observed behavior, such as the
ing 11 Variables) choice of mode by shippers, is the discretization process, also
Predicted by ANN known as the ‘‘classification threshold,’’ that is, the threshold
to be adopted to transfer the network output (a value between
Percent
0 and 1) into a discrete value (0 or 1). In a discrete model like
Observed Rail Truck Total correct
mode choice, the final results must be classified to belong to
(1) (2) (3) (4) (5)
one and only one category (i.e., mode). For example, a typical
Rail 651 254 905 71.9 neurofuzzy model result may be (0.54, 0.46), and another may
Truck 35 1,560 1,595 97.8
Total 686 1,814 2,500 88.4
be (0.96, 0.04). Although it may be fair to say that the second
result, that is, 96% probability of choosing mode 1 and 4%
28 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
 FIG. 5. Membership Functions

TABLE 7. Predicted Results of Neurofuzzy Model Using Four (percent correctly predicted) for an ANN and the neurofuzzy
Variables model is shown in Table 8.
Predicted by Neurofuzzy Model
Percent DISCUSSION
Observed Rail Truck Total correct
(1) (2) (3) (4) (5) A comparison of the prediction results of the four models
shows that their performances are relatively similar. Although
Rail 608 297 905 67.2
Truck 16 1,579 1,595 99.0 in terms of accuracy no approach shows a clear advantage, the
Total 624 1,876 2,500 87.5 neurofuzzy approach uses the smallest number of input vari-
ables to obtain the same performance. Therefore, it could be
considered as the most efficient among these approaches. In
probability of choosing mode 2, can certainly be classified to addition, the fact that it is capable of building such a model
belong to mode 1, it is not apparent whether the first result automatically is another advantage, while other models require
should end up with the same classification. Furthermore, the a manual, and sometimes subjective, intervention. The trans-
prediction of mode choice in the two examples given above parency of the knowledge gained by the neurofuzzy approach
are obviously not made with equal degrees of certainty. For is also an advantage. The use of linguistic variables makes it
example, if one requires 50% confidence in the prediction re- relatively easy to interpret the rules and if necessary change
sults, then both cases can be classified as mode 1. If one re- them.
quires 90% confidence, then only the (0.96, 0.04) result can Although all models achieved similar performance in the
be classified as mode 1, and the (0.54, 0.46) can only be clas- example used in this study, it is expected that in applications
sified to belong to a fuzzy state, somewhere between the two where the input/output relationship is more complex and in-
modes. The above examples illustrate that the predictive abil- volves higher nonlinearity that the ANN and neurofuzzy ap-
ity of a model decreases with an increasing level of confi- proaches would achieve a higher performance. This is not only
dence. due to the fact that conventional approaches (logit and probit)
In conventional mode choice models, the rate of success, or use linear regression to approximate potentially nonlinear re-
predictive ability of the model is assessed by summing up all lationships but also because in these models each input vari-
probabilities of choosing a particular mode over all observa- able can only affect the output in a univariate manner. This
tions. The sum of probabilities is then compared with the ob- would only be suitable for cases where the output f (x1, x2,
served share of that mode in the testing data set, and a success . . .) can be approximated by an additive (and linear) combi-
rate (or percent correctly predicted) is calculated. With the nation of f (x1), f (x2), etc. In neurofuzzy models, however, each
introduction of neurofuzzy and ANN models, it is now pos- subnetwork may consist of various combinations of as many
sible to fine tune the discretization process to reflect any de- input variables as is necessary up to a maximum number set
sired level of confidence. For comparison purposes, the vari- by the user. The same is implicitly true, of course, for ANN,
ation of the predictive ability, measured by the success rate but it comes at the price of the nontransparency that it causes.
Table 9 provides a comparison of the three approaches.
TABLE 8. Comparison Success Rates by Confidence Level
TABLE 9. Comparison of Three Approaches
Success Rate (% Correctly Predicted)
Comparison criteria Logit ANN Neurofuzzy
Confidence Neurofuzzy
(1) (2) (3) (4)
level Logit model ANN model model
(%) (12 variables) (11 variables) (4 variables) Constraining assumptions pres- IID & error No No
ent distribution
(1) (2) (3) (4) Possibility of usage for more Yes Yes Yes
50 86.3 88.4 87.5 than two modes
60 79.6 85.7 86.6 Ability to model nonlinearity No Yes Yes
Ability to model multivariability No Yes Yes
70 71.8 81.5 80.0
Automatic exclusion of irrele- No No Yes
80 61.2 72.1 69.4 vant inputs
90 43.4 55.6 48.8 Transparency of model To some degree Not transparent Very good
95 32.5 43.3 39.2 Ability to insert expert knowl- No No Yes
99 16.5 32.1 33.8 edge in model

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 29

CONCLUSIONS Hausman, J., and Wise, D. (1978). ‘‘A conditional probit model for qual-
itative choice: Discrete decisions recognizing interdependence and het-
In this paper, a neurofuzzy approach was proposed for be- erogeneous preferences.’’ Economitrica, 52, 1219–1240.
havior mode choice modeling. A large data set taken from the Jang, J. (1993). ‘‘ANFIS: Adaptive-network-based fuzzy inference sys-
U.S. freight transport market was used as an example to study tem.’’ IEEE Trans. on Sys., Man and Cybernetics, 23, 665–685.
Kavli, T. (1993). ‘‘ASMOD—An algorithm for adaptive spline modeling
the feasibility and advantages of the proposed approach. It was of observation data.’’ Int. J. Control, 58(4), 947–967.
shown that the neurofuzzy approach is capable of producing Kosko, B. (1992). Neural networks and fuzzy systems. Prentice-Hall, En-
the accuracy obtained by the conventional methods and ANNs. glewood Cliffs, N.J.
Most importantly, it was observed that the model developed Masters, T. (1993). Practical neural recipes in C⫹⫹. Academic, San
by the proposed approach uses the least amount of information Diego.
in terms of the number of inputs. In addition, the neurofuzzy McFadden, D. (1978). ‘‘The theory and practice of disaggregate demand
forecasting for various modes of urban transportation.’’ Emerging
approach automates the selection of the input variables whose transportation planning methods, Rep. No. DOT-RSPA-DP8-50-78-2,
contribution is significant enough to be used in the model. The U.S. Department of Transportation, Washington, D.C.
neurofuzzy model is also the easiest model to interpret because Neuframe manual. (1997). Neural Computer Sciences, Southampton,
of the linguistic variables that are used. The transparency of U.K.
the model achieved by the proposed approach can be very Ortuzar, J., and Willumsen, L. (1994). Modeling transport. Wiley, En-
useful in many applications. This enables the users to directly gland.
Rumelhart, D. E., and McClelland, J., eds. (1986). Parallel distributed
add their own expertise on the subject before or after the processing. Vol. 1. MIT Press, Cambridge, Mass.
model is built. This study shows that the neurofuzzy has a Zadeh, L. A. (1965). ‘‘Fuzzy sets.’’ Information and Control, 8, 338–
great potential to be successfully applied to a wide range of 353.
transportation problems.
APPENDIX II. NOTATION
APPENDIX I. REFERENCES
Abdelwahab, W., and Sayed, T. (1999). ‘‘Artificial neural networks: A The following symbols are used in this paper:
new approach to behavioral mode choice modeling.’’ Civ. Engrg. and
Envir. Systems. Ai = multivariate fuzzy linguistic input set that forms part of
Ben-Akiva, M., and Lerman, S. (1987). Discrete choice analysis: Theory ith multivariate rule;
and application to travel demand, 2nd Ed., Massachusetts Institute of
Technology, Cambridge, Mass.
a(x) = vector of basis function outputs;
Bossley K., Brown M., and Harris C. (1995). ‘‘Parsimonious neurofuzzy Bj = univariate fuzzy linguistic output set that forms part of
modeling.’’ Tech. Rep., Dept. of Electronics and Comp. Sci., University ith multivariate rule;
of Southampton, Southampton, U.K. ci j = rule confidence relating ith fuzzy input set jth fuzzy out-
Brown, M., and Harris, C. (1994). Neurofuzzy adaptive modeling and put set;
control. Prentice-Hall, New York. Pk(i) = probability that individual i chooses mode k;
Brown, M., and Harris, C. (1995). ‘‘A perspective and critique of adaptive Uik = utility of individual i choosing mode k;
neurofuzzy systems used for modeling and control applications.’’ Int. uA(x) = membership of x in fuzzy set A;
J. Neural Sys., 6(2), 197–220. Vik = deterministic utility of individual i choosing mode k;
Daganzo, C. F. (1979). Multinomial probit: The theory and its applica-
tions to demand forecasting.’’ Academic, New York.
w = vector of network weights;
Finney, D. (1971). Probit analysis, 3rd Ed., Cambridge University Press, ␣k = constant for mode k;
Cambridge, England. εi = random error term;
Ghareib, A. H. (1996). ‘‘Evaluation of logit and probit models in mode- ␴ = standard deviation of (εk ⫺ εj); and
choice situation.’’ J. Transp. Engrg., ASCE, 122(4), 282–290. ⌽(⭈) = standardized cumulative normal distribution function.

30 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 GENETIC PROGRAMMING-BASED APPROACH FOR
STRUCTURAL OPTIMIZATION
By Chee Kiong Soh1 and Yaowen Yang2

ABSTRACT: The main purpose of this paper is to introduce genetic programming into civil engineering problem
solving. This paper describes a genetic programming-based approach for simultaneous sizing, geometry, and
topology optimization of structures. An encoding strategy is presented to map between the real structures and
the genetic programming parse trees. Numerical results for two examples reveal that the proposed approach is
capable of producing the topology and shape of the desired trusses and the sizing of all the structural components.
Hence, this approach can potentially be a powerful search and optimization technique in civil engineering
problem solving.

INTRODUCTION strings, GP’s encoding of 2D parse trees of varying size and

complexity is more natural and flexible in representing the real
Optimization is a critical and challenging activity in struc- structures. There are other advantages over GAs as well. This
tural design. Designers are able to produce better designs while leads us to believe that GP may be more efficient in search
saving time and money through optimization. However, many than GAs, or at least it could be an alternative to GAs in
optimization problems in structural engineering are very com- solving structural optimization problems. This is the principal
plex in nature and quite difficult to solve by conventional op- motive in employing a GP approach in this study. Another
timization techniques, such as the mathematical programming motive is that GP is an original and novel methodology in
method, optimality criteria method, and so on. Since the civil engineering problem solving as it has yet to be applied
1960s, there has been an increasing interest in imitating living to this area.
beings (botany, animal, human) to solve such difficult opti- This study introduces GP into civil engineering problem
mization problems. Simulating the natural evolutionary pro- solving as a search and optimization method and investigates
cess of human beings results in random optimization tech- the applicability of a GP-based approach to structural opti-
niques called evolutionary algorithms. Evolutionary algo- mization problems. Although the proposed approach is illus-
rithms is a general term encompassing a number of related trated by truss structures, it is a general optimization meth-
methodologies, all of which are based on the natural evolu- odology that can be conveniently applied to other types of
tionary paradigm. Genetic algorithms (GAs), evolution strat- structures, such as frames, planar structures, plates, and shells.
egies, and evolutionary programming are the historically
prominent approaches, with genetic programming (GP) rapidly GENETIC PROGRAMMING
emerging in recent years.
In the last decade, GAs have proved to be successful and Genetic programming is an extension of John Holland’s ge-
powerful in solving various structural optimization problems netic algorithm (1975), in which the population consists of
(Hajela 1990; Jenkins 1991, 1992; Rajeev and Krishnamoorthy computer programs of varying sizes and shapes (Koza 1992).
1992, 1997; Adeli and Cheng 1993; Koumousis and Georgiou The GP paradigm follows the trend of dealing with the prob-
1994; Hajela and Lee 1995; Soh and Yang 1996; and Camp lem of representation in GAs by increasing the complexity of
et al. 1998). Compared to conventional optimization methods, the structures undergoing adaptation. In particular, the struc-
GAs have the advantages of imposing fewer mathematical re- tures (usually the parse trees) undergoing adaptation in GP are
quirements for solving the problems and of being very effec- general, hierarchical computer programs of dynamically var-
tive at performing global search. However, their lack of effi- ying sizes and shapes. Genetic programming breeds a popu-
ciency in search leads to high cost of computing time (Hajela lation of rooted, point-labeled trees (that is, graphs without
1990; Haftka et al. 1996; and Jenkins 1996). cycles) with ordered branches.
On the other hand, GP has been empirically shown to be a The search space in GP is the space of all computer pro-
powerful program-induction methodology and has been suc- grams composed of functions and terminals appropriate to the
cessfully applied to a large number of difficult problems, such problem domain. The functions may be standard arithmetic
as automatic design, pattern recognition, robotic control, syn- operations, standard programming operations, standard math-
thesis on neural architectures, symbolic regression, factory job ematical functions, logic functions, domain-specific functions,
scheduling, electronic circuit design, signal processing, and or any other functions defined by the developers. The termin-
music and picture generation. Recent works are described in als may be constants, variables, vectors, matrices, symbols, or
Kinnear (1994), Angeline and Kinnear (1996), and Koza et al. any other quantities defined by the developers. In GP, the func-
(1996, 1997). Although GP is an extension of GAs, it is in tion set consists of all kinds of functions and the terminal set
some sense much more powerful than GAs. Compared to consists of all kinds of terminals.
GAs’ encoding of individuals as 1D fixed-length binary In GP, populations of hundreds or thousands of computer
programs are genetically bred. This breeding is done using the
1
Assoc. Prof., School of Civ. and Struct. Engrg., Nanyang Technol.
Darwinian principle of survival and reproduction of the fittest
Univ., Singapore 639798. along with genetic operations (crossover and mutation) appro-
2
Res. Scholar, School of Civ. and Struct. Engrg., Nanyang Technol. priate for mating computer programs. A computer program that
Univ., Singapore 639798. solves (or approximately solves) a given problem may emerge
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. from this combination of Darwinian natural selection and ge-
To extend the closing date one month, a written request must be filed netic operations.
with the ASCE Manager of Journals. The manuscript for this paper was
submitted for review and possible publication on November 18, 1998.
According to Koza (1992), GP starts with an initial popu-
This paper is part of the Journal of Computing in Civil Engineering, lation of randomly generated computer programs. Functions
Vol. 14, No. 1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0031– are selected from the function set and denoted as roots or inner
0037/$8.00 ⫹ $.50 per page. Paper No. 19669. nodes of the parse trees, and terminals are selected from the
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 31
terminal set and denoted as leaves of the parse trees. The cre- of point-labeled trees found in the world of genetic program-
ation of this initial random population is, in effect, a blind ming and the node-element-labeled diagrams employed in the
random search of the search space of the problem. Each in- analysis of structures.
dividual computer program in the population is measured in We introduce three definitions, that is, S-expressions, leaf,
terms of how well it performs in the particular problem en- and function.
vironment. This measure is called fitness measure or evalua-
tion. It is usually accomplished by running each program on • Definition 1: An S-expression is a symbolic expression of
a collection of inputs called fitness cases and by assigning a GP parse tree. It consists of functions selected from a
numerical fitness values to the output of each of these runs; function set and terminals selected from a terminal set.
the resulting values are then combined to produce a single • Definition 2: A leaf in a GP parse tree is an end node of
fitness value for the program. the tree. It denotes a joint/node in the structure that the
Unless the problem is so trivial that it can be easily solved parse tree represents. In an S-expression, leaves are ter-
by blind random search, the computer programs in generation minal, that is, arguments of functions.
0 will have exceedingly poor fitness. Nonetheless, some in- • Definition 3: A function in a GP parse tree is an inner
dividuals in the population will turn out to be somewhat fitter node of the tree. A function connects leaves or output of
than others. These differences in performance are then ex- other functions with a member specified and outputs one
ploited. of its argument as result.
The Darwinian principle of reproduction and survival of the
fittest and the genetic operations of crossover and mutation are
Using the above definitions, the proposed encoding method
used to create a new offspring population of individual com-
is illustrated by the following example.
puter programs from the current population of programs.
The truss shown in Fig. 1(a) consists of four nodes, N1, N2,
The reproduction operation involves selecting a computer
N3, and N4, and six members, i, j, k, l, m, and n. It can be
program from the current population of programs based on
represented by the GP parse tree shown in Fig. 1(b). The tree
fitness (that is, the better the fitness, the more likely the in-
is composed of two kinds of nodes, the leaves and the inner
dividual is to be selected) and allowing it to survive by copy-
nodes, which are connected by lines. The leaves are the joints
ing it into the population.
selected from the truss. The inner nodes Ap ( p = i, j, k, l, m,
The crossover operation is used to create new offspring
n) are the functions representing the cross-sectional areas of
computer programs from two parental programs selected based
the members. The inner nodes can either connect two leaves,
on fitness. The parental programs are typically of different
or one leaf and one output of another function, or two outputs
sizes and shapes. The offspring programs are composed of
of the other two functions. If there is a connection line at the
subexpressions (subtrees, subprograms, subroutines, and build-
upper left of Ap, the output of Ap is the leaf at its lower right;
ing blocks) from their parents, and typically of different sizes
conversely, if there is a connection line at the upper right of
and shapes from their parents. The creation of offspring from
Ap, the output is the leaf at its lower left. For example, function
the crossover operation is accomplished by swapping random
Aj connects two leaves, N1 and N2, with member j. Because
fragments of the parents. An important improvement that GP
there is a connection line at its upper left, the output of Aj is
displays over GAs is its ability to create two new solutions
the leaf at its lower-right, leaf N1. Function An connects the
from the same solutions.
output of function Aj (leaf N1) and leaf N4 with member n; its
Mutation is implemented by randomly removing a subtree
output is leaf N4.
at a selected point and replacing it with a randomly generated
The S-expression of this parse tree can be represented as
subtree. Koza (1992) has argued that mutation is in fact useless
in GP because of the position independence of GP subtrees Am(Ai(Al(N3, N4), N1), Ak(An(N4, Aj (N2, N1)), N2))
and because of the larger number of chromosome positions in
typical GP populations. However, recent research (Luke and where function Ap(Nr, Ns), ( p = i, j, k, l, m, n; r, s = 1, 2, 3,
Spector 1997) indicated that mutation can in fact have use- 4) connects two joints, Nr and Ns, with member p, and with
fulness and that crossover does not consistently have a con- outputs Nr if there is a comma following the function, or Ns
siderable advantage over mutation, especially in the case of
small populations and large generations.
After the genetic operations are performed on the current
population, the population of offspring (the new generation)
replaces the old population (the old generation). Each individ-
ual in the new population of computer programs is then mea-
sured for its fitness, and the process is repeated over many
generations. The force driving this process consists only of the
observed fitness of the individuals in the current population in
grappling with problem environment. This algorithm will pro-
duce populations of computer programs that, over many gen-
erations, tend to exhibit increasing average fitness in dealing
with their environment.
The best individual that appeared in any generation of a run
(that is, the best-so-far individual) is typically designated as
the result produced by GP.

MAPPING BETWEEN STRUCTURES AND GP

PROGRAMMING PARSE TREES
How to encode a solution of the problem into a chromosome
(in GAs) or a parse tree (in GP) is the key issue for application
of genetic methods. Genetic programming can be applied to FIG. 1. Illustration of GP Parse Tree of Truss: (a) A Six-Bar
structural design if a mapping is established between the kind Truss; (b) GP Parse Tree; (c) Alternative GP Parse Tree

32 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

if there is a comma ahead of the function. In fact, the S- Find the particular set of sizing variables A, geometric var-
expression can be viewed as a computer program that consti- iables, R, and topological variables T such that structural
tutes certain functions and terminals. weight or material volume function W(R, A, T)
Note that both the parse tree shown in Fig. 1(b) and the
above S-expression refer to the same structure, that is, the six- W(R, A, T) → min (1)
bar truss shown in Fig. 1(a). The parse tree representation is
a graphic expression of the structure, while the S-expression subject to G ⱕ Gj(R, A, T) ⱕ G ,
L
j
U
j j = 1, 2, . . . , q (2)
is a symbolic expression of the structure. The graphic expres- in which A and R are the vectors of member cross-sectional
sion is a materialized encoding of a structure, and the symbolic areas and joint coordinates, respectively. T represents the set
expression is applied as a computer program that can be of existing members. L and U are superscripts denoting the
evolved in GP. lower and upper bounds, respectively. Gj(R, A, T) are the con-
Applying the above encoding approach, the six-bar planar straints that limit the relevant design variable domains, and q
truss can also be represented by the parse tree shown in Fig. is the total number of constraints.
1(c) or any other suitable trees conforming to the above en- The upper and lower bounds on the constraints functions of
coding rule. It should be noted that the encoding of the truss (2) include the following: (a) Member cross-sectional areas
is not unique, that is, the mapping from parse trees to struc- (A Li ⱕ Ai ⱕ A Ui , i = 1, 2, . . . , NM); (b) joint coordinates (R Li
tures is an n-to-1 mapping. Although this n-to-1 mapping is ⱕ Ri ⱕ R Ui , i = 1, 2, . . . , NJ); (c) member allowable stresses
applicable in practice, it enlarges the search space of GP at the (␴ Li ⱕ ␴i ⱕ ␴ Ui , i = 1, 2, . . . , n1 ⱕ NM); (d) joint allowable
same time and thus lowers the efficiency of the search. To displacements (⌬ Li ⱕ ⌬i ⱕ ⌬ Ui , i = 1, 2, . . . , n2 ⱕ NJ); and
overcome this limitation, an improved encoding approach is (e) member buckling strength (␴ ib ⱕ ␴i ⱕ 0, i = 1, 2, . . . , n3
employed as follows (without lost of generality, assume that ⱕ NM). Here NM, NJ, n1, n2, n3 are the total number of mem-
N1 < N2 < N3 < N4 and i < j < k < l < m < n): bers, joint coordinates, member allowable stress constraints,
joint allowable displacement constraints, and member buckling
1. Select the root of the parse tree as the minimum member strength constraints, respectively.
in the truss, member i. Write down function Ai and mem-
ber i’s two nodes (N1 and N3) as Ai’s arguments, with the
TERMINAL AND FUNCTION SETS
smaller node at below left. See Fig. 2(a). Remove mem-
ber i from the truss. The function set, F, for connection nodes in a GP parse tree
2. From left to right, find the minimum numbered members contains NM two-argument connecting functions. That is
(MNMs) among all members associated with the nodes.
For example, the MNMs for nodes N1 and N3 are j and F = {A1, A2, . . . , ANM} (3)
l, respectively. Replace the nodes with their MNMs and
write down the MNMs’ arguments (nodes N1 and N2 for where function Ai (i = 1, 2, . . . , NM) connects its two argu-
Aj and nodes N4 and N3 for Al). Remember to keep the ments with a member whose cross-sectional area is Ai, and Ai
former below-left node (node N1) and below-right node is a random number between A Li and A Ui . The constraints of a
(node N3) at the same side as their former positions. See member cross-sectional area can be included in this type of
Fig. 2(b). Once a member has appeared in the tree, re- function definition. The output of Ai is one of its arguments,
move it from the truss. as defined in the encoding approach.
3. Repeat step 2 until no member is left in the truss. See The terminal set, T, for connecting functions consists of
Fig. 2(c).
T = {N1(x1, y1, z1), N2(x2, y2, z2), . . . , NNJ(xNJ, yNJ, zNJ)} (4)
With the above improvement, mapping between structures where (xj, yj, zj), j = 1, 2, . . . , NJ are coordinates of joint Nj
and GP parse trees has become an ideal 1-to-1 mapping. and are random numbers selected from [x Lj , x Uj ], [y Lj , y Uj ], and
[z Lj , z Uj ], respectively. The constraints of joint coordinates can
PROBLEM DEFINITION also be included in this type of terminal definition.
The encoding approach shows that the layout of a GP parse
The structural model of a simultaneous sizing, geometry, tree reflects the topology of the truss considered. If we select
and topology design problem is itself allowed to vary during function values, that is, select cross-sectional areas of members
the optimization process. In this study, the process can be from the function set F, and select different values of coor-
stated as follows: dinates from terminal set T, the evolution of a population of
such a kind of parse trees is in fact a combined sizing, ge-
ometry, and topology optimization problem of the given truss.

FITNESS MEASURE
The fitness measure is what directs the evolutionary process
to a parse tree that constructs the optimal truss. In other words,
the optimal truss arises from fitness.
The penalty approach has been employed to couple with the
constraints of member allowable stresses, joint allowable dis-
placements, and member buckling strength in (2). The reason
for using the penalty approach instead of the rejecting strategy
is that, in constrained optimization problems, the optimum typ-
ically occurs at the boundary between feasible and infeasible
areas, and the penalty approach will force the genetic search
to approach the optimum from both feasible and infeasible
FIG. 2. Improved Encoding Approach: (a) First Step; (b) Sec- regions. The evaluation function of fitness of the following
ond Step; (c) Final Form of Parse Tree form is adopted
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 33
 冘 冘冋 册
p nj 2
兩Gij兩 The penalty approach is employed to cope with the
⌽(R, A, T) = W(R, A, T) ⫹ cj ⫺1 (5) constraints. The GP run is terminated if the termination
j=1 i=1 兩G aij兩
criterion is satisfied. Here, the maximum number of gen-
G aij = G Lij, when Gij < 0 erations is set as the termination criterion.
4. Encode the structures into the GP parse tree and perform
G aij = G Uij , when Gij ⱖ 0
a genetic operation, that is, reproduction, crossover, and

冋 兩Gij兩
兩G aij兩
⫺1 册 = max 冉 兩Gij兩
兩G aij兩
⫺ 1, 0 冊 mutation.
5. Return to step (2) and repeat the process until the ter-
mination criterion is satisfied.
where cj = the jth penalty parameters ( j = 1, 2, . . . , p) and p
is the total number of constraints. As GP seeks to maximize ILLUSTRATIVE EXAMPLES
the objective function, a rescaling evaluation objective func-
tion can be defined as follows: The two examples presented here are the two well-tested
examples used in structural optimization research. English
fitness units are used in references and in our computing, while the

= 再
Cavg (generation) ⫺ ⌽(R, A, T)
0
when
when
⌽(R, A, T) < Cavg
⌽(R, A, T) ⱖ Cavg
results given in the tables have been converted to SI units.

Example 1: 10-Member Planar Truss

(6)
The parameter Cavg is used to prevent a vector of the design The truss whose initial shape and topology are shown in
variables with fitness less than or equal to Cavg from entering Fig. 4 is subjected to two loads acting simultaneously. The
a new ‘‘pool’’ of candidate solutions that will be used for objective is to minimize its weight with varying shape and
reproduction in the next generation. It also gives high fitness topology. The design data are as follows: modulus of elasticity
values to ‘‘good’’ candidate solutions and low fitness to those is 206.7 GPa (30,000 ksi), material density is 2.768 g/cm3 (0.1
that violate the constraints or simply are ‘‘poor’’ solutions.
When ⌽(R, A, T) converges to a minimum value, the fitness
will tend to a maximum value. Therefore, such a fitness eval-
uation objective function can adequately measure the perfor-
mance of a candidate solution.

SOLUTION PROCEDURE
Fig. 3 gives an overview of the GP-based search process
for simultaneous sizing, geometry, and topology optimization.
The process can be divided into the following steps.
1. Initialize the population of the generation.
An individual of the initial population is generated by
randomly selecting functions from the function set and FIG. 4. 10-Member Planar Truss
terminals from the terminal set, constituting a GP parse
tree using functions and terminals selected by the ramped
half-and-half method (Koza 1992).
2. Decode the individuals of the GP parse tree into real
structures and perform FEM analysis of the structures.
If the structure decoded from its GP parse tree is not
a real structure and may not be analyzed by FEM, pe-
nalize the fitness of this individual in the next step and
skip the FEM analysis.
3. Evaluate the fitness value of each individual.

FIG. 5. Progressive Evolution of Topology and Geometry of

10-Bar Truss in Case 1: (a) Generation 0 (Weight = 10.587 kN); (b)
Generation 7 (Weight = 8.641 kN); (c) Generation 16 (Weight =
8.170 kN); (d) Generation 24 (Weight = 7.441 kN); and (e) Gener-
FIG. 3. Solution Procedure for Proposed GP-Based Approach ation 32 (Weight = 6.932 kN)

34 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

pci), the allowable tensile and compressive stresses are 172.25 tion by performing a 10% reproduction operation, 80%
MPa (25 ksi), and the allowable displacement at Joint 2 is crossover operation, and 10% mutation operation.
5.08 cm (2.0 in.). The bound on the member cross-sectional 3. Perform the GP run as described in the section on ‘‘So-
areas is 0.452–64.516 cm2 (0.07–10.0 in.2). The bounds on lution Procedure.’’
the joint coordinates are as follows: 10.16–20.32 m (400–800
in.) for x1 and 5.08–12.7 m (200–500 in.) for x3, y1, and y3. For Case 1, the progressive evolution of the topology and
The following two cases were investigated: Case 1, only stress shape of the truss during the run is shown in Fig. 5. The trusses
constraints; and Case 2, both stress and displacement con- shown in Figs. 5(a) to 5(e) possess the best fitness values of
straints. the generation. The optimal solution was obtained after a run
The solution procedure of this problem is as follows: of 32 generations. For Case 2, the minimum weight truss was
obtained in Generation 38.
1. Determine the function set and the terminal set. Comparison of results for the independent sizing optimiza-
Eqs. (3) and (4) have been used to express the function tion, for both the geometry and sizing optimization, for both
set F and the terminal set T of the problem, respectively, the topology and sizing optimization, and for simultaneous
where NM = 10 and NJ = 6. In (3), Ai (i = 1, 2, . . . , sizing, geometry, and topology optimization using GA and this
10) is a random number between 0.07 and 10.0. In (4), work for both Cases 1 and 2 is presented in Tables 1 and 2,
x1 is a random number between 400 and 800, x3, y1, and respectively. It is apparent that results obtained using the pro-
y3 are random numbers from 200 to 500, and x2, y2, x4, posed GP-based approach are better optimized.
y4, x5, y5, x6, and y6 are constants of 720, 0, 360, 0, 0,
360, 0, and 0, respectively. Example 2: 25-Bar Transmission Tower Space Truss
2. Select the control parameters for the GP run.
The control parameters for the GP run are as follows: Fig. 6 shows a 25-bar transmission tower space truss with
The population size is 500, and the maximum generation an initial geometry and topology that have been investigated
of the run is 50. Each subsequent generation of the run in sizing optimization using numerical optimization methods
is created from the population of the preceding genera- (Vanderplaats and Moses 1972), combination of geometry and

TABLE 1. Result of 10-Bar Truss for Case 1

2
Functions Ai (cm ) Fixed geometry Fixed Geometry Variable Geometry and Topology
and terminals (Xj, Yj) and topology Fixed topology
(m) (Kirsch 1990) (Yang and Soh 1995) Kirsch 1990 Hajela et al. 1993 Yang 1996 This work
(1) (2) (3) (4) (5) (6) (7)
A1 51.55 51.48 51.61 52.26 51.84 52.00
A2 0.645 0.710 — — — —
A3 52.00 53.87 51.61 51.61 53.89 53.29
A4 25.42 27.23 25.81 27.10 28.08 27.48
A5 0.645 1.161 — — — —
A6 0.645 0.645 — — — —
A7 37.03 37.10 36.52 37.42 36.52 36.52
A8 35.94 32.32 36.52 40.65 32.01 31.03
A9 35.94 37.05 36.52 39.35 38.17 37.68
A10 0.645 0.774 — — — —
X1 18.288@ 18.288@ — — — —
Y1 9.144@ 6.096 — — — —
X3 9.144@ 9.144@ 9.144@ 9.144@ 9.144 9.855
Y3 9.144@ 8.517 9.144@ 9.144@ 8.407 7.925
Weight (N) 7,089.42 7,022.88 7,043.01 7,367.10 6,968.65 6,932.16
Note: @ denotes coordinate is stationary.

TABLE 2. Result of 10-Bar Truss for Case 2

2
Functions Ai (cm ) and Fixed geometry and Variable Geometry and Topology
terminals (Xj, Yj) topology Fixed topology Fixed geometry
(m) (Kirsch 1990) (Yang and Soh 1995) (Kirsch 1990) Yang 1996 This work
(1) (2) (3) (4) (5) (6)
A1 63.87 59.35 58.06 57.42 64.39
A2 0.645 0.581 — — —
A3 51.61 54.45 51.61 51.84 52.26
A4 31.87 30.26 31.87 30.06 30.45
A5 0.645 0.484 — — —
A6 0.645 1.161 — — —
A7 36.77 37.10 36.52 36.58 36.52
A8 45.16 40.97 45.16 41.40 40.58
A9 45.16 49.23 45.16 48.71 43.68
A10 0.645 1.290 — — —
X1 18.288@ 18.288@ — — —
Y1 9.144@ 7.633 — — —
X3 9.144@ 9.144@ 9.144@ 10.160 9.855
Y3 9.144@ 8.466 9.144@ 8.052 8.230
Weight (N) 8,161.02 8,018.47 7,959.93 7,873.56 7,870.23
Note: @ denotes coordinate is stationary.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 35

 sizing optimization using GA approach (Hansen and Vander-
plaats 1990; Soh and Yang 1996), and simultaneous sizing,
geometry, and topology optimization using also the GA ap-
proach (Yang 1996). For this study, the two load cases and the
details of the member groups are given in Tables 3 and 4,
respectively. The truss is also required to remain symmetric
with respect to both the X–Z plane and the Y–Z plane. Thus,
only five coordinates, that is, x4, y4 and z4 of Joint 4 and x8
and y8 of Joint 8, are considered as geometric variables for the
sake of symmetry of the structure. The material properties are
as follows: The modulus of elasticity is 68.9 GPa (10,000 ksi),
the density of material is 2.768 g/cm3 (0.1 pci), the member
allowable stress in tension and compression is 275.6 MPa (40
ksi), and the allowable buckling strength is ⫺100␲EAi /8L 2i (i
= 1, 2, . . . , 25). The lower and upper bounds on the sizing
and geometric variables are as follows: 0.06452–6.452 cm2
(0.01⫺1.0 in.2) for Ai, 25.4–76.2 cm (10–30 in.) for x4 and
x8, 101.6–203.2 cm (40–80 in.) for y4, 152.4–254 cm (60–
FIG. 6. 25-Bar Transmission Tower Space Truss 100 in.) for y8, and 203.2–304.8 cm (80–120 in.) for z4.
In the function set and terminal set of this truss, NM = 25
and NJ = 10; Ai(i = 1, 2, . . . , 25) is a random number in
TABLE 3. Load Cases for 25-Bar Truss [0.01, 1.0]; x4(= ⫺x3 = x5 = ⫺x6) and x8(= ⫺x7 = x9 = ⫺x10)
Force at joints are random numbers in [10, 30]; y4(=y3 = ⫺y5 = ⫺y6) is a
Load Loading random number in [40, 80]; y8(=y7 = ⫺y9 = ⫺y10) is a random
cases (kN) 1 2 3 6
number [60, 100]; z4(=z3 = z5 = z6) is a random number in [80,
(1) (2) (3) (4) (5) (6)
120]; x1(= ⫺x2), y1(=y2), and z1(=z2) and z8(=z7 = z9 = z10) are
1 Fx 0.0 0.0 0.0 0.0 constants of 37.5, 0, 200, and 0, respectively.
Fy 88.9 ⫺88.9 0.0 0.0
Example 2 is similar to Example 1 in that the optimum
Fz ⫺22.225 ⫺22.225 0.0 0.0
sizing, geometry, and topology design of the structure was
2 Fx 4.445 0.0 2.2225 2.2225 sought by the proposed GP-based approach. The control pa-
Fy 44.45 44.45 0.0 0.0
Fz ⫺22.225 44.45 0.0 0.0
rameters for the GP run are the same as for Example 1, except
that the maximum number of generations is set at 80. Com-
parisons of the results for the independent sizing optimization,
for optimization under fixed topology, and for simultaneous
TABLE 4. Member Groups for 25-Bar Truss sizing, geometry, and topology optimization using GA and this
work using GP are presented in Table 5. The GP approach
Groups Members (end joints)
resulted in the deletion of four members (2–4, 1–6, 6–7, and
(1) (2)
5–8), as shown in Fig. 7. From Table 5, an optimal truss
A1 1–2 structure was achieved by the approach after a GP run of 52
A2 1–3, 2–4, 2–5, 1–6
A6 1–4, 2–3, 2–6, 1–5
generations. Again, the GP-based approach offered a more op-
A10 3–6, 4–5 timized result than all the other methods.
A12 3–4, 5–6
A14 6–7, 3–10, 5–8, 4–9 CONCLUSION
A18 3–8, 4–7, 6–9, 5–10
A22 3–7, 4–8, 5–9, 6–10
How to represent real structures using GP parse trees is the
key issue in applying this hierarchical search methodology to

TABLE 5. Optimal Results for 25-Bar Truss

Fixed geometry Fixed Topology Variable Geometry and Topology
Functions AI (cm2) and and topology
terminals (Xj, Yj) (Vanderplaats and Hansen and
(cm) Moses 1972) Vanderplaats 1990 Soh and Yang 1996 Yang 1996 This work
(1) (2) (3) (4) (5) (6)
A1 0.058 0.065 0.581 1.290 0.581
A2 5.045 3.142 2.839 3.613 3.548
A6 4.865 5.394 5.806 5.548 5.484
A10 0.006 0.155 0.323 0.710 0.581
A12 0.839 0.794 0.710 0.968 0.452
A14 3.600 0.542 1.355 1.290 1.290
A18 6.335 4.503 4.516 4.581 4.323
A22 5.168 3.535 3.613 4.645 4.710
X4 95.25@ 60.20 55.83 53.06 56.13
Y4 95.25@ 125.22 110.67 109.52 106.17
Z4 254.0@ 248.16 245.97 245.82 249.43
X8 254.0@ 69.85 35.89 37.85 38.35
Y8 254.0@ 244.86 206.10 212.06 204.22
Weight (N) 1,019.73 569.89 590.87 559.09 539.98
Note: @ denotes coordinate is stationary.

36 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 Angeline, P. J., and Kinnear, K. E. Jr., eds. (1996). Advances in genetic
programming II. MIT Press, Cambridge, Mass.
Camp, C., Pezeshk, S., and Cao, G. (1998). ‘‘Optimized design of two-
dimensional structures using genetic algorithm.’’ J. Struct. Engrg.,
ASCE, 124(5), 551–559.
Goldberg, D. E. (1989). Genetic algorithms in search, optimization and
machine learning. Addison-Wesley, Reading, Mass.
Haftka, R. T., Le Riche, R., and Harrison, P. (1996). ‘‘Genetic algorithms
for the design of composite panels.’’ Emergent computing methods in
engineering design: Application of genetic algorithms and neural net-
work, D. E. Grierson and P. Hajela, eds., Springer, Berlin, 10–29.
Hajela, P. (1990). ‘‘Genetic search: An approach to the non-convex op-
timization problem.’’ AIAA J., 28(7), 1205–1210.
Hajela, P., Lee, E., and Lin, C. Y. (1993). ‘‘Genetic algorithms in struc-
tural topology optimization.’’ Topology design of structures, NATO
ASI Series, M. P. Bendsoe and M. Soares, eds., Kluwer Academic,
Dordrecht, The Netherlands, 117–133.
Hajela, P., and Lee, E. (1995). ‘‘Genetic algorithms in truss topological
optimization.’’ Int. J. Solids and Struct., 32(22), 3341–3357.
Hansen, S. R., and Vanderplaats, G. N. (1990). ‘‘An approximation
method for configuration optimization of trusses.’’ AIAA J., 28(1),
FIG. 7. Final Topology and Geometry of 25-Bar Truss 161–172.
Holland, J. H. (1975). Adaptation in natural and artificial systems. Uni-
structural optimization problems. The encoding approach pro- versity of Michigan Press, Ann Arbor, Mich.
Jenkins, W. M. (1991). ‘‘Towards structural optimization via the genetic
vided by this study is an ideal 1-to-1 mapping, which makes algorithm.’’ Comp. and Struct., 40(5), 1321–1327.
it easy to translate the trusses into the GP parse trees and vice Jenkins, W. M. (1992). ‘‘Plane frame optimum design environment based
versa. This thus makes it possible to apply GP to structural on genetic algorithms.’’ J. Struct. Engrg., ASCE, 118(11), 3103–3112.
optimization and guarantees feasibility of GP runs. Although Jenkins, W. M. (1996). ‘‘A genetic algorithm for structural design opti-
the presented encoding approach is illustrated using trusses, it mization.’’ Emergent computing methods in engineering design: Ap-
is a general encoding strategy and can be applied conveniently plication of genetic algorithms and neural network. D. E. Grierson and
P. Hajela, eds., Springer, Berlin, 30–53.
to other types of structures, such as frames, planar structures, Kinnear, K. E. Jr., ed. (1994). Advances in genetic programming. MIT
plates, and shells. Press, Cambridge, Mass.
The two illustrative examples reveal that the proposed GP- Kirsch, U. (1990). ‘‘On singular topologies in optimum structural de-
based search strategy is capable of solving simultaneous siz- sign.’’ J. Struct. Optimization, 2, 133–142.
ing, geometry, and topology optimization problems. The re- Koumousis, V. K., and Georgious, P. G. (1994). ‘‘Genetic algorithms in
sults obtained using the proposed GP approach are the best discrete optimization of steel truss roofs.’’ J. Comp. in Civ. Engrg.,
ASCE, 8(3), 309–325.
among all the previous investigations of the same problems Koza, J. R. (1992). Genetic programming: On the programming of com-
using the conventional optimization methods and the GA- puters by means of natural selection. MIT Press, Cambridge, Mass.
based search approach. Koza, J. R., Goldberg, D. E., Fogel, D. B., and Riolo, R. L., eds. (1996).
The proposed GP-based method is a general search and op- Genetic programming 1996: Proc., 1st Ann. Conf. MIT Press, Cam-
timization approach. The main purpose of this study is to in- bridge, Mass.
troduce GP into general civil engineering problem solving and Koza, J. R., et al., eds. (1997). Genetic programming 1997: Proc., 2nd
Ann. Conf., Morgan Kaufmann Publishers, San Francisco, Calif.
not just confine its application to structural optimization. The Luke, S., and Spector, L. (1997). ‘‘A comparison of crossover and mu-
writers believe that this interesting and powerful methodology tation in genetic programming.’’ J. R. Koza et al., eds. Genetic pro-
has potentially wide application in areas such as conceptual gramming: Proc., 2nd Ann. Conf., Morgan Kaufmann Publishers, San
design, project scheduling, and system management. Francisco, Calif., 240–248.
It is worth mentioning that the emphasis of this paper is to Rajeev, S., and Krishnamoorthy, C. S. (1992). ‘‘Discrete optimization of
introduce a new search method and to illustrate how to use it. structures using genetic algorithms.’’ J. Struct. Engrg., ASCE, 118(5),
1233–1250.
Related issues, such as why we need GP when we already Rajeev, S., and Krishnamoorthy, C. S. (1997). ‘‘Genetic algorithm-based
have GAs, will be addressed in a separate paper. In this forth- methodologies for design optimization of trusses.’’ J. Struct. Engrg.,
coming paper, we will demonstrate that GP is in fact better at ASCE, 123(3), 350–358.
solving those problems in which neither the model nor the Soh, C. K., and Yang, J. P. (1996). ‘‘Fuzzy controlled genetic algorithm
parameters of solutions are known in advance: for example, for shape optimization.’’ J. Comp. in Civ. Engrg., ASCE, 10(2), 143–
the optimum design of structures. The investigation of GP’s 150.
Vanderplaats, G. N., and Moses, F. (1972). ‘‘Automated design of trusses
applications in structural optimum design actually shows that for optimum geometry.’’ J. Struct. Div., ASCE, 98(3), 671–690.
GP is more flexible and has a higher search efficiency than Yang, J. P. (1996). ‘‘Development of genetic algorithm based approach
GAs. for structural optimization,’’ PhD thesis, Nanyang Technological Uni-
versity, Singapore.
APPENDIX. REFERENCES Yang, J. P., and Soh, C. K. (1995). ‘‘An integrated shape optimization
approach using genetic algorithms and fuzzy rule-based system.’’
Adeli, H., and Cheng, N. T. (1993). ‘‘Integrated genetic algorithm for Proc., 4th Int. Conf. on Application of Artificial Intelligence in Civ. and
optimization of space structures.’’ J. Aerosp. Engrg., ASCE, 6(4), 315– Struct. Engrg., B. H. V. Topping, ed., Civil-Comp Press, Edinburgh,
328. Scotland, 171–177.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 37

 FINITE STATE VERIFICATION OF INTELLIGENT
TRANSPORTATION SYSTEMS
By John W. Baugh Jr.,1 Associate Member, ASCE, and GopalaKrishna Reddy Kakivaya2

ABSTRACT: Intelligent transportation systems (ITS) represent a major initiative to improve roadway and travel
conditions, not by building roads, but by drawing on new technologies in computing, sensing, and communi-
cations. However, these technologies are extremely complex and, while some demonstration projects have tested
stand-alone systems, the broader architectural and standardization issues are still under development. The use
of complex technologies in a safety-critical system such as ITS and at the scale envisioned requires a very
careful study of the interaction between its various subsystems. This paper describes an approach for modeling
the temporal aspects of ITS systems as well as their logical coordination using finite state verification tools. The
approach is demonstrated by verifying safety properties of an automated traffic signal at a road intersection.

INTRODUCTION emissions testing and mitigation. The full collection of user

services is the basis for the National ITS Architecture (ITS
Traffic congestion has become an everyday phenomenon of 1998), a framework developed for U.S. DOT by Lockheed
major urban areas and the problems are only getting worse. Martin and Odetics that provides a common structure for the
Conventional approaches, such as building and widening design of intelligent transportation systems.
roads, are both economically and environmentally costly, and To implement user services, the architecture defines a struc-
are often impractical in dense metropolitan areas. The goals tured set of functions that must be performed, where functions
of the Intermodal Surface Transportation Efficiency Act of are organized by the physical entities or subsystems in which
1991 (ISTEA) are clearly presented in its policy statement they reside, such as in vehicles or on the roadside. It also
(ITS 1994). defines the information flow between subsystems, together
with any communication requirements, e.g., wireline or wire-
to develop a National Intermodal Transportation System less. The architecture is organized as follows:
that is economically efficient and environmentally sound,
provides the foundation for the Nation to compete in the
global economy, and will move people and goods in an • Logical Architecture—Defines the functions or processes
energy efficient manner. that are needed to perform ITS user services along with
data flows that characterize the information exchanged be-
tween processes. The logical architecture is not dependent
As there is no single solution for achieving the above objec-
on specific technologies, and therefore supports widely
tive, a group of technologies, joined together and collectively
varied system designs.
known as Intelligent Transportation Systems (ITS), forms the
• Physical Architecture—Provides a physical representa-
basis for improving safety, mobility, and efficiency, and re-
tion of the architecture by structuring the processes and
ducing congestion and environmental impacts. Implementation
data flows into subsystems and architecture flows. Phys-
of such systems is based on new technologies in computing,
ical entities outside of the definition, e.g., drivers, pedes-
sensing, and communications.
trians, and traffic, are referred to as ‘‘terminals.’’ While
subsystems collect related processes into large groups,
Architecture equipment packages collect them into smaller, deploy-
After careful planning at the national level, a number of ITS ment-sized pieces.
capabilities—‘‘user services’’—have been identified by U.S. • Implementation Strategy—Consists of various market
DOT and ITS America (ITS 1998). The user services have in packages, e.g., surface street control, that provide the de-
turn been bundled into six broad categories: sired service through a collection of various subsystems,
equipment packages, terminators, and architecture flows.
• Standards Requirements—Define the data exchange pro-
• Travel and Transportation Management
tocols for each physical architecture interface.
• Public Transportation Operations
• Electronic Payment
• Commercial Vehicle Operations The National ITS Architecture provides overall guidance in
• Emergency Management ITS design that results in a number of benefits. It allows dif-
• Advanced Vehicle Control and Safety Systems ferent implementations of systems, products, and services that
adhere to the interfaces to be integrated seamlessly. It also
For example, the first category, Travel and Transportation fosters competition among different system vendors, thereby
Management, contains nine user services, including traffic reducing overall costs and offering choice in the implemen-
control, en-route driver information, incident management, and tation and usage of the system.

1
Assoc. Prof., Dept. of Civ. Engrg., North Carolina State Univ., Ra- Verification of Safety Properties
leigh, NC 27695.
2
Microsoft Corp., Redmond, WA 98052. A well-defined architecture is precise, unambiguous, and
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. complete. It further guarantees that a system following the
To extend the closing date one month, a written request must be filed prescribed protocols does not deadlock, is safe and fair, and
with the ASCE Manager of Journals. The manuscript for this paper was
submitted for review and possible publication on October 22, 1998. This
meets specified timing requirements. The activities required to
paper is part of the Journal of Computing in Civil Engineering, Vol. prove such properties about a given system are collectively
14, No. 1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0038–0050/ termed ‘‘verification,’’ and cannot be performed manually be-
$8.00 ⫹ $.50 per page. Paper No. 19502. cause of the large number of possible system states to be con-
38 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
sidered. Even computational approaches to finite state verifi- der information (Bryant 1986; Clarke et al. 1989; Kurshan
cation can be intractable, although recent advances have 1994). Only more recently have tools based on these algo-
enabled the verification of very large systems. Unlike tradi- rithms become available to researchers outside this community
tional discrete-event simulation, finite state verification is com- of theoretical computer scientists. Some of the more prominent
plete in the sense that all possible behaviors, or traces, are tools include the Concurrency Workbench, developed at Ed-
explored and shown correct. In other words, every combina- inburgh and North Carolina State University (Cleaveland and
tion of values for undetermined variables, e.g., timing values, Sims 1996), SMV at Carnegie Mellon University, HyTech at
is implicitly checked, as opposed to only a subset of combi- Cornell University, and SPIN at AT&T Bell Labs.
nations à la Monte Carlo simulation. As might be expected of a relatively new area of study, most
A mathematical statement of the verification problem is as of the case studies in finite state verification have been per-
follows: formed by its researchers. A good overview of these studies,
as well as those in formal methods generally, is given by
Given a system I and its specification S, show that I 兩= S, Clarke and Wing (1996). In civil engineering, a number of
where 兩= is read ‘‘satisfies.’’ verification studies have been performed over the last decade
at North Carolina State University. These studies have con-
Informally, a computing (and communication) system is mod- tributed to an understanding of the role of finite state verifi-
eled as a ‘‘machine’’ that is compared with a behavioral spec- cation in practical applications, as well as approaches such as
ification using a verification tool, as shown in Fig. 1. This semantic minimization and observer processes for verifying
approach is completely automatic in the sense that it is un- complex systems with as many as 1019 states (Elseaidy et al.
necessary to ‘‘give hints’’ to make a proof go through; i.e., 1997). These include the following:
finite state verification is decidable. The verifier simply returns
‘‘true’’ when the specification is satisfied or an error trace
showing actions by the machine that violate the specification. • Active Structural Control—Numerous structural control
Two of the more common approaches to finite state verifi- studies have been performed since the idea was first in-
cation are based on the form of the specification: troduced in the 1970s. The intent is to enhance the ability
of buildings, bridges, transmission towers, and other
• Equivalence-based verification—Both the system and the structures to respond to natural or other disasters, such as
specification are represented as finite state machines. I 兩= strong earthquakes, severe wind, or even terrorist attacks.
S when the system has the same behavior as the specifi- In a series of studies, fault-tolerant multiprocessor sys-
cation; thus, the 兩= relation is a notion of behavioral tems for active structural control have been modeled and
equivalence. Equivalence-based approaches accommodate verified using finite state approaches (Baugh and Elseaidy
a modular approach to verification, where equivalent-but- 1995; Elseaidy et al. 1996, 1997).
simpler machines can be substituted one for another. • Product and Process Data Modeling—Several groups in
However, the specification must capture in its entirety all the Architecture, Engineering, and Construction (AEC) in-
of the properties to be realized by the system. dustry are studying issues in the integration, exchange,
• Model checking—The system is represented as a finite and standardization of information generated during a
state machine and the specification as a formula in a tem- product’s life cycle. Information models are an attempt to
poral logic. I 兩= S when the system satisfies the temporal represent products and processes in a computer-readable
logic formula. In contrast to equivalence-based ap- form that is clear and precise, and that can be subjected
proaches, model checking accommodates the verification to syntactic and semantic checks. Like activity networks,
of ‘‘pieces of a specification’’ at a time, but makes no finite state models can be used for project planning and
provision for modular verification. management, with the added benefit that they readily sup-
port repetitive cycles, synchronization, communication,
In this study we use an equivalence-based approach to veri- and the coordination of events (Chadha et al. 1994; Baugh
fication, overcoming difficulties in giving a complete specifi- and Chadha 1997).
cation through the use of observer processes (Elseaidy et al. • Distributed Systems for Finite Element Analysis—The
1996, 1997), which is subsequently illustrated. prevalence of local area networks has made the combined
processing power of interconnected workstations a com-
Background and Related Work pelling alternative to mainframe computers for compute-
intensive analyses. Managing computation in distributed
In the mid-1980s and early 1990s, finite state modeling systems, however, is considerably more complex than it
emerged as a viable approach to verification with several is in conventional sequential programming. Finite state
breakthroughs that address the so-called ‘‘state explosion’’ models have been used to design and verify the correct-
problem of parallel composition. These include the use of bi- ness of distributed algorithms for structural analysis
nary decision diagrams, localization reduction, and partial or- (Chadha et al. 1999; Baugh and Chadha 1999).

FIG. 1. Finite State Verification

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 39

 In addition to the above literature, a more general view of the replacement of an LTS with an equivalent one so that ref-
verification for control systems in civil engineering can be erential transparency is maintained, allowing one to substitute
found in an article by Baugh (1998). ‘‘equals for equals.’’ The traditional language equivalent of
In this paper, our goals are to introduce the fundamental automata theory does not preserve the nondeterministic nature
aspects of finite state verification and to illustrate its use in of an LTS, with the result that two language-equivalence LTSs
modeling and verifying a simple ITS component. In doing so, can behave differently within the same environment. For ex-
we hope to inform the civil engineering community not only ample, an LTS may offer an action at some point of compu-
of modern approaches for verification, but of the role they can tation that another language-equivalent LTS is incapable of
play in an important area of information technology with offering. Bisimulation equivalence (Milner 1989), as defined
which we must be prepared to work. below, remedies this situation by comparing the branching
Before describing the verification of a fully actuated traffic structure of the two LTSs.
signal we give an introduction to finite state modeling using
labeled transition systems and process algebras. Definition 2 Let 具Q, Act, →典 be a transition system. The
binary relation R 僆 Q ⫻ Q is a bisimulation if q1Rq2 implies
LABELED TRANSITION SYSTEMS (LTSs) AND that
BISIMULATION EQUIVALENCE
a a
An LTS is defined by a finite number of states and a set of 1. if q1 → q⬘, 2 such that q2 → q⬘
1 then there is a q⬘ 2 and
actions that cause a transition between states. When repre- q⬘Rq⬘
1 2
a a
sented as a directed graph, nodes denote states, and edges, 2. if q2 → q⬘, 1 such that q1 → q⬘
2 then there is a q⬘ 1 and
labeled by actions, denote transitions. Formally: q⬘Rq⬘
1 2

Definition 1 A labeled transition system is a triple 具Q, Act, Hence, two states are bisimulation equivalent when they have
→典 where: transitions to derivative states that are themselves bisimulation
equivalent. As an example, consider system A in Fig. 3: states
• Q is the set of states q and r are bisimilar because they both have c-derivatives to
• Act is the set of actions states s and t, which are bisimilar as they both have d-deriv-
• → 僆 Q ⫻ Act ⫻ Q is the transition relation atives to state u, which is bisimilar to itself. States q and r can
be collapsed, as can states s and t, leading to system B, which
a
A transition a from state q1 to state q2 is written q1 → q2 , is bisimulation equivalent to the original system [Fig. 3(a)].
where q2 is referred to as an a-derivative of q1. An LTS with When states q1 and q2 in a transition system 具Q, Act, →典
a distinguished ‘‘start’’ node is referred to as a rooted LTS are bisimulation equivalent, we write q1 ⬃ q2 . Two rooted
and is defined by the quadruple 具Q, Act, →, q0典, where q0 僆 labeled transition systems 具Q1, Act, →1, q1典 and 具Q2 , Act, →2 ,
Q is the start state. q2典 satisfying Q1 艚 Q2 = 0/ are bisimulation equivalent when
An LTS describes a finite state process by showing the se-
quencing of ‘‘steps’’ in the computation. For example, con-
sider the simple LTSs of Fig. 2. System A begins in state p
and is capable of moving to state q on a and then back to p
on x, yielding the behavior a, x, a, x, etc. Likewise, system B
begins in state r and is capable of moving to state s on y and
then back to r on b, yielding the behavior y, b, y, b, etc.
Although an LTS is required to be in exactly one state at any
given time, systems can be combined with an interleaving se-
mantics as illustrated in system C, which is the composite of
systems A and B. By interleaving, we simply mean that tran-
sitions cannot happen simultaneously; hence, there is no direct
transition between state 具 p, r典 and state 具q, s典.
Most verification techniques for finite state systems rely on
the ability to compare systems for equivalence. When doing
so, the notion of equivalence should be strong enough to allow FIG. 3. Bisimulation Equivalent Systems

FIG. 2. Labeled Transition Systems

40 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

their respective start states, q1 and q2 , are bisimulation equiv- agents. Systems may be described by defining a set of com-
alent in the labeled transition system 具Q1 傼 Q2 , Act, →1 傼 puting agents ᏼ, and a set of actions Act, that can be per-
→2典. formed. For example, the notation

REPRESENTATION OF PROCESSES P = a.Q

Most automatic equivalence-checking algorithms work on denotes an agent P that first performs an action a and then
labeled transition systems, as described above. Direct appli- behaves like the agent Q. For this system, ᏼ = {P, Q} and
cation of this approach would imply the sequence of steps Act = {a}. More interesting behavior can be obtained through
shown in Fig. 4 for verification: Envision a control system, recursion. For example, the agent
devise a finite state machine to model it, enumerate all possible
system states (presumably using a verification tool), and show R = a.b.R
that the traces thereby generated are acceptable. performs the (potentially) infinite sequence of actions a, b, a,
Modeling directly with labeled transition systems, however, b, a, b, . . . , . Given the direct relationship between CCS and
can be both unwieldy and error prone. An alternative is to use labeled transition systems, agent R can be shown to correspond
a more abstract language for representing processes, where a to the LTS in Fig. 5, which is structurally similar to system A
precise semantics is given to each construct so that an equiv- in Fig. 2. By using relabeling to rename actions, a convenient
alent LTS can be generated automatically from the process notational device, agent R can be transformed into an agent
definition. Researchers have devised a number of such lan- that is actually equivalent with system A:
guages, e.g., Calculus of Communicating Systems (CCS) (Mil-
ner 1989), CSP (Hoare 1985), and LOTOS (Bolognesi and R[x/b] ⇔ R = a.x.R
Brinksma 1987), for representing processes. Of these, CCS is
perhaps the best known. where the notation [x/b] substitutes x for b in the definition of
In CCS, processes perform actions and are referred to as R.
Note that in using finite state systems we abstract from some
behaviors and focus on others. In the foregoing example, R
might be an iterative server that accepts and then processes a
connection repeatedly, where a stands for accepting a connec-
tion and b stands for processing it. Alternatively, R might rep-
resent the familiar instructions for using a shampoo: wash,
rinse, repeat. In either case, we abstract from the details of the
actions and focus on their coordination, i.e., a is followed by
b in a repetitive manner.
Of course, languages like CCS allow one to describe and
reason about concurrent systems. Two agents can act in par-
allel, e.g., P兩Q, which is referred to as composite behavior.
Parallel agents can communicate using an action and its com-
plement, represented as a and ā, using the notation (P 兩 Q)\{a}.
For example, the agents Server and Client in Fig. 6 commu-
nicate synchronously over channel x; we adopt the convention
for communication that an action (x) represents reading from
a channel, and that its complement (x̄) represents writing to a
channel. Note the reduction in available transitions from those
of Fig. 2(c) due to agent communication. When two agents
communicate over a channel they synchronize, creating what
is referred to as an internal action, or ␶. This internal action
allows two systems to be compared based only on their ex-

FIG. 4. Finite State Verification FIG. 5. Agent R as Labeled Transition System

FIG. 6. CCS Agents

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 41

ternal actions, e.g., a and b, and ignoring any internal com- alence (Cleaveland and Hennessy 1993; Cleaveland et al.
munication. 1993) on appropriately transformed transition systems.
Intuitively, the server performs some work a, the results of
which are output to the client over channel x; the client then Observational Equivalence
performs some work b, and the process repeats itself resulting
in an infinite number of possible traces, for example For our purposes, it is desirable to compare systems based
only on their observable actions, ignoring any internal com-
munication ␶ between components. Thus, a specification is
• a, ␶, b, a, ␶, b, a, ␶, b, a, . . .
considered to be an abstraction of desired behavior and, con-
• a, ␶, a, b, ␶, a, b, ␶, a, b, . . .
sequently, does not contain any ␶ that might arise in imple-
• a, ␶, b, a, ␶, a, b, ␶, b, a, . . .
menting the specification via communicating processes. A
• . . . and so on.
weaker approach to equivalence that captures this notion is
referred to as observational equivalence (Milner 1989), or
More formally, the set of atomic actions Act in CCS consists weak bisimulation. Before defining weak bisimulation, we de-
of observable actions ⌳ and a distinguished silent (internal) fine a weak counterpart to the transition relation.
action ␶ that represents process synchronization. The observ-
able actions are divided into a set of input actions Ꮽ and
Definition 3 A weak counterpart to the transition relation,
output actions Ꮽ̄ that produce an internal action ␶ when syn-
⇒, is defined as
chronizing. Assume that ␣ ranges over Act and a ranges over
⌳. Then, given agent expressions E and F and an agent vari- a ε a ε
1. pε ⇒ q if p⇒→⇒q
able X, the language of CCS is defined by the following gram- ␶
2. ⇒ ⬅ (→)*
mar:
ε
where epsilon reachability
a
(⇒) is a series of zero or more ␶,
E := nil termination and the transition p ⇒ q means there may be intermediate ␶
兩 ␣.E (␣ 僆 Act) prefixing transitions that are simply ignored. The definition of weak bi-
兩 E ⫹ E external choice simulation follows directly from these.
兩 E兩E parallel composition
兩 E \ L (L 債 ⌳) restriction Definition 4 Let 具Q, Act, →典 be a transition system. The
兩 E[ f ] ( f :⌳ → ⌳) relabeling binary relation R 僆 Q ⫻ Q is a weak bisimulation if q1Rq2
兩 X process invocation implies
兩 fix(X = E ) recursive definition
a a
1. if q1 → q⬘,
1 then there is a q⬘2 such that q2 ⇒ q⬘2 and
Several operations in CCS warrant further comment. External q⬘Rq⬘.
1 2
choice, as in P = a.Q ⫹ b.R, denotes a system that either a
2. if q2 → q⬘,
2
a
then there is a q⬘1 such that q1 ⇒ q⬘1 and
performs a and behaves like Q, or performs b and behaves q⬘Rq⬘.
1 2
like R; the criterion on which the decision is made is left
unmodeled. The relabeling function f is required to satisfy two Given a labeled transition system 具Q, Act, →典, the trans-
conditions: f (␶) = ␶ and f (ā) = f (a). The fix expression allows formed transition system 具Q, Act, ⇒典 is computed by applying
the definition of recursive processes: X is a bound variable in the following transformations in order:
the fix expression and can be replaced with any other variable
␶ ε
through ␣-conversion. An agent expression containing no free 1. Replace p (→)* q by p ⇒ qa
ε a≠␶ ε
variables is an agent. ᏼ is the set of agents over which P and 2. Replace p ⇒ → ⇒ q by p ⇒ q
Q range.
ε
The agent expressions are given ␣
an operational semantics in where ⇒␶ can be computed as the transitive and reflexive clo-
terms of the transition relation →, where ␣ is an action. The sure of →, and ⇒
a
by the relational composition of ⇒
ε
and
a
transition relation is defined inductively based ␣on the structure →. We then apply the standard algorithms for computing bi-
of
␣
the agent expression. For example, if E → E⬘, then E兩F simulation on the transformed transition system 具Q, Act, ⇒典;
→ E⬘兩F. Formally, this transition is defined by the following hence, determining the observational equivalence of the orig-
rules: inal transition system 具Q, Act, →典.
␣ ␣ ␣ ␣
¯ Two states, q1 and q2 , in a transition system 具Q, Act, →典 are
E → E⬘ F → F⬘ E → E⬘ F → F⬘
; ; observationally equivalent if there is a weak bisimulation re-
␣ ␣ ␶
E兩F → E⬘兩F E兩F → E兩F⬘ E兩F → E⬘兩F⬘ lating them. When q1 and q2 are observationally equivalent we
write q1 ⬇ q2 . Two rooted labeled transition systems 具Q1, Act,
where the top of each rule is a hypothesis and the bottom a →1, q1典 and 具Q2 , Act, →2 , q2典 satisfying Q1 艚 Q2 = 0/ are
conclusion. For example, the first rule states that the parallel observationally equivalent when their respective start states q1
composition of agents E and F has an ␣-derivative E⬘兩F if E and q2 are observationally equivalent in the labeled transition
has an ␣-derivative E⬘. The original work by Milner (1989) system 具Q1 傼 Q2 , Act, →1 傼 →2典.
contains a complete description of transition rules for the lan- When agents P and Q are strong bisimulation equivalent we
guage, which can be used to generate an equivalent labeled write P ⬃ Q, and when they are observationally equivalent
transition system for a given agent. we write P ⬇ Q.

APPROACH TO MODELING AND VERIFICATION Timed CCS

Bisimulation equivalence between labeled transition sys- Because we wish to reason about both coordination and tim-
tems is isomorphic to the notion of strong bisimulation in ing properties, a timed process algebra, TCCS (Moller and
CCS, meaning that systems modeled in CCS can also be tested Tofts 1990), is used to model and verify the system’s correct-
for equivalence. Other behavioral equivalences, including ob- ness. TCCS is a variant of CCS that enables the specification
servational equivalence, language equivalence, and testing of discrete time as actions, resulting in an expressive set of
equivalence, can be computed in terms of bisimulation equiv- timed behaviors. In this variant, an action d 僆 Nat represents
42 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
a delay of d time units. For example, P = 1.Q is an agent that
idles for one unit of time before behaving like Q.
More generally, TCCS constructs may be understood in
terms of the communication actions and units of delay (or
idling) they may engage in. Nondelay actions are assumed to
be durationless; so time passes only when processes are ca-
pable of ‘‘idling.’’ If a 僆 Nat, then a.P is a process that im-
mediately engages in action a and then behaves like P; it is
incapable of idling. If a 僆 Nat, then a.P idles for a time units FIG. 8. Comparison of Commands
and then behaves like P. A process that may either delay for
an arbitrary amount of time or behave as P is denoted ␦P, i.e., CWB-NC and the foregoing mathematical notations include
␦P = P ⫹ 1.␦P. For example, the process 5.␦c̄.P can output the ASCII symbols t, (’), and $, which replace ␶, (⫺), and ␦,
on channel c after at least five time units have elapsed (pos- respectively. The examples shown in Fig. 8 highlight some of
sibly longer) and then behave like P. As it does in CCS, P1兩P2 these differences.
represents the parallel composition of P1 and P2 . For the com-
posite system to idle, both components must be capable of TRAFFIC SIGNAL CONTROL
idling.
The following set of timed behaviors can be described in Traffic control is a user service contained in the Travel and
TCCS in a straightforward manner: Transportation Management bundle. The National ITS Archi-
tecture addresses traffic control with processes residing pri-
• Basic time delays: The agent P = 1.c.P performs an in- marily in the traffic management and roadway subsystems of
finite sequence of c actions separated by a unit time delay, the physical architecture. Within these subsystems are several
or 1.c.1.c.1.c . . . equipment packages related to signal control, including:
• Time-outs: The agent P ⫹ 1.Q enables Q after one unit
of time if P cannot execute immediately. • Roadway Signal Controls—Controls traffic signals at ma-
• Nondeterministic durations: The agent a.(␶.1.b.nil ⫹ jor intersections and on main highways in urban areas.
␶.2.b.nil ) enables b between 1 and 2 time units after a is • TMC Signal Control—Allows personnel in the Traffic
executed. Management Center (TMC) to monitor and manage the
traffic flow at signalized intersections.
A subtle use of ␶ in the prior example should be noted: As • TMC Regional Traffic Control—Provides capabilities for
a prefix in both legs of the nondeterministic choice operator, analyzing, controlling, and optimizing area-wide traffic
␶ defeats a maximum progress assumption (Elseaidy et al. flow.
1996) of TCCS that is otherwise useful in modeling timed
systems. The notation of ‘‘maximum progress’’ states that a The roadway equipment package provides basic traffic sig-
TCCS agent capable of making progress should not idle in nal control (Orcutt 1993), which may be either pretimed or
time, that is, ␶.P ⫹ 1.Q ⬇ ␶.P holds. The original work by actuated. Pretimed traffic signals assign right-of-way to differ-
Moller and Tofts (1990) gives a detailed description of the ent phases using a fixed sequence with fixed intervals. Actu-
language’s semantics. ated traffic signals have the ability to alter their sequences and
to skip phases on which no demand is registered by traffic
sensors. Actuated traffic signals that have flexibility in altering
Concurrency Workbench
timing on all their phases are termed fully actuated, while
To define and verify TCCS models we use the Concurrency those having at least one fixed-timing phase are termed semi-
Workbench of North Carolina (CWB-NC) (Cleaveland and actuated.
Sims 1996). The CWB-NC is an automated tool for verifying Fully actuated traffic signals are suitable for intersections
finite state systems. A number of process algebraic formalisms that experience large variations in traffic volumes on all ap-
are supported by the tool, including TCCS, and decision pro- proaches during the day. The following terminology is useful
cedures are also implemented for calculating various behav- in understanding the operation of a fully actuated traffic signal.
ioral equivalences and preorders.
The CWB-NC commands shown in Fig. 7 are used in the • Minimum green time: The minimum amount of time the
following discussion. Other syntactic differences between the controller allows a phase to remain green before changing

FIG. 7. Concurrency Workbench Commands

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 43

 the phase. It should be sufficient to allow stopped vehicles package, which allows us to verify its operation with respect
time to move forward and clear the intersection; rear-end to safety-critical and other requirements. The model includes
and right-angle accidents may result otherwise. traffic flow, vehicle detection, a controller, and a signal indi-
• Maximum green time: The maximum amount of time the cator. Other features of the roadway package such as pedes-
controller allows a phase to remain green after changing trian sensors, equipment fault monitors, and variable message
the phase. This is not a safety parameter, but is intended signs are omitted for brevity.
to mitigate sensor failures.
• Unit extension: The amount of green time extended by Traffic Flow
the controller after the last vehicle movement. This also
is not a safety parameter. Rather, it controls the amount Four different types of traffic flow correspond to the four
of time that is tolerated between successive vehicles. A phases of a traffic signal, as shown in Fig. 9. Because traffic
high value for unit extension makes the signal operate flow represents infinite behavior in time, its model should per-
sluggishly. mit both continuous and intermittent flow. The TCCS agent
• Amber time: The amount of time the signal displays am- Flow shown in Fig. 10 models a typical traffic flow by mak-
ber before displaying red, which is intended to allow ve- ing a nondeterministic choice of either having or not having
hicles time to clear the intersection. It should allow a vehicle approach the intersection at the current time. An
enough time for a vehicle entering the intersection while approaching vehicle is modeled by three successive actions
green to clear it. that synchronize with other components in the model:
• All-red interval: The amount of time the controller dis-
plays red on all approaches between phase changes. It • ’wait—trips the sensor to indicate its presence at the
should be sufficient to allow a vehicle traveling at design approach
speed to come to a stop at the intersection after seeing the • green—models the driver waiting on the controller to
phase change from green to amber. make the signal green
• ’move—indicates that the vehicle is passing through the
Control of an actuated traffic signal requires that the pres- intersection
ence or absence of a vehicle be determined. Pulse sensors issue
a pulse of short duration upon detection of a vehicle. The An arrival spacing of 1 unit of time between successive
duration of the pulse is independent of vehicle speed, length, vehicles is imposed on the approach. Absence of a vehicle is
and continued presence within the loop. In contrast, continu- modeled by simply having time pass, i.e., 1.1.1.1 . . . . The
ous sensors issue a continuous signal as long as a vehicle is flows corresponding to the four phases of the traffic signal are
within the loop. instantiated using the TCCS relabeling operator on the agent
As an example, consider the operation of a fully actuated, Flow. As noted above, including t as a prefix in both legs of
two-phase traffic signal at an intersection with four ap- the nondeterministic choice operator defeats the maximum
proaches. Let phase A be assigned the north-south direction progress assumption of TCCS. Without t, then, MakeChoice
and phase B the east-west direction. If phase A is green, it would always release a vehicle into the intersection because a
remains green until the minimum green time for the north- vehicle is capable of synchronizing with the sensor, making
south direction has expired and an approaching vehicle in the progress possible.
east-west direction actuates the sensor for phase B. If there is
no further traffic on the north-south direction, the controller Vehicle Detection
changes phase A from green to amber and finally to red. It
then changes phase B to green, which remains so until the Each of the four signal phases receives input from a sensor,
minimum green time for the east-west direction has expired. which detects a vehicle waiting to proceed in the direction
At the expiration of the minimum green time for the east-west being sensed. The TCCS agent Sensor shown in Fig. 11
direction, the controller attempts to change green from phase models a typical sensor as a one-place buffer; it is instantiated
B to phase A only if a sensor for phase A has registered activity
during the minimum green time for the east-west direction. If
the sensor for phase A has not registered any activity, phase
B remains green until either the sensor for phase A registers
activity, or the maximum green time for the east-west direction
has expired, whichever occurs first. At this point, the controller
changes phase B from green to amber and finally to red. It
then changes phase A to green, and the cycle continues.

ROADWAY SIGNAL CONTROLLER

The equipment package for roadway signal control provides
control that is implemented only by local roadside means, i.e.,
without input from the TMC. Even this basic component, how-
ever, carries with it a rather stern warning in its functional
requirements:

provide output data in a form which . . . appears in a safe

sequence, is unambiguous and does not provide conflicting
instructions to drivers and travelers that are likely to result
in circumstances which are life threatening (ITS America
1998)

In this section we present a finite state model of a controller

that implements some of the features present in the roadway FIG. 9. Traffic Signal

44 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 FIG. 10. Traffic Flow Model

FIG. 11. Sensor Model

for each of the four signal phases using relabeling. After de- • Fairness: The controller design should be fair to the driv-
tecting a vehicle, it waits for the controller to synchronize with ers approaching the intersection from different directions.
it. For example, a vehicle waiting on the east-west approach It should be noted that efficiency and fairness are com-
trips SensorEW by synchronizing with wew; the controller peting objectives. In other words, it is possible to improve
samples the sensor by synchronizing with sew. If additional throughput through the intersection by giving excessive
vehicles begin lining up, they are not ‘‘detected’’ until the first priority to major street approaches over minor street ap-
gets a green signal and clears the intersection; in other words, proaches. However, such a design may be considered un-
the agent Sensor models a continuous type sensor. fair. One notion of fairness can be characterized by the
statement: A vehicle at an approach should wait at most
Controller x units of time for a green signal, where x is a design
parameter. In contrast to safety properties, fairness can be
The controller is the ‘‘brain’’ behind the operation of a traf- much more difficult to verify.
fic signal and hence implements its design philosophy. Objec-
tives for traffic controller design include: The traffic signal controller is modeled by the TCCS agent
shown in Fig. 12. Control waits for a minimum green time
• Safety: The controller design should guarantee that a traf- of 2 units before sampling the sensors of other phases in a
fic signal implementing its program is free of conflicts. round-robin manner with increasing delay. For example, when
Freedom from conflict can be characterized by the state- the signal becomes green in the east-west direction, the control
ment: A time interval of x units should separate two well- unit starts sampling the sensors for the north-south, east-south,
behaved vehicles entering the intersection from different and south-west directions after 2, 3, and 4 units of time delay,
approaches, where x is a design parameter. respectively. If a north-south vehicle approaches within 2 time
• Efficiency: The controller design should be efficient in the units, the actions below will synchronize with other compo-
sense that the intersection throughput is maximized. nents:

FIG. 12. Controller Model

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 45

 • sns—senses the presence of a vehicle on the north-south cess is their composition, as shown in Fig. 14. The agent TS
approach represents the traffic signal system; TSChannels is the set
• ’cgns—initiates a change to green in the north-south di- of channels over which the traffic signal components synchro-
rection nize to achieve the desired functionality. The agent Sys rep-
resents the traffic signal in operation; ComChannels is the
After 5 time units the controller is then placed in the set of communication channels over which vehicles interact
ControlNS state. with the signal.
The round-robin strategy of the controller ensures that an
upper bound can be given on the amount of time that a stopped VERIFICATION OF MODEL
vehicle waits for a green signal. The values adopted by the
agent Control for some of the parameters of a fully-actuated Having modeled a roadway signal controller as a finite state
traffic signal are given below: machine in TCCS, we show that the following functional prop-
erties of the system are satisfied:
• Minimum green time = 2
• Maximum green time = ⬁ 1. A red signal separates conflicting movements of vehicles.
• Unit extension = 0 2. A time lag of at least 5 units separating conflicting move-
ments of vehicles.
Signal Indicator 3. The amount of time a vehicle waits for a green signal is
bounded.
This actuator unit responds to commands from the control-
ler, changing the indicator colors from green to amber, red, where the first two properties ensure safety, and the third en-
green, etc. The TCCS agent Signal, shown in Fig. 13, models sures fairness.
this behavior. Upon receiving a command to change the green We wish to use the theory of observational equivalence for
phase from east-west to north-south, for example, the indicator verifying the above properties; doing so requires that a suitable
takes the east-west phase from green to amber and finally to specification be devised in TCCS. Recall that time is repre-
red. It then changes the north-south phase to green and awaits sented in TCCS as a global clock that ‘‘ticks,’’ and these clock
further commands from the controller. As before, these tics are visible actions. Consequently, two TCCS agents are
changes are made via synchronization with other components; observationally equivalent if and only if they have observa-
it should be noted that additional synchronizations are made tionally equivalent action derivatives and time derivatives.
with sensors (e.g., using sew, sns, ses, and ssw) so that Thus, the definition of observational equivalence in TCCS
their registered activity is dropped. The values adopted by Sig- complicates writing specifications that either abstract from
nal for some of the parameters of a fully actuated traffic signal time completely or specify only bounds in time. For instance,
are as follows: an upper bound of four clock ticks between two actions would
be satisfied intuitively by a system with exactly two clock ticks
• Amber time = 3 between those actions, yet these behaviors are not observa-
• All red interval = 2 tionally equivalent.
In the case of Sys, the only visible actions apart from clock
Roadway Signal Controller
ticks are the signal red and vehicle movements mew, mns,
Given TCCS models of the various components of a fully mes, and msw. A specification of the first property restricts
actuated traffic signal, the final step in the development pro- the allowable sequences of visible actions obtained after re-

FIG. 13. Signal Indicator Model

FIG. 14. Traffic Signal Model

46 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

moving clock ticks and, consequently, completely abstracts translates into a deadlock in the agent (Sys 兩 Checker). The
from time. The second property specifies a lower bound on Timer agent monitors the passage of time and helps
the number of intervening clock ticks between the actions rep- Checker in verifying the second property by raising the flag
resenting vehicle movements, i.e., mew, mns, mes, and ’accident when it is violated. The third party is verified by
msw. The third property specifies an upper bound on the modifying the Vehicle agent definition as shown in Fig. 15.
number of clock ticks a stopped vehicle waits for a green The modified vehicle waits for at most 29 time units for a
signal. Thus, the second and third properties merely specify green signal before raising a flag ’tv; the presence of ’tv in
bounds in time without concern for how the bounds are sat- the system trace indicates that a vehicle is waiting for a longer
isfied. Consequently, it is not possible to devise a TCCS ex- period of time.
pression as a specification for Sys.
The inability to write a TCCS specification for Sys could RESULTS
preclude the use of observational equivalence for verification. The complete system, including the checker and specifica-
However, an idea due to Elseaidy et al. (1996) can be ex- tion processes, is shown to be correct using the CWB-NC, i.e.,
ploited in which a Checker process runs in parallel with Sys, (Sys 兩 Checker) ⬇ Spec. Because agent Spec merely mod-
converting its allowable traces (as defined by a specification) els the passage of time, this equivalence proves that Sys is
into t-sequences by synchronizing with certain visible actions. without deadlocks and raised flags. Consequently, our model
Thus, Checker can be thought of as a ‘‘specification ma- of the traffic signal satisfies all three functional properties pre-
chine’’ that looks for allowable sequences of visible actions viously given. Below we describe the steps followed in prov-
with all possible interleavings of clock ticks in between. Tim- ing the system correct.
ing violations can be modeled by having Checker raise a The composition of parallel systems can result in a combi-
flag and/or create a deadlock when such a violation is detected. natorial number of states, making finite state verification dif-
Using this approach, Sys satisfies its specification if and only ficult. Because of this concern we use symmetry arguments to
if (Sys 兩 Checker) is deadlock free. Because the only visible avoid modeling bounded-time vehicles (with BVehicle) on
part of a correct Sys is the passage of time, we are required each approach; rather, all approaches other than FlowEW use
only to show that (Sys 兩 Checker) ⬇ Spec where Spec = the prior definition of Vehicle. Despite this simplification,
1.Spec. agent Sys is too large for the verification of (Sys 兩 Checker)
The TCCS agent Checker that verifies the first and second ⬇ Spec in a single step. The CWB-NC command size Sys
properties is shown in Fig. 15. Violation of the first property gives the following output:

FIG. 15. Checker Model

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 47

 FIG. 16. Order of Composition

command: size Sys 1.2.7.2—Monitor Roadside Equipment Operation for Faults

Building automaton . . . TMS Signal Control
Done. 1.1.2.2—Process Traffic Data
185796 states 1.1.4.2—Provide Traffic Operations Personnel Traffic Data
522475 transitions. Interface
1.2.1—Select Strategy
The above command executes for more than 2 h on a Sun 1.2.2.2—Determine Indicator State for Road Management
workstation and uses in excess of 300 MB of memory. 1.2.4.1—Output Control Data for Roads
Fortunately, verification based on observational equivalence
can be performed in a piecewise manner using semantic min- To enable TMC control, our roadway signal, originally de-
imization (Elseaidy et al. 1997). Using this incremental ap- signed for independent operation, must be modified with re-
proach to verification, the state spaces of the components of spect to its internal behavior and its external interface. Recall
Sys are minimized at intermediate steps of the parallel com- that the controller’s interface to the outside world currently
position. Because minimization is performed with respect to includes eight communication channels: Four to vehicle de-
observational equivalence, which is not a congruence, care tection sensors and four to signal indicators. For coordinated
must be exercised to ensure that the final minimized system operation, the controller must be prepared in addition to com-
is congruent to the unminimized system. Agents simulating municate with the TMC, forwarding traffic sensor data and
flows, for example, have initial t actions, and thus are not processing a predetermined set of ‘‘commands.’’ With the ex-
minimized before composition so that congruence with the un- tended interface, the controller is ready for an update to its
minimized system is preserved. Because these agents interact internal operation that uses command input along with sensor
only with TS and not among themselves, however, they can data to control signal operation.
be composed with TS in any order. The composition order Consider a controller like before, for example, except that
shown in Fig. 16 minimizes the state space while preserving it can run one of two possible ‘‘programs,’’ each with its own
congruence. These commands first minimize TS, compose it set of timing values, e.g., minimum green time, that is appro-
with FlowEW, minimize the result, and so on; thus, we al- priate for a certain part of the day. Let those programs be
ternate between composition and minimization in an incre- denoted by the TCCS agents ControlDay and Control-
mental manner. By doing so, the computational requirements Night, and let commands for executing those programs be
for verifying the entire system are reduced from 2 h to just modeled as communication channels p1 and p2, respectively,
over 1 min. in TCCS. Then, a roadway controller that now accepts com-
mands from the TMC can be defined simply as
TMC SIGNAL CONTROL
proc Control = p1.ControlDay ⫹ p2.ControlNight
The modeling techniques presented in this paper can be ap-
plied to other ITS subsystems and components; here we out- By sending an output on p2, for example, the TMC controller,
line some of the issues in doing so. modeled by an agent TMCcontrol, causes Control to
As an example, a straightforward extension of this study is behave as ControlNight for nighttime operation. Then,
the design and verification of a coordinated system of traffic internal operations of ControlDay and ControlNight can
signals, a capability provided by the equipment package for be used to poll the TMC intermittently to see if a change in
TMC signal control. Fig. 17 shows a data flow diagram for program is desired. The expression p1.ControlDay ⫹
both TMC (shaded) and roadway (unshaded) signal control in 1.ControlNight appearing repetitively within a definition of
which some of the data flows have been omitted for clarity. ControlNight allows the TMC to reset the roadway con-
The relevant processes in the National ITS Architecture are as troller for nighttime operation; if there is no communication
follows: within one time unit from the TMC, the controller continues
with its nighttime program until it again polls the TMC. Al-
Roadway Signal Controls though not modeled, communication costs between the TMC
1.1.1.1—Process Traffic Sensor Data and roadway controllers can be easily accommodated using
1.2.7.1—Process Indicator Output Data for Roads the idling expressions of TCCS.
48 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
 FIG. 17. TMC Signal Control: Data Flow Diagram

In addition to modeling a single system-wide TMC, as signal. In addition, the approach used for modeling and veri-
above, architectural variations can be proposed and compared fication could serve as a template for verifying other safety-
with respect to system performance, communication band- critical aspects of intelligent transportation systems.
width and latency, and component-level bottlenecks before im-
plementation. These variations might include, for example, dif- ACKNOWLEDGMENTS
ferences in the number and topological arrangement of TMCs. The helpful comments of this paper’s anonymous reviewers signifi-
Finally, we note that the models presented in this paper ab- cantly improved the exposition of our ideas. Support for this research by
stract from the details of data transmission. That is, while syn- the North Carolina Department of Transportation is gratefully acknowl-
chronization and communication cost are captured, the form edged.
of the data with respect to handshaking protocols and the mar-
shalling and unmarshalling of data is ignored; finite state ap- APPENDIX. REFERENCES
proaches are amenable to the verification problems that arise Baugh, J. W., Jr. (1998). ‘‘Design and verification of real-time systems.’’
in such contexts (Clarke and Wing 1996). Artificial intelligence in structural engineering, lecture notes in artifi-
cial intelligence 1454, Ian Smith, ed., Springer, New York, 30–47.
CONCLUSIONS Baugh, J. W., Jr., and Chadha, H. S. (1997). ‘‘Semantic validation of
product and process models.’’ J. Comp. in Civ. Engrg., ASCE, 11(1),
Recent advances in finite state verification have made prac- 26–36.
tical a quantitative approach to the design of ITS components Baugh, J. W., Jr., and Chadha, H. S. (1999). ‘‘Distributed computing
and other real-time control systems (Baugh 1998). Its appeal systems.’’ High performance computing, ASCE, Reston, Va.
lies in its similarity to programming, its high degree of auto- Baugh, J. W., Jr., and Elseaidy, W. M. (1995). ‘‘Real-time software de-
mation, and its ability to scale. With respect to the latter, note velopment with formal models.’’ J. Comp. in Civ. Engrg., ASCE, 9(1),
that a complex system can often be replaced by its specifica- 73–86.
Bolognesi, T., and Brinksma, E. (1987). ‘‘Introduction to the ISO speci-
tion, etc. fication language LOTOS.’’ Comp. Networks and ISDN Sys., 14, 25–
In this study, the technical notion of bisimulation equiva- 59.
lence is used to prove the correctness of a roadway signal Bryant, R. E. (1986). ‘‘Graph-based algorithms for boolean function ma-
controller. The controller design in this traffic signal model nipulation.’’ IEEE Trans. on Comp., C-35(8), 677–691.
avoids starvation by using a round-robin approach to sensor Chadha, H. S., Baugh, J. W., Jr., and Wing, J. M. (1994). ‘‘Formal spec-
sampling with increasing delays. Also, the model incorporates ification of AEC product models.’’ Comput. in Civ. Engrg.: Proc., 1st
Congr., K. Khozeimeh, ed., Vol. 1, ASCE, New York, 571–578.
timing behavior and demonstrates the use of observer pro- Chadha, H. S., Baugh, J. W., Jr., and Wing, J. M. (1999). ‘‘Formal spec-
cesses in verifying properties that abstract from time. Because ification of concurrent systems.’’ Adv. in Engrg. Software, 30(3), 211–
the roadway signal controller satisfies the desired properties, 224.
it could serve as a basis for designing a fully-actuated traffic Clarke, E. M., Long, D. E., and McMillan, K. L. (1989). ‘‘Compositional

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 49

 model checking.’’ Proc., 4th Annu. Symp. on Logic in Comp. Sci. (LICS Elseaidy, W. M., Cleaveland, R., and Baugh, J. W., Jr. (1997). ‘‘Modeling
’89), IEEE Computer Society Press, New York, 353–362. and verifying active structural control systems.’’ Sci. of Comp. Pro-
Clarke, E. M., and Wing, J. M. (1996). ‘‘Formal methods: State of the gramming, 29(1–2), 99–122.
art and future directions. Report by the working group on formal meth- Hoare, C. A. R. (1985). Communicating sequential processes. Prentice-
ods for the ACM workshop on strategic directions in computing re- Hall, Englewood Cliffs, N.J.
search.’’ ACM Computing Surveys, 28(4), 626–643. ITS America (1994). ‘‘ITS architecture development program: Phase 1.’’
Cleaveland, R., and Hennessy, M. C. B. (1993). ‘‘Testing equivalence as Tech. Rep., Washington, D.C.
a bisimulation equivalence.’’ Formal Aspects of Computing, 5, 1–20. ITS America (1998). Federal programs: National ITS architecture doc-
Cleaveland, R., Parrow, J., and Steffen, B. (1993). ‘‘The concurrency umentation (Version 2.0). Washington, D.C.
workbench: A semantics based tool for the verification of finite-state Kurshan, R. P. (1994). Computer-aided verification of coordinating pro-
systems.’’ ACM Trans. on Programming Languages and Sys., 15(1),
cesses. Princeton University, Princeton, N.J.
36–72.
Cleaveland, R., and Sims, S. (1996). ‘‘The NCSU concurrency work- Milner, R. (1989). Calculus of communicating systems. Prentice-Hall, En-
bench.’’ Computer-aided verification (CAV ’96), lecture notes in com- glewood Cliffs, N.J.
puter science 1102, R. Alur and T. Henzinger, eds., Springer, New Moller, A., and Tofts, C. (1990). ‘‘A temporal calculus of communicating
York, 394–397. systems.’’ Proc., CONCUR’90, Lecture Notes in Comp. Sci. 458,
Elseaidy, W. M., Baugh, J. W., Jr., and Cleaveland, R. (1996). ‘‘Verifi- Springer, New York, 401–415.
cation of an active control system using temporal process algebra.’’ Orcutt, F. L. (1993). The traffic signal book. Prentice-Hall, Englewood
Engrg. with Comp., 12, 46–61. Cliffs, N.J.

50 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 INTELLIGIS: TOOL FOR REPRESENTING AND REASONING
SPATIAL KNOWLEDGE
By Xudong Jia,1 Member, ASCE

ABSTRACT: This paper presents a method for the representation and reasoning of spatial knowledge. Spatial
knowledge is important to decision making in many transportation applications that involve human judgment
and understanding of the spatial nature of the transportation infrastructure. Unfortunately, conventional Knowl-
edge-Based Expert Systems (KBESs) do not have a robust spatial inference engine to construct spatial knowledge
and deduce solutions. The method discussed in the paper integrated a KBES with geographic information system
(GIS) functions using Transmission Control Protocol/Internet Protocol and Remote Procedure Calling technol-
ogies. A software system (called IntelliGIS) implementing the method has been developed. As a case study, a
KBES-based pavement management system was developed using the IntelliGIS. The case study demonstrated
how IntelliGIS represents and reasons about spatial knowledge and how it provides graphical display capabilities
for representing spatial information and derived solutions.

INTRODUCTION lations, and spatial reference actions involved in spatial knowl-

edge.
Many knowledge-based expert systems (KBESs) have been Integration of geographic information system (GIS) tech-
developed to solve various transportation problems that are nology with a KBES has indicated a promising effort in han-
heavily dependent on human judgment and experience (Richie dling the spatial nature of various problems (Djokic 1991;
and Prosser 1991; Cohn and Harris 1993; Prechaverakul and Johnson and Kanonier 1991; Laurini and Thompson 1992:
Hadipriono 1995; Sarasua and Jia 1995; Spring and Hummer
McKinney et al. 1993; Spring and Hummer 1995). Currently,
1995). These systems use symbolic knowledge to construct
there are two types of integration between a GIS and a KBES:
human understanding of problems in the areas of transporta-
(1) Linking KBES functions to a GIS (KBES-enhanced GIS);
tion planning, design, operation, and maintenance. Because
and (2) linking GIS functions to a KBES (GIS-enhanced
symbolic knowledge is not well suited to describe the spatial
KBES). Researchers at the University of California at Santa
nature of transportation problems, automated reasoning sys-
Barbara carried out research on the former type and developed
tems often lack a mechanism to derive solutions based on spa-
a knowledge-based GIS (Pazner 1986). The system improves
tial knowledge (or knowledge about positional and topological
a standard GIS by incorporating several reasoning capabilities
characteristics) of transportation infrastructure.
in the system. The purpose of the knowledge-based GIS is to
Spatial knowledge is critical to spatial reasoning and deci-
respond intelligently to users’ queries on large spatial data-
sion making in many transportation applications. An ITS in-
cident management system for example may use the following bases stored in the GIS. Other research activities have also
knowledge (or IF-THEN rule) to determine which response involved the development of KBES-enhanced GISs for digital
team the traffic management center should call: image processing and pavement management (Johnson and
Kanonier 1991; Sarasua and Jia 1995). The effectiveness of
enhanced GISs is limited because data between KBES func-
IF an incident is located at the boundary of Los
Angeles county tions and a GIS are often exchanged using text files. In ad-
THEN the Los Angeles county response team should dition, enhanced GISs do not have core modules (or a spatial
be called. inference engine) to handle spatial knowledge representation
and reasoning; even limited KBES functions are linked to the
systems. Djokic focused his research on linking GIS functions
Another application may use the following knowledge to de-
scribe a spatial policy in a transit service planning program: to a KBES and developed a GIS-enhanced KBES for storm-
water and drainage network assessment (Djokic 1991). Later,
other researchers adopted his enhanced KBES for water plan-
IF network distance from Cal Poly, Pomona to its
ning (McKinney et al. 1993). Djokic’s work has created a new
closest bus station is within a half mile
way to handle problems with spatial consideration. However,
THEN access mode is walking
the inference engine in his enhanced KBES is not improved.
The engine may not effectively handle locators, spatial rela-
Clearly, the above two pieces of knowledge contain spatial
tions, and spatial reference actions involved in spatial knowl-
characteristics that are expressed by locators (boundary of Los
edge.
Angeles County, Cal Poly, Pomona, and bus station), spatial
This paper presents a method to develop a GIS-enhanced
relations (located at and within), and spatial inference (walk
KBES with a spatial reference engine for the representation
and call). Unfortunately, current conventional KBESs cannot
and reasoning of spatial knowledge. It begins by summarizing
handle spatial knowledge. They do not have an appropriate
method to encode and represent the spatial nature of knowl- primary components of a KBES in the context of knowledge
edge. Furthermore, they cannot deal with locators, spatial re- representation and reasoning. It also examines various require-
ments of spatial knowledge in a KBES. This paper then dis-
1
Asst. Prof., Dept. of Civ. Engrg., California State Polytechnic Univ., cusses IntelliGIS, an operational system that implements the
Pomona, Pomona, CA 91768. method. The discussion is centered on how the conceptual
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. framework of IntelliGIS supports the inclusion of GIS func-
To extend the closing date one month, a written request must be filed tions in a spatial reference engine. Descriptions of the tech-
with the ASCE Manager of Journals. The manuscript for this paper was nical detail of IntelliGIS implementation can be found in other
submitted for review and possible publication on December 2, 1998. This
paper is part of the Journal of Computing in Civil Engineering, Vol. papers (Jia 1996; Jia and Sarasua 1996). As a case study, the
14, No. 1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0051–0059/ paper also describes the use of the IntelliGIS for the devel-
$8.00 ⫹ $.50 per page. Paper No. 19774. opment of a pavement management system (PMS). The final
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 51
section of the paper concludes by summarizing the study find- how the spatial knowledge is employed to solve a particular
ings and outlining future research. problem.
The inference engine provides a mechanism in which
KBES TECHNOLOGY knowledge can be applied to infer solution(s) to the problem.
The inference usually involves two tasks: (1) Examining the
A typical KBES consists of three primary components: (1) existing knowledge stored in the knowledge base and existing
The knowledge base that contains rules that represent knowl- facts stored in the context (the context is discussed later); and
edge in a particular field; (2) the context that maintains the (2) determining in what order the inferences should be made.
problem conditions, the input data, and relevant information; The inference engine is the ‘‘brain’’ of a KBES. However, the
and (3) the inference engine that matches knowledge with in- brain is not designed to handle spatial knowledge and perform
formation in context. spatial reasoning.
The knowledge base contains two types of knowledge: do- The context is a global database containing information
main knowledge and control knowledge. The domain knowl- about the problem being considered. The information includes
edge represents an expert’s experience and understanding of problem data, solution status, and action history. In a KBES,
domain problems, and the control knowledge dictates how the the information is represented by a set of facts. These facts
domain knowledge is used for the search of solutions. Typi- can be spatially oriented because problems to be solved may
cally, these two types of knowledge are the key to the success involve spatial knowledge whose representation schemes de-
of a KBES. An adequate, reliable, and comprehensive set of cide how to encode the spatial facts in the context. Because
knowledge can provide a KBES with powerful knowledge rea- there are few representation mechanisms available for spatial
soning capabilities to derive effective solutions to problems. knowledge, the spatial facts cannot be encoded by a suitable
Knowledge can be acquired from a multitude of sources scheme for the purpose of spatial reasoning.
such as interviews with domain experts and examinations of
manuals, research reports, and textbooks. Once collected, the
INTELLIGIS
knowledge is organized and put into the knowledge base by a
knowledge representation or encoding method. A conceptual framework designed for the representation and
Spatial knowledge is a special kind of domain knowledge reasoning of spatial knowledge is shown in Fig. 1. The frame-
used for solving problems that involve spatial consideration. work, called IntelliGIS, adds two additional components into
Currently, encoding spatial knowledge into a knowledge base the conventional KBES discussed in the previous section. The
is a challenging issue in artificial intelligence and GIS com- two new components, GIS server and KBES-GIS interface,
munities, because the encoding scheme is heavily related to provide various functions and utilities for dealing with spatial
facts and for representing and reasoning about spatial knowl-
edge.
IntelliGIS allows spatial knowledge to be encoded and
stored in the knowledge base along with other knowledge. It
enhances the context component to contain various spatial
facts that describe spatial conditions of problems to be solved.
Furthermore, it strengthens the inference engine by adding
new modules to handle spatial reasoning. If a KBES applica-
tion contains a set of spatial knowledge, the inference engine
supports the pattern match of the spatial knowledge with in-
formation in the spatial context.
The KBES-GIS interface in IntelliGIS interacts with the
knowledge base, the context, the inference engine, and the GIS
server. It converts requests of spatial operations from the in-
ference engine and hands them over to the GIS server for
FIG. 1. Primary Components in IntelliGIS execution. The interface also manages the execution results

FIG. 2. (a) Spatial Knowledge Representation in IntelliGIS; (b) Spatial Knowledge Reasoning in IntelliGIS

52 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

back from the GIS server and converts them to spatial facts developers can update or add functions into the inference en-
for further spatial reasoning. gine of the CLIPS. This flexibility allowed the KBES-GIS in-
The GIS server provides KBES applications with GIS func- terface and a number of reasoning modules to be developed
tions. The functions include the manipulation of spatial and to handle spatial knowledge.
attribute databases, as well as spatial analysis operations re- ARC/INFO performs as the GIS server. There are several
quired by KBES applications in their reasoning process. reasons that ARC/INFO was chosen for this study. First, ARC/
INFO is a common commercial GIS that has powerful spatial
INTELLIGIS IMPLEMENTATION analysis functions for spatial reasoning in various transporta-
tion system analyses. One of the important functions in ARC/
IntelliGIS was implemented by enhancing CLIPS (a rule INFO is Inter-Application Communications (IAC) function.
and object-based expert system shell developed by the Soft- Use of the IAC function is critical to the development of the
ware Technology Branch at the Lyndon B. Johnson Space spatial reference engine because the IAC function allows direct
Center of the National Aeronautics and Space Administration) ‘‘talk’’ between CLIPS and ARC/INFO, and it does not re-
with ARC/INFO GIS and the KBES-GIS interface developed quire text files for the talk. Second, ARC/INFO can run in a
in the research (Jia and Sarasua 1996). CLIPS was selected server mode. It has capabilities to customize and automate GIS
because it is the only KBES shell claimed to have a certified processes using Arc Macro Language (AML). An AML pro-
correct inference engine (Wentworth 1997). CLIPS was writ- gram, like any other script program, groups a set of GIS com-
ten in C and its source code is available to the public. Devel- mands and functions to accomplish a spatial analysis task.
opers therefore can access and understand the internal infer- Third, ARC/INFO has been adopted by many state depart-
ence mechanism through the examination of the code. Also, ments of transportation as their primary GIS package for var-
ious projects.
The KBES-GIS interface used Transmission Control Pro-
TABLE 1. Rehabilitation and Maintenance Strategies in PMS tocol/Internet Protocol (TCP/IP) and Remote Procedure Call-
ing technologies to integrate CLIPS and ARC/INFO (Jia and
Number Strategy Sarasua 1996). The interface is seamless and effective in sup-
(1) (2) porting the communications between the KBES pattern match-
0 Do nothing ing process and the GIS spatial analysis operations. The de-
1 Crack seal tailed description of the interface implementation is provided
2 Base repair
in other papers (Jia 1996; Jia and Sarasua 1996).
3 Stripe seal
4 Slurry seal
5 Chip seal
SPATIAL KNOWLEDGE REPRESENTATION
6 Mill or level The representation of spatial knowledge in the IntelliGIS is
7 Resurface the same as that in a conventional KBES. Because of the na-
8 (Mill or level) and resurface
9 Chip seal and resurface
ture of the IntelliGIS and the object-oriented programming
features of CLIPS, the IntelliGIS can handle rule-based and

FIG. 3. Decision Tree in PMS

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 53

object-oriented spatial knowledge. It has features such as clas- clude the server identifier generated by the connect㛭to㛭
ses, message-handlers, encapsulation, and inheritance to rep- arcinfo function, an ARC/INFO command or an AML
resent and deduce solutions from spatial knowledge. program that contains a set of ARC/INFO commands, and
The IntelliGIS provides four KBES-GIS functions for the a series of parameters required by the ARC/INFO com-
support of spatial knowledge representation. The syntax for- mand or the AML program.
mats of the four functions are as follows: • disconnect㛭from㛭arcinfo The function disconnects a
communication channel established by the connect㛭to㛭
arcinfo function and releases GIS resources to other users.
• connect㛭to㛭arcinfo Arguments are not required. This A server identifier created by the connect㛭to㛭arcinfo func-
function activates the connection between IntelliGIS and tion is required.
ARC/INFO GIS server. The result of the function is a
server identifier sent back from the ARC/INFO GIS server The four functions are similar to CLIPS original functions for
to the context of a KBES application. the presentation of knowledge. They can be used at any place
• arcinfo㛭rhs 具server-id典 具ARCINFO command兩AML pro- where the original CLIPS functions are allowed. An example
gram典 {ai㛭arg1 . . . ai㛭argn} The function submits a rule highlighting how to use the connect㛭to㛭arcinfo function
message to the GIS server to request the server to perform in a KBES application is as follows:
spatial analysis functions involved at the right-hand side
(rhs) of a spatial rule. The arguments in the function in-
clude the server identifier generated by the connect㛭to㛭
arcinfo function, an ARC/INFO command or an AML
program that contains a set of ARC/INFO commands, and
a series of parameters required by the ARC/INFO com-
mand or the AML program. This example indicates that the connect㛭to㛭arcinfo function is
• arcinfo㛭lhs 具server-id典 具ARCINFO command兩AML pro- used at the left hand side of the setup rule. Also the example
gram典 {ai㛭arg1 . . . ai㛭argn} The function submits a shows that each time the application starts to run, it fires the
message to the GIS server to request the server to perform setup rule and invokes the connect㛭to㛭arcinfo function.
spatial analysis functions involved at the left-hand side An example spatial rule containing the arcinfo㛭rhs and
(lhs) of a spatial rule. The arguments in the function in- arcinfo㛭lhs functions is as follows:

FIG. 4. Firing Process of Spatial Rule for Defining Study Area

54 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

This rule represents spatial knowledge used in a prototype This rule represents a piece of knowledge saying ‘‘IF reason-
KBES-based PMS that will be described in later sections. The ing process is complete and its associated ARC/INFO server
spatial knowledge is ‘‘IF an ARC/INFO server is connected is still connected to wait for requests, THEN the communi-
and databases for pavement distresses, road characteristics, cations channel between the KBES application and the ARC/
and accidents are available for a route to be evaluated, THEN INFO server should be disconnected.’’ The ARC/INFO server
an event table should be generated for segments of that route.’’ is specified by the server identifier that is sent to the KBES
The rule uses the arcinfo㛭rhs function to request a connected application’s knowledge base previously by the connect㛭to㛭
ARC/INFO server to run an AML program called check.aml. arcinfo function.
The program checks whether a route selected for pavement
evaluation has condition data associated with it. If the route SPATIAL KNOWLEDGE REASONING
has the data, the program sends a confirmation flag (or value
>1) back to the rule. The rule uses test function, an original Spatial knowledge reasoning in the IntelliGIS is different
CLIPS function, to compare the returned confirmation flag from that in a conventional KBES. Take the reasoning process
with 0. If the conditions at the right-hand side of the rule are (or pattern matching) of the spatial knowledge discussed in the
true, the rule infers to the execution of the arcinfo㛭lhs func- above section as an example [Figs. 2(a and b)]. When a KBES
tion. The arcinfo㛭lhs function requests the ARC/INFO server application that contains the spatial knowledge fires the rule,
to run another AML program called create㛭event.aml. The pro- it talks with the ARC/INFO server and asks the server to run
gram generates an event table and assigns the events to seg- a program walking.aml. The ARC/INFO server receives the
ments of a route. request, runs the AML program, and sends the network dis-
An example rule indicating how to use the disconnect㛭 tance (in this case, 126 ft) back to the KBES application. Once
from㛭arcinfo function in a KBES application is as follows: the application obtains the network distance, it compares the

FIG. 5. Client/Server Interactions within Spatial Rule for Selecting Route

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 55

returned network distance with 2,640 ft and makes a decision The development of the PMS involved two major tasks:
‘‘access mode is walking.’’ knowledge acquisition and representation. The knowledge ac-
Clearly, the spatial knowledge reasoning is performed in quisition was a process of obtaining knowledge about pave-
two components: the inference engine and the GIS server. The ment treatments from a set of sources and storing the knowl-
inference engine is responsible for pattern matching of spatial edge into the knowledge base of the PMS. The sources
knowledge and flow control of matching order. The GIS server included a Federal Highway Administration district pavement
is responsible for various spatial actions requested by the in- engineer and Georgia DOT engineers. The pavement engineers
ference engine. initially provided a decision tree used for pavement manage-
ment. The decision tree was modified in this research through
KBES-BASED PMS interviews and consultation with the pavement engineers to
A prototype KBES-based PMS was developed in the consider the spatial nature of pavement. Fig. 3 shows a part
IntelliGIS to evaluate pavement distresses and determine seg- of the decision tree that organizes the obtained knowledge in
ment treatments. The PMS considers 10 possible rehabilitation a logical order. The decision tree provides a useful method-
and maintenance strategies obtained from the Georgia Depart- ology for selecting a feasible and effective segment treatment
ment of Transportation (DOT) (Table 1). It determines a treat- strategy from the 10 predetermined strategies.
ment strategy for a pavement segment based on a set of knowl- There were 255 rules acquired for the PMS. These rules can
edge (including spatial knowledge). be classified as follows:

FIG. 6. Firing Process of Spatial Rule for Event Overlay

FIG. 7. Firing Process of Spatial Rule for Obtaining Overlay Results

56 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 • Knowledge for handling spatial databases and generating KBES-GIS function by which the overlay results related to a
spatial facts segment can be organized by an AML program read㛭
• Knowledge for determining potential segment treatments events.aml. The AML program sends the overlay results to the
for various types of pavement distresses and finalizing the context of the PMS.
treatment for a segment When the overlay results are in the PMS, the PMS follows
• Knowledge for displaying segment treatments along a the decision tree to derive pavement segment treatments. Fig.
route and generating a report for decision makers 8 shows the segment treatments and the treatment legend of a
route.
The segment treatments of a route can be displayed graph-
There were 12 spatial rules created to handle spatial oper-
ically in the PMS (Fig. 9). The PMS uses a spatial rule to
ations involved in the spatial knowledge of the PMS. Fig. 4
request the ARC/INFO server to run an AML program called
shows how a spatial rule is fired to define a study area. In
report㛭interface.aml. This program provides an interface that
specifying a study area, the PMS sends a request to the ARC/
allows the ARC/INFO server to arrange treatments along a
INFO server and asks the server to run a customized spatial
route and to generate a report that can be printed as a hard
analysis function defined by an AML program study㛭area.aml.
copy.
The ARC/INFO server receives the request and sends a user
The pavement segment treatments were compared with the
interface (or a graphic form) back to the PMS. The user of the
treatments produced from the decision tree used in the Georgia
PMS then selects a county through a scroll window on the
DOT. It was found that the treatments from the KBES-based
interface. The area defined by the selected county is the study
PMS and the Georgia DOT’s decision tree were the same;
area used for the PMS because the spatial databases and pave-
however, the use of PMS is much quicker in making repair
ment-related data were organized on a county basis.
decisions on pavement segments because the PMS uses the
Fig. 5 shows the interactions between a spatial rule in the
automated decision tree as the source of knowledge. Also, the
PMS and the ARC/INFO server for selecting a route for study.
PMS provides the capabilities to graphically display treatment
Because the route has to be selected from the roadway cen-
decisions against distresses on same pavement segments. This
terline map (or the spatial databases) in the defined area, the
facilitates the validation of inference results.
spatial rule contains a KBES-GIS function sending a request
to the ARC/INFO server to execute an AML program select㛭
CONCLUSIONS AND FUTURE RESEARCH
a㛭route.aml. The ARC/INFO receives the request and displays
the route system in the study area. The user then selects a This paper demonstrates a new method to represent and han-
route graphically from the route network. dle spatial knowledge. A piece of spatial knowledge in this
Once a route is selected, the PMS assigns pavement dis- research is defined as a collection of facts and their relation-
tresses, road characteristics, and on-road accidents to each seg- ships (expressed by locators, topological relations, and spatial
ment of the route (Fig. 6). The PMS contains a spatial rule for inference actions) that can be used to solve a location-oriented
that purpose. The spatial rule involves requesting the ARC/ problem. Using the above definition, the IntelliGIS implements
INFO server to perform event overlay operations (defined by the method and represents a piece of spatial knowledge by a
the AML program create㛭event.aml ) in the server for the spatial rule. The spatial inference engine in the IntelliGIS uses
route. In Fig. 6, the average daily traffic (ADT) volume and KBES-GIS functions to perform spatial queries, spatial
the load cracking severity are overlaid onto the segments of a searches, and spatial operations.
route. Once the overlay results are ready in the server for each A useful feature of IntelliGIS is its capabilities to display
segment of a route, the PMS asks the server to send all of the input and output data graphically. The IntelliGIS allows spatial
overlay results to the PMS. Fig. 7 indicates that the results are facts to be entered into the context of a KBES application
sent from the server to the PMS, segment by segment. The using graphical capabilities provided by the GIS server. Fur-
spatial rule involved in the above data transfer contains a thermore, the decisions derived from a KBES application can

FIG. 8. Determination of Segment Treatments

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 57

 FIG. 9. Firing Process of Spatial Rule for Displaying Treatments

58 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

be represented in a graphical format, thus providing decision APPENDIX. REFERENCES
makers with clear and understandable information. This fea- Cohn, L. F., and Harris, R. A. (1991). ‘‘Knowledge-based expert system
ture is significant in successful use of KBES technology in the in transportation.’’ NCHRP Rep. 359, National Academy Press, Wash-
transportation field. ington, D.C.
This paper shows that the IntelliGIS system is an effective Djokic, D. (1991). ‘‘Urban stormwater drainage network assessment using
development environment for pavement management. Fur- an expert geographic information system,’’ PhD dissertation, Univer-
sity of Texas, Austin, Tex.
thermore, the system can also be used for developing KBES Jia, X. (1996). ‘‘A client/server-based intelligent GIS shell for transpor-
applications in other civil engineering areas (such as incident tation,’’ PhD dissertation, Georgia Institute of Technology, Atlanta.
management, accident analysis, bridge management, environ- Jia, X., and Sarasua, W. (1996). ‘‘A client/server-based intelligent GIS
mental analysis, and transportation network design) where en- for transportation.’’ Proc., 1996 GIS-T Symp., AASHTO, Kansas City,
gineering judgment and human experience can be formulated Mo.
Johnson, K., and Kanonier, J. (1991). ‘‘Knowledge-based land use clas-
and represented by a set of spatial rules. sification.’’ Int. GeoScience and Remote Sensing Symp. Digest, 3,
The research was successful in implementing a method for 1847–1850.
representing and reasoning about spatial knowledge, and there Laurini, R., and Thompson, D. (1992). Fundamentals of spatial infor-
are still a number of potential research areas that could be mation systems. National Academic Press, San Diego.
explored. The prototype PMS needs further improvements. McKinney, D. C., Maidment, D. R., and Tanriverdi, M. (1993). ‘‘Expert
geographic information system for Texas water planning.’’ J. Water
Currently, the PMS focuses on segment level analysis. How- Resour. Plng. and Mgmt., ASCE, 119(2), 170–183.
ever, a complete PMS should add methodologies for life-cycle Pazner, M. I. (1986). ‘‘Geographic knowledge base design and imple-
cost estimation and procedures for project- and network-level mentation,’’ PhD dissertation, University of California, Santa Barbara,
analyses. It will be worthwhile developing spatially heuristic Calif.
Prechaverakul, S., and Hadipriono, F. C. (1995). ‘‘Using a knowledge-
approaches to implement pavement prioritization schemes.
based expert system and fuzzy logic for minor rehabilitation projects
The current PMS also has limitations on the knowledge ac- in Ohio.’’ Transp. Res. Rec. 1497, Transportation Research Board,
quired from pavement engineers. The existing set of knowl- Washington, D.C., 19–26.
edge is only applicable to pavement segments where ADT Ritchie, S. G., and Prosser, N. A. (1991). ‘‘Real time expert system ap-
volume ⱖ4,000. To make the PMS operational for use in proach for freeway incident management.’’ Transp. Res. Rec. 1320,
Transportation Research Board, Washington, D.C.
the GDOT, further knowledge for the case of ADT volume Sarasua, W., and Jia, X. (1995). ‘‘A framework for integrating GIS-T with
ⱖ4,000 should be acquired from pavement engineers. KBES: A pavement management system example.’’ Transp. Res. Rec.
In addition, the current PMS does not have a knowledge 1497, Transportation Research Board, Washington, D.C., 153–163.
explanation module to describe how pavement treatments Spring, G. S., and Hummer, J. (1995). ‘‘Identification of hazardous high-
are derived and why certain information is needed. Adding way locations using knowledge-based GIS: A case study.’’ Transp. Res.
Rec. 1497, Transportation Research Board, Washington, D.C., 83–90.
this module in IntelliGIS is helpful in understanding the un- Wentworth, J. A., Knaus, R., and Aougab, H. (1997). Verification, vali-
derlying reasoning process involved in the treatment plan gen- dation and evaluation of expert systems, Version 1.2, 1st Ed., Vol. 1,
eration. Federal Highway Administration, Washington, D.C.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 59

 AUTOMATIC ROAD DETECTION IN GRAYSCALE AERIAL IMAGES
By Katherine Treash1 and Kevin Amaratunga2

ABSTRACT: Digital aerial photography provides a useful starting point for computerized map generation. Fea-
tures of interest can be extracted using a variety of image-processing techniques, which analyze the image for
characteristics such as edges, texture, shape, and color. In this work, we develop an automatic road detection
system for use on high-resolution grayscale aerial images. Road edges are extracted using a variant of the
Nevatia-Babu edge detector. This is followed by an edge-thinning process and a new edge-linking algorithm
that fills gaps in the extracted edge map. By using a zoned search technique, we are able to design an improved
edge-linking algorithm that is capable of closing both large gaps in long, low-curvature road edges and smaller
gaps that can occur at triple points or intersection points. An edge-pairing algorithm, which is subsequently
applied, takes advantage of the parallel edges of roads to locate the road centers. Results demonstrate that the
current system provides a simple yet effective first stage for a more sophisticated map-generation system.

INTRODUCTION the grayscale level of the surrounding areas is closer to that

of the roads. The road boundaries are sometimes broken by
With more and more of the world’s information becoming the presence of driveways. Fig. 1(c) illustrates the difficulty
digitized and with the popularity of geographical information of viewing roads through treetops and their shadows. The
systems (GIS), the detection and recording of digital map data edges of these roads are highly occluded, making it hard even
from aerial images has become increasingly important. In the for a human observer to make out the exact position of the
past, analog information was recorded manually, requiring edges. Fig. 1(d) shows a more densely populated area in which
large amounts of time and effort to catalog the precise details. the building edges are often sharper than those of the roads.
As well as speeding up this process, having a computer per- Shadows, this time from houses, also complicate the analysis
form the task allows for rapid updating, displaying, and trans- of such an image. Furthermore, although all four of these im-
mission of the information. ages are of approximately the same resolution, the road widths
There are various topographical features that need to be ex- and lengths vary.
tracted from aerial images: roads, lakes, rivers, railroads, The disparity in characteristics from image to image is the
buildings, and so forth. This paper focuses on the detection of primary source of difficulty in the road detection process. Sys-
roads from grayscale aerial images because of their importance tems that attempt to deal with all the variations can become
in most maps and because there is less variability in the char- quite complicated. It is the goal of this work to develop a
acteristics of roads as compared to the characteristics of other simple and practical method for extracting roads from a wide
features. Furthermore, an attempt is made to completely au- variety of aerial images. The only assumptions that are made
tomate the process. In general, roads have several character- are that roads are bounded by two roughly parallel edges and
istics that help automatic detection: They tend to be long and that they are either straight or change direction very slowly.
continuous, with small curvature, sharp, parallel boundaries, Then, despite the broken or faded boundaries, the roads are
and relatively constant grayscale and texture values between found. In a sense, this work will determine how far road de-
the boundaries. However, the problem is made difficult by the tection can proceed while ignoring many of the above-men-
fact that these characteristics do not always hold for individual tioned image variations, and therefore how important it is to
roads. The boundaries can be blurred, the background can be take into account the diversity of road characteristics.
very close in grayscale value to the road itself, there can be Methods for automatic road detection can be classified as
trees or shadows breaking the continuity of the boundaries, or either local or global in nature. Global procedures attempt to
the road can be a short, dead-end road or a driveway. The develop accurate models for the roads, incorporating the many
absence of color information in the images also adds to the characteristics into as few parameters as possible (Barzohar
complexity of the problem. and Cooper 1996). Within the category of local methods, many
Fig. 1 illustrates some examples of the variations found techniques exist. Edge detectors and linkers are often used to
within aerial images. (The test images used here have been find road seeds, which are then grown into the full road net-
provided courtesy of MassGIS and MIT, through the MIT Dig- work (Zhu and Yeh 1986; Zlotnick and Carnine 1993). The
ital Orthophoto Browser, http://ortho.mit.edu.) Fig. 1(a) shows textural and spectral properties of roads are employed to group
a close-to-ideal image for road detection. The roads are wide, road pixels together. Similarly, the difference between the
their boundaries are clear and continuous, and the background grayscale values of roads and the background is exploited in
is significantly darker than the roads. The only problem that order to determine the boundaries of the roads. In addition,
this image demonstrates is the cars on the roads, which break methods that attempt cooperation between local and global
up the relatively smooth texture of the highway and disturb methods also exist (Airault et al. 1994). Many of these refer-
the edge detection by creating short ‘‘noisy’’ edges. Fig. 1(b) ences give further lists of road-detection methods.
also has continuous roads with few occlusions; however, here In this work, we develop a ‘‘best of breed’’ road-detection
1 system for high-resolution aerial images. This combines (a) a
Grad. Res. Asst., Dept. of Civ. and Envir. Engrg., Massachusetts Inst.
of Technol., 77 Massachusetts Ave., Cambridge, MA 02139. variation of the Nevatia-Babu (1980) edge detector; (b) an
2
Asst. Prof., Dept. of Civ. and Environ. Engrg., Massachusetts Inst. of edge-thinning process; (c) a new edge-linking scheme that im-
Technol., 77 Massachusetts Ave., Cambridge, MA. proves on the technique of Heipke et al. (1994); and (d) an
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. edge-pairing process similar to that of Zlotnick and Carnine
To extend the closing date one month, a written request must be filed (1993). The linking scheme of Heipke et al. (1994) is confined
with the ASCE Manager of Journals. The manuscript for this paper was
to small search areas, predefined by the direction of an edge,
submitted for review and possible publication on May 4, 1999. This paper
is part of the Journal of Computing in Civil Engineering, Vol. 14, No. and can only fill small gaps in an edge map. The linking tech-
1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0060–0069/$8.00 nique described here performs a more extensive zoned search
⫹ $.50 per page. Paper No. 20894. that can fill in both small and large gaps.
60 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
 AUTOMATIC DETECTION OF ROADS
Edge Detection
Assuming that edges represent sharp changes in grayscale
value, and that roads have a significantly different grayscale
value than the surrounding background, edge detection is a
natural first step in extracting roads from aerial images. The
edge-detection process is based on a masking procedure,
which produces edge-magnitude and direction information that
can be used as a basis for selecting candidate edge points. An
edge-thinning procedure is then applied to these candidate
edge points. This ensures that the output of the edge detector
consists of sharp, well-defined edges.

Determination of Angles and Edge Values

The first step involves labeling each pixel of an image with
an edge value (magnitude) and a direction (angle). Several
masks are chosen, each corresponding to the digital values of
an ideal edge oriented in a certain direction. These masks are
convolved with the given image. At each pixel, the mask pro-
ducing the largest value is found. The corresponding direction
and this largest value are recorded as the angle and edge value
of that pixel.
Two factors must be considered when choosing the masks:
how many masks to use and how large to make them. The
number of masks corresponds to the number of angle values
possible. There is no front or back to the edges, so only angles
from 0 to 180⬚ need be considered, with 0 and 180 being the
same orientation. Based on experiments, six masks were found
to work well, with angle values 0, 30, 60, 90, 120, and 150.
The size of the masks is a detail considered closely by Nevatia
and Babu (1980). Small masks will have problems discrimi-
nating between noise and true edges, while large masks will
miss the finer details of the edges. A set of 5 ⫻ 5 masks was
found to work well for the test images used in the present
work. Fig. 2 illustrates one such mask, which corresponds to
an angle of 30⬚ from the vertical.

Selection of Edge Points

The direction and edge value information is next used to
determine which image pixels are truly part of an edge. Nev-
atia and Babu (1980) use three criteria to test for an edge pixel:
(1) A local maximum test checks that the central pixel has an
edge value larger than the edge values of its neighbors in the
direction normal to that of the edge; (2) an angle test checks
that the pixel has an angle close to the angles of its neighbors;
and (3) a thresholding test checks that the edge value is larger
than some threshold. Edge images resulting from the local
maximum test are generally not useful since the edges are
broken into many small pieces, which are hard to distinguish
from noise. The results from using the angle test by itself also
are not useful, as the edges in this case are quite wide and
blurred. Combining the angle and threshold tests yields good
results, as does the threshold test by itself. The combination

FIG. 2. Example of Mask for Finding Edge Values and Angles,

FIG. 1. Examples of Test Images Oriented at 30ⴗ to Vertical

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 61

 FIG. 3. Edge-Thinning Masks

FIG. 4. Example of Edge Thinning

of tests leads to more noise and broken edges, but at times Fig. 3, which need to be applied to each edge pixel in order
also includes parts of some true edges that the threshold test to thin the edges. One pass of the thinning algorithm compares
does not. However, it is judged that having more continuous, the neighborhoods of each edge pixel with the original, trans-
unbroken edges as output from the edge-detection process pose, and rotations of each of the 17 masks. Also, they must
leads to better final results. Thus, the threshold test is used by be compared starting from the left, right, top, and bottom in
itself to detect edge points. The threshold value is not a fixed order to take full advantage of symmetry. Fig. 4 shows the
value, but is found for a particular image as the value that thinning example from Heipke et al. (1994), illustrating the
retains a certain percentage of the edge energy after thres- process of removing an edge pixel that does not belong in the
holding. The user can adjust this percentage, such that a rea- skeleton. In dealing with the images used for this paper, one
sonable balance between continuous edges and unwanted noise pass through all the checks was enough to satisfactorily thin
is achieved. Edge energy is defined as the sum of the squares the edges. However, if edges are thick enough, the thinning
of the edge values. may have to be repeated several times.

Edge Thinning Edge Linking

Following the thresholding, any single points are deleted No edge-detection method is perfect, in that it finds all the
from the edge image. The remaining edges vary in thickness. edge points and only edge points. Usually, unwanted noise is
Before proceeding to the linking stage of the system, it is present in the form of short, erratic edges, and many edge
helpful to thin the edges to a one-pixel thickness. The goal is pieces remain disconnected from other pieces on the same
to remove edge pixels one by one until only the ‘‘skeleton’’ edge by gaps. Both these problems are addressed in the edge-
remains (Heipke et al. 1994). A pixel in the skeleton can have linking process. As with the problem of edge detection, spe-
only two other edge pixels in the 3 ⫻ 3 neighborhood sur- cifically road detection, linking methods are either local or
rounding it, unless it is the end of an edge or at a crossing global systems. Local methods deal with a neighborhood
point. The 3 ⫻ 3 neighborhood of each edge pixel is inspected around the pixel to be linked, analyzing the characteristics of
in order to determine if the center pixel is part of the skeleton these neighboring pixels to determine where the links should
or not. There are many possible patterns that indicate the cen- be. The simplest of local schemes focus on a small neighbor-
ter pixel is not part of the skeleton and can thus be removed. hood and use information from earlier edge detection in order
However, certain conditions on properly thinning edges to find pixels with similar characteristics. As the size of the
(Heipke et al. 1994) reduce the number of such cases dramat- neighborhoods gets larger, searches are introduced based on a
ically. These conditions guarantee that (1) thick edges are not variety of path metrics (Farag and Delp 1995). Fuzzy reason-
split into parallel thin edges, but are thinned as a single edge; ing has also been used to deduce which pixels in the search
(2) no holes are created in edges, and (3) edges are not short- area should be linked together (Law et al. 1996). Global meth-
ened at all, but remain the same length. The additional con- ods look at the overall pattern of edges and try to describe the
sideration of symmetry determines 17 3 ⫻ 3 masks, shown in features using a few variables. Examples of global methods
62 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
include modeling the edge image as a potential function (Zhu Determination of Primary and Secondary Directions
et al. 1996) and using the Hough transform (Gonzalez 1987),
Markov random fields (Geman et al. 1990), and least-square- As the basis for the linking process, primary and secondary
error curve fitting (Porill 1990). directions are associated with each end point. This is done by
The linking procedure that follows is a local method and first growing the tree with the end point as the root and then
specifically focuses on linking edges that represent roads. determining which combination of branches, beginning at that
Roads are usually long and relatively straight. These facts help end point, produces the straightest and most continuous edge.
determine which neighborhoods are searched for points to link Each combination of branches that starts at the root point and
together. The basic idea revolves around finding the direction ends at another end point is considered. An overall direction
in which the edge is headed and using this to outline a search is determined for each of these combinations by calculating
area. Although the general idea of search areas for edge linking the angle between the first point in the combination (the root
has been used before (Heipke et al. 1994), the details of the point) and the last point (another end point). The individual
searches performed by this method are new. In particular, our directions of each branch in the combination are also found
method uses a more extensive zoned search, which allows both as the angle between the beginning and end points of the
small and large gaps in edges to be filled. Because many dif- branch. The combination with the smallest sum of absolute
ferent edges are often connected together (at corners or triple differences between individual branch directions and the over-
points or because the edge detection is not perfect), steps are all combination direction is determined to be the ‘‘best’’ com-
taken to determine which edge points are actually a part of the bination.
true edge leading away from an end point. Once this is ac- The ‘‘best’’ combination of branches is used to determine
complished, an estimate is made as to the direction in which the direction in which the edge is heading by looking at the
the edge is heading. Primary and secondary directions, found last 20 steps taken within this edge. These steps are unit steps,
by looking at the last few step directions within an edge, are which limits the number of possible directions to 8. The last
assigned to each end point as candidates for the true edge 20 directions are ‘‘polled,’’ with the first and second most
direction. Because we expect the road edges to continue in the prevalent directions becoming the primary and secondary di-
direction they are heading, searching is done away from the rections associated with that end point. The numbers of times
end point in the primary and secondary directions. In this man- the primary and secondary directions occur in these last 20
ner, links are made that fill the gaps in road edges. steps are called the primary and secondary confidences.

Determination of End Points Searching for Links and Edge Linking

The first step in the linking process is to establish which Finally, once all end points have their primary and second-
edge points are end points. As defined by Zhu et al. (1996), ary directions, the search begins for links to make. Assuming
an edge point is also an end point if and only if it has (1) only road edges do not change direction suddenly, we expect to be
one other edge pixel in its 3 ⫻ 3 neighborhood, or (2) two able to continue a broken edge in the direction it is heading
edge pixels in its neighborhood that are directly adjacent. Due in order to find appropriate edge pieces to link it to. Therefore,
to the thinning of the edges, end points that are classified using as mentioned above, search areas for a particular end point are
the second test are not useful for linking, but instead tend to based on that point’s primary and secondary directions.
be glitches in smooth edges. Therefore, end points are deter- For a given end point, the search area is determined as fol-
mined using the first test alone. lows: Say x is the primary confidence and y is the secondary
confidence. A basic step pattern is determined consisting of x
Tracing the Tree
⫹ y steps, with x of those steps being in the primary direction
The next step involves forming a tree to represent each and y in the secondary direction, and with the y steps in the
edge. An end point is chosen as the root for the tree, and a secondary direction being evenly spaced throughout the x ⫹
series of branches are traced from this point. Initially, because y steps. For example, suppose that the primary step direction
of the way the end points are found, there will only be one was down and to the right (labeled as 8), with a confidence
branch leading away from the root end point. At each step of 13, and the secondary direction was to the right (labeled as
along this branch, a search is done for other edge pixels in the 5), with a confidence of 6. Then the basic step pattern would
3 ⫻ 3 neighborhood of the last recorded edge point. If no consist of 13 ⫹ 6 = 19 steps and would look like ‘‘8 8 5 8 8
others are found, the center point is another end point and the 5 8 8 5 8 8 5 8 8 5 8 8 5 8.’’ This is shown in Fig. 5(a).
branch ends there. If only one other is found, it is added to Starting at the given end point, the step pattern is repeated
the branch and becomes the new center point. If two or more until a certain number of steps have been taken away from the
edge pixels are found, the center point becomes a branch point, end point, in this case 50. Then the search returns to the start-
the end of the branch, and the edge pixels that are found in ing end point, takes one step to the left and one to the right,
its neighborhood are recorded as the beginnings of new and uses these two points as new starting points for the step
branches. The searches continue in this manner, tracing out a pattern. The left and right step directions are determined by
tree in which each branch begins and ends at either an end the primary and secondary directions in order to ensure that
point or a branch point. A stack is used to ensure that no no pixels within the bounds of the search area will be left
branches are missed. Each branch has a number, and branches unchecked and no pixel is checked twice. Continuing the ex-
that are connected are kept track of by a listing of pairs of ample from above, the ‘‘left’’ step would be taken upwards
branch numbers. The length of each branch is also registered. and the ‘‘right’’ step downwards, as in Fig. 5(b). Ten steps are
This allows some of the noise to be eliminated by deleting taken away from the end point, both to the left and the right,
trees (edges) with a total length of, say, 10 or fewer pixels. with the step pattern search beginning at each new point and
All the endpoints that form the tips of the tree are also re- continuing for 50 steps outward. The search area then is an
corded as belonging to this particular tree so that the tracing irregular box/parallelogram with the end point in the middle
is not repeated with these points as roots. of one of the shorter ends.
In this way, each edge is cataloged as a series of branches. The links are made between end points, so all end points
The representation of the edges as trees has the advantage of in the given search area are recorded. We call the end point
providing the needed edge information in a simple and acces- that is the base of the search the root point, and any end
sible format. point found within the search area is labeled as a candidate
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 63
end point. Once all candidate end points are located, a becomes the target point for linking. Zone two consists of all
series of tests is performed to determine which points to points 1, 2, or 3 steps off center. If the target point lies in this
link. zone, the link is made. Zone three contains all points between
The tests revolve around two measures: (1) The distance 4 and 6 steps off center, inclusive. Zone four consists of all
between the root and the candidate end point; and (2) the num- remaining points in the search area (7⫹ steps off center). If
ber of steps that the candidate point lies off the center path. the target point is in either of zones three or four, a check is
The values of these two measures determine four zones within made on the alignment of the root and the target. The align-
the search area where linking is possible (Fig. 6). The first ment test succeeds if the difference in their primary directions
zone is the center path leading away from the root point (steps is less than or equal to 90⬚. If the point is in zone three and
off center = 0). If a candidate point is in this zone, linking the alignment test succeeds, linking occurs. If the point is in
always takes place. If no points lie in zone one, the candidate zone four and the alignment test succeeds, linking is done if
end point with the shortest distance to the root is found and the candidate point has at least a distance of 20 from the root.
This final distance check prevents linking with points that are
in the bottom corners of the search area, which tend to be bad
links.
After a link is made, a final test is used to determine whether
the linking process should be repeated for the current root
point. If the point that has just been linked to the root is suf-
ficiently close (less than or equal to a distance of 10 units
away), then the linking process will be repeated for the root
point, considering the same candidate points minus the one
that has been linked already. This implies that several links
can be made with the same root point as the starting point.
These roots tend to be triple points or intersection points. The
distance test that determines whether the process repeats or
not allows for several shorter links to be made, but only one
longer link. Once a point further than 10 units away is linked
to the root, the process stops and the linking procedure moves
onto the next root point. The test is designed this way because
usually only one long link makes sense, while multiple short
links are often needed.
The linking method can be further improved if special cases
are accounted for. Currently, no specific care is taken with
corners or branches in the roads. Linking only takes place
between end points, ignoring the fact that other edge points
may provide a better starting point for linking. The links are
made with straight lines, neglecting the possibility that the
edge may be in the process of curving. Overall, some edge
gaps may be missed and some false links may be made. None-
theless, provided that enough edges are kept in the edge-de-
tection stage, this linking process will fill most of the gaps that
should be filled.

Edge Pairing and Determination of Roads

FIG. 5. Example of Finding Search Area from Primary and

Edge detection and linking are first steps toward finding the
Secondary Directions: (a) First 19 Steps of Center Path; (b) roads in an aerial image. Although these methods, especially
Paths Followed after Taking One Step to ‘‘Left’’ and One to the linking, are designed to work best with road edges, there
‘‘Right’’ remain edges that are not parts of roads. The goal of this final

FIG. 6. Zones for Edge Linking

64 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

phase is to focus on which of the remaining edges do corre- erative method that assigns scores to pairings based on lengths
spond to roads and where those roads are located within the and overlaps of edges. Zhu and Yeh (1986) expand on this
image. method, including tests for the intensity and width between
As discussed earlier, one characteristic that applies to all two edges.
roads is that they are bounded by two parallel, or very close In the present work, a simplified version of these methods
to parallel, sides. Even when occlusions, shadows, or other is used. The user inputs the minimum and maximum widths
details break the continuity of the road boundaries, enough of the roads to be found. At each edge point, a search is done
pieces remain after edge detection and linking to be able to in a direction normal to that of the edge for other edges that
pair together edges that represent the two sides of a road. are between the minimum and maximum distances away. A
Many of the local methods for road detection include a step check is done to see if the two edges are close to parallel. This
in which pairing of parallel lines occurs. After the edges are is done by making a local comparison of the directions of the
extracted, they are usually approximated by linear segments. edges to make sure their difference is below a certain thresh-
The straight lines that represent the edges also have orienta- old, also chosen by the user. If the two local directions are
tions, determined somewhere during the preceding steps of the close enough, a point is drawn half way between the edges
system. Sophisticated algorithms have been developed to and a record is made of the pairing. The output is an image
find the best sets of pairings between antiparallel lines. showing the linked edges, with points located half way be-
For example, Scher, Shneier, and Rosenfeld (1982) use an it- tween paired edges.

FIG. 7. Results from Test Image #1: (a) After Edge Detection and Thinning; (b) Zoom of 100 ⴛ 100 Block of (a); (c) After Edge Linking;
(d) Zoom of 100 ⴛ 100 Block of (c); (e) Linked Edges and Road Midpoints; (f) Road Midpoints Only

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 65

RESULTS Fig. 7(e) shows the result of edge pairing for this test image.
The quality of individual results can vary, depending on the The pairing step has plotted points along almost the entire
characteristics of the image. Fig. 7 shows the edge image after length of the horizontal road. Along the vertical road, the extra
each of the three main steps for a ‘‘well-behaved’’ test image. lane edges hamper this process somewhat, but there are still
The original test image is shown in Fig. 1(a) as having clear two fairly complete road lines, one for each lane. In order to
and continuous road edges that are easily picked out by the better illustrate the points that represent the middle of the
edge detection, as can be seen in Fig. 7(a). Fig. 7(b) is a zoom roads, they are plotted without the road edges in Fig. 7(f). One
of a 100 ⫻ 100 block of the edge image. These edges are could, in principle, process this image further to link the road
achieved by keeping 40% of the edge energy of the image. centerline segments and to remove short, noisy edges.
The texture of the background, the lines dividing the lanes, Similar sets of images are shown for three other, less ‘‘well-
and the cars on the roads all lead to extra, undesirable edges. behaved’’ test images. Fig. 8 displays the results of the image
Fig. 7(c) shows the same edge image after the linking process. shown in Fig. 1(b). The possible complication for this image
Many of the short ‘‘noisy’’ edges have been removed in this is that the background grayscale value is not as different from
step, and many obvious gaps have been filled. When viewing that of the roads, especially where the driveways meet the
the entire 500 ⫻ 500 pixel image, it is hard to see the links road. The crossroads could also pose problems for detecting
that have been made. A closeup of a 100 ⫻ 100 block of this roads. However, despite these difficulties, the edge detection
image is shown in Fig. 7(d) to make the links clearer. Finally, yields a fairly complete set of edges and the linking tidies up

FIG. 8. Results from Test Image #2: (a) After Edge Detection and Thinning; (b) Zoom of 100 ⴛ 100 Block of (a); (c) After Edge Linking;
(d) Zoom of 100 ⴛ 100 Block of (c); (e) Linked Edges and Road Midpoints: (f) Road Midpoints Only

66 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

the image. A couple of false links are also made that cut across tainly do not correspond to roads. The edge images in Figs.
the roads near the crossroads, as seen in the zoomed image 9(c) and 9(d), after edge detection and linking, are hardly rec-
shown in Fig. 8(d). The final results indicate that indeed the ognizable as the original image shown in Fig. 1(c). When the
crossroads do pose a problem when it comes to the pairing of first two steps fail, there is little chance of the third actually
edges. Perhaps if a larger width of road were allowed, more finding the roads, and indeed in this case the final result gives
road points would be detected here. However, the road leading no insight into the position of the roads in this image.
upward from the crossroads is detected quite nicely, despite The last test image is that of Fig. 1(d), with results exhibited
the gaps in the road edges due to driveways. in Fig. 10. This is a scene of a residential area, with a grid-
The third test image has highly occluded roads. The many like road structure and many buildings. The edge detection
trees and their shadows break the continuity of the roads and picks out the sharp edges of the houses and some broken road
obscure many of their edges. The results after edge detection, edges. In this case, the roads are more visible as the long,
shown in Figs. 9(a) and 9(b), reflect this. There are some straight areas containing no edges rather than areas bordered
places where the edges of the roads have been detected, but by edges. The linking step completes some of the houses and
most of the edges belong to trees or houses. The linking step roads. However, as can be seen in Fig. 10(e), the lack of road
does little to improve the image. Because so few road edges edges does not prevent the pairing step from finding most of
were found in the previous step, and because so may nonroad the third vertical road, and a good portion of the second ver-
edges were found, many of the links make no sense and cer- tical road. This demonstrates that, although there are many

FIG. 9. Results from Test Image #3: (a) After Edge Detection and Thinning; (b) Zoom of 100 ⴛ 100 Block of (a); (c) After Edge Linking;
(d) Zoom of 100 ⴛ 100 Block of (c); (e) Linked Edges and Road Midpoints; (f) Road Midpoints Only

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 67

FIG. 10. Results from Test Image #4: (a) After Edge Detection and Thinning; (b) Zoom of 100 ⴛ 100 Block of (a); (c) After Edge Link-
ing; (d) Zoom of 100 ⴛ 100 Block of (c); (e) Linked Edges and Road Midpoints; (f) Road Midpoints Only

cases of images with difficulties that cause the program to fail either as faded edges or occlusions due to trees or shadows,
or do poorly, there are also some difficult circumstances in the quality of the results declines. This suggests that using a
which this system can still perform quite well. simple three-step process such as that described in this paper
does not always do an adequate job of dealing with the variety
CONCLUSIONS of image characteristics mentioned in the introduction. How-
We have presented a simple and practical scheme for de- ever, this provides a good beginning to a more complex sys-
tecting roads in grayscale aerial images. This scheme combines tem. For example, in Zlotnick and Carnine (1993), the centers
the best features of previous methods (Nevatia-Babu edge de- of roads found by pairing edges are analyzed to find clusters,
tection, edge thinning, and edge pairing) with a new edge- which are then used as seeds to grow the road network. It
linking technique. The edge-linking technique improves on the appears that such a step is necessary if the goal is a complete
scheme presented by Heipke et al. (1994) by using a more road network.
extensive zoned search, which is capable of closing both large Another possible extension of this process is to incorporate
gaps in long low-curvature road edges and smaller gaps that information based on texture or grayscale value. For instance,
can occur at triple points or intersection points. in most cases the grayscale of the roads is significantly dif-
The overall road-detection system was found to perform ferent from that of the background. By looking at the variance
well for images with sharp, continuous edges, such as the first in grayscale value within a small neighborhood of each pixel,
test image. Once discontinuities in the edges are introduced, another tool for extracting road edges is created.
68 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
 A further way to incorporate other types of information may International Society for Optical Engineering, Bellingham, Wash.,
be to make the process semiautomatic instead of fully auto- 2315, 508–518.
Barzohar, M., and Cooper, D. B. (1996). ‘‘Automatic finding of main
matic. The system described here already requires the user to roads in aerial images by using geometric-stochastic models and esti-
input some parameters at the beginning of the program, but mation.’’ IEEE Trans. on Pattern Anal. and Machine Intelligence,
after the initial inputting, the computer proceeds on its own. 18(7), 707–721.
Semiautomatic methods have been devised that allow the com- Farag, A., and Delp, E. J. (1995). ‘‘Edge linking by sequential search.’’
puter and human operator to interact (Heipke et al. 1994). Pattern Recognition, 28(5), 611–633.
Usually a starting point and direction are inputted, as well as Geman, D., Geman, S., Graffigne, C., and Dong, P. (1990). ‘‘Boundary
detection by constrained optimization.’’ IEEE Trans. on Pattern Anal.
widths and lengths of the roads to be found. One can imagine and Machine Intelligence, 12(7), 609–628.
the program may also allow the user to click the mouse on a Gonzalez, R. C. (1987). Digital image processing, 2nd Ed., Addison-
point somewhere inside the road, from which the computer Wesley, Reading, Mass., 344–349.
could determine that road’s texture and/or grayscale. Other Heipke, C., Englisch, A., Speer, T., Stier, S., and Kutka, R. (1994).
road points could be found by comparing texture and grayscale ‘‘Semiautomatic extraction of roads from aerial images.’’ Proc., SPIE
information. This interaction between human and computer ISPRS Commission III Symp., SPIE—International Society for Optical
Engineering, Bellingham, Wash., 2357, 353–360.
may also allow the program to be interrupted if it runs off Law, T., Itoh, H., and Seki, H. (1996). ‘‘Image filtering, edge detection,
course. and edge tracing using fuzzy reasoning,’’ IEEE Trans. on Pattern Anal.
In conclusion, it has been shown that simple steps can form and Machine Intelligence, 18(5), 481–491.
a basis for an effective road detection system. However, ad- Nevatia, R., and Babu, K. R. (1980). ‘‘Linear feature extraction and de-
ditional information must be incorporated in order to develop scription.’’ Comp. Graphics and Image Processing, 13(3), 257–269.
a procedure that satisfactorily deals with the wide variety of Porrill, J. (1990). ‘‘Fitting ellipses and predicting confidence envelopes
using a bias corrected Kalman filter.’’ Image and Vision Computing,
aerial images that need to be processed. 8(1), 37–41.
Scher, A., Shneier, M., and Rosenfeld, A. (1982). ‘‘A method for finding
ACKNOWLEDGMENTS pairs of antiparallel straight lines.’’ IEEE Trans. on Pattern Anal. and
Machine Intelligence, 4(4), 316–323.
This work was partially supported by a grant from the Suksapattana Zhu, M. L., and Yeh, P. S. (1986). ‘‘Automatic road network detection
Foundation, Bangkok, Thailand, to the Intelligent Engineering Systems on aerial photographs.’’ Proc., CVPR’86 IEEE Comp. Soc. Conf. on
Laboratory, Massachusetts Institute of Technology. Comp. Vision and Pattern Recognition, IEEE Computer Society Press,
Washington, D.C., 34–40.
APPENDIX. REFERENCES Zhu, Q., Payne, M., and Riordan, V. (1996). ‘‘Edge linking by a direc-
tional potential function (DPF).’’ Image and Vision Computing, 14(1),
Airault, S., Ruskone, R., and Jamet, O. (1994). ‘‘Road detection from 59–70.
aerial images: A cooperation between local and global methods.’’ Zlotnick, A., and Carnine Jr., P. D. (1993). ‘‘Finding road seeds in aerial
Proc., SPIE Image and Signal Processing for Remote Sensing, SPIE— images.’’ CVGIP: Image Understanding, 57(2), 243–260.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 69

 INTEGRATING DOCUMENT MANAGEMENT WITH PROJECT
AND COMPANY DATA

By Dany Hajjar1 and Simaan M. AbouRizk,2 Member, ASCE

ABSTRACT: Electronic document management (EDM) systems have revolutionized the construction document
storage and retrieval process. An electronically managed document consists of a computer representation of the
main document body, usually in the form of a bitmap, and a reference or indexing structure used to retrieve the
document. In construction, the reference data includes information that relates the document to certain aspects
of the constructed facility and the construction company. It is argued that this relationship needs to be represented
explicitly and in an integrated manner as opposed to the approach used by traditional EDM systems. This paper
presents an integrated approach based on the concept of specialized construction data models. A detailed account
of how this approach was followed to integrate a quality assurance document management system within a steel
fabrication company is also presented.

INTRODUCTION Certain advanced systems do allow for database lookups,

which could partly solve this problem by accessing the latest
Construction document management is an essential com- company information.
ponent of the overall project management function. Construc- Complex data structures with custom data entry screens are
tion documents include drawings, design specifications, sched- almost impossible to implement with commercial EMDS sys-
ules, quality control reports, and others. The growing tems. A common example of such requirements in the steel
complexity and magnitude of construction projects have re- industry is a material listing. Each material is a complex data
sulted in an increase in the problems associated with manual type in the sense that it contains many attributes and subattri-
document management and retrieval. It is estimated that the butes with a large set of validation rules and custom user in-
total number of documents relating to a single building struc- terfaces. Such a structure is generally available as part of most
ture is about 10,000 (Turk et al. 1994). This trend, coupled steel project management systems. Although advanced EMDSs
with the increased affectability of computer systems, has led do allow such customization to a certain extent, it is possible
companies to invest in electronic document management sys- only through proprietary macrolanguages. Even then, main-
tems (EDMS). Typical features found on most commercial taining two parallel representations in two separate systems
EDMSs provide means to capture, edit, index, distribute, and becomes a maintenance nightmare.
retrieve various types of electronic and hardcopy documents. It is argued that there should be an explicit definition of the
Many different technologies are involved in EDMS, including relationship between document management systems and proj-
computer networks, servers, workstations, scanners, printers, ect information. Explicit integration allows for the automated
and databases. retrieval of electronic documents related to a given project data
The wide use of EDMS in the construction industry can element as well as the ability to automatically validate docu-
assist in managing construction documentation. However, the ment indexing information as they are entered.
complexity of construction projects and the amount of infor- State-of-the-art research dealing with this issue focuses on
mation involved present a unique problem to the implemen- utilizing a common product data representation scheme for
tation of commercial EDMSs. The issue is how to efficiently project and document data (Rezgui and Debras 1996). In this
and effectively relate documents to both construction project paper, the term ‘‘data model’’ is used to denote the way data
and company information at a detailed level. A review of sev- is represented and related within a given system. For example,
eral commercial EMDSs has revealed that this feature is not the product data model of a steel fabrication project defines
directly supported. Products like Insight (1999) and Optix the way drawing and material information is represented pos-
(1999) require specialized external interfaces typically devel- sibly in the form of a relational database diagram or an object
oped by dedicated consultants. Such other products as hierarchy. Similarly, the data model of a document manage-
AutoEDMS (1999) do provide an open architecture for doc- ment system determines the internal representation of a given
ument design. However, the process typically involves expert document. Numerous standard construction product data mod-
designers using specialty tools and proprietary programming els such as COMBINE (Dubois et al. 1995) are being devel-
languages.
oped. A comprehensive list of current work can be found in
Static and ‘‘hard-coded’’ integration schemes used by all
Froese (1998). The standard model approach where a single
currently available commercial EMSs lead to several draw-
data model is designed and developed to represent information
backs. Validation rules are defined on a field-by-field basis and
in the entire construction industry represents the ideal and per-
sometimes at the document level. The main disadvantage of
fect document management scenario. However, standard in-
this approach is that the resultant business rules are static in
formation models have yet to be adopted at a large scale and
the sense that they do not change as the organization evolves.
their proliferation in the industry is hindered by many orga-
1 nizational, cultural, and technological barriers.
Res. Engr., Dept. of Civ. and Envir. Engrg., Univ. of Alberta, Ed-
monton, Alberta, Canada T6G 2G7. E-mail: [email protected]. This paper proposes a more practical approach, where com-
2
Prof., Dept. of Civ. and Envir. Engrg., Univ. of Alberta, Edmonton, pany, project and document data are defined based on a com-
Alberta, Canada. E-mail: [email protected] mon data model customized to a single sector of the construc-
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. tion industry. The specialized modules are developed for each
To extend the closing date one month, a written request must be filed sector of the industry incrementally toward industry-wide stan-
with the ASCE Manager of Journals. The manuscript for this paper was
dardization. Benefits of CDM systems can be immediately re-
submitted for review and possible publication on July 6, 1998. This paper
is part of the Journal of Computing in Civil Engineering, Vol. 14, No. alized as implementation takes advantage of specialization.
1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0070–0077/$8.00 Such a model was successfully implemented by the writers at
⫹ $.50 per page. Paper No. 18704. a steel fabrication and erection company.
70 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
OVERVIEW OF METHODOLOGY formation as the organizational structure, departments, and
employee data. The structure is customized to fit the needs of
Overall Document Management Process the target construction sector.
Electronic document management is typically a two-step ap- The project data model is a specialized product data model
proach. First, templates representing the structure of the man- that encapsulates all related project information. Most con-
aged documents are designed. This consists of the structure of struction companies possess a minimal project database in the
the document content (i.e., image, text, CAD file, . . . ) and form of an estimating and control system. The project orga-
the structure of the indexing information. Document owners nizational structure is typically hierarchical, based on a func-
can then create document instances based on the available tem- tional, process, or product breakdown.
plates and define the context and index data. Various types of The document data model provides means to define the
editors are typically used to define the content data. For ex- structure of typical construction documentation in a given con-
ample, an ‘‘image’’ content type requires an external graphics struction sector. This consists of the actual data and the user-
editor where users can open an existing file or capture an im- defined indexing information (sometimes referred to as the
age from a scanner. document envelope, reference, or meta data). Document con-
Clients can then use a search engine to locate and retrieve tent will typically consist of text, image, or other data in a
electronic copies of required documents. The search param- variety of custom formats. Indexing information contains data
eters can be based on the type of document required and the such as date, type, reference number, and other information
values of the index data. used to retrieve the document.
By using this shared model, the structure of the document
Integration Defined indices can be defined as a function of other model elements
using direct references to existing objects or the ability to cre-
For the purposes of this paper, integration will be defined ate project or company model elements as part of the docu-
as the ability to efficiently and automatically retrieve the set ment data structure. Fig. 2 provides an example of this con-
of all documents related to a given project or any of its com- cept. ‘‘Index 1,’’ for example, consists of a direct reference to
ponents. For example, a paint specification document for a a project division. ‘‘Index 2’’ links to a specific supplier on
given steel box columns should be retrieved based on the iden- the indicated project. ‘‘Index 3’’ links to a specific employee
tification of the column and not through explicit reference to from the company data model. ‘‘Index 4’’ consists of a steel
the document. This type of integration involves the explicit section material object. Although the structure of this object
definition of links between project and document index data is defined as part of the project data model, the document data
by the user and will be referred to as ‘‘Type A’’ integration. model can utilize and embed it as part of its indexing infor-
The other type of integration (which will be referred to as mation. This feature is important for ‘‘Type B’’ integration.
‘‘Type B’’) provides the ability to retrieve potentially related
documents based on project-level properties. For example,
consider a 10 m beam consisting of a 350 mm by 137 Kg/m
welded wide flange (WWF). Type B integration allows for the
automated retrieval of all documents containing WWF sections
ranked according to the difference between the section dimen-
sions and the desired values (i.e., 350 ⫻ 137 and 10 m long).
This implies a certain compatibility between the representation
of project and document data.
Type A document integration is implemented during the
definition of the template—specifically during the definition
of the document indexing structure. The index section of a
document can contain references to project/company elements.
Conversely, project/company elements can make a direct ref-
erence to a given document instance as well. These relation-
ships are set up during the document design phase.
FIG. 1. Common Data Model Elements
Integration through Shared Data Model
The main issue to be dealt with when attempting to integrate
two parts of the information system is the development of a
shared data model capable of representing all relevant infor-
mation in both components. This process facilitates the design
of the integration mechanism, as relationships can be embed-
ded in the common model. The development of a special pur-
pose data module involves a careful analysis of all aspects of
the product lifecycle from the point of view of a given sector
of the construction industry. This includes a definition of the
common business processes involved—estimating, scheduling
and control, typical product information such as material list-
ing, standard support information such as drawings and spec-
ifications, and a variety of company level information, such as
the company organizational structure. The overall model can
generally be divided into three categories, as illustrated in Fig.
1. Each category is defined in an object-oriented manner, tak-
ing full advantage of inheritance, encapsulation, and code-
sharing functionality. FIG. 2. Establishing Document Relationships between Infor-
The company data model defines such company level in- mation Systems

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 71

Definition of Document Templates for copies of all quality reports related to a specific project or
shop supervisor requests for all UERs related to a specific
One method of implementing template design capabilities welder. The amount of work associated with using the manual
would be to have design experts define an object-oriented class system caused the quality department to give low priority to
(C⫹⫹ class for example) and embed references to, or in- such requests and, as a result, valuable information was not
stances of, exposed project classes. This class can then be delivered when needed. A need was identified for an electronic
compiled and utilized by end users. The drawback of this ap- quality control document management system that is accessi-
proach is that it would require extensive programming exper- ble company wide. The quality department becomes respon-
tise. Furthermore, it becomes a requirement to recompile all sible for capturing, indexing, and integrating paper-based qual-
templates each time the project or company classes are revised. ity reports. Users are then able to retrieve documents on their
The approach advocated by the writers allows for the cre- own workstations as needed.
ation of objects or object references using a runtime object
creation technique and dynamic link libraries. The method fol-
Implementation Platform
lowed can be supported with most commercial object-oriented
development systems, such as Microsoft Visual C⫹⫹. A more Document content data was stored centrally and managed
detailed account of how this is implemented will be presented by the operating system as regular application files. Document
in the sample implementation section. index and project data was stored on a central relational da-
tabase management system (RDMS). The first step in imple-
Integrated Document Searching menting the integrated EDMS was the design and implemen-
tation of the required data models. A subset of the developed
By using a common data model and establishing the struc-
C⫹⫹ class hierarchy is shown in Fig. 3. The complete hier-
tures for linking documents to other data model elements, it
archy includes other classes that are not relevant in this paper
becomes possible to design a generic and integrated document
and are used for scheduling, purchasing, accounting, and ship-
search facility. ‘‘Type A’’ integration allows for searches based
ping purposes.
on explicit links in the document index data. For example, all
documents referencing a given project or company element,
such as an employee, can be retrieved by obtaining all docu- Company Data Model
ment instances that contain index data referring directly to the The three company-level classes shown are CDepartment,
desired element (i.e., the employee). ‘‘Type B’’ integration, CEmployee, and CSupplier. They are included to illustrate
which was also made possible through the use of a common how they can be linked to document-indexing information.
data model, allows for fuzzy searches based on ‘‘object match- Each class encapsulates all relevant information. For example,
ing’’ rather than ‘‘object referencing.’’ the CSupplier class contains properties to represent the sup-
plier’s name, address, designation, and phone number.
SAMPLE IMPLEMENTATION
The discussed approach was used in the development of an Project Data Model
integrated document management system for application in the Structural steel fabrication and erection projects are pre-
structural steel fabrication and erection industry. The system sented hierarchically at the collaborating company. A project
was implemented with the help of a collaborating company is divided into many jobs at the highest level and followed by
for the management of quality control reports as a starting further subdivisions. At the lowest level is a collection of steel
point. The following sections provide an overview of the pre- material such as welded flanges, pipes, plates, and angles. The
vious manual document management processes leading into classes used to represent this information are CProject, CJob,
an identification of the needs for integrated electronic man- CDivision, CPiece, and CMaterial. The generic CCollection
agement. This is followed by a detailed description of the inner class is used by most classes to maintain relationships between
workings of the developed system. class instances. A collection can hold a list of objects or object
Typical EDMS components such as scanning, image manip- references.
ulation, and distribution were implemented using third-party A graphical input dialog box is created for each class to
plug-ins. The main focus of the development effort, as well as allow for the manipulation of its properties in a graphical and
this section, is not on the system itself, but rather on the design user-friendly manner. The dialog box for the CMaterial class
and implementation issues of the integration-related compo- is shown in Fig. 4. It allows for the definition of the material
nents. The objective is to illustrate how the discussed approach type (plate, welded wide flange, etc.) and the material prop-
can be accomplished in a real world setting. erties (depth, weight, grade, length, etc.).
Overview of Manual Process Document Data Model
The previous process included a manual paper-based index- Document-level classes provide the capability to create
ing scheme for various types of quality control documentation. and manage document templates and the means to instantiate
Examples of such documents are mill test reports (MTR), these templates to create the actual document objects. The
magnetic particle inspection reports (MPIR), and ultrasonic ex-
amination reports (UER). Most of these documents include
information related to a given structural steel piece, a shop
employee, and an inspector. These reports are often required
by engineers, project managers, and owners of the constructed
facility. An example request from within the company is an
engineer requesting the MTR for a certain piece while design-
ing a connection. In this case, the manual index is searched
for the specified beam, then the reference number is used to
locate the paper copy from the quality department’s report li-
brary. Once the MTR is found, a photocopy is made and for- FIG. 3. Developed Class Hierarchy for Representation of Proj-
warded to the engineer. Other examples include client requests ect, Company, and Document Information

72 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 Document Template Instances
Instantiation support for document templates is provided by
the CDocument and the CDocumentIndex classes. The graph-
ical dialog box shown in Fig. 6 is used to manipulate a doc-
ument instance. The user first selects from a list of available
document types that is obtained from the CDocument-
TCollection. Selecting a given type causes the content section
and the indexing information section to be initialized to cor-
respond to the selected template definition. Indexing values
can be manipulated using the shown ‘‘Edit’’ button. The dia-
log box presented depends on the index type and the refer-
FIG. 4. Graphical User Dialog Box for CMaterial Class encing method illustrated in Table 1.

CDocumentT and CDocumentIndexT classes are used to Integrated Search Module

manage document templates. The CDocument and the
A search module that allows for the retrieval of documents
CDocumentIndex classes are used to manage document tem-
based on project or company specifications was developed.
plate instances.
The main user interface is shown in Fig. 7. A project or com-
pany element such as a division, drawing, or a piece is first
Document Templates selected. At this point, users can perform a search using any
of the two buttons shown. The ‘‘Find Linked Documents’’
Document templates are defined by the document designer button performs a search based on ‘‘Type A’’ integration (i.e.,
using the CDocumentT dialog box shown in Fig. 5. The tem- documents that contain an explicit reference to the selected
plate name is used to identify a given document type. The object in their indexing information section). The ‘‘Find Re-
document content can be set to any available OLE (object lated Documents’’ button initiates an ‘‘object matching’’
linking and embedding) type. OLE is an application interfac- search based on ‘‘Type B’’ integration. First, all documents
ing technology, developed by Microsoft, that allows for one that contain class instances of the same type as a user-specified
application to manage another application’s data. Depending
on the content of the document, designers will typically select
an OLE object type that is easily accessible on all company
desktops. OLE technology provides electronic document man-
agement with the ability to manage document content in a
generic and uniform manner for such operations as saving,
loading, and printing.
Document template indexing information is accessed using
the bottom portion of the shown dialog box in Fig. 5. Each
row represents a single index. The index ‘‘Type’’ column can
be set to any standard data type or a data model class such as
CSupplier, CDivision, and CMaterial. The index’s ‘‘Reference
Type’’ column establishes the type of connection between the FIG. 6. Graphical User Dialog Box for CDocument Class
document object and the selected index type. The ‘‘is an in-
stance of’’ relationship is used when the document needs to TABLE 1. Selection Rules of User Dialog Boxes Based on
embed and own the class instance, such as the case of a doc- Index Definitions
ument reference number or a document date. The ‘‘is a ref-
erence to’’ type of relationship is used to define a link to an Reference type Class Displayed dialog box
existing object that is contained in another object or collection. (1) (2) (3)
The other type of referencing provided is ‘‘is a collection of Is an instance of Standard data type simple data entry text box
instances of’’ and ‘‘is a collection of references.’’ These be- Data model class dialog box associated with se-
have in the same manner as the previous two types but instead lected class
Is a reference to Standard data type invalid (not allowed)
manage a collection rather than a single object. Data model class dialog box associated with
source collection; certain
collection such as CMain-
Collection present a hierar-
chical selection view, others
display a simple list or grid
based views
Is a collection of Standard data type grid-based view with each row
instances of containing data value
Data model class grid-based view with each row
containing custom represen-
tation of contained object
data; from grid view, indi-
vidual rows can be selected
to manipulate properties of
objects
Is a collection of Standard data type invalid (not allowed)
references to Data model class grid-based view with each row
containing custom represen-
tation of referenced object
data
FIG. 5. Graphical User Dialog Box for CDocumentT Class

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 73

 underlying algorithms for the two search procedures are sum-
marized in Fig. 8.

Example Document Templates and Template

Instances
An example will now be provided to illustrate the process
of creating a document template, instantiating the defined tem-
plate, establishing the links to the project data, and then per-
forming an integrated search. The actual system includes the
complete set of quality document types, including ultrasonic
examination reports, fitting inspection reports, and welding in-
spection reports.
MTRs, or mill test reports, are provided by the supplier
upon steel delivery and indicate the various alloy properties.
The information content includes supplier information, a ref-
FIG. 7. Graphical User Dialog Box for Search Module erence number, a material listing, and the steel properties.
Users will typically search MTRs based on supplier informa-
tion or material specifications. As a result, the indexing section
class will be selected. The documents found are then ranked of the document template should contain this information at a
according to a comparison between the property values of the minimum.
selected class instance and the property values of the matched A template is first built to represent an MTR document,
instances from each document. A percentage rank is then as- using the editor previously shown (Fig. 5). The file type is set
signed based on the number of properties equal. The list of to ‘‘Image Document.’’ The indexing information section con-
documents found can also be filtered based on the document tains four types of indices: (1) a reference number used for
type and any user-defined constraints on the index values. The filing; (2) a date field to maintain the document creation date;

FIG. 8. Search Algorithms for Two Supported Methods

74 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

 FIG. 9. CDocumentT Class Instance for Defined MTR Template

(3) a reference to a supplier object; and (4) a collection of

material object instances representing the material on the MTR
document. The underlying data structure representation for the
template is illustrated in Fig. 9.
Once the template is defined, instances can be created by
the user as shown in Fig. 10. The document content can be
enabled by clicking on the shown button. This activates the
OLE document server which in this case is the user’s default
image editing program. Pressing the ‘‘Edit’’ button in the in-
dexing information section displays the corresponding dialog
box, as outlined in Table 1. Clicking the edit for the material
instance collection index displays a list of materials. Individual
material entries can then be manipulated using the definition FIG. 10. Mill Test Report Template Instantiate

FIG. 11. Instantiation Representation of MTR Document Template

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 75

dialog for the CMaterial class, which was shown in Fig. 4. ity assurance department. After the implementation of the sys-
The resultant data structure for the document instance is illus- tem, they gained the ability to search and view the MTR
trated in Fig. 11. The first two indices (Number and Date) are documents directly from their workstations. The results sum-
based on a standard data type so their value representation is marized in Table 2 indicate that marginal improvements were
simple. The ‘‘Supplier’’ index representation is a reference to observed for the basic types of document access (reference
an existing ‘‘CSupplier’’ object maintained by the number, supplier, and heat number). This is because the man-
‘‘CSupplierCollection’’ object. The ‘‘Material List’’ index is ual system had been optimized to use these fields for locating
a collection of ‘‘CMaterial’’ objects. documents quickly. However, for document searches based on
Once an MTR document is scanned and indexed it becomes such complex indices as material, the true benefits of the sys-
available to all computers connected to the company network. tem were observed to be significant. The estimated yearly ben-
Engineers requiring the steel properties for a given material efits are close to 444 person-hours.
need only start the integrated search module, select the ma- The mentioned improvements reflect the expected benefits
terial in question, filter the resultant documents reports to associated with any kind of automation. However, it is the
‘‘Mill Test Report,’’ and search for related documents. A list writers’ feeling that a major portion of the improvement—as
of matched MTRs will be displayed ranked according to the much as an estimated 20%—would not have been possible if
material properties (section type, dimensions, grade, and heat the integration approach presented in this paper was not fol-
number). lowed. This can be proven only if a parallel team had at-
tempted to solve the same problems with a commercial system.
Process Improvements due to New Approach It is estimated that excessive and extra setup costs as well as
further maintenance costs would definitely be incurred as new
The improvements observed after the implementation of document types become required or as the structures of the
electronic quality document management were recorded based company and product data models evolve. Further, three major
on a number of interviews with the employees utilizing the features of the system (automated validation, complex refer-
system. A breakdown of the realized annual savings for MTR ence data, and object-based searching) would not have been
(mill test report) related activities is provided in Table 2. The easily implemented, as explained in Table 3.
first column (1) lists the various document management activ-
ities performed. The second and third columns (2, 3, 4, 5) System Implementation Issues
provide the observed duration for the listed activities before
and after the implementation of the system for both the quality The implementation of the described approach with the col-
assurance (QA) and the engineering (ENG) departments. The laborating company would not have been possible without the
frequency of these activities was also obtained and used in support of upper management. This helped in overcoming
calculating the yearly savings. Prior to the system’s imple- some of the initial resistance from the employees. Implemen-
mentation, engineers obtained MTRs indirectly from the qual- tation was further facilitated through a policy of open, two-

TABLE 2. Process Improvements due to Implementation of Integrated Document Management of MTRs

Frequency of
Average Duration Average Duration Operation (events
Estimated yearly savings
before (min)a after (min)b per work year)
in person-hours
Specific operation QA ENG QA ENG QA ENG (2 ⫺ 4) ⫻ 6 ⫹ (3 ⫺ 5) ⫻ 7
(1) (2) (3) (4) (5) (6) (7) (8)
Search for and obtain hardcopy of MTR
By reference number 1 10 0.2 0.2 83 50 8.4
By supplier 1 1 0.2 0.2 50 17 0.9
By material 6 15 0.2 0.2 250 25 30.3
By heat number 1 10 0.15 0.15 25 30 5.3
Generate client turnover package 2,400 0 5 0 10 0 399.2
a
Durations based on subjective assessment of previous process.
b
Durations based on time studies of current computer system.

TABLE 3. Difference in Implementation between Commercial EDMS and Presented Approach for Three Major Features

1. Validation Static validation in the form of user-defined rules or Validation rules are automatically inherited from the business
code. Rules are static in the sense that they don’t rules defined as part of the underlying data model. In other
change as the organization evolves. words, validation is a function of the structure of the doc-
ument itself. As company and project data are altered, so are
the validation rules. More importantly, as the structure of the
data itself changes, the structure of the validation is auto-
matically updated.
2. Complex reference Only advanced systems allow for customization of the With a common data model, all the business objects (employee
data document reference structure. For those systems, the information, material, drawing, etc.) are available for docu-
process usually requires programming skills in a pro- ment designers during the template definition stage.
prietary macrolanguage.
3. Searching based on Object-based searching is almost impossible to imple- With an integrated approach, changes to any portion of the
‘‘object matching’’ ment with EDMS. The structure of the product overall data model—which include the document, project
and company data model would have to be com- and the company information—are automatically reflected
pletely understood by the document designer and in all client systems.
complex data manipulation code would have to be
written.

76 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

way communications regarding the system features between defined makes it possible for integrated project- or company-
potential users and the research team. The quick adoption of based searching to be performed.
unforeseen user requirements and suggestions for enhance- The presented approach was followed during the implemen-
ments eliminated any mistrust and concerns. Minimal formal tation of a quality document management system for a steel
training was required for document owners and information fabricating company. Observed results indicated that a clear
clients because they were continually updated on the progress improvement was realized not only because of automation, but
of the development and provided with the necessary user man- also because of the way relationships were represented.
uals. This was also due to a robust and user-friendly graphical
user interface. Several rollout sessions and demos were given APPENDIX. REFERENCES
to inform potential users about the information now accessible ‘‘AutoEDMS product overview.’’ (1999). 具http://www.acssoftware.com/
from their workstations. html/autoedms.html典.
It was discovered that the document design process was Dubois, A. M., Flynn, J., Verhoef, M. H. G., and Augenbroe, F. (1995).
difficult for novice users. As a result, the research team closely ‘‘Conceptual modelling approaches in the COMBINE project.’’ Proc.,
trained and assisted in the initial design of the presented qual- 1st ECPPM Conf., European Association for Product and Process Mod-
ity assurance reports. The trained users are now able to design eling in the Building Industry, Dresden, Dublin.
Froese, T. (1998). ‘‘Information modules of construction processes.’’
new documents as required with minimal assistance. 具http://www.civil.ubc.ca/⬃tfroese/rsch/models/index.html典.
‘‘In-sight product overview.’’ (1999). 具http://in-site.bidcom.com/html/
CONCLUSIONS overview.html典.
‘‘Optix product overview.’’ (1999). 具http://www.blueridge.com/optix-
EDMS can significantly improve the overall document man- faq.html典.
agement process. However, the unique aspects of the construc- Rezgui, Y., and Debras, P. (1996). ‘‘An integrated approach for a model
tion industry make it a requirement that relationships between based document production and management.’’ Electronic J. Info.
documents and the construction process be explicitly defined. Technol. in Constr., 具http//itcon.fagg.uni-lj.si/1996/1/index.html典.
Turk, Z., Bjork, B. C., Johansson, K., and Sevensson, K. (1994). ‘‘Doc-
In this paper, an approach was presented based on specialized ument management systems as an essential step towards CIC.’’ Proc.,
product data models designed in an object-oriented fashion. CIB W78 Workshop on Computer Integrated Construction, Interna-
This approach requires that designers first create document tional Council for Building Research Studies and Documentation, Hel-
templates. The way in which these document templates are sinki, Finland.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 77

 ARTIFICIAL NEURAL NETWORK FOR MEASURING
ORGANIZATIONAL EFFECTIVENESS
By Sunil K. Sinha1 and Robert A. McKim2

ABSTRACT: An artificial neural network based methodology is applied for predicting the level of organizational
effectiveness in a construction firm. The methodology uses the competing value approach to identify 14 variables.
These are conceptualized from four general categories of organizational characteristics relevant for examining
effectiveness: structural context; person-oriented processes; strategic means and ends; and organizational flexi-
bility, rules, and regulations. In this study, effectiveness is operationalized as the level of performance in con-
struction projects accomplished by the firm in the past 10 years. Cross-sectional data has been collected from
firms operating in institutional and commercial construction. A multilayer back-propagation neural network based
on the statistical analysis of training data has been developed and trained. Findings show that by applying a
combination of the statistical analysis and artificial neural network to a realistic data set, high prediction accuracy
is possible.

INTRODUCTION of an industry that is a conglomeration of diverse fields and

The assessment of organizational effectiveness represents an participants. Distinctive features that influence the organiza-
important step in the performance improvement of a construc- tional effectiveness of a construction firm are the contractual
relationships that are established when forming project orga-
tion firm. At the present time, many techniques and tools are
nizations (i.e., general contracting, subcontracting, joint-ven-
being advocated by practitioners and theorists to improve or-
turing, partnering, alliances, and consortiums). These relation-
ganizational effectiveness. Concepts such as total quality man-
ships result as a consequence of how the different types of
agement, re-engineering, partnering, conformance to ISO stan-
firms in the industry organize to procure services and deliver
dards, and other emerging management strategies are making
the constructed product. This study focuses on developing a
headlines. Some of these techniques may have little effective-
methodology based on an artificial intelligence technique to
ness when applied by construction firm, whereas others may
predict organizational effectiveness of construction firm.
have more lasting effects. However, all of the techniques stress
Artificial neural network (ANN) techniques open new pos-
measurements and continuous assessments of how the firm is
sibilities in classification and generalization of the available
organized as important steps in improvement. Therefore, it is data, by learning from examples that predict organizational
clearly becoming essential for construction firms to develop effectiveness of a construction firm. If such a mapping can be
valid methods of assessing and predicting their level of orga- effectively modeled in a neural system, in spite of data com-
nizational effectiveness. Utilizing a valid assessment method plexity, incompleteness, and incoherence, it might be useful
will enable construction firms to maintain their effectiveness for a construction firm as a new tool that supports the decision
and, hence, achieve consistency in the project’s performance. process and improves its efficiency. Many other successful ap-
Robbins (1990) defined organizational effectiveness as the plications of this technology are found in civil engineering
degree to which an organization attains its short-term and uses as follows: Karunanithi et al. (1994) successfully used an
long-term goals, the selection of which reflects strategic con- ANN to predict the flow characteristics of a complex river
stituencies, the self-interest of the evaluator, and the life stage system; Adeli and Wu (1998) used a regularization network to
of the organization. Nadler et al. (1992) advocated an emerg- predict construction cost estimation; Sinha and McKim (1997)
ing management tool that they called organizational architec- used an ANN model for forecasting construction cost escala-
ture. They proposed an organizational model that underlines tion; and Flood and Katim (1994) used an ANN based model
the importance of achieving two types of fits for the organi- to perform structural analysis problems. Moselhi et al. (1991)
zation to achieve organizational effectiveness: an internal and indicated that ANNs can be applied to models representing
external fit. Traditional contingency theory, according to Miller such areas as soil analysis, optimization, construction dispute
and Fresien (1984), found that organizations must change their resolution, and real-time applications.
internal attribute structure, strategies, and processes, to cope Nonetheless, a question can be posed: How to find and train
with changes in the environment. Hitt (1988) discussed the an effective architecture of the neural network? To answer this
measurement of effectiveness and its importance in the crea- question, which is of primary importance for improving ANN
tion and design of effective organizations. accuracy in engineering domains, neural network architecture
Organizational structure, strategy, and cultural factors is applied that has at the hidden layer(s) nodes corresponding
mainly determine the level of organizational effectiveness of to ‘‘significant’’ factors determined by statistical analysis of
the construction firm (Adas 1996). These factors influence the data. This is particularly important when neural networks are
adaptation to the environment of the construction industry. used to perform the classification tasks. The dimensionality
Construction firms operate in a series of overlapping markets reduction of the input not only removes the redundancy of the
1 data but also enables the use of a smaller size network struc-
Grad. Student Res. Asst., Constr. Engrg. and Mgmt., Dept. of Civ.
Engrg., Univ. of Waterloo, 200 University Ave. West, Waterloo, ON, Can- ture, which can be trained easier and has improved generali-
ada N2L 3G1. zation capability. This paper concerns a rather simple case, in
2
Prof., Constr. Engrg. and Mgmt., Dept. of Civ. Engrg., Univ. of Wa- terms of the dimensions of the decision space, but because the
terloo, 200 University Ave. West, Waterloo, ON, Canada N2L 3G1. experimental results achieved so far look encouraging, the
Note. Editor: Sivand Lakmazaheri. Discussion open until June 1, 2000. method will be further developed.
To extend the closing date one month, a written request must be filed First, the concepts of neural network are described and for-
with the ASCE Manager of Journals. The manuscript for this paper was
submitted for review and possible publication on November 4, 1997. This
mulated mathematically. Next, the methodology for measuring
paper is part of the Journal of Computing in Civil Engineering, Vol. effectiveness of construction organizations and statistical anal-
14, No. 1, January, 2000. 䉷ASCE, ISSN 0887-3801/00/0001-0009–0014/ ysis of data are described. Then, the synthesis of network and
$8.00 ⫹ $.50 per page. Paper No. 16906. results are presented.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 9
LEARNING SYSTEM
The most fundamental problem that neural networks have
been used to solve is that of ‘‘learning.’’ But it is very difficult,
if not impossible, to define learning precisely. To model learn-
ing computationally, however, it has to be defined in a prag-
matic manner rather than as an abstract concept. Learning can
be defined as a self-organizing process, a mapping process, an
optimization process, or a decision-making process. The last
definition is based on the observation that given a set of ex-
amples and a stimulus, one makes a decision about the re-
sponse to the stimulus.
Consider the special case of supervised learning, where the
system is first trained by a set of input-output examples. Then,
given a new input the learner decides the output. This is an
ill-posed problem because the answer can be any multitude of
values. The selected answer depends on the generalization cri-
teria or constraints chosen for the decision process. The ad- FIG. 1. Typical Feedforward Back-Propagation Neural Net-
vantage of viewing the learning process as a decision-making work Architecture
process is its explicit representation of the generalization cri-
teria. ⌬w [s]
ij (k) = ␩(t)␦ pj x j
[s] [s⫺1]
⫹ m⌬w [s]
ij (k⫺1) (2)
A prediction problem can be formulated as a supervised
learning problem. Consider a one-input–one-output noisy sys- Second, for output nodes it holds that
tem with input x and output y. The system can be expressed
␦ [o]
pj = (dj ⫺ oj ) f ⬘
[s]
j Ij (3)
mathematically as
and third, for the remaining nodes it holds that

冘
y = f (x) ⫹ e (1)
where f (x) = function of the input variable x; and e = inde- ␦ [s]
pj = f ⬘
[s]
j (I j ) ␦ pk
[s⫹1]
w [s⫹1]
jk (4)
k
pendent noise function with zero mean. A set of input-output
examples xi and yi (i = 1, 2, . . . , N ) is given. The prediction where x [s]
j = actual output of node j in layer s; w ij[s] = weight
problem is: For any given input x find the best estimate of the of the connection between node I at layer (s ⫺ 1) and node j
ouput y. The best estimate can be defined as the estimate that at layer (s); ␦ [s] [s]
pj = measure for the actual error of node j; I j =

minimizes the average error between the real and estimated weighted sum of the inputs of node j in layer s; ␩(t) = time-
outputs. Thus, such a supervised learning problem becomes dependent learning rate; f ( ) = transfer function; m = momen-
equivalent to a mapping or curve-fitting problem in a multi- tum factor (between 0 and 1); and dj, oj = desired and actual
dimensional hyperspace. The traditional statistical approaches activity of node j, respectively (for output nodes only).
to curve fitting, such as the regression analysis, fail to repre- A misclassification measure between the network output and
sent problems with no explicit mathematical model accurately the expert’s is defined as
in a multidimensional space of variables. On the other hand,
冋冘 册
c 1/2

the neural network approach can solve such problems more E= (uik ⫺ u⬘)
ik
2
(5)
effectively. i=1

ANN where E = measure of classification error; c = number of

classes; uik = membership value of the kth observation in the
A neural network consists of a number of nonlinear com- ith class as assigned by the experts; and u⬘ik = membership
putational processing elements (PEs), arranged in several lay- value of the kth observation in the ith class as estimated by
ers including an input layer, an output layer, and one or more the network.
hidden layers in between. Every layer contains one or more The problem with the multilayer feedforward back-propa-
PEs, and the output of every PE is usually fed into all or most gation neural network is that the generalization properties (un-
of the inputs of the PEs in the next layer. The input layer derfitted, overfitted, or properly fitted) depend on many factors
receives input signals, which are then transformed and prop- including the architecture (number of hidden layers and num-
agated simultaneously throughout the network. A PE accepts ber of nodes in the hidden layers), initial weights of the links
one or more input signals and produces one output, which is connecting the nodes, and number of iterations for training the
a nonlinear function of the weighted sum of inputs. system. The performance of the algorithm depends highly on
Back-propagation, which is a gradient-decent learning al- the selection of these parameters. The problem of arbitrary
gorithm, is one of the most popular supervised learning al- trial-and-error selection of the learning and momentum ratios
gorithms (Robel 1994). It back-propagates the error signals encountered in the momentum back-propagation learning al-
from the output layer to all the hidden layers, so that their gorithm can be circumvented by the adaptive conjugate gra-
weights can be adjusted accordingly. Back-propagation is a dient learning algorithm developed by Adeli and Hund (1994).
generalization of the least-mean square procedure for feedfor- However, that algorithm does not address the issue of deter-
ward, multilayered networks with hidden layers (Haykin mination of the number of hidden layers and number of nodes
1994). It uses a gradient descent technique, which changes the in each layer for satisfactory network performance. In this
weights in its original and simplest form by an amount pro- study, neural network architecture based on the statistical anal-
portional to the partial derivative of the error function with ysis of the construction data is synthesized to obtain the prop-
respect to the given weight. As shown in Fig. 1, the back- erly fitted approximation function.
propagation network is a layered, feedforward network com-
prised of one input layer, one or more hidden layers, and one METHODOLOGY FOR MEASURING EFFECTIVENESS
output layer. The back-propagation algorithm can be described
in three equations. First, weight connections are changed in The neural network based prediction model is developed in
each learning step k with three stages. First, organizational characteristics’ categories
10 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
and variables relevant for examining effectiveness of the con- TABLE 1. Categories and Variable of Organizational Effective-
struction firm are identified. Second, a field study is designed ness
and carried out. In this study, identified variables are opera- Independent and dependent
tionalized along with a measure of organizational effective- Category of variables variables
ness, scales of measurement are constructed, and cross-sec- (1) (2)
tional data are collected from a sample of construction firms Structural context Level of integration of services
using self-administered questionnaires. Different statistical Level of joint-venturing
techniques such as principal component analysis (PCA), linear Level of subcontracting
regression, and correlation are then performed to find the sig- Level of multiple project-handling
nificant factors in a data set. Synthesis, training, and validating Level of coordination
a neural network using the collected data represent the third Level of information flow
Flexibility, rules, and regulations Attitude toward change
stage. Level of rules and regulations
Level of adherence to rules
Level of process control
Organizational Effectiveness of Construction Firm Person-oriented processes Strength of culture
Level of workers’ participation in
Criteria of organizational effectiveness emphasized by the decision making
four models of the competing value approach have been used Strategic means and goals Level of planning
to identify variables and are applied as input in the developed Level of goal setting
network. The approach was proposed by Quinn and Rohor- Organizational effectiveness Measure of organizational effec-
baugh (1983) and is chosen here as a valid approach on which tiveness
to identify criteria of organizational effectiveness for the con-
struction firm. The reasons being, first, its four configurations
or models represent the integration of most effectiveness cri- Measuring Variables
teria already used by researchers and managers. Second, Ma- After all variables are operationalized and measurement
loney and Federle (1993) already validated it for construction scales, with different items to score their levels, were con-
organizations. structed, a field study is designed to collect the relevant data.
Each of the four models emphasized different criteria. The Raw scores for all variables except strength of culture are
rational goal model emphasizes control and organizational fo- based on responses to a seven-point ordinal Likert scale (‘‘very
cus as dominant effectiveness values, with planning and goal high’’ 7 to ‘‘very low’’ 1) (Furnham and Gunter 1993). Or-
setting as means, and productivity and efficiency as ends. The ganizational culture is measured using the constant method
open system model emphasizes flexibility and organizational scale that was used by Maloney and Federle (1993) to measure
focus as dominant effectiveness values, with readiness and their six major aspects of culture. The resulting scores are
flexibility as means, and growth and external support as ends. converted to a seven-point Likert scale to be compatible with
Dominant effectiveness values for the internal process model the scores of other variables.
are control and internal focus, stressing communication pro-
cesses as means and control as ends. The human-relations
Measuring Organizational Effectiveness
model emphasizes flexibility and internal focus, with cohesion
and morale as means and skilled workers as ends. Level of overall performance is used to operationalize ef-
An organization could pursue the criteria emphasized by fectiveness. It is determined by the three most commonly used
one or more of the four models. A construction firm that pur- indicators in the construction industry: time, cost, and quality
sues strategic flexibility must have a structure that is better of performance in completed work projects. In this study, the
suited for contractual arrangements that provide flexibility average level of performance over the last 10 years of oper-
such as subcontracting, joint venturing, partnering, and alli- ation is calculated from the data records collected from the
ances. The firm’s style of management of its workers, its cri- sampled firms. It is calculated as the average percent of
teria of success, and type and strength of organizational culture projects completed within scheduled time or less, percent of
that exists in the organization influences its focus. Internally projects that are completed within budgeted costs or less, and
focused construction firms that value flexibility in the assess- percent of projects that are completed without any claims and
ment of effectiveness are more sensitive to their workers. Ex- within acceptable level of compliance to clients specifications.
ternally focused construction firms emphasize adaptability The calculated level of performance is given by values be-
through a strong culture that promotes readiness and flexibility. tween 0 and 1, where a value of 0 indicates no effectiveness
In summary, the analysis identifies 14 variables that are and a value of 1 indicates 100% effectiveness.
deemed related to organizational effectiveness in four general
categories. As seen in Table 1 these categories are structural DATA COLLECTION METHODOLOGY
context; flexibility, rules, and regulation; strategic means and
goals; and person-oriented processes. In the category of struc- Data may be collected about the causes of the problem and
tural context, the study hypothesizes six variables that can be its seriousness from a variety of sources including interviews,
used to indicate the effectiveness of the structural component observations, questionnaire surveys, and archival data. All data
of the construction firm in dealing with its external and inter- collection methods need to satisfy two basic requirements: The
nal environment. In the second category of flexibility, rules, data must be reliable, and the data must be valid. Each method
and regulations, the study hypothesizes four variables that in- of data collection contains advantages and disadvantages. This
fluence the organizational structure in the flexibility-control study uses the questionnaire method of data gathering in its
dimension. All these variables together reflect the effort of the methodology. Two types of self-administered questionnaires
firm to mediate between how flexible it should be and how are used in this study: A questionnaire that targeted manage-
much control it should exercise over its processes to achieve ment of the construction firm, and a second one that consists
effectiveness. The third category is based on the dimension of of only parts of the first one and is targeted at the workers
focus, and it represents person-oriented processes in the firm, level in the firm. The reason that a single questionnaire is not
and the fourth category represents strategic means used by the used in this survey is because a number of questions regarding
firm to achieve its effectiveness. data concerning the firm could not be answered at the worker’s
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 11
level. For example, the scales that measured the level of sub- tau, with their significance levels. For quantitative, normally
contracting, level of integration in services offered, level of distributed variables, Pearson’s correlation coefficient method
planning, and level of goal setting are omitted from the is useful. If data are not normally distributed or have ordered
worker’s level questionnaire. categories, using Spearman’s rho or Kendall’s tau is the best
In designing these questionnaires to collect the necessary option that measures the association between the rank orders.
data for the variables of the study, as discussed in the previous Pearson’s correlation was conducted for construction data set.
section, Likert scales are used. Each scale is composed of var- It is observed that six variables are the major predictors for
ious items that attempt to measure the underlying attributes of the output in the correlation analysis.
the particular variable. Open-ended questions are strictly
avoided, which eliminated the need for coding of answers by Regression Analysis
respondents, and the raw data collected are transferred directly
into a computer format. In the interest of being unambiguous Linear regression estimates the coefficients of the linear
and precise, items are made as short as possible so that the equation, involving one or more independent variables, that
rater could understand without any misinterpretation and re- best predict the value of the dependent variable. The linear
spond quickly without any difficulty. regression analysis is carried with 14 independent variables
The firms considered in this study are comprised from in- and a dependent variable. The stepwise regression analysis be-
dustrial and commercial firms operating in the construction gins with all of the candidates’ variables, removes the least
industry. Approximately 100 firms are contacted to receive significant predictor at the first step, and continues until no
their assessments as far as the measurement of organizational ‘‘insignificant’’ variables remain. It is observed that eight var-
effectiveness is concerned. This study only considered those iables are the major predictors for the output in the stepwise
firms that are officially registered as such and that have the regression analysis.
appropriate license to operate in such capacities and have been
licensed for the last 10 years to handle contracts in industrial SYNTHESIS OF NETWORK AND RESULTS
and commercial construction. To address the specific task of developing a model that will
predict the effectiveness of construction organizations using
STATISTICAL ANALYSIS OF DATA the 14 dependent variables, five successive networks are syn-
thesized. Each network represents a stage in the evolution of
The ability of an ANN is to learn from examples and gen- the final model. The various combinations of significant pre-
eralize the knowledge to new situations. Effective generali- dictors, obtained by using different statistical techniques, are
zation, however, is difficult to achieve and is mainly deter- used to evaluate the performance of the network. Only the
mined by measuring accuracy of ANN mapping to an testing results are included in this paper because the testing
independent testing set. Many techniques have been developed data are unknown to the network, and hence the error pro-
for improving the network performance. One of the most suc- duced with it can exhibit the network’s mapping ability. The
cessful methods is the minimization of complexity in network performance of an ANN is evaluated on the basis of the root-
architectures by preprocessing the data. In this paper an at- mean-square error (RMSE). RMSE is defined according to the
tempt has been made to reduce training time and improve the following equation:
performance of an ANN by performing statistical analysis of
data. Different statistical techniques such as PCA, linear re-
gression, and correlation are performed to find the significant
factors in a data set. The predicted significant factors are used
RMSE = 冑冘 N

i=1
{Y ⫺ Y *}2/N (6)

to construct the back-propagation network architecture.

where Y = actual output; Y * = predicted output; and N = num-
ber of observations. Network weights and the amount of time
PCA needed to perform the training time are also computed. The
PCA is a statistical technique that linearly transforms an amount of time needed to perform one training cycle is given
original set of variables into a substantially smaller set of un- as (Skapura 1996)
correlated variables that represents most of the information in Tsec = {[(W ⭈ Tf ) ⫹ (W ⭈ Tb) ⫹ (H ⭈ Tt)]Nt}Ne /M (7)
the original set of variables. A smaller set of uncorrelated var-
iables is much easier to understand and use in further analysis where W = total weights in the network; Tf = time for forward
than a large set of correlated variables. Through PCA many propagation; Tb = time for backward propagation; Tt = time
variables can be represented by a few components; thus PCA for transfer function; H = number of hidden PEs; Nt = number
can be considered as a feature extraction of a data compression of training examples; Ne = number of epochs or iterations; and
technique. In the PCA, it is always important to choose how M = machine operating time.
many components are needed to adequately represent the data. The input data to the neural network consist of the 14 de-
How many and which principal components to retain depend, pendent variables. The neural network is set up to predict the
in part, on the goals of the analysis. Kaiser (1960) recom- effectiveness of construction organizations. In this study five
mended dropping those principal components of a correlation different network architectures are constructed to predict the
matrix with an eigenvalue <1. Another criterion is to retain effectiveness of a construction firm. Initially, the network is
enough principal components to account for a given percent- set up with 14 PEs in input and two hidden layers, and the
age of variance (e.g., 90%). The analysis of construction firm network output is represented by a single neuron. This network
data resulted in four (eigenvalue 1.0 and more) and eight (ei- model is denoted as GEN (general model). The second net-
genvalue 0.5 and more) components representing 70 and 90% work model is set up based on PCA (principal component
of variance. analysis model). Thus, the network comprises 8 PEs in input
and the first hidden layer and 4 PEs in the second hidden layer.
Correlation Analysis In the third model, the regression analysis results are used to
construct the network architecture. Input and two hidden layers
Correlation is used to measure how variables or rank orders consisting of eight neurons each are required. The third model
are related. Normally the correlation procedure computes Pear- is denoted as REGR (regression analysis model). The fourth
son’s correlation coefficient, Spearman’s rho, and Kendall’s model is denoted as CORR (correlation analysis model) and
12 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000
is based on correlation analysis results. This model is set up It is interesting to note that the most complex general net-
with 6 PEs in input and two hidden layers each. The fifth work was not the most accurate for predicting organizational
model is constructed using the concept of the GEN and PCA effectiveness. In comparison with the developed networks, the
models, the appropriate number of neurons in the hidden layer PCA-ANN provides better control to what the network has
are chosen according to the information obtained from PCA learned, and reduction in training time is also observed for the
of the construction data set. Thus, the network architecture is selected training cases. The results show that the ANN devel-
14 PEs in input layer, 8 PEs in first hidden layer, and 4 PEs oped can be used with a very good degree of reliability to
in the second hidden layer. This model is denoted as PCA- predict the level of effectiveness of the construction firm as
ANN (principal component neural network model). operationalized in the study. The use of this method allows
From the outset it is decided that hyperbolic tangent (TanH) the construction firm to predict its level of effectiveness and
transfer function and normal cumulative delta (NormCum) hence the level of overall performance independent of other
learning rule would be adopted for all five models. These net- performance measures. At the same time, the use of this meth-
works are tested for the level of error in prediction using odology can provide important feedback about the levels of
25,000 learning cycles. The parameters of the best performing significant organizational attributes that influence the firm’s ef-
networks are summarized in Table 2. The testing results of the fectiveness. It should be noted that the developed network is
neural networks in each category are graphically presented in intended as a tool that gives management an idea about how
Figs. 2 and 3. Fig. 2 shows the classification error of the five organizational characteristics, measured by the identified var-
developed networks, and Fig. 3 presents the comparison of the iables, influence the overall performance and not specifically
training time for different network architectures. rate performance on a project by project basis.
In this experimental study, it has been found that the neural
RESULTS networks are sensitive to changes of the learning rate. The
networks also appeared sensitive to changes in the momentum
Though theoretically possible, finding a good back-propa- and learning coefficient terms. The different initial conditions
gation network architecture capable of learning a particular set have influenced the network solution, but still most networks
of data may not be easy and may even fail sometimes. In this moved to a very similar solution. It is possible to deal with
study, the statistical analysis method has consistently shown the sensitivity of an ANN if groups of networks are trained
itself worthy to be considered as a technique for building good under various conditions. It is likely that the best network can
back-propagation network architecutre, particularly when the be found in this group. Further investigation should determine
collected data set is small and noisy. whether a better optimal model exists, but this goes beyond
the scope of this paper. One lesson learned from this study is
TABLE 2. Parameters of Developed Neural Networks that the exact learning coefficient and momentum terms of the
Hidden Hidden
processing element are not critical elements in the network
Input layer layer Output design. The exact size of the network architecture is very im-
ANN layer number 1 number 2 layer Learning Network portant for any given problem.
model PEs PEs PEs PEs cycle weights
(1) (2) (3) (4) (5) (6) (7) CONCLUSIONS
GENR 14 14 14 1 50,000 406
PCA 8 8 4 1 50,000 100 A neural network for predicting organizational effectiveness
REGR 8 8 8 1 50,000 136 of the construction firm is developed. Fourteen organizational
CORR 6 6 6 1 50,000 78
PCA-ANN 14 8 4 1 50,000 148
variables are identified and hypothesized to predict effective-
ness. The network is a valuable practical tool that can give
management of a construction firm operating in institutional
and commercial construction a preproject view of their per-
formance.
To improve the performance of the network, significant var-
iables through statistical analysis are identified, and then these
variables are used in the input and hidden layers of the ANN.
This has resulted in obtaining a simple and computationally
efficient network that can predict organizational effectiveness
of the construction firm.
The predictive capabilities of the neural network based
model developed in this study have only been compared with
other network based models. Further work in this area may
FIG. 2. Comparison of Classification Error for Five Developed look at comparing the predictive capabilities of the ANN based
Networks models with the predictive capabilities of the more traditional
statistically based models such as regression analysis.

APPENDIX. REFERENCES
Adas, A. A. (1996). ‘‘The level of organizational effectiveness of the
construction firm: A multivariate model-based prediction methodol-
ogy,’’ PhD thesis, University of Waterloo, Waterloo, Ont., Canada.
Adeli, H., and Hung, S. L. (1994). ‘‘An adaptive conjugate gradient learn-
ing algorithm for efficient training of neural networks.’’ Appl. Math.,
62(1), 81–102.
Adeli, H., and Wu, M. (1998). ‘‘Regularization neural network for con-
struction cost estimation.’’ J. Constr. Engrg. and Mgmt., ASCE, 124(1),
18–23.
Flood, I., and Kartam, N. (1994). ‘‘Neural networks in civil engineering.
FIG. 3. Comparison of Training Time for Five Developed Net- II: Systems and application.’’ J. Comp. in Civ. Engrg., ASCE, 8(2),
works 149–165.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000 / 13

Furnham, A., and Gunter, B. (1993). Corporate assessment: Auditing a in construction.’’ J. Constr. Engrg. and Mgmt., ASCE, 117(4), 606–
company’s personality. Routledge, London. 625.
Haykin, S. (1994). Neural networks: A comprehensive foundation. Mac- Nadler, D. A., Gerstien, M. S., and Shaw, R. B. (1992). Organizational
millan, New York. architecture: Designs for changing organizations. Jossey-Bass Publish-
Hitt, A. M. (1988). ‘‘The measuring of organizational effectiveness: Mul- ers, San Francisco.
tiple domains and constituencies.’’ J. Mgmt. in Rev., 28, 28–39. Quinn, R. E., and Rohorbaugh, J. (1983). ‘‘A spatial model of effective-
Kaiser, H. F. (1960). ‘‘The application of electronic computers to factor ness: Criteria towards a competing values approach.’’ J. Mgmt. Sci.,
analysis.’’ J. Educational and Psychological Measurement, 20, 141– 29, 363–377.
151. Robbins, P. S. (1990). Organization theory: Structure, design, and ap-
plications. Prentice-Hall, Englewood Cliffs, N.J.
Karunanithi, N., Grenney, W. J., Whitley, D., and Bove, K. (1994). ‘‘Neu-
Robel, A. (1994). ‘‘Dynamic pattern selection: Effectively training back-
ral networks for river flow prediction.’’ J. Comp. in Civ. Engrg., ASCE, propagation neural networks.’’ Proc., Int. Conf. on Artificial Neural
8(2), 201–220. Networks, Vol. 1, Springer, New York, 643–646.
Maloney, W. F., and Federle, M. O. (1993). ‘‘Practical models for orga- Sinha, S. K., and McKim, R. A. (1997). ‘‘Forecasting construction cost
nizational assessment.’’ J. Mgmt. in Engrg., ASCE, 9(1), 64–81. escalation using artificial neural networks.’’ Proc., IEEE Artificial Neu-
Miller, D., and Friesen, H. P. (1984). Organizations: A quantum view. ral Network in Engrg. Conf., Vol. 1, 829–836.
Prentice-Hall, Englewood Cliffs, N.J. Skapura, D. M. (1996). Building neural networks. Addison-Wesley, New
Moselhi, O., Hegazy, T., and Fazio, P. (1991). ‘‘Neural networks as tools York.

14 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 2000

JOURNAL OF
COMPUTING IN CIVIL ENGINEERING
Volume 14 Number 2 April 2000

TECHNICAL PAPERS
Transient Dynamic Nonlinear Analysis Using MIMD Computer Architectures
J. Sziveri and B. H. V. Topping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Simplified Fuzzy ARTMAP as Pattern Recognizer
S. Rajasekaran and G. A. Vijayalakshmi Pai . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Java-Based Regulation Broker
Vei-Chung Liang and James H. Garrett Jr.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Reducing Prediction Error by Transforming Input Data for Neural Networks
Jonathan Jingsheng Shi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Stochastic Construction Time-Cost Trade-Off Analysis
Chung-Wei Feng, Liang Liu, and Scott A. Burns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Computer Graphics for Nonlinear Strut-Tie Model Approach
Young Mook Yun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Computer Imagery and Visualization in Civil Engineering Education
N. Bouchlaghem, W. Sher, and N. Beacham . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Bridge Damage Assessment through Fuzzy Petri Net Based Expert System
Weiling Chiang, Kevin F. R. Liu, and Jonathan Lee. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
 TRANSIENT DYNAMIC NONLINEAR ANALYSIS USING MIMD
COMPUTER ARCHITECTURES
By J. Sziveri1 and B. H. V. Topping2

ABSTRACT: This paper presents research on the use of homogeneous parallel and heterogeneous distributed
computers for finite-element analysis of transient dynamic problems using a message passing interface. The
appropriate computer architectures are discussed, and this review is the basis for the development of a new
definition of computational efficiency for heterogeneously distributed finite-element analysis. A code for the
transient nonlinear analysis of reinforced concrete plates is profiled. It is demonstrated that although the code
is efficient for homogeneous computing systems a new message passing procedure must be developed for
heterogeneously distributed systems. A new algorithm is developed and tested on a heterogeneous system of
workstations.

INTRODUCTION found in engineering offices. The use of these platforms rep-

resents a considerable change in performance and program-
The computational load of a transient dynamic nonlinear ming environment compared with the distributed memory
finite-element analysis of reinforced concrete structures may multiprocessor transputer system. For this reason the charac-
become very large for many problems. Parallel and distributed teristics of performance of the algorithm described by Sziveri
processing (Topping and Khan 1996), which permits the en- et al. (1996, 1997) is much altered.
gineer to undertake the finite-element analysis in a consider- The characteristics of the parallel finite-element program are
ably shorter time, is therefore of increasing significance. first investigated. Efficiency tests are given for the version of
A parallelization scheme using transputer systems for the the program that is a direct translation of the original Parallel
temporal analysis of reinforced concrete plates subjected to C code into MPI (Lämmer et al. 1997). Then, with the aid of
severe dynamic loads is discussed by Sziveri et al. (1996, profiling information, a new parallel scheme is devised and
1997). This scheme utilizes an explicit Newmark iteration introduced at source code level. This new parallel algorithm
scheme (Hughes and Belytschko 1994) as a temporal discret- exploits the functionalities of the full MPI implementation
ization of the finite-element formulation with six-noded tri- (MPICH n.d.) and provides more efficiency. Speedup tests and
angular facet elements that account for a complex nonlinear efficiency comparisons are made on the different hardware
material model (Beshara 1991) represented in a layered man- with different finite-element meshes.
ner (Hinton and Owen 1984). The example runs (Sziveri et al. Two different hardware platforms were selected to carry out
1996, 1997) show very high speedups and efficiency results, the parallel analyses of this finite-element problem described
which proves the importance of using parallel implementations in this paper. These platforms are the computer architectures
for efficient solution of the problem. (Flynn 1972) most likely to be found in engineering offices
These high efficiencies were achieved on transputer-based today. They are described in the next subsections.
parallel hardware. In research described by Sziveri et al. (1996,
1997) the parallel program was written using 3L Parallel C
(Parallel 1991), which is a C language environment developed Multiprocessor Workstations
particularly for the transputer-based systems by adding mes-
sage-passing functionalities to the standard C run-time func- Currently, many large workstations may have more than one
tion library. Lämmer et al. (1997) showed that with the intro- central processing unit (CPU), but when running sequential
duction of a message-passing interface (MPI) implementation programs each of these processors executes instruction streams
as a low-level C library the Parallel C program may be con- that are unrelated. Therefore, such systems still should be re-
verted to the MPI standard (‘‘MPI’’ 1995) compliant in its garded as (multiple) single instruction single data (SISD) ma-
source code. This gives the direct advantage that the parallel chines acting on different data spaces. Examples of this kind
program can be recompiled for virtually every current parallel of SISD machine are, for instance, most workstations includ-
hardware platform. ing those of DEC, Hewlett-Packard, and Sun Microsystems.
In the research described in this paper two other hardware But with the aid of thread libraries a running process can
platforms were selected to investigate further the parallel prop- spawn a thread or child process that with these machines can
erties of this dynamic finite-element problem. These platforms be executed on a different processor. Thus the machine exe-
are the homogeneous multiprocessor workstations (shared cutes several instruction streams in parallel on different data.
memory) and the heterogeneous networks of workstations (dis- The significance of the multiprocessor SISD machines men-
tributed memory). These were selected for consideration be- tioned above is that the instructions and data are related be-
cause they are the two classes that are the most likely to be cause they represent different parts of the same task to be
executed. Thus, these systems may run many subtasks in par-
1
allel in order to shorten the time-to-solution for the complete
Res. Assoc., Heriot-Watt Univ., Riccarton, Edinburgh, EH14 4AS,
U.K.
main task to be executed. Thus with the aid of such a tech-
2
Prof. of Computational Mech., Heriot-Watt Univ., Riccarton, Edin- nique the multiprocessor SISD machine becomes a multiple
burgh, EH14 4AS, U.K. instruction multiple data (MIMD) parallel computer where the
Note. Editor: Sivand Lakmazaheri. Discussion open until September processes on different CPUs share the same address space.
1, 2000. To extend the closing date one month, a written request must Hence it is the shared memory model of communication (Top-
be filed with the ASCE Manager of Journals. The manuscript for this ping 1999). Because of physical hardware constraints in a mul-
paper was submitted for review and possible publication on December 8,
1998. This paper is part of the Journal of Computing in Civil Engi- tiprocessor workstation the CPUs are always of the same type,
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- and so performancewise these shared memory MIMD envi-
0079–0091/$8.00 ⫹ $.50 per page. Paper No. 19797. ronments are homogeneous.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 79
Networks of Workstations allel code. With this conversion the parallel program can now
be compiled on virtually every parallel hardware that has an
The role of distributed computing is increasing as existing MPI library implementation.
personal computers and workstations are all networked, and
thus their integration provides engineers with more powerful
systems, without additional expense. This takes the distributed MPI Implementation on Selected Hardware Platforms
memory MIMD concept one step further: instead of many in-
tegrated processors in one or several boxes, workstations are
In the present research two types of parallel hardware are
connected and set to work concurrently on tasks in the same
used: (1) The homogeneous multiprocessor workstations; and
program. Thus with the aid of the network as communication
(2) the heterogeneous networks of workstations. Although
media ‘‘virtual’’ distributed memory MIMD computers may
many other MPI library implementations exist for these hard-
be built from ordinary computers. Conceptually, this is not
ware platforms, the most widely known MPICH implemen-
different from the purpose built distributed memory MIMD
tation (MPICH n.d.) is used here, which was developed at the
high performance or supercomputers, but the communication
Argonne National Laboratory. This implementation supports
between the processors is often several orders of magnitude
slower. many different types of hardware including special high per-
These networks may be local area networks or sometimes formance supercomputers to Unix workstations integrated us-
even wide area networks. They usually utilize the TCP/IP pro- ing their network connectivity to perform as a parallel ma-
tocol with various types of network carrier (e.g., thin ethernet chine.
coaxial cable, fiberoptics, or twisted pair cable). By realizing In the present research, the multiprocessor workstations’
the potential of these techniques, certain vendors (e.g., NEC, shared memory functionalities and thread library are used as
SGI, and DEC) offer optional high bandwidth direct intercon- an underlying communication layer by the MPICH implemen-
nect facilities for their workstations. With the support of pro- tation. This setup provides a homogeneous shared memory
prietary software libraries these interconnects tightly couple MIMD parallel machine as it was described previously in the
the workstations into clusters and provide the possibility of Multiprocessor Workstations section. The communication be-
low latency (direct to memory) communication between the tween the processors is fast, generally resulting in high speed-
nodes. ups and efficiencies, but the level of parallelization is limited
These distributed systems are usually heterogeneous in the by the actual number of processors in a given machine.
sense that all the connected processors will not usually be of For the networks of workstations that overcome this limi-
the same type or power. Some nodes of the network may them- tation the p4 parallel programming system (‘‘The p4’’ n.d.)
selves be MIMD or SIMD computers that represent high per- was used by the MPICH implementation as a communication
formance compute nodes within the network. It is important device. It utilizes Unix sockets, and the connection is realized
to ensure that any computational task is allocated to the ap- via the TCP/IP protocol over the network carrier. The down-
propriate processor within the heterogeneous network. The side of this type of connection is that the message latency can
problem of initial and dynamic load balancing with regard to be very high. For the system used in the present study, which
parallel and distributed finite-element computations will be is the typical configuration of TCP/IP over thin ethernet
discussed later in this paper. (10Mb/s), the latency is approximately 200 ␮s, which is two
orders of magnitude greater than for the shared memory com-
BASIC MPI COMMUNICATION munication device. The advantage of this hardware is that the
number of processors comprising the virtual parallel machine
A parallel algorithm for the explicit dynamic analysis of is scalable with the network, and the latency problem can be
reinforced concrete plates was presented by Sziveri et al. eased somewhat by employing different network technology.
(1996, 1997), and the program was tested on transputer-based This hardware and software realizes a virtual distributed mem-
parallel hardware. The communication scheme presented for ory MIMD parallel computer as previously described in the
this explicit transient finite-element problem was called the Networks of Workstations section. However, this MIMD par-
pair-multiple (PM) method. Lämmer et al. (1997) showed how allel computer is generally heterogeneous in the sense that the
the program implementation of the PM scheme (written in nodes connected by the network are usually of a different type
Parallel C) was converted into an MPI standard compliant par- and capacity.

FIG. 1. A2 Finite-Element Idealization of Reinforced Concrete Panel Divided into Two, Four, and Eight Subdomains

TABLE 1. Details of Size of A2 Mesh

Number Total number Number of intersubdomain Number Number of

of subdomains of nodes boundary nodes of elements elements per subdomain
(1) (2) (3) (4) (5)
2 449 17 208 104
4 449 33 208 52
8 449 65 208 26

80 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

Example Meshes Speedup and Efficiency Calculation with
A reinforced concrete panel subjected to blast loading was Heterogeneous Computing
modeled by Sziveri et al. (1996) and idealized using approx- When performancewise the parallel hardware is homoge-
imately 200 triangular finite elements. This original idealiza- neous, the value of speedup and the parallel efficiency may be
tion, which is denoted by A2, was decomposed into two, four, calculated using the following two simple formulas (Topping
and eight subdomains in such a way that the number of ele- and Khan 1996):
ments in each subdomain are equal, and the cuts are optimized.
The domain decompositions were carried out using the ESGM tseq
Speedup = S = (1)
procedure described by Sziveri et al. (1999). The A2 ideali- tpar
zation is shown in Fig. 1 with the three decompositions illus-
trated. Details of the A2 mesh size and other parameters are S
Efficiency = E = (2)
given in Table 1. np
The same model was further refined into a finite-element
idealization comprising approximately 800 triangular elements where tseq = sequential execution time for a problem; tpar =
of the same type. This mesh is denoted as the A8 mesh and parallel execution time for the same problem; and np = number
was decomposed into two, four, and eight subdomains with an of (identical) processors used for the parallel analysis. When
equal number of elements using the same ESGM partitioner assessing a heterogeneous computing run, these equations be-
(Sziveri et al. 1999). The two- and four-subdomain decom- come inappropriate. First, the definition of the speedup is con-
positions of the A8 mesh are shown in Fig. 2. The mesh de- fusing because in a heterogeneous environment the sequential
composed into eight subdomains is shown in Fig. 3 with the time tseq might be different for each participating processor. It
numbers of each subdomain illustrated. Details of the A8 mesh might be suggested that the attributes of the fastest processor
size and other parameters are given in Table 2. The numbering should be used as a yardstick because a user would run the
of these subdomains will be important when assessing the with program on that machine if only the sequential version was
the executions of the parallel program on the heterogeneous available. This not only results in low efficiency figures but
system. does not represent the utilization of the parallel hardware cor-
These partitions were determined to ensure load balancing rectly either. On the other hand, it is difficult to implement the
(i.e., equal numbers of elements per subdomain) and minimum notion of speedup when the processors are of different capac-
interprocessor communication (i.e., minimum subdomain in- ities, assuming that a heterogeneous environment consists of
terface nodes). two processors: a fast one and a very slow one, which is only
a tenth of the speed of the first one. Then, given ideal load
balancing and no communication requirements, the program
would only run 1.1 times faster regardless of the general ex-
pectation of 2.0 when two processors of equal speed were
used. For these reasons the calculation of parallel computa-
tional efficiency has to be revised to account for the hetero-
geneousness of the hardware. This is achieved by dropping the
notion of speedup and generalizing the efficiency calculation
based on (1) and (2) as follows.
The true meaning of parallel computational efficiency lies
with the utilization of the parallel computing hardware. When
FIG. 2. A8 Mesh Decomposed into Two and Four Subdomains investigating a heterogeneous system this may be broken down
to consider the utilization of its individual components (i.e.,
each processor). Hence for a given processor in a parallel (het-
erogeneous) environment this efficiency may be described as
the ratio of the time it should ideally spend on doing a certain
part or piece of the overall job to the actual time it took to do
it. For the ith processor in the system this is represented by
Pi
E ih = (3)
tpar
where tpar = overall parallel execution time as above, which is
the same for each processor; P i represents some kind of pro-
portional sequential load on the ith processor, also expressed
as a time quantity. Due to the nature of the parallel algorithm
(i.e., geometric parallelism) (Topping and Khan 1996), this
proportional load is very easy to determine. This time P i rep-
resents that proportion of the sequential running time that the
ith processor would ideally spend working on a part of the
FIG. 3. A8 Mesh Decomposed into Eight Subdomains total problem. For the ith processor this may be calculated as

TABLE 2. Details of Size of A8 Mesh

Number Total number Number of intersubdo- Number Number of

of subdomains of nodes main boundary nodes of elements elements per subdomain
(1) (2) (3) (4) (5)
2 1,753 41 840 422/418
4 1,753 91 840 211/211/209/209
8 1,753 181 840 105/106/106/105/104/105/105/104

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 81

 P i = f li ⭈ t seq
i
(4) 1
冘
tpar
f il
where f il = load factor describing the ratio of the job of the
冘
Ehw = (9)
1
ith processor to the whole problem; and t iseq = total sequential
time required to execute the whole problem on processor i. In t iseq
the case of this explicit parallel finite-element analysis the load By noting that
factor may initially be considered as the proportion of ele-
ments given to the processor (i.e., the size of its subdomain
compared to the whole structure)
冘 f il = 1 (10)
the overall weighted efficiency of the whole heterogeneous
NE i system may be represented as
f il = (5)
NE 1

冘
Ehw = (11)
1
where NE = total number of finite elements in the whole struc- tpar i
t
ture; and NE i = number of elements in the ith subdomain that seq

is processed on the ith processor. On a homogeneous system t iseq is the same for each process,
In the case of a homogeneous system of np processors, t iseq and (11) becomes
is the same for each processor and with the assumption of

冘
np
perfect load balancing when f il = 1/np, (3) becomes equivalent
to (2). But E ih in (3) represents only the efficiency (i.e., the tseq tseq
utilization, of one single processor within the heterogeneous Ehw = = (12)
np tpar np tpar
system). To devise an overall measure of the efficiency for the
whole network, the average of the individual efficiency values which is the same as (2).
for each processor participating in the heterogeneous system
may be calculated as Timing on Unix Platforms

冘 冘
np np
1 1 f il ⭈ t iseq
The timing of program executions on Unix platforms is a
Eh 㛭 ave = E ih = (6) topic of constant debate in the computational literature.
np i =1 np i =1 tpar Whether to measure ‘‘pure’’ CPU time or if wall clock time
is sufficient is an important question. When measuring wall
The only difficulty with this approach is that it does not rep- clock time the main problem is how to deal with ‘‘distur-
resent the overall performance of the whole network well, par- bances’’ created by other processes consuming CPU time
ticularly when the processors have very different capacities. In slices and taking up system resources. This can be caused by
this latter case the most important factor when designing and standard or additional system activities including a wide range
deploying heterogeneous systems of very different compo- of processes (e.g., logging, printing, mail, etc.), daemons, or
nents is that the strongest processors should be utilized the other user application programs. Some researchers prefer to
most; they should run at the highest possible efficiency. The measure execution times when the Unix system excludes these
slow ones just ‘‘assist’’; their full utilization is of little or no disturbances by booting the machine into a different ‘‘init
concern as they do not really affect the overall performance. state’’ where they can restrict general system activities and
This principle may be represented in the efficiency calculation totally exclude network and user processes.
of the whole system by employing a weighting factor to the However in this study these concepts are not followed. The
contributions of the individual components. This weighting Unix workstations are considered ‘‘as is’’ in a state as they
factor would be directly proportional to the power of each are working in a general engineering office. Execution times
processor. Thus instead of (6), the overall efficiency of the are always measured as wall clock times in the concept of
whole heterogeneous system would be written as follows: time to solution values and the machines are in a ‘‘normal

冘冉 冊
state.’’ Thus users can log in and out, other processes can be

冘
f il ⭈ t iseq launched, and system daemons activated at any time. The only
i i pp i
( pp E ) tpar

冘 冘
h provision that was made to achieve some consistency among
Ehw = = (7) the timing result was that the test runs were made at times
pp i pp i when the machines were relatively silent in terms of user ac-
tivity and that there was no great system allocation either. Sev-
where pp i = appropriate measure of the power or capacity of eral measurements were taken in such conditions, and the ob-
the ith processor. viously extreme ones were omitted. The average of the
By concentrating on the finite-element problem to be solved remainder of the values gives the appropriate ‘‘time to solu-
here, the power of each processor can be most easily assessed tion.’’
by looking at the sequential execution times it takes to run the What happens when these conditions are disturbed by some
problem (or a representative problem) on each processor (i.e., arbitrary processes is discussed in the last section on load bal-
the t iseq values). Hence the power or capacity of the processor ancing.
to solve the problem is inversely proportional to these se-
quential time values, and a good implementation for the pp i
Timing Results of PM Parallel Scheme
measures would be
First of all the sequential version of the program was run
i 1 on all the machines that are involved in the present study.
pp = (8)
t iseq These machines are listed in Table 3. They were all running
SUN’s Solaris operating system version 2.5.1, which is a Sys-
The relative power of the workstations need only be assessed tem V Unix. The PCs had the Solaris x86 version installed of
once, when the heterogeneous system is installed. By substi- the same release. The network communication was 10Mb/s
tuting (8) into (7) and reducing the expression with t iseq, the ethernet cable.
overall efficiency formula may be rewritten as follows: The numerical analysis was carried out for a small part of
82 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
TABLE 3. Summary of Unix Workstations Used in This Study half of the sequential time. This would be the ideal case for
Host name Machine type CPU all parallel runs.
(1) (2) (3) These excellent results are due to the characteristics of the
hardware and its good utilization by the software environment.
carme SUN Sparc 10 4 ⫻ TI TMS390Z50 at 50 MHz
(superSparc chip)
The shared memory employed here is a very fast communi-
thebe SUN Sparc 20 1 ⫻ TI TMS390Z50 at 50 MHz cation device; hence it does not impose any serious penalty or
(superSparc chip) overhead on the parallel computation.
iapetus SUN Sparc 4 1 ⫻ Fujitsu MB86904 at 110 MHz
(microSparc II chip)
Running on Distributed Memory Workstation Cluster
elara ALR PC 2 ⫻ Intel Pentium at 90 MHz
(i586 chip)
Using the general-type network connections, a theoretically
rhea ABIT PC 1 ⫻ AMD K5 at 100 MHz
(i586 compatible chip) arbitrary number of processors may be connected to form one
virtual machine. Some of these workstations may themselves
be multiprocessor machines, and thus they can accommodate
TABLE 4. Sequential Execution Times for A2 and A8 Meshes several subdomains. In this case the communication may be
tseq
faster between these subdomains if the software environment
takes advantage of the fast shared memory access. This is not
Host for A2 mesh for A8 mesh the case in this study. When linking the parallel MPI program
name (s) (s) with that version of the MPI library that utilizes Unix sockets
(1) (2) (3) (via the p4 subsystem) as a communication device, every pro-
carme 11.0 49.0 cess instance of the program works according to this version
thebe 9.4 42.0 regardless of whether they are placed on the same machine or
iapetus 12.1 54.0
not.
elara 11.4 51.0
rhea 8.7 39.0 Table 6 shows how the subdomains were mapped onto the
eight processes and which one of these processes was launched
on which machine. The machine names refer back to Table 3
the simulation process using the original sequential version of for the full description of the hardware.
the program, and the execution time tseq was measured. Table The eight subdomain decompositions for the A2 and A8
4 shows these execution times achieved by running the pro- meshes, shown in Figs. 1 and 3, respectively, were used in the
gram for 25 iterations as a benchmark for the A2 mesh and eight processor heterogeneous runs. Fig. 3 shows the gray
the same for mesh A8. scales to identify each subdomain. Again, the same blocks of
These timing results do not include the time required for 25 iterations were measured, and the parallel execution times
the preprocessing of the finite-element mesh (i.e., the domain and the calculated efficiency values are detailed in Table 7.
decomposition ⬇2–3 s), the preparation of subdomain con-
nectivity data for the communication routine (⬇0.1 s), the de- TABLE 5. Parallel Execution Times Measured on Multiproces-
livery of the subdomains to the actual processing nodes sor Workstation Utilizing Shared Memory for Communication
(⬇0.2–0.3 s), nor the setting up time before the iteration com- for A2 and A8 Meshes Using PM Algorithm
mences (⬇0.1 s). However these times are negligible com- Number of tpar E
pared with the long time marching iteration of the transient Mesh subdomains np (s) S (%)
dynamic analysis. (1) (2) (3) (4) (5) (6)
A2 2 2 5.7 1.93 96.5
Running on Shared Memory MP Workstation 4 4 3.3 3.33 83.3
A8 2 2 ⬇24.5 ⬇2 ⬇100
First the MPI program was linked with that version of the 4 4 12.6 3.89 97.2
MPI library that utilizes the shared memory as a communi-
cation device. The program was run on the multiprocessor
workstation ‘‘carme’’ by launching two and four instances of TABLE 6. Placement of Processes (Corresponding to Subdo-
the program to operate on two and four subdomains, respec- mains) onto Different Machines When Using Heterogeneous
tively, of both the A2 and A8 meshes. The running times were Network of Workstations for Eight Subdomains
measured for the same blocks of 25 iterations as quoted in Subdomain number Host name (from Table 3)
Table 4 at the initial stage of the analysis when the material (1) (2)
is mainly in a linear state. The parallel execution times and
1 thebe
the calculated speedup and efficiency values are summarized 2 carme
in Table 5. 3 carme
The terms of np, tpar, S, and E in Table 5 are as defined by 4 carme
(1) and (2). In Table 5 the general trends of parallel processing 5 elara
may be clearly seen as follows: 6 elara
7 rhea
8 iapetus
• When using more processors to work on the same prob-
lem, the efficiency drops due to the increased communi-
cation requirements. TABLE 7. Parallel Execution Times Measured on Network of
• When analyzing larger problems higher computational ef- Workstations Utilizing Unix Sockets for Communication for A2
ficiencies may be achieved as the proportion of commu- and A8 Meshes Using PM Parallelization Scheme
nication shrinks compared to the total computational load. Number of tpar Ehw
Mesh subdomains np (s) S (%)
It is worth noticing that these efficiency values are quite high (1) (2) (3) (4) (5) (6)
even for the smaller problem of the A2 mesh. In the case of
A2 8 8 2.9 3.0–4.2 45.8
the A8 mesh analyzed on two processors the speedup is vir- A8 8 8 8.2 4.8–6.6 72.4
tually 2 as the parallel execution time was almost identical to
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 83
 The terms np, tpar, and S in Table 7 are as defined by (1) ues show that the processors are not utilized properly. This is
and (2). As speedup is not clearly defined for heterogeneous the result of changing the underlying hardware layer of com-
parallel runs, two bounding values are calculated based on the munication from the fast shared memory to the distributed
fastest and the slowest sequential runs measured on the cor- network connection. To study the effect of this in the pro-
responding fastest and slowest processors in the heterogeneous gram’s run-time profile, several time measurement instructions
network. The weighted efficiency of the heterogeneous net- were embedded into the source code. With the aid of these
work Ehw is determined as defined by (11). measurement points, the work of each processor was traced in
It may be noted by comparing Tables 5 and 7 that even detail and the timing results combined into charts.
though the number of processors used in the latter case was Sziveri et al. (1996, 1997) gave full details of this dynamic
increased from four to eight, it did not result in significant explicit analysis algorithm including the PM parallelization
shortening of the parallel execution times. The efficiency val- scheme. This scheme enables interprocessor communication

FIG. 4. Subtasks of PM Parallelization Scheme

FIG. 5. Percentage of Time Spent on Each Subtask of PM Parallelization Scheme by Every Processor on A2 Mesh

84 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 FIG. 6. Percentage of Time Spent on Each Subtask of PM Parallelization Scheme by Every Processor on A8 Mesh

between pairs (P) of subdomains and multiple (M) subdo- while the communication proceeds. Given the relatively high
mains. In Fig. 4 the major subtasks involved with this algo- communication bandwidth in both the transputer-based hard-
rithm are plotted, and the timing instructions clearly define the ware and the shared memory multiprocessor workstations
borders of these subtasks that constitute an iteration step. compared to their processing capabilities, the presented PM
In Fig. 4 a notation is introduced to reference each subtask communication scheme implemented with this kind of point-
that make up an iteration step. This notation is used in Figs. to-point-type blocking instructions was sufficient to achieve
5 and 6, which are detailed profiling charts of the PM parallel very high computational efficiencies.
algorithm with respect to meshes A2 and A8, respectively. Unfortunately, porting the algorithm to the distributed mem-
These two figures correspond to the parallel runs of Table 7. ory workstation cluster highlighted the weaknesses of this type
Figs. 5 and 6 show the amount of time spent for each sub- of parallel hardware system. The relative communication
task by each processor within one iteration. It is expressed in bandwidth is much less over the standard network facilities.
terms of percentage where 100% is the total time spent on one Moreover, the communication throughput is severely penal-
complete iteration step. ized by very high latency values.
This representation assumes that there is a synchronization These conditions imposed by the hardware environment re-
at the end of each iteration step by the energy communication. duced the efficiency of the parallel application. Clearly a new
This assumption is correct as the length of that communication parallel strategy is required to overcome these limitations.
is very short, and the collective nature of it forms a barrier
within the flow of the program for every processor. New Parallel Algorithm
It is apparent from Figs. 5 and 6 that a very large proportion
of the time is spent on communication (subtasks PM㛭st㛭2, The most obvious strategy when attempting to overcome
PM㛭st㛭3, and PM㛭st㛭4) in each iteration. It is as high as latency is to minimize the actual number of messages to be
40–50% in terms of the smaller A2 mesh and still 20–35% sent by packing as much information into each message buffer
with the bigger A8 mesh. This large communication overhead as possible. The PM communication scheme was developed to
results in the low efficiency values and thus prevents the avoid communication deadlock in the transputer router system
speedup to grow with the number of processors. The next sec- as a first priority. The number of messages to be sent were not
tion details the solution to this problem by introducing a new strictly minimized. When working with shared memory mul-
parallel communication scheme. tiple processor machines or on workstation clusters, message
routing is not a problem. Arbitrary connections are possible
NEW PARALLEL COMMUNICATION SCHEME between any two nodes within the parallel system, and the
MPI software library utilizes these connections. This gives an
The PM communication scheme was first presented by Szi- opportunity to rethink the message layout of the PM scheme.
veri et al. (1996, 1997) and was implemented in Parallel C for The message structure is the simplest between the subdo-
transputer-based hardware. Lämmer et al. (1997) showed how mains if every subdomain sends only one single message to
the Parallel C calls were substituted with standard MPI func- each subdomain to which it is connected. This is best illus-
tion calls. In both implementations these message passing trated in Fig. 7 where the notion of P and M nodes is replaced
functions were point-to-point-type blocking instructions, by groups of intersubdomain nodes that represent the actual
meaning that the current thread of execution could not con- messages to be sent.
tinue until the called communication function had been re- The MPICH library used in this study is a full implemen-
turned. Hence the computational part of the program stops tation of the MPI specification, which contains nonblocking
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 85
message passing functions as well as blocking ones. This al-
lows the introduction of a special strategy that results in very
significant efficiency improvements of the parallel system on
many parallel hardware platforms. When calling nonblocking
communication functions, the main thread of execution is re-
turned to the caller program before the communication, a send,
or more importantly a receive is completed. It is usually im-
plemented by launching background threads that operate on
reserved buffer spaces in an asynchronous way; that is, these
threads wait for the delays of the communication (caused by
latency and bandwidth characteristics of the hardware)
whereas the main program can continue its execution.
This is where the potential for these kinds of techniques
lies. Because the communication overhead (i.e., the time spent
on communication in a parallel environment) can be ‘‘cov-
ered’’ with useful sequential calculations, it is often referred
to as the ‘‘red-black’’ method in the computational literature
and will be denoted by the abbreviation of RB.
When implementing an RB scheme for the explicit finite-
element problem in question, the first feature to look at is the
profiling figures of the parallel runs. It is apparent from Figs.
5 and 6 that although a significant percentage of time is spent
on communication, the longest subtask within one iteration is
the one denoted by PM㛭st㛭1 and described in Fig. 4. With FIG. 8. Grouping of Elements and Intersubdomain Nodes for
an even more detailed timing it has been determined that ap- New (RB) Parallel Scheme
proximately 99% of the time spent on PM㛭st㛭1 is actually
spent on the calculation of the vector of internal forces N. In sisting of internal subdomain elements, will form the second
the source code of the program this calculation is a loop for group. This concept is illustrated in Fig. 8.
each element contained in the subdomain. This includes the
processing of very complex (nonlinear) material rules as ap- Implementation of RB Parallel Scheme
plied to each integration point of each finite element.
This calculation has to be completed before the exchange The main steps in the implementation of the RB scheme is
of contributions of N can take place between the neighboring described in Fig. 9. In this figure the major subtasks compris-
subdomains for each intersubdomain boundary node. However ing an iteration step of the RB algorithm are plotted and the
it is clear that it is sufficient to calculate the local contributions place of timing instructions in between these subtasks are
to N from those elements that are connected to these intersub- marked. Fig. 9 also introduces a notation to reference these
domain boundary nodes. At this moment the communication subtasks. This notation will be used subsequently in all pro-
may commence, and the integration of the rest of the elements filing figures of the RB parallel communication scheme.
may be continued independently. As may be seen in Fig. 9 after the integration of elements
Thus the overlapping of the communication and some cal- in the first group, which are connected to intersubdomain
culation can be realized in the explicit finite-element problem boundary nodes (within subtask RB㛭st㛭1), all communication
by dividing the elements within each subdomain into two is launched to sum the local contributions of the subdomains
groups. Those elements will be placed into the first group, to the vector N. This means that there are two types of oper-
which are connected to intersubdomain boundary nodes, and ations for each group of intersubdomain nodes (i.e., for each
the rest of the elements, the ‘‘core’’ of the subdomain con- subdomain relation shown in Fig. 8):

• The subdomain has to send its own contribution to the

neighboring subdomain.
• The subdomain has to receive the values from the other
subdomain as well.

These operations are implemented by MPI㛭Isend( ) and

MPI㛭Irecv( ) function calls, respectively, in which the ‘‘I’’
stands for ‘‘incomplete.’’ These are nonblocking functions;
thus they initiate the communication and return immediately,
so that the program may continue its work on the core of the
subdomain.
To avoid memory access conflicts between the main pro-
gram (which keeps putting values into N) and the communi-
cation threads for the values that are to be sent and received,
independent memory buffers are allocated separately. The send
routine makes a copy of the local contributions to the vector
N corresponding to the intersubdomain nodes in the send buf-
fer space before the call to MPI㛭Isend( ). Thus the back-
ground thread launched by MPI㛭Isend( ) only operates on
this send buffer memory area. Similarly, the background thread
launched by MPI㛭Irecv( ) retains the received values in an-
FIG. 7. Grouping of Intersubdomain Nodes to Form Messages other receive buffer, which is empty before receipt.
86 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 FIG. 9. Subtasks of RB Parallelization Scheme

To maximize program efficiency all of these communication TABLE 8. Parallel Execution Times Measured on Network of
buffers are dynamically preallocated in advance at program Workstations Utilizing Unix Sockets for Communication for A2
and A8 Meshes Using RB Algorithm
start-up by each instance of the program, and they correspond
to the neighboring subdomains of that instance. Number of tpar Ehw
When the main program is finished with the integration of Mesh subdomains np (s) S (%)
the elements within the core of the subdomain (i.e., subtask (1) (2) (3) (4) (5) (6)
RB㛭st㛭3 is complete), then the execution cannot continue any A2 8 8 2.1 4.1–5.8 63.3
further if the communication has not, as yet, been completed. A8 8 8 6.8 5.7–7.9 87.3
Thus synchronization takes place here, and it is implemented
by an MPI㛭Waitall( ) function call for the sends and a series
of MPI㛭Waitany( ) function calls for the receives. sults presented in Table 8. The notation of the subtasks of the
This function call is different because, for the threads RB communication scheme is as described in Fig. 9.
launched by the send routines, it is simply a check for com- It is apparent in Table 8 and Figs. 10 and 11 that the RB
pletion, and no other work has to be done. However for the communication scheme presented above is much more effi-
threads launched by the receive routines, further action is re- cient than the PM scheme. However, a significant proportion
quired. The MPI㛭Waitany( ) function call returns when any of the execution time is still spent on communication on cer-
of the receives are completed, and the content of the corre- tain processors. If that could be reduced, the program might
sponding receive buffer has to be added to the local system run almost as ideally parallel as for the shared memory plat-
form. An investigation to consider this ideal load balancing is
vector N of the subdomain. In such a way, after all receives
presented in the next section.
are returned for each subdomain relation, the vector of internal
forces N is synchronized between the subdomains. LOAD BALANCING
Timing and Profiling Result of RB Scheme With the introduction of the RB parallel communication
scheme in the previous section, the communication overhead
The program was recompiled and relinked with the MPI (i.e., the time spent on communication in the parallel system)
library as described earlier in the sections entitled Running on has been reduced considerably. However by examining the
Shared Memory MP Workstation and Running on Distributed parallel run-time profiling information in Figs. 10 and 11, it
Memory Workstation Cluster. Running the same examples on is apparent that some processors can still spend as much as
the shared memory MP workstation gives virtually the same 30% of their time on communication, while other nodes in the
results as described in Table 5. But these parallel runs were network spend much less.
almost ‘‘ideal’’; thus no further improvements were expected. As these processors are the fast ones (‘‘thebe’’ and ‘‘rhea,’’
However, running the same examples on the distributed which are P1 and P7 in Figs. 10 and 11) it indicates that this
memory workstation cluster with the same heterogeneous con- is not real work; rather these processors are waiting for the
figuration of workstations as described in Tables 3 and 6, the others to finish. The synchronization of their different speeds
timing results change significantly. In comparison with Table appear as extra communication load on the parallel system.
7 the new timing results for the RB parallel communication Clearly, this problem may be addressed by distributing the
scheme are presented in Table 8. All of the variables in this load onto the processors so that it is proportional to their pro-
table are as described for Table 7. cessing capabilities.
The parallel program runs were examined in detail, and the
profiling information is shown in Figs. 10 and 11. These fig- Energy Communication
ures were generated in a similar way as Figs. 5 and 6. It appears in Figs. 10 and 11 that a significant proportion
The timing ratios in Figs. 10 and 11 correspond to the re- of time is spent on the energy communication on processor
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 87
 FIG. 10. Percentage of Time Spent on Each Subtask of RB Parallelization Scheme by Every Processor on A2 Mesh

FIG. 11. Percentage of Time Spent on Each Subtask of RB Parallelization Scheme by Every Processor on A8 Mesh

P1. This is a collective type of communication as described sults hardly changed from those presented in Table 8. The
by Sziveri et al. (1996, 1997), and for such a course grained characteristics of the profiling shown in Fig. 12 had not
parallel system it should not take that long. To assess the effect changed from the original given in Fig. 11. This was expected
of this communication the source code of the program was as the computational load represented by the energy commu-
modified in such a way that the energy communication and nication, and the energy balance check is negligible compared
the energy balance check was cut out temporarily. The pro- to the load arising from the rest of the iteration. However,
gram was recompiled and run again with the A8 example mesh looking at the profiling data of processor P1 (in Fig. 12) the
on the distributed memory workstation cluster. The timing re- amount of time that was spent on the energy communication
88 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
FIG. 12. Percentage of Time Spent on Each Subtask of RB Parallelization Scheme by Every Processor on A8 Mesh When Omitting
Energy Communication

(Fig. 11) is now spent on waiting for the intersubdomain data As can be seen in Fig. 3, subdomain No. 4 is the one with
exchange to be completed. This clearly indicates that the as- the largest number of elements on the intersubdomain bound-
sumption above is correct (i.e., the difference of the processor ary; consequently the time bar corresponding to subtask RB㛭
speeds appears as an extra communication load). This is noth- st㛭1 is the longest in Fig. 11. Similarly, subdomain No. 3 is
ing else but the faster processor waiting for the slower to fin- the one with the largest number of elements in the core of the
ish. Consequently the load has to be split between the proces- subdomain; consequently the time bar corresponding to sub-
sors in such a way that they can complete their tasks at the task RB㛭st㛭1 is the shortest while the time bar corresponding
same time. to subtask RB㛭st㛭3 is the longest in Fig. 11.
Similar conclusions may be drawn by looking at other pro-
Load Balancing Decomposition cessors and the figure corresponding to the A2 idealization.
When taking the timing measurements shown in Figs. 10
and 11, the material was mainly in a linear state; consequently Adjustment of Decomposition
the computational load on each processor is approximately The computational power of a processor may be estimated
proportional to the number of elements within the subdomain on the basis of sequential run times for a given idealization
allocated to the processor. Thus the decomposition of the fi- (as was suggested earlier in the Speedup and Efficiency Cal-
nite-element mesh may be adjusted on the basis of the com- culation with Heterogeneous Computing section). Table 9
putational power of each processor participating in the het- shows the calculation of the number of elements to be placed
erogeneous network. onto each processor in the heterogeneous network when work-
ing with the A8 mesh. The calculation of the elements required
Intersubdomain Boundary versus Core Elements to load balance the system is based on the sequential execution
Other features shown in Figs. 10 and 11, which confirm times for the same A8 mesh as presented in Table 4.
their correctness and the conclusions above, may be deduced
by looking at the columns corresponding to processors P2, P3, TABLE 9. Redistribution of Elements of A8 Mesh When Using
Heterogeneous Network of Workstations for Eight Subdomains
and P4 in Fig. 11. These three columns correspond to processs
running on the machine carme, which obviously has identical tseq Number of
processors. The load is totally balanced as the number of el- Subdomain Host name (from Table 4) elements for
ements are equal in each subdomain. These conditions result number (from Table 3) (s) processor
in equal execution times for the sequential part of the com- (1) (2) (3) (4)
putation (i.e., RB㛭st㛭1 and RB㛭st㛭3). 1 thebe 42.0 118
This is clearly the case as the sum of the total combined 2 carme 49.0 101
time spent on subtasks RB㛭st㛭1 and RB㛭st㛭3 on each pro- 3 carme 49.0 102
cessor is almost equal. However, it is interesting to notice that 4 carme 49.0 101
the ratio between subtasks RB㛭st㛭1 and RB㛭st㛭3 changes 5 elara 51.0 98
6 elara 51.0 98
according to the ratio of the number of elements in the first 7 rhea 39.0 129
group (which are connected to the intersubdomain boundary 8 iapetus 54.0 93
nodes) to the second group (the core of the subdomain) within Total number of elements in mesh A8 840
each subdomain.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 89
 run is shown in Fig. 14. It is apparent from Table 10 and Fig.
14 that when some load balancing is introduced into the het-
erogeneous parallel system the execution became very effi-
cient. Working on the given example, the processors spent
<5% of their time on communication, which enables the good
speedup. It also means that the combined computational re-
sources of this parallel system are very well utilized, and this
utilization is expressed by the high efficiency value shown in
Table 10.

Disturbances to Computational Load Balance

Unfortunately a computationally well-balanced parallel sys-
tem, such as the one represented in Fig. 14, may not be able
to sustain its load balance.
For instance, running the same program on the same ide-
alization as that of Fig. 14 with the increased presence of non-
FIG. 13. A8 Mesh Decomposed into Eight Subdomains after
Adjustment linear effects [as shown by Sziveri et al. (1996, 1997)] the
performance decreases again, and the picture becomes similar
to the one shown in Fig. 11. Obviously, the reasons of this
The original subdomain decomposition, shown in Fig. 3,
symptom have to be reviewed to address its correction.
was adjusted to represent the values listed in Table 9. This
Generally, this performance decrease may be due to changes
new partitioning is displayed in Fig. 13.
internal to the parallel program and its execution, such as with
strong nonlinear effects or contact-impact problems where the
Timing Results
program code, which tries to simulate them, involves more
The complete program was run again on the same hetero- and more computations on certain regions of the domain. The
geneous configuration of workstations with this new, reparti- location of these regions in the idealization may not be bal-
tioned A8 mesh shown in Fig. 13. The timing results are pre- anced in terms of their mapping onto the computing nodes of
sented in Table 10. The profiling of this load balanced parallel the parallel system.
There may also be several other reasons external to the pro-
TABLE 10. Parallel Execution Time Measured on Network of gram that may affect the parallel performance. These reasons
Workstations for Adjusted Decomposition of A8 Mesh Using RB may include hardware failure or other system degradation on
Algorithm certain nodes (e.g., due to the activities of other processes or
Number of tpar Ehw users in a multitasking, multiuser environment) or communi-
Mesh subdomains np (s) S (%) cation network congestion and are generally much more typ-
(1) (2) (3) (4) (5) (6) ical of distributed network based hardware.
A8 8 8 6.2 6.3–8.7 95.7
The common feature of all of these phenomena is that they
manifest themselves as being computationally in-balance on

FIG. 14. Percentage of Time Spent on Each Subtask of RB Parallelization Scheme by Every Processor on A8 Mesh after Rebalance
of Subdomains

90 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

the parallel system causing a significant decrease in perfor- described was also supported by the Systems Architecture Committee of
mance. This also gives the opportunity to treat them similarly the U.K. Engineering and Physical Sciences Research Council through
the research contract ‘‘Domain Decomposition Methods for Parallel Finite
and compensate for their adverse effect. It is found that the Element Analysis’’ (GR/J51634).
measure of parallel computational efficiency is an extremely
good indicator when trying to overcome computational in-bal- APPENDIX. REFERENCES
ance.
Beshara, F. B. A. (1991). ‘‘Non-linear finite element analysis of reinforced
concrete structures subjected to blast loading,’’ PhD thesis, City Uni-
CONCLUSIONS
versity, London.
A program for the dynamic explicit transient analysis of Flynn, M. J. (1972). ‘‘Some computer organizations and their effective-
ness.’’ IEEE Trans. on Computing, 21, 948–960.
reinforced concrete plates was implemented in different par-
Hinton, E., and Owen, D. R. J. (1984). Finite element software for plates
allel environments, on a shared memory and a distributed and shells. Pineridge, Swansea, U.K.
memory MIMD architecture. Hughes, T. J. R., and Belytschko, T. (1994). ‘‘Nonlinear Finite Element
A parallel communication system that was developed for Analysis.’’ Course Notes, Paris.
transputer-based (blocking) hardware was revised and replaced Jimack, P. K. (1999). ‘‘An overview of dynamic load-balancing for par-
by a new approach. With this new communication strategy the allel adaptive computational mechanics codes.’’ Parallel and distrib-
uted processing for computational mechanics: Systems and tools, B.
parallel explicit finite-element analysis was made very efficient
H. V. Topping, ed., Saxe-Coburg Publications, Edinburgh, U.K., 350–
even on the distributed memory heterogeneous workstation 369.
cluster. In the implementation of this new scheme several fea- Lämmer, L., Sziveri, J., and Topping, B. H. V. (1997). ‘‘MPI for the
tures of a full MPI implementation were used. transtech parastation.’’ Advances in computational mechanics with par-
As demonstrated by the example runs above even such a allel and distributed processing, Civil-Comp Press, Edinburgh, U.K.,
coarse-grained parallel system, running the appropriate parallel 219–232.
‘‘MPI: A message passing interface standard.’’ (1995). Version 1.1, MPI
algorithm, can be very sensitive to computational in-balance.
Forum, 具http://www.mcs.anl.gov/mpi/mpi-report-1.1/mpi-report.html典.
A relatively small effort toward restoring this computational ‘‘MPICH: A portable implementation of MPI.’’ (n.d.). Argonne National
balance can make a big difference as was demonstrated. This Laboratory, Mathematics and Computer Science Division, 具http://
can be enough to maximize efficiency, thus making full use www.mcs.anl.gov/mpi/mpich典.
of the available computing resource. Parallel C user guide (compiler version 2.2.4), reference manual. (1991).
Consequently, one of the main areas of future work is the 3L Ltd., Livingston, U.K.
‘‘The p4 parallel programming system.’’ (n.d.) Argonne National Labo-
development of automated tools for dynamic load balancing
ratory, Mathematics and Computer Science Division, 具http://www.mcs.
and the integration of these tools into the parallel programs. anl.gov/home/lusk/p4/index.html典.
There are several theoretical algorithms for dynamic load Sziveri, J., Seale, C., and Topping, B. H. V. (1999). ‘‘Enhanced parallel
balancing of explicit finite-element problems in the literature subdomain generation.’’ Int. J. Numer. Methods in Engrg., accepted for
(Jimack 1999). These include algorithms for (re)distributing publication.
the computational load within the parallel system. Unfortu- Sziveri, J., Topping, B. H. V., and Iványi, P. (1996). ‘‘Parallel transient
non-linear analysis of reinforced concrete plates.’’ Advances in com-
nately, very little information is still available on how to de-
putational structures technology, Civil-Comp Press, Edinburgh, U.K.,
termine the timing of the load balancing and how frequently 383–395.
it is viable to apply these algorithms during a long parallel Sziveri, J., Topping, B. H. V., and Iványi, P. (1997). ‘‘Parallel dynamic
run. non-linear analysis of reinforced concrete plates.’’ Proc., NAFEMS
World Congr. ’97 on Des., Simulation and Optimisation, Reliability and
ACKNOWLEDGMENTS Applicability of Computational Methods, Vol. 2, University of Stuttgart,
Stuttgart, Germany, 806–817.
The research described in this paper was supported by Marine Tech- Topping, B. H. V., ed. (1999). Parallel and distributed processing for
nology Directorate Ltd research contracts ‘‘High Performance Computing computational mechanics: Systems and tools. Saxe-Coburg Publica-
for Marine Technology Research’’ (GR/J22191) and ‘‘High Performance tions, Edinburgh, U.K.
Adaptive Finite Element Computations for CAD of Offshore Structures Topping, B. H. V., and Khan, A. I. (1996). Parallel finite element com-
using Parallel and Heterogeneous Systems’’ (GR/J54017). The research putations. Saxe-Coburg Publications, Edinburgh, U.K.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 91

 SIMPLIFIED FUZZY ARTMAP AS PATTERN RECOGNIZER
By S. Rajasekaran,1 Member, ASCE, and G. A. Vijayalakshmi Pai2

ABSTRACT: Pattern recognition has turned out to be an important aspect of a dominant technology such as
machine intelligence. Domain specific fuzzy-neuro models particularly for the ‘‘black box’’ implementation of
PR applications have been recently investigated. In this paper, Kasuba’s simplified fuzzy adaptive resonance
theory map (ARTMAP) has been discussed as a pattern recognizer/classifier for image processing problems.
The model inherently recognizes only noise free patterns and in the case of patterns with noise or perturbations
(rotation/scaling/translation) misclassifies the images. To tackle this problem, a conventional moment based
rotation/scaling/translation invariant feature extractor has been employed. However, since the conventional fea-
ture extractor is not strictly invariant to most perturbations, certain mathematical modifications have been pro-
posed that have resulted in an excellent performance by the pattern recognizer. The potential of the model has
been demonstrated on two problems, namely, prediction of load from the yield patterns of elastoplastic analysis
of clamped and simply supported plates and prediction of modes from mode shapes.

INTRODUCTION Therefore, the aim of this paper is to build a pattern rec-

ognizer based on a fuzzy-neuro model and exhibit the follow-
Pattern recognition (PR) is a science that concerns the de- ing characteristics:
scription or classification (recognition) of measurements and
is often an important component of intelligent systems. Thus,
PR has turned out to be a useful and rapidly developing tech- 1. A less complex but powerful fuzzy-neuro architecture
nology with cross-disciplinary interest and contribution. 2. Significant reduction in training
PR techniques overlap with areas such as structural mod- 3. Recognition of both monochrome and color images
eling, optimization and estimation theory, and neuro-fuzzy sys- 4. Tolerance to pattern perturbations (e.g., rotation, trans-
tems. PR applications include seismic analysis, radar signal lation, and scaling)
analysis, computer vision, face recognition, speech recogni- 5. Tolerance to random noise
tion, electrocardiogram signal analysis, and handwritten and
character recognition. In the area of civil engineering, diag-
nostic analysis of crack patterns (Chao and Cheng 1998), Kasuba’s simplified fuzzy ARTMAP (SFAM) (1993), which
structural damage assessment (Ishizuka et al. 1982; Watada is a vast simplification of Carpenter et al.’s fuzzy ARTMAP
1984), and examination of soil cracking patterns are some of (1992), was chosen to function as the core architecture. When
the problems on which PR techniques have been applied. compared to its predecessor, SFAM has reduced computational
Although neural networks (NN) and fuzzy logic have in- overhead and architectural redundancy. Also, the model em-
dividually presented themselves as potential candidates of sev- ploys simple learning equations with a single user selectable
eral approaches well suited for pattern classification applica- parameter and can learn every single training pattern with only
tions, hybrid systems that combine NN and fuzzy logic have a handful of training iterations.
also been investigated in PR. Kosko’s fuzzy associative mem- Although SFAM inherently exhibits characteristics (1) and
ory (1992) has been used to represent fuzzy association. Ya- (2) listed above, experiments (Rajasekaran and Vijayalakshmi
makawa and Tomoda (1989) proposed a simple fuzzy neuron Pai 1997; Rajasekaran et al. 1997) revealed that the charac-
model for application in character recognition but failed to teristics (4) and (5) were almost nil, and the model displayed
describe the specific learning algorithm. Takagi et al. (1990) the capability to recognize/classify patterns that are only ab-
suggested a structured NN with the help of inference rules, solutely free of noise or perturbations.
but the network demands complicated training despite its bet- To overcome this problem, a careful choice of invariant fea-
ter performance. Machada and Rocha (1992) constructed a tures extracted from the patterns needs to be fed into SFAM
fuzzy connectionist expert system using a combinatorial NN as ‘‘preprocessed’’ input. Digital approximations of moment
model. Kwan and Cai (1994) presented a four-layer feedfor- invariants (Schalkoff 1989) have been employed for this pur-
ward fuzzy NN but with limited recognition capability partic- pose. However, these approximations, which are invariant to
ularly with regard to translation of patterns. translation (T) of the patterns, are not strictly invariant to ro-
A general survey of the models in the literature indicate that tation (R) and scaling (S) changes. The properties therefore
most models are applicable or have been discussed only for had to be mathematically modified before being put to use.
monochrome (binary) images or patterns and employ either a In this paper, a neuro-fuzzy pattern recognizer is discussed.
complex architecture or adopt lengthy training sessions. Fur- First, Kasuba’s SFAM and the learning equations are presented
thermore, the capability of the system to exhibit tolerance to in brief. Next, the conventional moment invariants and the
noise and pattern perturbations leaves much to be desired. modifications carried out for the successful feature extraction
are discussed. Finally, the structure of the pattern recognizer
1
Prof. and Head, Dept. of Civ. Engrg., PSG Coll. of Technol., Coim- and application of the architecture to two problems chosen
batore, Tamil Nadu 641 004, India. E-mail: [email protected] from civil engineering, namely, prediction of load from the
2
Sr. Grade Lect., Dept. of Mathematics and Comp. Applications, PSG yield patterns of elastoplastic analysis of clamped and simply
Coll. of Technol., Coimbatore, Tamil Nadu 641 004, India. E-mail: supported plates and prediction of modes from mode shapes,
[email protected] is discussed.
Note. Editor: Sivand Lakmazaheri. Discussion open until September
1, 2000. To extend the closing date one month, a written request must
be filed with the ASCE Manager of Journals. The manuscript for this
SFAM REVIEW
paper was submitted for review and possible publication on June 24,
1998. This paper is part of the Journal of Computing in Civil Engi-
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- SFAM is essentially a two layer net containing an input and
0092–0099/$8.00 ⫹ $.50 per page. Paper No. 18628. an output layer (Fig. 1).
92 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 enough match to a particular output node to be encoded by
that output node, or instead, if a new output node should be
formed to encode the input pattern.
If the match function value is greater than the vigilance
parameter, the network is said to be in a state of resonance.
Resonance means that output node j is good enough to encode
input I provided that output node j represents the same cate-
gory as input I.
A network state called ‘‘mismatch reset’’ occurs if the match
function is less than vigilance. This state indicates that the
current output node does not meet the encoding granularity
represented by the vigilance parameter and therefore cannot
update its weights even if the input patterns’ category is equal
to the category of the winning output node.
Once a winning output node j has been selected to learn a
particular input pattern I, the top down weight vector Wj from
FIG. 1. Architecture of SFAM Network the output node is updated according to the equation

Input Normalization Wnew

j = ␤(I ∧ Wjold) ⫹ (1 ⫺ ␤)Wjold (6)

NN are in general very sensitive to absolute magnitude, and where 0 < ␤ ⱕ 1. Once SFAM has been trained, the equivalent
fluctuations in inputs may tend to swamp the performance of of a ‘‘feedforward’’ pass for an unknown pattern classification
the network while predicting the desired outputs. Hence the consists of passing the input pattern through the complement
need for normalization of inputs so that the inputs correspond coder and into the input layer. The output node activation func-
to the same range of values. tion is evaluated and the winner is the one with the highest
Complement coding is a normalization rule for preserving value. The category of the input pattern is the one with which
amplitude information. Thus it represents the on-response and the winning output node is associated.
the off-response to an input vector. It also represents the pres-
ence of a particular feature in the input pattern and its absence. WORKING OF SFAM
For example, if a is the given input pattern vector of d fea-
tures, the complement coded vector a represents the absence The circle in the square problem requires a system to iden-
of each feature, where a⬘ is defined as tify which pockets of a square lie inside and which lie outside
a circle whose area equals half that of the square. This prob-
a⬘ = 1 ⫺ a (1) lem, specified as a benchmark problem for system perfor-
mance evaluation in the defense advanced research projects
The previous equation is valid because, just as in fuzzy logic, agency (DARPA) artificial neural network technology pro-
all SFAM input values must be within the range of 0–1. gram, has been used to illustrate the workings of SFAM. In
Therefore, the complement coded input vector I internal to this, SFAM is trained with a set of points for a definite number
SFAM is given by the 2D vector of training epochs. The inputs during training are the points
I = (a ⭈ a⬘) = (a1, a 2, . . . , ad , a⬘,
1 a⬘,
2 . . . , a⬘)
d (2) (x, y) and the category to which they belong, namely, inside
the circle (IN) or outside the circle (OUT). The training of
Output Node Activation SFAM has been illustrated in the following examples.

When SFAM is presented as an input pattern whose com- Example 1

plement coded representation is I, all output nodes become
active to some degree. This output activation denoted by Tj Consider the point (0.7, 0.7) as input I and whose category
for the jth output node, where Wj is the corresponding weight, is IN.
is given by Complement of I = (0.3, 0.3)
Augmented input I = (0.7, 0.7, 0.3, 0.3)
兩 I ∧ Wj 兩
Tj (I ) = (3)
␣ ⫹ 兩Wj 兩 Because I is the first input seen by SFAM in the IN category,
Here, ∧ indicates the fuzzy AND operator defined by the top down weights Wi = (Wx, Wy, Wxc, Wyc) are set to the
augmented input values and are therefore given by W1 = (0.7,
a ∧ b = min(a, b) (4) 0.7, 0.3, 0.3). The activation function value T1(I ) of the top
down weight node W1 for the input I shown is set to null and
where a and b = fuzzy vectors. Here, ␣ is kept as a small is set to point to the category IN in the category layer.
value close to 0, usually about 0.0000001. The winning output
node is the node with the highest activation function Tj . If
more than one Tj is maximal, the output node j with the small- Example 2
est index is arbitrarily chosen to break the tie. The category
associated with the winning output node is described as the I = (0.3, 0.8) and category = IN
network’s classification of the current input pattern. Augmented input I = (0.3, 0.8, 0.7, 0.2)
Activation function T1(I ) = 0.7499 (choosing ␣ =
The match function given below helps determine if learning
0.0000001)
should occur
Match function value = MF(I ) = 0.75
兩 I ∧ Wj 兩
兩I兩
(5) Here, MF(I ) is greater than the vigilance parameter ␳ (chosen
as 0.5), and because the category of I is the same as that
When used in conjunction with the vigilance parameter, the pointed to by W1, W1 is fit enough to learn the current input
match function value states whether the current input is a good I. This is accomplished by updating the weights of W1 as il-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 93
lustrated in (6), choosing ␤ = 1 [i.e., W new
1 = (0.3, 0.7, 0.3,
0.2)].

Example 3

I = (0.9, 0.9) and category = OUT

Because this is a new category, W2 = (0.9, 0.9, 0.1, 0.1) and

T2(I ) = null.

Example 4

I = (0.7, 0.9) and category = OUT FIG. 2. Example Pattern Perturbations

Augmented input I = (0.7, 0.9, 0.3, 0.1)

冕冕
⬁ ⬁

Because there are two weight nodes to decide which node is mpq = x py qf (x, y) dx dy, p, q = 0, 1, 2, . . . , ⬁ (7)
⫺⬁ ⫺⬁
fit enough to learn the new input I, the activation function
values of the two nodes are computed [i.e., T1(I ) = 0.9333 and However, in the case of a spatially discretized M ⫻ N image
T2(I ) = 0.8999]. The node with the highest activation function denoted by f (i, j), (7) is formulated using an approximation
value, W1 in this case, is chosen to learn the new input. Also, of double summations

冘冘
the match function of I with W1 is greater than the vigilance M N
parameter [i.e., (MF(I ) = 0.7) > (␳ = 0.5)]. However, due to
mpq = i pj qf (i, j) (8)
category mismatch, W1 is not fit to learn the input, and hence i=0 j=0
the next node is to be considered. Before proceeding to the
next node, match tracking is done by updating the vigilance The so-called central moments are given by

冘冘
parameter by a small quantity (i.e., ␳ = 0.501). The next node M N
W2 gives MF(I ) = 0.9. Because MF(I ) > ␳ and the categories ␮pq = (i ⫺ ˆi ) p( j ⫺ ˆj )qf (i, j) (9)
are also the same, learning occurs in W2 given by the updating i=0 j=0
of W2 as W new
2 = (0.7, 0.9, 0.1, 0.1).
where
Example 5 m10 m01
iˆ = ; jˆ = (10a,b)
m00 m00
I = (0.1, 0.3) and category = IN
Augmented input I = (0.1, 0.3, 0.9, 0.7) The central moments are still sensitive to R and S transfor-
Activation function values of W1 and W2 are T1(I ) = 0.5999 mations. The S invariant may be obtained by further normal-
and T2(I ) = 0.33 izing ␮pq or by forming

Choosing the highest activation function value, namely, W1, ␮pq

␩pq = , p ⫹ q = 2, 3, . . . (11)
the match function yields MF(I ) = 0.45, which is less than ␮ [( p⫹q)/2⫹1]
00

␳ = 0.501, rendering the node to be unfit to learn the pattern

From (11), constraining p, q ⱕ 3, and using the tools of in-
I. Choice of W2 also results in a similar case with MF(I ) =
variant algebra, a set of seven RST invariant features may be
0.3, which is less than ␳ = 0.501. In such a case, a new top
derived (Schalkoff 1989)
down weight node W3 pointing to IN is created with W3 =
(0.1, 0.3, 0.9, 0.7). ␾1 = ␩20 ⫹ ␩02 (12)
During inference, SFAM is presented with points (x, y)
alone to determine the category. In this case, the top down ␾2 = (␩20 ⫺ ␩02)2 ⫹ 4␩211 (13)
weight node that reports the highest activation function value ␾3 = (␩30 ⫺ 3␩12) ⫹ (3␩21 ⫺ ␩03) 2 2
(14)
for the given (x, y) is the winner, and the category pointed to
by the node is the category that (x, y) belongs to. ␾4 = (␩30 ⫹ ␩12) ⫹ (␩21 ⫹ ␩03) 2 2
(15)
␾5 = (␩30 ⫺ 3␩12)(␩30 ⫹ ␩12)[(␩30 ⫹ ␩12)2 ⫺ 3(␩21 ⫹ ␩03)2]
FEATURE EXTRACTION—MOMENT BASED
INVARIANTS ⫹ (3␩21 ⫺ ␩03)(␩21 ⫹ ␩03) (16)
The classification of 2D objects from visual image data is ␾6 = (␩20 ⫺ ␩02)[(␩30 ⫹ ␩12) ⫺ (␩21 ⫹ ␩03) ]
2 2

an important pattern recognition task. This task exemplifies

many aspects of a typical PR problem, including feature se- ⫹ 4␩11(␩30 ⫹ ␩12)(␩21 ⫹ ␩03) (17)
lection, dimensionality reduction, and the use of qualitative
descriptors. Moments are extracted features that are derived ␾7 = 3(␩21 ⫺ ␩03)(␩30 ⫹ ␩12)[(␩30 ⫹ ␩12) ⫺ 3(␩21 ⫹ ␩03) ] 2 2

from raw measurements. In practical imagery, various geo- ⫺ (␩30 ⫹ 3␩12)(␩21 ⫹ ␩03)[3(␩30 ⫹ ␩12)2 ⫺ (␩21 ⫹ ␩03)2] (18)
metric distortions or pattern perturbations may be observed in
the image to be classified. Fig. 2 illustrates some example Thus, for a common PR application, which involves rec-
pattern perturbations. It is therefore essential that features that ognition of 2D images, the moment based invariant functions
are invariant to orientations be used for the classification pur- ␾1 –␾7 are computed from the raw measurements of every
pose. For 2D images, moments have been used to achieve R, given shape. Those shapes with similar ␾i are classified as
S, and T invariants. Properties of invariance to R, S, and T belonging to the same class. In other words, shapes belonging
transformations may be derived using function of moments. to a class are the perturbed (rotated/scaled/translated) or exact
The moment transformation of an image function f (x, y) is version of the given nominal shape.
given by However, though the set of invariant moments shown are
94 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
invariant to T, despite their being computed discretely, the mo- as shown in Fig. 5. Here X *, Y * are the axes passing through
ments cannot be expected to be strictly invariant under R and the origin O. The dimension of any pixel is taken as a 1 ⫻ 1
S changes due to the ‘‘non 1:1 nature of general discrete square. Considering the unit thickness of the lamina, the image
geometric transformations (including rotation and scale function is analogous to mass density function. Total mass of
changes)’’ (Schalkoff 1989). an image is given as

冕冕
An investigation by the writers revealed that in the defini-
tion of ␮pq, the contribution made by a pixel had been over-
looked. The modified ␮pq definitions will be discussed in the M= f (x*, y*) dx* dy* (19)
succeeding section.
and the mass of pixel
␮pq REDEFINED
(i, j) = f (x *
i , y*
j ) (20)
Visual image data can be either binary (for monochrome

冘冘
n n
images) or real (for color images). For example, in Fig. 3 the
region R appears in the binary image as region of intensity M= f (x *
i , y*
j ) (21)
j=1 i=1
‘‘1.’’ If we represent an image function as f (x, y), then f (x, y)
is either 0 or 1 depending on whether the pixel is dark or assuming any figure is represented in a square region of n ⫻
bright. On the other hand, in Fig. 4 the intensity is represented n pixels. In general, the moment transform of an image func-
by various shades with 0 ⱕ f (x, y) ⱕ 1, indicating that the tion f (x, y) is given by
density lies anywhere between the ends of a spectrum: very
dark to very bright. This is how color images are represented.
However, the image function f (x, y) is constant over any pixel
region.
mpq = 冕冕 x*py*qf (x*, y*) dx* dy* (22)

Moments are descriptive techniques with an intuitive basis m10 which represents the first moment of the mass of the image
in the study of mechanics of bodies. Consider the real image about the Y *-axis is given by

m10 = 冕冕 x*f (x*, y*) dx* dy* (23a)

m10 of pixel (i, j) = f (x *

i , y*
j ) 冕冕 x* dx* dy* (23b)

From Fig. 6

x * = x i* ⫹ x (24)

where x *i is the distance of the center of gravity (CG) of the

pixel (i, j) from the Y *-axis. Because dx* = dx
FIG. 3. Monochrome Image

m10 of pixel (i, j) = f (x *

i , y*
j ) 冕 i ⫹ x) dx
(x * (25a)

m10 of pixel (i, j) = f (x *

i , y*
j )x *
i , ⬗ 冕 x dx = 0 (25b)

In the discrete case

冘冘
n n

m10 = f (x i*, y *
j )x *
i = Mx̄ *
C (26)
j=1 i=1

And similarly
FIG. 4. Color Image

FIG. 5. Pixel (i, j ) in Image FIG. 6. Pixel (i, j ) Represented in X *, Y *-Axes

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 95

 and physically expresses the ‘‘vertical centralness of the mass
of the image.’’

␮11: Derivation
Now

␮11 = 冕冕 XYf (X, Y ) dX dY (32)

␮11 of pixel (i, j) = f (xi , yj) 冕冕 (xi ⫹ x)( yj ⫹ y) dx dy (33)

冕 y dy = 0

␮11 of pixel (i, j) = f (xi , yj)(xi yj), ⬗ 冕 x dx = 0 (34)

冕冕 xy dx dy = 0

FIG. 7. (a) XY Axes Passing through Center of Mass (C ) of Im- ␮11 physically means the ‘‘diagonal’’ indication of the quad-
age; (b) Pixel (i, j ) Represented in X, Y-Axes rant with respect to the centroid where the region has more
mass.

冘冘
n n

m01 = f (x *
i , y*
j )y *
j = Mȳ *
C (27) ␮ 30, ␮ 03: Derivation

冕冕
j=1 i=1

where x̄ *
C and ȳ *
C denote the coordinates of the mass center of ␮ 30 = X 3f (X, Y ) dX dY (35)
the image with respect to the origin. Now consider the X, Y-

冕冕
axes passing through the center of the mass of the image as
shown in Fig. 7. Variables xi , yj represent the distance of the ␮ 30 for a pixel (i, j) = f (xi , yj) (xi ⫹ x)3 dx dy (36)
center of the pixel (i, j) from the centroidal axes. Now define
the central moments ␮pq as
␮ 30 for a pixel (i, j) = f (xi , yj)(x 3i ) (37)

␮pq = 冕冕 X pY qf (X, Y ) dX dY (28) Thus

冘冘
n n

where ␮ 00 = m 00 = M, the central mass of the image. Also, ␮10 ␮ 30 = f (xi , yj)(x 3i )
j=1 i=1
= ␮ 01 = 0, since the axes pass through the center of mass of
the image. and represents physically the ‘‘horizontal imbalance’’ of the
The derivations of ␮pq, p, q ⱕ 3 are presented next. location of the CG with respect to half horizontal extent. Sim-
ilarly

冘冘
␮ 20, ␮ 02: Derivation n n

␮ 03 =
冕冕
f (xi , yj)(y3j )
j=1 i=1
␮ 20 = X 2f (X, Y ) dX dY (29)
and represents the ‘‘vertical imbalance’’ of the location of the
CG of mass with respect to half vertical extent.
where X = xi ⫹ x.

冕 冕
1/2 1/2 ␮21, ␮12: Derivation

冕冕
␮ 20 of pixel (i, j) = f (xi , yj) dy (xi ⫹ x)2 dx (30)
⫺(1/2) ⫺(1/2)
␮ 21 = X 2Yf (X, Y ) dX dY (38)

␮ 20 of pixel (i, j) = f (xi , yj) 冉 冊 冕

x 2i ⫹
1
12
, ⬗ x dx = 0 (31)
␮ 21 for a pixel (i, j) = f (xi , yj) 冕冕 (xi ⫹ x)2( yi ⫹ y) dx dy
Thus

冘冘 冉 冊
n n
(39a)
1
␮ 20 = f (xi , yj) x ⫹
2
i ␮ 21 for a pixel (i, j) = f (xi , yj)(x y ) 2
i j (39b)
j=1 i=1 12
In the discrete case
and is physically expressed as the ‘‘horizontal centralness of

冘冘
n n
the mass of the image.’’ Similarly
␮ 21 = f (xi , yj)(x 2i yj)

冘冘 冉 冊
n n
1 j=1 i=1
␮ 02 = f (xi , yj) y2j ⫹
j=1 i=1 12 and similarly
96 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 冘冘
n n

␮12 = f (xi , yj)(xi y2j )

j=1 i=1

Here, ␮ 21 represents the ‘‘vertical divergence’’ indicating the

relative extent of the bottom of the region compared with the
top, and ␮12 represents the ‘‘horizontal divergence’’ indicating
the relative extent of the left of the region compared with the
right.
Summing up, the central moments are presented below
␮ 00 = M (Mass) (40)
␮10 = 0 (41)
␮ 01 = 0 (42)

冘冘 冉 冊
n n
1
␮ 20 = f (xi , yj) x 2i ⫹ (43)
j=1 i=1 12

冘冘 冉 冊
n n
1
␮ 02 = f (xi , yj) y2j ⫹ (44)
j=1 i=1 12 FIG. 8. Structure of SFAM Based Pattern Recognizer

冘冘
n n

␮ 21 = f (xi , yj)(x 2i yj) (45) define the palette. The image processor receives the images
j=1 i=1
that are to be trained (training patterns) or inferred by SFAM

冘冘 (inference patterns).
n n

␮12 = f (xi , yj)(xi y2j ) (46) The feature extractor obtains the RST invariant features for
j=1 i=1 each image, whether for training or inference. From the images

冘冘
n n surrendered for training or inference, the seven moment based
␮11 = f (xi , yj)(xi yj) (47) invariant functions, ␾1 –␾7, are extracted by the feature ex-
j=1 i=1 tractor for presentation to SFAM. Thus, SFAM works only on

冘冘
n n
the feature vectors of the image and not on whole images. The
SFAM activator functions as two modules: the training module
␮ 30 = f (xi , yj)(x 3i ) (48)
j=1 i=1
(SFAM training) and the inference model (SFAM inference).
The feature vectors of the training patterns and the categories

冘冘
n n
to which they belong (image category) are presented to the
␮ 03 = f (xi , yj)( y3j ) (49) SFAM’s training module. The only user selectable parameter
for the training session is the vigilance parameter ␳ (0 < ␳ <
j=1 i=1

The moments listed above are still sensitive to R and S trans- 1). Once the training is complete, the top down weight vectors
formations. Note that the S invariance may be obtained by represent the patterns learned. Next, the feature vectors of the
further normalizing ␮pq as given in (11). images that are to be recognized/classified are presented to the
From (40)–(49) it may be observed that ␮ 20 and ␮ 02 are inference module. SFAM now begins its classification of im-
different from their conventional definitions of central mo- ages by associating the feature vectors with the top down
ments. In the conventional definition of ␮ 20, for example, the weight vectors. The category of the image is output by SFAM.
term The system can handle both symmetric and asymmetric pat-

冘冘 冉冊
n n terns. However, in the case of symmetric patterns it is essential
1 that only distinct portions of the images be trained. This is
f (xi , yj)
j=1 i=1 12 because in the case of doubly symmetric patterns, their RST
invariant feature vectors ␾2 –␾7 acquire values that are very
has been neglected. This omission has resulted in a cascading
close to 0, and ␾1 tends to 1. This consequently results in
effect rendering ␩ 20 and ␩ 02 and the functions ␾1, ␾2, and ␾6
feature vectors, which are similar, leading to misclassification
incorrectly defined, leading to the misclassification of images
of patterns.
with RST orientations.
Using the normalized central moments and tools of invariant
algebra, a set of seven RST invariant features, similar to those APPLICATION 1: PREDICTION OF LOAD FROM YIELD
shown in (12)–(18), may be derived. Note that ␾7 is actually PATTERNS OF ELASTOPLASTIC, CLAMPED, SIMPLY
ST invariant and changes its sign for reflection. This is because SUPPORTED PLATES
for reflection about the Y-axis, let us say x becomes ⫺x, for
odd p, ␩ pq (i.e., ␩ 30, ␩12) changes sign, and hence ␾7 changes Benjamin Whang (1969) developed a finite-element dis-
sign for reflection. placement method for elastoplastic analysis of bilinear strain
hardening orthotropic plates and shells, also considering elastic
STRUCTURE OF SFAM BASED unloading. The two basic approaches adopted by Whang in
PATTERN RECOGNIZER the elastoplastic analysis are the initial stiffness approach and
tangent stiffness approach in conjunction with the Huber-
The overall structure of the pattern recognizer is as illus- Mises yield criterion, and the Prandtl-Reuss flow rule in ac-
trated in Fig. 8. The images (patterns) whether monochrome cordance with the strain hardening yield function.
or color are input through the image processor. In this work, Fig. 9 shows the uniformly loaded isotropic plates, one with
the images are engraved on a grid size of 40 ⫻ 40. Also, clamped edges and the other with simply supported edges. The
images may have their colors chosen from a fixed palette. Here following numerical values are used:
the colors that are fixed are white, black, green, yellow, and
magenta. However, any number of colors could be used to t = 1.0, E = 30,000.0, ␯ = 0.3, Ep = 300.0, and ␴0 = 30.0

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 97

 FIG. 9. Isotropic Plates: Clamped versus Simply Supported

FIG. 11. Peak Values of Displacements

FIG. 12. Pattern as Presented to SFAM (Prediction of Natural

Mode Shapes)

FIG. 10. Pattern as Presented to SFAM (Prediction of Load

from Yield Patterns) mode. The system is tested for noisy and noise free patterns.
In all cases, the predictions are quite correct.
where t = thickness; E = Young’s modulus; ␯ = Poisson’s ratio;
and ␴0 = yield stress. The formation of a plastic zone with CONCLUSIONS
respect to loading is also shown in the figure.
SFAM is trained with the patterns representing the plastic In this paper, a pattern recognizer based on SFAM has been
zones and their corresponding loading and tested for its infer- discussed. The model employs a moment based RST invariant
ence capability. Considering the doubly symmetric nature of feature extractor, which has been modified to accommodate
the patterns, only a quarter of the image is presented to SFAM strict invariance to R, S, and T changes.
for training. Fig. 10 illustrates a sample pattern for SFAM For classification of asymmetric patterns, the moment based
training. Observe that it is a quarter of the pattern correspond- feature extractor extracts unique feature vectors. But for dou-
ing to a loading of 1.8 of a simply supported plate. The pre- bly and multisymmetric patterns, the feature vectors turn out
diction of load by SFAM is done quite correctly. to be similar resulting in misclassification. Though the prob-
lem has been tackled by considering only a portion (quarter)
of the pattern, further investigations could be undertaken to
APPLICATION 2: PREDICTION OF NATURAL MODE
look for a feature extractor that is insensitive to symmetricity
SHAPES OF MULTISTORIED BUILDING FRAMES
of patterns as well as their perturbations.
In this case, the patterns representing the peak values of
displacements for a five-story shear frame are presented as ACKNOWLEDGMENTS
inputs to SFAM. Fig. 11 illustrates the peak values of dis- The writers express their thanks to the management and Dr. P. Rad-
placements. A sample pattern presented to SFAM has been hakrishnan, Principal, of PSG College of Technology, for providing the
shown in Fig. 12. The model presented one exemplar for every facilities to carry out this research work.

98 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

APPENDIX I. REFERENCES different geometry.’’ Proc., Nat. Conf. on Neural Networks and Fuzzy
Sys., Chennai, 107–116.
Carpenter, G. A., Grossberg, S., Markuzon, N., Reynolds, J. H., and Ro- Schalkoff, R. (1989). Digital image processing and computer vision. Wi-
sen, D. B. (1992). ‘‘Fuzzy ARTMAP: A neural network architecture ley, New York.
for incremental supervised learning of analog multidimensional maps.’’ Takagi, H., Konda, T., and Kojima, Y. (1990). ‘‘Neural network design
IEEE Trans. Neural Networks, 3(5), 698–712. on approximate reasoning and its application to the pattern recogni-
Chao, C.-J., and Cheng, F.-P. (1998). ‘‘Fuzzy pattern recognition model tion.’’ Proc., Int. Conf. of Fuzzy Logic and Neural Networks, 671–674.
for diagnosing cracks in RC structures.’’ J. Comp. in Civ. Engrg., Watada, J., Fu, K. S., and Yao, J. T. P. (1984). ‘‘Linguistic assessment of
ASCE, 12(2), 111–119. structural damage.’’ Rep. CE-STR-84-30, Purdue University, West La-
Ishizuka, M., Fu, K. S., and Yao, J. T. P. (1982). ‘‘A rule based inference fayette, Ind.
with fuzzy set for structural damage assessment.’’ Approximate rea- Whang, B. (1997). ‘‘Elasto-plastic orthotropic plates and shells.’’ Proc.,
soning in decision process, M. M. Gupta and E. Sanchez, eds., North- Symp. on Application of Finite Element Methods in Civ. Engrg., ASCE,
Holland, Amsterdam, 261–275. New York, 481–515.
Kasuba, T. (1993). ‘‘Simplified fuzzy ARTMAP.’’ AI Expert, Nov., 18– Yamakawa, T., and Tomoda, S. (1989). ‘‘A fuzzy neuron and its appli-
25. cation to pattern recognition.’’ Proc., 3rd Int. Fuzzy Sys. Assn. Congr.,
Kosko, B. (1992). Neural networks and fuzzy systems: A dynamical sys- 30–38.
tems approach to machine intelligence. Prentice-Hall, Englewood
Cliffs, N.J. APPENDIX II. NOTATION
Kwan, H. K., and Cai, Y. (1994). ‘‘A fuzzy neural network and its ap-
plication to pattern recognition.’’ IEEE Trans. on Fuzzy Systems, 2(3), The following symbols are used in this paper:
185–191.
Machada, R. J., and Rocha, A. F. (1992). ‘‘A hybrid architecture for fuzzy f (x, y) = image function;
connectionist expert systems.’’ Hybrid architecture for intelligent sys- mpq = moments;
tems, A. Kandel and G. Langholz, eds., CRC, Boca Raton, Fla.
Tj (I ) = activation function of jth node for input I;
Rajasekaran, S., and Vijayalakshmi Pai, G. A. (1997). ‘‘Application of
simplified fuzzy ARTMAP to structural engineering problems.’’ Proc., Wj = top down weights associated with jth output node;
All India Seminar on Application of Neural Networks in Sci., Engrg. ␩pq = normalized central moments;
and Mgmt., Bhubaneswar. ␮pq = central moments;
Rajasekaran, S., Vijayalakshmi Pai, G. A., and Jesmon, P. G. (1997). ␳ = vigilance parameter; and
‘‘Simplified fuzzy ARTMAP for determination of deflection in slabs of ␾i = RST invariant moment function.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 99

 JAVA-BASED REGULATION BROKER
By Vei-Chung Liang1 and James H. Garrett Jr.,2 Member, ASCE

ABSTRACT: Many engineering activities, such as building design and construction, site redevelopment, and
industrial operations, are governed by environmental regulations that are constantly being revised and published.
Many state environmental protection agencies are now starting to use Web sites to provide access to their
regulations. From the result of a preliminary investigation, many current state regulatory agency Web sites
provide no support, other than browsing, for finding relevant regulations. A Java-based environmental regulation
broker has been designed to help users in locating relevant environmental regulations. Eventually, such broker
software can be provided by regulatory agencies as a standardized support to access their regulations over the
Internet.

INTRODUCTION line regulations, some states used central Web sites, which
were often the state legislation Web sites. Thus, the agencies
The U.S. Environmental Protection Agency (EPA) and state only provided hyperlinks to the corresponding on-line regu-
departments of environmental protection define and execute lations. However, these state agencies lost access control to
environmental policies and regulations for individuals, com- their on-line regulations, which is needed to update, delete,
panies, and corporations to follow. These environmental reg- and classify their regulations. Most state agencies disclaimed
ulations, which are frequently revised, limit the nature and the accuracy of the on-line documents and regulations. The
amounts of air pollution, water pollution, and solid waste that second feature, search mechanism, examined how users found
can be generated from industrial activities. Members of the the relevant regulations on-line. Two major search mechanisms
regulated community request environmental regulations from, were found at the Web sites: classifier-based and full-text
and ask environmental policy questions to, the U.S. EPA and search mechanisms. Sites that provided classifications or cat-
the state agencies. Most often, the reason for asking such ques- egories would generally use air, water, land, and waste as the
tions and requesting regulations is that these people do not only classifiers. Full-text search was used for searching over
want to fail to consider, or violate, the current environmental the on-line regulations or over the entire HTML files of the
regulations and policies that are applicable to their operations. sites. The third feature, permitting service, examined the on-
To comply with the most up-to-date environmental policies line permitting process provided by the agency. The state agen-
and regulations, engineers and lawyers have to locate and de- cies, which were authorized to approve permitting applica-
termine which regulations, and provisions in these regulations, tions, usually put their application forms, frequently asking
apply to their companies’ activities. The regulatory agencies questions (FAQ), and guides on their sites. Several sites pro-
are trying to make their regulations more accessible by adopt- vided wizard-like interactive questions to help users find the
ing new technologies. It is our premise that Web-based support required permits for their business activities. The fourth fea-
for accessing and using regulations can provide an easier way ture, viewing support, examined the formats of the on-line
for users to access the most recent regulations, and for main- regulations. The formats of these regulations included: plain
tainers to maintain and update these regulations (Liang 1998). text, Adobe PDF files, WordPerfect files, Microsoft Word files,
We also suggest that support for searching, in addition to and HTML files. The regulations were sometimes divided into
browsing, is necessary to accelerate the process of locating the chapters and sections. Few state agencies were taking advan-
applicable regulations and provisions for the regulated public. tage of the hyperlink function in the HTML files for cross-
A preliminary investigation has been conducted of the 50 referencing the set of regulations. Several agencies provided
state EPA Web sites. The focus of the investigation was to downloadable PDF files or ZIP files of the regulations. The
discover the usability and the distribution of the on-line reg- last feature, notifying service, examined the ways that the
ulations at these Web sites. Five features were evaluated for agencies notified their community about the changes of the
each state agency’s Web site: regulation on-line service, search regulations. One state agency allowed users to submit their
mechanism, permitting service, viewing support, and notifying e-mail addresses to obtain updated information about their reg-
service. The results are summarized in Table 1. The degree to ulations. Others provided news hyperlinks or labeled the up-
which the area of a circle is filled in Table 1 represents the dated regulations with icons. Several states also provided on-
level that the feature is provided by the state agency’s Web line access to permitting statuses.
site. The first feature, regulation on-line service, represents From this preliminary investigation, we can see that most
whether the Department of Environmental Protection or state of the state agencies have put their regulations on the Web
agency put its regulations on-line. Some state agencies put with different levels of support. Different strategies and meth-
partial sets of regulations on-line because they were relevant ods must be utilized by the users of these systems to search
to permitting processes. Others only put office phone numbers and browse the regulations relevant to their business activities.
on the Web by which requests of information about their reg- However, to help and allow users to locate more easily and
ulations could be made. To provide a unified version of on- effectively the applicable and appropriate regulations and pol-
1 icy documents, a systematic approach for treating these doc-
Res. Asst., Dept. of Civ. and Envir. Engrg., Institute for Complex
Engineered Systems, Carnegie-Mellon Univ., Pittsburgh, PA 15213. uments is necessary. In this paper, we describe a Java-based
2
Prof., Dept. of Civ. and Envir. Engrg., Institute for Complex Engi- regulation broker system that the environmental agencies
neered Systems, Carnegie-Mellon Univ., Pittsburgh, PA. could use to provide on-line regulation search services to their
Note. Editor: Sivand Lakmazaheri. Discussion open until September regulated constituency. In this paper, we will focus on the tech-
1, 2000. To extend the closing date one month, a written request must nical feasibility of the regulation broker. Although important,
be filed with the ASCE Manager of Journals. The manuscript for this
the legal issues of using the broker system will not be ad-
paper was submitted for review and possible publication on December 9,
1998. This paper is part of the Journal of Computing in Civil Engi- dressed here, as the broker is not intended to replace lawyers
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- but rather augment the regulated public’s ability to seek out
0100–0108/$8.00 ⫹ $.50 per page. Paper No. 19824. relevant information before seeking legal advice. The history
100 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
TABLE 1. Investigation Results of 50 Environmental State Agency Web Sites

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 101

 TABLE 1. (Continued)

of the regulation broker is first reviewed. The Java-based bro- search engine inside the Inquery search engine was used for
ker system and its architecture are then discussed, and next, a both searching schemes. The regulations in the agency data-
pilot test of this system and the results of this test are pre- base were marked up with standard generalized markup lan-
sented. Finally, the future directions for a standardized regu- guage (SGML) tags for the SGML build program to create the
lation broker system are discussed. internal indexing codes. The SGML tags were used to store
the assigned classifiers with the text, the classification system,
HISTORY OF THE REGULATION BROKER and the logical structural information of the regulations and
provisions.
An environmental regulation broker is a piece of software Using this CGI-based broker, the regulation users were able
that assists engineers and companies to identify the applicable to retrieve applicable regulations from the broker, to traverse
environmental regulations and policies related to their business the hierarchical structure of a selected regulation, and to per-
or engineering activities. The broker takes advantage of the form classifier-based and full-text searches for specific provi-
connectivity of the Internet and delivers the applicable regu- sions. To browse a particular provision by using classifier-
lations to users’ machines. To assist users in identifying the based and full-text searches, the user had to request a list of
applicable regulations, a broker system has to provide search applicable regulations, select one of the applicable regulations,
schemes to locate the relevant regulations and browsing func- request the query form, fill out and send the query form, and
tionality for users to navigate through the contents of docu- then select the desired provision. Several disadvantages were
ments. Two different implementations of such a broker system found to exist with this previous approach: (1) Stateless op-
have been developed. In this section, a common gateway in- erations on client’s machine; (2) heavy workload of server;
terface (CGI) Web-based approach is reviewed. The more re- and (3) lack of editing tools. Each of these disadvantages are
cent system, the Java-based approach, is introduced in the next described in the following sections.
section.
Krofchik (1995) and Stasiak (1996) used CGI scripts and Stateless Operations on Client’s Machine
the Inquery search engine (Callan et al. 1992) to generate the
user interfaces and necessary searching components for a pre- The regulation users had to send HTTP requests to the bro-
liminary broker system. The CGI scripts were used to allow ker server when they wanted to navigate to the next level of
agencies to register their regulations to the broker system, to the classifiers. The users might need to send several HTTP
allow users to retrieve the applicable regulations, and to allow requests to find the document because the Web browsers could
users to do the full-text and classifier-based searches. The not store the navigation status. If users were accessing the
102 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
server via low-speed connections to the Internet, such as by a
modem, they could do nothing with the broker except wait for
results from the broker.
Heavy Workload of Server
Because every user had to send queries for retrieving the
documents and navigating the classification system, the work-
load of the server was made very heavy. The heavy workload
also came about because the broker had to keep track of the
states of the connected client machines. A better approach is
to have users navigate the classification systems by using the
resources of the users’ machines. Then, after users review and
select a set of classifiers, they can send the selected classifier
set to the broker to retrieve the applicable documents. Sup-
porting this usage pattern improves the usability and perfor-
mance of the broker system.
Lack of Editing Tools
In the CGI-based approach, the maintainer of the regulations
had to put SGML tags inside the documents for the search
engine to use when indexing the documents. The tags were
used for recording the associated classifiers and the logical
structure of the documents. The logical structural information
of the provisions was used to record the relationships between
the document files. Having the content of the classification
system mixed with the SGML tags in the document file could
lead to an inconsistency of the classification system. There
were no integrated editing tools available for the maintainers
to use in managing the document files, their structure, and their
classification.
JAVA-BASED VERSION OF BROKER SYSTEM
To address the limitations of the previous CGI-based broker,
a Java-based version of the broker system was developed that FIG. 1. New Scheme for Java-Based Broker System
provides a new user interface, a persistent data storage mech-
anism, and a self-contained search engine. The new scheme
for the broker system is presented in Fig. 1. The user interface to represent these items. The original form of most regulations
has been redesigned and allows users to navigate the entire is a paper hardcopy. These hardcopy documents must be trans-
classification system without sending requests to the Web formed into an electronic form so that the broker system can
server. The user downloads the user subsystem from the broker search through the regulations and present the relevant pro-
Web site. The user subsystem retrieves the whole classification visions to the users. Optical character recognition technology
system from the regulation database and displays it on the can be used to convert the hardcopy text into electronic form.
screen. The user can then perform classifier-based and full-text All the text, images, tables, and formulas must also be stored.
searches. The user subsystem sends the query to the broker Most of the current document formats, such as HTML, RTF,
server and displays the query result, which is the applicable Postscript, Microsoft Word, and PDF can be used to embed
regulation list, on the screen. the tables, images, and formulas within the documents. How-
A database management system (DBMS) is used to separate ever, most of these formats require a specific computer appli-
the classification systems and structural information from the cation to access them. Some applications are expensive and
document files. The DBMS is not only used for storing the some are not Internet-enabled. We chose the HTML format
classification systems and structural information about the doc- since it is supported by the HTTP protocol and the user is able
ument, but also for performing the classifier-based search. A to browse the documents by using the various Web browsers
commercial off-the-shelf search engine is used for performing that are freely available on the Internet. In addition, most word
full-text search. The search engine receives search queries, processors used to compose these documents now provide the
such as strings and keywords, from the user subsystem and functions to convert documents into HTML format. The math-
performs the full-text search over the regulation files. On the ematical equations within the regulations can be converted into
broker server, the bodies of the regulation files are not in the image files embedded in HTML files automatically. The users
database because both the database server and search engine can use their own text editors to modify the regulations with-
have to retrieve and locate these files. out digging into the HTML source files.
Several issues were considered during the design of the The broker maintainers can classify a regulation document
Java-based broker system prototype: (1) The representations with one or many classifiers. This document-classifier rela-
of the regulations and classification systems; (2) the platform tionship information should be stored in the broker system. If
of the prototype broker; and (3) the enabling technologies the information were stored inside the electronic format of the
needed to deliver the requirements of the broker concept. documents, it would increase the complexity of updating the
classifier assignment or the context of the regulations, which
Representations of Regulations and Classification had been discovered in the CGI-based version of this broker
Systems described previously.
Regulations contain text, images, tables, and formulas. We chose to use a relational DBMS to store both the doc-
Therefore, the representation of these documents must be able ument and classifier hierarchies. The hierarchical relationships
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 103
between the classifiers can be represented in a relational
DMBS (Liang 1998); the relationships between classifiers and
documents can be treated as ‘‘has a’’ or ‘‘has many’’ relation-
ships. Information that needs to be stored in the database con-
sists of locations of document files, classifiers, relations be-
tween classifiers and documents, the classification hierarchy,
the document hierarchy, and the organization hierarchy. We
decided to store the location of the document in the database
but save the document body as one or more files in a directory
of a file system on the broker server. Using this approach, we
need only save the HTML-based layout and format of the reg-
ulations. However, an external full-text search engine is
needed as the DBMS cannot perform string search over the
external HTML files.

Platform of Prototype Broker

We used a Microsoft NT workstation as the platform to
develop the broker system. The benefit of using NT is that it
provides a Peer Web Server that contains a component, the
Microsoft Index Server (Swank and Kittel 1997), which pro-
vides the full-text search functionality. We could have used
other search engines if we established the interfaces between
the broker system and these search engines. To date, the in-
terface for the MS Index Server was the only one that we have
developed. The DBMS used for the broker system was Mi-
crosoft Access 97 with the open database connectivity 32
driver (ODBC 32) for developers to use in writing applications
that communicate with MS Access.

Enabling Technologies Needed to Deliver

Requirements of Broker Concept
To have the broker system adopted by most organizations
or agencies, the broker system must run in most organizations
and allow users to access the broker server from different plat-
forms. Because the broker system is an Internet application, FIG. 2. Three-Tier Broker Architecture
the user can employ the broker system via a web browser from
different types of operating systems and machines, which ad- groups and authors of the regulations to maintain the regula-
dresses the heterogeneous platform problem from the users’ tion documents and the classification systems. Both subsys-
side. To meet this requirement, we decided that we have to tems use the same three-tier system architecture, as shown in
choose a language that operates on most platforms to reduce Fig. 2. In this section, we will explain the three-tier architec-
the cost of the development. ture starting with the user subsystem. The differences between
Several technologies were considered as the candidates for the maintainer subsystem and the user subsystem will be dis-
delivering the broker system: the Java platform (Flangan cussed later in this section.
1996), ActiveX controls (Homer 1998), and CGI scripts (Kim
1996). By examining the features of these options, we found System Architecture: Three-Tier Architecture
Java to be the best choice to fulfill the requirements of the
broker system. The Java Applets keep track of the states of Users download Java Applets from a Web server and launch
the operations at the client machines and reduce the workload the applets in their Web browsers. Due to the security con-
of the broker server. These applets overcome the stateless op- cerns, the Java Virtual Machine (VM) in the Web browser only
erations of the previous CGI-based broker and ActiveX con- allows an applet to communicate back to the host machine
trols. The frequency of the web server disruption and service from where it was sent. In some circumstances, the DBMS
degradation is reduced because part of the workload of the server accessed by the broker system might not be located
broker system is transferred to the client’s machine. Users necessarily in the same machine as the Web server. Thus, the
from different operating systems can use Java-enabled brows- remote Java Applet cannot communicate with the DBMS
ers to download the Java Applet for locating and retrieving server directly. Instead of using the traditional two-tier client/
the applicable regulations. For the prototype, it was decided server architecture, we decided to use a three-tier architecture
to use the Java language as our delivery language and devel- to overcome this constraint of the Java VM. The first tier is
opment environment and combine it with an off-the-shelf the Java Applet requesting information from the Web site. The
search engine and relational database management system run- second tier is the Web server, called the broker server, pro-
ning on an NT workstation. cessing and transmitting messages between the other two tiers.
The third tier is the back-end database server providing reg-
Architecture and System Details ulation information to the broker server.

The broker system consists of two primary subsystems: a First Tier

user subsystem and a maintainer subsystem. The user subsys-
tem allows the regulation user to retrieve and search the ap- The first tier is the broker client running on the user’s ma-
plicable regulations. The maintainer subsystem allows the IT chine. The broker client sends requests to the broker server
104 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
and displays the returned value set, in an appropriate interface, .IDQ file to store general query information for the MS Index
for the user to browse and manipulate. The client contains two Server. The .IDQ file informs the MS Index Server of the
components communicating with the second tier. The first physical location of the document collection, the attributes the
component is the DBClient component. The DBClient is MS Index Server should return (i.e., filename, document title,
responsible for sending structured query language (SQL) que- etc.), query types (i.e., phrase query, natural language query,
ries to, and receiving query results from, the DBServer com- etc.), and the location of the .HTX file. The .HTX file de-
ponent inside the second tier. The DBClient composes the scribes the format of the return results. The MS Index Server
SQL statements and sends statements to the broker server. It will search over the physical location, retrieve the required
saves the returned query results in the memory of the user’s attributes for each hit document, and send results back to the
machine that can be retrieved later by the broker client. The remote client in the format described in the .HTX file.
second component in the first tier is the FTQuestioner,
which stands for full-text questioner. This component is re- Third Tier
sponsible for sending full-text search queries to the index
server connection in the second tier, receiving the search re- The third tier of the system contains the back-end database
sults, and parsing them into a format that can be understood server of the broker, a repository of HTML documents, and
by the broker client. The broker client then joins the results the MS Index Server. The back-end database server, which
from the DBClient and FTQuestioner and displays the in- responds to requests from the second tier, does not necessarily
tersection of these two lists of applicable regulations on the have to run on the same machine as the Web server, since they
user’s Java Applet interface. are in different tiers. However, for the prototype of the broker
system at this time, it does run on the same machine.
Second Tier The MS Index Server updates the index file periodically.
We could not find any documents describing the internal
The second tier is a broker server residing on a machine search algorithm used by the MS Index Server. It is not a
accessible from the Internet. Due to the Java VM security pol- requirement to use the MS Index Server as the search engine
icy, a broker server was used as the second tier to dispatch the in the broker. The MS Index Server was selected because it is
requests to the corresponding back-end database and full-text a self-contained search engine and is freely available. The MS
search servers. Generally speaking, the second tier contains Index Server can perform full-text search over various file for-
two main components. One is the broker server and the other mats without modifying the original files. In order to use some
is the index server connection. The broker server is a com- other type of search engine, another interface for the second-
ponent embedded in the Web server that communicates with tier connection would have to be provided.
the back-end database system. The index server connection The HTML file repository is the external file directory that
communicates with the back-end full-text search engine. is used to store the regulations and documents in the HTML
The broker server was designed as a Java Servlet, which is format. The files are located on the same machine as the
a Java Applet running at the server site (Moss 1998). The first DBMS server. Thus, maintainers can easily save the file lo-
significant benefit of using the Java Servlet is that the admin- cations into the database. When document maintainers create
istrator does not have to launch the broker server as a Java or insert document information into the database, they also
application every time the server machine is rebooted. The have to save the information describing the physical location
Web Server launches the Java Servlet whenever it needs to of the documents to the database.
respond to a user query. The second benefit is that it uses the
same port number as the web server. During the pilot-test,
User Subsystem
Pennsylvania Department of Environmental Protection (PA-
DEP) users had to communicate with the broker server through The user subsystem provides the user with an interface to
their organizational firewall that filtered out all the TCP/IP port communicate with the remote broker server. The main func-
connections other than the standard HTTP port 80. Using a tions provided by the user subsystem are querying the database
Java Servlet solves the problem because it can be set up to for applicable regulations, browsing the classification system
listen to the same port as the Web server. The architecture of of the agency, browsing the regulations, browsing the classi-
the servlet is multithreaded in nature; thus, the servlet-enabled fication system of a specific regulation, browsing the docu-
Web server will start-up multiple services in response to dif- ments, performing a keyword search, and viewing the contents
ferent clients’ requests. of the document. The first tier of the user subsystem is down-
The broker server uses JDBC and ODBC to communicate loadable from the Web server. The Web server that serves as
with the back-end DBMS. JDBC is an application program- the second tier should be set up at the agency site. The third
ming interface for executing SQL statements in Java (Hamil- tier, the back-end database server and the full-text search en-
ton and Gattell 1997). JDBC can be used to communicate with gine, should also be set up at the agency site.
different commercial database servers as long as the JDBC The user subsystem contains two windows, as shown in
drivers are provided. In the second tier of the broker, the Figs. 3 and 4. The first window is the document list viewer
DBServer component receives SQL statements composed by and the second is the document viewer. The document list
the DBClient from the first tier applet, passes the queries to viewer allows users to identify the applicable regulations by
the appropriate JDBC drivers, and returns the query results using classifier-based searches. The applicable regulations are
from the JDBC drivers to the DBClient. The DBServer shown in the right panel of the window. The second window,
only needs to load the JDBC driver corresponding to the back- the document viewer, allows users to browse the contents of
end DBMS. The current broker uses MS Access as the back- selected applicable regulations from the document list viewer.
end DBMS server. There is no JDBC driver available for MS Users can perform classifier-based and full-text searches over
Access, so the JDBC-ODBC bridge driver is used to convert the current viewing regulations. The content of the selected
JDBC operations into ODBC operations, which can be rec- section of the regulations is shown in the content viewing area.
ognized by MS Access.
The last part of the second tier is the index server connec- Maintainer Subsystem
tion. The index server connection is used to communicate with
the MS Index Server to provide the full-text search capability The maintainer subsystem is intended to support maintainers
of the broker system. The index server connection uses an of the regulations as they update, delete and assign the regu-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 105
 tainer subsystem does not have a content viewing area. Be-
sides being able to update regulation document structure and
classifier relations, the viewer can access the local file system
information to store physical file locations in the back-end
database. The first tier of the maintainer subsystem must be
installed on the maintainer’s machine. The second and third
tiers must be set up at the agency site.

PILOT TEST
One objective of this specific research project was to con-
duct a field test with the PADEP in order to evaluate the us-
ability and performance of the broker system. The field test
started after the maintainer subsystem was built. We obtained
27 Air Quality Policy documents in MS Word format from the
PADEP. The documents were converted into HTML files by
using the ‘‘save as HTML’’ command provided by MS Word.
As each of these documents contained chapters and sections,
the documents were dissected based on their sections. Hence,
FIG. 3. Document List Viewer the table of contents of the document was used to represent
the logical structure of the document. For example, the doc-
lations, provisions, and classifications. The maintainer subsys- ument ‘‘Prevention of Significant Deterioration’’ contains 12
tem was built as a Java application. Unlike the Java Applet, a sections. Each section was saved as an HTML file in the di-
Java application can create a direct input and output stream rectory ‘‘prevention㛭of㛭significant㛭deterioration.’’ All the reg-
among the local machine and remote machines. The major ulation directories were saved under the parent directory
differences between a Java Applet and Java application is that ‘‘AirQualityPolicies.’’
the application is considered as a secure program by the Java Working with PADEP engineers, we developed a classifi-
VM. Thus, the Java application can communicate not only cation system for these documents. We used the same classi-
with the local machine and file system, but also with other fication system for classifying the policies and the provisions
remote database servers. In the maintainer subsystem, no index of these policies. The classification system was a four-level-
server connection component is needed since the MS Index deep hierarchy, with a total of 151 classifiers. Each classifier
Server has provided an administrator user interface for updat- was then assigned to the applicable sections of these 27 doc-
ing and indexing the files. The functionality of the DBServer uments used in the pilot study. It took about two complete
of the maintainer subsystem has been extended so that it can days to finish the assignment of classification. The result of
handle updating and deleting commands in SQL. this assignment process was then input into the broker system
The maintainer subsystem also has two user interface win- by using the maintainer subsystem. It took about 4 days to
dows for the maintainers to use in manipulating the regulation input the classifiers into the broker system. One way of in-
information. Additional pop-up menus are provided for main- creasing the document classification speed will be discussed
tainers to access maintenance commands. The maintainer’s in the next section. The following use case [see Jacobson
document viewer provides not only navigation functions but (1992) for the description of a use case] demonstrates how
also maintenance commands to maintain the list of regulations users locate and access the applicable regulations and docu-
known to the broker. The document list viewer of the main- ments in the PADEP Air Quality Policy collection.

FIG. 4. Document Viewer

106 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 Assume that the user wants to find the PADEP air quality similar query occurs in the future. There are no monitoring or
policy documents that relate to ‘‘volatile organic compounds’’ feedback subsystems in the current broker system for tracing
(VOCs). The user needs a Java-enabled web browser to launch and recording the satisfaction of users and their behaviors.
the broker system. The user selects classifiers from the clas- To associate a document or a provision with the proper clas-
sifier hierarchy to describe the air quality policies desired. The sifiers is a time-consuming process that increases exponen-
number of applicable documents in the list changes as the user tially in time as the number of provisions and classifiers in-
selects the classifiers. The user can select multiple classifiers creases. If the maintainer subsystem could learn from the
at the same time to narrow down the search result. To open existing document-classifier relationship data, it might be able
multiple regulations or documents simultaneously, the user se- to assist the maintainer in identifying the possibly relevant
lects multiple documents and launches the Document Viewer. classifiers for a given provision or document. Text categori-
To remove the nonrelevant sections from document list viewer, zation technologies have been developed for a decade. The
such as these related to ‘‘sulfur compound testing,’’ the user accuracy for classifying a document is about 80–90% (Koller
can do a classifier-based search and a full-text search to find and Sahami 1997; Yang 1994). Using automated text catego-
only those provisions that apply to VOCs. For example, there rization automation might help in reducing the time for main-
are only two applicable sections left after the user does the taining the broker system. This was tried by Stasiak (1996)
full-text search on ‘‘VOC.’’ The user then views the contents using a statistical correlation approach with limited success.
of a provision as shown in Fig. 4.
After inputting 151 classifiers and 27 air quality policy doc- SUMMARY
uments into the broker database, the time to start up the broker
client took more than 5 minutes. The overhead was due to the Using the Java platform to deliver the broker system is a
processing of the SQL queries that requested classifier and better approach than the CGI-based broker system. The exist-
document node information. To remedy this problem, we cre- ing technologies and framework supported in Java, such as
ated external cache files to save the node information in ad- Java Servlets, the JDBC/ODBC bridge, and platform-indepen-
vance since the files would not be changed unless the main- dent graphical user interfaces, saved development time for de-
tainer of the broker system updated the data. After making this livering the broker system. This approach based on a Java
cache, the communication overhead dropped by a factor of 7. Applet has transferred several parts of the work load from the
server machine to the remote client machine, such as display-
DISCUSSION OF THE JAVA-BASED BROKER ing the entire classification system and the contents of the
document hierarchy.
The use of Java brings several features to the broker system. The Web-based broker system extends the accessibility and
The use of the JDBC allows us to adopt the three-tier archi- saves the time and cost of accessing the regulations. The Web-
tecture for the broker system. The three-tier architecture allows based broker system allows the user to obtain the most up-to-
the regulatory agencies to use different types of DBMS servers date regulations and policy information faster than the tradi-
as the third tier of the broker system. Thus, the agencies can tional hardcopies and CD-ROMs. The classifier-based search
directly use their existing DBMS servers to reduce the instal- and full-text search provided by the broker system allows the
lation cost of the broker system. Users only need Java-enabled user to locate applicable regulations more easily. The Java-
web browsers to launch the first tier of the broker system. The based broker system provides a friendlier user interface and
graphical user interface components in the first tier allow the more easily maintained system architecture. The searching
user to visualize the entire classification systems and logical tools in the broker system are used to help the users locate the
structure of the documents in a single window. Unlike the applicable regulations.
CGI-based broker, no communication occurs between the bro- The broker system is currently running on a single machine
ker client and server when the user navigates the structure of and contains documents from one regulatory agency (PADEP).
the classification or the regulations. The Java-based broker Users will likely want to search for relevant environmental
system thus reduces the communication time between the regulations or documents over all 50 states and other nations.
browser and the server and provides a better visual environ- One direction in which the next version of the broker system
ment for supporting the user in navigating the classifiers and should be extended is to make it a distributed system of many
documents. different communicating agency brokers. Such a distributed
The broker system uses a classification system and full-text system could allow a user to send queries to a central broker
search to help the user to identify applicable regulations. Nev- that then sends out appropriate queries to these distributed da-
ertheless, the terms used in the classification might not convey tabases, collects and combines the results of these searches,
the same meaning as that understood by the users. Missing and sends them back to the user.
relevant regulations and returning irrelevant documents were
the two errors that could occur in the broker system. Such
ACKNOWLEDGMENTS
errors might come from misclassified documents and unclear
full-text search queries. As a result, users would not be able This paper describes research sponsored by the PADEP, the Institute
to find all the relevant regulations that apply to their business for Complex Engineered Systems at Carnegie-Mellon University (Pitts-
or engineering activities. They might violate some regulations burgh, Pa.), and the Pennsylvania Infrastructure Technology Alliance Pro-
gram, which is sponsored by the Pennsylvania Department of Community
or provisions that were not displayed during their search ses- and Economic Development. The writers would also like to acknowledge
sions. The broker system should provide other mechanisms to Bongjin Choi for his help in developing the Java-based version of the
assist the users in expressing their questions, such as wizard- broker.
like interface. As the broker is only intended as an early as-
sistant for finding potentially relevant regulations, other re- APPENDIX. REFERENCES
sources, such as agency officers and lawyers should also be
consulted before making any critical decision regarding these Callan, J. P., Croft, W. B., and Harding, S. M. (1992). ‘‘The INQUERY
environmental regulations. retrieval system.’’ Proc., DEXA ’92, Int. Conf. on Database and Expert
System Applications, Springer, Berlin, 78–83.
The broker system should also collect users’ degree of sat- Flangan, D. (1996). Java in a nutshell. O’Reilly & Associates, Inc., Se-
isfaction with the results of the broker searches. Using this bastopol, Calif.
relevant feedback (Salton and Buckley 1990), the broker could Hamilton, G., and Gattell, R. (1997). JDBC娂: A Java SQL API. Sun
possibly learn how to improve the query performance when a Microsystems Inc., Mountain View, Calif.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 107

Homer, A. (1998). Professional ASP techniques for webmasters. WROX Liang, V. C. (1998). ‘‘Java-based regulation broker.’’ MS thesis, Carne-
Press, Birmingham, U.K. gie-Mellon Univ., Pittsburgh.
Jacobson, I. (1992). Object-oriented software engineering. Addison-Wes- Moss, K. (1998). Java servlets. McGraw-Hill, New York.
ley, Reading, Mass. Salton, G., and Buckley, C. (1990). ‘‘Improving retrieval performance by
Kim, E. E. (1996). CGI developer’s guilde. Sams.net Publishing, Indi- relevance feedback.’’ J. Am. Soc. for Information Sci., 41(4), 288–297.
anapolis. Stasiak, D. M. (1996). ‘‘An environmental regulations broker: Classifi-
Koller, D., and Sahami, M. (1997). ‘‘Hierarchically classifying documents cation and co-word analysis,’’ MS thesis, Carnegie-Mellon Univ., Pitts-
burgh.
using very few words.’’ Proc., 14th Int. Conf. on Machine Learning
Swank, M., and Kittel, D. (1997). Designing and implementing Microsoft
(ICML ’97). Morgan Kaufmann Publishers, Inc., San Mateo, Calif., index server. Sams.net Publishing, Indianapolis.
170–187. Yang, Y. (1994). ‘‘Expert network: Effective and Efficient learning from
Krofchik, M. A. (1995). ‘‘A broker for delivering environmental regula- human decisions in text categorization and retrieval.’’ Proc., 17th Int.
tion over the World Wide Web,’’ MS thesis, Carnegie-Mellon Univ., Conf. on Res. and Devel. in Information Retrieval (SIGIR ’94), Vol.
Pittsburgh. 358, Springer, Berlin, 13–22.

108 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 REDUCING PREDICTION ERROR BY TRANSFORMING INPUT DATA FOR
NEURAL NETWORKS
By Jonathan Jingsheng Shi,1 Member, ASCE

ABSTRACT: The primary purpose of data transformation is to modify the distribution of input variables so that
they can better match outputs. The performance of a neural network is often improved through data transfor-
mations. There are three existing data transformation methods: (1) Linear transformation; (2) statistical stan-
dardization; and (3) mathematical functions. This paper presents another data transformation method using
cumulative distribution functions, simply addressed as distribution transformation. This method can transform a
stream of random data distributed in any range to data points uniformly distributed on [0,1]. Therefore, all neural
input variables can be transformed to the same ground-uniform distributions on [0,1]. The transformation can
also serve the specific need of neural computation that requires all input data to be scaled to the range [⫺1,1]
or [0,1]. The paper applies distribution transformation to two examples. Example 1 fits a cowboy hat surface
because it provides a controlled environment for generating accurate input and output data patterns. The results
show that distribution transformation improves the network performance by 50% over linear transformation.
Example 2 is a real tunneling project, in which distribution transformation has reduced the prediction error by
more than 13% compared with linear transformation.

INTRODUCTION formations is to modify the distribution of the input variables

so that they can better match the distribution of the predicted
There has been a growing interest in neural computing tech- output. Research shows that the performance of a neural model
nique in civil engineering in recent years (Flood and Kartam is often improved by transforming continuous numeric inputs
1994a,b) for its capabilities including the ability to learn and (NeuralWorks 1997). Data transformations have been widely
generalize from examples, to produce meaningful solutions to studied and applied for analyzing, exploring, and transforming
problems even when input data contain errors or are incom- raw data in statistics (Tukey 1977; Carroll and Ruppert 1988).
plete, and to adapt solutions over time to compensate for Logothetis (1988) reported the application of data transfor-
changing situations. The rapid advancement of neural com- mations for off-line control. Chiba and Smith (1994) discussed
putation and computer technology has also contributed to ap- the impact of data transformations on least-squares solutions.
plications of neural networks (NN) in science and engineering. Existing data transformations in NN can be categorized into
Its applications in construction go back to the late 1980s three groups: (1) Linear transformation; (2) statistical stan-
(Flood 1989) but already cover a large range of topics such dardization; and (3) various mathematical functions. Linear
as estimating construction costs and markup estimation (Mo- transformation is a simple and widely used method in many
selhi et al. 1991; Hegazy and Moselhi 1994), predicting con- NN systems (e.g., NeuralShell and Brainmaker), in which a
struction productivity (Chao and Skibniewski 1994; Shi 1999), data set is scaled to interval [0,1] or [⫺1,1] by using the orig-
predicting settlements during tunneling (Shi et al. 1998), and inal data range as a scalar. Statistical standardization computes
predicting the outcome of construction litigation (Arditi et al. the deviation from the mean and divides it by the standard
1998). deviation (Shanker et al. 1996). Mathematical function trans-
Most reported NN applications fall in the category of pre- formation applies a mathematical function (e.g., exponential)
diction, in which there is an unknown relationship between a to a data stream. Linear transformation is actually the simplest
set of input factors and an output. The objective is to find a one in this group.
formula that facilitates predicting the outcome from the input This paper reports a new transformation method by using
factors based on a representative set of historical examples for cumulative distribution functions (CDFs). This method can
this unknown mapping. Due to the well-developed generalized transform a stream of random data distributed in any range to
delta-rule (Rumelhart et al. 1986) and supervised training al- uniformly distributed data points on [0,1]. Therefore, all input
gorithm, back-propagation (BP) networks account for over data can be transformed to the same ground-uniform distri-
90% of the reported engineering applications. A typical BP butions on [0,1]. The transformations can serve two purposes:
network has an input layer, an output layer, and hidden layers. (1) To satisfy the neural computation requirement; and (2) to
A mapping relationship between inputs and outputs is explored improve the performance of NN.
during training (Rumelhart and McClelland 1986). Training
data represent the set of points that are used to fit the param- USE OF DISTRIBUTION FUNCTIONS TO TRANSFORM
eters of the model. Test data measure how well the model INPUT DATA SETS TO UNIFORM DISTRIBUTIONS
interpolates. Hence data are essential for the effectiveness of
a neural net. Because the observed data points of an input variable can
Due to the requirements of neural computation algorithm, be in any pattern and in any range, some neural systems such
raw data are usually normalized to the interval [0,1] or [⫺1,1] as NeuralWorks Predict have developed dozens of transfor-
through data transformations. The primary purpose of trans- mations (e.g., linear, log, exponential, etc.). After transforma-
tions, input data can be scaled to [⫺1,1] or [0,1] but still
1
Assoc. Prof., Dept. of Civ. and Arch. Engrg., Illinois Inst. of Technol., appear in unregulated patterns. Which transformation brings
3201 S. Dearborn St., Chicago, IL 60616. E-mail: [email protected] in the best result can only be determined by contrasting net-
Note. Editor: Sivand Lakmazaheri. Discussion open until September work performance under various transformations.
1, 2000. To extend the closing date one month, a written request must Is it possible to transform all inputs with different distri-
be filed with the ASCE Manager of Journals. The manuscript for this
paper was submitted for review and possible publication on May 24,
butions to a regulated pattern? Will doing so bring in any
1999. This paper is part of the Journal of Computing in Civil Engi- potential benefits to neural performance? The answer to the
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- first question is absolutely positive, and it is the intended work
0109–0116/$8.00 ⫹ $.50 per page. Paper No. 21075. of this paper. Like transformations in traditional statistical
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 109
analyses, it is very hard to prove the effectiveness of a trans- model should result in an R value close to 1 and small values
formation. The reasoning as follows may help to understand for error terms. On the contrary, a poorly trained model has
the benefit of transforming all inputs to the same ground. an R value close to 0 and large error values.
If a stream of data is transformed to [0,1] by linear trans-
formation, the data may concentrate on a small area. For in- Other Measures
stance, 90% of the data are in a small range [0,0.1]. Since
neural computation can only interpolate, it might be unreason- Mean Prediction Error (MPE)
able to expect a continuous mapping for this input variable on MPE is defined as the percentage of the averaged absolute
the entire input range [0,1]. On the other hand, if the high prediction error in proportion to the actual output value; that
frequency of occurrence in the small range is properly consid- is

冘
ered and the entire population is transformed to be uniformly N
distributed on [0,1], a continuous and smooth mapping may 兩Actuali ⫺ Predicti兩
MPE = 100 ⭈ (3)
be expected. Of course, the effectiveness of a transformation i=1 Actuali
can be validated by extensive examples.
where N = total number of data patterns; Actuali and Predicti
= actual and predicted values of pattern i, respectively.
CDF and Probability Distribution Function (PDF)
All NN inputs are independent random variables, and out- Error Difference
puts are dependent random variables. In statistics, the char- Error difference is defined as the difference of the absolute
acteristics of a random variable are defined by a PDF or CDF. prediction error between distribution and linear transforma-
It is interesting to notice that CDF is scaled between 0 and tions; that is
1. Inverse transformation method is widely used in Monte
Carlo simulation to generate random distributions from uni- EDi = 兩DErrori兩 ⫺ 兩LErrori兩 (4)
form distribution on [0,1] (Pritsker et al. 1997). Inverse trans- DErrori = Actuali ⫺ DPredicti (5)
formation uses the fundamental information of the CDF F(X).
The random variable R = F(X) is uniformly distributed on the LErrori = Actuali ⫺ LPredicti (6)
unit interval [0,1]. Thus, to generate a random sample of X, where EDi denotes the error difference between the two trans-
we generate a random number r and solve the equation r = formations for the same data pattern; DErrori and LErrori rep-
F(X) for the corresponding value of x = F ⫺1(r). resent the prediction errors of distribution and linear transfor-
The proof for the uniformality of r distribution is straight- mation, respectively; DPredicti and LPredicti = predicted
forward (Bratley et al. 1987) and is based on the following output values from the networks with distribution and linear
reasoning: Let R = F(X) have distribution G. Then, for 0 ⱕ r transformations, respectively.
ⱕ 1, we have The EDi measures the improvement (EDi < 0) or deterio-
G(r) = P[F(X) ⱕ r] = P[X ⱕ F ⫺1(r)] = F(F ⫺1(r)) = r (1) ration (EDi > 0) of distribution transformation over linear
transformation. The values of EDi on the entire data set can
Thus, R is uniformly distributed on [0,1]. be plotted on a scatter graph to enable the reader to visualize
the effect as detailed in the following examples. The averaged
Distribution Transformation EDi value indicates the overall improvement or deterioration
of distribution transformation over linear transformation

冘
Using the relationship between PDF and CDF, any distri- N
bution in any range can be transformed to a uniform distri- 1
MED = EDi (7)
bution on [0,1]. Therefore, this provides a method that can N i=1
transform all NN input random variables to uniform distribu-
where MED is the mean value of the EDi .
tions on [0,1].
Let X be an input variable. Observations corresponding to Example 1: Cowboy Hat Surface
N training patterns are x1, x2, . . . , xN. Using statistical methods
(Decision 1997), a random distribution can be fitted to best Mathematical functions provide a controlled environment to
match the sample with a CDF F(X). A new random variable generate accurate training and test data. They have been
is defined as widely used as a means for validing the performance of an
NN. The first example attempts to fit a cowboy hat surface as
Y = F(X) (2) shown in Fig. 1 and the corresponding function is given as
Then Y is a uniform distribution on [0,1]. The corresponding follows (Choueiki and Mount-Campbell 1999):
transformed data stream can be calculated from yi = F(xi). z = sin(兹x 21 ⫹ x 22) (8)
The term Y is the transformed input variable, and F(X) is the
transformation function, addressed as distribution transforma- Training Network Models
tion.
A BP NN is chosen for this example with three layers: 2
neurons on the input layer corresponding to two inputs (x1 and
PERFORMANCE MEASURES x2), one hidden layer with 22 neurons, and 1 neuron on the
Distribution transformation will be contrasted with linear output layer.
transformation. To examine its effect on the performance of Both x1 and x2 are limited on [⫺3,3]. To guarantee the qual-
an NN, the following traditional and nontraditional measures ity of training data, 500 data points are independently gener-
are used. ated for x1 and x2, respectively, using Monte Carlo simulation
with @Risk (Decision 1997) from the same triangular distri-
Traditional Measures bution (minimal = ⫺3, mode = ⫺3, and maximal = 3) to create
biased input distributions. The values of Z are then calculated
Traditional measures include (1) Correlation coefficient R; from (8).
(2) mean squared error; (3) mean absolute error; (4) minimal The 500 data patterns are treated as the training set.
absolute error; and (5) maximal absolute error. A well trained NeuralShell is employed for all training purpose with net-per-
110 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 FIG. 1. Cowboy Hat Surface

fect function to avoid overfitting (NeuralShell 1998). Two TABLE 1. Performance Measures for Fitting Cowboy Hat Sur-
training scenarios are designed. Scenario 1 scales the 500 input face (Training Set)
data to the range [⫺1,1] with linear transformation. Since Linear Distribution Improvement
NeuralShell has a built-in linear transformation function, the Item transformation transformation (%)
raw data can be directly presented with specifying linear trans- (1) (2) (3) (4)
formation for input data. Scenario 2 is presented with the data Correlation coefficient R 0.993 0.999 —
transformed using the triangular distribution function as fol- Mean squared error 0.003 0.001 66.7
Mean absolute error 0.043 0.018 58.1
lows:

再
Minimal absolute error 0.000 0.000 —
Maximal absolute error 0.233 0.122 —
(X ⫺ min)2 MPE (%) 18.57 7.89 57.5
, min ⱕ X ⱕ mod
(mod ⫺ min)(max ⫺ min)
F(X) =
(max ⫺ X)2
ues, sorted in ascendant order for the 500 patterns, are plotted
1⫺ , mod ⱕ X ⱕ max
(max ⫺ min)(max ⫺ min) versus the predicted output values from distribution trans-
(9) formation Predict(D) and linear transformation Predict(L) in
Fig. 2.
where min = ⫺3; mod = ⫺3; and max = 3. In Fig. 2, it is evident that the predicted values of distri-
Apply (9) to 500 data points of x1 and x2 separately. Trans- bution transformation are much closer to the actual values than
formed input data can be obtained with uniform distribution that of linear transformation. The error differences between
on [0,1]. They constitute the transformed training set for Sce- distribution and linear transformation are computed from (4)
nario 2 together with the original output values z. and are plotted in Fig. 3. The majority of the dots in Fig. 3
Experiments of the two training scenarios are performed lay under 0, which means the absolute predicted errors from
independently. It should be mentioned that the two scenarios distribution transformation are less for the majority of the 500
may need different learning rates, initial weights, and momen- training patterns. This also concludes the error reduction effect
tum values for the best performance even under the same net- of distribution transformation.
work architecture. Therefore, each training scenario must in-
dependently search for the appropriate values of the three Test Network Performance
parameters to achieve the best performance. The two best per- A test data set of 300 patterns is generated again using
forming models, one from each of the two scenarios, are then Monte Carlo simulation from a uniform distribution on [⫺3,3]
contrasted with the performance measures summarized in Ta- for both x1 and x2 with actual z values calculated using (8). It
ble 1. should be mentioned that the same distribution transformation
Table 1 shows that Scenario 2 has a significant improvement function (9) must be used to transform the test data before
(approximately 60%) over Scenario 1. Linear transformation they can be presented to the trained network. Two correspond-
results in an average prediction error of 18.57%, and distri- ing test scenarios are formed with the contrasting results de-
bution transformation results in an average prediction error of tailed in Table 2.
7.89%—an improvement of 57.7%. The mean absolute error Table 2 indicates that distribution transformation can sig-
also shows an improvement of 58.1%. The actual output val- nificantly reduce prediction errors with an overall improve-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 111
 FIG. 2. Actual Outputs versus Predicted Outputs for Cowboy Function (Training Set)

ment of more than 50% for both the mean absolute error and
MPE. The actual outputs can also be plotted versus the two
sets of predicted outputs corresponding to the two test scenar-
ios.
Fig. 4 also suggests reduced prediction errors resulted from
distribution transformation. The error differences are plotted
in Fig. 5 to show the individual improvement for each pre-
dicting pattern. It is evident that much smaller prediction errors
are generated for the majority of the 300 test patterns.
EXAMPLE 2: A TUNNELING EXAMPLE
Reducing prediction errors has always been the interest of
an NN network application. This example applies the proposed
FIG. 3. Error Differences between Two Transformations for distribution transformation to a real-world construction prob-
Cowboy Hat Surface lem. The Brası́lia Tunnel (6.5-km-long, 9.6 m in diameter) was
excavated in shallow soft ground, as described in detail by
TABLE 2. Performance Measures for Cowboy Hat Surface
Ortigao et al. (1996). The tunneling method had the following
(Training Set) stages: excavation, placement of lattice girders, shotcreting the
primary layer of the lining, closing the invert, and finally plac-
Linear Distribution Improvement ing the final or secondary layer of the lining. In all cases a
Item transformation transformation (%)
clay core was always left to improve the stability of the face.
(1) (2) (3) (4)
Excavation encompassed three construction methods as
Correlation coefficient R 0.909 0.985 —
Mean squared error 0.041 0.008 80.9
listed in Table 3. Method A is the least expensive and is em-
Mean absolute error 0.095 0.046 51.6 ployed in favorable conditions of face stability and negligible
Minimal absolute error 0.000 0.000 — damage to nearby structures. On the other hand, if excavation
Maximal absolute error 1.328 0.522 — conditions deteriorate (such as decreasing thickness of soil
MPE (%) 72.1 34.3 52.4
cover, poor soils are encountered, sensitive nearby structures
exist, or deformation observation indicates that stability may
be decreasing), methods B or C could be selected. An addi-
tional construction method, D, consisting of side drift with
temporary invert, was not included in the analyses, because it
was only used for a very short extension of 20 m in length.
Surface settlement observations took place in 50-m intervals
along the tunnel length at three different excavation stages: at
the face passage, at the closing of the inverted arch, and the
final value after stabilization (Ortigao et al. 1996). In addition,
single settlement marks were deployed at 10-m intervals,
among the fully instrumented stations. The measured stabilized
settlement is plotted in Fig. 6, which appears discontinuous
due to missing or disqualified data points.
Training Network Models
FIG. 4. Actual Outputs versus Predicted Outputs for Cowboy BP neural networks with nine input variables, as listed in
Hat Surface (Test Set) Table 4, were attempted for predicting settlements during tun-
112 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 FIG. 5. Error Differences between Two Transformations for Cowboy Hat Surface (Test Set)

TABLE 3. Construction Methods TABLE 4. NN Input Variables

Distance to close Symbol Description
invert behind (1) (2)
excavation face (X) Dr Length of excavation from drive start
Method Construction method (m) H/A Depth of soil cover above tunnel crown/area of tunnel
(1) (2) (3) section
A Full face 4.8–7.2 X Delay for closing invert
B Heading excavation with temporary 2.4–5.4 WL Water level depth
invert; bench excavated afterwards RA Rate of advance of excavation of face for each tunnel
in 3-m-long steps drive, taken as average value for each tunnel drive
C Side-drift employing sidewall; en- 2.4–5.4 Nc Mean SPT N value at tunnel crown level
largement to final section, with Ns Mean SPT N value at tunnel springline level
demolition of sidewall Ni Mean SPT N value at tunnel inverted arch level
CM Construction method

neling (Shi et al. 1998). The 6.5-km tunnel is divided into

eight excavation stations. A total of 340 valid data patterns uniform distributions on [0,1], we need to fit the data points
were collected: 312 patterns from seven stations (chainage 0– of each input variable with an appropriate distribution. There
6,000 m) are used as the training set; and 28 patterns from are 312 observations for each of the nine input variables. Us-
Station Gal (chainage 6,000–6,500 m) are used as the test set. ing a statistical analysis software [e.g., BestFit (Decision
Linear transformation, which is still defined as Scenario 1, 1997)], each input variable can be best fitted by a distribution.
is the reference for evaluating the effect of distribution trans- The histogram versus the fitted Beta distribution for input var-
formation, which is defined as Scenario 2. The two scenarios iable Dr is plotted in Fig. 7 as an example. The best-fitted
have the same network architecture including three layers: 9 distributions of the nine inputs are summarized in Table 5.
neurons on the input layer, 11 neurons on the hidden layer, Applying the fitted distribution functions to corresponding
and 1 neuron on the output layer. input data, the original data are transformed to uniformly dis-
To transform the data points of the nine input variables to tributed data points on [0,1], which are the training set for

FIG. 6. Brası́lia Tunnel and Stabilized Settlement

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 113

Scenario 2. The performance indicators of the two scenarios TABLE 5. Fitted Distributions of Nine Input Variables
are summarized in Table 6 for the training set. Fitted
Table 6 shows that the two scenarios result in a mean ab- Symbol Minimal Maximal Mean distribution Parameters
solute error of 22.8 and 19.9 mm and a mean prediction error (1) (2) (3) (4) (5) (6)
of 27.6 and 23.8, respectively. In conclusion, distribution Dr 0 440 169 Beta ␣ = 0.773, ␤ = 1.24
transformation can reduced prediction error by approximately H/A 0.584 1.681 1.063 Gamma ␣ = 21.97, ␤ = 0.05
X 3.15 11.63 7.35 Beta ␣ = 2.54, ␤ = 2.584
13%. WL ⫺1.3 12.7 1.57 Gamma ␣ = 1.71, ␤ = 1.68
RA 1.05 2.3 1.64 Beta ␣ = 0.75, ␤ = 0.91
Nc 0.8 8.5 3.1 Gamma ␣ = 5.31, ␤ = 0.58
Ns 3 34 7.88 Weibull ␣ = 1.58, ␤ = 8.48
Ni 6 50 28.44 Gamma ␣ = 0.7, ␤ = 0.67
CM 1 3 1.42 Gamma ␣ = 5.35, ␤ = 0.25

TABLE 6. Performance Measures for Brası́lia Tunnel (Training

Set)
Linear Distribution Improvement
Item transformation transformation (%)
(1) (2) (3) (4)
Correlation coefficient R 0.901 0.931 —
Mean squared error 1,191 854 28.3
Mean absolute error 22.8 19.9 12.7
Minimal absolute error 0.023 0.000 —
Maximal absolute error 175.2 140 —
MPE (%) 27.6 23.8 13.9
FIG. 7. Histogram versus Fitted Beta Distribution (Dr)

FIG. 8. Actual Settlements versus Predicted Values (Training Set)

FIG. 9. Actual Settlements versus Predicted Values (Training Set)

114 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 FIG. 10. Error Differences between Two Transformations for Brası́lia Tunnel (Training Set)

dicted values are plotted in Fig. 11. A closer relationship be-

tween Predict(D) and actual values can be observed. The er-
rors differences plotted in Fig. 12 also confirm reduced
prediction errors from distribution transformation. In sum-
mary, distribution transformation can reduce the prediction er-
ror by 27.3% for the test set.

CONCLUSIONS

This paper reported distribution functions for transforming

unregulated input data sets into uniform distributions on [0,1].
Such transformations can serve two purposes: (1) To meet neu-
ral computation requirement; and (2) to improve NN perfor-
mance. The distribution transformation demonstrates an im-
FIG. 11. Actual Settlements versus Predicted Values (Test provement of 50% over linear transformation for fitting the
Set) cowboy hat surface. Comparatively, it has reduced the predic-
tion error by more than 13% for a real world civil engineer
case—predicting tunneling settlement.
A general reasoning can conclude that transforming input
data sets into uniform distributions should improve NN per-
formance, although it is hard to theoretically prove such a
conclusion. Generally, distribution transformation is superior
to linear transformation. However, the improvement may be
minimal for simple applications whereas linear transformation
can result in high accuracy. When distribution transformation
is used, the goodness of fit of a distribution function to the
data points of an input variable affects the performance. The
goodness of fit can be evaluated by traditional statistic meth-
FIG. 12. Errors Differences between Two Transformations ods. Moreover, it can be judged by whether or not it can trans-
(Test Set) form the data set into a uniform distribution. Our experience
shows that many input data patterns can be fitted by appro-
The actual settlement versus predicted values are plotted priate distribution functions given a wide variety of distribu-
along chainage in Fig. 8. It is evident that Predict(D) values tion functions and their flexibility to fit any shapes.
are closer to the actual values than Predict(L) are. The closer The writer is very interested to see that this transformation
relationship is more evident whereas the three settlement val- method be applied to more engineering applications to further
ues are plotted along data patterns in an ascendant order of test its validity.
the actual values as in Fig. 9. The scatter graph in Fig. 10
details the error differences between the two scenarios. It also
confirms the reduced prediction errors from distribution trans- APPENDIX. REFERENCES
formation. Overall, distribution transformation reduces the av-
erage prediction error by 2.86 mm, an improvement of 13% Arditi, D., et al. (1998). ‘‘Predicting the outcome of construction litigation
over linear transformation. using neural networks.’’ Comp.-Aided Civ. and Infrastructure Engrg.,
13(2), 75–81.
Test Network Performance Bratley, P., Fox, B. L., and Schrage, L. E. (1987). A guide to simulation,
2nd Ed., Springer, New York.
As mentioned previously, the 28 data patterns from Station Carroll, R. J., and Ruppert, D. (1988). Transformation and weighting in
‘‘Pal’’ are used for testing. The actual settlement versus pre- regression. Chapman and Hall, Boca Raton, Fla.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 115

Chao, L.-C., and Skibniewski, M. J. (1994). ‘‘Estimating construction Moselhi, O., Hegazy, T., and Fazio, P. (1991). ‘‘Neural networks as tools
productivity: Neural-network-based approach.’’ J. Comp. in Civ. in construction.’’ J. Constr. Engrg. and Mgmt., ASCE, 117(4), 606–
Engrg., ASCE, 8(2), 234–251. 625.
Chiba, S., and Smith, D. L. (1994). ‘‘Impacts of data transformations on NeuralWorks Predict: Quick start guide. (1997). Aspen Technology, Inc.,
least-squares solutions and their significance in data analysis and eval- Cambridge, Mass.
uation.’’ J. Nuclear Sci. and Technol., 31(8), 770–781. Ortigao, J. A. R., Kochen, R., Farias, M. M., and Assis, A. A. (1996).
Choueiki, M. H., and Mount-Campbell, C. A. (1999). ‘‘Training data ‘‘Tunnelling in Brasilia porous clay.’’ Can. Geotech. J., Ottawa, 33(4),
development with the D-optimality criterion.’’ IEEE Trans. on Neural 565–573.
Networks, 10(1), 56–63. Pritsker, A. A. B., O’Reilly, J. J., and LaVal, D. K. (1997). Simulation
Decision making tools: @Risk, BestFit. (1997). Palisade Corp., New with visual SLAM and AweSim. Wiley, New York.
York. Rumelhart, D. E., and McClelland, J. L. (1986). Parallel distributed pro-
Flood, I. (1989). ‘‘A neural network approach to the sequencing of con-
cessing. MIT Press, Cambridge, Mass.
struction tasks.’’ Proc., 6th Int. Symp. on Automation and Robotics in
Constr., Construction Industry Institute, Austin, Tex, 204–211. Shanker, M., Hu, M. Y., and Hung, M. S. (1996). ‘‘Effect of data stan-
Flood, I., and Kartam, N. (1994a). ‘‘Neural networks in civil engineering. dardization on neural network training.’’ Omega. Int. J. Mgmt. Sci.,
I: Principles and understanding.’’ J. Comp. in Civ. Engrg., ASCE, 8(2), 24(4), 385–397.
131–148. Shi, J. (1999). ‘‘A neural network based system for analysing earth-mov-
Flood, I., and Kartam, N. (1994b). ‘‘Neural networks in civil engineering. ing operations.’’ Constr. Mgmt. and Economics, London, 17(4), 483–
II: Systems and applications.’’ J. Comp. in Civ. Engrg., ASCE, 8(2), 471.
149–162. Shi, J., Ortigao, J. A. R., and Bai, J. (1998). ‘‘Modular neural network
Hegazy, T., and Moselhi, O. (1994). ‘‘Analogy-based solution to markup for predicting settlements during tunneling.’’ J. Geotech. and Geoenvir.
estimation problem.’’ J. Comp. in Civ. Engrg., ASCE, 8(1), 72–87. Engrg., ASCE, 124(5), 389–395.
Logothetis, N. (1988). ‘‘The role of data transformation in Taguchi anal- Tukey, J. W. (1997). Exploratory data analysis. Addison-Wesley, New
ysis.’’ Quality and Reliability Engrg. Int., 4(1), 49–61. York.

116 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 STOCHASTIC CONSTRUCTION TIME-COST TRADE-OFF ANALYSIS
By Chung-Wei Feng,1 Liang Liu,2 Associate Member, ASCE,
and Scott A. Burns,3 Member, ASCE

ABSTRACT: Traditional time-cost trade-off analysis assumes that the time and cost of an option within an
activity are deterministic. However, in reality the time and cost are uncertain. Therefore, in analyzing the time-
cost trade-off problem, uncertainties should be considered when minimizing project duration or cost. Simulation
techniques are useful for analyzing stochastic effects, but a general strategy/algorithm is needed to guide the
analysis to obtain optimal solutions. This paper presents a hybrid approach that combines simulation techniques
and genetic algorithms to solve the time-cost trade-off problem under uncertainty. The results show that genetic
algorithms can be integrated with simulation techniques to provide an efficient and practical means of obtaining
optimal project schedules while assessing the associated risks in terms of time and cost of a construction project.
This new approach provides construction engineers with a new way of analyzing construction time/cost decisions
in a more realistic manner. Historical time/cost data and available options to complete a project can be modeled,
so that construction engineers can identify the best strategies to take to complete the project at minimum time
and cost. Also, what-if scenarios can be explored to decide the desired/optimal time and/or cost in planning and
executing project activities.

INTRODUCTION considered when minimizing project cost and duration. This

leads to the so-called stochastic time-cost trade-off problem.
Time-cost trade-off analysis is one of the most important Although simulation techniques are useful for analyzing such
aspects of construction project planning and control. In gen- problems, a general strategy/algorithm is needed to guide the
eral, there are trade-offs between time and cost to complete analysis to obtain optimal solutions. The following sections
each activity of a project. The usual relationship is that the briefly review the stochastic time-cost trade-off problem, rel-
less expensive the resources used, the longer it takes to com- evant research in the area, and finally the new approach that
plete an activity. For example, using more productive equip- combines simulation with genetic algorithms (GAs).
ment or hiring more workers may save time but would be more
expensive. Fig. 1 presents a typical discrete relationship be- PROBLEM STATEMENT
tween time and cost for an activity. Generally, project planners
strive to find the most cost effective way to complete a project Construction time-cost trade-off problems are viewed as one
within time limits. Because of the time-cost relationship of the most important aspects of construction decision making
among activities, it usually takes several iterations to select the (Feng et al. 1997). Difficulties arise because, for the hundreds
proper methods, equipment, and crew sizes to obtain an ac- of activities of a project, there are various options of com-
ceptable overall project duration within the contractual time pleting these activities using different crew sizes or equipment.
limit. Because not all activities are critical, some activities can This creates the classic combinatorial search problem for con-
be performed more slowly, with less cost, without impacting struction engineers to identify the best selections of crew size
the overall project duration. Schedulers can perform the so- or equipment that produce the minimum cost possible to com-
called time-cost trade-off analysis to minimize the project cost plete the project. Traditionally, the time and cost for the op-
or to minimize the total project duration. A complete analysis tions to complete an activity are assumed to be deterministic
leads to (1) a time-cost trade-off curve showing the relation- in construction time-cost trade-off analysis (Harris 1978), but
ship between total project duration and cost; and (2) identifi- in reality they typically follow a certain kind of probabilistic
cation of the specific construction methods corresponding to distribution, as indicated by historical data. The stochastic na-
each solution along the project time-cost trade-off curve. ture of time and cost adds an additional dimension of com-
A common assumption is that the cost and duration of an plexity to the already hard to solve combinatorial problem.
option within an activity are deterministic. In addition, an op- Although most researchers recognize the fact that time-cost
tion is defined as a method or a mode to execute the activity trade-off problems are stochastic by nature, most working so-
with associated cost and duration. However, in reality they are lutions assume that time and cost for completing an activity
uncertain. The cost and time of the options within an activity are deterministic (De et al. 1995). The following discussions
are probabilistically distributed, as demonstrated in Fig. 2. highlight the limitations of solving construction time-cost
Some options are more risky than others and have a broad
distribution. In addition, it is hard to properly evaluate the
options using deterministic methods if the overlaps between
the options are significant. Therefore, uncertainties should be
1
Asst. Prof., Dept. of Civ. Engrg., Nat. Cheng Kung Univ., No. 1, Ta-
Hseuh Rd., Tainan, Taiwan. E-mail: [email protected]
2
Adjunct Asst. Prof., Dept. of Civ. Engrg., Univ. of Illinois at Urbana-
Champaign, 205 N. Mathews Ave., Urbana, IL 61801.
3
Assoc. Prof., Dept. of General Engrg., Univ. of Illinois at Urbana-
Champaign, 104 S. Mathews Ave., Urbana, IL. E-mail: [email protected]
Note. Editor: Sivand Lakmazaheri. Discussion open until September
1, 2000. To extend the closing date one month, a written request must
be filed with the ASCE Manager of Journals. The manuscript for this
paper was submitted for review and possible publication on October 23,
1998. This paper is part of the Journal of Computing in Civil Engi-
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- FIG. 1. Typical Discrete Relationship between Time and Cost
0117–0126/$8.00 ⫹ $.50 per page. Paper No. 19514. of Activity

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 117

 FIG. 2. Stochastically Distributed Time and Cost of Options

FIG. 3. Options without Significant Overlap

FIG. 4. Options with Significant Overlap

trade-off problems by assuming deterministic time and cost, significant overlap. It may be sufficient to use only the mean
at both the activity and the project levels. values of the duration and cost of these options to find the
optimal choice for each activity in such a situation.
Activity Level On the other hand, if there are significant overlaps between
the distributions, using only the mean values is no longer suf-
ficient. For example, in Fig. 4 we cannot say that the duration
Overlap between Distributions
of Option 1 is definitely shorter than the duration of Option
In the traditional deterministic time-cost trade-off problem, 2, even though this is true of the mean values. Therefore, we
only certain values of the duration and cost of the options need a comparison scheme that considers the whole distribu-
within an activity, usually the mean values, are used to eval- tion in order to evaluate the duration and cost of an option.
uate the options. This may be done with the stochastic time-
cost trade-off problem only if there are no significant overlaps Correlation between Duration and Cost
between the distributions of both duration and cost of the op- Because both the duration and cost of an option are deter-
tions. For example, in Fig. 3 both the duration and cost of the mined by the type of resource being used, it is reasonable to
options are distributed within their specific regions without assume that the duration and cost of an option are correlated.
118 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 FIG. 5. Example of Correlation between Duration and Cost of Activity Option

FIG. 6. Comparison of Two Solutions

Using only the mean values of the duration and cost does not the solution. As shown in Fig. 6, we can see that d1, 35% >
reflect the correlation between the duration and cost of an op- d2, 35% and c1, 35% < c2, 35%, which explains that neither Solution
tion. The duration and cost of an option can be obtained from 1 nor 2 is dominant. Having information describing the entire
historical data and so can their correlation, which may be ob- distribution allows us to assess risk associated with a particular
tained by using regression analysis. Examples of possible cor- solution.
relation between the cost and duration of an option are shown
in Fig. 5. The first graph in this figure shows that the cost and Correlation between Project Duration and Cost
duration of an option to finish an activity are independent of
each other; the second graph shows that they are negatively Because the project duration and cost are produced by the
correlated. same combination of options, it is reasonable to assume that
the project duration and cost are correlated. However, the real
Correlation between Activities values of the project duration and cost are now known until
the project is finished; the same is true of the correlation be-
Some activities within a project use the same resources, tween the project duration and cost. Nevertheless, we can still
such as heavy equipment; therefore, these activities may be estimate the mean values, variance, and correlation of the
affected by each other. In other words, these activities are cor- project duration and cost according to the results from simu-
related to each other. Such correlated relationships are not re- lation.
flected when using only the mean values of the options to
evaluate the project duration and cost. EXISTING TECHNIQUES
Project Level Existing techniques for construction time-cost trade-off
problems can be categorized into three areas: heuristics, math-
Overlap between Distributions of Project Duration and Cost ematical programming, and simulation. The following sections
briefly describe the strengths and weaknesses of the existing
In solving the traditional deterministic time-cost trade-off techniques.
problem, both the project duration and cost are determined by
the length of the critical path and the sum of the costs of the Heuristic Methods
activities, respectively. As a result, the project duration and
cost are evaluated based on certain values, which are usually Heuristic methods are based on rules of thumb that gener-
the mean values. However, such an approach is not sufficient ally lack mathematical rigor. They provide good solutions but
for stochastic time-cost trade-off problems. For example, in do not guarantee optimality. Examples include Fondahl’s
Fig. 6 both the project duration and cost of Solutions 1 and 2 method (1961), Prager’s structural model (1963), Siemens’s
are normally distributed. If we consider only the mean values effective cost slope model (1971), and Moselhi’s structural
of the project duration and cost, we would conclude that So- stiffness method (1993). In resource-constrained scheduling,
lution 1 dominates Solution 2, because Solution 1 finishes the intensive research has been conducted under the consideration
project more quickly and uses less money than Solution 2 of renewable resource constraints in terms of minimizing
does. However, the above comparison becomes insufficient if project duration. Heuristic methods have enjoyed the majority
the whole distribution of the project duration and cost are of the research effort because they can find good solutions with
taken into consideration. For instance, suppose we compare far less computational effort than other methods. Heuristics
two solutions based on 35 percentile values without consid- can be classified into two types: (1) Serial heuristics, in which
ering the correlation between the project duration and cost of processes are first prioritized and retain their values throughout
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 119
the scheduling procedure; and (2) parallel heuristics, in which ration or expenditures (Dobin 1985; Weiss 1986). Few studies
process priorities are updated each time a process is scheduled. have been conducted to optimize the project in a stochastic
Examples are Bell and Han (1991), Kahattab and Choobineh network. Examples are Wan’s perturbation analysis (1994) and
(1991), and Boctor (1993). These heuristics provide good so- Kidd (1987). They provide a good strategy to analyze the re-
lutions; however, the performance is problem-dependent, and sults of simulation; however, the above analyses do not at-
good solutions are not guaranteed. tempt to optimize the entire project time-cost trade-off curve.
Heuristic methods, mathematical models, and simulation
Mathematical Programming Models techniques show both strengths and weaknesses. The heuristic
approaches select the processes to be shortened or expanded
Mathematical programming methods convert time-cost based on certain selection criteria, but they do not guarantee
trade-off problems to mathematical models and utilize linear optimal solutions. On the other hand, mathematical models
programming, integer programming, or dynamic programming require great computational effort, and some approaches do
to solve them. Kelly (1961) formulated time-cost trade-off not provide the optimal solution either. In addition, for large-
problems by assuming linear time-cost relationships within scale networks, neither heuristic methods nor mathematical
processes. Other approaches, such as those by Hendrickson programming models can obtain optimal solutions efficiently.
and Au (1989) and Pagnoni (1990), also use linear program- Simulation techniques provide a good estimate for the optimal
ming to solve the time-cost trade-off problem. Linear program- solutions; however, a guide to analyze the result of the sim-
ming approaches are suitable for problems with linear time- ulation must be provided in order to find the solutions effi-
cost relationships but fail to solve those with discrete time-cost ciently. The state of research suggests the need to develop a
relationships. Meyer and Shaffer (1963) and Patterson and more efficient algorithm to conduct optimal time-cost trade-
Huber (1974) solved time-cost problems including both linear off analysis under uncertainties.
and discrete relationships by using mixed integer program-
ming. However, integer programming requires a prohibitive GAs
amount of computational effort once the number of options to
complete an activity becomes too large or the network be- GAs are search algorithms developed by Holland (1975),
comes too complex. Burns et al. (1996) took a hybrid approach which are based on the mechanics of natural selection and
that used (1) linear programming to find a lower bound of the genetics to search through decision space for optimal solutions
trade-off curve; and (2) integer programming to find the exact (Goldberg 1989). The metaphor underlying GAs is natural se-
solution for any desired duration. Robinson (1975), Elmagraby lection. In evolution, the problem each species faces is to
(1993), and De et al. (1995) used dynamic programming to search for beneficial adaptations to the complicated and chang-
solve time-cost trade-off problems for a special class of net- ing environment. In other words, each species has to change
works that can be decomposed to pure series or parallel sub- its chromosome combination to survive in the living world. In
networks. GAs, a string represents a set of decisions (chromosome com-
bination), a potential solution to a problem. Each string is eval-
Simulation uated on its performance with respect to the fitness function
(objective function). The ones with better performance (fitness
Finding the optimal solutions for construction time-cost op- value) are more likely to survive than the ones with worse
timization in a stochastic network has proven to be a difficult
problem to solve. For example, if the durations of the options
were random variables (continuous or discrete), the evaluation
process would involve excessive numerical integration, which
makes it practically impossible. Simulation techniques have
been used to enhance the study of stochastic project networks;
however, many studies focus only on estimating project du- FIG. 7. Example of GA Chromosome Representation

FIG. 8. Minimal Distance to Segments of Convex Hull

120 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

performance. Then the genetic information is exchanged be- that GA is very powerful in searching for optimal solutions
tween strings by crossover and perturbed by mutation. The of large-scale construction time-cost trade-off problems.
result is a new generation with (usually) better survival abili- However, for stochastic time-cost trade-off problems, a more
ties. This process is repeated until the strings in the new gen- complicated selection scheme must be used. The following
eration are identical, or certain termination conditions are met. sections describe how the deterministic method described
In the past two decades, GAs have been successfully for- above is modified for the stochastic time-cost trade-off prob-
mulated to solve many science and engineering problems and lem.
have proved to be an efficient means of searching for the op-
timal solutions of large-scale optimization problems. In addi- Use of Simulation at Activity Level
tion, GAs are suitable for problems with multiple objectives.
As a result, GAs provide an alternative other than weighting Overlaps between Distributions
for multiobjective problems. Furthermore, because there are Simulation utilizes the whole distribution of duration and
no restrictions of using GAs in terms of uncertainties, GAs cost to represent an option, which solves the problems caused
also serves as a good tool for searching the optimal solutions
under uncertainties.

GA FOR STOCHASTIC TIME-COST TRADE-OFF

PROBLEM OF CONSTRUCTION PROJECTS
Background
The application of GAs to the stochastic time-cost trade-off
problem is an extension of work presented recently, and ap-
plied to deterministic problems (Feng et al. 1997). In this pre-
vious work, each string in the GA population represents a se-
lection of options to complete the project, with associated time
and cost to finish the project. Fig. 7 presents an example of a
string representing a seven-activity network. The value in each
position within the string indicates which option is selected
for the activity. The fitness function is computed by measuring
the distance from the individual member of the population to
the convex hull separating the entire population from the cost
and duration axes. The smaller this distance, the more fit is
the individual (Fig. 8). Therefore, the solutions closer to the
convex hull reproduce more frequently than the solutions fur-
ther away from the convex hull. Consequently, there will be
a natural tendency for the new population to move toward the
convex hull, as well as toward the optimal trade-off curve
because there are no individuals between the convex hull and
the trade-off curve. The use of GA and convex hull shows

FIG. 11. Flowchart of GA

TABLE 1. Option of Test Example

Option Duration Mean SD Cost Mean SD

(1) (2) (3) (4) (5) (6) (7)
1 n 14 2 n 23,000 4,600
1 n 20 3 n 18,000 3,000
FIG. 9. Example of Correlation between Activity Duration and 1 n 24 5 n 12,000 2,000
Cost 2 n 15 3 n 3,000 100
2 n 18 5 n 2,400 400
2 n 20 5 n 1,800 60
2 n 30 7 n 1,200 360
2 n 60 6 n 600 120
3 n 15 5 n 4,500 750
3 n 22 2 n 4,000 533
3 n 33 5 n 3,200 853
4 n 12 2 n 45,000 1,500
4 n 16 4 n 35,000 10,500
4 n 20 3 n 30,000 7,000
5 n 22 6 n 20,000 6,000
5 n 24 4 n 17,500 1,750
5 n 28 5 n 15,000 1,500
5 n 30 5 n 10,000 1,000
6 n 14 2 n 40,000 4,000
6 n 18 3 n 32,000 2,133
6 n 24 4 n 18,000 2,400
7 n 9 2 n 30,000 2,000
7 n 15 2 n 24,000 7,200
7 n 18 3 n 22,000 4,400
FIG. 10. Convex Hull Formed by Mean Values of Solutions

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 121

by overlaps between the distributions. Because the duration of the duration and cost is bivariate normal. The conditional
and cost of an option are randomly generated by the simulation mean duration of a randomly generated cost Cr is
scheme, they are no longer represented by fixed values, and
␮Cr/Cr = ␮d ⫹ ␳(␴d /␴c)(Cr ⫺ ␮c) (1)
the whole distributions are accounted for.
and the conditional variance is
Correlation between Duration and Cost
␴2Dr/Cr = ␴d (1 ⫺ ␳2) (2)
The problem of correlation between the duration and cost
of an option can also be resolved by adding the correlation if the variance is assumed to be constant. Therefore, the as-
into the simulation scheme. That is, after randomly generating sociated duration of Cr is
a sample value of cost Cr, the associated duration Dr is deter-
Dr = ␮Dr/Cr ⫹ z ⭈ ␴Dr/Cr (3)
mined from the regression analysis based on existing historical
data (Fig. 9). Consider an option in which the joint probability where z is determined by another random number from [0, 1].

FIG. 12. Network Configuration of Test Example

FIG. 13. GA/Simulation Results of Test Example

122 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 TABLE 2. Optimal Solutions for Test Example Correlation between Activities
.
Option
Mean The interdependencies between activities can also be solved
duration Mean cost 1 2 3 4 5 6 7 by adding the correlation between activities in the simulation.
61.64722 158170.2 1 1 1 1 1 2 1 For example, if Activities A and B are dependent on each
61.66295 148170.2 1 1 1 2 1 2 1 other, the duration and cost of both activities will be simulated
62.36143 144162.5 1 1 1 1 1 3 1 with due regard to this correlation. The degree of correlation
63.68453 134162.5 1 1 1 2 1 3 1 among activities can be derived from historical project data,
64.88648 131662.5 1 1 1 2 2 3 1 if available, or from the assessment of a project planner from
66.35098 129169.2 1 1 1 3 1 3 1 his or her past experience.
66.64976 128533.7 1 2 1 3 1 3 1 Note that the distributions of the overall project duration
66.93237 126669.2 1 1 1 3 2 3 1
and cost can be obtained from the simulation. These distri-
67.36245 126033.7 1 2 1 3 2 3 1
68.21735 125433.7 1 3 1 3 2 3 1 butions serve as the genetic information for each individual
69.3065 124902 1 3 2 3 2 3 1 and are used to evaluate the performance of the individual, as
69.74045 124153.6 1 1 1 3 3 3 1 described in the next section.
70.2066 123518 1 2 1 3 3 3 1
70.4742 119169.2 1 1 1 3 4 3 1 Computing Fitness at Project Level
71.16872 118533.7 1 2 1 3 4 3 1
71.83073 118153.5 1 1 1 3 4 3 1 The convex hull is defined by the mean values of the project
72.58713 117933.7 1 3 1 3 4 3 1 durations and costs of the solutions in the population, as
74.28801 117402 1 3 2 3 4 3 1 shown in Fig. 10. The fitness value of an individual solution
76.03536 117041.8 1 1 1 3 3 3 2 is defined as the solution’s average minimal distance to the
76.50153 116406.3 1 2 1 3 3 3 2 convex hull. That is, all of the simulation results are used to
76.76911 112057.4 1 1 1 3 4 3 2
evaluate the distance to the convex hull. The goal of this ap-
77.46365 111421.9 1 2 1 3 4 3 2
78.88206 110821.9 1 3 1 3 4 3 2
proach is to determine which solutions are closer to the convex
79.76911 110057.4 1 1 1 3 4 3 3 hull and, consequently, which ones are better in a probabilistic
80.46365 109421.9 1 2 1 3 4 3 3 sense. The ellipses in Fig. 10 are used to graphically represent
80.59936 108712.3 3 1 1 3 4 3 1 the simulation results corresponding to a solution. Having
81.29388 108076.8 3 2 1 3 4 3 1 knowledge of these complete distributions allows the project
82.7123 107476.8 3 3 1 3 4 3 1 planner to assess risks associated with various solutions. The
83.58878 106965 2 2 1 3 4 3 2 flowchart in Fig. 11 summarizes the GA presented in this pa-
84.41318 106945.1 3 3 2 3 4 3 1 per. In addition, the algorithms for finding the pareto front and
85.00721 106365 2 3 1 3 4 3 2 the convex hull can be found in Feng et al. (1997).
85.89428 105600.6 2 1 1 3 4 3 3
86.58879 104965 2 2 1 3 4 3 3
86.89428 101600.6 3 1 1 3 4 3 2 TABLE 3. Best Choice for Each Option of Test Example
87.58879 100965 3 2 1 3 4 3 2
89.00722 100365 3 3 1 3 4 3 2 Probability of Option Being Optimal for Each Activity
89.89426 99600.57 3 1 1 3 4 3 3
Option 1 2 3 4 5 6 7
90.58879 98965.03 3 2 1 3 4 3 3
92.00722 98365.03 3 3 1 3 4 3 3 1 0.56 0.37 0.81 0.05 0.14 0.00 0.49
93.70808 97833.37 3 3 2 3 4 3 3 2 0.09 0.28 0.12 0.07 0.12 0.05 0.23
101.6719 97233.37 3 4 2 3 4 3 3 3 0.35 0.28 0.07 0.88 0.09 0.95 0.28
104.6375 97033.37 3 3 3 3 4 3 3 4 0.00 0.05 0.00 0.00 0.65 0.00 0.00
104.6528 96433.37 3 4 3 3 4 3 3 5 0.00 0.02 0.00 0.00 0.00 0.00 0.00
131.6719 95833.36 3 5 3 3 4 3 3 Best options 1 1 1 3 4 3 1

FIG. 14. PDF of Project Duration

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 123

 FIG. 15. CDF of Project Duration

FIG. 16. PDF of Project Cost

FIG. 17. CDF of Project Cost

124 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

EXAMPLE cost trade-off curve, as presented in Table 2. Conclusions can
be drawn regarding the best option for each activity using
Fig. 12 shows a test network configuration. Table 1 is the probabilistic analysis, as shown in Table 3. It can be seen that
option data of this test example. In Table 1, n indicates that Option 3 for both Activities 4 and 6 are the dominating options
the option cost and duration is normally distributed. Other dis- and can be selected as the best options without regard to proj-
tributions can be used, as provided in the simulation software ect cost or duration. Selection of the remaining options re-
being used. Mean and standard deviation values are provided quires that the project planner considers time and/or cost con-
for each of the normally distributed options in Table 1. This straints, as well as the level of acceptable risk. Having access
example is tested on a Pentium 133 PC with 48MB RAM to the entire distribution of simulation results for each solution
using a C-coded computer program to implement the GA writ- on the time-cost trade-off curve gives the project planner
ten by the first writer. The results of running the GA/simula- plenty of information to make these decisions. Some of the
tion method are shown in Fig. 13 at various stages of com- ways that this information can be represented are presented in
pletion. The initial population is scattered widely over the the next section.
solution space. There is no significant convergence after first
20 iterations. After 40 iterations, a significant convergence to- PROBABILISTIC ANALYSES
ward the optimal time-cost trade-off curve has occurred, and
a stable solution has been found. The series of white dots After finding the optimal solutions, several probabilistic
through the mid-region of the simulation results are the mean analyses can be performed according to the project planner’s
values of the optimal solutions. The optimal solutions are those needs. For example, the project planner may want to know the
with the better fitness values and constitute the optimal time- distributions of the project duration and project cost for a cer-

FIG. 18. Mean ⫾1 Standard Deviation of Project Duration versus Mean of Project Cost

FIG. 19. Mean of Project Duration versus Mean ⫾1 Standard Deviation of Project Cost

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 125

tain solution. The project planner can use simulation to find ACKNOWLEDGMENT
this. A simulation is performed by using the suggested ‘‘best This material is based on work supported by the National Science
options’’ presented in Table 3. Figs. 14 and 15 show the prob- Foundation under Award No. DDM-8957410.
ability density function (PDF) and cumulative distribution
function (CDF) of the project duration, respectively. We can APPENDIX. REFERENCES
estimate the project duration for this example by using the Bell, C. E., and Han, J. (1991). ‘‘A new heuristic solution method of
mean value, which is 71.83. In addition, the project planner resource constrained project scheduling problems.’’ Naval Res. Logis-
can look into the CDF of the project duration to find out the tics, 38, 315–331.
probability that the project can be finished within a certain Boctor, F. F. (1993). ‘‘Heuristic for scheduling projects with resource
restrictions and several resource-duration modes.’’ Int. J. Production
desired duration. For example, if the desired duration is 75 Res., 31, 2547–2558.
days, the probability would be around 78%. Figs. 16 and 17 Burns, S., Liu, L., and Feng, C. (1996). ‘‘The LP/IP hybrid method for
show the PDF and CDF of the project cost, respectively, for construction time-cost trade-off analysis.’’ Constr. Mgmt. and Econom-
the same example. ics, 14, 265–276.
In addition, the project planner may want to know how the De, P., Dunne, E. J., and Wells, C. E. (1995). ‘‘The discrete time-cost
overall project duration versus project cost is distributed. For tradeoff problem revisited.’’ Eur. J. Operational Res., 81, 225–238.
Dobin, B. N. (1985). ‘‘Bounding the project completion time distribution
example, the project planner can use the mean ⫾1 standard in PERT networks.’’ Operation Res. 33, 862–881.
deviation for both project duration and cost to find a desired Elmagraby, S. E. (1993). ‘‘Resource allocation via dynamic programming
region for probabilistic analyses. Fig. 18 shows the mean ⫾1 in activity networks.’’ Eur. J. Operational Res., 64, 199–215.
standard deviation of project duration versus the mean of Feng, C.-W., Liu, L., and Burns, S. A. (1997). ‘‘Using genetic algorithms
project cost for the test example. If the project planner is more to solve construction time-cost trade-off problems.’’ J. Comp. Civ.
Engrg., ASCE, 11(3), 184–189.
concerned with project cost, he or she may choose a solution Fondahl, J. W. (1961). ‘‘A non-computer approach to the critical path
with a desired project cost and see how the associated project method for the construction industry.’’ Tech. Rep. No. 9, Constr. Inst.,
duration varies. Fig. 19 shows the mean of project duration Dept. of Civ. Engrg., Stanford University, Stanford, Calif.
versus the mean ⫾1 standard deviation of project cost for the Goldberg, D. E. (1989). Genetic algorithms in search, optimization and
test example. Again, if the project planner is concerned more machine learning. Addison-Wesley, Reading, Mass.
with project duration, he or she may choose a solution with a Harris, R. B. (1978). Precedence and arrow networking techniques for
construction. Wiley, New York.
desired project duration and see how the associated project Henderickson, C., and Au, T. (1989). Project management for construc-
cost varies. It may be desirable to pick a solution with a nar- tion. Prentice-Hall, Englewood Cliffs, N.J.
row distribution. Holland, J. H. (1975). ‘‘Adaptation in natural and artificial systems.’’
University of Michigan Press, Ann Arbor, Mich.
Kahattab, M., and Choobineh, F. (1991). ‘‘A new approach for project
CONCLUSIONS scheduling with limited resource.’’ Int. J. Production Res., 29, 185–198.
Kelly, J. E., Jr. (1961). ‘‘Critical path planning and scheduling: Mathe-
matical basis.’’ Operations Res., 9(3), 167–179.
This research utilizes GAs along with simulation techniques Kidd, J. B. (1987). ‘‘A comparison between the VERT program and other
to imitate the probabilistic nature of project networks through- method of project duration estimation.’’ Omega—Int. J. Mgmt. Sci.,
out the search of optimal solutions. This approach provides 15(2), 129–134.
more realistic solutions for construction time-cost trade-off Meyer, W. L., and Shaffer, L. R. (1963). ‘‘Extensions of the critical path
problem under uncertainty. It also demonstrates that GAs can method through the application of integer programming.’’ Civ. Engrg.
be integrated with simulation techniques to provide an efficient Constr. Res. Ser. 2, University of Illinois, Urbana.
Moselhi, O. (1993). ‘‘Schedule compression using the direct stiffness
and practical means of assessing project time and cost risks. method.’’ Can. J. Civ. Engrg., Ottawa, 20, 65–72.
Although the traditional approaches for time-cost trade-off Pagnoni, A. (1990). Project engineering: computer oriented planning and
analysis, such as heuristic and mathematical programming, operational decision making. Spring, Berlin and Heidelberg, Germany.
have their roles, the future seems to favor combined GAs and Patterson, J. H., and Huber, D. (1974). ‘‘A horizon-varying, zero-one
simulation, particularly for time-cost trade-off analysis of con- approach to project scheduling.’’ Mgmt. Sci., 20(6), 990–998.
Prager, W. (1963). ‘‘A structural method of computing project cost poly-
struction projects under uncertainty. This new approach pro-
gons.’’ Mgmt. Sci., 9(3), 394–404.
vides construction engineers with a new way of analyzing con- Robinson, D. R. (1975). ‘‘A dynamic programming solution to cost-time
struction time/cost decisions in a more realistic manner. tradeoff for CPM.’’ Mgmt. Sci., 22(2), 158–166.
Historical time/cost data and available options to complete a Siemens, N. (1971). ‘‘A simple CPM time-cost tradeoff algorithm.’’
project can be modeled, so that construction engineers can Mgmt. Sci., 17(6), B-354–B-363.
identify best strategies to take to complete the project at min- Wan, Y. (1994). ‘‘Resource allocation for a stochastic CPM-type network
through perturbation analysis.’’ Eur. J. Operational Res., 79, 239–248.
imum time and cost. Also, what-if scenarios can be explored Weiss, G. (1986). ‘‘Stochastic bounds on distributions of optimal value
to decide the desired/optimal time and/or cost in planning and functions with applications to PERT, network flows and reliability.’’
executing project activities. Operations Res., 34(4), 595–605.

126 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 COMPUTER GRAPHICS FOR NONLINEAR STRUT-TIE MODEL APPROACH
By Young Mook Yun1

ABSTRACT: This paper presents an interactive computer graphics program for implementing the nonlinear
strut-tie model approach for the practical design and analysis of disturbed regions in structural concrete. The
graphics program consists of three parts. The first and second parts are considered as the pre- and post-processors
for a nonlinear finite-element analysis of 2D plain concrete and plane truss structures. The results of the first
part provide guidance in the development of strut-tie models and in the evaluation of effective strength levels
for concrete struts. The third part is a design routine for dimensioning and detailing the components of the
developed strut-tie model. This graphics program implementing the nonlinear strut-tie model approach has been
shown to be efficient in the development, analysis, and detailing of strut-tie models for the analysis and design
of structural concrete including disturbed regions.

INTRODUCTION preparation time for the finite-element analysis required in

complex analysis and design situations. A graphics program
The nonlinear strut-tie model approach has proved to be a
would also provide a very convenient tool for possible revi-
useful and consistent method for the analysis and design of
sions of the chosen strut-tie model. A clear visualization of the
reinforced and prestressed concrete members including dis-
flow of internal forces would allow the engineer to be more
turbed regions. Strut-tie models are discrete representations of
effective in providing a correctly detailed concrete structure.
actual stress fields resulting from applied load and support
Accordingly, the use of an interactive computer graphics pro-
conditions. These models represent the load-carrying mecha-
gram in the analysis and design of structural concrete in com-
nisms of structural members by approximating the flow of in-
plex situations is considered to be essential.
ternal forces by means of struts representing the flow of com-
In this paper, an interactive computer graphics program im-
pressive stresses and ties representing the flow of tensile
plementing the nonlinear strut-tie model approach is presented.
stresses. Strut-tie models are a generalization of the truss anal-
The basic principles and algorithms implemented in the pro-
ogy first proposed by Ritter (1899) and Morsch (1909) in the
gram are also outlined. The design of a squat shear wall with
early 1900s. The original truss analogy concept assumes that
openings is conducted with the aid of the proposed graphics
concrete after cracking is not capable of resisting tension and
program.
postulates that a cracked reinforced concrete beam acts as a
truss with parallel longitudinal chords and a web composed of NONLINEAR STRUT-TIE MODEL APPROACH
diagonal concrete struts and transverse ties. Although Morsch
recognized the possibility that the inclination angles of the Developing a strut-tie model using the nonlinear strut-tie
diagonals could vary from 45⬚, he proposed the use of a 45⬚ model approach proposed by Yun (1999) is an iterative pro-
truss for design simplicity. This method was later modified by cess, as shown in Fig. 1. The process starts with the selection
Leonhardt (1965) and others until Thurlimann et al. (1975), of the initial centerline truss model, followed by an evaluation
Muller (1976), and Marti (1985) furnished a scientific basis of the member forces and the dimensioning of the components
by tracing the concept back to the theory of plasticity. of the model based on the internal member forces and strength
Further work by Schlaich et al. (1987) extended the beam of the concrete and steel. Next, the joints of the truss model,
truss model with the use of uniformly inclined diagonals, or nodal zones in a strut-tie model, are formed and analyzed.
thereby enabling application to all parts of the structure in the If the dimensions of the struts and ties are not compatible with
form of generalized strut-tie systems. For the practical appli- the geometrical constraints of the structure, or a failure mech-
cation of strut-tie models in structural concrete, a lot of work anism occurs in a nodal zone, the truss model itself and/or its
has been done on the effective strengths of concrete struts and geometry must be modified and the procedure repeated until
strength of concrete in nodal zones by Ramirez and Breen a satisfactory solution is obtained.
(1983), Collins et al. (1985), MacGregor (1988), Schlaich and Initial strut-tie models can be formulated from experimental
Anagnostou (1990), Bergmeister et al. (1991), Jirsa et al. observations using crack failure patterns, recorded strains in
(1991), Alshegeir (1992), Adebar and Zhou (1993), and Yun the concrete and reinforcements, and actual specimen detail-
and Ramirez (1996). ing, as well as loading and support conditions. In straightfor-
Though the nonlinear strut-tie model approach has been ward situations, an experienced engineer is generally capable
proved to be useful for the analysis and design of disturbed of developing strut-tie models based on engineering common
regions in structural concrete, it has limitations in practical sense and knowledge of the behavior of structural concrete.
analysis and design applications since it requires a lot of time However, in more complex cases, this practical knowledge is
and effort. Selecting a final strut-tie model is an iterative pro- often not enough to develop adequate strut-tie models. In such
cess, and there is seldom a unique strut-tie model associated cases, the principal compressive stress trajectories are used to
with a particular analysis or design situation. An interactive select the orientation of the model struts. Thereafter, the strut-
computer graphics program that could implement the nonlinear tie system is completed by replacing the ties so as to furnish
strut-tie model approach would considerably reduce the data a stable load-carrying structure. The interactive computer
graphics program NL-STM (discussed in the next section) is
1
Asst. Prof., Dept. of Civ. Engrg., Kyungpook Nat. Univ., Taegu thus suggested as suitable for this approach.
702-701, Korea. In NL-STM, the principal stresses and their trajectories can
Note. Editor: Sivand Lakmazaheri. Discussion open until September be evaluated by either linear or nonlinear finite-element anal-
1, 2000. To extend the closing date one month, a written request must yses. In the present study, the finite-element nonlinear analysis
be filed with the ASCE Manager of Journals. The manuscript for this
paper was submitted for review and possible publication on December 1,
formulation by Yun (1993), which includes the substitution of
1998. This paper is part of the Journal of Computing in Civil Engi- the expressions of updated material properties extracted from
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- a curve-fitting of an equivalent uniaxial stress-strain relation-
0127–0133/$8.00 ⫹ $.50 per page. Paper No. 19770. ship of 2D unreinforced concrete, is employed because it fa-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 127
 FIG. 2. Algorithm for Determination of Strut and Tie Areas

the nodal zones that connect the elements of the strut-tie model
FIG. 1. Flow Chart for Nonlinear Strut-Tie Model Approach and redirect the internal forces are dimensioned, and the bear-
ing capacities of these zones are then verified by employing
the finite-element nonlinear analysis used for evaluating the
cilitates the accurate consideration of loading and boundary principal stresses and their trajectories.
conditions and permits the updating of the material properties
and geometry of a structure. Although a finite-element nonlin-
ear analysis represents an improvement over linear elastic ap- ANALYSIS AND DESIGN TOOL NL-STM
proaches, it is not intended to model the actual structure for
final conditions. The finite-element nonlinear analysis of 2D NL-STM is a special-purpose computer graphics program
unreinforced concrete at this stage is only intended to help in of about 20,000 lines of Fortran code that is designed for use
the development of a more realistic strut-tie model. with UNIX operating system and HOOPS graphics package
Once a centerline strut-tie model for a given situation is (HOOPS 1997). It was developed as a tool for studying the
developed, a preliminary finite-element analysis of the model behavior of reinforced and prestressed concrete members, as
is conducted to determine the internal forces in its individual well as dimensioning those members. NL-STM consists of
members. After the internal forces are determined, the strut three parts. The first part functions as a pre- and postprocessor
and tie areas are calculated using a linear or nonlinear ap- to the nonlinear finite-element program PLANE for solving
proach that requires a few iterations within the effective stress plane solid problems. The preprocessor is a simple finite-ele-
level limits. The algorithm for calculating strut and tie cross- ment mesh generator that enables the user to draw a finite-
sectional areas using a linear approach is shown in Fig. 2. In element model on the screen. The computer then extracts and
order to evaluate behavior and strength and obtain accurate sends the required input data to PLANE. Thereafter, the post-
strut and tie forces in the analysis and design of the structural processor receives the output data (principal stresses, principal
concrete, a finite-element nonlinear analysis is again conducted angles, and principal stress ratios) from PLANE, which it then
on the strut-tie model itself. In this nonlinear analysis, the displays on the screen. In some cases, the first part can be
actual dimensions of the concrete ties positioned with steel ties bypassed and a strut-tie model developed directly based on the
and steel struts positioned with concrete struts are included in user’s experience. The second part enables the user to develop
the model. In addition, every concrete strut is considered to strut-tie models interactively, based on the results of PLANE.
have its own stress-strain relationship with a peak stress (ef- This part is also considered as a pre- and postprocessor, yet
fective stress) and corresponding peak strain; accordingly, the this time to the nonlinear finite-element program TRUSS for
tangent modulus of elasticity of every strut under an incre- solving plane truss structures. In this part, the computer ex-
mental external load can be used. The stress-strain relationship tracts and sends the required data to TRUSS. The postpro-
of a strut is approximated by the curve-fitting equation (Saenz cessor receives the output data (member forces, member
1964) that describes the stress-strain behavior of concrete un- stresses, and strains) from TRUSS, which it then displays on
der uniaxial compression. The stress-strain relationship of the screen. The third part enables the user to dimension the
every steel tie is assumed to be bilinear, which considers the struts, ties, and nodes. The strut design is based on its internal
strain-hardening of steel. The tangent modulus of elasticity of force and concrete compressive strength, which is a function
a steel tie before and after yielding is obtained using the slopes of the multiaxial state of stress and disturbance of the cracks
of two straight lines. When the stresses of a strut and tie reach and reinforcement. The nodal zone is checked by examining
their peak and ultimate stress, respectively, the stiffness of the its bearing capacity. The node has to be capable of carrying
strut and tie is assumed to be very small, such that these mem- the applied stresses resulting from the struts and ties. The ap-
bers cannot carry any additional forces. plication of the graphics program will be illustrated through
As the last step in the nonlinear strut-tie model approach, the design example outlined in the following section.
128 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
COMPUTER-AIDED DESIGN OF SQUAT SHEAR WALL model was dimensioned and precisely detailed, based on the
internal forces in the struts and ties, material properties of the
A three-story squat shear wall with openings (Paulay and concrete and steel, and current design specifications.
Priestley 1992), as shown in Fig. 3, was designed using the The objective of the preliminary finite-element analysis is
nonlinear strut-tie model approach with the aid of NL-STM. to determine the direction of the compressive principal stresses
The procedure is presented below in three steps. First, the and principal stress ratio of each finite element. The finite-
direction of the compressive principal stresses was determined element types available in NL-STM are three-, four-, and nine-
using a finite-element nonlinear analysis of plain concrete. node isoparametric elements, which give the user more flexi-
Second, a strut-tie model was developed interactively using bility in modeling any type of plane solid problem. Data
the finite-element analysis results. Finally, the developed preparation for a finite-element analysis is a frustrating and
time-consuming task; however, NL-STM provides a finite-
element mesh generator that eliminates these difficulties and
indicates whether the input is right or wrong.
The finite-element model is shown on the screen by a set
of straight lines. The lines are drawn to scale using a mouse.
Thereafter, the element connectivity is automatically calcu-
lated. Since the wall has openings, the finite elements in the
corresponding regions of the openings are deleted, and the
element connectivity is then recalculated accordingly (Fig. 4).
Specifying the load and boundary conditions is the last step
in the preparation of input data for the finite-element analysis
of a plane solid structure. These conditions are defined in the
global coordinate system. The current loading option only in-
cludes concentrated loads at predefined nodes. If external loads
are applied with history, a nonlinear finite-element analysis is
required. Boundary conditions are specified using four types
of supports: hinge, horizontal roller, vertical roller, and roller

FIG. 3. Geometry of Squat Shear Wall with Openings

FIG. 4. Finite-Element Mesh of Shear Wall after Calculating FIG. 5. Display of Finite-Element Analysis Results of Shear
Element Connectivity Wall: (a) Model I; (b) Model II

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 129

with any slope angle. The insertion of external loads, either logue window. Fig. 6 shows the developed model along with
with or without history, and supports is performed interactively the magnitude of the forces, stresses, and strains in each mem-
through a dialogue window that is developed upon activating ber. In order to highlight the stress states of the struts and ties,
the loading and boundary options, respectively. the members were redrawn in different styles. The solid,
After completing the preliminary finite-element analysis, the dashed, and dotted lines indicate tension, compression, and
postprocessor manipulates the stress analysis results (magni- zero force members, respectively. The forces in each truss
tude of principal stresses, principal stress ratios, and principal member are listed in Table 1. These forces will be used to
angles) and displays them on the screen symbolically and nu- dimension the struts and ties.
merically, as shown in Fig. 5. Since the wall does not respond
symmetrically to a reversal of the horizontal forces, another TABLE 1. Steel Tie Forces in Strut-Tie Model for Shear Wall
finite-element model with the horizontal forces in the opposite
Member Force Member Force
direction is also considered. The length of the straight line at
number (kN) number (kN)
the center of each finite element, which represents the mag- (1) (2) (3) (4)
nitude of the minimum principal stress, depends on the ratio
between the stress magnitude at the element and the magnitude (a) Model I
of the maximum stress at some point in the structure. The units 5, 14 446.6 36 307.4
for external loads and principal stresses are determined in kips 9, 18, 28 243.3 37, 38 238.9
19 276.7 45 169.0
and pounds per square inch, respectively. If external loads are
33 139.2 — —
applied with history, subscreens for displaying the results are
automatically produced. (b) Model II
To develop a strut-tie model, the principal compressive 1, 10, 20 139.2 23, 25, 27, 29 398.1
stress trajectories are used to select the orientation of the struts. 3, 12, 22 331.8 32 169.9
After sketching the strut-tie model using a mouse, the truss 7, 16 618.7 33 75.62
9, 18, 28 13.79 42 185.0
geometric parameters (total number of nodes, total number of
elements, and element connectivity) are automatically calcu-
lated by selecting the ‘‘element no.’’ option. Loading and
boundary conditions are entered interactively through a dia-

FIG. 7. Dimensioned Strut-Tie Model for Shear Wall: (a) Model

FIG. 6. Strut-Tie Models for Shear Wall: (a) Model I; (b) Model II I; (b) Model II

130 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 The last step is the replacement of the model compression TABLE 3. Cross-Sectional Areas of Struts and Ties in Strut-
and tension members with dimensioned concrete struts and Tie Model for Shear Wall
concrete and steel ties, respectively. Using suitable values of Member Area Member Area
fc and ft for the effective strengths of the strut and tie, respec- number (cm2) number (cm2)
tively, and assuming that the reinforcement yields prior to the (1) (2) (3) (4)
crushing of the concrete, the required strut and tie sizes, Astrut (a) Model I
and Atie, respectively, can be determined by
1 105.1 36 10.3
C ⱕ Astrut fc (1) 5 13.9 37 8.0
7 1181.8 38 8.0
T ⱕ Atie ft (2) 9 9.08 39 1153.6
10 170.4 40 377.6
where C and T are the compression and tension forces, re- 14 13.9 41 666.3
18 9.1 42 12.3
spectively. If the dimensions of the struts and ties are not con- 19 9.2 43 429.5
sistent with actual dimensions of the structure, the model 20 114.3 44 448.3
should be modified. This procedure must be repeated until the 28 9.1 45 5.6
sizes of the struts and ties do not essentially change after the 29 318.6 46 63.7
last iteration and they fit inside the boundary of the structure. 30 695.1 48 284.2
The dimensioned geometry of the strut-tie models is shown in 31 685.5 49 103.2
32 101.0 50 407.2
Fig. 7. The effective strengths of concrete struts and the com- 33 4.6 51 1074.5
puted cross-sectional areas of struts and ties are listed in Tables 34 697.6 52 602.3
2 and 3, respectively. The cross-sectional areas of struts and 35 1046.9 53 500.9
ties that are not listed in Table 3 are very small. (b) Model II
Struts and ties correspond to 1D stress fields, whose inten-
1 4.6 32 5.7
sity should not exceed the uniaxial compressive strength of the 3 11.6 33 2.5
concrete and yield stresses of the steel, respectively. The in- 5 444.6 34 1016.2
tersections of three or more struts or ties form nodal zones 7 20.1 35 697.6
(the nodes of the truss). Since several stress fields meet in 9 0.46 36 13.3
nodal areas, a 2D stress field is developed for such plane struc- 10 4.6 37 312.9
tures, and the bearing capacity of these nodal zones must be 12 11.6 38 488.5
16 20.1 40 89.3
verified. In this study, the nodal zone is considered as a plane 18 0.46 41 82.0
stress problem whose boundaries are determined by the inter- 19 167.2 42 6.2
section of the stress fields framing the node. Steel ties are 20 4.6 43 391.3
considered as struts acting on the node in the opposite direc- 21 318.6 44 430.1
tion to the ties. They are also incorporated as forces/units of 22 11.6 45 40.9
length when developed through the node. 23 13.3 46 10.2
25 13.3 47 134.5
The bearing capacity of a nodal zone is verified using a 27 13.3 50 167.5
28 0.46 51 334.1
TABLE 2. Effective Strengths of Concrete Struts in Strut-Tie 29 13.3 52 1480.1
Model for Shear Wall 30 1153.6 53 907.3
31 312.8 — —
Member number fc /f ⬘c Member number fc /f c⬘
(1) (2) (3) (4)
(a) Model I finite-element nonlinear analysis with approximate failure cri-
1 1.00 31 0.17 teria that incorporate the different states of stress. Once the
3 0.56 32 0.92 geometry of the node is determined, a finite-element mesh is
4 0.39 35 0.10
laid out, and the axial forces in the individual struts are divided
6 0.12 39 0.10
7 0.10 40 0.17 into their components and applied as concentrated loads at the
10 0.62 41 0.10 nodes of the finite elements bordering each strut, as shown in
12 0.56 42 0.52 Fig. 8. The finite-element nodal forces in Fig. 8 were calcu-
13 0.32 44 0.14 lated as follows:
16 0.13 46 0.10
20 0.92 48 0.25
23 0.14 49 0.93
25 0.28 50 0.22
26 0.30 51 0.16
27 0.68 52 0.41
29 0.10 53 0.25
30 0.15 — —
(b) Model II
5 0.10 38 0.24
6 0.10 40 0.75
8 0.10 41 0.80
17 0.10 43 0.16
21 0.10 45 0.31
26 0.10 46 0.63
30 0.10 47 0.53
31 0.33 50 0.20
34 0.10 51 0.20
36 0.93 52 0.24
37 0.29 53 0.20
FIG. 8. Model of Nodal Zone

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 131

 Pi =
P
L 冋
xi⫹1 ⫺ xi
2
⫹
2 册
xi ⫺ xi⫺1
(3)
strength of the concrete. In the analysis, the tangent modulus
of elasticity of the concrete in the critical principal direction
was set to be very small when the maximum or minimum
principal stress of a finite element with a certain principal
where 1 ⱕ i ⱕ n; x0 = x1; and xn = xn⫹1. The nodal forces pi stress ratio reached the crushing or cracking stress. If the fail-
are also decomposed into horizontal and vertical components. ure mechanism in the nodal zone is developed, the strut-tie
The failure criteria employed in this study were determined model needs to be modified and reanalyzed.
from the biaxial stress failure envelope. The failure criteria In this example, the nodal zone formed by the members 3,
limit the effective concrete strength in the nodal zone to 1.25 50, and 52 in Fig. 7(a) was analyzed using a finite-element
f ⬘c when the nodal zone is fully compressed with a principal nonlinear analysis with 20 incremental loading steps. The
stress ratio of about 0.5, where f ⬘c is the uniaxial compressive width of the support-bearing plate was determined by consid-
ering a stress of 0.85 f ⬘c in the plate. The finite-element model
of the nodal zone with the total accumulated incremental nodal
forces is shown in Fig. 9. In this model, all the hinge boundary
conditions were imposed along the support-bearing plate. The
tie force of member 3 was distributed along the hinge supports
(not shown in the figure). The compressive principal stress
flow in the nodal zone was close to the stress flow of the same
region in Fig. 5. A maximum compressive stress level of 27.4
MPa (1.09 f ⬘c ) was calculated at the bottom left corner element
of the nodal zone, and no failure mechanism was expected,
even though the corner element was crushed.
For the reinforcement details of the wall, the required
amounts of steel bars with a yield strength of 275.8 MPa were
determined using the maximum member forces of the orthog-
onal net of steel ties from the two strut-tie models. The number
of steel bars was then calculated and placed within the effec-
tive widths of the steel ties. At the locations of the members
of the orthogonal net that were not replaced by steel reinforce-
ment bars, D10 steel bars were placed in two layers in order
to satisfy the requirements of maximum reinforcement spacing
and the minimum reinforcement ratio in a horizontal and ver-
tical direction (‘‘Building’’ 1995). Fig. 10 shows the reinforce-
ment details of the wall. In the detailing, the vertical reinforce-
ment bars surrounding all the horizontal reinforcement bars
were bent at the top of the wall without any discontinuities,
and all the horizontal reinforcement bars were provided with
standard hooks to enable the horizontal member forces to be
transmitted to the concrete. To prevent any possibility of pre-

FIG. 9. Analysis of Critical Nodal Zone in Strut-Tie Model for

Shear Wall: (a) Finite-Element Modeling; (b) Compressive Prin-
cipal Stress Flows FIG. 10. Reinforcement Details of Shear Wall

132 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

mature buckling due to plastic hinge formation in the bound- notes on CSA standard A23.3-M84.’’ Concrete design handbook, Ca-
ary regions adjacent to the openings, the R6 (6 mm in diam- nadian Portland Cement Association, Ottawa.
HOOPS Graphics System. (1997). Tech Soft America, 1301 Marina Vil-
eter) ties around the bars were provided with 10 cm centerline lage Parkway, Alameda, Calif.
spacing. The spacing and amount of tie reinforcements in the Jirsa, J. O., Breen, J. E., Bergmeister, K., Barton, D., Anderson, R., and
boundary regions were determined based on the criteria given Bouadi, H. (1991). ‘‘Experimental studies of nodes in strut-and-tie
in the design code NZS 3101 (‘‘New Zealand’’ 1982). models.’’ Rep., IABSE Colloquium, Structural Concrete, Stuttgart, Ger-
many, 525–532.
SUMMARY AND CONCLUSIONS Leonhardt, F. (1965). ‘‘Reducing the shear reinforcement in reinforced
concrete beams and slabs.’’ Mag. Concrete Res., 17.
This paper presents an interactive computer graphics pro- MacGregor, J. G. (1988). Reinforced concrete—Mechanics and design.
Prentice-Hall, Englewood Cliffs, N.J.
gram for implementing the nonlinear strut-tie model approach Marti, P. (1985). ‘‘Basic tools of reinforced concrete beam design.’’ J.
for the practical design and analysis of disturbed regions in Am. Concrete Inst., 82(1), 46–56.
structural concrete. The design of a squat shear wall with Morsch, E. (1909). Concrete-steel construction. E. P. Goodrich, Trans-
openings was conducted with the aid of the graphics program. lator, McGraw-Hill, New York.
The interactive computer graphics program considerably re- Muller, P. (1976). ‘‘Failure mechanism for reinforced concrete beams in
duces the data preparation for the finite-element analysis, in torsion and bending.’’ Publ. 36-II, International Association for Bridge
and Structural Engineering, Zurich, Switzerland, 147–163.
addition to providing a very convenient iteration tool for pos- ‘‘New Zealand Standard Code of Practice for the Design of Concrete
sible revisions of a selected finite element of a 2D plain con- Structures.’’ (1982). NZS 3101, Standard Association of New Zealand,
crete structure and a strut-tie model. The displaying of the Wellington, New Zealand.
results of the finite-element analysis on screen was found to Paulay, T., and Priestley, M. J. N. (1992). Seismic design of reinforced
be very helpful in developing strut-tie models, understanding concrete and masonry buildings. Wiley, New York.
the function of concrete and steel reinforcement, and discern- Ramirez, J. A., and Breen, J. E. (1983). ‘‘Proposed design procedure for
shear and torsion in reinforced and prestressed concrete.’’ Rep., Ctr. for
ing load transfer mechanisms. Transp. Res., University of Texas, Austin, Tex.
The design result showed that the nonlinear strut-tie model Ritter, W. (1899). ‘‘Die Bauweise Hennebique [Hennebique’s construction
approach combined with the graphics program produces sim- method].’’ Schweizerische Bauzeitung, Zurich, Switzerland, 17, 41–43,
ple and effective solutions through providing accurate details 49–52, and 59–61.
on structural concrete. Accordingly, the use of an interactive Saenz, L. P. (1964). ‘‘Equation for the stress-strain curve of concrete.’’
Discussion of the paper by P. Desayi and S. Kirshnan, J. Am. Concrete
computer graphics program in the design of structural concrete Inst., 61, 1229–1235.
in complex situations is considered to be essential. Schlaich, J., Schaefer, K., and Jennewein, M. (1987). ‘‘Towards a con-
sistent design of structural concrete.’’ J. Prestressed Concrete Inst.,
APPENDIX. REFERENCES 32(3), 74–150.
Schlaich, M., and Anagnostou, G. (1990). ‘‘Stress fields for nodes of strut-
Adebar, P., and Zhou, Z. (1993). ‘‘Bearing strength of compressive struts and-tie models.’’ J. Struct. Engrg., ASCE, 116(1), 13–23.
confined by plain concrete.’’ ACI Struct. J., 90(5), 534–541. Thurlimann, B., Grob, J., and Luchinger, P. (1975). Torsion, Biegung und
Alshegeir, A. (1992). ‘‘Analysis and design of disturbed regions with Schub in Stahlbetontragern [Torsion, flexure and shear in reinforced
strut-tie models,’’ PhD thesis, School of Civ. Engrg., Purdue University, concrete girders]. Institute of Structural Engineering, ETH, Zurich,
West Lafayette, Ind. Switzerland.
Bergmeister, K., Breen, J. E., and Jirsa, J. O. (1991). ‘‘Dimensioning of Yun, Y. M., and Ramirez, J. A. (1996). ‘‘Strength of struts and nodes in
the nodes and development of reinforcement.’’ Rep., IABSE Collo- strut-tie model.’’ J. Struct. Engrg., ASCE, 112(1), 20–29.
quium, Structural Concrete, Stuttgart, Germany, 551–564. Yun, Y. M. (1993). ‘‘An explicit finite element formulation for very large
‘‘Building code requirements for structural concrete (ACI 318-95) and deformations based on updated material reference frame.’’ Int. J. Finite
commentary (ACI 318R-95).’’ (1995). American Concrete Institute, Elements in Anal. and Des., 13(2), 209–224.
Farmington Hills, Mich. Yun, Y. M. (1999). ‘‘Nonlinear strut-tie model approach for structural
Collins, M. P., Mitchell, D., and MacGregor, J. G. (1985). ‘‘Explanatory concrete.’’ ACI Struct. J. (in press).

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 133

 COMPUTER IMAGERY AND VISUALIZATION IN CIVIL
ENGINEERING EDUCATION
By N. Bouchlaghem,1 W. Sher,2 and N. Beacham3

ABSTRACT: Higher education institutions in the United Kingdom have invested significantly in the implemen-
tation of communication and information technology in teaching, learning, and assessment of civil and
building engineering—with mixed results. This paper focuses on the use of digital imagery and visualization
materials to improve student understanding. It describes ways in which these materials are being used in the
civil and building engineering curriculum, and, in particular, how distributed performance support systems
(DPSS) can be applied to make more effective use of digital imagery and visualization material. This
paper centers on the extent to which DPSS can be used in a civil and building vocational and continuing
professional development context by tutors in the form of an electronic course delivery tool and by students in
the form of an open-access student information system. This paper then describes how a DPSS approach to
education is being adopted at Loughborough University as part of the CAL-Visual project. After highlighting
the main aims and objectives of the project and describing the system, this paper discusses some of the issues
encountered during the design and implementation of a DPSS and presents some preliminary results from initial
trials.

INTRODUCTION to reduce the need to physically visit construction sites. Al-

though the system is still at an early stage of development,
In the construction sector, students learning about construc-
Finkelstein (1998) outlined some of the unexpected outcomes
tion technology need to visualize materials and components
stemming from evaluating its features with students, architects,
and to have an appreciation of the total constructional form of
and nonprofessionals. These include its ability to be applied
a building. The traditional lecturing environment cannot re-
to other disciplines and as a methodology for digital ware-
place the benefits to be gained from access to construction sites
housing.
and completed buildings. However, such traditional ap-
Many CBL packages tend to be complete and autonomous
proaches are characterized by problems of cost, safety, avail-
ability of and access to a variety of suitable construction sites, pieces of course material containing clear learning objectives
and time. Such is the importance of using images to support and used to support or supplant lectures. In contrast, multi-
and enhance students’ understanding of building design and media teaching aids (MTAs) are small, discrete tools used to
construction that this issue has been recognized by accrediting help tutors convey complex material within lectures, such as
bodies as an important part of the learning process. The Char- the cycle for a self-climbing formwork system described by
tered Institute of Building also actively supports initiatives that Riley and Pace (1997).
promote such methods. Riley and Pace (1997) advocated the use of MTAs by ar-
To address the problems highlighted above, many educa- guing that ‘‘once they are developed, they have the potential
tional organizations have turned to communication and infor- to decrease the preparation time and classroom time needed to
mation technology, and, in particular, multimedia, computer- effectively convey course material to students.’’ Furthermore,
based learning (CBL) and computer-assisted assessment these teaching aids can be used to form the foundations of
(CAA) to introduce some of the experiences gained from con- CBL and CAA packages. In doing so, using teaching aids in
struction sites into the classroom. Examples of the successful this way can also reduce the time and cost of developing CBL
use of visual material (in the form of multimedia), to better and CAA packages.
educate civil engineering students, have been presented by Unlike CBL and CAA packages, which tend to dictate how
Aminmansour (1996), Echeverry (1996), Chinowsky (1997), a course will be implemented and delivered, MTAs do not
Riley and Pace (1997), and Finkelstein (1998). contain rigid learning objectives. Teaching aids not only allow
Although most methods adopt a combination of the different tutors to embed them into their course without having to mod-
types of digital visual material to reinforce the textual infor- ify the course’s structure but also to use conventional computer
mation, Finkelstein (1998) described the development and use tools and techniques familiar to tutors to create and deliver
of a simulated site visit system containing a wide variety of them. For example, many tutors create their presentations us-
visual materials. They include virtual reality (VR) models, ing Microsoft PowerPoint. PowerPoint is particularly effective
QuickTime VR, movies, photographs, CAD drawings, audio at creating teaching aids, because it is easy to embed various
sequences, video sequences, documents, and links of other In- forms of digital visual material. Furthermore, because the use
ternet sites. The system was designed to teach students about of MTAs is an evolutionary approach to traditional teaching
the process of constructing a building through multimedia and and learning, resistance to change is reduced.
Barker (1996a) also investigated the development and de-
1
Lect., Dept. of Civ. and Build. Engrg., Loughborough Univ., Lough- livery of MTAs using an ‘‘electronic course delivery’’ ap-
borough, Leicestershire, LE11 3TU, U.K. proach to present course material. A fundamental difference in
2
Sr. Lect., Dept. of Civ. and Build. Engrg., Loughborough Univ.,
Loughborough, Leicestershire, LE11 3TU, U.K.
this case is that Barker (1996a) described a methodology for
3
Res. Assoc., Dept. of Civ. and Build. Engrg., Loughborough Univ., creating, organizing, and delivering these MTAs from two per-
Loughborough, Leicestershire, LE11 3TU, U.K. spectives: a staff perspective and a student perspective. An
Note. Editor: Sivand Lakmazaheri. Discussion open until September important part of this methodology has been the development
1, 2000. To extend the closing date one month, a written request must of a distributed Performance Support system (DPSS) to facil-
be filed with the ASCE Manager of Journals. The manuscript for this itate creation, organization, and delivery of MTAs. DPSS and
paper was submitted for review and possible publication on March 10,
1999. This paper is part of the Journal of Computing in Civil Engi- their ability to facilitate the use of visual material in MTAs by
neering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/0002- staff and students are discussed in detail in the following sec-
0134–0140/$8.00 ⫹ $.50 per page. Paper No. 20538. tion.
134 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
WHAT IS A DPSS? port tools for use by students and staff in an academic setting
within a university environment.
In their day-to-day activities, civil engineering tutors and The attempts of Barker and Beacham (1999) and Barker et
students are often required to perform a wide range of different al. (1994) at building a DPSS involved using a computer net-
tasks. These vary quite considerably in their complexity. The work system (the Internet) to access different forms of subject
successful execution of many of the more complex tasks often expertise that resided at different geographical locations. This
requires the use of some sort of tool, aiding facility, or per- work has recently been extended and enhanced through the
formance support system (PSS). The basic rationale underlying use of ‘‘mobile computing’’ environments (Barker 1996b).
the use of a PSS is that individuals or groups of people (work- From the perspective of continuing professional development
ing together on a common project) are provided with appro- (CPD), this has particular importance in providing on-the-job
priate tools and techniques to support the tasks they have to support to CPD graduates on site. In the remaining part of this
perform. According to Barker and Hudson (1998), four major section we describe how a DPSS has been used in a vocational
objectives of a PSS are (1) to achieve increases in productivity; and CPD setting.
(2) to improve the overall quality of task/job execution; (3) to
improve the overall quality of the environments in which peo- DPSS Facilities to Support Teaching and Learning
ple work; and (4) from a human perspective, to reduce the
complexity of the processes involved in executing a task. An important application of DPSS is their use for the in-
Simple examples of performance support tools include a tegration of computer-based teaching and learning resources
typewriter, telephone, bicycle, and automobile. Increasingly, within education and training environments (Gery 1991; Ring
computer-based resources are being used to implement PSS. 1994; Collis and Verwijs 1995; Ruggles et al. 1995).
An electronic performance support system (EPSS) is therefore Because of a DPSS’s future importance as a tool for facil-
a facility that uses various types of computer technology to itating ‘‘learning-by-doing,’’ and sharing teaching and learning
realize each of the four previously listed objectives within a material, the remainder of this paper describes and discusses
given working environment or problem-solving domain. Soft- the evolution and future potential utility of DPSS within civil
ware packages such as word-processing systems and spread- and building engineering. The design and construction of a
sheets provide a good example of computer-based PSS. multiple-site, single-server distributed performance support fa-
Naturally, a fundamental requirement of an EPSS is that it cility as part of the CAL-Visual project is described, and an
should increase users’ on-the-job performance within a given outline is given of some of the ways in which we are intending
task domain. This usually involves improving skill levels, re- to use it. An attempt is currently being made to measure the
ducing task complexity, and/or providing appropriate training. potential utility of the system from the perspective of sup-
These requirements can be achieved in two basic ways. First, porting its end-users’ use of MTAs, CBL, and CAA packages.
through the provision of ‘‘automation aids’’; and, second, by
providing various mechanisms to support ‘‘on-the-job,’’ ‘‘just- DPSS WITHIN CAL-VISUAL
in-time’’ (JIT) training enabling users of any given system to
‘‘learn as they do.’’ CAL-Visual is a research project being undertaken by a con-
Many of the early developments in EPSS have been de- sortium of universities within the United Kingdom to inves-
scribed by Gery (1991), McGraw (1994), Varnadoe and Barron tigate how experiences on a construction site can be brought
(1994), Raybould (1995), and Banerji (1995). Currently, there into the classroom using computer-based imagery and visual-
are two important directions of development for EPSS facili- ization materials. The project is part of the Teaching and
ties: (1) The creation of integrated toolsets that are able to Learning Technology Programme initiative funded by the
meet the performance support needs of particular application Higher Education Funding Council for England.
domains (Barker and Hudson 1998; Hudson 1998); and (2) a The CAL-Visual project’s main aim is to implement the use
growing requirement to facilitate group working at a distance of computer imagery to support the teaching, learning, and
through the incorporation of network technology and the prin- assessment of subject areas where visualization of objects and
ciples of distributed computing environments (Beacham 1998). processes plays an important role. Images are usually readily
Indeed, the advent of relatively low-cost network technologies available in slide or paper form and used in an uncoordinated
and the widespread availability of facilities such as the Internet manner. This project is developing a digital-based framework
and the World Wide Web has meant that PSS can now become for an efficient use of existing and new imagery to aid teaching
highly distributed in nature; that is, different components can and learning technology, hence offering time and cost benefits
reside (or be replicated) at, and be accessed from, different to the process.
geographical locations. The rationale for using a DPSS approach within CAL-Vi-
As far as this paper is concerned, we therefore regard a sual is threefold. First, the aims of the project highlight the
DPSS as an integrated and globally accessible collection of need for an integrated repository of digital, visual material that
electronic aids, tools, and data that can be used (as and when is cross referenced and can be used to embed visual resources
required) at particular points of need (within a workgroup or within external teaching and learning aids—similar to the
an organization) to improve human performance within a ECD approach described above. Second, the framework spec-
given task domain. ified to accommodate existing visual materials, tools, learning
resources, and pedagogic methods needs to provide access to
resources as and when they are required. Third, the overall
Educational DPSS system needs to be accessible, portable, inexpensive, platform-
independent, and able to facilitate sharing of visual resources,
Some early work in developing DPSS facilities has been on-the-job support and JIT training.
described by Banerji (1995) and Barker (1995). Banerji
showed how the performance of students using a file transfer DESCRIPTION OF CAL-VISUAL SYSTEM
program (Kermit) could be improved through the provision of
PC-based, on-line help and interactive aiding and training fa- Since electing to develop a DPSS, we have created the
cilities. Similarly, Barker (1995) has shown how the emerging framework for organizing digital, visual material, as shown in
principles of performance support can be used to apply EPSS Fig. 1. This repository of visual material has formed the foun-
techniques to the design and development of integrated sup- dations upon which to create a searchable resource (in the form
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 135
 FIG. 2. System’s Architecture

first is in the form of an Internet hybrid search engine, and

the second is in the form of a VR environment. Both of these
pathways are accessible through a standard Web browser. Each
of these components is described in detail below.

FIG. 1. Framework for Indexing Imagery and Visualisation Ma- Database Design
terial
One of the project’s objectives is to produce a set of CD-
of a Web-based database), three integrated CBL modules, a ROMs that contains platform-independent visual information.
CAA module, and the development of a suite of electronic The rationale for publishing the database in a Web-based form
lectures containing MTAs. Furthermore, it is proposed to in- on CD-ROM was to produce a product that is maintenance
clude on-the-job support and JIT training for tutors who re- free and able to be transported to a server-based configuration
quire assistance embedding such visual material within their if required. This flexibility will also enable the database to be
electronic lectures. A detailed description of the design and adapted to the needs of end-users and institutions.
implementation of the framework, database, and additional One option considered was to develop the database using a
system components is given in this section. proprietary software tool, such as Microsoft Access. An ad-
vantage of Microsoft Access is that it has a feature that pub-
lishes information in the form of a Web site. This Web site
Detailed Framework Description
can then be accessed on a stand-alone computer and on a Win-
The framework development began by investigating a num- dows 95 Personal Web server or on an NT server.
ber of classifications systems for the indexing of visual ma- Another option was to develop a Windows-based database
terial. It was then decided to use a keyword system based on using a bespoke software tool. Such a tool can be developed
the CI/Sfb system (Construction Index/Samarbetskommittén using a language, such as C⫹⫹. The main advantage is that
för Byggnadsfrâgor; a standard classification system for in- the database could be developed so that it is portable to a
dexing technical and product information in the construction number of platforms. However, the software tool used to ac-
industry used in Europe and the United Kingdom), augmented cess the database would need to be re-engineered and/or re-
with user-defined keywords to form a generic index. Each set compiled to work on each platform. As the system will mainly
of keywords associated with a visual object is to be augmented be used on one type of platform, for example Windows, this
with administrative information to form the metadata. A num- solution would have been suitable. It does, however, require a
ber of indexes have also been created using the visual material large amount of development work to produce a working pro-
and metadata that arrange the visual material into a number of totype.
key subjects. A third option that could meet the platform-independence
This framework is also used as the foundation for the design requirement is using Internet technologies. While this tech-
and implementation of a Web-based database that is to store nology has many advantages over traditional systems, it often
the visual material and its meta-data. Once the metadata and uses a server configuration and common gateway interface
visual material is entered into a Web-based database, the da- scripts to process the information accessible from a database.
tabase can be automatically published along with its indexes. This would therefore be inappropriate if the database is to be
To publish the database as a Web site, on the Internet, intranet, published on a CD-ROM.
or CD-ROM, each image will be stored in an HTML page The actual solution has involved using a combination of the
template. Once the database is published as a Web site, staff options given above. A Microsoft Access database has been
and students can use a number of DPSS components to access created in order to facilitate the entering of metadata and pub-
the visual material through an integrated user interface. lishing the information in the form of a Web site. A hybrid
retrieval system has then been developed using Web-based
System’s Components technologies. This is due to the issue of retrieving this Web-
based information, regardless of whether it exists on a CD-
The system contains a number of integrated components ROM or a server.
(Fig. 2) to provide staff and students with different pathways
in which to access the visual material. Starting with the un- Hybrid Retrieval System
derlying Web-based database, a retrieval system is the main
method of accessing visual material. In doing so, the retrieval Although there are many propriety software tools and lan-
system provides two methods of browsing the database. The guages available for developing Internet and CD-ROM prod-
136 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 FIG. 3. CAL-Visual Retrieval System

ucts, only a few lend themselves to the creation of hybrid text- will be created in the form of a VR environment that uses the
retrieval search engines. Many of the tools (such as WebCD) retrieval system as a central facility for accessing the image
are themselves platform-dependent, difficult to learn and use, materials.
and expensive. Alternatively, the languages are nonstandard
and difficult to learn. For this reason, two prototypes have VR Environment
been developed that use two of the more suitable languages:
Java and JavaScript. This is because Java and JavaScript are Although the textual user interface is considered to be the
based on a common programming language; they are platform- main access point for searching for visual material, the other
independent and are freely available. Although there was a form of access will be through a number of VR models. Each
small number of compatibility problems, when using different model will form a navigational aid and will be developed as
browsers to run Java applets and JavaScript, the latest versions an alternative interface for the retrieval of images from the
of the main two browsers (Internet Explorer and Netscape database. These models will represent a type of building
Communicator) are now compatible. through which the user can navigate and access the informa-
A hybrid retrieval system solution was created for use on a tion in the database (Fig. 4). An initial prototype has been
CD-ROM and a Web server using Java and JavaScript (Fig. developed using QuickTime VR consisting of a model of a
3). Because the components are created and embedded within low-rise building with hot spots allowing the user to zoom
an HTML page, the amount of programming required is re- into particular areas of the building (Fig. 5). This will be used
duced. In the initial phase of the project, a number of retrieval as an interface to access information in the database. For ex-
system prototypes were developed using each of these Web- ample, information on roof construction or defects can be ac-
based approaches to test their feasibility. cessed by clicking on the appropriate hot spot.
Browser compatibility was the main reason for not using
ActiveX. Although ActiveX is widely used to embed objects
into Internet Explorer, it is not commonly used within Net-
scape browsers. This is because using ActiveX within a Net-
scape browser requires the installation of an ActiveX plug-in.
Furthermore, a different plug-in is needed for each version of
the Netscape browser. Unlike ActiveX, both Java and Java-
Script do not require plug-ins.
Overall, the Java-based, retrieval system performed the best
and as a result showed the most potential. The next stage of
the project is to develop a simple textual user interface in the
form of an HTML form to provide staff and students with easy
access to visual material. In addition, a graphical user interface FIG. 4. VR Environment

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 137

 alongside the input fields and used by end-users as a reference
aid. Whenever a keyword is used to identify part of an image
that is not in the index, it is also added to the list.
As well as the keywords field, each image will have five
other fields: image ID, title, date entered, author/owner, cop-
yright details, and process/description. Each of these additional
fields will be used for the purpose of managing and publishing
the resources. They will also be displayed as a complementary
aid when end-users are searching for specific materials.

IMPLEMENTATION STRATEGY

The CAL-Visual system has been designed to support tutors

in the development and delivery of their teaching and assess-
ment material and to support students in their self-centered
learning giving easy and quick access to visual materials.
Therefore the implementation strategy was designed to take
into account both types of users.
The implementation strategy identified areas where visual
material is essential to students’ understanding and where time
and cost savings can be made compared to traditional tech-
FIG. 5. QuickTime VR Model Showing Hot Spot niques. In building production for example site visits are es-
sential to expose students to site processes and techniques. The
Marking Up Images CAL-Visual system can bring this experience to the classroom
using images and videos visualizing processes through time,
Before any image can be entered into the database and ac- which can be costly and time consuming when using tradi-
cessed through the retrieval system, it must be described in a tional means (site visits). Similar benefits can be achieved in
textual form (otherwise known as marking up). This process the area of building defects where examples of buildings can
involves entering data into a number of fields on a Microsoft be shown to aid students’ practical understanding of the sub-
Access form (Fig. 6). This will allow the system to search the ject.
database for particular images based upon specific criteria. These two areas (building production and building defects)
A comprehensive list of keywords relating to the CI SfB have been selected for the implementation and initial evalua-
classification (mentioned above) is used to describe each im- tion of the CAL-Visual system. The implementation took the
age. This list of keywords is also displayed in the Access form following form:

FIG. 6. Screen Display Showing Microsoft Access Form for Marking Up of Images in Database

138 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 FIG. 7. Screen Display Showing Building Construction CBL Module

• Use of the database by tutors to develop electronic lec- vironment. The implementation of additional sets of key-
tures by searching through the database for visual material words for other disciplines could provide an enhancement
and embedding it into PowerPoint presentations. to the system.
• Use of the database to develop CBL modules. As an ex- • The classification system used to mark up visual material
ample a building construction prototype was developed was seen as limiting by some of the users, and as a result
using images from the database. This prototype links vi- an additional field has been added to the database to allow
sual material to a flowchart of construction activities to for the input of user-defined keywords.
visualize specific activities within the construction se- • Some technical aspects that could limit the use of the
quence (Fig. 7). system had to be considered. For example, viewing video
• Visual material for a specific subject was made available clips required a suitable World Wide Web browser con-
to students for access in their own time. This visual ma- taining a video plug-in or player, and the browser used to
terial was linked to textual information and was made access the system needs to support HTML frames and be
available to students through the Internet. Java compliant.
• Finally, there can be limitations on the amount of visual
INITIAL SYSTEM EVALUATION AND RESULTS material to be stored that are imposed by hardware spec-
ifications.
Initial system evaluation and students’ trials have been con-
ducted in order to establish the system’s usability and its ben- Students’ Evaluation
efits to the process of teaching and learning. This was done Initial trials have been conducted with students using the
through workshops involving tutors and evaluation sessions building construction prototype (Fig. 7) as well as some MTAs
with students. developed by tutors using images and videos from the system.
These trials showed improvements in the following areas
System’s Evaluation (based on observations by tutors and students’ feedback):
• Better participation from students during the lecture ses-
• The performance of the system depended on where its sions
different components are located. The system worked fas- • Better students’ motivation during the tutorial sessions
ter on a CD-ROM and an intranet as compared to being • Less time required to complete tutorial exercises
accessed over the Internet. • Availability of visual material for access through the In-
• Some interface improvements were introduced such as the ternet well received by students
use of thumbnails, scalable images, and better feedback • Better retention of information by students
messages during the navigation through the system. • Easier assimilation particularly for those students with
• The thesaurus currently used in the system as it stands is limited practical experience
only suitable for visual material related to the built en-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 139
 A more detailed evaluation involving a wider audience and Barker, P. G. (1995). ‘‘Interface design for learning.’’ Proc., Comp. Based
more prototypes will be undertaken in the final phase of the Learning in Sci. Int. Conf.
Barker, P. G. (1996a). ‘‘Living books and dynamic electronic libraries.’’
project. Electronic Library, Oxford, U.K., 14(6), 491–502.
Barker, P. G. (1996b). ‘‘Towards real information on demand.’’ Proc.,
CONCLUSIONS ONLINE ’96 Int. Conf., Learning Information (Europe) Limited, U.K.,
261–270.
Communication and information technology play an impor- Barker, P. G., and Beacham, N. A. (1999). ‘‘Making a case for distributed
tant role in the higher education sector encouraged by the in- performance support.’’ Proc., ED-Media ’99, Association for the Ad-
creasing availability and performance of computer technology vancement of Computing in Education, 311–316.
that facilitates the embedding of emergent technologies into Barker, P. G., and Hudson, S. R. G. (1998). ‘‘An evolving model for
multimedia performance support systems.’’ Trends in communication:
teaching and learning. In areas such as the built environment Part 4. Interface technology—Enhancing the quality of life, H. van
where visualization is important to convey concepts and pro- Oostendorp and A. G. Arnold, eds., Boom Publishers, The Netherlands,
cesses to students, the benefits offered by multimedia have 60–73.
been recognized and are being used to improve the process of Barker, P. G., Richards, S., and Banerji, A. (1994). ‘‘Intelligent ap-
teaching and learning. proaches to performance support, ALT-J.’’ J. Assn. of Learning Tech-
This paper discussed the potential of using DPSS as tools nol., Wales, U.K., 2(1), 63–69.
Beacham, N. A. (1998). ‘‘Distributed performance support systems,’’ PhD
for facilitating the development and organization of MTAs thesis, Interactive Systems Research Group, Human-Comp. Interaction
and, in particular, for delivering them using electronic course Lab., School of Computing and Mathematics, University of Teesside,
delivery and electronic open-access student information ser- Middlesbrough, U.K.
vice. Chinowsky, P. S. (1997). ‘‘Introducing multimedia cases into construction
Based upon the principles of DPSS and their implementa- education.’’ Proc., 4th Congr. held in conjunction with A/E/C Systems
tion into mainstream education, this paper discussed the ra- ’97, Computing in Civ. Engrg., Teresa M. Adams, ed., ASCE, New
York, 122–128.
tionale for using such systems within the CAL-Visual project. Collis, B., and Verwijs, C. (1995). ‘‘A human approach to electronic per-
The main aim of this project is to facilitate the use of multi- formance support systems: Hybrid EPSS.’’ British J. Educational Tech-
media in the teaching, learning, and assessment in the field of nol., Oxford, U.K., 35(1), 5–21.
the built environment. A DPSS-based methodology has been Echeverry, D. (1996). ‘‘Multimedia-based instruction of building con-
used within the CAL-Visual framework. This includes a Web- struction.’’ Proc., 3rd Congr. held in conjunction with A/E/C Systems
based generic database of visual aids and a search engine aug- ’96, Computing in Civ. Engrg., Jorge Vanegas and Paul Chinowsky,
eds., ASCE, New York, 972–977.
mented by specialized prototypes dealing with specific topics Finkelstein, J. (1998). ‘‘$D multimedia datascapes: The SSV program.’’
such as building construction and building defects. The system Proc., Int. Conf. on Information Visualisation, IEEE Computer Society
uses different interfaces to access information ranging from a Press, New York, 280–285.
textual-based keyword search to VR models using hot spots Gery, G. J. (1991). Electronic performance support systems—How and
giving access to information in the database. why to remake the workplace through the strategic application of tech-
The project is nearing its completion of the design and im- nology. Weingarten Publications, Boston, Mass.
Hudson, S. R. G. (1998). ‘‘Multimedia performance support systems,’’
plementation phase where a Web-based approach was used to PhD thesis (draft), School of Computing and Mathematics, University
make the system widely available, accessible, and platform- of Teesside, Middlesbrough, U.K.
independent. The main issues related to the visual material Irving, R. H., Higgins, C. A., and Safayeni, F. R. (1986). ‘‘Computerised
included the capturing of textual information to be attached to performance monitoring systems: Use and abuse.’’ Communications of
the images, for this purpose a Microsoft Access form was de- the ACM, 29(8), 794–801.
McGraw, K. L. (1994). ‘‘Performance support systems: Integrating AI,
veloped for the structuring of this information and including
hypermedia and CBT to enhance user performance.’’ J. Artificial In-
it as metadata with the images. telligence in Educ., 5(1), 3–26.
Initial user’s evaluation of the system and trials with stu- Raybould, B. (1995). ‘‘Performance support engineering: An emerging
dents have been conducted to test the usability and perfor- development methodology for enabling organisational learning.’’ Perf.
mance of the system. The final stage of the work will include Improvement Quarterly, Spec. Issue on Electronic Perf. Support Sys.,
a detailed evaluation to measure the effectiveness of embed- 8(1), 7–22.
Riley, D., and Pace, C. (1997). ‘‘Improving teaching efficiency with mul-
ding visual material within a construction-engineering curric- timedia teaching aids.’’ Proc., 4th Congr. held in conjunction with A/
ulum using the CAL-Visual system. E/C Systems ’97, Computing in Civ. Engrg., Teresa M. Adams, ASCE,
New York, 129–135.
APPENDIX. REFERENCES Ring, G. (1994). ‘‘Creating electronic performance support systems for
Australian graduate students in computer education.’’ Proc., Annu.
Aminmansour, A. (1996). ‘‘Utilizing information technologies to better Meeting of the Am. Educational Res. Assn.
educate engineers of tomorrow.’’ Proc., 3rd Congr. held in conjunction Ruggles, C., Underwood, J., and Walker, D. (1995). ‘‘STILE: A hyper-
with A/E/C Systems ’96, Computing in Civ. Engrg., Jorge Vanegas and media approach to the provision of a flexible interdisciplinary resource
Paul Chinowsky, eds., ASCE, New York, 965–971. base for teaching and learning.’’ Innovations in Educ. and Training
Banerji, A. K. (1995). ‘‘Designing electronic performance support sys- Int., London, 32(1), 209–219.
tems,’’ PhD thesis, HCI Lab., University of Teesside, Middlesbrough, Varnadoe, S., and Barron, A. (1994). ‘‘Designing electronic performance
U.K. support systems.’’ J. Interactive Instruction Devel., 6(3), 12–17.

140 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 BRIDGE DAMAGE ASSESSMENT THROUGH FUZZY PETRI NET BASED
EXPERT SYSTEM
By Weiling Chiang,1 Kevin F. R. Liu,2 and Jonathan Lee3

ABSTRACT: The reliability of the techniques adopted for damage assessment is important for bridge manage-
ment systems. It is widely recognized that the use of expert systems for bridge damage assessment is a promising
direction toward bridge management systems. However, several important issues need to be addressed, such as
the management of uncertainty and imprecision, the efficiency of fuzzy rule based reasoning, and the need of
an explanation facility to increase confidence about the assessment results. To address the issues arising from
using expert systems, this paper is aimed at developing an expert system for assessing bridges based on an
expert system shell, which is called the fuzzy Petri net based expert system (FPNES). Major features of FPNES
include the ability to reason using uncertain and imprecise information, knowledge representation through the
use of hierarchical fuzzy Petri nets, a reasoning mechanism based on fuzzy Petri nets, and an explanation of
the reasoning process through the use of hierarchical fuzzy Petri nets. Therefore, this expert system for assessing
bridges does not impose any restriction on the inference mechanism. Furthermore, this approach offers more
informative results than other systems. An application to the damage assessment of the Da-Shi bridge in Taiwan
is used as an illustrative example of FPNES.

INTRODUCTION (1990), Hadipriono and Ross (1991), Shiraishi et al. (1991),

Krauthammer et al. (1992), Miyamoto et al. (1993), and Kush-
In recent years, many countries have been aware of bridge
ida et al. (1997).
problems and have initiated the development of bridge man-
Several important issues need to be addressed in using
agement systems to assist their decision-makers in establishing
expert systems for damage assessment of bridges:
efficient repair and maintenance programs. Basically, bridge
management systems consist of several modules to record the
inventory data of bridges, store the inspection results, evaluate • The descriptions of heuristic damage assessment knowl-
the damage states of bridges, propose maintenance and retrofit edge by experts usually take the form of natural language
schemes, estimate the maintenance and retrofit cost, and al- that contains intrinsic imprecision. For example, experts
locate the available funds appropriately. The reliability of the may treat the extent, degree, and seriousness of a defect
technique adopted for evaluating bridge states is playing an as linguistic variables that have fuzzy values. Therefore,
important role in bridge management systems. Damage as- a mechanism that can cope with imprecise information in
sessment for a bridge is a process to evaluate the damage state expert system is essential.
of the bridge based on visual inspection and empirical tests. • Uncertain and imprecise information involved in the dam-
However, it is a difficult task due to a lack of complete un- age assessment usually makes the problem harder. There-
derstanding of the mechanism of bridge deterioration. Bridge fore, a reasoning mechanism that can deal with uncertain
structures are too complex to analyze completely, and there- and imprecise information is crucial for damage assess-
fore, numerical simulations require a host of simplified as- ment.
sumptions. Nevertheless, an experienced engineer who has • To consider the efficiency of fuzzy rule based reasoning,
closely studied these problems over the years could use his it is important that fuzzy facts (input or inferred) find the
heuristic knowledge to achieve the task. The use of expert matched fuzzy rules efficiently, rather than scanning
systems for damage assessment has been widely recognized as through all the fuzzy rules.
a promising solution for alleviating the lack of experts. • In order to increase the confidence about the assessment
An expert system is a computer program that emulates the results, an explanation facility that can describe how the
problem-solving ability of human experts. Separating the conclusions are derived is instrumental for a computer-
knowledge base from the inference procedure makes it easy aided tool designed especially for damage assessment.
to extend and modify knowledge collected from all possible
sources. Expert systems can perform task much more effi- To address the first two issues, the writers have proposed a
ciently than experts, and can even be accessed through the possibilistic logic based approach (Lee et al. 1998) to handling
Internet without time and space barriers to serve as a ‘‘knowl- uncertain and imprecise information. The writers have also
edge server’’ (Eriksson 1996). Recently, researchers have be- proposed a fuzzy Petri net approach to addressing the last two
gun to investigate the use of expert systems to assess damage issues. The writers have proposed a framework of integrated
assessment—for example, Yao (1980), Ishizuka et al. (1982), expert systems based on proposed fuzzy Petri nets, called
Ogawa et al. (1985a,b), Furuta et al. (1991, 1996), Ross et al. fuzzy Petri net based expert systems (FPNES) (Lee et al.
1
1999). This paper is aimed at developing an expert system for
Prof., Dept. of Civ. Engrg., Nat. Central Univ., Chungli, Taiwan assessing bridges based on FPNES. Unlike other researchers’
32054. E-mail: [email protected]
2
PhD, Dept. of Civ. Engrg., Nat. Central Univ., Chungli, Taiwan
systems, our expert system does not impose any restriction on
32054. E-mail: [email protected] the inference mechanism, that is, the intended meaning is not
3
Prof., Dept. of Comp. Sci. and Information Engrg., Nat. Central Univ., required to be intact; meanwhile, the confidence level can be
Chungli, Taiwan 32054. E-mail: [email protected] partially certain. Furthermore, it offers more informative re-
Note. Editor: Sivand Lakmazaheri. Discussion open until September sults because the explanation provided in the system and the
1, 2000. To extend the closing date one month, a written request must confidence level of the conclusions can be used as a way of
be filed with the ASCE Manager of Journals. The manuscript for this
paper was submitted for review and possible publication on September
justification on whether to take the recommendations into ac-
28, 1998. This paper is part of the Journal of Computing in Civil En- count or not. An application to the damage assessment of the
gineering, Vol. 14, No. 2, April, 2000. 䉷ASCE, ISSN 0887-3801/00/ Da-Shi bridge in Taiwan is used as an illustrative example of
0002-0141–0149/$8.00 ⫹ $.50 per page. Paper No. 19364. FPNES. The result through FPNES not only matches the ex-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 141
perts’ judgments, but also provides more information than ex- Hierarchical Fuzzy Petri Nets
perts do.
To increase the efficiency of rule based reasoning, two is-
sues are particularly relevant: the possibility of exploiting con-
FUZZY PETRI NET BASED EXPERT SYSTEM currency, and the use of smart control strategies. To achieve
these goals, the writers have proposed the use of fuzzy Petri
FPNES, a framework of integrated expert systems based on
nets to model fuzzy rule based reasoning (Lee et al. 1998).
the proposed fuzzy Petri net, has been proposed by Lee et al.
Furthermore, the explanation of how to reach conclusions is
(1999) and is briefly described in this section. Major features
expressed through the movements of tokens in fuzzy Petri
of FPNES include the ability to reason using uncertain and
nets.
fuzzy information, explanation of the reasoning process
There are several rationales behind basing a computational
through the use of hierarchical fuzzy Petri nets, and a reason-
paradigm for expert systems on Petri net theory. First, Petri
ing mechanism based on fuzzy Petri nets.
nets achieve the structuring of knowledge within rule bases,
which can express the relationships among rules and help ex-
REASONING FOR UNCERTAIN AND FUZZY perts construct and modify rule bases. Second, Petri nets’
INFORMATION graphic nature provides the visualization of dynamic behavior
of rule based reasoning. Third, Petri nets make it easier to
The distinction between imprecise and uncertain informa- design an efficient reasoning algorithm. Fourth, Petri nets’ an-
tion can be best explained by the canonical form representation alytic capability provides a basis for developing a knowledge
(i.e., a quadruple of attribute, object, value, and confidence) verification technique. Last, the underlying relationship of con-
proposed by Dubois and Prade (1988). Imprecision implies the currency among rules activation can be modeled by Petri nets,
absence of a sharp boundary of the value component of the which is an important aspect where real-time performance is
quadruple, whereas uncertainty is related to the confidence crucial.
component of the quadruple, which is an indication of the
reliability of the information.
Fuzzy Petri Nets: Definition
To represent uncertain imprecise information, a fuzzy prop-
osition with a fuzzy valuation has been chosen by Lee et al. A fuzzy Petri net FPN is defined as a quintuple:
(1998), denoted as
FPN = (FP, UT, A, W, M0) (3)
(r, ␶) (1)
where FP = {( p1, F1), ( p2, F2), . . . , ( pm, Fm)} is a finite set
where r is fuzzy proposition of the form ‘‘X is F ’’ [i.e., X is of fuzzy places, where pi represents a fuzzy condition and Fi
a linguistic variable (Zadeh 1975, 1976) and F is a fuzzy set is a fuzzy subset of Ui to represent the fuzzy set of the con-
in a universe of discourse U]; and ␶ is a fuzzy valuation. It dition; UT = {(t1, ␶1), (t2, ␶2), . . . , (tn, ␶n)} is a finite set of
should be noted that for every formula (r, ␶) (called a truth- uncertain transitions, where tj represents the causal relationship
qualified fuzzy proposition), we assume ␶ is larger than or of fuzzy conditions and ␶j is a fuzzy truth value to represent
equal to ␶(r兩␲) [i.e., ␶(r兩␲) is the real fuzzy truth value derived the uncertainty about the causal relationship of fuzzy condi-
from r and a possibility distribution ␲], which means ␮␶(t) is tions; A is a set of arcs; W is a weight function; and
the upper bound of the possibility that r is true to a degree t. M0 = {M( p1), M(p2), . . . , M( pm)} is the initial marking, where
The fuzzy set is to represent the intended meaning of imprecise M( pi) is the number of tokens in pi.
information, while the fuzzy truth value serves as the repre- Each token is associated with a pair of fuzzy sets (F⬘i , ␶i)
sentation of uncertainty for its capability to express the pos- (called an uncertain fuzzy token). Fuzzy places with uncertain
sibility of the degree of truth. fuzzy tokens can be interpreted as uncertain fuzzy facts related
An inference rule for truth-qualified fuzzy propositions has to the fuzzy conditions modeled by the fuzzy places.
been expressed as follows:
Fuzzy Rule Based Reasoning and Fuzzy Petri Nets
(r1 ⵩ r2 ⵩ ⭈ ⭈ ⭈ ⵩ rn) → q, ␶1
r 1⬘, ␶2 It is widely recognized that fuzzy Petri nets are a promising
modeling mechanism for formulating fuzzy rule based reason-
r 2⬘, ␶3 (2)
⭈⭈ ing. The three key components in fuzzy rule based reasoning
⭈ —fuzzy propositions, fuzzy rules, and fuzzy facts—can be
r n⬘, ␶n ⫹1 formulated as places, transitions, and tokens, respectively. The
q ⬘, ␶n ⫹2 mapping between fuzzy rule based reasoning and fuzzy Petri
nets has been made below.
where ri, r ⬘i (i = 1 ⬃ n), q, and q⬘ are fuzzy propositions and
characterized by ‘‘Xi is Fi,’’ ‘‘Xi is F ⬘’’
i and ‘‘Y is G,’’ ‘‘Y is • Fuzzy places: Fuzzy places correspond to fuzzy proposi-
G⬘,’’ respectively; ␶j ( j = 1 ⬃ n ⫹ 2) are fuzzy valuations for tions. The fuzzy sets, attached to the fuzzy places, rep-
truth values and defined by ␮␶ j (t). Fi and F i⬘ are the subsets resent the values of fuzzy propositions. Fuzzy input and
of Ui; G and G⬘ are the subsets of V. There are three major fuzzy output places of a truth-qualified transition are used
steps for deriving q⬘ and ␶n⫹2 of (2): to represent the antecedent and conclusions parts of a
truth-qualified fuzzy rule, respectively.
• The fuzzy rules and fuzzy facts with fuzzy truth-values • Uncertain fuzzy tokens: An uncertain fuzzy token repre-
are transformed into a set of uncertain classical proposi- sents a truth-qualified fuzzy fact. The fuzzy sets and fuzzy
tions with necessity and possibility measures. truth-values are attached to uncertain fuzzy tokens to rep-
• The possibilistic entailment is performed on the set of resent the values and our confidence level about the ob-
uncertain classical propositions. served facts, respectively.
• We reverse the process in the first step to synthesize all • Uncertain transitions: Uncertain transitions are classified
the classical sets obtained in the second step into a fuzzy into four types: inference, aggregation, duplication, and
set, and to compose necessity and possibility pairs to form aggregation-duplication transitions. The inference transi-
a fuzzy truth-value. tions represent the truth-qualified fuzzy rules; the aggre-
142 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000
 gation transitions are designed to aggregate the conclusion the new token are derived by three steps: transformation, ag-
parts of rules that have the same linguistic variables; the gregation, and composition.
duplication transitions are used to duplicate uncertain Type 3: Duplication Transition (t d ). The purpose of du-
fuzzy tokens to avoid the conflict problem; and the ag- plication transitions is to avoid the conflict by duplicating the
gregation-duplication transitions link the fuzzy proposi- token. For example, there are l truth-qualified fuzzy rules hav-
tions with the same linguistic variables. These have been ing the same linguistic variable in the antecedents, denoted as
formally defined below.
(r11 → q1, ␶1), (r12 → q2, ␶2), ⭈ ⭈ ⭈ , (r1 → q1l, ␶l) (6)
i
Type 1: Inference Transition (t ). An inference transition where r1i means ‘‘Xi is F1i.’’ They are linked by a duplication
serves as modeling of a truth-qualified fuzzy rule. A truth- transition shown in Fig. 3. After firing the duplication transi-
qualified fuzzy rule having multiple antecedents is represented
as

(r1 ⵩ r2 ⵩ ⭈⭈⭈ ⵩ rn) → q, ␶1 (4)

where ri and q are of the forms of ‘‘Xi is Fi’’ and ‘‘Y is G,’’
respectively. In Fig. 1, after firing the inference transition t i1,
the tokens will be removed from the input places of t il, a new
token will be deposited into the output place of t i1, and the
fuzzy set and the fuzzy truth value attached to the new token
are derived by three steps: transformation, inference, and com-
position.
Type 2: Aggregation Transition (t a ). An aggregation
transition is used to aggregate the conclusions of several truth-
qualified fuzzy rules that have the same linguistic variable, and
to link the antecedent of a truth-qualified fuzzy rule that also
has the same linguistic variable. For example, there are m
truth-qualified fuzzy rules having the same linguistic variable
in the conclusions, denoted as

(r1 → q11, ␶1), (r2 → q12, ␶2), ⭈ ⭈ ⭈ , (rm → q1m, ␶m) (5)

where q1i is ‘‘Y is G1i.’’ In Fig. 2, after firing the aggregation

transition t am⫹1, the tokens in input places of t am⫹1 will be re-
moved, a new token will be deposited into the output place of
t am⫹1, and the fuzzy set and the fuzzy truth value attached to
FIG. 2. Modeling Aggregation of Conclusions by Aggregation
Transition (a) before Firing t amⴙ1 ; (b) after Firing t amⴙ1

FIG. 1. Modeling Fuzzy Rule Based Reasoning through Fuzzy FIG. 3. Modeling Duplication of Uncertain Fuzzy Token
Petri Nets (a) before Firing Inference Transition t i1 ; (b) after Fir- through Fuzzy Petri Nets (a) before Firing Duplication Transition
ing t i1 t d1 ; (b) after Firing t d1

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 143

tion t d1, the tokens in the input place of t d1 will be removed, ing a same linguistic variable in the conclusions and l truth-
new tokens will be added into the output places of t d1, and the qualified fuzzy rules having the same linguistic variable in the
fuzzy sets and the fuzzy truth values attached to the new to- antecedents, denoted as
kens are not changed.
Type 4: Aggregation-Duplication Transition (t ad ). An (r1 → q11, ␶1), (r2 → q12, ␶2), ⭈ ⭈ ⭈ , (rm → q1m, ␶m) (7a)
aggregation-duplication transition is a combination of an ag-
(q1(m⫹1) → s1, ␶m⫹1), q(1(m⫹2) → s2, ␶m⫹2), ⭈ ⭈ ⭈ ,
gregation transition and a duplication transition. It is used to
link all fuzzy propositions that have the same linguistic vari- (q1(m⫹1) → s1, ␶m⫹1) (7b)
able. For example, there are m truth-qualified fuzzy rules hav-
where si means ‘‘Z is Hi.’’ They are linked by an aggregation-
duplication transition shown in Fig. 4. After firing the aggre-
gation-duplication transition t ad
1 , the tokens in the input places
of t ad
1 will be removed, new tokens will be deposited into the
output places of t ad
1 , and the fuzzy sets and the fuzzy truth
values attached to the new tokens are derived by three steps:
transformation, aggregation, and composition.

Hierarchical Fuzzy Petri Nets

To overcome the complexity arising from large sizes of rule
bases and fuzzy Petri nets, two important features, modular-
ized rule bases and hierarchical fuzzy Petri nets, have been
adopted in FPNES. Modularization is used to partition rule
bases into smaller parts for organizing rules. In a hierarchical
fuzzy Petri net, each hierarchy contains a fuzzy Petri net that
may or may not contain other hierarchies. The connections
between hierarchies are achieved by defining importing and
exporting fuzzy places (Fig. 5). That is, an exporting fuzzy
place with respect to a hierarchy is defined as a fuzzy place
that is connected to the hierarchy by an arc from the fuzzy
place to the hierarchy; an importing fuzzy place with respect
to a hierarchy is defined as a fuzzy place connected to the
hierarchy by an arc from the hierarchy to the fuzzy place. In
a graphical representation, a hierarchy is drawn as a double-
lined square to connect the importing or exporting fuzzy
places. A hierarchical fuzzy Petri net that contains a main hi-
erarchy H0 and hierarchy H1 is illustrated in Fig. 5(a). The
FIG. 4. Modeling Aggregation-Duplication of Uncertain Fuzzy status of the fuzzy place P1 in Fig. 5(a) is shown in Fig. 5(b).
Token through FPN (a) before Firing Aggregation-Duplication In this figure, the fuzzy Petri net in the middle window is the
Transition t ad ad
1 ; (b) after Firing t 1 main hierarchy at the top level of the hierarchical structure,

FIG. 5. (a) Hierarchical Fuzzy Petri Net; (b) Place Status for Pl in Ho

144 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

and the fuzzy Petri net in the bottom window is hierarchy H1 fuzzy rules efficiently, rather than scanning through all the
at the second level. In H0, fuzzy places P2 and P4 are the fuzzy rules. Fuzzy Petri nets offer an opportunity to achieve
exporting fuzzy place with respect to hierarchy H1, and fuzzy this goal by using transitions and arcs to connect fuzzy rules
place P5 is the importing fuzzy places with respect to hier- as a net based structure. A data-driven reasoning algorithm has
archy H1. In the hierarchy H1, fuzzy places P1 and P3 are been developed by the writers; for details about the reasoning
the importing fuzzy places with respect to H0, and fuzzy place algorithm, see Lee et al. (1998).
P5 is the exporting fuzzing place with respect to H0. When a
token is inserted into the fuzzy place P2 in H0, it will be Overview of FPNES Tool
transited into hierarchy H1 and added to place P1 in hierarchy FPNES has been implemented in Java with a client-server
H1. Similarly, once a token enters into place P4 in H0, it will architecture, encompassing four main parts: fuzzy Petri net
be sent into hierarchy H1 and reaches the fuzzy place P3 in system (FPNS), user interface, transformation engine, and
H1. After firing transition t1, t2, and t3 in hierarchy H1, the knowledge bases (Fig. 6). Java has been adopted as the pro-
token arrives the fuzzy place P5 in H1 and then enters the gramming language for the FPNES tool for its capability of
fuzzy place P5 in H0. running on multiple platforms through the Internet.
There are two main benefits by having a hierarchical struc- FPNS is a modeling and analysis tool for fuzzy Petri nets,
ture in the system: (1) the notion of hierarchy makes easy the and serves as an inference engine, and explanation facility in
handling of complex systems through decomposition; and (2) FPNES. FPNS mainly contains the simulator and analyzer for
a hierarchical Petri net facilitates the reusability, that is, each fuzzy Petri nets. It provides the basic constructs for hierarchi-
hierarchy can be considered as a reuse unit. cal fuzzy Petri nets (e.g., hierarchies, fuzzy places, uncertain
transitions, arcs, and uncertain fuzzy tokens). After judging the
Transforming Modularized Fuzzy Rule Bases into firing conditions, the simulator will compute the fuzzy sets and
Hierarchical Fuzzy Petri Nets move tokens. The analyzer performs the tasks of analyzing the
properties of fuzzy Petri nets, such as incidence matrix, reach-
The explanation of how to reach conclusions is expressed
ability trees, and state equations.
through the movement of tokens in hierarchical fuzzy Petri
Users can edit modularized fuzzy Petri rule bases in the
nets. The first step toward this goal is to transform modular-
client site, including the assignments of linguistic variables,
ized fuzzy rule bases into hierarchical fuzzy Petri nets. To
truth-qualified fuzzy rules, the relationship of modules, and
bridge the gap between fuzzy rule based expert systems and
modularized structures of input facts. When users finish edit-
fuzzy Petri nets, it is important to have a mechanism to au-
ing the modularized fuzzy rule bases and the corresponding
tomatically transform modularized fuzzy rule bases into hier-
facts, and decide to run them, the data is then sent to the
archical fuzzy Petri nets. In our earlier paper (Lee et al. 1999),
transformation engine and transformed to a hierarchical fuzzy
two algorithms have been involved in the transformation. One
Petri net in FPNS. After FPNS processes the hierarchical fuzzy
is to transform modularized fuzzy rule bases into a hierarchical
Petri net with the aid of the reasoning algorithm, it sends the
incidence matrix. The other is to transform the hierarchical
results back to users. The results are presented in a hierarchical
incidence matrix into a hierarchical fuzzy Petri net.
fashion to provide a flexible explanation facility.
Reasoning Mechanism Based on Fuzzy Petri Nets DAMAGE ASSESSMENT
To consider the efficiency of fuzzy rule based reasoning, it To develop our modularized rule bases for damage assess-
is crucial that fuzzy facts (input or inferred) find the matched ment of bridges, we define the scope of knowledge domain,

FIG. 6. Overview of FPNES Tool

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 145

describe the process of linguistic assessment, and discuss how capacity of bridges include (1) concrete discontinuities caused
to construct the modularized rule bases. by overloadings, such as shear or flexure cracks, delamina-
tions, and spalls; (2) component displacements such as rota-
Knowledge Domain tional movement and settlement; and (3) soil scours due to
rushing flood. Various patterns and locations of the defects are
The type of defects that could occur in a bridge depends used to determine the possible causes and effects on the func-
upon the construction materials, environmental conditions, and tionality, and the magnitude of each defect relates to the level
external loadings. Nowadays, prestressed concrete bridges are of damage.
widely used in Taiwan. However, some of them become im- The level of damage, severity of defects, and intensity of
plicit threats to the public because they have served for a long confidence about descriptions are expressed linguistically and
period of time and lots of defects are developing. To start with, are considered as linguistic variables. The severity of damage
the scope of the system is limited to (1) the concrete bridges in terms of linguistic variables is classified into seven fuzzy
that have prestressed concrete I-girders and hammerhead piers; levels, such as (1) very severe [i.e., ␮(x) = x 8]; (2) severe [i.e.,
and (2) the defects that occur most commonly and influence ␮(x) = x 4]; (3) fairly severe [i.e., ␮(x) = x 2]; (4) fair [i.e., ␮(x)
the load-carrying capacity. = x]; (5) fairly slight [i.e., ␮(x) = x1/2]; (6) slight [i.e., ␮(x) =
x 1/4]; and (7) very slight [i.e., ␮(x) = x1/8], where x is a real
Linguistic Assessment number between zero and one, and denotes the degree of dam-
In order to keep a bridge functioning, each component age. The intensity of confidence about rules or facts is also
serves several functions to meet demands such as traffic load- considered in terms of linguistic variables with values such as
ing, earthquakes, erosion, and environmental corrosion. Based (1) very true [i.e., ␮(t) = t 2]; (2) true [i.e., ␮(t) = t]; (3) fairly
on the observed defects, an expert can identify what kind of true [i.e., ␮(t) = t 1/2]; (4) fairly false [i.e., ␮(t) = (1 ⫺ t)1/2];
function is insufficient, judge the effect, and determine the (5) false [i.e., ␮(t) = 1 ⫺ t]; (6) very false [i.e., ␮(t) = (1 ⫺
damage level. The inference procedure for experts to assess t)2]; and (7) unknown [i.e., ␮(t) = 1], where t is a real number
bridge damage is described as follows: between zero and one, and denotes the degree of truth.
Modularized Rule Bases
1. A team of inspectors visually investigates each compo-
nent of a bridge and records the observed defects, such Crack, delamination, spalling, movement, and scour are five
as cracking, delamination, spalling, honeycomb, efflores- main defects in our system. Cracking is the most common
cence, corrosion, and movement. defect and can be discovered in any kind of concrete structure.
2. Based on the defect symptoms, such as locations and Its location and pattern are used to determine whether a crack
patterns, experienced bridge engineers can identify the affects the load-carrying capacity. The level of severity of a
possible causes of the defects. crack should not be determined directly without full consid-
3. The possible causes can induce what kinds of functions eration of the factors involved. the factors could include but
are eliminated due to the defects. are not necessarily limited to those shown in Fig. 7, and the
4. The damage level of each defect is evaluated according related fuzzy rules of module ‘‘crack in I-girder’’ are presented
to the quantitative description of its symptoms. in Fig. 8. Delamination is separation along a plane nearly par-
5. The levels of functional insufficiency are inferred based allel to the surface of the concrete. Spalling is a depression
on the damage levels of the defects. resulting from the dislodgment surface of concrete. It is typi-
6. The assessment of damage can be estimated by aggre- cally found in old concrete structures and not in young con-
gating both the functional insufficiency and its levels. crete. Before developing into a depression, spalling may be
presented as a delamination below the concrete surface. Some-
The expression of observed defects by inspectors is viewed times it is as deep as the reinforcing steel.
as the input of the system. Defects affecting the load-carrying Vertical movement such as differential settlement can pro-
duce serious distress in a bridge structure. Rotational move-
ment of substructure can be considered to be the result of
unsymmetrical settlement or lateral movements. Piers, abut-
ments, and walls may undergo this type of movement. The
degree of damage depends on the extent of vertical and rota-
tional movement. The recommendation of strengthening or re-
pairing is based on the possible cause. Scour of the substruc-
ture support is one of the more frequent factors that may cause
or lead to structural failure or foundation distress. Scour is the
removal of the streambed, backfill, slopes, or other supporting
material, by stream tidal action, dredging, propeller backwash,
etc. The degree of damage depends on the extent of exposure
FIG. 7. Factors Related to Cracking in Concrete of foundation and the extent of deterioration of foundation.

FIG. 8. Part of Rule Base for Crack in I-Girder

146 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 FIG. 9. Fuzzy Rule Petri Net Transformed from Rule Base of Crack in I-Girder

The whole rule base is subdivided into a lot of modules in kens were transformed into these nets in order to fire transi-
order to overcome the complexity arising from too large sizes tions and perform the reasoning mechanism. The results of
of the rule bases and to make the rules well organized. At damage assessment using FPNES for the Da-Shi bridge were
present there are 34 modules in this system, in which 216 expressed in a hierarchical fashion to serve as an explanation
truth-qualified fuzzy rules and 224 recommendations are in- mechanism to facilitate the retrieval of detailed information on
cluded. Seven modules related to the cracking in concrete, five damaged components from the top down to lower levels (Fig.
modules dealing with the delamination in concrete, five mod- 10). The recommendations embedded in rule bases were also
ules concerning the spalling in concrete, two modules about provided on an if-needed basis. As a result, the damage to the
the movement of foundation, and two modules handling to the Da-Shi bridge was evaluated to be severe with strong confi-
scour of support are constructed. Furthermore, two modules dence (i.e., very true) since the overall damage to the super-
are used to evaluate the insufficiency of shear or flexure re- structure was severe with common confidence (i.e., true) and
sistance, and 11 modules are used to assess the damage levels the overall damage to the substructure was also severe with
of components; 216 truth-qualified fuzzy rules and 224 rec-
strong confidence (i.e., very true). The severity of superstruc-
ommendations are accommodated in these modules. As was
ture results from the overall damage to decks was fairly slight
mentioned previously, fuzzy Petri nets’ graphical representa-
tion can help us construct and modify fuzzy rule bases. Fig. with fair confidence (i.e., fairly true), the overall damage to I-
9 shows the fuzzy Petri nets after the transformation of the girders was severe with common confidence (i.e., true), and
rule bases of a crack in an I-girder. the overall damage to diaphragms was very slight with strong
confidence (i.e., very true). Meanwhile, the severity of sub-
EXAMPLE structure as derived from the overall damage of piers was se-
vere with strong confidence (i.e., very true). The FPNES re-
To demonstrate the use of FPNES, the Da-Shi bridge lo- sults not only match the experts’ judgments, but also provide
cated in northern Taiwan was considered. It was rebuilt in more information than experts do, because the explanation
1960 as a simply supported, 12-spanned bridge, 550 m long provided in the system and the confidence level associated
and 7.8 m wide, across the Da-Han river. This bridge consists with the conclusions can be used to justify whether to take the
of 12 decks, 36 prestressed concrete I-girders, 120 diaphragms, recommendations into account.
11 piers, and two abutments. The following recommendations were made with strong
In 1997, this bridge was inspected by the Center of Bridge
confidence (i.e., very true). The observed defects or deficien-
Engineering Research at National Central University. Visual
cies may affect the load-carrying capacity of the bridge. For
inspection revealed that many minor cracks accompanied with
efflorescence spread over eight panels within deck 7. The PCI the safety of the traveling public, traffic on the bridge should
girders S9G1, S9G2, S9G3 in span 9 and S10G1, S10G2, be limited. A second inspection and load capacity evaluation
S10G3 in span 10 had severe flexure, shear cracks, and some should be made. The second inspection is used to identify the
spalls. There were two diaphragms where several spalls were compressive strength of concrete, the carbonation depth in
found. The foundations of piers P1 to P6 were exposed up to concrete, the content of chloride in concrete, the corrosion of
5 m above the ground level, and some of them suffered severe reinforcement, the crack depth in concrete, etc. The load ca-
spalling. Furthermore, pier 2 had rotational movement. pacity evaluation is performed to ascertain the safety load ca-
After executing FPNES for damage assessment of the Da- pacity for the current condition of the bridge. Based on the
Shi bridge, the hierarchical fuzzy Petri nets were constructed results of the load capacity evaluation, a plan can be made for
based on the modularized rule bases, and uncertain fuzzy to- rehabilitating, strengthening, or replacing the bridge.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 147
 FIG. 10. Results of Damage Assessment for Da-Shi Bridge Using FPNES

CONCLUSION damage assessment using genetic algorithms and neural networks.’’

Microcomputers in Civ. Engrg., 11(1), 37–45.
This research has resulted in the development of a fuzzy Furuta, H., Shiraishi, N., Umano, M., and Kawakami, K. (1991). ‘‘Knowl-
Petri net based expert system (FPNES) for bridge damage as- edge-based expert system for damage assessment based on fuzzy rea-
soning.’’ Comp. and Struct., 40(1), 137–142.
sessment, which contains a reasoning mechanism to deal with
Hadipriono, F. C., and Ross, T. J. (1991). ‘‘A rule-based fuzzy logic
uncertain and imprecise information, a fuzzy Petri nets ap- deduction technique for damage assessment of protective structure.’’
proach to modeling fuzzy rule based reasoning, and a tool Fuzzy Sets and Sys., 44(3), 459–468.
supporting the damage assessment of bridges. Unlike other Ishizuka, M., Fu, K. S., and Yao, J. P. T. (1982). ‘‘Inference procedures
researchers’ systems, our expert system does not impose any under uncertainty for the problem-related method.’’ Information Sci.,
restriction on the inference mechanism, that is, the intended 28(3), 179–206.
meaning is not required to be intact; meanwhile, the confi- Krauthammer, T., Muralidharan, R., and Schimdt, W. (1992). ‘‘Combined
symbolic-numeric explosion damage assessment for structures.’’ J.
dence level can be partially certain. Furthermore, it offers more Comp. in Civ. Engrg., ASCE, 6(4), 417–434.
informative results because the explanation provided in the Kushida, M., Miyamoto, A., and Kinoshita, K. (1997). ‘‘Development of
system and the confidence level of the conclusions can be used concrete bridge rating prototype expert system with machine learning.’’
to justify whether to take the recommendations into account. J. Comp. in Civ. Engrg., ASCE, 11(4), 238–247.
An application to the damage assessment of the Da-Shi bridge Lee, J., Liu, K. F. R., and Chiang, W. (1998). ‘‘Fuzzy Petri nets for
in Taiwan is used as an illustrative example of FPNES. The modeling rule-based reasoning.’’ Int. J. Artificial Intelligence Tools,
7(4), 463–485.
FPNES results not only match the experts’ judgments, but also
Lee, J., Liu, K. F. R., and Chiang, W. (1999). ‘‘A fuzzy Petri nets based
provide more information than experts do. expert system for damage assessment of bridges.’’ IEEE Trans. on Sys.,
Future research will consider issues related to damage as- Man., and Cybernetics, Part B: Cybernetics, 29(3), 350–370.
sessments that remain to be addressed further: (1) how to en- Miyamoto, A., Morikawa, H., Kushida, M., and Tokuyama, T. (1993).
hance FPNES to involve numerical data such as earthquake ‘‘A knowledge-based expert system application in structural safety as-
records, experimental results, and load testing information; and sessment of concrete bridges.’’ Bridge management: Inspection, main-
(2) how to integrate the numerical data and linguistic assess- tenance, assessment and repair 2, J. E. Harding, G. E. R. Parke, and
ment into an overall evaluation. M. J. Ryall, eds., E & FN Spon, London, 96–109.
Ogawa, H., Fu, K. S., and Yao, J. P. T. (1985a). ‘‘An inexact inference
for damage assessment of existing structures.’’ Int. J. Man-Machine
ACKNOWLEDGMENTS Studies, 22(2), 295–306.
Ogawa, H., Fu, K. S., and Yao, J. P. T. (1985b). ‘‘SPERIL-II: An expert
The writers would like to thank Sam Lin for his early work on this system for damage assessment of existing structures.’’ Approximate
project. This research is partially sponsored by the National Science reasoning in expert systems, M. M. Gupta, A. Kadel, W. Bandler, and
Council (Taiwan) under grants NSC 86-2213-E-008-008 and NSC 86- J. B. Kiszka, eds., Elsevier Science, Amsterdam, 731–744.
2211-E-008-015. Ross, T. J., Sorensen, C., Savage, J., and Carson, J. M. (1990). ‘‘DAPS:
Expert system for damage assessment.’’ J. Comp. in Civ. Engrg.,
APPENDIX I. REFERENCES ASCE, 4(4), 327–348.
Shiraishi, N., Furuta, H., Umano, M., and Kawakami, K. (1991). ‘‘An
Dubois, D., and Prade, H. (1988). Possibility theory: An approach to expert system for damage assessment of a reinforced concrete bridge
computerized processing of uncertainty. Plenum, New York. deck.’’ Fuzzy Sets and Sys., 44(3), 449–457.
Eriksson, H. (1996). ‘‘Expert systems as knowledge servers.’’ IEEE Yao, J. P. T. (1980). ‘‘Damage assessment of existing structures.’’ J.
Expert, 13(3), 14–19. Comp. in Civ. Engrg., ASCE 106(4), 785–799.
Furuta, H., He, J., and Watanabe, E. (1996). ‘‘A fuzzy expert system for Zadeh, L. A. (1975). ‘‘The concepts of a linguistic variable and its ap-

148 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000

 plication to approximate reasoning (Part 1).’’ Information Sci., August, p = fuzzy place;
199–249. pi = fuzzy condition;
Zadeh, L. A. (1976). ‘‘The concepts of a linguistic variable and its ap- q, r, s = fuzzy proposition;
plication to approximate reasoning (Part 2).’’ Information Sci., Septem- t = truth degree;
ber, 43 – 80.
tj = causal relationship of fuzzy conditions;
ta = aggregaton transition;
APPENDIX II. NOTATION
t ad = aggregation-duplication transition;
The following symbols are used in this paper: td = duplication transition;
ti = inference transition;
A = set of arcs; U, V = universes of discourse;
F, G = fuzzy set; UT = finite set of uncertain transitions;
FP = finite set of fuzzy places; W = weight function;
FPN = fuzzy Petri net; X, Y, Z = linguistic variable;
H = hierarchy; ␮(x) = membership function;
M0 = initial marking; ␲ = possibility distribution; and
M( pi) = number of tokens in pi ; ␶ = fuzzy truth value.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 149

JOURNAL OF
COMPUTING IN CIVIL ENGINEERING
Volume 14 Number 3 July 2000

TECHNICAL PAPERS
A Representation for Conceptual Design of Buildings
Hugues Rivard and Steven J. Fenves. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Application Framework for Development of Simulation Tools
Dany Hajjar and Simaan M. AbouRizk. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Evaluation of Mobile Mapping Systems for Roadway Data Collection
Hassan A. Karimi, Aemal J. Khattak, and Joseph E. Hummer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Structural Matrix Computations with Units
Mark A. Austin, Wane-Jang Lin, and Xiaoguang G. Chen. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Forecasting Cyanobacterial Concentrations Using B-Spline Networks
Holger R. Maier, Tarek Sayed, and Barbara J. Lence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
System Dynamics Modeling of Reservoir Operations for Flood Management
Sajjad Ahmad and Slobodan P. Simonovic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
Distributed Genetic Algorithm Model on Network of Personal Computers
M. C. Balla and S. Lingireddy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Computerized System for Application of Value Engineering Methodology
Sadi Assaf, Osama A. Jannadi, and Ahmed Al-Tamimi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
 A REPRESENTATION FOR CONCEPTUAL DESIGN OF BUILDINGS
By Hugues Rivard,1 Associate Member, ASCE, and
Steven J. Fenves,2 Honorary Member, ASCE

ABSTRACT: The building representation for storing design solutions adopted in SEED-Config consists of two
levels of abstraction models defined on top of the object-oriented data model. The first level is an information
model, called the building entity and technology model, which stores design data as they are generated during
conceptual design, supports case-based reasoning, and shares data among all design participants. This model
represents each building entity as a generic container that encompasses its properties, taxonomy, geometry,
composition, relationships, and design knowledge applied in generating it. The second level is a conceptual
model that defines the types of objects, relationships, and data needed to fully represent the information in a
given design domain. The conceptual model specifies the semantics of the design information for the domain
using the syntax defined in the information model. The proposed representation also includes a faceted classi-
fication scheme to define the controlled vocabulary from which indexes are obtained. The representation has the
following advantages: It integrates multiple views, supports design evolution, supports design exploration, and
is extensible.

INTRODUCTION lyzed until a satisfactory solution emerges. The relevant fea-

tures of the design problem manifest themselves as the design
SEED (Software Environment to support the Early phases proceeds and depend on the decisions taken. Hence, the build-
in building Design) is a multidisciplinary project intended to ing representation should support dynamic changes not just in
provide computer support for the conceptual design of build- terms of values and instances, but also in terms of model struc-
ings. The emphasis is on supporting early design exploration, ture (Eastman and Fereshetian 1994). Finally, the building
that is, the fast generation of alternative design concepts and design process evolves over time with the advent of new meth-
their rapid evaluation against a broad spectrum of relevant— ods, technologies, requirements, and products. Hence, a build-
and possibly conflicting—criteria (Flemming and Woodbury ing representation should be extensible because it is next to
1995). The SEED project is subdivided into three main mod- impossible to develop a representation that can satisfy every
ules: SEED-Pro, which supports the generation of an architec- need of every designer at the outset. Therefore, a conceptual
tural program; SEED-Layout, which supports the generation building design representation should satisfy the following four
of schematic layouts; and SEED-Config, which supports the requirements: (1) Integrate multiple views; (2) support design
design of a 3D building configuration in terms of spaces, sys- evolution; (3) favor design exploration; and (4) be extensible.
tems (e.g., structural, enclosure, and mechanical systems), sub- Several building design representations have been proposed
systems and physical components (Fenves et al. 1995; Wood- by both industry and academia. There are two international
bury and Chang 1995). This article documents the standardization efforts that address the representation of build-
representation for the conceptual design of buildings adopted ing designs: STEP (STandard for the Exchange of Product
in SEED-Config. model data) being developed by the International Organization
The first step necessary in solving a problem with a com- for Standardization (Burkett and Yang 1995); and the Industry
puter involves the development of a representation for the Foundation Classes, which are specifications for a set of stan-
problem (Simon 1996). Hence, the development of the SEED- dardized object definitions, being developed by the Interna-
Config design environment first required the definition of a tional Alliance for Interoperability (Kiviniemi 1999). Both
building design representation. The building design process standards address later stages of design, do not explicitly sup-
has some inherent characteristics that impose requirements on port design evolution, and define static building representa-
such a representation. The design of a building requires a team tions aimed at the transfer of information between applica-
of designers from numerous design disciplines, each contrib- tions. Other building or structural design representations have
uting a particular body of knowledge and expertise to the over- been proposed in the literature such as RATAS (Bjork 1989,
all effort. A building is considered successful when the con- 1992, 1994), EDM (Eastman 1992), multiple views and rep-
tributions of the designers and their respective views are resentations (Rosenman and Gero 1996), principal and joint
integrated into a harmonious whole. The building design pro- model (Dias 1996), and Generic Object-Oriented Detailed De-
cess produces design information that naturally grows in quan- sign of Buildings (GOOD㛭B) (Biedermann and Grierson
tity and quality as the design evolves. Thus, data that are only 1995). The first three are general building design representa-
outlined in the early stages of design are gradually modified, tions, and the last two are building structure representations.
enhanced, and detailed as the design unfolds. The building The drawbacks of these representations are lack of extensibil-
design process is also exploratory in nature. It is an iterative ity (RATAS, principal and joint model, GOOD㛭B), lack of
process in which design alternatives are synthesized and ana- support for integrating multiple views (EDM, principal and
1
joint model, GOOD㛭B), and lack of support for design explo-
Asst. Prof., Centre for Building Studies, Dept. of Build., Civ., and
Envir. Engrg., Concordia Univ., 1455 de Maisonneuve Blvd. West, Mon-
ration and for querying and browsing design information (RA-
treal, Quebec, Canada H3G 1M8. TAS, EDM, multiple views and representations, principal and
2
Univ. Prof. Emeritus, Dept. of Civ. and Envir. Engrg., Carnegie Mel- joint model, and GOOD㛭B). Therefore, none of these building
lon Univ., Pittsburgh, PA 15213. design representations specifically addresses the early design
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, stages and satisfies the four requirements stated earlier. Finally,
2000. To extend the closing date one month, a written request must be no commercially available software applications address the
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on October 5, 1998.
conceptual design phase of buildings.
This paper is part of the Journal of Computing in Civil Engineering, The proposed building design representation is an attempt
Vol. 14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0151– to overcome the above limitations. The representation consists
0159/$8.00 ⫹ $.50 per page. Paper No. 19383. of two levels of abstraction models defined on top of the ob-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 151
ject-oriented data model. The information model is a set of tain, extend, and understand by a single specialist (Howard et
basic constructs, built on top of a data model, enabling the al. 1992; Phan and Howard 1993). Furthermore, it cannot ad-
organization of data about a product in a logical and structured dress unexpected situations that require the addition of new
manner. An information model for a product carries richer con- attributes in an already instantiated object, because each at-
structs than a data model and spans the product’s life cycle tribute has to be hard-coded in advance in the structure (e.g.,
(Eastman and Fereshetian 1994). The information model deals a class or a relational table). In the third approach, at the other
with the syntax of the design information; it defines how this extreme, each attribute is stored separately. The attribute val-
information is structured. The conceptual model of an object ues are accessed by attribute names. This approach leads to
is a computer-based representation of the object whose defi- complex and obscure naming conventions (e.g., preliminary㛭
nition depends on the observer’s perception and conception thickness, nominal㛭thickness, and actual㛭thickness).
(Rosenman and Gero 1996). It refines an information model The approach selected for the BENT model is located be-
by specifying the categories of information used in a particular tween these two extremes. The attribute-value pairs character-
domain (Bjork 1992). The conceptual model specifies the se- izing a building entity are organized at two hierarchical levels,
mantics of the design information for a specific domain; it the unit level and the component level. At the unit level, data
assigns meanings by mapping structures from the information are grouped into three subsets, the functional unit (FU) the
model to knowledge, objects, and properties of the domain. design unit (DU), and the evaluation unit (EU). The FU de-
This paper is organized as follows. The information model fines intended purposes, requirements, and constraints on the
and its main constructs are described in the second section. entity. These requirements have to be satisfied to realize the
The third section discusses how the issues of classification and intended purpose. The FU can be seen as a design-problem
indexing are addressed. The conceptual model for conceptual statement (Gielingh 1988). The DU includes all the physical
structural design is presented in the fourth section. Finally, the and spatial characteristics that define the actual design of the
representation is evaluated in the fifth section. Although the entity. The DU can be seen as a solution alternative to the
intent of the information model is to be general and applicable design problem. The EU represents the behavior of the entity
to the overall building design process, the discussion focuses and includes the response to stimulation associated with dif-
on the domain of conceptual structural design. A proof of con- ferent design conditions. It can also be used to record the con-
cept prototype has been implemented, and readers interested sequences of a design that has been implemented. Examples
in additional details may refer to Rivard (1997). of data recorded in the EU are costs, stresses, heat flow, con-
densation rate, and in-service performance.
BUILDING ENTITY AND TECHNOLOGY At the second level of data aggregation, the attribute-value
INFORMATION MODEL pairs of an entity are combined into small cohesive subsets,
each of which is called a component. A component is defined
The information model, called the BENT (Building ENtity as a group of closely related attributes that are found together
and Technology) model, consists of a set of constructs that in a repository (access-cohesive), instantiated at the same time
provides the basic building blocks for representing any build- (time-cohesive), and that represent the same concept (concept-
ing design domain. The goal of this information model is cohesive). Cohesion is the only criterion used in decomposing
threefold: To store design data as they are generated during entities. It is defined as a measure that shows how closely the
conceptual design; to support case-based reasoning; and to attributes of an entity relate to one another. The component is
share data among all design participants. The case-based rea- based on the concept of the primitive proposed by Phan and
soning functionality of SEED-Config is described in (Rivard Howard (1993) in the primitive-composite approach (P-C ap-
and Fenves 2000). proach). The P-C approach is a data model and a structured
In the information model, the building is represented as an methodology for modeling facility engineering processes and
assembly of entities with relationships among them. An entity data to achieve integration. An example of a component is one
is something that can be distinctly identified in a building de- that collects together the attributes used to describe the section
sign and about which data are accumulated (Eastman and Fer- properties of structural members such as depth, width, cross-
eshetian 1994). Each building entity represents a concept section area, and moment of inertia. There can be only one
meaningful to design participants such as a beam, a room, or instance of a component class stored in a building entity.
a structural frame. An entity can be a system, a subsystem, a Hence, the name of the component class uniquely identifies
constituent, a part, a feature of a part, a space, or a connection one subset of attributes. Components allow entities to be re-
(Gielingh 1988). All building entities are modeled with a ge- fined in staged steps by adding sets of attribute-value pairs to
neric container that glues together the modular information of the entity as they are generated in the design process. Hence,
a building entity. This container provides a single standardized there is no need to predict all possible attribute-value pairs
interface to access information about any entity in a building. needed in a building entity at the outset. Components also
allow the integration of multiple views within one entity. For
Organization of Attribute-Value Pairs instance, a load-bearing exterior wall that has an enclosure-
A building entity is described by a set of attribute-value related view and a structural-related view can be represented
pairs that may be organized according to the following three with a building entity containing components that are specific
approaches: for the enclosure view (e.g., thermal properties) and for the
structural view (e.g., structural properties) as well as compo-
nents that are shared among the two views (e.g., material prop-
• The collection of all attributes defining an entity may be
erties) (Rivard et al. 1999).
grouped into one flat structure.
• The attributes may be divided into small cohesive subsets.
Building Entity and Its Elements
• Each attribute may be represented as a distinct structure.
A building entity includes the following elements: Compo-
The first and last approaches correspond to the two extremes nents (FU, DU, and EU components); classifier instances; ge-
of a scale. The first approach leads to the creation of exceed- ometric descriptions (primary spatial and secondary geometric
ingly complex entities that require values that may not be de- representations); relationships [containment and domain-spe-
fined at instantiation time, do not facilitate the display or cific relationships (DSRs)]; technology tags; and groups. The
exchange of a subset of the attributes, and are difficult to main- first element is described in the previous subsection. The other
152 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 several parts. A hierarchical decomposition is the result of
breaking down a complex design problem successively until
smaller and more manageable subproblems are obtained.
These subproblems or parts are designed with some degree of
independence, because each part affects the others only
through its function, regardless of the details of how this func-
tion is achieved within the part (Simon 1996). In the BENT
model, every node of the hierarchical decomposition is rep-
resented as a building entity linked by containment relation-
ships.
Other essential relationships are stored in the second cate-
gory of relationships, called domain-specific. Each building
domain has its own specific types of relationships that matter.
For instance, in structural engineering, a column supports a
girder, whereas in mechanical engineering, a thermostat con-
FIG. 1. Class Diagram of Building Entity in UML Notation trols the air mixture of an air handling unit. These relationships
are important and modeled with DSRs. A building entity may
five elements are described below. Fig. 1 shows the class di- have several DSRs with other entities. Examples are support,
agram of the building entity and the other constructs using the connect, and control. Each DSR is implemented as an object
Unified modeling language notation (Booch et al. 1999). that specifies the roles of the two building entities it relates.
Hence, only binary relationships among building entities are
supported.
Classification and Indexing
Spatial relationships—such as next to, above, spatially con-
A classified instance is a label assigned to a building entity tained in, and adjacent—are not stored explicitly in this in-
for the purpose of classifying and indexing it. Classifier in- formation model because they can be obtained directly from
stances are required to classify the generic building entities as the geometric modeler.
they are being refined during the design process. Classifier
instances are also used as indexes for querying the case library. Technology Nodes and Technology Tags
The classification and indexing approaches are described in
The BENT model separates design knowledge from design
the third section.
data. This partition allows design knowledge to evolve inde-
pendently from the information model. Design knowledge is
Geometry encapsulated in a finite number of technology nodes organized
The geometric descriptions of a building entity are grouped into a hierarchically structured technology graph. This tech-
into two categories. The primary spatial representation is its nology graph assists designers in designing a building entity
high level geometric description, which is used primarily for by offering various alternatives. A technology node is defined
reasoning about its geometry and topological relationships as ‘‘a collection of computational mechanisms that creates,
with other entities, whereas the secondary geometric represen- details and instantiates entities to satisfy the requirements de-
tation of an entity is its domain-specific geometric represen- fined in the functional unit of a building entity in a design
tation. Each entity has only one primary spatial representation, context based on a specific construction technology or form
but it may have several secondary representations. This rep- generation principle’’ (Woodbury and Fenves 1994). In the
resentation scheme relies on the premise that topological re- structural design domain, the technology graph represents the
lationships of physical entities are invariant with respect to various alternative structural systems, subsystems, and types
their domain-specific secondary representations (Zamanian of elements available to the designer. The root of a graph op-
and Fenves 1994). For example, a structural engineer may erates on an abstract building as a whole, whereas nodes at
work directly with the primary spatial representation of a beam succeeding levels operate on more and more specific building
as a line whereas an architect may use a secondary geometric elements. Hence, the technology graph may deal with elements
representation to view the beam as a mass element. A non- ranging from the most abstract (e.g., a full 3D building for
manifold boundary representation scheme is used, because which a tube structure may be an alternative structural system)
topological relationships can be represented without being af- to the most specific elements (e.g., individual beams or even
fected by the various dimensionalities used in representing the connections, reinforcement, etc.).
geometric entities. The selection of this modeling scheme is Technology nodes have constraints on their applicability
justified because a design session may have to deal with lines, based both on feasibility and on economy of design. Engi-
planes, and volumes at the same time (e.g., the primary spatial neering expertise and heuristics are used to determine appro-
representations of a beam, a slab, and a room). priate constraints for the more abstract technology nodes,
whereas constraints are obtained from product catalogs and/or
Relationships design standards for the more specific technology nodes. When
a designer attempts to apply a technology node to a building
A relationship is a qualified association between two enti- entity, the technology node’s constraints are verified. Con-
ties. Relationships among building entities are grouped into straints are assessed differently depending on their nature.
two categories: Containment relationships (also known as ag- When constraints take the form of geometric limitations, such
gregation relationships), which capture the links between an as a suitable span range or a required aspect ratio, the tech-
entity and its parts, and DSRs, which contain other relation- nology node verifies these values from the primary spatial
ships of interest. representation or from DU components stored in the current
Composite objects are modeled as building entities linked building entity. When constraints take the form of usabil-
by containment relationships. Consequently, the properties of ity requirements, such as loading, minimum deflections, and
the parts define the properties of the composite object (East- voids required, the technology node checks these values in
man and Fereshetian 1994). The composite object is treated as FU components associated with the current building entity.
a unit in many operations, although physically it is made of In other circumstances, when there are constraints on whether
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 153
a technology may be used in combination with other technol- of building entities. The collection of these building entities
ogies [e.g., a precast horizontal system should not be sup- represents an invaluable repository of design knowledge. This
ported by an unbraced frame (Schodek 1990)], the technology knowledge can be useful only if it is easy to locate. To facil-
node checks whether the related building entities are imple- itate the retrieval process and to maximize the value of such
mented using compatible technologies (Rivard et al. 1996). a collection, two complementary tasks are necessary, classi-
Once a technology node is applicable in the current context, fying and indexing. Classifying involves the elaboration of a
a designer may select it and apply it to the current building classification by arranging objects into classes, and indexing
entity. The technology node advances the design of the build- is the process of assigning meaningful labels to each of the
ing entity by adding DU components, classifier instances, and objects for retrieval purposes. These two tasks are intertwined
DSRs to the entity or by decomposing the current building because classification deals with the organization of knowl-
entity into subentities. Hence, a building entity is designed edge and provides a logical foundation for indexing. The ap-
through the selection and application of a sequence of tech- proach adopted for classifying and indexing is presented in the
nology nodes from the technology graph. For instance, a slab next two subsections.
element could be designed with the selection of three tech-
nology nodes: ‘‘Decking,’’ ‘‘steel decking,’’ and ‘‘steel and Classification
concrete.’’ This sequence results in what we call a design path SEED-Config uses a faceted classification scheme because
through the technology graph. A design path records a set of it has the advantage of being more specific than other schemes
design decisions which can be reenacted in the same context such as enumerative ones. In this scheme, an object is de-
or in new ones, thus supporting rapid adaptation and redesign scribed by a set of facets called categories. Each category de-
without having to start from ‘‘scratch.’’ A multifunctional fines a specific aspect of an object and consists of a hierarchy
building entity may have several design paths applied to it. of classifiers. A classifier is a label assigned to a group of
For instance, a building entity that performs an enclosure and objects sharing a common characteristic. An example of a cat-
a structural function will have at least two design paths. This egory in structural design is the structural material category.
design knowledge representation supports the generation of An example of a classifier is ‘‘timber,’’ which is assigned to
solutions in staged steps so as to allow backtracking and the beams made out of wood. Fig. 2 shows the ‘‘structural mate-
generation of alternative design solutions for the same sub- rial’’ category (in bold) and its partial hierarchy of classifiers,
problem (Fenves et al. 1995). A building entity can revert to where the root is the category, and each subsequent node is a
a previous design state (i.e., unapply a technology node) by classifier. Classifiers are arranged in a subsumption hierarchy.
removing all the refinements a technology node generated for Subsumption is the process of recognizing a narrower concept
it. Readers interested in additional information regarding the (or classifier) to be part of a broader one. Hence, classifiers
design knowledge representation and the development and im-
are more specific as one travels down and more general or
plementation of technology nodes may refer to Gomez (1998).
abstract as one travels up the hierarchy.
The information model keeps a trace of the design knowl-
A building entity type is a special type of classifier used to
edge that was used in designing a building entity with tech-
identify each generic building entity construct as one of the
nology tags. A technology tag uniquely identifies a technology
nodes of a hierarchical decomposition defined by the concep-
node in a technology graph that has been applied to a building
tual model (an example of a hierarchical decomposition is pre-
entity. Design paths are implemented as stacks of technology
sented in the next section). Hence, a structural frame is rep-
tags.
resented as a building entity construct classified with the
Grouping building entity type ‘‘frame.’’ Each building entity type has a
number of categories associated with it. This set of categories
The group construct is used to combine a set of building may differ from one type to another. Fig. 2 shows two building
entities for the purpose of designing them together. When sev- entity types along with their categories and partial hierarchy
eral building elements possess some degrees of similarity, de- of classifiers. The example shows that a category, here ‘‘struc-
signers typically consider only the most constraining one and tural material,’’ can be shared by more than one building entity
apply that design to all elements of the group. Designers do type. Therefore, based on the definition of a building entity
this because it facilitates the design, construction, and man- type, a building entity can be classified by an arbitrary number
agement of these elements. The BENT model allows a set of of classifiers but with at most one classifier from each of the
building entities to be grouped together and designed simul- categories associated to its assigned building entity type. This
taneously. Three different types of groups have been identified:
‘‘Same’’ means that all the entities are assigned the same de-
sign components according to the most constraining entity of
the group; ‘‘similar’’ means that all the entities are designed
with the same technology node, but each is assigned design
components according to its own constraints; and ‘‘identical’’
means that all the entities in the group have the same dimen-
sions and are subject to the same conditions and hence can be
designed just by designing one of the entities. As for groups
of type ‘‘same,’’ the entities in groups of type ‘‘identical’’ are
assigned the same DU components. The grouping is done at
the level of components because one building entity can be in
several different groups based on different components. For
instance, several slabs could be grouped in such a way that
they have identical depths, the same material, and similar re-
inforcements. Each characteristic (i.e., depth, material, and re-
inforcement) corresponds to a different DU component.
CLASSIFICATION AND INDEXING
A typical design firm may easily handle dozens of building FIG. 2. Two Building Entity Types, Their Categories, and Their
projects in a 5-year span, each project consisting of thousands Partial Hierarchy of Classifiers

154 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

restriction arises from the requirement that classifiers within a of which offers a different handle to retrieve it. Hence, a search
category are mutually exclusive at a given level of the clas- can be based on a unique classifier or on the conjuction of a
sifier hierarchy. It is important to note that the scope of a set of classifiers. The classification also supports complex rea-
category is limited to its associated building entity types. soning in the forms of queries such as ‘‘is a’’ queries (e.g., is
Hence, the category ‘‘type of 3D system,’’ in the example, is this building entity a beam?), queries involving subsumption
limited to the 3D system type and is not inherited by the frame inference (e.g., is a purlin a beam?), and class partitioning
type. (e.g., a column can be short, intermediate, or slender). The
classification is independent from the information model, thus
Indexing ensuring the consistency of the information model at all times.
It allows the ad hoc expansion of classifier hierarchies without
An index is a pointer to which a keyword is assigned and affecting the information model. Another advantage of having
that leads to information about a specific topic in a large col- classification orthogonal to the information model is that mul-
lection of data. The purpose of an index is to facilitate the tiple inheritance, with its conflict problems, is avoided in the
retrieval of specific information from a large repository. There information model but supported in the classification scheme.
are two useful functions for indexes in the design of buildings,
to facilitate the retrieval of specific information in an on-going
design project and to allow the retrieval of past design solu- CONCEPTUAL MODEL
tions in a library of cases. The BENT information model defines a set of basic con-
A typical problem faced in index assignment is what ter- structs that can describe a building configuration design at the
minology to use (e.g., are you designing a cinema, a movie conceptual stage. To be useful for a particular design domain,
theater, or a motion-picture theater?). To avoid indexing prob- however, the information model has to be specialized into a
lems, keywords need to be selected from a controlled vocab- conceptual model. The conceptual model defines the design
ulary. Such a vocabulary ensures that indexes will be assigned vocabulary of a particular domain. ‘‘The objective of a con-
in a consistent manner so that information about the same ceptual model is to describe a universe of discourse that allows
subject will be found with the same keyword. Hence, to avoid the accurate expression of the designer’s conception of some
terminology problems and to have indexes organized in a sys- relevant portion of the world’’ (Eastman and Fereshetian
tematic and logical order, the BENT model relies on the clas- 1994). In other words, the conceptual model defines the set of
sification scheme presented in the previous subsection as the concepts that the designer needs to effectively design in a par-
controlled vocabulary. ticular domain using the SEED-Config design environment.
Indexing in the BENT model is done using classifier in- The procedure for creating a conceptual model for a partic-
stances. A classifier instance represents a label that classifies ular design domain consists of three steps:
and indexes the building entity it is assigned to. Each classifier
instance corresponds to only one classifier or building entity
type. A classifier instance is a distinct construct that keeps a 1. Defining the principal types of building entities and ar-
reference to the classifier it pertains to and that contains ad- ranging them into a hierarchical decomposition
ditional information relating to the design process. Classifier 2. Defining the types of data in terms of classifiers; DSRs;
instances allow the taxonomy to be separated from the design and FU, DU, and EU components that capture the in-
data. The advantage of this separation is that more than one formation required to describe each building entity
classification can be associated with a given design. Each clas- 3. Defining the design knowledge used in a set of technol-
sification is defined by a different conceptual model for a spe- ogy nodes
cific design domain. Designers have the freedom to associate
only the taxonomy they need and the capability to modify or The technology nodes must be defined last because it is nec-
append to the classification. essary to have the types of data established before defining
When a building entity is classified to be of type ‘‘wall,’’ a the processes that instantiate them.
classifier instance corresponding to that building entity type is This section presents a conceptual model specific to the con-
attached to the entity. A multifunctional building entity may ceptual design of building structures. The next two subsections
be assigned to more than one building entity type (e.g., a load- correspond to the first two steps in defining a conceptual
bearing enclosure wall). Classifiers must be selected from one model. The first subsection describes the set of structural en-
of the categories associated with each of the building entity tities and their arrangement in a hierarchical decomposition.
types. Whenever a classifier instance is assigned to a building The second subsection outlines the second step, which defines
entity, the generic entity construct becomes a member of the components, classifiers, and DSRs. The creation of technology
class defined by the corresponding classifier. Classifier in- nodes is outside the scope of this paper, but readers may refer
stances have two important purposes; they categorize the ge- to Gomez (1998) for additional information.
neric building entities into domains and subdomains as they
are being refined during the design process, and they act as Hierarchical Decomposition of Structural System
indexes in case-based retrievals and in queries for data from
the on-going design. Every design domain is concerned with a set of specific and
Indexes must be assigned to design data before a query is meaningful entities. Because designers typically break down
attempted. Hence, in the proposed approach, indexing is per- complex problems into small subproblems, it is natural to ar-
formed automatically as designers select and apply technology range these entities in a hierarchical manner that parallels this
nodes to building entities. Technology nodes assign classifier ‘‘divide-and-conquer’’ strategy. Fig. 3 shows the hierarchical
instances according to the classifiers defined in their action decomposition of a generic building project. At the root level,
part. Thus, a design is automatically query-ready, and a case there is a unique building entity, called project, which refers
is already indexed when it is added to the case library. directly or indirectly to all other building entities. The project
There are several advantages to this classification and in- is the natural repository for the client brief, the architectural
dexing approach. The classifier instances provide access to program, and all other general information pertaining to the
building entities and design information independently from project as a whole. A building project is made of four major
the design sequence that generated them. A given building systems, which represent the second level of the hierarchy
entity may have several classifier instances assigned to it, each (Rush 1986): structure, enclosure, mechanical services, and in-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 155
 FIG. 3. Hierarchical Decomposition of Building Project and Structural System

terior. The AND arc in the figure indicates that all four systems allows each aspect to be designed and analyzed separately.
are necessary. Each system is, in itself, a hierarchical decom- These two subsystems are frequently addressed together by
position of entities organized to address its specific purpose one structural system, but sometimes they may be distinct, and
and provides a unique view of the overall building. The struc- the conceptual model should be able to support both situations.
tural system is the root of all building entities that have struc- For instance, a low-rise industrial building may be designed
tural functions in a building project. It is the repository for with columns for the vertical gravity subsystem and a few
specifications, functional requirements, and constraints that af- braced frames for the vertical lateral subsystem, or it may be
fect the overall structural design. A hierarchical decomposition designed with moment resisting frames that address the func-
for the enclosure system has been proposed in Rivard et al. tion of both subsystems.
(1999). Each of the four structural subsystems consists of a set of
Fig. 3 shows six of the seven levels identified in a structural structural assemblies, which represent the next level of the
system. At the structural massing level, below the systems hierarchy. Structural assemblies are collections of basic ele-
level, the structure of a building can be broken down into ments that represent meaningful concepts in structural engi-
independent structural volumes depending on the building’s neering and that are frequently analyzed independently from
size and complexity. An independent structural volume models the rest of the structure. The decomposition shows seven struc-
a self-contained structural skeleton. Long or irregularly shaped tural assemblies: The roof, which covers the building; the
buildings may have several independent structural volumes floor, which divides a building into stories; the column stack,
that are delimited by expansion joints. An independent struc- which encompasses every aligned column element from the
tural volume may be composed of one or more functional foundation to the roof; the 3D system, which defines a core
spaces. A functional space corresponds to a portion of the vol- or tubelike structure whose primary function is to resist lateral
ume that has a specific use or occupancy that may entail spe- loading by behaving as a 3D cantilever; the frame, which is a
cific structural requirements (e.g., swimming pool or office structural assembly of linear elements connected (simply or
space with minimum column spacing). rigidly) at a finite number of discrete points called nodes; the
The next level of the hierarchy defines the structural sub- truss, which is a structural assembly of linear elements where
systems. An independent structural volume has four essential each element is subject to axial forces only; and the wall,
functional subsystems—the foundation, the horizontal, the which is a vertical plane that may extend beyond several floors
vertical gravity, and the vertical lateral subsystems. The AND and that may resist shear and carry vertical loads. A structural
arc again indicates that all four are necessary. Each subsystem assembly can be made of other structural assemblies; for in-
deals with a specific structural function: The horizontal sub- stance, a roof can be made of trusses, or a core can be made
system is responsible for transferring live, snow, and dead of walls.
loads to the vertical gravity subsystem; the vertical gravity Structural assemblies are decomposed into basic elements,
subsystem is responsible for transferring vertical loads to the which are at the sixth level. The basic elements represent the
foundation subsystem; the vertical lateral subsystem is respon- smallest construction blocks (at the conceptual design stage)
sible for transferring horizontal loads, such as wind and earth- used to assemble the structural system. These elements must
quake loads, to the foundation subsystem; and the foundation be nonintersecting geometrically so that they can be assembled
subsystem is responsible for transferring all loads of the struc- into complex structural systems (Dias 1996). The hierarchical
ture to the ground. The foundation subsystem is not elaborated decomposition has four basic elements; the slab element,
further here. The separation of the two vertical subsystems which is a planar element that makes up a roof or a floor; the
156 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
wall element, which is a vertical planar element that makes up identified in the first step. In the BENT model, building enti-
a wall; the linear element, which is a 1D element that has a ties are described with classifiers, DSRs, and attribute-value
length much greater than its cross-sectional dimensions and pairs organized into components. A conceptual model defines
may be used to represent a column, a beam, or a bar of a truss; a specialization of these constructs. Table 1 provides for a
and the node, which is a zero-dimensional element that has number of structural entities a tentative list of components and
position, but no size or form (Meyer 1995). categories of classifiers organized in terms of functional and
A seventh level, not shown in the figure, comprises con- design unit. The behavioral aspect of structures has not been
stituents. A basic element may be broken down into constitu- addressed yet. Items in bold are classifier categories, and plain
ents such as webs and flanges, or concrete and reinforcement. items are components. The table shows that most of the func-
This last level is not discussed further here because it is too tional requirements are defined at the highest level of the hi-
detailed for conceptual structural design. erarchical decomposition, whereas design specifications appear
The hierarchical decomposition of the structural system fol- at the more detailed levels. Note that dimensions, locations,
lows the total-system approach promoted by Lin and Stotes- shapes, as well as topological relations, which are important
bury (1981), where the designer focuses first on the 3D im- design information, are not shown in the table, because they
plications of architectural space-form options, then on the are obtained either from the primary spatial representation or
more-or-less planar subsystems that make up the structure, and from the geometric modeler.
finally on the elaboration and refinement of individual ele- The design unit of the linear element is described here to
ments and connection details. First, 3D structural systems are illustrate the results of the second step. The building entity
decomposed into 2D structural elements such as frames, walls, type ‘‘linear element’’ has one category and three components
and slabs. Then, 2D structural elements are further decom- associated with its design unit. The category ‘‘type of linear
posed into 1D elements or 2D subelements (e.g., beams, col- element’’ classifies the element as a beam, a column, or a
umns, and slab elements). Hence, a complex structural prob- tension member. Each of these classifiers is further specialized
lem is decomposed into simpler subproblems that can be into a hierarchy of classifiers similar to those shown in Fig. 2
considered in a semi-independent fashion. This approach fol- (e.g., a beam is specialized into a joist, a girder, or a girt). The
lows the design philosophy that the whole of a design scheme three design unit components associated with a ‘‘linear ele-
should give rise to the need for details and not vice versa (Lin ment’’ are the end conditions, which defines whether the ends
and Stotesbury 1981). of the linear element are fixed or free to rotate; the material
The hierarchical decomposition can serve as an abstraction properties, which defines the characteristics of the material that
mechanism to present different levels of detail. The designer makes up the linear element; and the section properties, which
can browse through the graph and examine any level of ab-
straction simply by going to a given depth in the graph. Typ- TABLE 1. Structural Entities Described by Components and
ically, the higher levels of the hierarchy are especially useful Categories of Classifiers
in early design stages for defining functional requirements and
for making broad design decisions, whereas the lower levels Building entity FU DU
(1) (2) (3)
are more tangible and often contain more information that is
generated in the detailed design phase (Bjork 1989). The Structural system Seismic performance cate- Main material type
gory
BENT model supports the design process as it unfolds by al- Wind load category
lowing designers to populate the design space in an intuitive Ground snow load category
manner from global system entities through subsystem entities, Soil profile type
Erection time requirement
all the way down to constituent entities. As the design pro- Budget construction cost
ceeds, the description of the design problem becomes increas- Maintenance costs
ingly detailed. Independent structural Most restrictive use/occu- Structural grid
volume pancy
The hierarchical decomposition for a design domain is kept Fire-resistance ratings
in a classification where it is represented by building entity Geometric irregularities
types. An actual building project is modeled with building en- Load combinations
tities linked by containment relationships. Each generic build- Functional spaces Use or occupancy
Minimum column spacing
ing entity is categorized by a classifier instance that corre- Linear element Type of structural action Type of linear element
sponds to one of the building entity types in the hierarchical Distributed load End conditions
decomposition. Fig. 3 shows more building entity types than Material properties
Section properties
may be needed in an actual project. A typical building project Node Point load Type of connection
would be assigned all building entity types defined at the first Restraints
four levels of the hierarchy and possibly only a subset of the
building entity types at the structural assembly and basic ele-
ment levels.
Although the hierarchical decomposition in Fig. 3 represents
an attempt to define all the types of building entities required
in the conceptual structural design stage, some types of entities
may have been overlooked or some designers may want to
include additional types of entities. The BENT model allows
hierarchical decompositions to be extended and modified by
designers. Hence, additional building entity types can be in-
corporated within the hierarchical decomposition as new ones
are identified, or a different hierarchical decomposition can be
modeled altogether.
Components, Classifiers, and Relationships in
Structural Engineering
The second step in elaborating the conceptual model is to FIG. 4. Definition of Three Design Unit Components for Linear
define the type of data required by each building entity type Element

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 157

defines the dimensions and geometric properties for the cross lows structural design and reasoning to be done not just in
section of the linear element. The attributes of these compo- terms of elements, but also in terms of systems, subsystems,
nents are shown in Fig. 4. or assemblies. The model is not limited to the terms defined
Two DSRs are defined for structural systems, support and in this research, because it can be augmented and modified by
connect. The support DSR is used to define load paths at the designers or developers to support new kinds of solutions. This
basic element level. A load path consists of a network of ad- capability could support the design of proprietary artifacts as
jacent members that transfer the loads from their origins or well as a firm’s highly specialized design practice and style.
their supports (Meyer 1995). The support DSR relates two The representation approach presented here satisfies the four
building entities, one with the role ‘‘supported,’’ and the other requirements specified in the Introduction section:
one with the role ‘‘supporting.’’ The connect DSR is required
to specify connectivity between a linear element and a node. • The approach supports design evolution by allowing de-
The connect DSR has two roles, element and node. This DSR signers to populate the design space in an intuitive manner
represents the adjacency relations between the elements and from global system entities through subsystem entities, all
nodes of a frame (Meyer 1995). This DSR is not sufficient to the way down to constituent entities. As the design pro-
completely describe the relationship between a linear element ceeds, the description of each building entity becomes in-
and a node. This is why the DU component ‘‘end conditions’’ creasingly detailed through the addition of sets of attri-
is required to record the additional information concerning the bute-value pairs, indexes, geometry, and relationships.
relationship. Here, the component specifies whether each end • The approach supports design exploration by allowing dy-
of the element is free or fixed to its respective node. Note that namic changes to occur not just in terms of attribute-value
the connect and support DSRs have domain-related semantic pairs and building entities, but also in terms of model
meanings that differ from the topological relationships that can structure. Design exploration is also supported by how the
be obtained directly from the geometric modeler. Other DSRs approach represents design knowledge. The technology
may emerge as the conceptual model is refined. nodes support the generation of solutions in staged steps
so as to allow backtracking and the generation of alter-
CONCLUSIONS native design solutions. Furthermore, building entities are
richly indexed with classifier instances, which allow de-
The BENT information model specifically addresses the sign information to be easily queried and browsed.
early design stages by being capable of representing rapidly • The approach supports multiple views by allowing a
generated conceptual design alternatives, sharing data among building entity to pertain to more than one design domain.
all design participants, and serving as a case library of previ- A given building entity, which is described geometrically
ous designs so that relevant cases can be recalled and adapted by a single primary spatial representation, can be part of
as elements of the solution of new design problems. The more than one hierarchical decomposition and can encap-
BENT model has only one generic class that provides a stan- sulate components, classifier instances, secondary geo-
dardized interface for all the entities of a building. This generic metrical representations, DSRs, and technology tags that
container glues together the modular information of a building may originate from more than one design domain. Some
entity, including the properties, taxonomy, geometry, compo- of the components within the building entity may be
sition, relationships, and generating design knowledge of the shared among different views. Hence, the BENT model
building entity. The proposed representation provides the foun- provides a common underlying data structure that can be
dation for a novel design environment that separates design shared among designers, and applications.
representation, design knowledge, classification, and geometry. • The approach is also extensible because it defines a set
In the proposed approach, a classification is associated with of basic constructs that can be specialized into a concep-
each design project. The classification defines the model of the tual model that specifically addresses a particular design
hierarchical decomposition (for a given design domain) using domain. A conceptual model can be created or augmented
building entity types, each of which has an associates set of by designers through the definition of new building entity
categories with related hierarchies of classifiers. The classifi- types, components, classifiers, relationships, and technol-
cation is used as a controlled vocabulary from which indexes ogy nodes.
are obtained. The controlled vocabulary ensures consistent in-
dex assignment and avoids terminology problems. Indexing is Two important issues still need to be investigated before the
performed automatically as designers select and apply tech- building design representation of the SEED-Config prototype
nology nodes to building entities. Such an automatic indexing can support a fully operational design environment. The first
is a necessity due to the amount of data required to represent issue is concurrent use. Because changes are usually localized
buildings. The capability to classify and index design data with to the component levels, a concurrent system that applies locks
user-supplied symbolic classifiers has not been found in design at the component level can be envisioned. The second issue is
environments either in industry or in the literature. the management of design views. In a collaborative design
Support for a particular design domain is developed by spe- environment that encompasses the perspective of many design
cializing the BENT information model into a conceptual participants, there must be a functionality to provide and dis-
model. A conceptual model consists of the definition of the play only the data designers need. Designers should be able
classifications, technology nodes, and the types of entities, re- to focus on only one building system when viewing the 3D
lationships, and data to be used in that domain. The conceptual geometric model, the hierarchical decomposition of the build-
model is capable of representing complex design information ing, or the data of an entity that is of interest to them. Such
and relationships as well as providing a clear mapping between views should be defined dynamically by the designer through
the portion of the world of interest and its representation. The the selection of a particular building system.
solution space in a particular domain is limited by the expres- The proposed approach can be accepted industrywide only
siveness of the design vocabulary defined by the conceptual if a shared and standardized conceptual model is defined. This
model. A conceptual model has been defined that specifically standard would define the classifications and the types of en-
addresses the structural engineering view in the early stages tities, data, and relationships that would provide the basis for
of building design. This model can describe a wide variety of information exchange among the professionals of a particular
structures and can capture the more abstract concepts of struc- domain. Defining such a standardized conceptual model
tural systems as well as their refinements. This approach al- would represent a major endeavor, but would result in a much
158 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
leaner standard than STEP or IFC, because the BENT model erarchies.’’ PhD thesis, Dept. of Civ. and Envir. Engrg., Carnegie Mel-
provides the foundation upon which standardized conceptual lon University, Pittsburgh.
Howard, H. C., Abdalla, J. A., and Phan, D. H. D. (1992). ‘‘Primitive-
models can be shared. Furthermore, such standardized con- composite approach for structural data modeling.’’ J. Comp. in Civ.
ceptual models are not rigid and closed, because designers or Engrg., ASCE, 6(1), 19–40.
developers can expand them. Kiviniemi, A. (1999). ‘‘IAI and IFC—State of the art.’’ Proc., 8th Int.
Conf. on Durability of Build. Mat. and Components, M. A. Lacasse
ACKNOWLEDGMENTS and D. J. Vanier, eds., Vol. 4, NRC Research Press, Ottawa, 2157–
2168.
The writers wish to thank the Natural Sciences and Engineering Re- Lin, T. Y., and Stotesbury, S. D. (1981). Structural concepts and systems
search Council of Canada, the Fonds pour la Formation de chercheurs et for architects and engineers, Wiley, New York.
l’aide à la recherche (Fonds FCAR) of the province of Quebec, and the Meyer, S. (1995). ‘‘A description of the structural design of tall buildings
Engineering Design Research Center of Carnegie-Mellon University for through the grammar paradigm.’’ EDRC Tech. Paper 02-35-95, PhD
financially supporting this research. The writers would like to extend thesis, Dept of Civ. and Envir. Engrg., Carnegie Mellon University,
special thanks to Ulrich Flemming, Prof., School of Architecture, Car- Pittsburgh.
negie Mellon University and Nestor Gomez, Asst. Prof., Dept. of Civ. Phan, D. H. D., and Howard, H. C. (1993). ‘‘The primitive-composite (P-
and Envir. Engrg., Florida International University, for their support and C) approach—A methodology for developing sharable object-oriented
invaluable input to this work. data representations for facility engineering integration.’’ Tech. Rep. 85
(A and B), Center for Integrated Facility Engineering (CIFE), Stanford
APPENDIX. REFERENCES University, Stanford, Calif.
Rivard, H. (1997). ‘‘A building design representation for conceptual de-
Biedermann, J. D., and Grierson, D. E. (1995). ‘‘A generic model for sign and case-based reasoning.’’ PhD thesis, Dept. of Civ. and Envir.
building design.’’ Engrg. with Comp., London, 11(3), 173–184. Engrg., Carnegie Mellon University, Pittsburgh.
Bjork, B.-C. (1989). ‘‘Basic structure of a proposed building product Rivard, H., Bédard, C., Ha, K. H., and Fazio, P. (1999). ‘‘Shared con-
model.’’ Comp.-Aided Des., Oxford, U.K., 21(2), 71–78. ceptual model for the building envelope design process.’’ Build. and
Bjork, B.-C. (1992). ‘‘A unified approach for modelling construction in- Envir., Oxford, U.K., 34(2), 175–187.
formation.’’ Build. and Environment, Oxford, U.K., 27(2), 173–194. Rivard, H., and Fenves, S. J. (2000). ‘‘SEED-Config: A case-based rea-
Bjork, B.-C. (1994). ‘‘RATAS project—Developing an infrastructure for soning system for conceptual building design.’’ Accepted for publica-
computer-integrated construction.’’ J. Comp. in Civ. Engrg., ASCE, tion in Artificial Intelligence for Engrg. Des., Anal. and Manufacturing,
8(4), 401–419. Cambridge, U.K.
Booch, G., Rumbaugh, J., and Jacobson, I. (1999). The unified modeling Rivard, H., Gomez, N., and Fenves, S. J. (1996). ‘‘An information model
language user guide, Addison Wesley Longman Inc., Reading, Mass. for the preliminary design of buildings.’’ Proc., CIB W78/TG10 Work-
Burkett, W. C., and Yang, Y. (1995). ‘‘The STEP integration information shop ‘‘Constr. on the Information Hwy.,’’ Publ. 198, Z. Turk, ed., In-
architecture.’’ Engrg. with Comp., London, 11(3), 136–144. ternational Council for Building, Rotterdam, The Netherlands, 435–
Dias, W. P. S. (1996). ‘‘Multidisciplinary product modeling of buildings.’’ 448.
J. Comp. in Civ. Engrg., ASCE, 10(1), 78–86. Rosenman, M. A., and Gero, J. S. (1996). ‘‘Modelling multiple views of
Eastman, C. M. (1992). ‘‘A data model analysis of modularity and ex- design objects in a collaborative CAD environment.’’ Comp.-Aided
tensibility in building databases.’’ Build. and Envir., Oxford, U.K., Des., Oxford, U.K., 28(3), 193–205.
27(2), 135–148. Rush, R. D., (1986). The building systems integration handbook, Butter-
Eastman, C. M., and Fereshetian, N. (1994). ‘‘Information models for use worth-Heinemann, Stoneham, Mass.
in product design: A comparison.’’ Comp.-Aided Des., Oxford, U.K., Schodek, D. L. (1980). Structures, Prentice-Hall, Englewood Cliffs, N.J.
26(7), 551–572. Simon, H. A. (1996). The sciences of the artificial, 3rd Ed., MIT Press,
Fenves, S. J., Rivard, H., Gomez, N., and Chiou, S.-C. (1995). ‘‘Concep- Cambridge, Mass.
tual structural design in SEED.’’ J. Arch. Engrg., ASCE, 1(4), 179– Woodbury, R., and Chang, T.-W. (1995). ‘‘Massing and enclosure design
186. with SEED-Config.’’ J. Arch Engrg., ASCE, 1(4), 170–178.
Flemming, U., and Woodbury, R. (1995). ‘‘Software environment to sup- Woodbury, R. and Fenves, S. J. (1994). ‘‘SEED-Config requirements
port early phases in building design (SEED): Overview.’’ J. Arch. analysis.’’ SEED team, Carnegie-Mellon University and University of
Engrg., ASCE, 1(4), 147–152. Adelaide, 具http://seed.edrc.cmu.edu/SC/requirements-new/SC-req-new.
Gielingh, W. (1988). ‘‘General AEC reference model.’’ ISO TC 184/SC4/ book.html典 (June 1997).
WG1 doc 3.2.2.1, TNO Rep. BI-88-150, International Organization for Zamanian, K. M., and Fenves, S. J. (1994). ‘‘A referential scheme for
Standardization, Geneva. modeling and identifying spatial attributes of entities in constructed
Gomez, N. (1998). ‘‘Conceptual structural design through knowledge hi- facilities.’’ Res. in Engrg. Des., 6(3), 142–168.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 159

 APPLICATION FRAMEWORK FOR DEVELOPMENT OF
SIMULATION TOOLS
By Dany Hajjar1 and Simaan M. AbouRizk,2 Member, ASCE

ABSTRACT: Development of specialized construction simulation tools is a time-consuming and complex op-
eration. This has hindered the effective and wide use of simulation in the construction industry. An object-
oriented application framework is a promising software engineering principle, which can help in overcoming
this obstacle. Application frameworks encapsulate reusable code as well as common design patterns for a specific
class of applications so that development of new tools is greatly simplified. This paper describes this new
principle and applies it to the development of a construction simulation object library. A case study performed
to compare the efforts between framework- and nonframework-based simulation tool development revealed that
productivity can be dramatically improved.

INTRODUCTION from the ideas of frameworks is computer simulation. Com-

puter simulation allows for the modeling, analysis, and opti-
Computer-based drawing, cost estimating, scheduling, con- mization of construction operations. It is a proven technology,
struction, and project control systems have infiltrated from the and countless studies have been performed to demonstrate its
smallest to the largest of construction companies. As individ- applicability to the analysis of a wide range of construction
ual companies strive to be more innovative and competitive, operations including aggregate production, earthmoving, min-
custom software is continually being developed and refined to ing, tunneling, and precast concrete manufacturing. Previous
support the varying business processes. Due to the relatively studies have identified how customization of these simulation
large investment required for such software development proj- systems to a specific domain can greatly increase their appeal
ects, it has become important to greatly reduce the time and and utilization within the construction industry. However, the
effort involved. Traditional efforts to reduce the development actual development of such tools remains complex and time-
effort have focused on software code reuse, typically in the consuming (AbouRizk and Hajjar 1998).
form of functional libraries developed in-house or as com- This paper presents a framework for the development
mercial products developed by third parties. of construction simulation tools. The novel aspects of this
Recent trends in software engineering research have dem- contribution are the presentation of identified common design
onstrated that effective software reuse must include both com- patterns and components in construction simulation tools as
mon code as well as common designs. Design reuse simplifies well as the application of a proven software engineering
the design component of software development projects and principle to the construction software development arena. The
allows for more effective code reuse. proposed framework captures the recurring designs and
An object-oriented application framework, referred to as code components encountered during the development of sev-
‘‘application framework’’ or ‘‘framework’’ in the rest of this eral construction simulation tools. The advantages of this ap-
paper, is the term used to describe such a concept. It is im- proach include the reduced development time for new tools,
portant to note that this term is used in a very specific context and the standardization of both the internal structure and ex-
according to the software engineering definition. A framework ternal interface of new tools leading to increased user-friend-
is a set of object-oriented classes encapsulating common de- liness.
sign patterns and code components for a specific application This paper is organized as follows. A general overview on
domain (Deutsh 1989). computer simulation as well as construction simulation is pre-
Frameworks have been applied to the engineering discipline sented first followed by further background on object-oriented
to a limited extent. Forde and Stiemer (1989) used it in the application frameworks in general. A presentation of a vari-
development of a generic environment called GenApp for the ability analysis that reveals the identified common design pat-
construction of interactive and graphical engineering software. terns and components of construction simulation tools follows.
Although GenApp’s functionality is now available as part of This analysis was used in the design of the proposed frame-
today’s graphical operating systems such as Linux and Micro- work. Implementation issues are then briefly discussed, fol-
soft Windows, the research did demonstrate the effectiveness lowed by a presentation of the results of two case studies per-
of using frameworks for an engineering-oriented application. formed to validate and test the effectiveness of the framework
A framework called SEAF (‘‘SEAF’’ 1999) has been devel- approach.
oped for supporting worksheet-based engineering applications.
SEAF was used in the development of a tool for the sizing COMPUTER SIMULATION
and selection of pressure safety valves called SizeMaster
(‘‘SizeMaster’’ 1999). Computer simulation is defined as the process of designing
One class of computer applications that can greatly benefit a mathematical-logical model of a real-world system and ex-
perimenting with the model on a computer (Pristker 1986).
1
Res. Engr., Dept. of Civ. and Envir. Engrg., Univ. of Alberta, Ed-
Early simulation users were required to build a model by writ-
monton, AB, Canada T6G 2G7. E-mail: [email protected] ing programming code, mainly in FORTRAN, and experi-
2
Prof., Dept. of Civ. and Envir. Engrg., Univ. of Alberta, Edmonton, menting by directly manipulating the computer program. This
AB, Canada T6G 2G7. E-mail: [email protected] was followed by the invention of simulation specific program-
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, ming environments where users write simulation specific code
2000. To extend the closing date one month, a written request must be or access a provided function library. ‘‘Modeling’’ is the term
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on June 10, 1999. This
used to describe the process of specifying a given simulation
paper is part of the Journal of Computing in Civil Engineering, Vol. model.
14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0160–0167/ Simulation systems were later introduced that allowed for
$8.00 ⫹ $.50 per page. Paper No. 21178. alternative model development. This meant that modelers no
160 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 1. Sample Discrete Event Representation

longer had to write code directly. Graphical modeling made it CONSTRUCTION SIMULATION
possible to define the simulation model by creating, manipu-
lating, and linking a number of available basic building blocks. Halpin (1977) popularized the use of simulation in construc-
This meant that users no longer had to be proficient in pro- tion research with his development of a system called CY-
gramming. A detailed account of the history of simulation con- CLONE (CYCLic Operation NEtwork). CYCLONE allowed
cepts and systems is detailed in Kreutzer (1986). the user to build models using a set of abstract but simple
There are many ways of modeling a given problem. These constructs. The system became the basis for a wide range of
generally fall into two categories: continuous and discrete- construction simulation research efforts with the objective of
event. With continuous modeling, differential equations are enhancing the basic system functionality. Although CY-
used to describe the progress of a given activity. However, CLONE and its derivatives introduced a wider academic au-
when mathematical modeling is not possible, the discrete- dience to computer simulation, its use in the industry was very
event approach is utilized. limited. This was because CYCLONE proved practical for
Discrete-event simulation views a model as a set of events small-scale applications. Furthermore, the modeling process
and transitions. Transformations are processed as part of event required simulation training—something that the industry was
handlers, and no relevant transformations are assumed to occur not generally ready to invest in. As a result, research turned
during transitions. Entities represent the active elements of the to other concepts that could address these issues and simplify
model. They travel throughout the event network and trigger the modeling process even further.
transformations. An important component of a discrete-event AbouRizk and Hajjar (1998) found that effective transfer of
simulation is the simulation event monitor, which is respon- computer simulation knowledge to the construction industry is
sible for managing the event calendar. The described concept best done through specialization and customization of the
is now illustrated below with a simple example. modeling, analysis, and reporting components of the simula-
Assume that trucks arrive at a concrete batch plant and await tion systems. This approach, known as special purpose simu-
their turn for loading fresh concrete from a mixer. When they lation (SPS), was followed in the successful development and
finish loading, they proceed to their destination. The described implementation of three custom simulation tools. For the mod-
model is illustrated using an event diagram as shown in eling environment, this means that the modeling elements that
Fig. 1. are used to construct simulation models are tailored to the
The ‘‘truck arrival’’ event is first added to the event calendar specific target domain.
during the initialization of the simulation. When the event is The first tool, called AP2-Earth (Hajjar and AbouRizk
processed as part of the event handler, the arrival of the next 1996), allows for the analysis of large earthmoving projects.
truck entity is first scheduled to occur. Then the status of the AP2-Earth was developed with the objective of providing
mixer is checked. If it is busy, then the truck entity is added earthmoving contractors with a system that can automate the
to a queue. Otherwise, the ‘‘truck departure’’ event is sched- traditionally time-consuming and manual process of preparing
uled. During the processing of the Truck Departure event, earthmoving estimates for large and complex projects. The
which occurs when a loading operation has been completed, main interface form shown in Fig. 2 is where the simulation
the status of the waiting queue is checked. If there are any models can be built. Standard earthmoving modeling elements
waiting truck entities, then the first one is removed from the such as sources, trucks, placements, and road segments are
queue so that it can start loading. The transition duration be- used to build general earthmoving scenarios.
tween each truck arrival event is based on the desired inter- The second tool, called CRUISER (Hajjar and AbouRizk
arrival time of trucks. The transition duration between the 1998), can be used for modeling aggregate production plants.
truck arrival and truck departure events depends on the desired It allows users to define the plant layout by creating the crush-
truck loading time. ing components and visually connecting them with conveyers
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 161
 FIG. 2. Sample Simulation Model in AP2-Earth

FIG. 3. Sample Simulation Model in CRUISER

as shown in Fig. 3. The objective of CRUISER is to provide viewed in the form of 2D cross-sectional graphs or 3D surface
managers with a tool to experiment with several possible al- charts.
ternatives for plant design. With this program, ‘‘what-if’’ anal- Although each tool was successfully implemented, it was
ysis can be performed very easily by changing the desired observed that the relatively large initial investment required
parameters and resimulating. This allows plant managers to for their development would hinder the application of the SPS-
test the effect of changing screen sizes, using an alternative based approach to other construction operations. As a result,
pit and countless other design alternatives, by simply changing research efforts were shifted to the identification of software
the CRUISER model and resimulating. The application takes engineering principles that can lead to improvements in the
care of all the tedious analysis that would otherwise be done actual development process of such tools.
by hand.
The third tool was CSD (Hajjar et al. 1998), which allows OBJECT-ORIENTED APPLICATION FRAMEWORKS
for the optimization of construction site dewatering operations.
Overview
The visual modeling environment allows the user to enter var-
ious site data such as the site dimensions, layers, and other Object-oriented application frameworks encapsulate a given
properties. Pumping wells and watch wells are added using design philosophy and a set of components that can be used
simple click and drag mouse operations. Watch wells represent to develop specific applications for a given domain. This
the desired drawdown at various points in the site. The sim- means that overall development of systems becomes a two-
ulation engine of CSD can then be initiated to predict the step approach. Numerous expert developers, familiar with a
resultant water table level throughout the site. Results are specific application domain, first develop the application
162 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
framework. Examples of specific application domains are cost is typically an object-oriented library that encapsulates a basic
estimating systems, scheduling systems, and construction sim- tool architecture as well as various specialized components.
ulation systems. The developed application framework cap- These components, also referred to as services, are grouped
tures the broad and recurring business processes, transactions, by their function and represent the basic infrastructure of the
and transformations within the target domain. Once the appli- framework.
cation framework is developed, specific tool implementations Once the framework implementation is complete, it is tested
within the domain can then be developed. For example, a spe- through the implementation of several basic tools. Ideally,
cific implementation based on an estimating framework could some of the tools originally used for the variance analysis are
include an earthworks estimating systems, a bridge estimating reimplemented. This proves the initial practicality and expres-
system, or a multistory commercial building estimating sys- siveness of the developed application framework.
tem. Several issues are related to the deployment of application
Understanding of application frameworks has greatly ad- frameworks after their implementation and testing is com-
vanced in recent years. The following subsection will provide pleted. First, the framework needs to be introduced in a careful
an overview of the basic concepts involved in framework de- manner, so that developers can gradually learn its basic re-
sign, implementation, documentation, and deployment. The in- quirements, features, and components. This process can in-
formation provided is based on application framework design clude rollout sessions, example applications, extensive docu-
and development guidelines as outlined by Froehlich et al. mentation, and tutorials. Second, the distribution means of the
(1998). framework must be addressed. Possible options include source
code or binary code distributions. Third, a process must be
Framework Development established to capture important feedback from the tool de-
The first step in the design of an application framework is velopers. This includes general support information, bug re-
the determination of its intended domain. This helps in defin- ports, and enhancement requests.
ing the scope of the framework and the basic services that it
should provide. Variability analysis is a technique that can be
used to assist in this process. This involves assessing various CONSTRUCTION SIMULATION APPLICATION
components of applications across a number of scenarios and FRAMEWORK
identifying both variable and common features. Scenarios may
include general descriptions, formal process models, or, ide- Overview
ally, actual tools developed for the intended domain. Variabil-
ity analysis results in the identification of the recurring design A framework approach is followed in the development of
patterns and common features. an object-oriented library that can be used for the development
Common features translate into what is referred to as the of individual construction simulation tools. The design of this
‘‘frozen spots’’ of the framework. Frozen spots are framework library is based on a detailed analysis of the three stand-alone
components that tool developers have little or no control over. simulation tools previously described. This variability analysis
Distinct or variable features translate into what is referred to reveals that a large proportion of features were identical in all
as the ‘‘hot spots’’ of the framework. Hot spots are areas in tools; in fact, code had been reused for these features. Fur-
the framework that tool developers can customize for specific thermore, of all the remaining features, the majority do not
tool requirements. vary significantly, and it was possible to generalize their func-
Once the design of the framework is complete, the imple- tionality to standardize their implementation. These common
mentation can begin. Often, the process is iterative rather than features translate into the frozen spots of the framework. The
linear in that implementation may require the redesign of cer- remaining features are genuinely unique to the specific tools.
tain components, as the domain becomes more understood. They are used in the identification of the hot spots of the
During this process, the means by which developers will be framework. The following two subsections will discuss the
customizing the framework are also determined. The end result common and variable features in detail.

FIG. 4. General System Architecture of SPS Tools

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 163

Common Features Statistical Collection and Analysis

Overview Statistical collection and analysis involves the collection of

observed values during simulation execution followed by the
The recurring system architecture, which was observed to actual analysis at the end of the simulation. Analysis of col-
occur in the three construction SPS tools, is illustrated in Fig. lected observations depends on whether a given statistic is
4. The ‘‘modeling elements’’ are the domain specific model considered intrinsic or nonintrinsic. During the collection of
building blocks available to the user. These elements are used the observations, the simulation time also is tracked. A statistic
by the user when building simulation models. The ‘‘simulation declared as nonintrinsic will cause the analysis to ignore this
engine’’ is a discrete-event processor responsible for schedul- time component and perform simple analysis based only on
ing and processing of the model generated events. During the collected data values. Otherwise, the calculations will be
event processing, several services may be accessed including biased by the time value of the collected statistic. Intrinsic-
the ‘‘statistical collection and analysis’’ service and the ‘‘ran- type statistics are typically used for variables that indicate a
dom number sampling’’ service. The ‘‘trace manager’’ is used certain state of the model such as the current file length and
to provide feedback on simulation progress and encountered the number of available resources, whereas nonintrinsic statis-
errors. The ‘‘project plan information manager’’ provides a tics are used for performance-type variables such as cycle
mechanism for translating simulation results into a project plan time, travel time, and waiting time.
in a standard data format. A project plan includes a schedule,
a budget, resource utilization information, and cash flow pro- Trace Manager
jection. During the modeling or simulation execution pro-
cesses, external databases may be accessed for retrieval of This component is used to track simulation progress in de-
standard information, or generation of results into external sys- tail or for reporting certain runtime errors. Traced messages
tems. Simulation model data are persisted using a ‘‘model per- are typically grouped by various categories, which include
sistence manager,’’ which allows for the storage and retrieval simulation, warnings, integrity violations, and runtime errors.
of all model related data including model layouts, statistical Simulation messages are typically used by developers to mon-
analysis results, and generated project plans. The graphical itor the detailed workings of their simulation code and to de-
user interface (GUI) acts as the view and controller in a model- bug their applications. Warnings are noncritical errors, which
view-controller-type user interface pattern (Goldberg and Rob- provide information as to possible inconsistencies in the sim-
son 1983). The following subsections will provide more details ulation model. Integrity violations are used to report model
regarding each component. construction errors, which must be corrected prior to the sim-
ulation execution, and runtime errors are critical errors gen-
Modeling Elements erated from the specific tool code or from the framework code
itself.
The modeling elements of a given SPS tool are domain-
based simulation building blocks that map real physical or
logical components of the target process. For example, an ag- External Database Access
gregate production plant model is built using a set of modeling A great factor in the successful implementation of SPS tools
elements that include crushers, screen decks, and sieve anal- is their ability to integrate with existing systems. This reduces
ysis nodes. Although the general structure of a given modeling the overall data entry requirements and allows the simulation
element is uniform across various tools, the actual implemen- results to be viewed directly as part of systems to which users
tation of each modeling element varies considerably between are already accustomed. For example, to make AP2-Earth ap-
modeling elements within a given tool and across tools. This pealing to an earthmoving contractor, it was given the capa-
is the biggest and most fundamental variable feature that dis- bility to generate information for use by an estimating module.
tinguishes a given SPS tool from another. As a result, access to external databases is identified as a com-
mon and important feature of SPS tools.
Simulation Engine
The simulation engine encapsulates a discrete-event proces- Project Plan Information Generation
sor responsible for managing all entities, resources, files, and
scheduled events. At the core of a simulation engine is the It has been discovered, through efforts to integrate earth-
event calendar, which is used for the management of the sim- moving and aggregate production models with estimating and
ulation events. The simulation event loop first removes the scheduling systems, that a standard for data generation and
next event to be processed from the event calendar and ad- exchange is required. This standard should specify not only
vances the current simulation time to the time of the event. the format of the transferred data but also the classes of data
The event is then routed to the modeling element instance that that should be generated from every SPS tool. By following
originally scheduled it. During simulation processing, in- this strategy, a single extraction utility can be developed for
stances may schedule further events, which are in turn inserted each type of external system. The generated information con-
into the appropriate calendar position. This process continues sists of basic facts about the occurrence of planning related
until all events in the calendar have been processed or until events during the simulation. The interpretation and analysis
simulation time reaches a maximum specified by the user. of the data is the responsibility of external integrated systems.
As a result, a standard approach is proposed whereby the sim-
Random Number Sampling ulation engine will generate information that falls under eight
categories that are described in Table 1.
This component provides support for random number gen-
eration and transformation into several standard stochastic dis- Simulation Model Persistence
tributions. The random number generator produces uniformly
distributed values between 0 and 1. Several functions use this The persistence manager is responsible for the storage and
base value to produce common transformations such as beta, retrieval of simulation model representations, tracing mes-
exponentially, normally, triangularly, and uniformly distributed sages, and generated project plan information. Simulation
variates. model representations include the model layout, which is the
164 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 TABLE 1. Project Plan Information Categories interface controls; and line and bitmap primitives such as line,
Information text and bitmap routines.
category Description Tracing information, which is managed by the tracing man-
(1) (2) ager, is also viewed graphically using a tabular form. This
allows users to filter and sort the messages according to their
Activity listing Basic activities that will take place during
simulation categories.
Activity resources Resources for each activity that will be Generated project-planning information is designed for use
utilized mainly by external systems. However, it is believed that users
Activity cost centers Cost categories or centers for which cost can also benefit from the ability to quickly view a summary
will be incurred of the generated information in a simple format prior to the
Activity schedule Records that indicate start and stop times final analysis by external systems. As a result, the GUI in-
of each activity
Activity production Records indicating quantity of work rec- cludes a number of project planning display forms including
ognized for a given activity at a given the summary form shown in Fig. 5. The left side displays a
simulation time tree that contains a hierarchical view of the project activities.
Activity revenue Records indicating dollar value collected At the highest level is the project itself. Selecting a given
for revenue generating activity at a activity from this tree will display its related information on
given simulation time the sheet shown on the right side of the figure.
Activity resource utilization Records indicating change in utilization of
a particular resource by a given activity
at a given simulation time Variable Features
Activity cost Records indicating cost incurred by a par-
ticular cost center on a given activity at The main variable features across SPS tools are the set of
a given simulation time available modeling elements and the implementation of the
behaviors of each modeling element. This ability to identify
the exact nature of variation across tools is what makes frame-
set of modeling element instances and their internal state data, works such an ideal and natural approach to SPS tool devel-
and the current statistical analysis results. opment. Using framework terminology, the modeling element
Although the various tools utilized different means of per- is the hot spot. This hot spot is designed as a base generic and
sisting data including text, binary, and database approaches, it intelligent object-oriented class, which developers use as the
is decided that a relational database approach is most suitable. basis for specific modeling element developments. Using ob-
The advantages include the popularity and wide use of such ject-oriented language, the generic base modeling element is
systems, the native support for security and transaction man- a virtual base class that encapsulates data structures repre-
agement, and the seamless support for multiuser access to sim- senting a host of information including graphical representa-
ulation models. tion, parameters, outputs, results, and simulation resources.
Class methods provide the means to manipulate these struc-
Graphical User Interface tures and to perform other operations. Class events provide
feedback as to changes in the state of an element. Class prop-
This component provides all the required functionality to erties, methods, and events that are related are grouped to-
support a graphical user interface for controlling the modeling gether as part of a class behavior. Based on this approach,
and simulation system in a standard and consistent manner. construction simulation tool development is reduced to the
SPS tools require a wide range of dialog boxes; forms; user process of declaring the set of required modeling elements and

FIG. 5. Project Plan Summary Form

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 165

overriding the default implementation of the desired class implemented as an ActiveX library using Microsoft Visual
methods. Basic.

Framework Implementation VALIDATION OF DEVELOPED FRAMEWORK

Overview
The discussed variability analysis results are used in the
design of a framework and its implementation as an object To test the hypothesis that development time is dramatically
library called the Simphony Object Model (SOM). The iden- shortened as a result of the framework-based approach, two
tified common features translate into the main framework of the previously mentioned tools, AP2-Earth and Cruiser,
components and a definition of the generic base modeling el- were redeveloped by the original developer using the SOM.
ement. A simplified representation of the framework structure The original development duration was then compared with
is shown in Fig. 6. The previously identified common features the time it took for the redevelopment.
are grouped into the services of the framework. The generic
base modeling element encapsulates a default representation AP2-Earth Redevelopment
and is used by developers to create custom simulation tools.
The SOM is implemented as part of a complete construction A new SOM-based simulation tool, denoted CEM㛭EMS, is
simulation tool development and utilization platform called developed with most of the functionality provided by AP2-
Simphony. The overall platform includes several programs in- Earth. Users are able to model preparation, loading, hauling,
cluding a custom integrated development environment that dumping, and spreading operations and to generate results in
provides controlled access to the SOM. The SOM itself is a similar manner to AP2-Earth.
The CEM㛭EMS template consists of 16 modeling elements
that can be used to define the various components of an earth-
moving operation. Most of these elements have attributes that
users can manipulate to modify the outcome of the simulation.
A sample model based on this template is shown in Fig. 7.
Analysis of historical records showed that the original de-
velopment of the AP2-Earth project was equivalent to one per-
son year. This involved time spent learning the fundamentals
of earthmoving operations, site visits, company interviews, site
deployment, documentation, and integration. The actual design
and programming time, which involved C⫹⫹ coding, data-
base development, and report design and construction, was
approximately 400 h. This value is compared with the frame-
work-based approach, which resulted in a total tool develop-
ment duration of 28 h.

CRUISER Redevelopment
In a similar manner, a SOM-based tool, denoted CEM㛭
FIG. 6. Simplified Representation of Simphony Application CRUSH, was developed with the same functionality as
Framework CRUISER. Fourteen modeling elements were developed to al-

FIG. 7. Sample Model Layout Based on CEM㛭EMS SPS Tool

166 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

low for the modeling of aggregate production operations in- ACKNOWLEDGMENTS
cluding size separation and reduction. This work was funded by a number of construction companies in Al-
Historical records revealed that total project investment was berta and the Natural Science and Engineering Research Council of Can-
equivalent to one and a half person years of which 650 h were ada under Grant No. IRC-195558/96.
actual design and programming time. The remaining time was
APPENDIX. REFERENCES
spent on process discovery, documentation, and implementa-
tion. The development of the equivalent framework-based tool AbouRizk, S., and Hajjar, D. (1998). ‘‘A framework for applying simu-
lation in the construction industry.’’ Can. J. Civ. Engrg., Ottawa, 25(3),
took approximately 40 h. 604–617.
Deutsch, L. P. (1989). ‘‘Design reuse and frameworks in the Smalltalk
80 System.’’ Software reusability. Vol. II: Applications and experience,
CONCLUSIONS Ted J. Biggerstaff and Alan J. Perlis, eds., Addison-Wesley, Reading,
Mass., 57–71.
Forde, B. W., and Stiemer, S. F. (1989). ‘‘Development of engineering
This paper discussed the application of object-oriented ap- software within a generic application framework.’’ Microcomputers in
plication framework concepts to the construction simulation Civ. Engrg., 4(3), 205–216.
domain. An application framework approach greatly reduces Froehlich, G., Hoover, J., Liu, L., and Sorenson, P. (1998). ‘‘Designing
object-oriented frameworks.’’ CRC handbook of object technology,
the development effort and allows for the standardization of CRC, Boca Raton, Fla.
the generated tools. Application frameworks represent the nat- Goldberg, A., and Robson, D. (1983). Smalltalk-80, Addison-Wesley,
ural progression from archaic code reuse, to the development Reading, Mass.
of reusable code libraries, to the development of structured Hajjar, D., and AbouRizk, S. (1996). ‘‘Building a special purpose simu-
lation tool for earth moving operations.’’ Proc., 1996 Winter Simulation
object libraries that include reusable designs. It also is strongly Conf., ASCE, New York, 1313–1320.
believed that they can be equally effective in other construc- Hajjar, D., and AbouRizk, S. (1998). ‘‘Modeling and analysis of aggregate
tion management domains. production operations.’’ J. Constr. Engrg. and Mgmt., ASCE, 124(5),
The proposed framework was designed based on a vari- 390–401.
Hajjar, D., AbouRizk, S., and Xu, J. (1998). ‘‘Optimizing construction
ability analysis that identified the recurring design patterns and site dewatering operations using CSD.’’ Can J. Civ. Engrg., Ottawa,
components of special purpose simulation tools in construc- 25(3), 819–828.
tion. Based on the design, the implementation of new tools Halpin, D. W. (1977). ‘‘CYCLONE—Method for modeling job site pro-
was reduced to a simple process of modeling element decla- cesses.’’ J. Constr. Div., ASCE, 103(3), 489–499.
Kreutzer, W. (1986). System simulation programming styles and lan-
ration and implementation. guages, Addison-Wesley, Don Mills, Ontario, Canada.
Two case studies were performed and indicated that the Pristker, A. A. B. (1986). Introduction to simulation and SLAM-II, Wiley,
framework-based approach can lead to a dramatic simplifica- New York.
tion and shortening of the development effort of construction ‘‘SEAF project.’’ (1999). World Wide Web html document, 具www.cs.
ualberta.ca/⬃softeng/oldpages/SEAF/project.shtml典.
simulation tools. These two case studies, which involved the ‘‘SizeMaster Mark IV software overview.’’ (1999). World Wide Web
redevelopment AP2-Earth and CRUISER, showed how devel- html document, 具www.avrasoft.com/prothos/sizemaster.x/overview/
opment time was cut by a factor of 15. overview.htm典.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 167

 EVALUATION OF MOBILE MAPPING SYSTEMS FOR ROADWAY
DATA COLLECTION
By Hassan A. Karimi,1 Aemal J. Khattak,2 and Joseph E. Hummer3

ABSTRACT: Mobile mapping systems (MMSs) are based on advanced technologies for automatically collecting
roadway inventory data. A host of advanced technologies, such as the Global Positioning System for collecting
georeferencing data and digital cameras for collecting roadway data, are used in current MMSs. An MMS is
driven on a subject roadway, collecting positional data and digital images of the roadway. Through photogram-
metric software packages, roadway data are extracted from images. Although there have been studies evaluating
the accuracy of positional data collected by MMSs, there have been limited studies of the descriptive data
accuracy they provide. In this paper the results of an evaluation of the accuracy of descriptive inventory data
collected by three different MMSs are discussed. Each system was tested in three different roadway environ-
ments, and five different types of inventory elements were included in each test. The accuracy of collected data
was compared with ground truth data. Analysis of these data indicated that the roadway environment and the
type of inventory element affect the accuracy of descriptive inventory data collected by MMSs.

INTRODUCTION the vehicle is driven over the roadways of interest. MMSs

Roadway agencies across the United States collect and offer the promise of labor savings and greater accuracy when
maintain vast inventories that facilitate activities such as short- compared to manual data collection methods.
and long-term transportation planning, maintenance and re- The purpose of this paper is to present MMSs as an auto-
habilitation of roadways, and roadway performance monitor- mated method of collecting roadway inventory data and to
ing. Roadway inventories contain two types of data, positional discuss the results of an evaluation of the descriptive data ac-
and descriptive, about each element (feature) of interest. The curacy they provide. The paper has three main sections. A
positional data provide information on the location of an in- general overview of MMSs and specific technologies used to
ventory element in space (e.g., latitude, longitude, and alti- collect roadway data are provided in the first section. Evalu-
tude). The descriptive data provide a definition and list of the ation strategy and the results are discussed in the second sec-
key characteristics of the roadway element (e.g., length, width, tion. Finally, a summary of results and conclusions are pre-
height, condition, and age). sented in the last section.
A roadway inventory data collection method consists of an
appropriate means of transport for data collection equipment, MMSs
appropriate technologies for collection of positional data, and
appropriate technologies for collection of descriptive data. A GPS receiver, a DMI, an INS, and digital cameras are
Current technologies considered in roadway inventory practice common technologies used in an MMS. To obtain high posi-
for collecting positional data include the Global Positioning tional accuracies, the GPS receiver is high end and capable of
System (GPS), range or distance measuring instruments performing differential GPS techniques. The use of GPS tech-
(DMIs), and inertial navigation systems (INSs). Current tech- nology in an MMS is to collect positional data on inventory
nologies for collecting descriptive data include paper and pen- elements. The DMI and INS provide backups for positional
cil, electronic writing pads, keyboards and computers, voice data during GPS signal loss. The DMI triggers data capture
recognition systems, digital image capture systems, and auto- activities at regular distance intervals, and the INS provides
matic image processing systems. A comprehensive description data on the vehicle body roll, pitch, and heading. The GPS
of current positional and descriptive technologies applicable to positioning data also correct the growth of the INS errors,
roadway data collection is given in Karimi et al. (2000). Pro- whereas the INS high frequency measures are essential to de-
cessing and archiving of the field data completes the inventory tect and correct cycle slips (the most familiar drawback of GPS
data collection process. kinematic measures). The digital cameras mounted on the van
The qualities that roadway agencies seek in an inventory and pointed in different directions record images at regular
data collection system are low cost, low labor requirements, distance intervals. Once the images are georeferenced, opera-
and data accuracy sufficient to serve the agency’s needs. Man- tors use photogrammetric software packages to make digital
ual data collection methods are common but are labor inten- measurements (geometric measurements in the three dimen-
sive, and the data accuracy could vary widely from element sions of space on objects captured in the 2D images) of fea-
to element. Many agencies therefore are interested in mobile tures and extract descriptive data from images.
mapping systems (MMSs). Relatively new, highly automated Fig. 1 shows the basic process of collecting digital mea-
systems, MMSs are mounted in a vehicle that collects data as surements. On-board cameras capture overlapping images of
1
the roadway environment, while the GPS, DMI, and INS units
Asst. Prof., Dept. of Information Sci. and Telecommunications, Univ. provide spatial coordinates of the image capture points. Data
of Pittsburgh, Pittsburgh, PA 15260. from the GPS, DMI, and INS are integrated after application
2
Transp. Engr., Ctr. for Transp. Res. and Educ., Iowa State Univ.,
Ames, IA 50010-8615. of differential corrections to the GPS data. A differential GPS
3
Assoc. Prof., Dept. of Civ. Engrg., North Carolina State Univ., Ra- can be implemented in a real-time mode or a postprocessing
leigh, NC 27695-7908. mode. The real-time mode requires a radio communication
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, link between a reference (base) GPS station and the GPS re-
2000. To extend the closing date one month, a written request must be ceiver mounted in an MMS. In the postprocessing mode, the
filed with the ASCE Manager of Journals. The manuscript for this paper
data from a base station and from the GPS receiver (mounted
was submitted for review and possible publication on August 31, 1999.
This paper is part of the Journal of Computing in Civil Engineering, on the van in an MMS) are integrated at a later time. Because
Vol. 14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0168– of the radio link requirement, the real-time mode is more ex-
0173/$8.00 ⫹ $.50 per page. Paper No. 21747. pensive than the postprocessing mode. For roadway inventory
168 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
data collection, the postprocessing mode is adequate (Karimi image coordinates. Identification of conjugate points in two or
et al. 2000). more overlapping images allows positioning of the images rel-
Using special photogrammetric software, users indicate (us- ative to each other. This results in the creation of a scaled
ing the computer mouse) a point of interest in two or more model of the configuration from which the available images
overlapping images (Fig. 2). The indication of the same point were originally obtained. The position of the model is set to
captured in two images results in recording of two sets of the correct orientation and scale. The GPS coordinates col-
lected during the data collection process are used for orienta-
tion and scaling the model. Using the GPS coordinates of the
exposure positions, a bundle adjustment provides the 3D co-
ordinates of the actual points in space to which the image
points correspond. Alternatively, the process of scale correc-
tion may be performed by measuring control points. These
points are well-defined targets whose spatial coordinates are
known beforehand. Users may obtain object dimensions (e.g.,
lane width and traffic sign height) by indicating the appropriate
object end points. Software packages use a distance formula
to calculate the distance between the object endpoints.
Using digital measurements (extracting data from images
collected by MMSs) to obtain descriptive data reduces the time
spent in the field compared to other methods and eliminates
the need to frequently revisit the site. Another advantage of
digital measurements compared to manual methods is that data
collectors do not have to linger in unsafe environments while
collecting data. Table 1 summarizes some of the advantages
and disadvantages of MMSs for collection of roadway inven-
tory data relative to manual estimates made from a moving
vehicle.
The literature provides information on a number of tests
conducted to determine the positional accuracy of data ob-
tained by using different MMSs (Coetsee et al. 1994; Vaidya
et al. 1994; Whited and Kadolph 1995; Schwarz and El-
Sheimy 1996; Novak and Nimz 1997; Shaw and Guthrie
FIG. 1. Digital Measurement Process 1997). In general, positional accuracy of better than 3 m has

FIG. 2. Extraction of Data from Digital Images Collected by MMS

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 169

 TABLE 1. Advantages and Disadvantages of MMSs larity in crack classification (block, fatigue, transverse, longi-
Relative advantages Relative disadvantages tudinal wheelpath, and edge) between the two methods. In
(1) (2) another test, Roadware Corp. (1996) showed that their system
was able to automatically classify collected data on pavement
Can collect data at near highway Requires a crew of two personnel
speeds, thus reducing time spent for safe operation
cracks into the LTPP categories. However, there was no in-
in field formation if the classification was correct.
Rental of system may allow more Requires large initial investment Overall, one cannot base any conclusion on the results of
cost-effective data collection on part of purchasing agency these studies. The studies were not based on statistical data
Wide choice of technologies for Slow data collection due to ad- analyses, did not use appropriate measures of effectiveness,
collection of different types of verse weather conditions and did not provide information on the types of inventory el-
positional and descriptive data
Can carry redundant systems, thus Slow data collection due to traffic
ements evaluated or the types of roadway environments tested.
enhancing data accuracy congestion The work presented in this paper attempted to overcome the
Skilled crew required for operation weaknesses in the previous studies. In this work different in-
and in-office data processing ventory elements and roadway environments were examined,
and appropriate statistical tests on a valid measure of effec-
tiveness were conducted. The ground truth data were carefully
been reported. However, the literature provides information on designed and collected—manual measurements using a tape
relatively few tests of the accuracy of descriptive data obtained measure or similar instrument that represented the best pos-
by MMSs. Table 2 shows a summary of the tests conducted sible estimate of the dimension of interest.
to determine the accuracy of descriptive data collected by
MMSs. In a comparative study of data obtained by an MMS EVALUATION
and ground truth observations, Lee et al. (1991) concluded that
the data obtained by the MMS were of ‘‘reasonable’’ accuracy. To evaluate MMSs, data on different types of inventory el-
However, the study report did not provide any statistical anal- ements in different roadway environments were collected. This
ysis of the data (perhaps due to the limited number of obser- is because it was expected that roadway environments might
vations) or a definition of ‘‘reasonable’’ accuracy. Mastandrea affect the images captured, thus affecting the accuracy of data
et al. (1995) reported a descriptive data accuracy of 5–10 cm collected. Roadway environments and inventory elements that
on offsets, widths, and heights of various inventory elements. are common in the inventories of most roadway agencies were
They did not report on the data elements used in the evalua- chosen. Urban, two-lane rural, and rural interstate were chosen
tion, the evaluation methodology used in the study, any details as the three roadway environments in which data were col-
on the evaluation test, the sample data, or the statistical anal- lected. Within each roadway environment, lane width, traffic
ysis. El-Sheimy (1996) compared dimensions of inventory el- sign width, traffic sign height, traffic sign support height, and
ements estimated using digital measurements to ground truth lateral placement of traffic signs from lane edge were chosen
and reported errors in the range of 5–15 cm for objects 33 m as the types of inventory elements in the evaluation. These
directly in front of the camera. However, there is no infor- elements appear in most inventories, appear often on most
mation on the identity and size of the measured inventory el- roads, and represent a variety of sizes and distance from the
ements or on the number of observations made on the ele- MMS.
ments. Roadware Corp. (1994), Paris, Ontario, Canada, Three different MMSs available on the U.S. commercial
reported on the accuracy of pavement crack identification and market as of mid-1998 were used in the evaluation. The MMSs
classification using an MMS. The test indicated comparable had similar technologies but different integration strategies,
crack identification with the long-term pavement performance photogrammetric software, and input data. The MMSs tested
(LTPP) procedure as described in the Strategic Highway Re- used forward-facing, high mounted, charged-couple device
search Program (SHRP 1993). However, there was no simi- digital cameras to record images. Table 3 provides a summary

TABLE 2. Summary of Descriptive Data Accuracy Tests

Inventory Photogrammetric Roadway
Author MMS element tested Methodology software environment Reported accuracy
(1) (2) (3) (4) (5) (6) (7)
Lee et al. (1991) Pavedex Sign size and height Comparison with ground truth data Microscience SIS Unreported ‘‘Reasonable’’
Mastandrea et al. ARAN Offset, width, and height Absolute accuracy assessment Surveyer Unreported 5–10 cm
(1995) of objects
El-Sheimy (1996) VISAT Unreported Comparison with ground truth data VStation Rural and Urban 10–15 cm
Roadware Corp. ARAN Pavement cracks Comparison of crack identification WiseCrax Unreported Crack identification compara-
(1994) and classification with data col- ble to LTPP procedure, but
lected by LTPP procedure crack classification was not
comparable
Roadware Corp. ARAN Pavement crack classifi- Manual classification of collected WiseCrax Unreported Data successfully classified
(1996) cation data into LTPP categories into LTPP categories
Note: ARAN = Automatic Road Analyzer. VISAT = Video Inertial Satellite.

TABLE 3. Summmary of Hardware and Software Used in Selected MMSs

Experiment
(MMS GPS Digital camera Photogrammetric
number) receiver (pixel resolution) INS DMI software package
(1) (2) (3) (4) (5) (6)
1 Trimble 7400 Pulnix a (768 ⫻ 484) Littona Daytrona Feature Extraction
2 Ashtech Z-12 COHO 4980 RS-170 (640 ⫻ 480) Honeywell Laser Ref III Vehicle ABS VISAT Station
3 Novatel a Sony XC-007 (640 ⫻ 480) INS with 3 gyros and 2 accelerometers a Vehicle ABS RoadView Ruler
a
Indicates exact make or model is proprietary information.

170 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

of the hardware and software in the three MMSs used in the of the relation between a dependent variable and one or more
evaluation. Three comparisons were designed; in each com- independent variables interacting together. Unlike regression,
parison one of the MMSs was compared to ground truth data. the ANOVA models do not require making assumptions about
The results reported below are from over 500 inventory ele- the nature of the statistical relations between the dependent
ments, or an average of 12 observations for each combination and independent variables. The ANOVA models also do not
of element and environment in each experiment. Data collec- require that the independent variables be quantitative.
tion for the three experiments occurred in relatively flat terrain The ANOVA table provides information on the variation in
and in good weather. the phenomenon under investigation due to the factors and due
to random errors. The F-test (ratio of mean squares to mean
Measure of Effectiveness squared error) is appropriate for testing the significance of the
factors and the factor interactions. A significance level of ␣ =
To measure the accuracy of the data collected, we used the 0.05 was used for the tests. Thus, values of the F-statistic
percent measurement error (PME), which is defined ⱖ1.96 indicate statistical significance. The significance level
PMEi = 冉 MMSi ⫺ GTi
GTi 冊
⭈ 100 (1)
of a test indicates the probability of making a Type 1 error,
that is, rejecting the null hypothesis when it is true.
Note that because the three experiments were conducted in-
where PMEi = percent error in the ith observation on an in- dependently of each other and at separate locations, the results
ventory element; MMSi = ith observation on a particular type of the experiments are not directly comparable. In the remain-
of inventory element using MMSs; and GTi = ith ground truth der of this paper, an analysis of the data collected in the three
observation on that particular type of inventory element. experiments is discussed.
As the measure of descriptive data accuracy for data col-
lected by an MMS, the PME has several useful properties. Its
sign (positive or negative) allows the evaluator to determine Experiment 1
if a particular method of data collection is overestimating (pos-
itive sign) or underestimating (negative sign) the true dimen- Table 4 shows the ANOVA results from Experiment 1. The
sion of the inventory element. The use of the ground truth ANOVA results indicate that the environment factor was not
observation in the denominator of (1) normalizes for the size statistically significant, but the type of inventory element mea-
of an inventory element. Thus, errors in measurement of in- sured was statistically significant. This would indicate that the
ventory elements of different sizes are comparable. Further- type of element measured influenced the PME values. How-
more, the use of the PME is simple and straightforward. ever, a conclusion regarding significant effects of the type of
element on the PME values cannot be reached because of the
Analysis of Results significance of the interaction effects between the type of in-
ventory element and the roadway environment. The statistical
We conducted analysis of variance (ANOVA) to study the significance of the interaction term indicates that one should
effects of the independent variables (also called factors; type look at the individual levels of each factor.
of inventory element and roadway environment are the inde- Table 5 summarizes the mean and standard deviations of
pendent variables in this study) on the dependent variable (the PME values across the type of inventory element and the road-
PME in this case). Box et al. (1978), Neter et al. (1990), and way environments. The table indicates that collection of lane
Devore (1991) provide a wealth of insights on the ANOVA width data by the MMS was quite accurate across all three
data analysis technique. The ANOVA models allow the study roadway environments. However, measurements on traffic sign
width and lateral placement from lane edge were not very
TABLE 4. ANOVA Results for Experiment 1 accurate across all three roadway environments. Traffic sign
Source of Sum of Degrees of Mean height was accurately measured across all three environments,
variation squares freedom squares F-value and traffic sign support height was measured accurately on
(1) (2) (3) (4) (5) urban streets and two-lane rural roadways. Overall, the MMS
Between factors 35,922.22 14 7,465.01 — provided more accurate data on the rural interstate environ-
Environment 1,702.45 2 851.22 0.97 ment, compared to the two-lane rural and the urban environ-
Element 18,690.62 4 4,672.65 5.35 ments.
Environment and From the PME mean values in Table 5, it appears that
element 15,529.15 8 1,941.14 2.22 MMSs provide accurate horizontal measurements if the inven-
Error 198,931.10 228 872.50 —
tory element is directly in front of the cameras. However, the
Total 234,853.32 242 — — horizontal accuracy degrades as objects move away from the
Note: Environment consists of three levels: Urban, rural interstate, and front of the cameras. The accuracy of vertical measurements
two-lane rural. Element consists of five levels: Lane width, lateral place- appears quite high. Only in the case of traffic sign support
ment, support height, sign height, and sign width. Number of observations
height in the rural interstate environment was the accuracy
in data set = 243.
slightly lower.

TABLE 5. Experiment 1 PME Mean and Sandard Deviations

Inventory element type Urban street Two-lane rural Rural interstate Overall
(1) (2) (3) (4) (5)
Lane width ⫺1.32 (7.18) 0.75 (6.37) ⫺0.30 (6.64) ⫺0.31 (6.66)
Traffic sign width 16.87 (12.98) 38.58 (40.92) 8.49 (13.25) 18.98 (26.03)
Traffic sign lateral placement 38.18 (74.80) 10.35 (14.03) 6.97 (8.81) 16.01 (40.78)
Traffic sign height 2.44 (6.20) ⫺1.51 (4.34) ⫺4.95 (10.85) ⫺1.53 (8.06)
Traffic sign support height ⫺2.18 (6.57) ⫺1.95 (7.83) 8.86 (53.96) 2.98 (37.15)
Overall 10.41 (41.12) 6.12 (18.86) 4.65 (29.12) 6.61 (30.87)
Note: Standard deviations are in parentheses.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 171

 TABLE 6. ANOVA Results for Experiment 2 TABLE 8. ANOVA Results for Experiment 3
Source of Sum of Degrees of Mean Source of Sum of Degrees of Mean
variation squares freedom squares F-value variation squares freedom squares F-value
(1) (2) (3) (4) (5) (1) (2) (3) (4) (5)
Between factors 8,226.44 14 2,088.07 — Between factors 6,305.36 14 1,590.31 —
Environment 869.65 2 434.83 1.43 Environment 1,052.20 2 526.10 1.19
Element 5,869.17 4 1,467.29 4.84 Element 3,260.62 4 815.15 1.84
Environment and Environment and
element 1,487.62 8 185.95 0.61 element 1,992.54 8 249.06 0.56
Error 49,094.40 162 303.05 — Error 64,349.74 145 443.80 —
Total — 176 2,391.12 — Total — 159 2,391.12 —
Note: Environment consists of three levels: Urban, rural interstate, and Note: Environment consists of three levels: Urban, rural interstate, and
two-lane rural. Element consists of five levels: Lane width, lateral place- two-lane rural. Element consists of five levels: Lane width, lateral place-
ment, support height, sign height, and sign width. Number of observations ment, support height, sign height, and sign width. Number of observations
in data set = 177. in data set = 160.

TABLE 7. Experiment 2 PME Mean and Standard Deviations TABLE 9. Experiment 3 PME Mean and Standard Deviations

Mean Mean
Inventory element type (SD) Inventory element type (SD)
(1) (2) (1) (2)
Lane width ⫺1.62 (8.31) Lane width ⫺0.89 (2.84)
Traffic sign width 4.87 (15.36) Traffic sign width ⫺1.74 (39.75)
Traffic sign lateral placement 13.04 (21.68) Traffic sign lateral placement 3.86 (21.59)
Traffic sign height 9.71 (21.35) Traffic sign height ⫺10.71 (23.61)
Traffic sign support height 2.35 (17.37) Traffic sign support height ⫺5.20 (18.90)
Overall 5.25 (17.91) Overall ⫺1.90 (21.20)

Experiment 2 not significant in Table 7. Measurements on lane width, traffic

sign width, and traffic sign lateral placement appear fairly ac-
Table 6 presents the ANOVA results from Experiment 2. curate. Both the traffic sign height and traffic sign support
The ANOVA results show that the roadway environment had height were underestimated. Judging from the overall PME
little effect on the accuracy of inventory data, but the type of value, the MMS provided quite accurate results. Again, the
inventory element was statistically significant, indicating that accuracy pattern is somewhat similar to the first two experi-
descriptive data accuracy depends on the type of inventory ments.
element measured. The interaction between roadway environ-
ment and type of inventory element was not statistically sig- SUMMARY OF RESULTS AND CONCLUSIONS
nificant. The ANOVA results indicate that one should look
Analysis of the data collected by the three different MMSs
closely at the different levels of the type of inventory element.
in three separate experiments indicated that the accuracy of
Table 7 shows the PME values for the different types of
descriptive inventory data depends on the type of inventory
inventory elements investigated during Experiment 2, along
element measured. Specifically, the MMSs tested in this work
with the standard deviations. The roadway environment data
measured roadway elements more accurately that were closer
is not shown because that factor and its interactions were not
to the cameras or near a principal camera axis. This generally
significant in Table 6. The table indicates that lane width and
confirms previous research (El-Sheimy 1996). Although in two
traffic sign width were measured quite accurately. The data on
of the experiments the roadway environment did not signifi-
the lateral placement of traffic signs and on sign heights had
cantly affect the accuracy of the MMS, the effects of environ-
large errors. Measurements made on traffic sign supports in-
ment on accuracy cannot be readily discarded because of its
dicated fairly accurate measurement of heights. The pattern in
significance in the other experiment. Based on the results we
accuracy is somewhat similar to that of Experiment 1, in which
obtained, we conclude that the type of technologies used in
lane width, traffic sign height, and sign support height were
MMSs, the type of inventory elements collected, and the type
accurately measured.
of roadway environments affect the accuracy of descriptive
Experiment 3 data collected by MMSs.
Whether the accuracies of the tested inventory elements
Table 8 shows the results of the ANOVA from Experiment were acceptable to roadway agencies depends upon the appli-
3. The results show that the roadway environment factor was cations of interest to them. For example, lane widths are im-
not statistically significant, indicating that roadway environ- portant when developing crash models, with agencies typically
ment did not affect the PME values in the experiment. The desiring lane width measurements to the nearest 30 cm. This
type of inventory element measured during the experiment implies an error of no more than 8% or so. The mean and
showed partial statistical significance (at ␣ = 0.10). This in- standard deviation PME values in Tables 5, 7, and 9 show that
dicates that the type of element measured during the experi- the three MMSs tested in these experiments delivered most
ment influenced the PME values, although the statistical evi- measurements within those bounds. For other applications,
dence is not as compelling as in the previous experiments. The however, higher mean or standard deviation PME values in
results do not indicate any significant interaction effects be- certain combinations of element and environment would be
tween the type of inventory element and the roadway envi- unacceptable.
ronment. From a practical standpoint, the main decision a roadway
Table 9 shows the mean and standard deviations of PME agency has to make in this area is whether to invest in an
values obtained in Experiment 3. The roadway environment MMS given a set of data it would like to collect. The accuracy
data is not shown because that factor and its interactions were of the data an MMS can provide is central in that decision.
172 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
This research showed that agencies can expect an MMS to Lee, H., Weissman, M., and Powell, J. (1991). ‘‘Measuring highway in-
produce a range of accuracy, from very accurate for elements ventory features using stereoscopic imaging system.’’ Proc., 2nd ASCE
Int. Conf., ASCE, New York, 448–452.
near a camera axis to quite inaccurate for elements further Mastandrea, A., Swindall, B., and Klassen, G. (1995). ‘‘Development and
from a camera axis. The cost of the equipment and speed of implementation of an automated facility inventory system.’’ Transp.
collection, which is related to the cost of labor needed to op- Res. Rec. 1500, Transportation Research Board, Washington, D.C.,
erate the MMS, are other major factors in the decision. 127–133.
Neter, J., Wasserman, W., and Kutner, M. (1990). Applied linear statis-
ACKNOWLEDGMENTS tical models, 3rd Ed., IRWIN Inc., Boston.
Novak, K., and Nimz, J. (1997). ‘‘Transportation infrastructure manage-
This research was conducted as part of National Cooperative Highway ment: Seneca County creates a complete inventory of its transportation
Research Program (NCHRP) Project 15-15. The writers wish to thank the infrastructure.’’ Proc., Geographic Information Sys. for Transp. Symp.,
NCHRP, Washington, D.C., staff for their support. The writers also wish U.S. Dept. of Transp., Washington, D.C., 301–307.
to thank all the state DOTs, inventory data collection companies, and Roadware Corp. (1994). ‘‘WiseCrax: North Carolina test results.’’ Rep.
vendors who helped in this research. Submitted to U.S. Fed. Hwy. Admin., Washington, D.C.
The opinions expressed in this paper are entirely those of the authors. Roadware Corp. (1996). ‘‘WiseCrax test results.’’ Rep. Submitted to Kan-
Its contents do not necessarily reflect the official views or policies of the sas Dept. of Transp.
NCHRP. Schwarz, K., and El-Sheimy, N. (1996). ‘‘Kinematic multi-sensor systems
for close range digital imaging.’’ Int. Soc. for Photogrammetry and
APPENDIX. REFERENCES Remote Sensing, Vienna, 31(B2), 774–784.
Shaw, L., and Guthrie, M. (1997). ‘‘Eaton County—Michigan Global
Box, G., Hunter, W., and Hunter, S. (1978). Statistics for experimenters, Positioning Systems pilot study.’’ Proc., Geographic Information Sys.
Wiley, New York. for Transp. Symp., U.S. Dept. of Transp., Washington, D.C., 259–273.
Coetsee, J., Brown, A., and Bossler, J. (1994). ‘‘GIS data collection using SHRP. (1993). ‘‘SHRP distress identification manual for the Long-Term
GPS van supported by a GPS/inertial mapping system.’’ Proc., Institute Pavement Performance project.’’ Strategic Highway Research Program,
of Navigation, GPS-94, Alexandria, Va., 85–93. SHRP-P-338, National Research Council, Washington, D.C.
Devore, J. (1991). Probability and statistics for engineering and science, Vaidya, A., Baraniak, D., and Novak, K. (1994). ‘‘CONRAIL’s GIS in-
3rd Ed., Brooks/Cole Publishing Co., Pacific Grove, Calif. corporates kinematic GPS and stereo digital imaging.’’ Proc., AM/FM
El-Sheimy, N. (1996). ‘‘A mobile multi-sensor system for GIS application Int. Conf., AM/FM International, Aurora, Colo., 396–408.
in urban centers.’’ Int. Soc. for Photogrammetry and Remote Sensing, Whited, J., and Kadolph, S. (1995). ‘‘Automated collection of sign in-
Vienna, 31(B2), 95–100. ventory information by integrating Global Positioning System (GPS)
Karimi, H., Hummer, J., and Khattak, A. (2000). ‘‘Collection and pre- with videologging data collection activities.’’ Rep. DTEH71-92-DP85-
sentation of roadway inventory data.’’ Rep. No. 437, Nat. Cooperative IA-35, Ofc. of Plng. Services, Iowa Department of Transportation,
Hwy. Res. Program, National Research Council, Washington, D.C. Ames, Iowa.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 173

 STRUCTURAL MATRIX COMPUTATIONS WITH UNITS
By Mark A. Austin,1 Member, ASCE, Wane-Jang Lin,2 and Xiaoguang G. Chen3

ABSTRACT: Despite the well-known benefits of physical units, matrices, and matrix algebra in engineering
computations, most engineering analysis packages are essentially dimensionless. They simply hold the engineer
responsible for selecting a set of engineering units and making sure their use is consistent. Although this practice
may be satisfactory for the solution of established problem-solving procedures, identifying and correcting un-
intentional errors in computations developed by teams of engineers remains a difficult problem unless units are
an integral part of the computational procedure. This paper begins with a description of the data structures and
algorithms needed to represent and manipulate physical quantity variables and matrices of physical quantities.
The result is a high-level scripting language that offers enhanced type checking for expressions and assignments,
automatic conversion of systems of units, and program control structures for the solution of engineering prob-
lems. These ideas have been implemented in Aladdin, a new computational environment for matrix and finite-
element calculations. The capabilities of Aladdin are demonstrated through the deflection analysis of a cantilever
beam.

INTRODUCTION sure their use is consistent. For example, MATLAB supports

matrix calculations but does not support units of measurement
Now that high-speed computers with Internet connectivity at all (The student 1995). Among the engineering analysis pro-
are readily available, many structural engineering companies grams that do incorporate physical units, most handle them at
see opportunities for using this technology to expand their do- either the program input and output stage (i.e, the input and
main of business activities and enhance business productivity output will be presented and displayed in a certain set of units)
via computer-based support for engineering analysis (including or at the physical quantity level. A number of engineering
design code rule checking, optimization, and interactive com- analysis programs allow a user to specify a set of units to be
puter graphics), contract negotiations, procurement, manage- used (e.g., SI or US) and then assumes that all of the subse-
ment of project requirements, and access to engineering ser- quent calculations will conform to the selected units set. Pro-
vices. The participating computer software programs and gramming environments such as MathCad (MathCad 1998)
networking systems must be integrated, fast and accurate, flex- and Mathematica (Wolfram 1999), which support variables
ible, reliable, and, of course, easy to use. The pathway from with physical units (e.g., x = 2 in), discard the same units when
automated information management and ease-of-use to pro- families of equations are written in matrix format. Indeed, we
ductivity gains is well defined (Goudreau 1994; Austin et al. have not been able to find any computer programs that sys-
1995). tematically integrate units into the definition of physical quan-
With these opportunities in mind, a fundamental tenet of tities, matrices, and finite-element analysis.
our work is that matrices, matrix algebra, and physical units Although this practice may be satisfactory for the solution
are key ingredients of structural engineering systems integra- of self-contained and established problem-solving procedures,
tion. Matrices and linear matrix algebra are an essential part the ability to identify and correct unintentional errors in com-
of present-day structural analysis because they enable large putations developed by teams of engineers remains a difficult
families of equations and their solution procedures to be spec- problem unless units are an integral part of the computational
ified at a relatively high level of abstraction and because ma- procedure. In the evaluation of a new equation, for example,
trix operations are ideally suited for step-by-step solution pro- an environment that supports handling of units can check the
cedures on computers. In traditional approaches to problem dimensional consistency of all subexpressions before proceed-
solving, engineers employ units of measure to help relate the ing with the equation evaluation. Failure to identify and cor-
properties of an object to the real world, thereby generating rect an inappropriate use of units can lead to embarrassing
problem descriptions that are easier to understand, maintain, system failures. A case in point is NASA’s $125,000,000 Mars
and validate than if the units were omitted. Because the selec- Climate Orbiter that recently plowed into the Martian atmo-
tion of a system of units is most often based on convenience sphere because engineers failed to make a simple conversion
and tradition, units of measure are just as important as the from English to metric units in their navigation calculations
numerical quantity itself (Karr and Loveman 1978; Gehani (Sawyer 1999).
1982; Cmelik and Gehani 1988).
Despite the well-known benefits of matrices, matrix algebra, OBJECTIVES AND SCOPE
and physical units, most engineering analysis packages are es- The objectives of this research are to use system develop-
sentially dimensionless. They simply hold the engineer re- ment methodologies and ideas from computer science to de-
sponsible for selecting a set of engineering units and making sign and implement a scripting language and computational
1
toolkit for engineering matrix and finite-element analysis, with
Assoc. Prof., Dept. of Civ. Engrg. and Inst. for Sys. Res., Univ. of units as an integral part of the scripting language. The inte-
Maryland at College Park, College Park, MD 20742.
2
Des. Engr., Bridge Des. Div., Maryland State Hwy. Admin., Balti-
gration of units into a scripting language offers enhanced type
more, MD 21203. checking for expressions and assignments and allows for prob-
3
Chf. Technologist of Risk Mgmt., Tuttle Decision System, Inc., 655 lem-oriented scaling of variables and automatic conversion of
Redwood Hwy., Ste. 200, Mill Valley, CA 94941. systems of units. The target application areas for this work are
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, static and dynamic analysis of multistory buildings and high-
2000. To extend the closing date one month, a written request must be way bridge structures, including earthquake resistant struc-
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on May 14, 1999. This
tures.
paper is part of the Journal of Computing in Civil Engineering, Vol. This paper describes the architecture of a computational
14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0174–0182/ toolkit called Aladdin (Austin et al. 1995) and the data struc-
$8.00 ⫹ $.50 per page. Paper No. 20966. tures, algorithms, and high-level scripting language needed to
174 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
represent and manipulate units, physical quantity variables, (2) the kernel, implemented as a stack machine; and (3) soft-
and matrices of physical quantities in an efficient manner. In ware libraries of units, matrix, and finite-element functions.
Aladdin, finite-element computations are viewed as a special- Matrix and finite-element problems are written as blocks of
ized form of matrix computations, matrices are viewed as rec- statements in an input file.
tangular arrays of physical quantities, and numbers are viewed The lower half of Fig. 1 shows that the development of
as dimensionless physical quantities. The capabilities of Alad- Aladdin is a composition of two problems: (1) The develop-
din are demonstrated in the second half of this paper by work- ment of the matrix and finite-element libraries; and (2) their
ing step-by-step through the displacement analysis of a can- integration into a working system. Solutions to matrix/finite-
tilever beam. element problems are computed using a combination of code
interpreted from the input file and compiled code located in
ALADDIN TOOLKIT the units/matrix/finite-element software libraries. To be useful
to engineers, the problem description language should be rea-
Our implementation of Aladdin has been inspired in part by sonably high level and yet precise, perhaps resulting in hun-
the systems integration methods developed for the European dreds/thousands of low-level native machine instructions per
ESPRIT Project (Kronlof 1993), namely, that system specifi- Aladdin statement. It is important to keep in mind that as the
cations should contain four key components: speed of desktop computers increases, the time needed to pre-
pare a problem description increases relative to the total time
1. A Model—The model will include data structures for the needed to work through an engineering analysis. Hence, clarity
information to be stored (i.e., physical quantities, matri- of an input file’s contents is of paramount importance. The
ces, and finite elements) and a stack machine for the Aladdin scripting language achieves these goals with (1) lib-
execution of the matrix and finite-element algorithms. eral use of comment statements (as with the C programming
2. A Language—The language acts as an interface between language, comments are inserted between /* ⭈ ⭈ ⭈ */); (2) con-
the engineer and the underlying computational model. sistent use of function names and function arguments; (3) use
For our purposes, this means describing matrices, finite- of physical units in the problem description; and (4) consistent
element meshes, and numerical solution procedures. It use of variables, matrices, and structures to control the flow
should be a composition of data, control structures, and of program logic. Aladdin’s scripting language is C-like in the
functions (Salter 1976). sense that it uses only a small number of keywords and sup-
3. Defined Steps and Ordering of the Steps—The steps will ports a variety of familiar looping and branching constructs.
define the transformations that can be carried out on sys- However, unlike C, the scripting language is designed from
tem components (e.g., nearly all engineering processes the bottom-up with definition and manipulation of physical
will require iteration and branching). quantities and matrices of physical quantities in mind.
4. Guidance for Applying the Specification—Guidance in- The functional components of Aladdin provide hierarchy to
cludes factors such as easy-to-understand problem de- the solution of our matrix and finite-element processes, and
scription files, documentation, and example problems. they are located in libraries of compiled C code shown along
the right-hand side of Fig. 1. The strategy we have followed
The upper half of Fig. 1 is a high-level schematic of the Alad- in Aladdin’s development is to keep the number and type of
din architecture; its three main parts are (1) the input file(s); arguments employed in library function calls small. Whenever

FIG. 1. High Level Architecture for Aladdin

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 175

possible, a function’s return type and arguments should be of
the same data type, thereby allowing the output from one or Aladdin parses and stores the character sequence ‘‘20 kg’’ as
more functions to act as the input to following function calls. the physical quantity twenty kilograms. The units name is
Aladdin has library functions for basic matrix operations, the ‘‘kg,’’ the units scale factor is 1, and the [␣, ␤, ␥, ␦, ε] ex-
numerical solution to linear matrix equations, and the sym- ponents are [0, 1, 0, 0, 0]. This quantity is discarded once the
metric eigenvalue problem. The finite-element library contains statement has finished executing. The block of statements
functions for adding nodes, elements, and external loads to the
Aladdin database, specifying boundary conditions, forming the
global stiffness matrix, and printing the system displacements
and stresses.

PHYSICAL QUANTITIES
shows how physical quantities can be assigned to identifiers.
Aladdin supports three data types: ‘‘character string’’ for Fig. 3 shows the physical quantity content settings for each
variable names, physical quantities, and matrices of physical variable. Variable t represents three seconds. The numerical
quantities for engineering data. A physical quantity is a mea- value for this quantity is 3 and the units name is ‘‘sec.’’ Al-
sure of some quantifiable aspect of the modeled world. In though variables x and y both represent distances of two me-
Aladdin, basic engineering quantities, such as length, mass, ters and have a numerical value of 2, their units names are m
and force, are defined by a numerical value (number itself) and cm, respectively. When Aladdin prints a quantity, the out-
plus physical units. Fig. 2 is a subset of units presented in the put numerical value is the quantity numerical value divided
Unit Conversion Guide (Unit 1990) and shows the primary by the scale factor associated with the units name. Therefore,
base units, supplementary units, and derived units that occur the statements
in structural analysis. The four basic units needed for engi-
neering analysis are length unit L; mass unit M; time unit t;
and temperature unit T. Planar angles are represented by the
supplementary base unit rad.
Derived units are expressed algebraically in terms of base will generate the output
and supplementary units by means of multiplication and di-
vision, namely
units = k ⭈ L␣M ␤t ␥T ␦ ⭈ radε (1) For variable x, the printed value is 2 because 2/1 = 2. For
where ␣, ␤, ␥, ␦, and ε = exponents; and k = scale factor. The variable y, the printed value is 200 because 2/0.01 = 200.
four basic units play the primary role in determining compat- Finally, when Aladdin parses the assignment statement for
ibility of units in physical quantity and matrix operations. Be- gravity, the character sequence ‘‘m/sec⵩2’’ is translated
cause a radian represents the ratio of two distances (i.e., dis- to the set of unit exponents [␣, ␤, ␥, ␦, ε] = [1, 0, ⫺2, 0, 0].
tance around the perimeter of a circle divided by its radius),
most software implementations deal with radians as if they Special Names for Derived Units
were dimensionless entities. Aladdin departs from this trend
by explicitly representing radians and employing a special set A number of derived units have special names that may
of rules for their manipulation during physical quantity and themselves be used to express other derived units in a simpler
matrix operations. The result is enhanced readability of pro- way than in base units. For example, Aladdin uses the symbol
gram output, especially for computations involving displace- N to represent one newton (i.e., [␣, ␤, ␥, ␦, ε] = [1, 1, ⫺2,
ments and rotations of structural components. 0, 0]), and so the force
In addition to a set of units defined by (1), each physical
quantity contains a character string storing the unit’s name and
a flag for representation of various systems of units (e.g., SI, can be simply written 25 N (Fig. 2).
US, and combined SI㛭US). SI㛭US units are those shared by
both US and SI systems, the most important examples being Simplification and Casting of Physical
time, frequency, and planar angle. Quantity Output
The Aladdin units package provides facilities for dynamic
allocation of units, units copying, consistency, checking and Because Aladdin stores all quantities internally with SI
simplification, and units printing. Operations for units conver- units, quantities specified in US units must be converted into
sion are provided, as are facilities to turn units operations on/ their SI counterparts before they can be stored and used in
off. In an effort to keep the scripting language usage and im- calculations. Fortunately, for most sets of units (the important
plementation as simple as possible, all physical quantities are exception being temperature units), only a single scale factor
stored as floating point numbers with double precision accu- is needed to convert a quantity from US into SI and vice versa
racy plus units. Floating point numbers are viewed as physical (e.g., 1 in = 25.4 mm = 25.4E-3 m). Aladdin automates con-
quantities without units (i.e., [␣, ␤, ␥, ␦, ε] = [0, 0, 0, 0, 0]). version between different, but equivalent, units and provides
There are no integer data types in Aladdin. features in its language for physical quantities to be cast and
displayed in a desirable set of units. For example, with the
Physical Quantity and Variable Declarations definition of gravity in place, the statements
Aladdin stores physical quantities as ‘‘unscaled’’ values with
respect to a set of SI units—meter ‘‘m’’ is the reference for
length, kilogram ‘‘kg’’ for mass, second ‘‘sec’’ for time, and
‘‘deg㛭C’’ for temperature. For example, the Aladdin statement generates the output

generates the output

176 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 2. Primary Base and Derived Units in Structural Analysis

Both sets of units have the same dimension (i.e., length di- Physical quantity expressions are evaluated in two steps. In
vided by time squared), with the numerical value of the phys- the first step, units within the expression are examined for
ical quantity differing only by a scale factor. The stream of compatibility with respect to the operation to be performed.
characters between (. . .) indicates the desired format for the Incompatible units will trigger the printing of an error message
physical quantity units. and termination of the program execution. Expressions con-
Of course, casting of physical quantity output also works taining compatible units proceed to the second step—evalua-
with special names for derived units. This feature is useful for tion of the numerical expression with appropriate units for the
situations where a particular set of unit exponents has more result.
than one interpretation. For example, the unit exponents [␣, To see how operations of physical quantities works in prac-
␤, ␥, ␦, ε] = [2, 1, ⫺2, 0, 0] can be interpreted as a moment tice, let q1 be a physical quantity with unit scale factor k1 and
(force times distance) and as work done (force times distance). unit exponents [␣1, ␤1, ␥1, ␦1, ε1]. Let q2 be a physical quan-
Casting of units allows for these cases to be distinguished, as tity with unit scale factor k2, and unit exponents [␣2, ␤2, ␥2,
in ␦2, ε2].

Expressions Involving Arithmetic Operators

Table 1 contains a summary of arithmetic expressions to-
gether with the rules for computing the scale factor and unit
exponents. The operations of addition and subtraction, q1 ⫾
q2, are defined when the units of q1 and q2 are equivalent. The
The program output is rules for evaluation of units are as follows:

1. If the operands all have the same units type (e.g., let us
say all are SI or US units), then the result units will be
PHYSICAL QUANTITY ARITHMETIC set according to the first operand.
2. If the operands have different units type, then the result
Aladdin supports the construction and evaluation of physical units will be set according to the operand, which has the
quantity expressions involving arithmetic, relational, and log- same units type as the environmental units type.
ical operators, thereby enabling a powerful check for the di- 3. When there are two or more operands that have the same
mensional consistency of formulas. Units and physical quan- units type as the environmental units type, the first one
tities are integrated into the Aladdin scripting language via (1). will be the basis for the result units.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 177
 FIG. 3. Data Structure Contents for x, y, t, and gravity

For example, the script of code TABLE 1. Units Arithmetic in Arithmetic Operations
Scale
Description Expression factor Unit exponents
(1) (2) (3) (4)
Addition q1 ⫹ q2 k1 [␣1, ␤1, ␥1, ␦1, ε1]
Subtraction q1 ⫺ q2 k1 [␣1, ␤1, ␥1, ␦1, ε1]
Multiplication q1 * q2 k1 ⭈ k2 [␣1 ⫹ ␣2, ␤1 ⫹ ␤2, ␥1 ⫹ ␥2, ␦1
generates the output ⫹ ␦2, ε1 ⫹ ε2]
Division q1/q2 k1/k2 [␣1 ⫺ ␣2, ␤1 ⫺ ␤2, ␥1 ⫺ ␥2, ␦1
⫺ ␦2, ε1 ⫺ ε2]
⵩
Exponential q q
1 2 k 1N a [N␣1, N␤1, N␥1, N␦1, Nε1]a
a
N is the value of q2.

and shows Rule 1 in action. For multiplication and division,

the units of q1 ⭈ q2 may be obtained by simply adding the ex- TABLE 2. Expressions Involving Relational and Logical Oper-
ponents in the units of q1 and q2. The units of q1/q2 are com- ators
puted by subtracting the exponent values in q2 from the ex-
Operator Description Example Result
ponent coefficients in q1.
(1) (2) (3) (4)
The expression q1q2 only makes sense when q1 and q2 are
dimensionless, and cases where one of the two operands, but < less than x<y true
not both, has units. For example, if variable y has units and > greater than x>y false
<= less than or equal to x <= u true
x = y0.5, then the unit exponents of x will be half those of y. >= greater than or equal to x >= y false
== identically equal to x == y false
Expressions Involving Relational and != not equal to x != y true
Logical Operators && logical and (x < y) && (x <= y) true
㛳 logical or ( y < x) 㛳 (x <= y) true
Table 2 summarizes the logical and relational operators that ! logical not !y false
can be applied to physical quantities. Briefly, expressions in-
volving the relational operators (e.g., q1 < q2, q1 ⱕ q2) are
defined only when the units of q1 and q2 are equivalent. A
a pointer to the matrix body of double precision floating point
units compatibility check is made before the operation pro-
elements. The matrix element units are stored in two 1D arrays
ceeds, and the result of the operation is either true (1) or false
of type units, as defined by (1). One array stores the column
(0). Assuming that variables x and y are as defined in the
units, and a second array the row units. The units for matrix
previous section, the Aladdin statements
element at row i and column j are simply the product of the
i-th element of the row units buffer and the j-th element
of column units buffer.
Our use of row and column units buffers means that this
generate the output model does not support the representation of matrices of quan-
tities having arbitrary units. For most engineering applications,
matrices are simply a compact and efficient way of describing
families of equations of motion and equilibrium, and collec-
Similar Aladdin statements can be written for the logical op- tions of data. Engineering considerations usually dictate that
erands, and expressions involving combinations of arithmetic, the terms within an equation be dimensionally consistent. Sim-
relational, and logical operands (Table 2). ilarly, consistency of dimensions in large collections of engi-
neering data also tends to hold. Hence, in practical terms, the
MATRICES OF PHYSICAL QUANTITIES assumptions made by this model not only have minimal impact
on our ability to solve engineering problems with matrices,
Aladdin’s engineering units module supports operations on but the model requires much less memory than individual stor-
engineering quantities and matrices for both US and SI sys- age of units for all matrix elements. More precisely, for an (n
tems of units. The matrix module supports storage of double ⫻ n) matrix, our storage scheme requires memory for 2n units.
data types, the specification of units for each matrix element, Individual storage of units for all matrix elements requires
and a wide range of matrix operations. Indirect storage patterns memory for n2 units.
are used for small general matrices. Skyline storage techniques
are employed for large symmetric matrices.
Fig. 4 shows a high-level layout of memory for the matrix Dynamic Allocation of Matrices
data structure. Memory is provided for a character string con-
taining the matrix name, integers for the number of matrix The step-by-step procedure for constructing a matrix with
rows and columns, the storage scheme of matrix elements, and units is as follows:
178 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 4. Layout of Memory in Matrix Data Structure

1. Dynamically allocate memory for MATRIX and a char-

acter string for the matrix name.
2. Dynamically allocate memory for the matrix body.
3. Dynamically allocate memory for the row and column
units matrices.
4. Initialize values for the matrix elements.

For small matrices, Steps 2 and 3 can be combined by explic-

itly declaring a complete matrix of physical quantity elements. FIG. 5. Activities and Output in ‘‘for-looping’’ Construct
A comprehensive example of this feature is presented at the
end of this paper. = aib2c2dj = ⭈ ⭈ ⭈ = aibqcqdj. Matrix element zij is assigned the
units aib1c1dj. For matrix Z, the units buffers are written as a
MATRIX OPERATIONS row units buffer [a1c1, a2c1, . . . , amc1], and a column units
buffer is [d1b1, d2b1, . . . , drb1]. It is important to notice that
Aladdin supports a wide variety of matrix operations, in- although the units for matrix Z are unique, the solution for the
cluding addition and subtraction, matrix multiplication, units buffers is not unique.
solution of linear systems of equations, and the symmetric
eigenvalue problem. Units are carried along with all matrix PROGRAM CONTROL
operations.
Program control is the basic mechanism in a programming
Addition and Subtraction language for using the outcome of logical and relational ex-
pressions to guide the pathway of a program execution. Alad-
Addition and subtraction of matrices is a direct extension din supports the if and if-then-else branching con-
of addition and subtraction of physical quantities. The opera- structs, and the while and for looping constructs, with
tion will proceed if and only if the matrices have the same logical and relational operations being computed on physical
number of rows and columns and the row and column buffers quantities. For example, the syntax for the for-loop construct
of both matrices are dimensionally consistent. Now let matrix is as follows:
Z be the result of X ⫾ Y. Matrix Z will assume the same row
and column units as X, with element values of Z scaled ac-
cordingly.

Multiplication The initializer, condition, and increment state-

The handling of units in multiplication of 2D matrices needs ments can be zero or more statements separated by a comma.
special attention. Let X be a (m ⫻ n) matrix with row units Similarly, zero or more statements may be located in the for-
buffer [a1, a2, . . . , am] and column units buffer [b1, b2, . . . , loop body. For example, the block of statements
bn]. Let Y be a (n ⫻ r) matrix with row units buffer [c1, c2,
. . . , cn] and a column units buffer [d1, d2, . . . , dr]. The units
for matrix elements xik and ykj are ai * bk and ck * dj, respectively.
Moreover, let Z be the product of X and Y. From basic linear
algebra we know that generates the sequence of activities and output shown in Fig.

冘
n
5. Execution of the looping construct begins with the ini-
tializer statement, where we assign 1 cm to variable x.
zij = xik *ykj (2) The condition statement is then evaluated to see if it is
k=1
true or false. In this particular example, x <= 5 cm evaluates
Due to the consistency condition, all terms in the matrix ele- to true and so statements inside the body of the loop are ex-
ment summation must have the same units. That is, the ex- ecuted (i.e., we print out the value of x). When all of the
ponents of the units must satisfy the product constraint aib1c1dj statements inside the for-loop body have finished execut-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 179
 FIG. 6. Single Span Cantilever Beam with Supported Endpoint

ing, the increment statement is executed (i.e., we increase factor corresponding to the beam-end lateral displacements
the value of x by 2 cm) and condition is re-evaluated to and rotations.
see if another loop of computations is required. For this ex-
ample, looping continues for four iterations until the con- Definition of Element Stiffness Matrix
dition evaluates to false. At the end of the for-loop exe- It is well known that if cubic interpolation functions are
cution, the value of x will be 7 cm. You should notice that all used to describe displacements along the beam element, then

冋 册
of the elements in the for-loop statement are dimensionally the beam element stiffness matrix will be
consistent.
⫺3/L
冋 册
2 3/L 1
ALADDIN STACK MACHINE 2EI 3/L 6/L2 3/L ⫺6/L2
k= ⭈
L 1 3/L 2 ⫺3/L
The heart of the Aladdin kernel is a finite-state stack ma- ⫺3/L ⫺6/L2 ⫺3/L 6/L2
chine model that follows in the spirit of work presented by
Kernighan and Pike (1983). Stack machines are suitable for where EI = flexural rigidity of the beam element; and L = its
modeling systems that have a finite number of internal config- length. The abbreviated fragment of Aladdin code
urations or states, and whose behavior evolves as a linear se-
quence of discrete points in time. The stack machine reads
blocks of command statements from either the input file, or
keyboard, and processes them in two phases:

1. High-level input statements are converted into an array defines variables for the beam material and section properties
of equivalent low-level stack machine instructions. and its length. The block of statements
2. The stack machine walks along the array of machine in-
structions and executes the functions pointed to by the
machine instructions.

The scripting language specification and conversion process is

enabled by the UNIX tool YACC, an acronym for Yet Another
Compiler Compiler (Johnson 1975). YACC takes a language
description (i.e., a Bacus-Naur form grammar) and automati- dynamically allocates the layout of memory shown in Fig. 4,
cally generates C code for a parser that will match streams of initializes all matrix elements to zero and the matrix name to
input against the rules of the language. For a detailed step-by- ‘‘stiff,’’ and sets the row and column units matrices. Now the
step demonstration of the stack machine processing the as- layout of memory and contents of the units buffers and matrix
signment x = 2 in (Austin et al. 1999). body are
EXAMPLE: DEFLECTION ANALYSIS OF
CANTILEVER BEAM
In this example we use Aladdin’s scripting language to com-
pute displacements in the cantilever beam shown in Fig. 6. We
have deliberately selected a problem with a small number of
unknowns so that step-by-step details of the problem formu-
lation and solution procedure can be illustrated using functions
from the matrix library alone. Of course, the same example You should observe that while all of the matrix elements have
can be implemented in far fewer scripting statements by call- a numerical value of zero, the matrix elements are not equal
ing built-in functions from the finite-element library. because they have different sets of units. For example,
The cantilever has a length of 12 m, modulus of elasticity stiff[1][1] equals 0.0 N ⭈ m/rad and stiff[2][1]
E = 200 GPa, and constant moment of inertia I = 15.5 ⫻ 106 equals 0.0 N/rad. With the exception of the radian units in
mm4 along its length. We will assume that the cantilever is the first and third columns, the matrix units are symmetric.
rigid in its axial direction (i.e., beam cross section area ⬇⬁ The block statements
m2) and that all other deformations are small. The boundary
conditions are full-fixity at the base of the cantilever. Overall
behavior of the cantilever will be modeled with three beam
finite elements and six global degrees of freedom, as numbered
on the left-hand side of Fig. 6. Displacements along each beam
element will be expressed as the superposition of four cubic
interpolation functions, each weighted by a nodal displacement
180 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
initializes the stiffness matrix element values. Before the arith- allocates memory for the (6 ⫻ 6) global stiffness matrix, as-
metic expression on the right-hand side of each statement is signs appropriate row and column units, and then systemati-
assigned to the matrix element, the units of the evaluated ex- cally assembles the global stiffness matrix.
pression are compared to the matrix element units, as defined Although blocks [d] and [e] do a nice job of illustrating
by the row and column units buffers. Radians are excluded the algorithm for assembling the global stiffness matrix, asking
from the units compatibility check. The assignment will pro- a programmer to manually assemble the global stiffness matrix
ceed if and only if units of the right- and left-hand sides of via the LD matrix is tedious and error prone. Moreover, exe-
the assignment are compatible. (Otherwise, an error message cution of the global stiffness assembly will be slow because
will be printed, indicating an incompatibility of units and the the corresponding input file instructions are interpreted. All of
file line number where the error occurs.) The element stiffness these problems can be mitigated by calling the compiled C
matrix is function, Stiff( ), in Aladdin’s finite-element library, The
single statement

replaces blocks of code [d] and [e] shown above.

External Load Vector/Solution of Linear
Matrix Equations
An external load of ⫺1.5 kN is applied to degree of free-
dom five. The required Aladdin code, with appropriate units,
Rows 1 and 3 of stiff represent equations of equilibrium is as follows:
when a moment is applied at the beam end to cause a unit
rotation, and rows 2 and 4 represent equations of equilibrium
for unit lateral displacements of the beam element. We distin-
guish these cases by applying units of m to rows 1 and 3.

Destination Array
A straightforward way of assembling these matrices is with Now we can compute the solution to the linear equations
the use of destination arrays, namely stiff*displ = eload, and print the cantilever displace-
ments by simply writing

Compatibility of units in the linear matrix equations dictates

that

冋 册冋 册
Element LD[i][j] contains the global degree of freedom [N/rad N/m N/rad N/m N/rad N/m]
for degree of freedom j of beam finite element i. A zero entry
indicates that the beam element degree of freedom is fully displ[1][1] units N* m
fixed and will not be mapped to the global stiffness matrix. displ[2][1] units N
displ[3][1] units N* m
⭈ =
Assembly of Global Stiffness Matrix displ[4][1] units N
displ[5][1] units N* m
The assembly process for the global stiffness matrix may be displ[6][1] units N

冋 册冋册
written k1 ⫹ k2 ⫹ k3, where ki represents the stiffness matrix Equating and rearranging terms in the units equation gives
for beam element i mapped onto the global degrees of free-
dom. The following block of Aladdin code displ[1][1] units rad
displ[2][1] units m
displ[3][1] units rad
=
displ[4][1] units m
displ[5][1] units rad
displ[6][1] units m
In other words, the column units of stiff multiplied by the
row units of displ must match the row units of eload. By
explicitly representing units for the third column as N/rad
(versus N), radian units can be assigned to displ[3][1].
Including radians in the solution vector enhances its readabil-
ity and allows for appropriate conversion among units for pla-
nar angles. The displacements are

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 181

The vertical deflection of the cantilever endpoint is 27.87 MathCad 8 user’s guide. (1998). Mathsoft Inc., Cambridge, Mass.
cm, downward. Salter, K. G. (1976). ‘‘A methodology for decomposing system require-
ments into data processing requirements.’’ Proc., 2nd Int. Conf. on
Software Engrg.
CONCLUSIONS AND FUTURE WORK Sawyer, K. (1999). ‘‘Engineers’ lapse led to loss of Mars spacecraft:
Lockheed didn’t tally metric units.’’ Washington Post, Oct. 1.
Aladdin (Version 2.0) is a computational toolkit for the in- The student edition of MALAB; version 4 user’s guide. (1995). MATLAB
teractive matrix and finite-element analysis of large engineer- Curriculum Series, Prentice-Hall, Englewood Cliffs, N.J.
ing structures. The toolkit views finite-element computations Unit conversion guide. (1990). Fuels and Petrochemical Division of
as a specialized form of matrix computations, matrices as rec- AIChE, American Institute of Chemical Engineers, New York.
tangular arrays of physical quantities, and numbers as dimen- Wolfram, S. (1999). The Mathematica book, 4th Ed., Cambridge Univer-
sionless physical quantities. The data structures and algorithms sity Press, New York.
described in this paper have been used to solve engineering
problems containing as many as 2,500 matrix rows and col- APPENDIX II. NOTATION
umns. For a wide range of application examples and the pro- The following symbols are used in this paper:
gram source code, see the Aladdin Web site at 具http://www.
isr.umd.edu/⬃austin/aladdin.html典. a = element of row units buffer;
Our original goal for Aladdin was a computational environ- b = element of column units buffer;
ment that would help engineers set up and solve matrix prob- c = element of row units buffer;
lems in structural analysis and mechanics and finite-element d = element of column units buffer;
problems in building and highway bridge analysis. However, deg-C = degrees Celsius;
it is now evident from the large number of Aladdin source displ = matrix of structural displacements;
code downloads that Aladdin is useful for the solution of a distance = length physical quantity;
much wider range of engineering and scientific applications. E = Young’s modulus of elasticity;
To enhance the usefulness of Aladdin across engineering and eload = matrix of external loads;
business domains, the units package should be extended to gravity = gravitational acceleration;
include units of electric charge and nonphysical units, such as gstiff = global stiffness matrix;
money/currency. F = degrees Fahrenheit;
We are currently working on finite elements and numerical force = force physical quantity;
algorithms for the analysis of structures undergoing large ge- I = moment of inertia;
ometric displacements, problems involving fluid-structure in- K = global stiffness matrix;
teraction, and numerical algorithms for the hybrid control of k = units scale factor, and element stiffness matrix;
large structural systems. The results of this work will be re- L = length units, and length of beam element;
LD = destination array;
leased in Aladdin (Version 3).
M = mass units;
m = number of matrix rows/columns;
ACKNOWLEDGMENTS n = number of matrix rows/columns;
This work was supported in part by the National Science Foundation’s q = physical quantity;
Engineering Research Centers Program (NSFD CDR 8803012) and by a r = number of matrix rows/columns;
series of Federal Highway Administration Fellowship grants to Wane- rad = radians;
Jang Lin. This support is gratefully acknowledged. The views expressed SI = International System of Units;
in this paper are those of the writers and not necessarily those of the SI㛭US = units common to SI and US systems of units;
sponsors.
stiff = element stiffness matrix;
T = temperature units;
APPENDIX I. REFERENCES t = time units, and physical quantity;
Austin, M. A., Chen, X. G., and Lin, W.-J. (1995). ‘‘Aladdin: A com- US = English system of units, used in United States;
putational toolkit for interactive engineering matrix and finite element X = matrix;
analysis.’’ Tech. Res. Rep. TR 95-74, Inst. for sys. Res., University of x = physical quantity, and an element of matrix X;
Maryland, College Park, Md. Y = matrix;
Austin, M. A., Lin, W.-J., and Chen, X. G. (1999). ‘‘Structural matrix y = physical quantity, and an element of matrix Y;
computations with units: Data structures, algorithms, and scripting lan-
Z = matrix; and
guage design.’’ Tech. Res. Rep. TR 99-55, Inst. for Sys. Res., University
of Maryland, College Park, Md. z = element of matrix Z.
Cmelik, R. F., and Gehani, N. H. (1988). ‘‘Dimensional analysis with
C⫹⫹.’’ IEEE Software, 21–26. Superscripts
Gehani, N. H. (1982). ‘‘Databases and units of measure.’’ IEEE Trans.
on Software Engrg., 8(6), 605–611. ␣ = exponent for length units;
Goudreau, G. L. (1994). ‘‘Computational structural mechanics: From na- ␤ = exponent for mass units;
tional defense to national resource.’’ IEEE Computational Sci. and ␥ = exponent for time units;
Engrg., 1(1), 33–42. ␦ = exponent for temperature units; and
Johnson, S. C. (1975). ‘‘YACC—Yet Another Compiler Compiler.’’ ε = exponent for radian units.
Comp. Sci. Tech. Rep. 32, AT&T Bell Laboratories, Murray Hill, N.J.
Karr, M., and Loveman, D. B. (1978). ‘‘Incorporating units into program-
ming languages.’’ Communications of the ACM, 21(5), 385–391. Subscripts
Kernighan, B. W., and Pike, R. (1983). The UNIX programming environ-
ment, Prentice-Hall Software Series, Prentice-Hall, Englewood Cliffs, i, j, k = matrix row and column indices, and indices of
N.J. units in matrix row and column units buffers; and
Kronlof, K. (1993). Method integration: Concepts and case studies, Wi- 1, 2, 3, 4 = matrix row and column, and physical quantity in-
ley, New York. dices.

182 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

 FORECASTING CYANOBACTERIAL CONCENTRATIONS
USING B-SPLINE NETWORKS
By Holger R. Maier,1 Tarek Sayed,2 and Barbara J. Lence3

ABSTRACT: Artificial neural networks have been used successfully in a number of areas of civil engineering,
including hydrology and water resources engineering. In the vast majority of cases, multilayer perceptrons that
are trained with the back-propagation algorithm are used. One of the major shortcomings of this approach is
that it is difficult to elicit the knowledge about the input/output mapping that is stored in the trained networks.
One way to overcome this problem is to use B-spline associative memory networks (AMNs), because their
connection weights may be interpreted as a set of fuzzy membership functions and hence the relationship between
the model inputs and outputs may be written as a set of fuzzy rules. In this paper, multilayer perceptrons and
AMN models are compared, and their main advantages and disadvantages are discussed. The performance of
both model types is compared in terms of prediction accuracy and model transparency for a particular water
quality case study, the forecasting (4 weeks in advance) of concentrations of the cyanobacterium Anabaena spp.
in the River Murray at Morgan, South Australia. The forecasts obtained using both model types are good. Neither
model clearly outperforms the other, although the forecasts obtained when the B-spline AMN model is used
may be considered slightly better overall. In addition, the B-spline AMN model provides more explicit infor-
mation about the relationship between the model inputs and outputs. The fuzzy rules extracted from the B-spline
AMN model indicate that incidences of Anabaena spp. are likely to occur after the passing of a flood hydrograph
and when water temperatures are high.

INTRODUCTION model developed by Maier et al. (1998) for the same case
study, in terms of prediction accuracy and model transparency.
In recent years, artificial neural networks (Flood and Kartam
1994) have been used successfully in a number of areas of COMPARISON OF MLP WITH B-SPLINE AMNs
civil engineering. In the disciplines of hydrology and water MLPs and AMNs are able to approximate any continuous
resources engineering, they have been applied to the prediction function, given sufficient degrees of freedom (Brown and Har-
and forecasting of rainfall, rainfall-runoff, water levels, and a ris 1994; Hornik et al. 1989). They can be mapped onto three-
number of water quality parameters (Maier and Dandy 2000), layer networks, use feedforward connections, and are trained
as well as the management of water quality systems [e.g., using supervised algorithms. However, there are differences in
Takyi and Lence (1999)]. In the majority of cases, multilayer the structure of the input and hidden layers, resulting in dis-
perceptrons (MLPs) that are trained with the back-propagation tinct modeling capabilities.
algorithm are used (Maier and Dandy 2000). One of the major
shortcomings of this approach is that the knowledge that is Model Structure and Operation
contained in the trained networks is difficult to interpret, be-
The p inputs to MLPs are represented by individual nodes
cause it is distributed across the connection weights in a com-
in the input layer (Fig. 1). The input signals to MLPs are
plex manner. weighted and passed to nodes in one or more hidden layers.
One way to overcome this problem is to use the class of The weighted inputs connected to each hidden layer node are
associative memory networks (AMNs) that have a fuzzy in- summed and passed through a transfer function. The latter is
terpretation and are thus more transparent (Brown and Harris generally a bounded ridge function, such as the logistic or hy-
1995). It has been proven that B-spline AMNs and certain perbolic tangent function. Generally, MLPs are fully connected
types of fuzzy models are learning equivalent, i.e., model co- and their geometry and complexity are defined by the number
efficients are estimated in the same way for both model types of hidden layers and the number of nodes in each of these.
(Brown and Harris 1994). Consequently, B-spline AMNs may In contrast, the input space of AMNs is normalized by a p-
be viewed as a particular type of neurofuzzy model. The ob- dimensional lattice (Fig. 2) (Brown and Harris 1994). Each
jectives of this paper are to: (1) provide a theoretical compar- cell of the lattice represents similar regions of the input space.
ison between AMNs and MLPs; (2) develop a B-spline AMN The geometry of AMNs is limited to one hidden layer. The
model for forecasting (4 weeks in advance) a particular species hidden layer nodes consist of basis functions, which are de-
group of blue-green algae (cyanobacteria) in the River Murray fined on the lattice formed by normalizing the input space.
at Morgan, South Australia; and (3) assess the suitability of The basis functions can take a number of forms, including
the model by comparing its performance with that of the MLP threshold, B-spline, or Gaussian functions. Model structure
1
Postdoctoral Res. Fellow, Dept. of Civ. Engrg., Univ. of British Co-
lumbia, 2324 Main Mall, Vancouver, BC, Canada V6T 1Z4; and Lect.,
Dept. of Civ. and Envir. Engrg., Univ. of Adelaide, Adelaide, S.A. 5005,
Australia. E-mail: [email protected]
2
Asst. Prof., Dept. of Civ. Engrg., Univ. of British Columbia, 2324
Main Mall, Vancouver, BC, Canada V6T 1Z4.
3
Assoc. Prof., Dept. of Civ. Engrg., Univ. of British Columbia, 2324
Main Mall, Vancouver, BC, Canada V6T 1Z4.
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1,
2000. To extend the closing date one month, a written request must be
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on June 18, 1999. This
paper is part of the Journal of Computing in Civil Engineering, Vol.
14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0183–0189/
$8.00 ⫹ $.50 per page. Paper No. 21288. FIG. 1. Structure of Typical MLP Model

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 183

 FIG. 3. Additive ANOVA Decomposition of AMNs [after Brown
and Harris (1995)]

FIG. 2. Structure of Typical AMN Model [after Brown and Har-

ris (1994)]
and two 2D subnetworks. It should be noted that each of these
subnetworks represents a separate AMN, the outputs of which
are summed to produce the overall model output.
and complexity are determined by the size, shape, and overlap
of the basis functions.
For both network types, the number of nodes in the output Optimization (Training)
layer corresponds to the number of model outputs. The output MLP and AMN models also differ in the way in which they
of MLPs is obtained by summing the weighted outputs of the are calibrated (trained). Calibration of MLPs is a nonlinear
nodes in the last hidden layer and passing them through a optimization problem. A number of first-order (e.g., back-
transfer function (Fig. 1). In contrast, the output for AMNs is propagation), second-order (e.g., Levenberg-Marquardt, con-
produced by summing the weighted outputs from the basis jugate gradient), or global (e.g., genetic algorithms) methods
functions (Fig. 2) can be used for this purpose (Maier and Dandy 1999, 2000).

冘
p Optimization of the connection weights of AMNs, on the other
y= ai wi (1) hand, is a linear optimization problem. Consequently, conven-
i=1 tional linear learning theory can be used, and convergence re-
sults can be derived (Brown and Harris 1994). In on-line ap-
where y = model output; ai = output from the pth basis func- plications, the weights of AMNs are generally updated
tion; and wi = connection weight associated with ai. iteratively using normalized least mean squares learning rules
As the input space of AMNs is normalized into a p-dimen- (Brown and Harris 1994).
sional lattice, similar network inputs only activate a small
number of basis functions (Brown and Harris 1994). In other Model Development
words, only a small number of weights contribute to the model
output for a given set of model inputs. Consequently, infor- Another difference between MLP and AMN models is the
mation is stored and adapted locally. Only those weights that way in which appropriate model structures are selected. The
contribute to the model output are adjusted during training, type of nonlinear function used in the hidden layer(s) can have
resulting in a fast rate of initial convergence and an ability to a significant impact on the performance of both types of mod-
incorporate new information into the model without overwrit- els. In MLPs, sigmoidal functions are generally used, whereas
ing information that is stored in other regions of the network. the two most commonly used AMN basis functions are B-
In contrast, MLPs have a global response. Adjustment of any splines and Gaussian radial basis functions. A detailed discus-
weight will have an impact on the output obtained for every sion of the relative merits of these two types of functions is
input. In addition, the basis functions may be interpreted as a given by Brown and Harris (1995). The major difference be-
set of fuzzy linguistic rules, making the knowledge that is tween Gaussian and B-spline functions is that the latter have
contained in the model transparent. bounded supports (i.e., the output of the functions is only non-
The major disadvantage of AMNs is that the number of zero if inputs lie in the subset of the input space on which
connection weights is exponentially dependent on the dimen- they are defined), whereas the former do not. Consequently,
sion of the input space (Brown and Harris 1995). Conse- B-spline networks are less flexible than Gaussian networks.
quently, they are not suitable for problems that have a large However, they are more efficient to implement, have greater
number of inputs. However, if it is assumed that the desired transparency, and are more appropriate for generating fuzzy
input/output mapping can be additively decomposed, the di- algorithms.
mensionality of AMN models can be reduced significantly by Parsimony is a major consideration when choosing appro-
modeling the desired relationship as a linear combination of a priate structures for MLP and AMN models. It is desirable to
number of subnetworks. The analysis of variance (ANOVA) use models that have the smallest number of adjustable param-
representation of multivariate functions provides a framework eters (e.g., connection weights), while still being able to ade-
for achieving this, because it represents an n-dimensional func- quately map the desired input/output relationship. Larger net-
tion as follows (Brown and Harris 1995): works have higher computational costs and need more training

冘 冘冘
n n⫺1 n
data to ensure that overfitting does not occur and make it more
difficult to extract knowledge from trained networks (Maier
f (x) = f0 ⫹ fi (xi) ⫹ fi, j (xi, xj ) ⫹ ⭈ ⭈ ⭈ ⫹ f1,2,...,n(x) (2)
i=1 i=1 j=i⫹1
and Dandy 2000). To minimize the number of connection
weights in MLP models, a number of construction and pruning
where f0 represents a constant (the function bias); and the other algorithms have been proposed for optimizing the number of
terms represent univariate, bivariate, and higher-order subfunc- hidden layer nodes (Reed 1993; Kwok and Yeung 1997). In
tions. In many instances, the majority of the higher-order terms addition, several approaches have been suggested for ensuring
are negligible, and the input/output mapping can be approxi- that only relevant model inputs are included (Maier and Dandy
mated by a limited number of subnetworks of reduced dimen- 1997a) as the inclusion of redundant inputs also increases net-
sion. An example of such an additive decomposition is shown work size.
in Fig. 3, where a 5D function is decomposed into one 1D The number of connection weights in B-spline AMNs is a
184 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
function of the number of basis functions. The latter is depen-
dent on the dimensionality of the input space, and conse-
quently it is desirable to use the smallest number of model
inputs and the smallest number of basis functions for each of
these. The adaptive spline modeling of observation data
(ASMOD) algorithm (Kavli 1993) is an automatic method for
obtaining the optimal structure of B-spline AMNs. It starts
with a simple model and iteratively refines the complexity of
the model by introducing new input variables, forming mul-
tivariate subnetworks (based on the ANOVA approach) and
adding new basis functions for individual input variables. Fi-
nally, the impact of pruning the network (e.g., removing basis
functions and splitting subnetworks) is investigated in an effort FIG. 4. Schematic Representation of Typical Fuzzy System
[after Bossley et al. (1995)]
to arrive at the simplest model structure. Thus, the ASMOD
algorithm uses the training data to automatically determine the
model inputs, the number and spacing of the basis functions where xk = kth real-valued input; y = output; ri, j = ijth fuzzy
(i.e., the optimum partitioning of the input space), and the rule with corresponding rule confidence ci, j; Aik = univariate
number and dimension of the subnetworks. However, the order membership function for xk, which forms part of the ith mul-
of B-spline functions has to be determined a priori. Higher- tivariate rule premise (or antecedent); and Bj = corresponding
order functions result in smoother model outputs, but increase jth univariate fuzzy membership function for the output. The
computational cost and can lead to overfitting of the data rule confidence is an indication of the strength with which a
(Brown and Harris 1994). Consequently, it is desirable to use particular rule contributes to the output. Examples of rules of
lower-order functions, provided that they are able to model the the above type in the fields of hydrology and water resources
desired relationship adequately. engineering could include
As part of the ASMOD algorithm, models are trained using
IF ‘‘Rainfall Intensity’’ is High AND ‘‘Rainfall Duration’’
a stochastic approximation version of the normalized least
mean squares rule until certain stopping criteria are met at each is Long THEN ‘‘Runoff’’ is Large, c = 0.9
stage of the refinement process (Kavli 1993). The performance
of models of varying complexity is then compared using mea- IF ‘‘Temperature’’ is High AND ‘‘Flow’’ is Low
sures that balance model size, the number of training data, and
THEN ‘‘Dissolved Oxygen’’ is High, c = 0.1
model error. Such measures, including the Bayes information
criterion (Schwarz 1978) and Akaike’s information criterion Fuzzy algorithms are formed by the union (use of the fuzzy
(Akaike 1974), are commonly used in the development of time OR operator) of individual fuzzy rules. The way in which the
series models and have recently been applied to artificial neu- fuzzy operators (IF, THEN, AND, OR) are implemented can have
ral network models [e.g., Doering et al. (1997)]. Alternative a significant impact on model performance. Traditionally, trun-
approaches for automatically determining the optimum struc- cation operators are used for this purpose (Zadeh 1973). However,
ture of B-spline AMNs have also been proposed [e.g., Menken
the use of algebraic operators has recently gained popularity [e.g.,
(1998)]. However, the ASMOD algorithm is used in this re-
Wang and Mendel (1992)] because they generally produce
search because it is well known, has good convergence prop-
smoother outputs (Brown and Harris 1994). If a real-valued out-
erties (Weyer and Kavli 1997), and has been found to perform
well compared with other methods for a wide variety of mod- put is required, a defuzzification process has to be carried out.
eling problems (Brown and Harris 1994). This is generally achieved by either using a mean of maximums
or a center of gravity defuzzification strategy (Brown and Harris
FUZZY INTERPRETATION OF B-SPLINE AMNs 1994).
A schematic of a complete fuzzy system is shown in Fig. 4.
Fuzzy logic was introduced by Zadeh (1965) to enable the
An input/output mapping is obtained by presenting a real-valued
mathematical representation of vague, linguistic concepts such
as ‘‘very high,’’ ‘‘high,’’ ‘‘medium,’’ ‘‘low,’’ etc. This is input to the model, which is fuzzified, operated on using the
achieved by allowing a particular variable to be a partial mem- knowledge stored in a fuzzy rule base, and defuzzified to produce
ber of a particular set. This is in contrast to classical logic, a real-valued output. It should be noted that once all of the com-
where a variable is either a member of a set or it is not. ponents of a particular fuzzy system have been defined, it operates
Generally, a small number of linguistic terms (e.g., high, in a deterministic manner (Procyk and Mamdani 1979). In other
medium, low), referred to as fuzzy sets, are assigned to each words, the input/output relationship is no longer uncertain, but
variable (e.g., river temperature). These fuzzy sets overlap and well defined. Brown and Harris (1994) proved that if fuzzy sys-
cover the necessary range of variation for that variable. The tems use B-spline membership functions, algebraic operators, and
degree of membership (from 0 to 1) of a real-valued input x the center of gravity defuzzification algorithm, they are equivalent
(e.g., temperature of 22⬚C) to a particular fuzzy set A (e.g., to B-spline AMNs. Consequently, the B-spline basis functions
high) is given by a ‘‘membership function’’ uA(x). This trans- may be viewed as fuzzy membership functions and can be used
formation of real-valued inputs into a degree of membership to elicit fuzzy rules. It should be noted that B-spline AMNs are
to a particular fuzzy set is called ‘‘fuzzification.’’ Membership only one type of neurofuzzy system (Nauck et al. 1997). Consid-
functions can take on a variety of forms, including Gaussian eration of other neurofuzzy approaches is beyond the scope of
and B-spline functions (Brown and Harris 1995). this paper, as its objective is to compare the performance of back-
Fuzzification of the variables enables the relationship be- propagation MLPs and B-spline AMNs.
tween the input variable(s) and the output variable to be ex-
pressed in terms of a set of fuzzy rules of the form (Brown
FORECASTING CYANOBACTERIAL
and Harris 1995)
CONCENTRATIONS IN RIVER MURRAY
ri, j: IF (x1 is Ai1 AND x2 is Ai2 AND
The River Murray is South Australia’s major surface water
⭈ ⭈ ⭈ AND xn is Ani ) THEN ( y is Bj ), ci, j (3) resource. Water is pumped several hundred kilometers to major
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 185
 iables are dominant (Maier and Dandy 1997b). Consequently,
model transparency is an important consideration.
The available data include weekly values of concentrations
of the cyanobacterium Anabaena spp., turbidity, color, tem-
perature, total phosphorus, soluble phosphorus, oxidized nitro-
gen, and total iron at Morgan as well as weekly flows at Lock
7 (Fig. 5). All data are available from 1985/1986 to 1992/1993.
It should be noted that no data are available for some factors
that can influence the size of cyanobacterial populations in-
cluding the amount of incoming radiation, the concentrations
of algal predators and competing algal species, the thermal
structure of the river, pH, and dissolved oxygen. A time series
plot of Anabaena spp. concentrations at Morgan from 1985/
1986 to 1992/1993 is given in Fig. 6. It can be seen that the
time series exhibits irregular seasonal variations but does not
contain any trends. The same applies for the time series for
the other variables.

DEVELOPMENT OF B-SPLINE AMN MODEL

The B-spline AMN model for forecasting (4 weeks in ad-
vance) concentrations of Anabaena spp. in the River Murray
at Morgan is implemented using the commercially available
FIG. 5. South Australian Reaches of River Murray software package NEUframe, version 3.0. The software makes
use of the ASMOD algorithm, described in the second section,
to determine appropriate model inputs and the number and size
cities via a number of pipelines. There are various off-take of the basis functions. Model performance is assessed by in-
points along the length of the river, including Morgan (Fig. specting plots of actual and predicted values of cell densities
5). As a result of activities in the basin, the river has been of Anabaena spp. in terms of the model’s ability to forecast
prone to outbreaks of toxic cyanobacterial (blue-green algal) the onset, peak, and duration of growth events of Anabaena
blooms. Anabaena spp. is the most common genus among the spp., because these are the three most important characteristics
nuisance cyanobacteria that occur in the lower reaches of the describing such phenomena. It is considered that a composite
River Murray and is of the most concern from a water supply error measure (e.g., RMS error) would not describe model per-
perspective. The occurrence of high numbers of cyanobacteria formance adequately. Model validation is restricted to the pe-
at Morgan is a significant water supply operational problem, riod from November to April (i.e., spring–summer), based on
because a water treatment plant is located there and water ab- the higher incidence of blooms of cyanobacteria at this time
stracted at this point is delivered to a number of South Aus- of the year (Maier et al. 1998).
tralian cities. Forecasts of concentrations of Anabaena spp. can To simulate a real-time forecasting situation and to make
be used by water management authorities to make a provision use of the maximum amount of training data available, data
to alter treatment plant operation in case of impending blooms. from 1985/1986 to 1991/1992 are used for training and data
The factors affecting the growth of Anabaena spp. include from 1992/1993 are used for validation. This is in agreement
temperature, flow, pH, dissolved oxygen, the availability of with the way the data are used by Maier et al. (1998), enabling
nutrients, the amount of available light, and the state of the a fair comparison between the results obtained using the back-
aquatic ecosystem in the river. However, the interaction of the propagation MLP and B-spline AMN approaches.
above factors is not well understood, making it difficult to Although the ASMOD algorithm automatically determines
implement management strategies aimed at alleviating the cy- which model inputs are most significant, a choice still has to
anobacterial bloom problem. The use of predictive models re- be made about how many candidate model inputs to consider.
lating cyanobacterial concentrations to various environmental This is particularly important when time series data are used,
variables can provide a better understanding about which var- as a decision is needed as to how many potential lagged values
to include for each input time series. For the sake of consis-
tency, the model inputs that are found to be significant by
Maier et al. (1998) are used as candidate model inputs in this
research (Table 1, column 2).
In addition, a second set of potential model inputs, including
lagged flow values and the latest known values of each of the
other input variables (Table 1, column 3), is considered. This
permits the investigation of whether the B-spline AMN model

TABLE 1. Lagged Inputs Used (Weeks)

Variable Input set 1 Input set 2
(1) (2) (3)
Turbidity — 1
Color 1, 2, . . . , 4 1
Temperature 1 1
Flow 1, 2, . . . , 15 1, 2, . . . , 15
Total phosphorus — 1
Soluble phosphorus — 1
Oxidized nitrogen — 1
FIG. 6. Cell Densities of Anabaena spp. at Morgan (1985/1986 Total iron — 1
to 1992/1993)

186 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

is able to detect any significant relationships between concen-
trations of Anabaena spp. and turbidity, total phosphorus, sol-
uble phosphorus, oxidized nitrogen, and total iron that are not
found to be significant by the model developed by Maier et
al. (1998). It is considered reasonable to only include lagged
flow values, because the dynamic behavior of the other vari-
ables is linked to changes in the flow regime (Maier et al.
1998).
All of the training and validation data are scaled to range
from 0.1 to 0.9 to ensure each variable receives equal repre-
sentation during the training process. B-spline basis functions
are used to achieve maximum model transparency. As the
training data are considered to be noisy, second-order B-spline
functions and the Bayes information stopping criterion are
used (Bossley et al. 1995).

RESULTS AND DISCUSSION

A comparison of the 4-week forecasts of cell densities of
Anabaena spp. in the River Murray at Morgan for the vali-
dation period obtained using the B-spline AMN and back-
propagation MLP models is shown in Fig. 7. Overall, the per-
formance of both models may be considered good, given that FIG. 7. Four-Week Forecasts of Anabaena spp. Concentra-
process-based and traditional statistical approaches have not tions at Morgan (November 1992–May 1993)
been successful in predicting the magnitude and timing of the
incidence of a particular species or species groups of algae
(French and Recknagel 1994) and that the precision of the cell
count data can be ⫾20% or more (Maier et al. 1998). The B-
spline AMN forecasts the onset, peak, and duration of the first,
smaller incidence of Anabaena spp. (weeks 5–11) very well.
The model slightly underpredicts the magnitude and duration
of the peak, and the predicted onset lags the actual onset by
approximately 4 days. However, this is much less than the
forecasting period of 28 days. In contrast, the back-propaga-
tion MLP fails to predict this incidence of Anabaena spp.
altogether. Both models provide good forecasts of the on-
set and peak of the second, major incidence of Anabaena
spp. (weeks 14–28), although the B-spline AMN performs
slightly better with regard to these two performance measures.
It gives a better representation of the sharp increase in con- FIG. 8. Schematic Representation of B-spline AMN Model for
centrations of Anabaena spp. and only underpredicts the peak Forecasting Concentrations of Anabaena spp. in River Murray
concentration of Anabaena spp. by 7%, compared with 21% at Morgan
for the back-propagation MLP model. However, the back-
propagation MLP model forecasts the duration of the major times of the year. During the critical periods, less than one-
incidence of Anabaena spp. very well, whereas the B-spline third of the rules make significant contributions to the model
AMN underpredicts it significantly. Overall, neither model output. At the time of the peak concentration of Anabaena
clearly outperforms the other. The B-spline AMN model pro- spp., rules 10, 12, 18, and 19 are most significant. Their rule
duces better forecasts of the smaller incidence of Anabaena activations are 0.19, 0.28, 0.23, and 0.10, respectively, com-
spp. and the onset and peak of the larger incidence of Ana- pared with a combined level of activation of 0.20 for the re-
baena spp., whereas the back-propagation MLP model pro- maining 15 rules. The overall activation of the flow inputs is
duces a better forecast of the duration of the larger incidence 0.52, and the combined activation of the temperature inputs is
of Anabaena spp. 0.28. Consequently, flow is the most important variable, and
By using the ASMOD algorithm, the number of model in- temperature is the second most important variable. This is in
puts to the B-spline AMN model is reduced from 20 to 4. In agreement with the results obtained when the back-propaga-
other words, the back-propagation MLP model uses all 20 tion MLP model is used (Maier et al. 1998). In general, the
model inputs (Table 1, column 2), whereas the B-spline AMN fuzzy rules indicate that incidences of Anabaena spp. occur
model only uses a subset of 4 of these values. Consequently, after the passing of a flood hydrograph and when temperatures
the B-spline AMN model is more parsimonious than the back- are high. This corresponds with the results obtained by Maier
propagation MLP model, using 16 connection weights com- et al. (1998), who carried out a sensitivity analysis on the
pared with 357. Fewer model inputs also result in a smaller trained back-propagation MLP model. However, the form of
fuzzy rule base, which is easier to interpret. The inputs that the information obtained using this approach is not as explicit
are found to be significant by the ASMOD algorithm include as that extracted from B-spline AMN models.
flow at times t ⫺ 1, t ⫺ 8, and t ⫺ 10, as well as temperature Even though the trained B-spline AMN model produces
at time t ⫺ 1. A schematic representation of the B-spline AMN good 4-week forecasts of concentrations of Anabaena spp. for
model is given in Fig. 8. the validation period considered and the fuzzy rules obtained
The fuzzy rule base extracted from the B-spline AMN are in agreement with the physical understanding of the sys-
model is given in Table 2. It can be seen that there are a total tem, it is unlikely that it provides a general representation of
of 19 fuzzy rules. However, before attempting to interpret the complex relationship between Anabaena spp. and the en-
these rules, it is useful to consider which rules fire at different vironmental variables affecting their population size. For ex-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 187
 TABLE 2. Fuzzy Rules Extracted from B-Spline AMN Model
Subnetwork Rule
number number Rule
(1) (2) (3)
1 1 IF ‘‘Flow (t ⫺ 1)’’ is Low AND ‘‘Flow (t ⫺ 8)’’ is Very Low
THEN ‘‘Anabaena Conc.’’ is Low (0.73) OR ‘‘Anabaena Conc.’’ is High (0.27)
2 IF ‘‘Flow (t ⫺ 1)’’ is High AND ‘‘Flow (t ⫺ 8)’’ is Very Low
THEN ‘‘Anabaena Conc.’’ is Low (0.69) OR ‘‘Anabaena Conc.’’ is High (0.31)
3 IF ‘‘Flow (t ⫺ 1)’’ is Low AND ‘‘Flow (t ⫺ 8)’’ is Medium Low
THEN ‘‘Anabaena Conc.’’ is Low (0.75) OR ‘‘Anabaena Conc.’’ is High (0.25)
4 IF ‘‘Flow (t ⫺ 1)’’ is High AND ‘‘Flow (t ⫺ 8)’’ is Medium Low
THEN ‘‘Anabaena Conc.’’ is Low (0.55) OR ‘‘Anabaena Conc.’’ is High (0.45)
5 IF ‘‘Flow (t ⫺ 1)’’ is Low AND ‘‘Flow (t ⫺ 8)’’ is Medium
THEN ‘‘Anabaena Conc.’’ is High (1.00)
6 IF ‘‘Flow (t ⫺ 1)’’ is High AND ‘‘Flow (t ⫺ 8)’’ is Medium
THEN ‘‘Anabaena Conc.’’ is Low (0.89) OR ‘‘Anabaena Conc.’’ is High (0.11)
7 IF ‘‘Flow (t ⫺ 1)’’ is Low AND ‘‘Flow (t ⫺ 8)’’ is Medium High
THEN ‘‘Anabaena Conc.’’ is Low (0.82) OR ‘‘Anabaena Conc.’’ is High (0.18)
8 IF ‘‘Flow (t ⫺ 1)’’ is High AND ‘‘Flow (t ⫺ 8)’’ is Medium High
THEN ‘‘Anabaena Conc.’’ is Low (1.00)
9 IF ‘‘Flow (t ⫺ 1)’’ is Low AND ‘‘Flow (t ⫺ 8)’’ is Very High
THEN ‘‘Anabaena Conc.’’ is Low (0.79) OR ‘‘Anabaena Conc.’’ is High (0.21)
10 IF ‘‘Flow (t ⫺ 1)’’ is High AND ‘‘Flow (t ⫺ 8)’’ is Very High
THEN ‘‘Anabaena Conc.’’ is Low (0.14) OR ‘‘Anabaena Conc.’’ is High (0.86)
2 11 IF ‘‘Temperature (t ⫺ 1)’’ is Low
THEN ‘‘Anabaena Conc.’’ is Low (0.81) OR ‘‘Anabaena Conc.’’ is High (0.19)
12 IF ‘‘Temperature (t ⫺ 1)’’ is High
THEN ‘‘Anabaena Conc.’’ is Low (0.61) OR ‘‘Anabaena Conc.’’ is High (0.39)
3 13 IF ‘‘Flow (t ⫺ 10)’’ is Extremely Low
THEN ‘‘Anabaena Conc.’’ is Low (0.68) OR ‘‘Anabaena Conc.’’ is High (0.32)
14 IF ‘‘Flow (t ⫺ 10)’’ is Low
THEN ‘‘Anabaena Conc.’’ is Low (0.81) OR ‘‘Anabaena Conc.’’ is High (0.19)
15 IF ‘‘Flow (t ⫺ 10)’’ is Low Medium
THEN ‘‘Anabaena Conc.’’ is Low (0.73) OR ‘‘Anabaena Conc.’’ is High (0.27)
16 IF ‘‘Flow (t ⫺ 10)’’ is Medium
THEN ‘‘Anabaena Conc.’’ is Low (0.89) OR ‘‘Anabaena Conc.’’ is High (0.11)
17 IF ‘‘Flow (t ⫺ 10)’’ is Medium High
THEN ‘‘Anabaena Conc.’’ is Low (0.45) OR ‘‘Anabaena Conc.’’ is High (0.55)
18 IF ‘‘Flow (t ⫺ 10)’’ is Fairly High
THEN ‘‘Anabaena Conc.’’ is Low (0.69) OR ‘‘Anabaena Conc.’’ is High (0.31)
19 IF ‘‘Flow (t ⫺ 10)’’ is Very High
THEN ‘‘Anabaena Conc.’’ is Low (0.70) OR ‘‘Anabaena Conc.’’ is High (0.30)
Note: Conc. = Concentration.

ample, it has been shown that water transparency and nutrients dence of Anabaena spp. However, the reverse applies in re-
can have a significant impact on the growth rate of Anabaena lation to the ability of the models to forecast the duration of
spp., although the exact relationship is not well known (Maier this incidence of Anabaena spp.
et al. 1998). However, even when turbidity, soluble phospho- The back-propagation MLP model uses 20 model inputs,
rus, total phosphorus, oxidized nitrogen, and total iron are in- and whereas the same values are used as potential input can-
cluded as model inputs (Table 1, column 3), only flow and didates for the B-spline AMN model, only 4 of these are found
temperature inputs are found to be significant when the to be significant when the ASMOD algorithm is used. Con-
ASMOD algorithm is used. The most likely reason for this is sequently, the B-spline AMN model is more parsimonious than
that there are insufficient training data to cover the full range the back-propagation MLP model, and the transparency of the
of possible environmental conditions. For example, during the former is increased. The model inputs that are found to be
time period covered by the training data, nutrient levels are significant include river flow at times t ⫺ 1, t ⫺ 8, and t ⫺
relatively high and thus unlikely to be limiting (Maier et al. 10, and water temperature at time t ⫺ 1. The fuzzy rules ob-
1998). tained indicate that incidences of Anabaena spp. are most
likely to occur after the passing of a flood hydrograph and
CONCLUSIONS when water temperatures are high. Although this is in agree-
The results of the case study carried out indicate that B- ment with the current level of physical understanding of the
spline AMN and back-propagation MLP models are suitable system, the fuzzy rules do not provide any information about
for obtaining 4-week forecasts of concentrations of the cyano- the relationship between the concentration of Anabaena spp.
bacterium Anabaena spp. in the River Murray at Morgan, and other environmental variables that are considered to be
South Australia. Overall, the forecasts obtained when the B- important (e.g., nutrients, color, and turbidity). The most likely
spline AMN model is used may be considered to be slightly reason for this is that these relationships are not significant in
better than those obtained when the back-propagation MLP is the data that are available for training the model. Conse-
used. The former is able to forecast the first (although smaller) quently, although B-spline AMN models offer a great deal of
of the two incidences of Anabaena spp. that occur during the model transparency by representing the relationships between
validation period accurately, whereas the latter fails to forecast model inputs and outputs in terms of fuzzy rules, the generality
this incidence of Anabaena spp. altogether. The B-spline AMN of the rules obtained is a function of the quality of the avail-
model also outperforms the back-propagation MLP in terms able data. However, if certain relationships are well known a
of forecasting the onset and peak of the second (major) inci- priori (which is not the case in the application considered in
188 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
this paper), they can be incorporated into the model with rel- Menken, G. J. (1998). ‘‘Structural adaptation of B-spline networks using
ative ease. genetic algorithms.’’ Engrg. Intelligent Sys. for Electr. Engrg. and Com-
munications, 6(3), 147–152.
Nauck, D., Klawonn, F., and Kruse, R. (1997). Foundations of neuro-
APPENDIX I. REFERENCES fuzzy systems, Wiley, Chichester, England.
Akaike, H. (1974). ‘‘A new look at the statistical model identification.’’ Procyk, T. J., and Mamdani, E. H. (1979). ‘‘A linguistic self-organising
IEEE Trans. on Automatic Control, 19, 716–723. process controller.’’ Automatica, 15, 15–30.
Bossley, K., Brown, M., and Harris, C. (1995). ‘‘Parsimonious neurofuzzy Reed, R. (1993). ‘‘Pruning algorithms—A review.’’ IEEE Trans. on Neu-
modelling.’’ Univ. of Southampton Res. J., Dept. of Electronics and ral Networks, 4(5), 740–747.
Comp. Sci., University of Southampton, Southampton, England. Schwarz, G. (1978). ‘‘Estimating the dimension of a model.’’ Ann. of
Brown, M., and Harris, C. (1994). Neurofuzzy adaptive modelling and Statistics, 6(2), 461–464.
control, Prentice-Hall, Englewood Cliffs, N.J. Takyi, A. K., and Lence, B. J. (1999). ‘‘Surface water quality management
Brown, M., and Harris, C. (1995). ‘‘A perspective and critique of adaptive using multiple-realization chance constraint method.’’ Water Resour.
neurofuzzy systems used for modelling and control applications.’’ Int. Res., 35(5), 1657–1670.
J. Neural Sys., World Scientific, Singapore, 6(2), 197–220. Wang, L. X., and Mendel, J. M. (1992). ‘‘Fuzzy basis functions, universal
Doering, A., Galicki, M., and Witte, H. (1997). ‘‘Structure optimization approximation, and orthogonal least squares learning.’’ IEEE Trans. on
of neural networks with the A*-algorithm.’’ IEEE Trans. on Neural Neural Networks, 3(5), 807–814.
Networks, 8(6), 1434–1445. Weyer, E., and Kavli, T. (1997). ‘‘Theoretical properties of the ASMOD
Flood, I., and Kartam, N. (1994). ‘‘Neural networks in civil engineering. algorithm for empirical modelling.’’ Int. J. Control, Taylor and Francis
I: Principles and understanding.’’ J. Comp. in Civ. Engrg., ASCE, 8(2), Group, London, 67(5), 767–789.
131–148. Zadeh, L. A. (1965). ‘‘Fuzzy sets.’’ Information and Control, 8, 338–353.
French, M., and Recknagel, F. (1994). ‘‘Modeling of algal blooms in Zadeh, L. A. (1973). ‘‘Outline of a new approach to the analysis of com-
freshwater using artificial neural networks.’’ Computer techniques in plex systems and decision processes.’’ IEEE Trans. of Sys. Man and
environmental studies V, P. Zannetti, ed., Computational Mechanics Cybernetics, 3, 28–44.
Publications, Southampton, England, 87–94.
Hornik, K., Stinchcombe, M., and White, H. (1989). ‘‘Multilayer feed- APPENDIX II. NOTATION
forward networks are universal approximators.’’ Neural Networks, 2,
359–366. The following symbols are used in this paper:
Kavli, T. (1993). ‘‘ASMOD—An algorithm for adaptive spline modeling
of observation data.’’ Int. J. Control, Taylor and Francis Group, Lon- A = fuzzy membership function;
don, 58(4), 947–967. Aik = univariate membership function for xk that forms part of
Kwok, T.-Y., and Yeung, D.-Y. (1997). ‘‘Constructive algorithms for the ith multivariate rule premise;
structure learning in feedforward neural networks for regression prob- ai = output from ith basis function;
lems.’’ IEEE Trans. on Neural Networks, 8(3), 630–645. ap = output from pth basis function;
Maier, H. R., and Dandy, G. C. (1997a). ‘‘Determining inputs for neural Bj = jth univariate fuzzy membership function for model out-
network models of multivariate time series.’’ Microcomputers in Civ.
Engrg., 12(5), 353–368.
put;
Maier, H. R., and Dandy, G. C. (1997b). ‘‘Modelling cyanobacteria (blue- cij = rule confidence associated with rij;
green algae) in the River Murray using artificial neural networks.’’ f1,2,...,n = function of order n;
Mathematics and Comp. in Simulation, 43(3-6), 377–386. p = number of nodes in input layer of MLP neural network;
Maier, H. R., and Dandy, G. C. (1999). ‘‘Empirical comparison of various rij = rule relating ith fuzzy input set to jth fuzzy output set;
methods for training feedforward neural networks for salinity fore- t = time;
casting.’’ Water Resour. Res., 35(8), 2591–2596. uA(x) = membership of x in fuzzy set A;
Maier, H. R., and Dandy, G. C. (2000). ‘‘Neural networks for the pre- w = connection weight;
diction and forecasting of water resources variables: A review of mod- wi = connection weight associated with ai;
elling issues and applications.’’ Envir. Modelling and Software, 15(1),
101–123.
wjn = connection weight associated with nth input to node j;
Maier, H. R., Dandy, G. C., and Burch, M. D. (1998). ‘‘Use of artificial xjn = nth input to node j;
neural networks for modelling cyanobacteria Anabaena spp. in the xk = kth model input;
River Murray, South Australia.’’ Ecological Modelling, 105(2/3), 257– y = model output; and
272. yj = output from node j.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 189

 SYSTEM DYNAMICS MODELING OF RESERVOIR OPERATIONS FOR
FLOOD MANAGEMENT
By Sajjad Ahmad1 and Slobodan P. Simonovic,2 Member, ASCE

ABSTRACT: There exists a strong need to explore simulation techniques that not only represent complex
dynamic systems in a realistic way but also allow the involvement of end users in model development to increase
their confidence in the modeling process. System dynamics, a feedback-based object-oriented simulation ap-
proach, is presented for modeling reservoir operations. The increased speed of model development, the trust
developed in the model due to user participation, the possibility of group model development, and the effective
communication of model results are main strengths of this approach. The ease of model modification in response
to changes in the system and the ability to perform sensitivity analysis make this approach more attractive
compared with systems analysis techniques for modeling reservoir operations. The proposed approach is applied
to the Shellmouth reservoir on the Assiniboine River in Canada. Operating rules are developed for high flow/
flood years to minimize flooding. Alternative operating rules are explored by changing reservoir storage allo-
cation and reservoir outflows. Impacts on the flood management capacity of the reservoir are investigated by
simulating a gated spillway in addition to an existing unregulated spillway. Sensitivity analysis is performed on
the reservoir levels at the start of the flood season and the outflow from the reservoir.

INTRODUCTION reservoir operations, which is efficient and simple to use com-

The application of systems analysis techniques for reservoir pared with traditional systems analysis techniques and does
management and operations has been a major focus of research not require complex mathematical description of the system.
in water resources engineering during the past four decades. System dynamics, a feedback-based object-oriented simulation
Numerous models have been reported in the literature for siz- approach, is becoming increasingly popular for modeling wa-
ing storage capacity and establishing release policy, both at ter resource systems. Palmer and colleagues (Palmer 1994;
the project planning stage and for real-time operations. Most Palmer et al. 1993, 1995) have done extensive work in river
of the techniques that have been developed or adapted to res- basin planning using SD. Keyes and Palmer (1993) used SD
ervoir operations are described in several textbooks, e.g., simulation model for drought studies. Matthias and Frederick
Loucks et al. (1981) and Mays and Tung (1992). Wurbs and (1994) have used SD techniques to model sea-level rise in a
Tibbets (1985), listing over 700 references, produced a state- coastal area. Fletcher (1998) has used system dynamics as a
of-the-art review and an annotated bibliography of systems decision support tool for the management of scarce water re-
analysis techniques applied to reservoir operations. Yeh (1985) sources. Simonovic et al. (1997) and Simonovic and Fahmy
provided an excellent review on various approaches to reser- (1999) have used the SD approach for long-term water re-
voir optimization and simulation and pointed out that, despite sources planning and policy analysis for the Nile River basin
considerable progress, research related to reservoir systems in Egypt.
analysis has been very slow in finding its way into practice, System dynamics, a feedback-based object-oriented simu-
particularly at the level of the actual operators. He attributes lation paradigm, is presented in this paper as a powerful ap-
this partly to the fact that operators usually have not been proach for modeling reservoir operations. The study deals with
involved in the formulation and development of computer simulation of reservoir operations for flood management pur-
models; partly to the fact that most applications deal with sim- poses. Operating rules have been developed for high flow
plified reservoir systems and are difficult to adapt to real sys- years to minimize flooding. Impacts on flood management ca-
tems; and partly to institutional constraints. Also, operators pacity of a reservoir have been explored by simulating a gated
and their managers are not comfortable with the degree of spillway in addition to an existing unregulated spillway. Al-
abstraction necessary for efficient application of simulation ternative operating rules have been investigated by changing
techniques to very complex systems (Russell and Campbell reservoir storage allocation, reservoir levels at the start of
1996). flooding season, and reservoir outflows.
This literature review suggests that systems analysis has its This paper outlines a general framework for modeling res-
own place in the field of reservoir management, and simulation ervoir operations using the SD approach. The theoretical back-
is an essential tool for developing a quantitative basis for res- ground of the SD simulation approach is given. Then, a gen-
ervoir management decisions (Simonovic 1992). However, eral modeling approach for reservoir simulation is presented
there is a strong need to explore simulation tools that can by introducing model structure and complex relationships
represent the complex systems in a realistic way and where among its components. The benefits of the proposed approach
operators can be involved in model development to increase are demonstrated by application to a case study of a single
their confidence in the modeling process. A promising alter- multipurpose reservoir. Finally, a discussion on results is pre-
native is the system dynamics (SD) approach for modeling sented and conclusions are drawn. Suggestions for possible
model application and extension conclude the paper.
1
Res. Asst., Dept. of Civ. and Geological Engrg., Univ. of Manitoba,
Winnipeg, MB, Canada R3T 5V6.
2
SYSTEM DYNAMICS MODELING OF RESERVOIR
Prof., Natural Res. Inst. and Dept. of Civ. and Geological Engrg., OPERATIONS
Univ. of Manitoba, Winnipeg, MB, Canada R3T 5V6.
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, System dynamics has a long history as a modeling paradigm
2000. To extend the closing date one month, a written request must be with its origin in the work of Forrester (1961), who developed
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on August 3, 1999.
the subject to provide an understanding of strategic problems
This paper is part of the Journal of Computing in Civil Engineering, in complex dynamic systems. SD models, by giving insight
Vol. 14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0190– into feedback processes, provide system users with a better
0198/$8.00 ⫹ $.50 per page. Paper No. 21629. understanding of the dynamic behavior of systems. Areas of
190 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
application of SD have always been very wide, however, with
an emphasis on socioeconomic applications. In recent years
there has been a tendency to model small-scale systems in
greater detail with more emphasis on quantitative results
(Fletcher 1998). Continuity of mass, one of the basic concepts
in SD, is also an important concept for reservoir representa-
tion. Thus, reservoir simulation problems are well suited for
application of SD solution techniques.
The SD tool used in this study to model reservoir operation
has four basic building blocks: stock, flow, connector, and con-
verter. Stocks (levels) are used to represent anything that ac-
cumulates; an example of stock would be water stored in a
reservoir. Flows (rates) represent activities that fill and drain
stocks; an example of flow includes reservoir releases or res-
ervoir inflows. Connectors are used to establish the relation-
ship among variables in the model. They are represented
graphically by the software as arrows, and the direction of the
arrow indicates the dependency relationships. They carry in-
formation from one element in a model to another element.
This information can be a quantity, a constant, an algebraic
relationship, or a graphical relationship. Converters transform
input into output. Converters can accept input in the form of
algebraic relationships, graphs, and tables. The concept of the
stocks and the flows in SD is very appropriate to deal with a
reservoir problem in water resources.
The SD approach is based on theory of feedback processes.
A feedback system is influenced by its own past behavior. This
system has a closed-loop structure that brings results from past
actions of the system back to control future actions. One class
of feedback system—negative feedback—seeks a goal and
responds as a consequence of failing to achieve the goal, e.g.,
thermostat. It is the negative feedback or goal-seeking struc-
ture of the system that causes balance and stability. A second
class of feedback system—positive feedback—generates FIG. 1. Example of Causal Loop Diagram for Reservoir
growth processes where action builds a result that generates
still greater action, e.g., population growth. A positive feed- the spillway, will positively influence the availability of water
back system structure amplifies or adds to change and thus for other purposes and will reduce flooding downstream of
causes the system to diverge or move away from the goal. reservoir, but at the cost of increased flooding upstream of
A feedback loop is the basic structure within the system, reservoir. Causal loop diagrams play an important role in de-
and levels and rates are fundamental variables within a feed- scribing the system behavior and the relationship between its
back loop. The level variables describe the condition of the components.
system at any particular time, and rate variables tell how fast The model of reservoir simulation presented in this paper,
the levels are changing. The feedback loops are also known using the SD approach, has been implemented in the STELLA
as causal loops. The diagrams representing feedback loops and environment (Tutorial 1992). The modeling tool, which is an
their variables are known as causal loop diagrams or influence object-oriented simulation environment, allows the develop-
diagrams. These diagrams show how levels and rates are in- ment of complex water resources models with significantly
terconnected to produce the feedback loops and how the feed- less effort than using traditional programming languages. It
back loops interconnect to create the system. The reading of has a user-friendly graphical interface and supports modular
these diagrams indicates how a given system might behave program development. Using this tool, the modeler defines ob-
because of its internal feedback loops and the effect that pos- jects representing physical or conceptual system components
itive and negative feedback loops have on a system. The con- and indicates the functional relationships among these objects.
vention for the diagram is that arrows depicting both the di- This mode of construction is analogous to drawing a flow chart
rection and sign of influence indicate links between state or schematic of the system to be simulated. Building on these
variables. Figs. 1(a and b) are a part of causal loop diagram strengths, the general architecture of a reservoir simulation
for a reservoir. The causal loop disgram in Fig. 1(a) indicates model is described in the next section.
that the inflow has a positive influence on reservoir storage.
The increase in reservoir storage causes an increase in the MODEL ARCHITECTURE
reservoir level, which leads to increase in area covered by the This study deals with modeling reservoir operation using
reservoir, thus causing flooding upstream of the reservoir. In- the SD approach. The model is developed for a single multi-
creased flooding upstream of the reservoir will trigger more purpose reservoir with a focus on flood management role of
release from the reservoir, thus reducing the reservoir storage. the reservoir. Then, the model is used to develop a reservoir
Similarly, increase in releases will increase river flow, thus operational policy for high-flow years to minimize flooding.
causing downstream flooding that will lead to a decrease in The model also serves as a tool for studying impacts of chang-
releases from the reservoir. The causal loop diagram in Fig. ing reservoir storage allocation and temporal distribution of
1(b) indicates that storage reallocation, e.g., increasing the reservoir levels and outflows. The general architecture of the
flood storage zone in the reservoir, will reduce flooding; how- model is presented in this section, and model sectors and the
ever, it will also reduce the supply of water for other uses. complex dynamic relationships among these sectors are also
Increasing the maximum reservoir level, e.g., putting gates on discussed.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 191
 The SD model of a reservoir can be constructed graphically Storage(t) = Storage(t ⫺ 1) ⫹ (Qin ⫺ Qout) ⭈ dt (1)
on the screen by employing basic building blocks, i.e., stocks,
flows, connectors, and converters available in the model de- Storage at time step t is equal to the storage at a previous time
velopment tool. In the reservoir model the storage is repre- step plus the difference of inflow and outflow. The solution
sented as a stock. Varying inflows and outflows cause changes interval (dt) is selected to ensure stability within the compu-
in storage volume over time. Inflows and outflows are repre- tation process.
sented by building block ‘‘flow.’’ Converters are provided to Conduit flow and spillway modules govern the flow through
extend the range of calculations that can be performed on the conduit and the spillway, respectively. System constraints,
flows, and they house data and logical/mathematical functions spillway curves, and conduit outflow capacity at different gate
to operate the system. Reservoir operating rules are also im- openings are provided to the model as part of its knowledge
plemented through converters. Connectors (directed arrows) base. Reservoir operating rules are captured in this sector using
link various elements of the model, i.e., converters, flows, and IF-THEN-ELSE statements. A screen dump of the reservoir
stock, to indicate relationships and influence. The simulation sector is shown in Fig. 2(b).
model uses differential and difference equations to describe
the complex dynamic systems. These equations are solved Upstream Flooding
with Euler’s method. Due to the modular nature of the simu-
lation tool, the reservoir model is developed in sectors that are This sector calculates the area flooded upstream of the res-
described in the next section. ervoir. Upstream flooding is triggered by a combination of
reservoir inflow, reservoir level, and reservoir outflow. The
Model Sectors number of days when the upstream area is flooded is also
counted in this sector. A screen dump of this sector is shown
A general reservoir simulation model for flood management in Fig. 2(a).
purpose can be divided into three main sectors: the reservoir,
the upstream area, and the downstream area. A schematic di-
agram of reservoir and its three sectors is shown in Fig. 2. Downstream Flooding
Details on these sectors follow. This sector calculates individual and total flooded area and
duration of flooding due to the reservoir operation at selected
Reservoir
locations between the dam and the final disposal point of the
This is the core sector of the reservoir model. Inflows and river. All sources and sinks affecting the flow in the river are
outflows from the reservoir are the main components of this introduced in this sector. A part of the downstream flooding
sector. Flow from all tributaries directly contributing to the sector is shown in Fig. 2(c).
reservoir is considered as inflow to the system. Inflow data To set up a general reservoir simulation model for flood
files, one for each flood year, are provided to the model as management purpose inflows, system constraints and operating
input. Total reservoir outflow consists of reservoir releases, rules are required. Additional data might be required depend-
spill, evaporation, and seepage losses. Reservoir storage can ing on specific objectives of the study. The data requirements
be described in terms of mass balance equation: for the model are discussed in detail in the following section.

FIG. 2. Schematic Diagram of Reservoir with Different Sectors

192 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

As output, the model provides information on variation of the Red River in Winnipeg. Thus, flooding on the Assiniboine also
reservoir levels, area flooded upstream and downstream of the contributes towards flooding of Winnipeg City. At Portage, a
reservoir, and duration of flooding. Once all sectors are de- portion of the Assiniboine River discharge can be diverted to
veloped and model relationships and operating rules are de- Lake Manitoba through a diversion channel of capacity 710
fined, the user can simply run the simulation and evaluate the m3/s. Headingly is the last station on the Assiniboine River
impacts of alternate operating rules. To demonstrate the ap- where discharge is measured before the river joins the Red
plicability of the proposed reservoir simulation approach, a River in Winnipeg.
case study is presented. The Shellmouth Dam is a zoned earth-fill embankment, ap-
proximately 1,319 m long with an average height of 19.8 m.
CASE STUDY APPLICATION A gated concrete conduit with a maximum discharge capacity
of 198.2 m3/s on the east abutment and a concrete chute spill-
The proposed SD approach for modeling reservoir opera- way on the west abutment control outflow from the dam (Wa-
tions has been applied to the Shellmouth reservoir, located on ter Resources Branch 1992). The reservoir is 56 km in length,
the Assiniboine River, close to the Manitoba/Saskatchewan 1.28 km in average width, and covers a surface area of 61 km2
border in Canada. The flooding in the Assiniboine River, when full. The elevation of the top of the dam is 435 m above
mainly caused by heavy spring runoff, has resulted in exten- mean sea level with a dead storage elevation of 417 m. The
sive damage to residential, agricultural, and industrial property. spillway elevation is 12 m higher, at 429 m. The volume of
The Shellmouth dam and reservoir were developed primarily the inactive pool below the conduit invert elevation is 12.3 ⫻
to protect the cities of Brandon and Winnipeg from floods on 106 m3. The difference between volume of reservoir at active
the Assiniboine River. Supplementary benefits of the project storage (370 ⫻ 106 m3) and crest level of natural spillway (477
include flood control to agricultural land in the river valley. ⫻ 106 m3) is the flood storage capacity of reservoir, i.e., 107
The problem of flooding upstream of the Shellmouth reservoir ⫻ 106 m3. Current operating rules specify that the reservoir
is caused by a combination of high water levels in the reservoir should be brought to 185 ⫻ 106 m3 by March 31 to accom-
and high inflows during flood season. Releases from the res- modate floods and a reservoir volume of 370 ⫻ 106 m3 is a
ervoir that exceed the channel capacity cause flooding at sev- goal during the summer months. Maximum reservoir outflow
eral locations along the river downstream of the reservoir. Cur- is limited to 42.5 m3/s to prevent flooding downstream, and
rently, there is no control structure on the spillway to regulate the outflow must be greater than 0.71 m3/s to avoid damage
spill from the reservoir. Considering these aspects of flood- to fish and aquatic life in the river system (Water Resources
ing, the objectives of the simulation modeling study were de- Branch 1995).
fined as: The data sets that were used to set up the reservoir simu-
lation model include: (1) reservoir volume curve; (2) reservoir
• Developing a reservoir operational policy for high flow area curve; (3) reservoir inflow (daily); (4) reservoir water lev-
years to minimize flooding. els (daily); (5) reservoir operating rules; (6) spillway rating
• Exploring the impacts on the reservoir flood management curve; (7) conduit rating curve; (8) relationship between depth
capacity by installing gates on an existing unregulated of water and area flooded at all points of interest upstream and
spillway. downstream of the reservoir; (9) additional flows joining the
• Developing a tool for evaluating alternative operating Assiniboine at different downstream locations; and (10) evap-
rules by changing the reservoir storage allocation, the res- oration and seepage losses from the reservoir.
ervoir levels at the start of the flood season, and the res-
ervoir outflows. MODELING APPROACH
The main objective of the study was to develop a reservoir
A schematic diagram of the study area is shown in Fig. 3. The operating policy for high-flow years/floods using the SD ap-
Assiniboine River, on which the Shellmouth reservoir is lo- proach. The five largest flood events in the history of the res-
cated, flows along the towns of Russell, St. Lazare, Miniota, ervoir, occurring in 1974, 1975, 1976, 1979, and 1995, were
Griswold, Brandon, Holland, and Portage and finally joins the selected for simulation. Only inflow was considered as input

FIG. 3. Study Area

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 193

to the reservoir. Contribution of rain over the reservoir was
not taken into account considering its insignificant influence
on the reservoir operation during a flood. Outflow through the
conduit and spills were considered as total outflow from the
reservoir. The model was developed in three main sectors, i.e.,
the reservoir, upstream flooding, and downstream flooding, as
described in general model architecture. After defining con-
nections between model sectors and components, operating
rules were incorporated in the model using logical statements
of IF-THEN-ELSE structure:

IF (Res㛭level > 429.3) AND (Spillway㛭Control = 0) AND

(TIME > 120) AND (Reservoir㛭Inflow > Unregulated㛭Spillway)
THEN (198) (2)

This statement explains that if the reservoir is full (429.3 m),

unregulated spillway is selected for simulation, it is the flood-
ing season (May), and inflow is more than outflow through
the unregulated spillway, then the conduit must be operated at
its maximum discharge capacity (198 m3/s). Similarly, if for
simulation, a gated spillway option is selected, the reservoir
level has reached between 430.5 and 431.2 m, and it is the
flooding season (late April to middle of June), then outflow
should be equal to the inflow to the reservoir.

IF (Spillway㛭Control = 1) AND (Res㛭level > = 430.5) AND

(Res㛭level < = 431.2) AND (TIME > 110) AND
(TIME < 165) THEN (Reservoir㛭Inflow) (3)

The option is provided in the model to route floods through

the reservoir using natural spill or gated spill scenarios. The
model uses the spillway rating curve and information on cur- FIG. 4. System Dynamics Model Development Life Cycle
rent reservoir level, inflows, and time of the year to make
decision about discharges through the spillway. The conduit MODEL USE
flow module defines the flow through the gated conduit. Based
on which spillway option is active, there are two different sets The model’s main control screen to run the reservoir oper-
ation simulations is shown in Fig. 5. There are five separate
of operating rules for conduit flow. Once spillway selection is
input data files for the five largest flood years. The user can
made, this information is automatically passed to conduit con-
select the flood year for simulation using a graphical tool
trol and appropriate conduit operating rules are fired. Current
(slider). Choices from 1 to 6 on the slider correspond to dif-
reservoir level, inflows, time of the year, and safe channel
ferent flood years. The spillway module has a slider that pro-
capacity downstream of the reservoir are criteria on which
vides the user with an option to choose either unregulated or
quantity of the releases through the conduit is based. At Por- gated spillway for simulation. Warnings linked to minimum
tage a part of the Assiniboine River flow can be diverted to and maximum reservoir levels have been provided in the
Lake Manitoba; this diversion is a function of levels in the model in the form of text messages and sounds. A text message
lake and capacity of the diversion channel. ‘‘Spillway will start operating soon’’ prompts the user when
Several tests recommended by Forrester and Senge (1980) the reservoir level reaches the spillway crest level. A sound
and Raymond (1985) including behavior replication test, be- warning in the model is activated when the reservoir reaches
havior sensitivity test, and behavior prediction tests are per- the minimum or the maximum level. While simulation is run-
formed to validate the model and to confirm that the model ning, the operator has control over the flow through the con-
response matches the response of the system being modeled. duit and can increase or decrease the discharges as the need
These tests are equally valid for other simulation techniques arises. As output, the model provides information on variation
and are independent of the problem domain. of the reservoir levels. The model also calculates the number
The model development process is summarized in the sche- of days when the reservoir is full or at the minimum level and
matic diagram of a model life cycle (Fig. 4). The modeling the number of days the spillway is operated. Other model out-
process starts with defining the purpose/goal of the system. put includes number of days of downstream/upstream flooding
Then, boundaries of the system to be modeled are specified. and number of days when channel capacity is exceeded due
This is followed by identification of key variables in the sys- to reservoir operation. The model also calculates total and in-
tem that affect the system the most. Then, behavior of the key dividual areas flooded at several locations along the river due
variables is described, the stocks and flows are identified, and to the reservoir operation.
their structure is mapped in the modeling tool using basic After developing the model and ensuring that it replicated
building blocks. Quantitative information, i.e., equations and the actual system behavior, several model runs were carried
data, is included in the model structure. The model is run to out. Following each run, the reservoir levels and area flooded
test the behavior. The model is evaluated and adjustments are due to the reservoir operation were carefully studied. Then
made. Once the model is replicating system behavior it is modifications of operating rules were made with a target to
ready for simulation modeling. improve the reservoir performance for flood management. The
194 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 5. Main Control Screen of Reservoir Model

calculation of the flooded area and duration of flooding due to as flow takes five days to reach Winnipeg once released from
the reservoir operation provides information on the effective- dam. With several trials it was found that a time step of one
ness of different operating policies for flood management. day provides the best trade-off between speed of calculation
Simulation techniques are not capable of generating directly and accuracy of results. Similarly, a variation of the delay
an optimal solution to a reservoir operation problem; however, function, used for flood routing, affects the timing and the
by going through several runs of a model with alternative pol- duration of flooding at downstream locations.
icies, an optimal or near optimal operating policy can be iden-
tified. RESULTS AND DISCUSSION
Simulations of the Shellmouth reservoir operation were
Daily variations of the reservoir levels for four major flood
made for the five largest historic floods with natural and gated
events (1974, 1975, 1976, and 1979) are shown in Fig. 6. A
spill scenarios. Model inputs were annual series of daily in-
summary of selected results is provided in Tables 1–3. Re-
flows to the reservoir during the five major flood events.
vised operating rules with natural spill and gated spill are
Model output includes daily variation of the reservoir level,
shown in Fig. 7 along with existing operating rules. Results
daily discharges from the reservoir, total flooded area upstream
show that with revised operating rules it was possible to op-
of the reservoir, discharges and flooded area at seven down-
erate the reservoir with only minor flooding upstream and
stream locations, and diversion to Lake Manitoba at Portage.
downstream for four out of five major flood events. While
Discharges at Headingly were used to estimate the contribution
simulating flood years 1975 and 1976, the spillway was not
of the Assiniboine River towards the flooding of Winnipeg
City. Policy alternatives were explored by changing initial res-
ervoir storage level, both at the start of simulation and at the
start of flooding season. Trials were also made to explore the
effects of changing outflow through the conduit on the varia-
tion of the reservoir level.
Models are developed to answer questions related to rela-
tively uncertain conditions; this is especially true in water re-
sources modeling. SD modeling provides a very convenient
and powerful tool to explore how changes in one system var-
iable impact other variables. For this study, a sensitivity anal-
ysis was performed on Time Step and Delays. Time Step, also
called Delta Time (DT), is the interval of time between model
calculations; thus, DT represents the smallest time interval
over which a change in numerical values of any element in
the model can occur. The Delay function returns a delayed
value of input, using a fixed lag time of delay duration. This FIG. 6. Water Levels in Shellmouth Reservoir for Flood Years
function is important to capture the timing of the flood peak, 1974, 1975, 1976, and 1979

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 195

TABLE 1. Flood Management with Revised Operating Rules
for Selected Flood Years
Down-
Reservoir Upstream stream Area
Flood Operating full flooding flooding flooded
year rules Spill (days) (days) (days) (ha)
(1) (2) (3) (4) (5) (6) (7)
1974 Existing Natural 101 4 11 630
Revised Natural 119 6 5 70
Revised Gated 125 1 0 0
1976 Existing Natural 120 2 7 370
Revised Natural 158 0 0 0
Revised Gated 158 0 0 0
1979 Existing Natural 106 5 19 1,067
Revised Natural 121 11 12 151
Revised Gated 129 0 0 0
1995 Existing Natural 161 7 47 24,530
Revised Natural 193 5 38 21,537 FIG. 8. Comparative Graph of Water Levels in Shellmouth
Revised Gated 250 30 23 16,234 Reservoir (Flood Year 1995) with Natural Spill and Gated Spill

operated and there was no flooding upstream or downstream.

TABLE 2. Impacts on Flooding by Changing Reservoir Levels In flood year 1974, the spillway was operated for only five
at Start of Year (A) and Start of Flooding Season (B) for 1976 days with a maximum discharge of 9.35 m3/s and 70 hectares
Flood Year without Using Gated Spillway
were flooded. Similarly in 1979, the spillway was operated
Initial with a maximum discharge of 37.97 m3/s and 151 hectares of
reservoir Reservoir Upstream Downstream Total area land were flooded. Simulations were made again for the flood
level full flooding flooding flooded events of 1974 and 1979 with the gated spillway option and
(m) (days) (days) (days) (ha) it was found that downstream flooding can easily be avoided
(1) (2) (3) (4) (5) without increasing flooding upstream of the reservoir.
A The flood in 1995 has a return period of 100 years and
422.2 163 0 0 0 inflows were well over three times the volume usually expe-
425.2 163 0 0 0 rienced. However, this flood event provided an opportunity to
428.3 176 2 0 152 look into the advantage of having a gated spillway. With the
429.2 195 5 0 152
B
free spill option, 166 hectares upstream and 21,371 hectares
422.2 163 0 0 0 downstream were flooded for 5 and 38 days, respectively (Ta-
425.2 174 12 17 4,790 ble 1). Peak discharge was reduced from its natural 660.92
428.3 195 13 26 20,160 m3/s to 359.45 m3/s through the reservoir operation. By routing
429.2 195 13 31 21,030 the flood of 1995 through the reservoir with the gated spillway
option, there was a reduction of about 5,000 hectares in
flooded area and flood days were reduced to 23. Maximum
TABLE 3. Impacts on Flooding by Changing Flow through outflow was reduced to 223.85 m3/s, an almost 40% improve-
Conduit ment over the unregulated spillway option. With gated spill-
Reservoir
way the maximum discharge at Headingly was 5.66 m3/s,
Conduit Reservoir Upstream Downstream level at end which is equal to the minimum required flow, as compared
outflow full flooding flooding of year with 172.2 m3/s with the free spill option. This means that
[m3/s (cfs)] (days) (days) (days) (m) Assiniboine River’s contribution towards flooding of Winni-
(1) (2) (3) (4) (5) peg City was zero. Fig. 8 shows a comparative graph of the
reservoir level variation with and without gates on the spillway
45.3 (1,600) 157 0 0 422.64
34.0 (1,200) 164 0 0 422.79 for the 1995 flood. The reservoir levels with gated spill are
19.8 (700) 196 10 0 424.37 higher compared with levels with free spill, as more water is
11.3 (400) 247 10 0 427.33 stored during the flood. The reservoir with gated spill reaches
0 (0) 250 10 0 429.25 a lower level at the end of simulation because discharges
through the conduit, over the falling limb of the hydrograph,
are higher to release water stored during the flood.
Reservoir Level at Start of Simulation
The 1976 flood year was selected to investigate how initial
water levels in the reservoir, at the start of simulation, and at
the start of flood season, affect the spillway operation and the
reservoir levels during the flood. 1976 was selected because
this year was the 2nd largest flood in terms of volume of
inflow and the spillway was not operated during the simulation
with the free spill option. Several simulations were carried out
by considering different levels at the start of the year and at
the start of flooding season. The simulations considered a
range of reservoir levels between empty (422.5 m) and full
(429.5) reservoir.
The effects of varying the reservoir level at the start of year
are shown in Fig. 9. It can be noted that if simulation starts
FIG. 7. Existing and Revised Rule Curves with Natural Spill in January, the reservoir levels at the start of simulation do not
and Gated Spill have any serious impact on the reservoir levels during the
196 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
flood and total area flooded as a result of reservoir operation level, the number of days when the reservoir is full, the num-
(Table 2). As floods arrive typically in late April or early May, ber of days of upstream and downstream flooding, and flooded
there is sufficient time to bring the reservoir to a lower level area increase as well (Table 2).
to accommodate incoming floods. Release rules are written in
a way that they adjust outflow based on the information on Flow through Conduit
inflow, time of the year, and the reservoir level. If the reservoir
simulation starts in April, the impact of initial reservoir level The variation of the reservoir level, by changing outflow
on the reservoir levels during the flood is significant (Fig. 10). through the conduit for a range between no flow and maximum
As the flood arrives soon after simulation starts, there is not flow, is shown in Fig. 11. Data in Table 3 and Fig. 11 support
enough time to bring the reservoir to a lower level to accom- that both reservoir levels at the end of simulation and the num-
modate incoming flood. By increasing the initial reservoir ber of days when reservoir is full are very sensitive to outflow
through the conduit.

Summary of Results
System dynamics proved to be a very time efficient, user
friendly, and successful approach for modeling the reservoir
operations. The simulation of the reservoir operation verified
that with the revised operating rules, the capability of the
Shellmouth reservoir for flood management can be improved.
For four out of the five largest historic floods, the reservoir
was operated without causing any significant flooding down-
stream or upstream. Due to revision of the operating rules, the
contribution of the Assiniboine River towards the flooding of
Winnipeg City is negligible (discharge at Headingly, which
once reached 170 m3/s, can be reduced to 5.66 m3/s with the
gated spillway option). Number of days when reservoir is full
or at the minimum level is very sensitive to reservoir outflows,
FIG. 9. Reservoir Levels by Varying Initial Reservoir Level at especially over the falling limb of the flood hydrograph. Res-
Start of Simulation: Curve 1 = 422.2 m; Curve 2 = 425.2 m; Curve
ervoir levels during the flood, flooded area upstream as well
3 = 428.3 m; Curve 4 = 429.2 m
as downstream, and duration of flooding are sensitive to res-
ervoir level at the start of the flooding season.
Simulation of the Shellmouth reservoir operation, consid-
ering both gated and unregulated spillway options, suggests
that installation of gates on the spillway will improve the flood
management capacity of the reservoir, especially for large
floods. Installation of gates may be justified, as flooding up-
stream due to the reservoir level rise is insignificant compared
with the flood losses downstream. However a detailed study
incorporating economic analysis and environmental impacts is
recommended before finalizing the decision to put gates on
the spillway.

CONCLUSIONS
The research reported in this paper focused on the simula-
tion of a single multipurpose reservoir for flood management
FIG. 10. Reservoir Levels by Varying Initial Reservoir Level at
purposes using the SD approach. Operating rules were revised
Day 90: Curve 1 = 422.2 m; Curve 2 = 425.2 m; Curve 3 = 428.3 m; for high-flow/flood years to minimize flooding. Impacts on
Curve 4 = 429.2 m flood management capacity of the reservoir were explored by
simulating a gated spillway in addition to the existing unreg-
ulated spillway. Alternative operating rules were explored by
changing the reservoir storage allocation, reservoir levels at
the start of flooding season, and the reservoir outflows.
The proposed SD-based simulation approach is a valuable
alternative to conventional simulation techniques. The in-
creased speed of model development, ease of model structure
modification, ability to perform sensitivity analysis and effec-
tive communication of model results are the main strengths of
the reservoir simulation model based on the SD approach.
However, one limitation is the simplified flood routing as com-
pared with sophisticated hydrodynamic models.
Because of its ease of construction and modification, SD
simulation environments facilitate rapid prototyping and
greatly reduce programming effort. Modeling effort can be di-
FIG. 11. Reservoir Levels by Varying Discharges through
rected to important tasks such as system conceptualization,
Conduit (Falling Limb of Hydrograph at Elevation 429.2–428 m): data collection, gaining input from system operators, and in-
Curve 1 = 0 m3/s; Curve 2 = 11.3 m3/s; Curve 3 = 19.8 m3/s; Curve volving stakeholders.
4 = 34.0 m3/s; Curve 5 = 45.3 m3/s The Shellmouth reservoir simulation model can be fine-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 197
tuned easily in light of operating experience, or with the help Keys, A. M., and Palmer, R. (1993). ‘‘The role of object oriented simu-
of insight provided by an expert. The SD approach offers a lation models in the drought preparedness studies.’’ Proc., 20th Ann.
Nat. Conf., Water Resour. Plng. and Mgmt. Div., ASCE, New York,
way for operators to participate in the model building process, 479–482.
thus increasing their trust in the model. The operator’s feed- Loucks, D. P., Stedinger, J. R., and Haith, D. A. (1981). Water resources
back provides direction for follow-up simulations and modi- systems planning and management, Princeton Hall, Englewood Cliffs,
fications in the model structure. N.J.
The entire modeling process is open, interactive, and trans- Matthias, R., and Frederick, P. (1994). ‘‘Modeling spatial dynamics of
parent. The model is easy to expand and modify in terms of sea-level rise in a coastal area.’’ Sys. Dyn. Rev., 10(4), 375–389.
Mays, W. L., and Tung, Y.-K. (1992). Hydrosystems engineering and
its internal structure and the amount, type, and resolution of management, McGraw-Hill, New York.
data used. Modifications in both structure and values of pa- Palmer, R. (1994). ‘‘(ACT-ACF) River basin planning study.’’ 具http://
rameters can be made easily as new information that can affect maximus.ce.washington.edu/⬃actacf/典.
the model becomes available. The model can then be rerun, Palmer, R., Hilton, K., and Hahn, M. (1995). ‘‘Willamette river ba-
the results analyzed, and the model improved as necessary. sin reauthorization feasibility study.’’ 具http://www.ce.washington.edu/
Thus, the entire dynamic modeling process facilitates adapta- ⬃willamet/典.
Palmer, R., Keyes, A. M., and Fisher, S. (1993). ‘‘Empowering stake-
tion of a general model to a particular problem. holders through simulation in water resources planning.’’ Water man-
The architecture of the Shellmouth reservoir model is ge- agement for the ’90s, ASCE, New York, 451–454.
neric in nature and can be applied to other reservoirs by re- Raymond, C. S. (1985). ‘‘Dynamic simulation models: how valid are
placing data and operating rules for that particular reservoir. they?’’ Self-report methods of estimating drug use: meeting current
Numerous simulation scenarios, in addition to what has been challenges to validity, A. B. Rouse, J. N. Kozel, and G. L. Richards,
demonstrated in this study, can be tested using the existing eds., National Institute on Drug Abuse, Bethesda, Md.
Russell, S. O., and Campbell, P. F. (1996). ‘‘Reservoir operating rules
framework. As the current model provides information on ex- with fuzzy programming.’’ J. Water Resour. Plng. and Mgmt., ASCE,
tent and duration of flooding, another sector can be added to 122(3), 165–170.
calculate damage to crops or economic losses due to lost op- Simonovic, S. P. (1992). ‘‘Reservoir systems analysis: closing gap be-
portunity of seeding. The model can be extended from a single tween theory and practice.’’ J. Water Resour. Plng. and Mgmt., ASCE,
multipurpose reservoir to a system of reservoirs. 118(3), 262–280.
Simonovic, S. P., and Fahmy, H. (1999). ‘‘A new modeling approach for
water resources policy analysis.’’ Water Resour. Res., 35(1), 295–304.
ACKNOWLEDGMENTS Simonovic, S. P., Fahmy, H., and El-Shorbagy, A. (1997). ‘‘The use of
This work was made possible by an operating grant from the Natural object-oriented modeling for water resources planning in Egypt.’’ Water
Sciences and Engineering Research council of Canada (NSERC). The Resour. Mgmt., 11(4), 243–261.
writers would like to thank the Water Resource Branch, Manitoba De- Tutorial and technical documentation STELLA威 II. (1992). High Perfor-
partment of Natural Resources, for providing the necessary data. mance Systems, Inc., Hanover, N.H.
Water Resources Branch. (1992). ‘‘Shellmouth reservoir study: prelimi-
APPENDIX. REFERENCES nary study of spillway control gates.’’ Internal Rep., Manitoba De-
partment of Natural Resources, Winnipeg, Canada.
Fletcher, E. J. (1998). ‘‘The use of system dynamics as a decision support Water Resources Branch. (1995). ‘‘Assiniboine river flooding & operation
tool for the management of surface water resources.’’ Proc., 1st Int. of Shellmouth dam.’’ Rep. Prepared for the Shellmouth Flood Review
Conf. on New Information Technolo. for Decision-Making in Civ. Committee, Manitoba Department of Natural Resources, Winnipeg,
Engrg., University of Quebec, Montreal, Canada, 909–920. Canada.
Forrester, J. W. (1961). Industrial dynamics, Productivity Press, Portland Wurbs, R. A., and Tibbets, M. N. (1985). ‘‘State-of-the-art review and
Oreg. annotated bibliography of systems analysis techniques applied to res-
Forrester, J. W., and Senge, P. M. (1980). ‘‘Tests for building confidence ervoir operations.’’ Tech. Rep. No. 136, Texas Water Resources Insti-
in system dynamics models.’’ TIMS studies in management science, tute, College Station, Tex.
Vol. 14, A. Legasto, J. Forrester, and J. Lyneis, eds., North Holland Yeh, W. W.-G. (1985). ‘‘Reservoir management and operation models: a
Press, Amsterdam. state-of-the-art-review.’’ Water Resour. Res., 21(12), 1797–1818.

198 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

 DISTRIBUTED GENETIC ALGORITHM MODEL ON NETWORK OF
PERSONAL COMPUTERS
By M. C. Balla1 and S. Lingireddy2

ABSTRACT: Despite recent advances in desktop computing power, certain civil engineering problems continue
to depend on supercomputers and powerful workstations. Although supercomputers and workstations helped in
field testing complicated mathematical models, there appears to be a significant gap in widespread technology
transfer to the industry. The sluggish progress in transfer of technology to the industry may be attributed to the
inability to implement these models on PCs coupled with high costs associated with supercomputers. The paper
reports results from an exploratory research that implemented a complicated optimization model based on a
distributed genetic algorithm on a network of PCs. PCs that formed the network were hardwired using 16bit
10Base-T Ethernet cards and were made accessible using Peer-to-Peer networking capability, which is a built-
in feature of Microsoft Windows 95/NT operating system. The inherent parallelism associated with genetic
algorithms coupled with relatively small data exchange between the computers resulted in a significant reduction
of computational time. The proposed generalized optimization framework, which can be adopted to model several
water resources related problems, is expected to accelerate the transfer of technology to the industry.

INTRODUCTION though GAs are very robust and powerful in obtaining a good
optimal solution, the excessive computational time require-
There was a significant increase in desktop computing ment (a characteristic of a GA) has been a serious hindrance
power beginning in the early 1990s. Affordable PCs with sig- in applying this technology for large-scale real-life problems.
nificant processing speed helped most civil engineering prob- Lack of affordable computational power has significantly
lems become less dependent on mainframe computers. Simul- slowed down the transfer of technology to the industry. This
taneously, there has been a growing interest in developing new prompted the researchers to look into the options of improving
or improved computational tools to solve complicated prob- the computational efficiency of the GA-based optimization
lems. However, because of these modern computational tools, models. Distributed implementation of GAs on an affordable
certain civil engineering problems continue to require large network of PCs is one such option.
computational requirements that cannot be realistically pro- Parallel implementation of GAs is not a new concept. How-
vided by even the most powerful desktop PCs. Unfortunately, ever, most efforts in the past have been focused at imple-
not every user has access to supercomputers or, in some cases, menting the GAs on supercomputers (Hung and Adeli 1994;
even to powerful workstations. The drinking water supply in- Lin et al. 1995; Foo et al. 1997), transputers (Fukuyama et al.
dustry is one such example, wherein lack of affordable com- 1996; Pollmeier et al. 1996; Oliveira et al. 1997; Yang et al.
putational resources prevented optimal utilization of treated 1997), or a network of distributed computing workstations
water and electricity resources. (Davis et al. 1994; York et al. 1994; Areibi and Vannelli 1998;
Past research concerning water supply systems has resulted Vatsa and Wedan 1999). These machines are either beyond the
in abundant technology on optimal utilization of scarce re- reach of most water utilities or too complicated for the per-
sources (Coulbeck et al. 1988a,b; Jowitt and Germanopoulos sonnel in charge of the utilities’ computational resources. In
1992; Ormsbee et al. 1992; Lingireddy 1994; Ormsbee and addition, most of the available literature on parallel computing
Lansey 1994; Ormsbee and Lingireddy 1995a,b). However, addresses noncivil engineering issues.
transfer of technology to the industry has been slow, due pri- This paper reports results from a preliminary study that im-
marily to a lack of affordable computational resources. Al- plemented a GA-based calibration model on a network of PCs.
though recent advances in desktop computational power have The focus of this article is on implementing the calibration
encouraged even small water utilities to utilize simple com- model on a network of PCs for distributed processing, and not
puter models, there appears to be a significant gap in available on details of the calibration model itself. The calibration model
and required desktop computational resources for widespread is described primarily to show the necessity and usefulness of
utilization of advanced computer models. Optimal control of distributed implementation of the model on a network of PCs.
water distribution systems and optimal calibration of distri- The described approach is suitable to other GA-based com-
bution network models are two problems where there is a need puter models as well. Specific examples in water resources
for significant computational power. engineering areas that can benefit from this technology include
A growing interest in the use of genetic algorithms (GAs) the efficient real-time operation of water distribution systems,
for many engineering problems was witnessed in the 1980s the optimal design of transient-pressure suppression devices,
and 1990s. Several GA-based models for solving complex and the optimal operation of multipurpose reservoirs.
civil engineering problems were proposed during this period
(Goldberg and Kuo 1987; Lingireddy 1994, 1996, 1998; CALIBRATION OF WATER DISTRIBUTION NETWORK
Ormsbee and Lingireddy 1995a, 1998; Sen et al. 1992). Al- MODELS
1
Grad. Student, Dept. of Civ. Engrg., Univ. of Kentucky, Lexington, Network models play an important role in the day-to-day
KY 40506-0281. operation of drinking water distribution systems. Recent en-
2
Assoc. Prof., Dept. of Civ. Engrg., Univ. of Kentucky, Lexington, KY. actment of the Safe Drinking Water Act further enhances the
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, importance of network models as they form the basis for water
2000. To extend the closing date one month, a written request must be quality modeling in pipe networks. Whereas network models
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on March 8, 1999.
are indispensable for safe and efficient operation of water
This paper is part of the Journal of Computing in Civil Engineering, distribution systems, the practical difficulties associated with
Vol. 14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003-0199– data collection introduce uncertainty in two major parame-
0205/$8.00 ⫹ $.50 per page. Paper No. 20422. ters of a network model—pipe roughness coefficients and wa-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 199
ter demand at nodes. The most economical way to avoid this mally, the explicit decision variables will include the rough-
problem is to seek indirect estimates for these parameters ness coefficient of each pipe or a group of pipes and the de-
by resorting to a calibration process. Calibration of water dis- mands at each node or a multiplication factor for a group of
tribution systems typically involves adjusting the uncertain pa- nodes. For a given vector of pipe roughness coefficients C
rameters to match measured flow and pressure data collected these constraints can be expressed
during a fire hydrant flow test with model predicted values.
Manual adjustment of uncertain parameters is highly ineffi- Lc ⱕ C ⱕ Uc (5)
cient and often unsuccessful. Use of optimization models is where Lc is lower bound; and Uc is upper bound on decision
considered appropriate for efficient calibration (Lingireddy and variable C. Likewise for a given vector of nodal demands D,
Ormsbee 1998). these constraints can be expressed
An optimization model typically comprises an objective
function that needs to be minimized/maximized subject to a LD ⱕ D ⱕ UD (6)
set of constraints. In the present study, the objective function
is formulated to minimize the square of the difference between where LD is lower bound; and UD is upper bound on decision
observed and predicted values of pressures and flow rates. variable D. Eqs. (1)–(6) constitute the optimization model for
Mathematically, the objective function (fitness function in ge- the optimal calibration of a water distribution system.
netic algorithms terminology) may be expressed The nonlinear nature of the objective function and some of
the constraints warrant the use of a nonlinear solution tech-

冘 冘
I J
nique such as a gradient-based method or a GA-based method.
minimize Z = ␣ (P io ⫺ P ip)2 ⫹ ␤ (Q jo ⫺ Q jp)2 (1) A growing interest in adopting underlying principles of natural
i=1 j=1
biological processes into engineering design tools was wit-
where P oi and P pi = observed and predicted pressures at node nessed in the 1980s and 1990s, thereby increasing the robust-
i; Q oj and Q pj = observed and predicted flow rates through pipe ness and efficiency of design tools (Goldberg and Kuo 1987;
j; I and J = number of nodes and number of pipes for which Lingireddy 1998). GAs are one such class of computational
measurements were taken; and ␣ and ␤ = normalizing factors. tools developed to aid mathematical programming problems.
State variables P and Q are implicit functions of decision var- These tools attempt to mimic the mechanics of natural selec-
iables—pipe roughness coefficients and nodal demands. The tion and genetics. In other words, GAs are a class of optimal
objective function in (1) needs to be minimized subject to a search techniques modeled after the process of natural selec-
set of implicit system constraints and a set of explicit bound tion, which performs a number of simultaneous searches in the
constraints as follows. most promising regions of the decision space. Genetic opti-
mization offers a significant advantage over the traditional
Implicit System Constraints nonlinear optimization approaches, in that it attempts to obtain
an optimal solution by continuing to evaluate multiple solution
The implicit system constraints include nodal conservation vectors simultaneously. In addition, GA methods do not re-
of mass and conservation of energy. The nodal conservation quire gradient information. Further, genetic optimization meth-
of mass requires that the sum of flows into or out of any ods employ probabilistic transition rules as opposed to deter-
junction node i minus any external demand Di must be equal ministic rules, which have the advantage of ensuring a robust
to zero. For each junction node i this may be expressed solution methodology. The present study adopted a GA-based

冘
Ni method because of its inherent advantage over a traditional
Qj ⫺ Di = 0 (2) gradient based method (Goldberg 1989).
j=1

Solution Technique
where Ni = number of pipes connected to junction node i.
The conservation of energy constraint requires that the sum The use of a GA-based method assures that the explicit
of the line loss HL and the minor losses HM over any path or bound constraints [(5) and (6)] are satisfied throughout the
loop k, minus any energy added to the liquid by a pump EP, optimization process (Lingireddy and Ormsbee 1998). There-
minus the difference in grade between two points of known fore, the optimization model essentially reduces to (1)–(3).
energy DE is equal to zero. For any loop or path k this may Because (2) and (3) are constraints to the objective function
be expressed [(1)], (1)–(3) constitute a constrained optimization problem.

冘
Nk However, the solution based on a GA may only be applied to
(HLj ⫹ HMj ⫺ EPj) ⫺ DEk = 0 (3) unconstrained optimization problems. One of the popular ways
j=1 of transforming a constrained optimization problem into an
unconstrained one is via the use of penalty functions (Lingi-
where Nk = number of pipes associated with loop or path k. reddy and Elango 1989). In this approach, the objective func-
Herein, it should be emphasized that HLj, HMj, EPj, and DEk tion is augmented (penalized) with the appropriate penalty
are all nonlinear functions of the pipe discharge Q and pipe functions calculated based on constraints that were not satis-
roughness coefficients. For example, when the pipe roughness fied either explicitly or implicitly. However, the penalty func-
is expressed in terms of the popular Hazen-William roughness tion is zero when all constraints are satisfied. The present study
coefficient, the expression for line loss HL is given by employs a bilevel solution approach in which the constraints
10.675LQ1.852
[(2) and (3)] are satisfied outside the optimization framework,
HL = (4) consequently eliminating the need for the explicit transfor-
d1.852C 4.87 mation of the constrained optimization problem into an un-
where L = length of pipe; d = diameter; and C = Hazen-Wil- constrained problem. In the bilevel approach, (2) and (3) are
liam roughness coefficient. removed from the optimization model, solved using a hydrau-
lic simulation model, and the results are linked back to the
Explicit Bound Constraints optimization model. Fig. 1 shows a schematic for the bilevel
approach. A brief description on the methodology that a GA
The explicit bound constraints are used to set limits on the employs in obtaining an optimal solution is provided in the
explicit decision variables of the calibration problem. Nor- following.
200 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 1. Bilevel Optimization Framework

In general, GAs start with a randomly generated set of de- tion evaluation is where most of the computational time is
cision vectors referred to as population. With regard to net- expended. Although evaluation of objective functions requires
work model calibration, each decision vector comprises a set a great deal of data, most of the data remain unchanged from
of decision variables, which may include group roughness val- one function evaluation to the next. The input variables that
ues and/or demand adjustment factors. Each decision variable change from one evaluation to the next are primarily the de-
in the decision vector is coded as a binary number. With this cision variables, and the number of such variables is typically
binary coding, the entire decision vector is reduced to a string <100 for most optimization models. For example, when cali-
of binary numbers (i.e., 0 and 1). Such strings are analogous brating a water distribution system comprising 10 groups of
to biological chromosomes. From the current set of decision pipes and 10 groups of demands, the number of decision var-
vectors, a subset of decision vectors is selected based on the iables associated with each objective function evaluation is 20.
value of their fitness function (i.e., objective function). Once The objective function evaluations not only require negligible
this subpopulation of decision vectors is selected, a new pop- amounts of variable input data, but they are also independent
ulation of decision vectors (i.e., offspring) is generated by per- of each other. That is, the output generated by one objective
forming the sequential operations of crossover and mutation. function does not influence the output generated by another
The crossover and mutation operations are performed on any objective function. Such independent computer operations
randomly chosen pair of strings from the set of already se- coupled with negligibly small amounts of input requirements
lected best decision vectors. A detailed discussion on GA- and output generated by the objective function make GAs
based optimization models may be found in Goldberg (1989) amenable to distributed computing. Some of the independent
and Lingireddy (1996, 1998). operations may be performed on the other computers, leaving
the primary computer performing important tasks related to
NECESSITY FOR DISTRIBUTED COMPUTING the optimal genetic search. The external computers can pro-
vide essential information to the primary computer with a min-
Although genetic optimization models are more robust than imum of interruption.
the conventional gradient-based nonlinear optimization mod- In the case of the optimal calibration model, after the GA
els, they require a large number of function evaluations. If the generates the decision vectors, some of the decision vectors
optimization model searches through 100 generations and if may be transferred to other computers for distributed process-
each generation has a population of 30 decision vectors, then ing. The results accumulated from all the processors in the
the model has evaluated the objective function 3,000 times. distributed computing network may then be used to generate
Calibration of water distribution systems usually involves the the next set of decision vectors through a genetic search. Such
simultaneous evaluation of several fire hydrant flow tests, thus an approach would speedup the genetic search process of iden-
requiring many hydraulic simulation evaluations to evaluate tifying an optimal solution to a given problem. The process of
the objective function. Even a powerful simulation model such transmitting decision variables to external computers and re-
as KYPIPE (Wood 1995) typically requires about 0.5 CPU ceiving results from external computers may be automated if
second for a distribution system of about 1,000 pipes on an the primary and external computers are networked. The pres-
IBM compatible PC with a Pentium 133-MHz processor. If ent study involved fabrication of a network of PCs, which was
the calibration exercise involves 10 hydrant flow tests, then then used to implement a GA-based optimization model as
the optimization model uses the simulation model 30,000 explained next.
times for 100 generations of search requiring a total CPU time
of about 4 h. Although this may not be a great concern for a METHODOLOGY
calibration exercise, it certainly is a serious drawback when
Network of PCs
genetic optimization models are used for other purposes such
as real-time operation of water distribution systems. The network of PCs used for the present study was designed
Close observation of the proposed calibration model reveals based on Microsoft Windows 95/NT native peer-to-peer net-
that it is highly amenable to distributed computing. Each gen- working protocol. Contrary to a server-based network (such as
eration of the genetic search comprises several decision vec- Novell Network), a Peer-to-Peer network does not demand a
tors (usually between 20 and 50) that need to be evaluated for dedicated server and sets no hierarchy among computers. This
their objective functions independently. In serial computing, is a major advantage of using a peer-to-peer network instead
the objective functions are evaluated considering one decision of a server-based network. Absence of a designated server also
vector at a time. In models based on GAs, the objective func- eliminates the requirement for more expensive components
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 201
 FIG. 2. Parallel Implementation of Optimal Calibration Model

such as a large hard disk space, extra memory (RAM), etc. the main-processor or on the sub-processors. It may be noted
The only additional components required to build a peer-to- that a preference was neither intended nor warranted when
peer network are the network cards, connecting cables, and designating a particular PC as a main-processor.
preferably a network hub. Because a peer-to-peer networking When the distributed calibration program is initialized on
capability comes standard with the popular Microsoft Win- the PCNet, all N computers read the basic data pertaining to
dows 95/NT operating system, the cost of buying network the distribution system model being calibrated and perform a
software was also eliminated. preliminary hydraulic analysis that initializes certain variables
The present study employed a ‘‘star’’ type topology in set- for subsequent optimization runs. Next, the main-processor
ting up the computer network. Topology refers to the arrange- randomly generates an initial set of decision vectors, retains
ment or physical layout of computers, cables, and other com- some of the decision vectors, and sends the remaining to the
ponents in the network. If the computers are connected to sub-processors. The main-processor then evaluates the objec-
cable segments that branch out from a single point or hub, the tive function values for the decision vectors that it retained.
topology is known as a star. The advantage of star topology At the same time, the sub-processors evaluate the objective
is that if one of the computers or the cable that connects it to function values for the decision vectors provided by the main-
the hub fails, only the failed computer will not be able to send processor. After completing an objective function evaluation
or receive network data. The rest of the network continues to for all the decision vectors, the sub-processors transmit the
function normally. A 16bit 10Base-T Ethernet network inter- objective function values back to the main-processor for fur-
face card was installed in each computer to be networked. This ther processing. The main-processor, after evaluating its own
was the only hardware modification needed for an ordinary share of decision vectors, gathers the output provided by the
PC to make use of it in a network. The computers with Eth- sub-processors and uses the collective information to produce
ernet cards were then connected in a star type topology using a new generation of decision vectors via crossover and mu-
standard 10Base-T connecting cables and a network hub. The tation operations. This process continues until a prespecified
following section illustrates the implementation of the distrib- convergence criterion has been satisfied. Fig. 2 illustrates the
uted calibration model on a network of N PCs. The network above process in the form of a flowchart.
of PCs hereinafter, will be referred to as the PCNet. One issue that has not been completely explored yet is load
sharing; that is, how to split the decision vectors among the
Implementation Strategy computers in the PCNet? In other words, how many decision
vectors should be retained by the main-processor, and how
Taking advantage of the built-in peer-to-peer networking ca-
many decision vectors should be sent to each sub-processor in
pability of the Windows 95/NT operating system, one of the
the PCNet? The present study, however, divided the decision
PCs in PCNet was designated as a main-processor and the
vectors based on the processor’s speed of each computer while
remaining N-1 as sub-processors. The PCs designated as sub-
allowing extra computational time to the main-processor to
processors receive input (decision vectors) from the main-pro-
perform the crucial crossover and mutation operations. More
cessor, process the data (evaluate the objective function for
information on load balancing issues and strategies may be
each decision vector), and transmit data back to the main-
found in Vatsa and Wedan (1999) and Woo et al. (1997).
processor. On the other hand, the PC designated as a main-
processor not only performs the genetic search to generate
APPLICATIONS
decision vectors but may also evaluate objective function val-
ues for some of the decision vectors. In addition, the main- Applicability and robustness of the proposed parallel opti-
processor sends the decision vectors to the sub-processors and mal calibration model is demonstrated by applying the model
receives the output corresponding to those decision vectors to calibrate three different water distribution systems, A, B,
from the sub-processors. This process was automated by writ- and C. The computer network fabricated for the present study,
ing/reading information (decision vectors, objective function PCNet, was composed of three PCs, X, Y, and Z. PC X had
values, etc.) to and from a few scratch files located either on a 166-MHz Pentium processor, and Y and Z had 133-MHz
202 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
Pentium processors. The primary operating system on all three applying the calibration model. Pipelines in the network were
machines was Microsoft Windows 95. Computer Y was des- divided into four groups based on their pipe material and di-
ignated as a main-processor and the remaining two as sub- ameter. A roughness value was designated for each group, and
processors. Based on processing speed alone, computer X all pipes in the group were assigned that roughness value (Ta-
(166-MHz) will get 38% of the decision vectors, while com- ble 1). Fire flows were simulated making use of the KYPIPE
puters Y (133-MHz) and Z (133-MHz) get 31% each. To make hydraulic simulation model, and the corresponding residual
allowance for the extra computational effort required by the pressures were noted. Table 2 shows the simulated data. As-
main-processor, the present study reduced the burden to the suming group roughness values as the unknown decision var-
main-processor (computer Y) by 4% and apportioned that iables, the calibration model was applied to predict them using
equally to the other computers. This resulted in 40, 27, and the simulated flow-pressure data (Table 2). The predictions
33% share of decision vectors for the computers X, Y, and Z, from a good calibration model should be close to the actual
respectively. Although this division reduced waiting time for roughness coefficients with which the flow-pressure data were
the main-processor and sub-processors in sending/receiving simulated. Upper and lower bounds were imposed on group
data, further investigation is warranted in automating this pro- roughness coefficients as listed in Table 1. Group roughness
cess so that there will be a minimum unwanted waiting time values to be predicted by the calibration program that were
for the main-processor and sub-processors. The calibration already known did not have any influence on the optional
model based on a bilevel optimization framework developed search process other than in setting the upper and lower
in this study was installed on all three computers in the PCNet. bounds. The calibration model was designated to have 30 de-
First, the calibration model running stand-alone on individ- cision vectors in each generation.
ual computers was applied to calibrate the distribution sys- First, the calibration model, running on one computer at a
tems. The corresponding computational times were then com- time (i.e., all 30 decision vectors were processed by a single
pared to those required when the model implemented on computer), was applied to obtain the roughness coefficients of
PCNet was applied for calibration. The three water distribution the distribution system. The roughness coefficients predicted
systems employed differed in size and complexity. Distribution by the calibration model are presented in Table 1. As expected,
system A was composed of 52 pipes, system B was composed each run resulted in identical group roughness predictions after
of 435 pipes, and system C was composed of 850 pipes (Figs. going through 35 generations of genetic research. Residual
3, 4, and 5). System A represents a highly skeletonized model pressures computed by the calibration model are presented in
of an existing water distribution system. Systems B and C
represent existing water distribution systems, one serving a
rural community and the other serving a suburban community,
respectively.
The calibration exercise as applied to distribution system A
is described in detail whereas the others follow the procedure.
For distribution system A, being an example network model,
the data required for calibration (such as hydrant flow rates
and corresponding residual pressures) were simulated prior to

FIG. 5. Schematic for Distribution System C

FIG. 3. Schematic for Distribution System A TABLE 1. Group Roughness Coefficients for Distribution Sys-
tem A
Roughness Coefficients

Used for Calibration

Bounds Calibrated
Number obtaining
Group of pipes simulated Individual
number in group dataa Upper Lower computers PCNet
(1) (2) (3) (4) (5) (6) (7)
1 10 118 125 100 119 119
2 25 72 85 60 72 72
3 10 89 110 75 89 89
4 7 102 110 90 98 98
a
As found in Table 2.
FIG. 4. Schematic for Distribution System B

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 203

 TABLE 2. Simulated and Calibrated Flow/Pressure Values for Distribution System A

Tank Levels Residual Pressure (psi)

Calibrated
Simulated fire
Simulation hydrant flow Demand multiplication Tank A Tank B Individual
number (gal./min) factor (ft) (ft) Simulated computers PCNet
(1) (2) (3) (4) (5) (6) (7) (8)
1 1,500 1.20 1,090 1,090 41.7 41.7 41.7
2 1,200 1.50 1,110 1,110 33.4 33.2 33.2
3 1,300 0.50 1,090 1,090 23.7 23.6 23.6
4 1,500 1.00 1,080 1,080 21.7 21.6 21.6
5 1,400 0.75 1,100 1,090 29.1 29.3 29.3
Note: 1 ft = 0.3048 m; 1 psi = 6.895 kPa; 1 gal/min = 0.0631 L/s

TABLE 3. Summary of Computational Time Requirements DISCUSSION OF RESULTS

Computational Time (s) As evident from Table 3, significant reduction in computa-
Distribution Distribution Distribution tional time may be achieved by implementing the optimal cal-
PC(s) system A system B system C ibration model on a network of PCs. Furthermore, the reduc-
(1) (2) (3) (4) tion in computational time is more significant when calibrating
X (166-MHz) 70.96 1,404.67 1,940.00 large distribution systems. It may be noted that the number of
Y (133-MHz) 89.31 1,756.46 2,473.82 data items transferred across the network during the optimi-
Z (133-MHz) 89.00 1,730.95 2,455.72 zation process is small and remains practically at that level
PCNet 35.15 638.24 851.78 irrespective of the network size. However, large distribution
systems require more computational time per hydraulic sim-
ulation compared to smaller systems. A typical calibration run,
Table 2 and are very close to the simulated values, implying based on a genetic optimization model, requires several hun-
a good calibration. Table 3 shows the computational time dred of such hydraulic simulations. Therefore, the use of
needed by individual computers. PCNet to calibrate larger distribution systems results in sig-
Subsequently, the calibration model running on PCNet was nificant gains in computational time. It may be of interest to
applied to calibrate distribution system A. The decision vectors some of the readers to know how fast a distribution system
were divided in the ratio of 40:27:33 (X:Y:Z) that resulted in may be calibrated making use of PCNet compared to a single
12 decision vectors for sub-processor X, 8 decision vectors for PC. Table 4 shows the computational advantage (CA), defined
main-processor Y, and 10 decision vectors for sub-processor as a measure of quickness, in calibrating distribution systems
Z. Because there are four decision variables (group roughness of different sizes. If TP is time required to calibrate a distri-
coefficients) in each decision vector, the number of values (real bution system on PCNet and TI is time required to calibrate
numbers) provided by the main-processor to sub-processor X the same distribution system using one of the computers that
as input is 48. Similarly, the number of values provided by formed the PCNet, CA may be computed as follows:
the main-processor to sub-processor Z as input is 40, Sub- TI
processors X and Z, upon receiving the decision vector data, CA = ⭈ 100 (7)
evaluate the corresponding objective function values. This in- TP
volves performing hydraulic simulation of the distribution sys- The average CA for distribution system C (large system) was
tem [solution to (2) and (3)] five times per decision vector 269%, indicating that the calibration process was 2.69 times
(one for each fire flow reading). That is, sub-processor X upon faster on PCNet compared to individual PCs. The average CA
receiving data from the main-processor performs 60 (12 times for distribution system A (small system) was only 236%. How-
5) hydraulic simulations on distribution system A before re- ever, the necessity for distributed processing is more when
turning the objective function values to the main-processor. dealing with large distribution systems, and a higher CA for
Upon completion, the sub-processors X and Z return 12 and such systems is an added advantage.
10 objective function values, respectively, to the main-proces- Although the authors have used a calibration model to dem-
sor. Clearly, the time required for data exchange is small com- onstrate the efficacy of the PCNet, the underlying GA frame-
pared to the time required in evaluating the objective func- work is what is responsible for the gain in computational ef-
tions. In addition, the time required for data exchange may ficiency. Any other optimization model based on GAs, in
practically remain the same irrespective of the size of the water which the function evaluation takes most of the computational
distribution system being calibrated. The resulting group time, can make use of the advantages offered by PCNet. GAs
roughness coefficients are presented in Table 1. As expected, may be used advantageously to develop efficient optimization
these values are the same as those predicted by the individual
computers. Similarly, residual pressures obtained by the par- TABLE 4. CA of Using PCNet Compared to Individual Com-
allel calibration model (Table 2) are the same as those obtained puters
by the individual computers. Time taken by PCNet to calibrate
CA (%)
distribution system A is shown in Table 3. This process was PC that was
repeated for the other two distribution systems, B and C. Table compared to Distribution Distribution Distribution
3 summarizes the corresponding computational time require- network system A system B system C
ments. In all these cases, a probability of crossover of 0.6, a (1) (2) (3) (4)
probability of mutation of 0.03, and a gene length of 5 were X (166-MHz) 201 220 228
employed. Based on a sensitivity analysis, these values were Y (133-MHz) 254 275 290
Z (133-MHz) 253 271 288
found to give better results for hydraulic network calibration Average 236 255 269
problems (Lingireddy and Ormsbee 1998).
204 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
models for many of the optimal control problems in the water Goldberg, D. E. (1989). Genetic algorithms in search, optimization and
resources field (e.g., pump scheduling in water distribution machine learning, Addison-Wesley, Reading, Mass.
Goldberg, D. E., and Kuo, C. H. (1987). ‘‘Genetic algorithms in pipeline
systems, ground-water pumping, and multipurpose reservoirs). optimization.’’ J. Comp. in Civ. Engrg., ASCE, 1(2), 128–141.
Successful implementation for such models on PC networks Hung, S. L., and Adeli, H. (1994). ‘‘Parallel genetic/neural network learn-
for distributed processing will significantly reduce the com- ing algorithm for MIMD shared memory machines.’’ IEEE Trans. Neu-
putational time, thus facilitating a rapid transfer of technology. ral Networks, 5(6), 900–909.
The simplicity associated with setting up of a network of PCs Jowitt, P. W., and Germanopoulos, G. (1992). ‘‘Optimal pump scheduling
as described in the paper encourages the utilities to embrace in water-supply networks.’’ J. Water Resour. Plng. and Mgmt., ASCE,
118(4), 406–422.
this technology without hesitation. Although preliminary re- Lin, J., Bartal, Y., and Uhrig, R. E. (1995). ‘‘Predicting the severity of
sults from this study are highly encouraging, further study is nuclear power plant transients using nearest neighbors modeling opti-
warranted to investigate the performance of the models when mized by genetic algorithms on a parallel computer.’’ Nuclear Technol.,
implemented on a network comprising more than three PCs. 111(1), 46–62.
In this context, it may be worthwhile exploring the viability Lingireddy, S. (1994). ‘‘Genetic algorithms in groundwater manage-
of a custom-built PC with multiple motherboards that accesses ment.’’ 1994 Nat. Conf. on Hydr. Engrg., ASCE, Reston, Va.
Lingireddy, S. (1996). ‘‘Optimal land grading based on genetic algo-
(read/write) data from a single hard drive in place of a network rithms.’’ J. Irrig. and Drain. Engrg., ASCE, 122(4), 183–188.
of several PCs. Lingireddy, S. (1998). ‘‘Aquifer parameter estimation using genetic al-
gorithms and neural networks.’’ Civ. Engrg. Sys., 15, 125–144.
CONCLUSIONS Lingireddy, S., and Elango, K. (1989). ‘‘Optimal design of water distri-
bution networks with head-dependent outflows using augmented La-
There are several mathematical models in the field of water grangian functions.’’ J. Instn. of Engrs., Envir. Engrg. Div., India,
73(EN 2), 27–33.
resources engineering that require expensive computational Lingireddy, S., and Ormsbee, L. E. (1998). ‘‘Neural networks in optimal
hardware such as super computers or powerful workstations. calibration of water distribution systems.’’ Artificial neural networks
Lack of affordable computational power has hampered transfer for civil engineers: Advanced features and applications, I. Flood and
of technology from research level to the industry. The present N. Kartam, eds., ASCE, Reston, Va., 277, 53–76.
study demonstrates that a mathematical model based on a GA Oliveira, V., Pina, A., and Proenca, A. (1997). ‘‘Parallel genetic algo-
implemented on a network of PCs may significantly reduce rithms in a networked workstation environment.’’ Applications of high
performance computing in engineering, Santiago de Compostela,
the computational requirements. The study employed a net- Spain.
work of three PCs that were networked using the built-in Peer- Ormsbee, L. E., Chase, D. V., and Lingireddy, S. (1992). ‘‘Optimal op-
to-Peer networking protocol of the popular Microsoft Win- eration of water distribution systems.’’ Final Rep., National Science
dows 95/NT operating system. One of the applications, Foundation, Washington, D.C.
involving calibration of a 850-pipe distribution system, shows Ormsbee, L. E., and Lansey, K. E. (1994). ‘‘Optimal control of water-
supply pumping systems.’’ J. Water Resour. Plng. and Mgmt., ASCE,
that the calibration process could be 2.70 times faster when 120(2), 237–252.
implemented on a network comprising three PCs compared to Ormsbee, L. E., and Lingireddy, S. (1995a). ‘‘Optimal control of water
using a single computer. Although the study focused on an supply pumping systems using genetic algorithms and artificial neural
optimal calibration model based on a GA, the proposed opti- networks.’’ Int. Fedn. for Automatic Control Symp. on Large Scale Sys.,
mization framework may easily be adapted to many other wa- ’95, Elsevier Science, Inc., London.
ter resources engineering problems. Ormsbee, L. E., and Lingireddy, S. (1995b). ‘‘Nonlinear heuristic for
pump operations.’’ J. Water Resour. Plng. and Mgmt., ASCE, 121(4),
302–309.
APPENDIX. REFERENCES Pollmeier, K., Burrows, C. R., and Edge, K. A. (1996). ‘‘Mapping of
large scale fluid power system simulations on a distributed memory
Areibi, S., and Vannelli, A. (1998). ‘‘Distributed advanced search tech- parallel computer using genetic algorithms.’’ Fluid Syst. Technol. Div.
niques for circuit partitioning.’’ Proc., 11th Can. Conf. on Electr. and Publ., ASME, New York, 3, 83–91.
Comp. Engrg., IEEE, Piscataway, N.J., Vol. 2, 553–556. Sen, M. K., Gupta, A. D., Stoffa, P. L., Lake, L. W., and Pope, G. A.
Coulbeck, B., Brdys, M., Orr, C., and Rance, J. (1988a). ‘‘A hierarchial (1992). ‘‘Stochastic reservoir modeling using simulated annealing and
approach to optimized control of water distribution systems: Part I. genetic algorithm.’’ Proc., 1992 SPE Annu. Tech. Conf. and Exhibition.
Decomposition.’’ J. Optimal Control Applications and Methods, 9, 51– Vatsa, V. N., and Wedan, B. W. (1999). ‘‘Parallelization of a multiblock
61. flow code: An engineering implementation.’’ Comp. and Fluids, 28(4),
Coulbeck, B., Brdys, M., Orr, C., and Rance, J. (1988b). ‘‘A hierarchial 603–614.
approach to optimized control of water distribution systems: Part II. Woo, S., Yang, S., Kim, S., and Han, T. (1997). ‘‘Task scheduling in
Lower level algorithm.’’ J. Optimal Control Applications and Methods, distributed computing systems with a genetic algorithm.’’ Proc., 2nd
9, 109–126. Conf. High Perf. Computing on the Information Superhighway, HPC
Davis, M., Liu, L., and Elias, J. G. (1994). ‘‘VLSI circuit synthesis using Asia ’97, IEEE, Los Alamitos, Calif., 301–305.
a parallel genetic algorithm.’’ Proc., IEEE Conf. Evol. Comput., Vol. Wood, D. J. (1995). KYPIPE hydraulic network analysis software, user’s
1, IEEE, Piscataway, N.J., 104–109. manual, Civil Engineering Software Center, University of Kentucky,
Foo, H. Y., Song, J., Zhuang, W., Esbensen, H. K., and Ernst, S. (1997). Lexington, Ky.
‘‘Implementation of a parallel genetic algorithm for floorplan optimi- Yang, H., Yang, P., and Huang, C. (1997). ‘‘Parallel genetic algorithm
zation on IBM SP2.’’ Proc., 2nd Conf. High Perf. Computing on the approach to solving the unit commitment problem: Implementation on
Information Superhighway, HPC Asia ’97, IEEE, Los Alamitos, Calif., the transputer networks.’’ IEEE Trans. Power Sys., 12(2), 661–668.
456–459. York, T. A., Patrick, D., Stinchcombe, J., Szolkowski, M. S., Tyrell, S.
Fukuyama, Y., Nakanishi, Y., and Chiang, H. (1996). ‘‘Parallel genetic G., and Green, P. R. (1994). ‘‘SPARC-GAP: A parallel genetic algo-
algorithm for generation expansion planning.’’ Int. J. Eng. Intell. Syst. rithm processing platform.’’ IEE Colloquium on High Perf. Applica-
Electric Eng. Commun., 4(1), 49–56. tions of Parallel Arch., IEE, Stevenage, U.K., No. 26.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 205

 COMPUTERIZED SYSTEM FOR APPLICATION OF VALUE
ENGINEERING METHODOLOGY
By Sadi Assaf,1 Osama A. Jannadi,2 and Ahmed Al-Tamimi 3

ABSTRACT: This paper aims to present an easy, quick, and accurate way of implementing value engineering
(VE) studies by introducing an integrated computerized system for the VE technique and life-cycle cost as a
checkout system. The VE computer program was designed using a database computer program, FoxPro Window.
The computerized version of the VE study consists of several interactive panels, representing the systematic
approach (job plan) of a VE study, through which all the information and data of the VE study will be input.
The program will help the user to come up with the best selected idea, based on certain calculations that will
be done automatically by the VE program. Moreover, the VE case studies that will be investigated by the
program will be stored, and thus the program will also serve as a data bank for these case studies as references
for future studies. The objective of this paper is not to create a smart VE system, but it can be considered a
major step toward this objective.

INTRODUCTION Therefore, the main objective of this study is to introduce the

principle of computerization to help eliminate some of these
Saving money and at the same time providing better value complications. The proposed integrated computerized system
is a concept that everyone can support. The benefits are for the VE technique and life-cycle cost (LCC) system will
spreading out investment, building more for less money, in- have the following advantages:
creasing efficiency, and cutting down the dependency on en-
ergy. Intensive buildings, which mean high energy, high cost,
• It will make the VE study easy to use, which in turn will
and more plant facilities, need to be recognized today and
encourage an increasing number of engineers to apply the
pursued in the future. People are interested in saving money
VE technique to their projects.
and looking for a sound investment with a high return for their
• It will save time and money for the project.
invested money. In the constant battle to find a better way to
• It will eliminate any possible human error.
fight inflation, the application of value engineering (VE) is
• It will serve as a database for keeping a record of the case
coming to the forefront. The VE methodology is a proven
studies that undergo VE analysis.
technique used to combat escalating costs. Constraints on
• It will serve as a teaching tool for new users of VE.
spending, the tightening of investment money, and a desire to
retain the maximum value for money spent has all helped in-
crease the acceptance of VE. The initial success of VE on The main objectives of the study are
construction projects within the governmental agencies in
Saudi Arabia has brought about an awareness of the merits of • To develop an integrated computerized system for the VE
VE by architects, engineers, and contractors. technique and the LCC system in the Kingdom of Saudi
The Kingdom of Saudi Arabia has a stable and growing Arabia
economy. Since the discovery of oil in the early 1930s, the • To build a computerized project file history that contains
kingdom’s economy has kept growing at a rapid rate. Con- records of various types of projects and to serve as a
struction activity, which is the second largest component of model for similar projects in the future
gross domestic product after the oil industry, has concentrated
on infrastructure development and is growing steadily (Cor- SCOPE
coran 1981). It is for this reason that knowledge of the VE
technique is of the utmost importance to optimize expenditure The scope of the study will be limited to the following pa-
and maintain the quality of these projects. rameters:
Due to the large number of projects currently undertaken in
the kingdom and the number of projects planned for the future, • The VE study will be applied only to construction projects
VE is a necessary tool and a key to success in any project. conducted in the Kingdom of Saudi Arabia.
Lack of knowledge about the VE concept and its financial and • Although an expert system will be required to replace the
technical benefits was found to be the main reason for the manual work completely, developing this expert system is
limited application of VE in Saudi Arabia (Sadi et al. 1996). beyond the scope of this study. The main difference be-
tween this program and an expert system is that the latter
1
Prof., Dept. of Constr. Engrg. and Mgmt., Coll. of Envir. Des., Box should have the ability to assign the various alternatives
680, King Fahd Univ. of Petr. and Minerals, Dhahran 31261, Saudi Ara- for different types of projects. The expert system, with
bia. minimum user interface, would be able to select the pre-
2
Assoc. Prof. and Dean, Coll. of Envir. Des., Box 1978, King Fahd
ferred alternative.
Univ. of Petr. and Minerals, Dhahran 31261, Saudi Arabia. E-mail: jan-
[email protected]
3
Grad. Student, Dept. of Constr. Engrg. and Mgmt., Coll. of Envir. Theory and Fundamentals of VE
Des., King Fahd Univ. of Petr. and Minerals, Dhahran 31261, Saudi Ara-
bia. The main objective of VE analysis is to achieve the required
Note. Editor: Sivand Lakmazaheri. Discussion open until December 1, performance and efficiency of a product or service at the low-
2000. To extend the closing date one month, a written request must be est possible cost. However, there are other reasons that neces-
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on November 18,
sitate the use of the VE technique. These reasons are the ones
1997. This paper is part of the Journal of Computing in Civil Engi- that produce high cost and lower efficiency or quality, and they
neering, Vol. 14, No. 3, July, 2000. 䉷ASCE, ISSN 0887-3801/00/0003- include a lack of information, a lack of creative ideas, honest
0206–0214/$8.00 ⫹ $.50 per page. Paper No. 17026. but wrong beliefs, misconceptions resulting from past experi-
206 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
ences, temporary circumstances, and habits and attitudes systems, services, or designs. During planning and design,
(Mudge 1988). choices are made between reasonable estimates of alternative
As defined by the Society of American Value Engineers, VE courses of action.
is the systematic application of recognized techniques that An organized and systematic approach to the VE job plan
identify the function of the product or service, establish a mon- is the key to success in the VE study. The job plan is the road
etary value for that function, and provide the necessary func- map for defining the required task and determining the most
tion reliably at the lowest possible cost. The purpose of VE is economical combination of functions to achieve the task. It is
to provide each individual with a means of skillfully, delib- through the job plan that the study identifies the key areas of
erately, and systematically analyzing and controlling the total unnecessary cost and seeks new and creative ways of perform-
cost of a product or service. Total cost means the sum of all ing the same function.
efforts and expenditures made in the development of a product Zimmerman and Hart (1982) defined VE by what is true
or service. The systematic analysis and development of alter- and what is not true about the VE concept. They stated that
native means of achieving the functions that are desired and VE is a systematic and multidisciplined management tech-
required accomplish this total cost control. The purpose of a nique. On the other hand, they stated that VE is not a design
VE systematic approach is well demonstrated when the user reviewing, cost lowering, or quality control process.
is able to define and segregate the necessary from the unnec-
essary and thereby develop alternative means of accomplishing VE Job Plan
the necessary at a lower total cost (Miles 1972).
There are three basic terms to be used during the discussion The job plan is a systematic approach that follows several
of the VE systematic approach, namely, value, function, and basic phases, including the information phase, function anal-
product. Value is defined as the lowest cost to reliably provide ysis phase, creative phase, evaluation phase, development
the required functions or service at the desired time or place phase, and recommendation phase. The following paragraph
and with the essential quality. The term function as used in summarizes these phases.
VE is defined as that which makes the product or service work The information phase (Tables 1 and 2) involves a definition
or sell. Work directly relates to the use value, and sell relates of the project, the obtaining of background information lead-
to the personal or prestige value. The third term, product, is ing to product design, limitations on the project, and sensitivity
anything produced or obtained as a result of some operation to the cost involved in owning and operating. Some of the
or work such as generation, growth, labor, study, or skill. most relevant areas of information needed to conduct the VE
studies are
VE Systematic Approach
• Design criteria (system requirement)
To maximize investment, VE has to be undertaken as early • Site conditions (topography, soil condition, surrounding
as possible before the commitment of funds and approval of area, etc.)
• Regulatory requirements
TABLE 1. Performance Criteria
• Elements of the design (process parts, construction com-
Study Title: VE of Road Median Paving ponents)
(a) Objectives or Required Data • History of the project
• Constraints imposed on the project
1. To cover road median with nice-looking landscape features
• Available utilities
2. Minimum performance as required by ASTM
3. Treatments to be distinct color treatment • Requirements resulting from public participation
4. Stay with reasonable budget, not to exceed SR 100.00/m2 • Design computations
5.
6. The function analysis phase (Tables 3 and 4) is the phase
(b) Goals, Desired Criteria or Features where the work involved and the requirements of the project
1. Materials to be in-kingdom product are clearly defined. Functions are defined by using two words,
2. Patterns are to be rectangular in shape with short dimensions perpen- an action verb and a measurable noun. The two-word descrip-
dicular to traffic direction tion should be general so that it does not imply a solution but
3. Thickness of 50 mm only the required function. As an example, pumping equip-
4. Red color ment has the function of conveying fluid.
5. Size to be 200 ⫻ 50
The function of a project is separated into basic and sec-
6.
7. ondary elements to distinguish between the components of the
project that are actually doing the required job and those that
Note: List required attribute/criteria that any proposed idea must meet
(quantity if possible).
support the main function. Then, the cost and worth related to
each function are identified. Worth is defined as the least cost

TABLE 2. Cost Data

Particular Number Material Labor Install Unit price Total
(1) (2) (3) (4) (5) (6) (7)
Bricks 1,000 500 600 500 1,600 1,600,000
Precast colored pavers 1,000 500 500 500 1,500 1,500,000
Cast-in-place concrete 1,000 200 500 100 800 800,000
Patterned colored cast-in-place concrete 2,000 20 20 10 50 100,000
Asphalt with base 20,000 10 10 5 25 500,000
Artificial grass 17,000 50 30 20 100 1,700,000
Grass including fertilizers 2,000 100 100 50 250 500,000
Ceramic tiles 2,000 20 30 20 70 140,000
Mosaic tiles 20,000 10 10 10 30 600,000
Marble 20,000 100 70 50 220 4,400,000
Mulch stone 2,000 60 40 20 120 240,000

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 207

 TABLE 3. Function Worksheet interrelate. The FAST diagrams aid in the identification of ba-
Why Function How sic function and scope (Parker 1985).
(1) (2) (3) The creative phase (Table 5) is intended to encourage the
VE participants to think with regard to the defined function.
For beautification of median paving Cover road Cover road
that meets ASTM standard median paving median
They may have several alternatives that have worked well for
them in the past and that they traditionally use as solutions to
designs.
TABLE 4. Identify Present Functions In the evaluation phase, as well as in previous phases, the
function and functional approach is the guiding force, which
Verb Noun leads to low-cost alternatives. The evaluation phase, which is
(1) (2)
directed toward the development of preliminary alternatives,
(a) What Does It Now Do? (Item Description) can be considered a continuation of the creative phase. The
Protect Surface major difference, however, lies in the application of careful
Resist Load appraisal and creative judgment to the ideas developed on each
Drain Water specified or required function. It is at this stage that the best
Provide Cover possibilities are selected and researched first.
(b) What Must It Do? (Objectives or Required Criteria) In the development phase (Tables 6–8), selected ideas re-
Improve Appearance sulting from the evaluation phase are developed into workable
Provide Color and sellable solutions. The ideas selected are thoroughly re-
Minimize Dust searched. Preliminary designs, sketches, and LCC estimates
Note: State item description as verb/noun functions. Use only two are prepared. The development phase has the following steps:
words, verb and noun.
1. Design comparison

TABLE 6. LCC Results

Project Annual amount
(1) (2)
Proposed project 106,521.25
Original project 286,743.83
Savings in LCC 180,222.58
Note: Savings = 63% reduction.
FIG. 1. Fast Diagram

TABLE 5. Creative Phase TABLE 7. Development Phase

General Ideas Value Study Summary (Proposed)

1. Cover with bricks Original design brick system for road median paving can be very well
2. Cover with precast colored pavers accomplished without jeopardizing basic or secondary functions by using
3. Cover with cast-in-place concrete colored patterned cast-in-place concrete. The proposed method will
4. Cover with patterned colored cast-in-place concrete • Improve appearance
5. Cover with asphalt with base • Provide color (contrast)
6. Cover with artificial grass • Minimze dust
7. Cover with grass including fertilizers • Protect surface
8. Cover with ceramic tiles • Resist load
9. Cover with mosaic tiles • Drain water
10. Cover with marble • Provide cover
11. Cover with mulch stone • Utilize in-kingdom products
Proposed method will meet all requirements and will cost much less.
Note: Generate many ideas, processes, or methods to fulfill function Changes needed to implement new proposal are very minor engineering
that item under study must perform. Do not evaluate ideas during this work, which can be done internally with available staff (Approximately
phase. 12 man-hours).

required to perform the function (i.e., cost related to the basic TABLE 8. Development Phase
function). For example, for a power plant the basic function
Implementation Plan
is to generate power. Expenses related to this will be worth,
and all other expenses will be just cost. The next step is to 1. How should it be implemented?
The proposed case e will be discussed with the manager of engineering
compare the overall system costs to the sum of the worth of and construction for his acquaintance and overcoming any road blocks
the basic functions. The result is defined as the cost-to-worth prior to management presentation
ratio. Based on experience in comparing cost, a cost-to-worth 2. What should be changed and in what sequence?
ratio >2 will usually indicate possibilities for removing cost The proposed case e will be discussed with the manager of engineering
within the project. Having completed the function analysis and and construction for his acquaintance and overcoming any road blocks
cost model for the project, the VE team is now ready to ven- prior to management presentation
3. Who should do it?
ture into the creative phase. In addition, the functional analysis Engineering department will amend required changes prior to package
technique (FAST) is developed in this phase. submittal to procurement department
The diagram shown in Fig. 1 was first used by an engineer 4. How long should it take?
in 1965 (Miles 1972). The FAST diagram is a tool that helps One month
organize the random listing of functions by answering the 5. Any deadline required?
questions How?, Why?, What does it do?, and What must it As per management report forecast
6. Implementation cost, internal hours?
do? This helps the VE team to develop many verb-noun func- 250 man-hours, total cost SR 5,000
tion solutions, which the FAST diagrams can help sort out and
208 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 2. Sketch of the original and proposed design ings. This idea led to the development of VE incentive clauses
3. Description of the recommendation and what it involves (VEICs).
4. LCC analysis comparison A VEIC provides a method for the contractor to propose
5. Advantages and disadvantages that each recommenda- changes in contract plans and specifications that will lower
tion entails total costs without degrading performance, maintainability, or
6. Implications and requirements to implement the VE rec- reliability.
ommendation The description and application paragraph of a VEIC must
define the basic aims of the change proposal so that they are
The recommendation phase is the final step, and it is per- clear to the parties.
haps one of the hardest steps in the VE job plan. It is the
phase that involves the presentation of ideas, solutions, or VE in Saudi Arabia
changes in terms of benefits to the user. Recommendations are
essentially a challenge to the original design. The recommen- The technology of VE studies was transferred into the King-
dation phase can be summarized by the word ‘‘salesmanship.’’ dom of Saudi Arabia in the early 1980s. The Military Works
at the Saudi Arabian Ministry of Defense and Aviation was
LCC the first organization in the kingdom to apply the VE concept
to its projects. The general director of the General Directorate
The term LCC refers to the cost implications over the life of Military Works and a member of the Larry Miles Founda-
of the facility; i.e., owning, operating, and maintenance must tion Trust introduced the technique into the kingdom (Al-You-
be provided with the greatest value. It is defined by the Amer- sef 1992).
ican Institute of Architects as any technique that allows as- The year 1981 was the year that VE started to be imple-
sessment of a given solution, or choice among solutions, on mented in Saudi Arabia and began to cease to be an imported
the basis of considering all relevant economic consequences tool in the hands of foreign companies to be used in projects.
over a given period of time (or life cycle). It is further defined Now, internal regulations provide a limited service to other
as the total economic cost of owning and operating a facility, sections and offices of departments, and the signing of a con-
manufactured process, or product. tract for six American VE specialists to instruct and create VE
The LCC analysis reflects the present and future cost of the experts among Saudi engineers is proceeding.
project over its useful life. The two methodologies used to Value analysis should be applied to every government proj-
calculate the LCC are the annualized costs and the present ect, and it must be mandatory—VE may be performed in-
worth cost. house and/or by VE experts of VE private companies under
contract (Al-Otaishan 1985).
VE in Construction Projects The most effective tool for improving the quality of life and
the economy of the country, VE is the discipline of the future.
Designs of construction projects are complex. Regardless of
One cannot afford not to participate in it and to make it part
how capable a designer is, there will always be unnecessary
of our daily life.
cost hidden in his design. The very nature of construction de-
sign demands that countless variables be considered and pulled
together by a certain date. The designer is under pressure to RESEARCH METHODOLOGY
meet that deadline and must ensure that a viable, reliable plan The main objective of this study is to develop a computer
is formulated with that set of circumstances and that the design program to perform the VE analysis. This section discussed
grows and takes form. Once the project is formed after con- the methodology used in designing and building the VE com-
siderable effort, the designer becomes wed to the project as it puter program.
is and is perhaps unable to review it for unnecessary cost. The
VE team has an objective viewpoint, not having participated
in the original design. The VE team may lack the knowledge Selecting Standard VE Approach
of the original designer of the project, but they can still im- To date, various techniques have been used with VE studies.
prove it. Some studies use only five phases, whereas others might use
The designer must understand that unnecessary cost in a as many as nine phases. Analyzing each phase of the job plan
design is not a reflection of his abilities as a professional but in turn shows that some 22 individual techniques are employed
is rather a management problem that needs to be addressed. (Mudge 1988). However, the main VE job plan phases are the
Although there are numerous reasons for poor value, the fol- information phase, function phase, generation phase, and eval-
lowing seem to recur most often (Zimmerman and Hart 1982): uation phase.
To avoid any discrepancy, this study selected a standard VE
• Lack of time systematic approach and job plan. A flowchart of the job plan
• Lack of information used throughout the whole study is shown in Fig. 2. The flow-
• Meaning not clear chart shows a standard job plan that consists of five phases,
• Misconceptions including the information phase, functional phase, creative
• Temporary circumstances that inadvertently become phase, evaluation phase, and development phase.
permanent
• Habits Computerized VE System
• Attitudes
• Politics The computerized version of the VE study consists of sev-
• Lack of fee eral interactive panels through which all the information and
data of the VE study can be input. The computer panels layout
Contractors influence value with the quality of workman- of the VE program is shown in Fig. 2. This program is de-
ship. Their skill in construction and the integrity of their prod- signed for the evaluation of and comparison between one or
uct affects the cost of the facility for years into the future. The more projects to select the most appropriate one in terms of
resources and talents of the suppliers and contractors doing cost and other considerations selected by the user.
work especially for the management is an area for cost sav- In this program, the required information is distributed in
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 209
 isting Project’’ to do any revision, modification, editing, etc.
The ‘‘Projects’’ screen is made up of a list that contains a list
of the previously saved projects.
The two following screens will be ‘‘Objectives Or Required
Criteria’’ and ‘‘Goals, Desired Criteria or Features.’’ They are
quite similar in design; each has 12 character fields and one
output area that displays the project title that is found on all
screens. Users are required to enter required and desired cri-
teria for the project. The data entered will be stored in the
corresponding data files.
The ‘‘Cost Data’’ screen (similar to Table 2) is designed to
represent information about the methods suggested for each
project in which the cost analysis of the original design as
well as all alternative designs is performed. The total cost con-
sists of unit price, material cost, labor cost, and installation
cost. Once the user modifies any of the editable data he has
to press the ‘‘Update’’ button, which activates a subprogram
to recalculate the unit price and the total price. Pressing
‘‘Done’’ ensures writing and saving data in their place in the
database files.
The next two screens, namely, ‘‘Present Function’’ and ‘‘Fu-
ture Function,’’ are exactly the same in their design. They ac-
tually are linked to two database files. The first is to add,
display, or edit data in that file, and the other is where some
verbs and nouns are stored.
The ‘‘Generate Ideas’’ screen is designed to provide the
right method or action to be taken. In this screen, two kinds
of displayed fields are shown. These are the ‘‘Study Title’’ and
‘‘Alternatives’’ that were originally entered in the ‘‘Cost Data’’
screen. The user has to add only one term such as ‘‘cover
with’’ or ‘‘build with,’’ and the program will copy it to all
alternatives in their corresponding boxes. The program was
designed to be as flexible as possible by allowing the user to
add more ideas manually on top of the ideas copied from the
‘‘Cost Data’’ screen. The additional ideas will be copied au-
FIG. 2. Job Plan Panels Used in Study
tomatically to the ‘‘Cost Data’’ screen. It is worth mentioning
here that determining the right number of ideas generated ac-
cording to the number of alternatives is done by using the
three groups of database files linked together by a common immediate ‘‘if’’ statement as follows:
field representing the project number.
The first group of database files includes basic information IIF (Empty (Cons-mg 1.particular)
about the project (information phase) such as data regarding
the project team members (study premise), objectives to be Then the ‘‘Feasibility Ranking’’ screen (Fig. 3) is displayed,
accomplished, materials, detailed cost for each material, sum- where each idea that is suggested for every alternative (partic-
mary, and implemented ideas. ular method) is evaluated by giving it a rank out of 10, the
The second group of database files handles the detailed total of these values is calculated (summed up), a ranking ma-
ideas, their advantages and disadvantages, and their rank com- trix is created and sorted (using a sort command), and the rank
pared to each other. The third group of database files is created (1) is given to the maximum value, and so on. This program
for the evaluation of the original and proposed methods using of creating a matrix, sorting, and giving an order or a rank is
weights assigned to each one. performed once the user presses the ‘‘Update’’ command but-
A fourth database file, which is made up of ‘‘memo’’ fields, ton, which is a subprogram included in the program that was
provides on-line help for each screen to guide the user to the developed for the screen.
function and details of each screen. Another file was also cre- The ‘‘Idea Comparison’’ screen (Fig. 4) is designed to write
ated to store information related to function identification, and up to three advantages and three disadvantages for each par-
it includes two character fields. One field is for the verb and ticular case. The study title, the different particulars, and their
the other is for the noun. corresponding rank are displayed as output fields (boxes), and
Upon execution of the VE program, a ‘‘Title’’ screen is dis- only editing is allowed for advantages and disadvantages (in-
played for a few moments. Then the ‘‘Main Menu’’ screen is put fields). One additional key button is added here to allow
displayed. Selection from the ‘‘Main Menu’’ is made through the user to select the number of best choices (ideas) to be
the use of three radio buttons. It has three options, ‘‘New Proj- displayed. Up to seven ideas can be selected. Activating this
ect,’’ ‘‘Revise An Existing Project,’’ or ‘‘Quit.’’ The program key displays a message on the screen, which is created by the
is interactive and all the screens are self explanatory. ‘‘wait window’’ command, and the user enters the number of
It should be pointed out that the screens that follow the main best choices so that the corresponding ideas are displayed.
screen might differ according to the user’s choice. For exam- The ‘‘Weights’’ screen [Fig. 5(a and b)] is designed to per-
ple, if the user selects ‘‘New Project,’’ the ‘‘Screens’’ panel form a comparison among all selected ideas or particulars. The
will be displayed for a few seconds to show the user the var- preferences are to be selected from a pop-up list with three
ious screens he is about to encounter for his VE study. Then fixed entries—major, medium, and minor preference—and
the ‘‘Study Premise’’ screen will be displayed. The ‘‘Projects’’ they are repeated between each of two ideas. This information
screen, however, will appear upon selecting ‘‘Revise An Ex- is stored in database files as an initial step for determining
210 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 3. Feasibility Ranking

FIG. 4. Idea Comparison

their score on the ‘‘Scoring Matrix’’ screen [Fig. 5(a)]. In this 2, and 1, respectively. Each rank is multiplied by its weight,
screen all of the selected ideas are displayed, along with their and the subtotal is displayed in the total rank boxes. These
scores that were calculated from the ‘‘Assigned Weights’’ subtotals are summed up and listed in the ‘‘Total Ranking’’
screen [Fig. 5(b)], by assigning different scores corresponding column.
to each selected preference displayed as raw scores. Then nor- The next screen is the ‘‘Saving and Life-Cycle Cost’’ screen
malization of these scores is done and displayed in the as- (Fig. 7), and it is designed to compare the cost of the original
signed weight column. So, this is only an output screen, and and proposed ideas and to determine the percentage of savings.
the user does not add or modify any entry, because it is a It also performs the LCC analysis of both original and pro-
result of what he has decided and selected in previous screens. posed methods and demonstrates that savings can be achieved.
The final stage of this phase is in the ‘‘Evaluation Matrix’’ Both the original and proposed methods can be selected from
screen (Fig. 6), where all selected particulars, ideas, and their a ‘‘list from array’’ box, where all methods are copied into an
weights are displayed as output fields, and the user decides array, and the user selects the proper one from this array. The
the rank of each criterion against each of the particulars or ‘‘Calculate’’ button is then pressed to display the correspond-
attributes. Each selection is also made out of a pop-up list ing cost of each method by performing a subprogram that se-
including the choices ‘‘Excellent,’’ ‘‘Very Good,’’ ‘‘Good,’’ lects the value of cost corresponding to the method and stores
‘‘Fair,’’ and ‘‘Poor,’’ which are assigned the values of 5, 4, 3, it in its new field as an original or proposed cost by a series
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 211
 FIG. 5. Determining Weights: (a) Scoring Matrix; (b) Assigned Weights

of ‘‘if’’ statements. Also the percentage of the difference be- 4. The expected life time (in years) n is to be input by the
tween costs is displayed, and the percentage of the difference user, as well as the salvage cost (if any).
is calculated and displayed as savings in a percentage. 5. The program then calculates the LCC in two steps:
Upon activating the ‘‘Life Cycle Cost Analysis’’ button, a a. Present work value
subprogram is executed to analyze the LCC for the selected
idea. The LCC analysis is based on the summation of equiv- PW = initial cost ⫹ annual fixed cost(P/A, i%, n)
alent worth of installation, labor, operation, and other costs for ⫹ escalated annual cost(P/A, i1%, n) ⫺ salvage cost
each alternative over its useful life and permits the alternative
having the minimum LCC to be identified. Calculating the b. Annual LCC
LCC in the subprogram is performed as follows (Degarmo et
al. 1984): AW(SR/year) = PW(A/P, i%, n)
The LCC is calculated twice, once for the original method
1. Initial project cost is calculated (installation, material, la- and once for the proposed method. All terms of these equations
bor, etc.). This is obtained from the database file. are expressed as memory variables, and when the LCC per
2. Additional cost is calculated for annual expenditures year is calculated, it is stored in the database file. Two values
(e.g., maintenance cost) with the suitable (determined) are saved, one for the original method and the other for the
interest rate. Both the value and the interest rate are to proposed one.
be input by the user. The program is well designed to give the summary of the
3. Other costs that may have an escalation rate are to be cost analysis with its ‘‘Summary’’ screen and suggestions for
determined. Their interest rate and the escalation rates the implementation of the project with its ‘‘Implementation
are also expected from the user. Plan’’ screen. They are two simple edit screens with scrollable
212 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000
 FIG. 6. Evaluation Matrix

FIG. 7. Savings Screen

edit boxes to enable the user to write the required information. A VE study was applied to the ground-covering system. The
Finally, using the ‘‘Print All’’ screen, all the information en- information phase included objectives or required criteria,
tered and other details can be printed. goals (i.e., desired criteria or features), and cost data. The
The information included in all database files is shown functional phase consisted of answers to two questions, What
through a series of panels representing the different phases, does it now do? and What must it do? Eleven alternatives were
which includes the information phase, functional phase, crea- proposed in the creative phase. In the judicial phase, the al-
tive phase, judicial phase, and development phase. Each phase ternatives have been tested and ranked against six factors.
is divided into several panels. Those factors were aesthetics, cost to develop, probability of
implementation, time to implement, potential cost, and re-
Case Study: VE of Road Median (Al-Jubail) placement cost. Six of the alternatives, which had high rank-
The Royal Commission in Jubail (Royal Commission 1988), ings, were again evaluated by comparing their advantages and
Saudi Arabia, decided to do modification toward beautification disadvantages. Weights were assigned to each goal (i.e., de-
of some major roads. The work included landscaping, irriga- sired criteria), which were then evaluated against each other
tion, and sidewalk and median paving. The designers recom- using the evaluation matrix. The evaluation matrix showed that
mended the following: Paving the median from the nose of the colored patterned cast-in-place concrete was the best al-
the median to a distance of 150 m with bricks and the rest ternative and that it would save about $600,000, equivalent to
with aggregate; and that sidewalks be designed to be covered 38% of the original cost.
with aggregate. The design has specified paving to be common bond paving
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000 / 213
system, red in color, and meeting the following ASTM stan- This paper provides information to develop an easy and
dard: smart system that will help in VE studies. This study recom-
mends a continuation in this direction. The proposed smart
• Concrete building bricks—C55 system must take into account the various factors that have to
• Sampling and testing bricks and structural tiles—C67 be investigated when performing a VE study. The smart sys-
tem must be so flexible that it can accommodate various types
Trees are designed to be 8.0-m apart in the median and of projects in various fields. In other words, the smart system
sidewalks. The total number of bricks to be used in this project must overcome all the limitations that this program suffers.
is estimated to be 20,000/m2. This report will examine the ACKNOWLEDGMENT
ground-covering system and recommend the most suitable sys-
The writers would like to thank King Fahd University of Petroleum
tem using VE techniques. The VE application to the above
and Minerals for providing various facilities for conducting this study.
project is shown in detail in Tables 1–8. Computer screens
showing ‘‘Feasibility Ranking,’’ ‘‘Idea Generation,’’ ‘‘Scoring APPENDIX. REFERENCES
Matrix,’’ ‘‘Assigned Weights,’’ ‘‘Evaluation Matrix,’’ and
Corcoran, K. R. (1981). Saudi Arabia, keys to business success, McGraw-
‘‘Savings’’ are given in Figs. 3–7, which are part of the above Hill, New York, 16–30.
example. All the screens are self-explaining and interactive. Degarmo, E. P., Sullivan, U. G., and Canada, J. R. (1984). Engineering
economy, 7th Ed.
CONCLUSIONS Miles, L. D. (1972). Techniques of value analysis and engineering, 2nd
Ed., McGraw-Hill, New York, 1–24.
The VE computerized model can help greatly in teaching Mudge, A. E. (1988). Value engineering, a systematic approach, 5th Ed.,
30.
the concept and application of VE. It will guide users of the Al-Otaishan, A. T. (1985). ‘‘Establishing a Saudi Arabian value engi-
technique, and it will speed up the process by handling all neering program.’’ Proc., 2nd Saudi Engrs. Conf., KFUPM, Dhahran,
calculations automatically and in interaction with each other. Saudi Arabia, 373–396.
The VE computerized model will encourage people to use the Parker, D. E. (1985). Value engineering theory, Lawrence D. Miles Value
VE technique, because it will simplify the required procedure. Foundation, Washington, D.C., 62–64.
Another, and probably the most important, feature of this pro- Royal Commission at Jubail. Rep. on Value Engrg. Study, Jubail, Saudi
Arabia.
gram is that it will serve as a database of previous VE case Sadi, A., Al-Musallami, A. I., and Al-Sughaiyer, M. (1996). ‘‘Value en-
studies. The intent was not to develop a smart system, because gineering in public construction projects in Saudi Arabia.’’ Build. Res.
it is beyond the scope of this study; however, the system was and Information, 24(3), 152–156.
designed in such a way that it could be modified and enhanced Al-Tamimi, R. A. (1996). ‘‘A computerized system for the value engi-
in the future into a smart system. The VE computer program neering technique in Saudi Arabia.’’ MS thesis, King Fahd University
was designed using a database computer program, FOXPRO of Petroleum and Minerals, Dhahran, Saudi Arabia.
Al-Yousef, A. (1992). Value engineering, concept and technique, 1st Ed.,
Window. The computerized version of the VE study consists 27.
of several interactive panels, representing the systematic ap- Zimmerman, L. W., and Hart, G. C. (1982). Value engineering, a practical
proach (job plan) of a VE study, through which all information approach for owners, designers and contractors, Van Nostrand Rein-
and data of the VE study will be input. hold, New York, 1–31.

214 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JULY 2000

JOURNAL OF
COMPUTING IN CIVIL ENGINEERING
Volume 14 Number 4 October 2000

TECHNICAL PAPERS
Interactive Analysis of Spatial Subsurface Data Using GIS-Based Tool
R. L. Parsons and J. D. Frost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Design Processing of Regular and Nonstandard Structural Components
Mustafa K. Badrah, Iain A. MacLeod, and Bimal Kumar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Parametric Coordinator for Engineering Design
Chee-Kiong Soh and Zhonghui Wang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Estimating Labor Productivity Using Probability Inference Neural Network
Ming Lu, S. M. AbouRizk, and Ulrich H. Hermann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Fuzzy Logic Integrated Genetic Programming for Optimization and Design
Yaowen Yang and Chee Kiong Soh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Multireservoir Systems Optimization Using Genetic Algorithms: Case Study
Mohd Sharif and Robin Wardlaw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Computer-Based Application for the Processing of Clients’ Requirements
J. M. Kamara, C. J. Anumba, and N. F. O. Evbuomwan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264

TECHNICAL NOTE
Applications of Mathematical Software Packages in Structural Engineering
Education and Practice
Primus V. Mtenga and Lisa K. Spainhour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
END-OF-YEAR INDEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
 INTERACTIVE ANALYSIS OF SPATIAL SUBSURFACE DATA USING
GIS-BASED TOOL
By R. L. Parsons1 and J. D. Frost,2 Members, ASCE

ABSTRACT: Much of the spatial information that is an inherent part of geotechnical and geoenvironmental
investigation programs has been underutilized due to the unavailability of analysis tools designed to take ad-
vantage of the spatial attributes of the information. This paper describes a system developed to provide a
quantitative estimate of the adequacy of coordinate-based information used to make particular decisions, with
particular emphasis on geotechnical and geoenvironmental site investigation information. The system was de-
veloped using a geographic information system (GIS) as a platform that was integrated with an existing com-
mercial geostatistical software package to allow for more comprehensive analysis and optimization capabilities.

INTRODUCTION This paper describes the development of the software tool

GIS-ASSESS, which is a GIS-based tool designed to permit
Geographic information systems (GIS) are increasingly assessment of the adequacy, or thoroughness, of an investi-
viewed as a key tool for the storage, display, and analysis of gation. Although developed in a research environment, GIS-
spatial data (Bowman 1998). The rapid development of high- ASSESS was designed with the goal of transferring the tech-
speed, high-storage computer hardware has made it possible nology to the private sector. GIS-ASSESS was developed
to develop GIS-based tools that can perform these functions. using ArcInfo, a GIS system developed by the Environmental
Such tools can help analyze the adequacy of coordinate-based Systems Research Institute, Inc., and was designed to be com-
information used to make particular decisions in a spatially patible with existing geostatistical software. It is capable of
sensitive manner. The tool described in this paper, GIS- analyzing multidimensional data from multiple tools. The re-
ASSESS, is designed specifically for use with the practical sult is a quantitative, objective, spatial measure of investiga-
evaluation of the adequacy of geotechnical and geoenviron- tion adequacy displayed in an easily understood format. The
mental data sets, although the tool may be used with any co- benefits of such a system include (1) optimization of site in-
ordinate-based data set. A quantitative, spatially sensitive mea- vestigation resources through selection of optimal sampling
sure of investigation adequacy is one in which variations in locations and appropriate tools for use at a given site; (2) abil-
subsurface conditions are acknowledged to exist and are in- ity to compare and rank the adequacy of ongoing investiga-
cluded in subsequent analyses. For geotechnical engineering, tions at multiple locations so that available investigation re-
inclusion of the spatial attributes of the data is itself a signif- sources can be prioritized; and (3) assistance in setting
icant advance over common practice, where the spatial com- numerical targets for determining when sufficient site inves-
ponent of these data is often ignored through the averaging of tigation work has been completed. It has also been designed
multiple data values from different locations and the applica- to accept data in the form of ASCII files from a spreadsheet
tion of a safety factor to account for variability among other program, enabling the user to easily transfer data to GIS-
factors. ASSESS without dedicating significant person-hours to the
An additional benefit of such tools is that they provide a task. Writers have successfully used GIS-ASSESS with several
means of addressing the problem that there are no commonly data sets and have performed an extensive evaluation of a large
accepted procedures for determining the amount of data re- data set from Treasure Island (Parsons 1998). (This artificial
quired or for selecting the tools that should be used to ade- island, constructed in San Francisco Bay in the 1930s, was the
quately characterize the area (or volume) to be investigated. subject of a large geotechnical study for the U.S. Navy after
For many fields such determinations are made based largely the Loma Prieta earthquake in 1989.)
on the engineering judgment of the investigator and ‘‘rules of
thumb’’ that are based on past experience. This approach is IMPORTANCE OF GIS
reflected both in the modern geotechnical curriculum and in
published standards for site investigation, for example, ASTM A GIS is an information system designed to work with data
D 420-93 (1998). This results in investigations that vary referenced by spatial or geographic coordinates. It may be
widely in scope and tool selection, and therefore adequacy. thought of as a higher-order map (Star and Estes 1990). Such
While differences clearly exist among investigations, there is systems may be manual, although they are almost exclusively
currently no commonly accepted measure of variations in in- thought of as computer software packages today. An example
vestigation adequacy. This lack of a standard procedure for of a manual GIS would be a base map with a series of overlays
evaluating the adequacy of site investigation data can lead to containing spatial information with various themes. These
varying levels of investigation adequacy that may be signifi- themes might include, but are not limited to, spatial informa-
cant, particularly in sensitive areas such as contaminated sites tion about roads, drainage, utilities, soil type, soil unit thick-
or waste disposal facilities. ness, depth to water, land use, and zoning. A computer-based
system contains the same type of information but is much
1
Asst. Prof., Dept. of Civ. and Envir. Engrg., Univ. of Kansas, more powerful in that it can store much greater amounts of
Lawrence, KS 66045. information, perform engineering analyses on the data set, and
2
Assoc. Prof., School of Civ. and Envir. Engrg., Georgia Inst. of Tech- enable the user to interact with the data set much more effec-
nol., Atlanta, GA 30332. tively. Within ArcInfo the different types of spatial information
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, are stored in a set of files called a coverage. The advantages
2001. To extend the closing date one month, a written request must be of such a computerized system are enormous and include the
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on September 24, following:
1999. This paper is part of the Journal of Computing in Civil Engi-
neering, Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/ • Generation of output (maps) is quicker and less expensive
0004-0215–0222/$8.00 ⫹ $.50 per page. Paper No. 21931. • Output may be previewed on screen prior to production
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 215
FIG. 1. Using Engineering Judgment and GIS to Bridge ‘‘Is-
lands of Automation’’ [Adapted from Parsons (1998)]

• Extensive data analysis may be performed easily

• Sensitivity and optimization analysis may be performed
easily
• System has compatibility with data already in digital form
FIG. 2. Relationship between Investigation Thoroughness
and Cost [Adapted from Tsai and Frost (1996)]
A GIS enables the user to store, retrieve, sort, and display
information according to its spatial characteristics. As such, it
is a suitable environment when used by an engineer to inte- 0 and 100 that quantified the thoroughness of the investigation
grate the data management, analysis, and visualization func- for the site as a whole, or global thoroughness.
tions that are typically performed as independent, albeit re- The system developed by Tsai (1996) combined a CLIPS
lated, actions in practice today. By integrating these (expert system software) coded system with ArcInfo to cal-
traditionally distinct functions in a GIS framework, the engi- culate the thoroughness for discrete targets. It used an inverse
neer using the spatially based system is clearly the ‘‘bridge’’ distance method to account for some spatial correlation within
between the traditional ‘‘islands of automation,’’ as shown in areas with preexisting information and generated a thorough-
Fig. 1, and can thus focus on using the data efficiently rather ness value representing a global or sitewide assessment of the
than spending significant levels of effort in just ‘‘preparing’’ quality of the investigation. The thoroughness was recalculated
the data for use by the various functions. for varying confidence levels in the prior information and for
Selected efforts that have been made to use GIS technology different sampling plans to show how the quality of the in-
for the analysis of site investigation include but are not limited vestigation could be affected by these factors.
to those by Adams (1994), who developed a workbench for For GIS-ASSESS, the concept of thoroughness has been
managing subsurface data and supporting site characterization augmented such that the assessment of investigation adequacy
activities; Luna and Frost (1998), who used the UNIX-based is location specific and can be displayed in a visual format.
ArcInfo as a platform to evaluate liquefaction potential using This improved approach is referred to as continuous thor-
the Liquefaction Potential Index; Tsai (1999), who also used oughness (T) and is based on a probabilistic analysis. The
ArcInfo together with a knowledge-based system to evaluate advantage of this approach is that not only can the ‘‘global’’
site investigation quality; and SimSite, an education package quality of the investigation be obtained, but areas of high and
developed by McBrayer et al. (1998), a Windows-based edu- low investigation quality are immediately apparent as well. For
cational package used for the evaluation of ground-water con- cases where the sample density is not great enough for the
tamination investigations. continuous thoroughness method to be applicable, an alterna-
tive method called discontinuous thoroughness was developed
PROCEDURE FOR QUANTIFICATION OF SITE and integrated into GIS-ASSESS, as described later.
INVESTIGATION ADEQUACY To generate a value of investigation quality that was site
specific, the geostatistical methods of ordinary kriging and in-
The primary method used in GIS-ASSESS for evaluating dicator kriging were used to generate probability values in
investigation adequacy is continuous thoroughness, which was accordance with those used in other geotechnical investiga-
proposed by Parsons (1998) and based on the original concept tions, such as those by Honjo and Kuroda (1991) and Wild
of thoroughness as a measure of site investigation quality pre- and Rouhani (1996). For ordinary kriging, an estimate of the
sented by Tsai (1996) and Tsai and Frost (1996a,b, 1999), value of the parameter of interest may be calculated using a
combined with the use of probability values generated using weighted average of sample values, as shown in (1).
the geostatistical methods of ordinary kriging and indicator
x = w1 x1 ⫹ w2 x2 . . . wn xn (1)
kriging.
Tsai (1996) related investigation quality to cost and showed where xi = sample value; wi = sample weight; and x = esti-
that different sampling plans would result in different levels mated value.
of thoroughness for the same investigative effort (cost), or that The values of the weights are determined by minimizing the
the same thoroughness could be achieved for a lower cost with error of the estimate in the manner described by Journel and
a better sampling strategy (Fig. 2). He described how the Huijbregts (1978). The standard deviation of the error of the
‘‘level of knowledge’’ gained from use of a particular inves- estimate may also be calculated by a similar procedure.
tigative tool changed with tool type, depth, noise, geology, and For ordinary kriging calculations, using the estimate and the
test arrangement (overlap). In developing an estimate of thor- standard deviation of the error and assuming a normal distri-
oughness, he described how thoroughness was dependent on bution of the error, the probability that the field value is below
the effective measurement area of the tool used, spatial cor- some limiting or boundary value (b), where b is a design value
relation, prior knowledge of the site and nature of the contam- or regulatory value, may be calculated according to (2) and
ination, the target object’s characteristics, and noise effects. (3), in the manner described by Dowdy and Weardon (1983).
This level of knowledge was represented by a value between Examples of boundary values include the allowable settle-
216 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 PLATE 1. (a) Example of Contaminated Site with Source at Top Center of Figure (Storage Tanks); (b) Contours of Contamination as Predicted by Kriging; (c) Probabilities That Con-
centration Is Less Than Threshold Value of 30 ppb at Any Potential Sampling Point; (d) Band of Low Thoroughness Is Shown Surrounding Threshold Contour of 30 ppb

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 217

218 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
PLATE 2. (a) Thoroughness Image after Removal of Two Existing Sampling Locations (Lower Corners) and Addition of Four Virtual Sampling Locations to Data Set; (b) Change in
Investigation Thoroughness Caused by Addition of Four Virtual Sampling Locations and Removal of Two Actual Sampling Locations; (c) Estimated Investigation Thoroughness with
Two Existing Sampling Locations Removed and with Four Virtual Samples Added with Assigned Data Values One Standard Deviation Higher than Estimated; (d) Change in Investiga-
tion Thoroughness Caused by Removal of Two Sampling Locations and by Addition of Four Virtual Sampling Locations with Data Values One Standard Deviation Higher than Estimated
ment for a structure or the maximum allowable contaminant
concentrations for a given site. Probability values are appli-
cation dependent because the calculated value for probability
is dependent on the boundary value, which is specified by the
user. For example, as the design capacity for a pile is reduced,
the probability that the actual pile capacity meets or exceeds
the design pile capacity is increased:
b⫺x
z= (2)
␴

冕
z 2
e⫺z /2

P [xtrue < b] = (3)

⫺⬁ ␴兹2␲
where b = boundary or threshold value; x = expected value
from kriging at location of interest; xtrue = actual value at lo-
cation of interest; and ␴ = standard deviation associated with
estimate of x.
The kriging error variance is a function of the spatial ar- FIG. 3. Relationship between Throughness and Probability
rangement of the tests and the model variogram used to rep-
resent the spatial correlation. However, the variability of many located at the center of the top of the figure, and the regulatory
geotechnical parameters is closely related to the mean of the threshold concentration for this example is assumed to be 30
parameter, as summarized by Harr (1987). For this reason, ppb. The effect of dispersion is highly exaggerated in this fig-
GIS-ASSESS was designed to enable use of the relative var- ure for purposes of clarity. Plate 1(b) shows a contour map of
iogram to calculate the relative standard deviation, according contamination concentrations generated using kriging. While
to the method proposed by Isaaks and Srivastava (1989). The Plate 1(b) shows the most likely location for the 30 ppb thresh-
relative standard deviation may be multiplied by the local ex- old, no information is conveyed about the uncertainty associ-
pected value to obtain an equivalent standard deviation of the ated with its predicted location, and it therefore may in fact
error. convey a false impression of certainty with regard to the lo-
For indicator kriging, the sample value is compared with cation of the actual 30 ppb threshold. Plate 1(c) shows that
the boundary value and assigned an indicator value of 1 or 0, the probability that the concentration is below the threshold
depending on whether it is higher or lower than the boundary value increased as the distance from the source increased. Plate
value. The resulting indicator values are then kriged, and the 1(d) shows the thoroughness image for the same data set. This
expected values generated are considered to be probability val- image also shows the expected location of the 30 ppb contour
ues directly. By using a series of boundary values, it is possible line, but in addition it also shows a zone of uncertainty about
to get a series of images that, taken collectively, can show the the expected location, giving the user an easily understood,
sensitivity of the adequacy of the investigation to changes in quantitative estimate of the quality of the estimate of the lo-
the boundary value. cation of the threshold contour line. This approach is related
The concept of continuous thoroughness was developed to to confidence intervals for contour lines, as described by My-
facilitate evaluation of the adequacy of site characterization ers (1997).
activities. Values of thoroughness are calculated by applying Thoroughness functions may take other forms. For example,
a transform function to kriged probability values to yield an a design may call for the use of shallow footings if there is a
index that allows for graphical representation of the zone or 95% probability that there will be less than 1 in. of settlement.
zones where the variable of interest crosses a specified thresh- If this condition cannot be met, a pile foundation should be
old. Thoroughness is intended to complement traditional data used. For this case, the thoroughness may be calculated in
analysis procedures by allowing enhanced visual analysis of accordance with (5), as illustrated in Fig. 4.
the data.
The thoroughness function used in GIS-ASSESS is related If P[x ⱕ b] ⱖ 95%, T = 100% (5a)
to the probability of the outcome as shown in (4).
If P[x ⱕ b] < 95%, T = 0% (5b)
T = 兩100 ⫺ 2P 兩 (4)
Under these conditions, an increase in T due to more or
where P = probability of a given outcome; and T = thorough- better investigation planning is of no practical value unless the
ness. resulting probability is greater than 95%. For P > 95%, thor-
As illustrated in Fig. 3, if the probability of the occurrence oughness is at a maximum because no further benefit can be
of an event is low or high, the thoroughness will be high (the achieved at that location by additional sampling. While no new
probability of a false positive or false negative is low). Con- information is created by this transform, it has the advantage
versely, low thoroughness values reflect intermediate proba- of making the data more visually apparent, and it is more
bilities that an event will occur (the probability of a false pos- compatible with expected cost calculations such as those de-
itive of false negative is high). For example, if the analysis of scribed by Gilbert and McGrath (1998).
investigation data yields the estimate that the probability that The preceding examples have illustrated in a general man-
the parameter of interest is within the specified value (b) is ner how spatial data can be efficiently interacted with through
51%, then the probability that it is not within the specified a customized GIS-based system. Additional details of the sys-
value is 49%. This is only a slight improvement over random tem development are provided below.
chance, and therefore the investigation is calculated to have a
low thoroughness value (T = 2). SYSTEM OVERVIEW
The added information conveyed by this approach is clearly
shown in Plates 1(a–d). Plate 1(a) shows a simulated contam- GIS-ASSESS was originally developed by Parsons (1998)
inated site with concentrations of a particular contaminant in for use with the UNIX version of ArcInfo and the geostatistical
parts per billion (ppb) for 20 sampling points. The source is package Geo-EAS (Englund and Sparks 1991). It has recently
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 219
 LUT is accessed and coordinated with the test locations stored
in the point coverage through the use of a ‘‘relate,’’ which is
created using a menu in GIS-ASSESS specifically coded for
that purpose. Relates are used within GIS-ASSESS as rela-
tional joins between coverages and separate files that contain
information related to the features within the coverages.

Interaction with Geostatistical Package

Once the point coverage, LUT, and relate are created, an
input file for the geostatistical package may be generated. This
ASCII file consists of header information followed by a series
of lines containing the data. Each line contains three values:
the x and y locations and the data value for the test. For those
FIG. 4. Alternate Relationship between Thoroughness and
Probability
types of data where there is a single data value at each loca-
tion, that value is written directly to the file. Where multiple
data values are recorded at a given location, the default pro-
cedure is for the program to average all values that meet the
selection criteria and to write the average value to a Geo-EAS
input file. Functions other than an arithmetic average could
also be used to generate a representative value. If no data
values meet the selection criteria for a particular data location,
that data location is not included in the analysis. Selection
criteria include the maximum and minimum z-values and an
assigned characteristic. For geotechnical applications, the
z-values may be in the form of elevation or depth, and the
assigned characteristic is usually the soil unit. Use of appro-
priate selection criteria to automatically screen the data values
is a powerful attribute of GIS-ASSESS because it enables the
user to select subsets of the data for analysis and prevents
‘‘smearing’’ or skewing of the data caused by the inclusion of
data that do not meet the selection criteria.
A geostatistical analysis may then be performed within Geo-
EAS using the resulting ASCII input file. Use of the geosta-
tistical package Geo-EAS was preferred over the use of the
geostatistical functions built into ArcInfo because of the
greater flexibility in developing the spatial correlation func-
tion. The expected values and absolute or relative standard
deviation values may be generated for a grid of cells covering
the coordinates of interest and stored in an ASCII file for use
by GIS-ASSESS.

Output Generation
The ASCII file generated by Geo-EAS may be converted
into ArcInfo line coverages containing expected value or stan-
dard deviation contours, or into polygon coverages containing
probability or thoroughness polygons. For probability and
FIG. 5. Flowchart Shows Procedure for Evaluating Sensitivity thoroughness coverages, the initial ASCII file is read and the
of Data Set to Changes in Data Values and for Optimizing Addi- probability and thoroughness values are calculated. For each
tional Sampling Resources [after Parsons (1998)] grid cell, after conversion of the relative standard deviation
value to the equivalent absolute standard deviation value (if
been converted for use on Windows NT. The process of anal- required), the normal distribution defined by the expected
ysis is shown in the flowchart in Fig. 5. The system is designed value and standard deviation value is converted to the standard
to operate using menus and virtually eliminates the use of the normal distribution (Dowdy and Weardon 1983). The proba-
ArcInfo command line. GIS-ASSESS is designed to accept bility the actual value is greater than (or less than) the bound-
any type of data that can be represented as a numerical value ary value, b, set by the user is then calculated by solving the
with x, y or x, y, z coordinates. The data may be imported into integral shown in (3) using the regular polygon method (Swo-
GIS-ASSESS from pairs of text files generated in a spread- kowski 1984). If a thoroughness coverage is desired, the val-
sheet or word processor. Specific options have been included ues are calculated at this time. The resulting probability or
to automate the importation of data in selected formats. thoroughness values are then written to an ASCII file com-
The procedure for importing the data into ArcInfo begins patible with ArcInfo and converted to a polygon coverage. The
by generating a point coverage using the first text file. The system is designed such that values are grouped into 10 classes
point coverage is a set of files that contains the information with values 0–10, 10–20 . . . 90–100. Each class is repre-
about the test locations and identification information. The sec- sented by a different color or shading.
ond text file is then imported into a lookup table (LUT). Cre- The polygon coverage that is generated begins as a rectan-
ation of a LUT is automatic for data in one of several standard gular coverage due to restrictions imposed by the geostatistical
formats; LUT tables may also be generated manually for other software. Therefore a module has been included that enables
data sets with relatively little effort. The information in the the user to clip the rectangle using a polygon coverage con-
220 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
taining the investigation boundary, such that the coverage of approach to provide meaningful results. For sites where the
investigation adequacy conforms to the boundaries of the in- target parameter may not exist throughout the investigation
vestigation area. area, Parsons (1998) showed that for a site with targets that
were approximately circular, the sample spacing for a grid pat-
System Output tern must be approximately 45% of the target diameter to pro-
duce continuous thoroughness images of moderate quality, and
The coverages generated may be viewed from within the 23% of the target diameter to produce images of good quality.
ARCPLOT module. Menus are provided that enable the user This sampling density is not economical for most geotechnical
to display and overlay polygon, line, and point coverages. projects.
Control of line color, legend display, and text font is provided. For those sites where the sample spacing is too great for the
Grid files containing the actual probability or thoroughness continuous thoroughness approach to be used, GIS-ASSESS
values may also be viewed in 3D from any angle and elevation
at any magnification.

Analysis and Optimization

Geostatistics by its nature is dependent only on data values
and data locations. A module in GIS-ASSESS was developed
to take advantage of this characteristic by enabling the user to
select virtual sampling locations with the mouse and to assign
data values to those locations. New probability and thorough-
ness images may then be developed based on the combination
of existing and simulated data. A menu is provided that enables
the user to assign the expected value to the location, add or
subtract one standard deviation from the expected value, or FIG. 6(a). Volume of Investigation for Cylindrical Tool
specify a different value using the input field. By experiment-
ing with different sampling locations and patterns and a range
of sample values, the user can determine the impact of addi-
tional sampling points and the sensitivity of the eventual anal-
ysis to a range of potential values at the sampling locations.
An example of the value of this approach is shown in Plate
2(a). For this analysis, samples at four additional locations
were added to the data set shown earlier. Three were added in
areas with low thoroughness values, and one was added in a
location with a thoroughness value greater than 90%. Each of
these samples was assigned the expected data value for that
location from the first analysis, so that the benefit of additional FIG. 6(b). Area of Investigation at Surface and at Selected Tar-
samples could be evaluated in near isolation from changes in get Depth for Spherical Tool
the expected values. Additionally, two samples were removed
from the data set to determine their impact on the analysis.
Plate 2(b) shows the magnitude and areal extent of the im-
provement in investigation quality for the three samples in the
low thoroughness areas. It also shows that the additional sam-
ple in the high thoroughness area had a negligible impact, and
that the impact of the two samples that were removed was
quite different, even though the boundary contour was ap-
proximately symmetrical.
This same module can also be used to evaluate the sensi-
tivity of the analysis procedure to additional data values that
are unexpectedly high or low. Plate 2(c) shows the same data
configuration and the original 30 ppb contour, but for this anal-
ysis the simulated data points were assigned values one stan-
dard deviation above the expected values for their respective
locations. Plate 2(d) shows how such a sampling event would
affect the estimated location of the boundary contour and the
magnitude and areal extent of the effect of additional sampling
on the thoroughness of the investigation.

DISCONTINUOUS THOROUGHNESS
Geotechnical investigations are normally performed to de-
termine the value of geotechnical parameters that exist
throughout the subsurface: shear strength and permeability, for
example. The parameter values may vary, but values for them
exist throughout the subsurface. Geotechnical investigations
may also include searches for targets that have discrete bound-
aries such as karst features, perched water, and steel drums.
When attempting to locate this type of target, the sample den- FIG. 7. Calculation of Effective Detection Area Given Tool In-
sity must be sufficiently high for the continuous thoroughness vestigation Radius and Critical Target Size Radius

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 221

 mobilization of expensive field testing equipment and person-
nel. In addition, it provides a method for evaluating and com-
paring investigations and for setting targets for investigation
quality, which may allow for better optimization of the use of
limited site investigation resources.

ACKNOWLEDGMENTS
The research described in this paper has been supported by NSF Grant
No. CMS 9457549. This support is gratefully acknowledged.

APPENDIX. REFERENCES
Adams, T. M. (1994). ‘‘GIS-based subsurface data management.’’ Micro-
computers in Civ. Engrg., 9(4), 305–313.
ASTM. (1998). ‘‘Standard guide to site characterization for engineering
design and construction purposes.’’ D 420-93, West Conshohocken, Pa.,
FIG. 8. Effective Sampling Area for Grid Pattern (Circles) (Tar-
1–7.
get Centroids Are Shown as Dots)
Bowman, D. (1998). ‘‘Civil engineering data meets GIS.’’ J. Comp. in
Civ. Engrg., ASCE, 12(1), 5–7.
has an alternative procedure, referred to as discontinuous thor- Dowdy, S., and Wearden, S. (1983). Statistics for research, Wiley, New
oughness, which is based on site investigation work by Tang York.
(1987) and the original work on thoroughness performed by Englund, E., and Sparks, A. (1991). ‘‘Geo-EAS 1.2.1 user’s guide,’’ EPA
600/8-91/008, U.S. Environmental Protection Agency, Washington,
Tsai (1996) and Tsai and Frost (1996a,b; 1999). Under this D.C.
approach, the adequacy of the investigation is based on the Gilbert, R. B., and McGrath, T. C. (1998). ‘‘Design of site investigation
percentage of the area of interest that is investigated. programs for geotechnical engineering.’’ Uncertainty in modeling and
For some investigations, the depth of the targets has an im- analysis in civil engineering, CRC Press, Boca Raton, Fla.
pact on the adequacy of the investigation because the area Harr, M. E. (1987). Reliability-based design in civil engineering, Mc-
investigated changes for some tools with depth, while for oth- Graw-Hill, New York.
Honjo, Y., and Kuroda, K. (1991). ‘‘A new look at fluctuating geotech-
ers it is constant. For example, tools such as borings or cone nical data for reliability design.’’ Soils and Found., Tokyo, 31(1), 110–
penetrometers sample a cylindrical volume with a constant di- 120.
ameter and a depth that is slightly greater than the maximum Isaaks, E. H., and Srivastava, R. M. (1989). An introduction to applied
depth of the tool. These tools are referred to within GIS- geostatistics, Oxford University Press, New York.
ASSESS as cylindrical tools [Fig. 6(a)]. Other tools, such as Journel, A. G., and Huijbregts, C. J. (1978). Mining geostatistics, Aca-
a magnetometer, sample a volume that can be approximated demic, San Diego.
Luna, R., and Frost, J. D. (1998). ‘‘Spatial liquefaction analysis system.’’
as a sphere (Tsai 1996). For these tools, the area sampled J. Comp. in Civ. Engrg., ASCE, 12(1), 48–56.
changes with depth, as shown in Fig. 6(b). For situations McBrayer, M. C., Wiley, K. E., and Gilbert, R. B. (1998). ‘‘SimSite:
where the target size is known or where targets smaller than Computer model to simulate site investigation for groundwater reme-
a given size may be ignored, an effective sampling radius may diation.’’ Proc., 1st Int. Conf. on Site Characterization, P. K. Robertson
be determined by combining the tool search radius and the and P. W. Mayne, eds., Balkema, Rotterdam, The Netherlands, 189–
minimum target radius, as shown in Fig. 7. 194.
Myers, J. C. (1997). Geostatistical error management, Van Nostrand
Using the test locations, sampling radius, depth of interest, Reinhold, New York.
and minimum target radius, GIS-ASSESS automatically gen- Parsons, R. L. (1998). ‘‘Assessment and optimization of site characteri-
erates an image showing the area investigated and the thor- zation and monitoring activities using geostatistical methods within a
oughness of the investigation. An example of the output gen- geographic information system environment.’’ PhD thesis, Georgia In-
erated using this method is shown in Fig. 8. For this example, stitute of Technology, Atlanta.
the main radius of 133 targets was averaged and combined Parsons, R. L., Frost, J. D., Chameau, J.-L. A., and Tsai, Y. C. (1998).
‘‘GIS-ASSESS: A spatial analysis tool for site investigation planning
with the search radius of the tool to determine the effective and evaluation.’’ Proc., 1st Int. Conf. on Site Characterization, P. K.
detection area for an individual sample. The sample locations Robertson and P. W. Mayne, eds., Balkema, Rotterdam, The Nether-
(circles) are shown with the centroids of the targets (dots). A lands, 251–256.
target was determined to be ‘‘found’’ if the effective detection Star, J., and Estes, J. (1990). Geographic information systems: An intro-
area for a given sample intersected the centroid of a target. duction, Prentice-Hall, Englewood Cliffs, N.J.
Swokowski, E. W. (1984). Calculus with analytic geometry, Prindle, We-
ber & Schmidt, Boston.
CONCLUSIONS Tang, W. H. (1987). ‘‘Updating anomaly statistics—Single anomaly
case.’’ Structural safety, Elsevier Science, Amsterdam, 151–163.
The spatial analysis tool GIS-ASSESS represents a signifi- Tsai, Y. C., and Frost, J. D. (1996a). ‘‘A spatial thoroughness quantifi-
cant contribution to many areas of civil engineering by pro- cation system for environmental site characterization.’’ Proc., 3rd Int.
viding a tool for the analysis of the adequacy of characteri- Symp. on Envir. Geotechnol., Technomic, Lancaster, Pa., 967–976.
zation data sets. The analysis is based on geostatistical Tsai, Y. C., and Frost, J. D. (1996b). ‘‘Reducing uncertainty in environ-
mental site characterization.’’ Proc., 1996 Conf. on Uncertainty in Ge-
methods, and the results are displayed in an easily understood ologic Environment, ASCE, New York, 58(2), 1019–1033.
visual format using the capabilities of GIS. This more sophis- Tsai, Y. C. (1996). ‘‘A quantitative spatial thoroughness methodology for
ticated analysis tool can enable the user to evaluate the sen- environmental site characterization.’’ PhD thesis, School of Civ. and
sitivity of the investigation to changes in data values or the Envir. Engrg., Georgia Institute of Technology, Atlanta.
effect of additional sampling, which may yield unexpected val- Tsai, Y. C., and Frost, J. D. (1999). ‘‘Using GIS and KBS technology for
ues. By performing such analyses, the user may gain a better real-time planning of site characterization activities.’’ Can. Geotech. J.,
Ottawa, 36(2).
understanding of what conclusions may be drawn from the Wild, M. R., and Rouhani, S. (1996). ‘‘Effective use of field screening
data and a measure of how reliable those conclusions are. This techniques in environmental and geotechnical applications.’’ Geostatis-
approach also enables the user to better determine the number, tics for environmental and geotechnical applications, ASTM STP 1283,
type, and location of additional tests in the office prior to the ASTM, West Conshohocken, Pa., 88–101.

222 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 DESIGN PROCESSING OF REGULAR AND NONSTANDARD
STRUCTURAL COMPONENTS
By Mustafa K. Badrah,1 Iain A. MacLeod,2 and Bimal Kumar3

ABSTRACT: A design standard can be complemented with a case base of nonstandard designs to form a richer
knowledge source that can be used in the development of a system for processing of both regular and nonstandard
structural designs. This leads to the question of classification of a design context to regular or nonstandard,
which relates to the interpretation of the standards scope. This paper describes a software system, called SADA-
II, which has been developed to address the issues of standards scope characterization and nonstandard com-
ponent design processing. SADA-II also introduces a method for processing of standards provisions that are
written in a generic format (e.g., pertaining to different section shapes) and explores the potential of integrating
the system with a graphics package and other CAD applications. Illustrative examples for the application of
SADA-II to detail structural design are provided.

INTRODUCTION computer code, so it would fully model a generic provision.

In the same time, the complexity of geometrical data entry and
Structural design standards prescribe procedures and re-
processing should not hinder the usability of this method.
quirements that govern the design of buildings, bridges, and
Different approaches for the representation and process of
other structures. While design standards can be used in a
design standards and building codes have been developed.
straightforward manner for handling regular components (i.e.,
These include approaches based on decision tables (Fenves
those within the scope of the standard), this is not the case for
and Garrett 1986), logic programming (Rasdorf and Lakma-
nonstandard components (those which can not be handled di-
zaheri 1990), production rules (Dym et al. 1988), object mod-
rectly or completely by the standard). However, a design stan-
eling technology (Garrett and Hakim 1992; Yabuki and Law
dard can be complemented with a case base of nonstandard
1992), and distributed architectures (Kiliccote et al. 1996).
designs (for which solutions have been established and ap-
However most computer-aided component design applications
proved). This will form a richer knowledge source that can be
(developed commercially and in research projects) have fo-
used in the development of a system for design processing of
cused on regular components. Such applications do not address
both regular and nonstandard structural components.
the issue of nonstandard design processing. Also, they do not
An issue related to the previous discussion is the classifi-
provide a means to help the user determine whether a design
cation of a design context to regular or nonstandard. In order
context is within the scope of a standard before the design
to correctly use a design standard for a design context, the
process takes place; such a task is generally left to the user’s
user should first determine that this context is within the scope
judgment. Neilson (1997) developed a facility for nonstandard
of the standard. A design standard as a whole, plus its parts,
component design processing incorporated in the Standards
sections, and provisions, should ideally include clear state-
Automated Design Processing (SADA) system. Design pro-
ments of their scope. However, design standards are often am-
cessing of nonstandard components was further investigated,
biguous specifically regarding their scope. So, a designer may
and other issues related to the modeling of the design stan-
not be able to infer the scope of a standard without consulting
dards and design rules related to nonstandard components
other sources of information. Therefore, in the development of
were addressed in SADA-II (Badrah et al. 1998).
a standards processing system, it is desirable to incorporate a
In this paper, a more complete view of SADA-II is pre-
facility for characterizing the scope of a standard and its pro-
sented and the issues of standards scope characterization, ge-
visions. This facility will be based on several sources of in-
neric provision processing, expressing nonprocedural infor-
formation, e.g., experienced engineers and code writers in ad-
mation of design cases, and integration into the CAD
dition to the standard itself.
environment are addressed.
Additionally, in a design standard, some provisions are writ-
ten in a generic format to be applicable to a variety of situa-
tions (e.g., pertaining to different section shapes); these are SYSTEM DESCRIPTION
called ‘‘generic provisions.’’ Such provisions involve substan- Architecture
tial use of geometrical data, especially if they are applied to
irregular sections. A method for generic provision processing Fig. 1 shows the overall architecture of SADA-II, which
is desirable to be incorporated in a standards processing sys- includes the Standards Model, the Case Model, and the Prod-
tem. This method should be capable of handling various sit- uct Model. The Standards Model (SM) adopts a formal rep-
uations including new ones without any modification to the resentation of design standards based on object-oriented soft-
ware technology and intended for the processing of regular
1
Lect., Facu. of Civ. Engrg., Univ. of Aleppo, Aleppo, Syria; formerly, components. Standards requirements are organized in the Stan-
Grad. Student, Univ. of Strathclyde, 107 Rottenrow, Glasgow, G4 0NG, dards Requirement Object Hierarchy. Objects in this hierarchy
U.K.
2
Prof., Dept. of Civ. Engrg., Univ. of Strathclyde, 107 Rottenrow, Glas-
include data and methods to represent requirements implied
gow, G4 0NG, U.K. by standards provisions. The Standards Model can help both
3
Sr. Lect., Dept. of Civ. Engrg., Univ. of Strathclyde, 107 Rottenrow, to identify appropriate provisions, by firing the scope rules,
Glasgow, G4 0NG, U.K. and to view the standards requirements. The SM objects are
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, associated with relevant scope rules and can access external
2001. To extend the closing date one month, a written request must be functions for generic provision processing. The Case Model
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on August 20, 1999.
(CM) is a formal representation of design rules for nonstan-
This paper is part of the Journal of Computing in Civil Engineering, dard components characterized as design cases. A design case,
Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/0004-0223– here, is a set of approved and detailed rules for the design
0232/$8.00 ⫹ $.50 per page. Paper No. 21710. processing of conceptual entities representing nonstandard
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 223
 1. The Standards Applicability Module, whose purpose is
to help the user find out whether a context is within the
scope of the standard and identify the applicable provi-
sions.
2. The Design Rules Establishing Module, which helps the
user establish the relevant design issues and rules for a
new nonstandard problem context.
3. The Component Design Processing Module, which per-
forms design synthesis and conformance checking for
both regular and nonstandard components.

In the implementation of these modules, special attention was

given to the issue of human-computer interaction to make sure
that the implemented approach can potentially support the de-
velopment of a user-friendly application. Each module is pro-
vided with a set of sequentially numbered clickable buttons
describing the steps that the user should follow to run a mod-
ule. Additionally, the processing results are arranged in a re-
port format, and dialogue boxes are extensively used.

DEVELOPMENT STRATEGIES
Standards Model (SM)
A design standard can be viewed as a collection of provi-
sions given for specific design issues regarding different types
of design entities. A design issue, here, is a topic in the design
standards involving specific requirements that need to be sat-
isfied during the design synthesis or conformance checking
processes and that may be relevant to a number of design
entities. A design standard is organized in the Standard Re-
quirement Object Hierarchy as a number of classes at different
levels (Fig. 2). Member consideration is an abstract class for
the different member types in a standard, e.g., beam, column,
slab, and base. Each member will have provisions regarding
FIG. 1. SADA-II Architecture
different Design Issue Groups (DIGs), e.g., the ultimate limit

components (Badrah 1998). The CM is based on object-ori-

ented software technology and case-based reasoning (Kolod-
ner 1993). It is intended to provide processing of existing non-
standard designs by modeling their requirements and support
for establishing design rules for future nonstandard design
contexts by specifying metadata describing the existing cases
and their possible reuse. Objects in the CM that include the
requirements of nonstandard designs can access the Standards
Model requirements and the external generic functions; thus,
the Standards Model and the Case Model are integrated.
The Product Model is a symbolic representation of the de-
sign data related to structural components in an attribute-value
format. The PM adopts an object-oriented hierarchical struc-
ture and includes objects representing design data for regular
components, which can be accessed by the Standards Model;
and objects for nonstandard components, which can be ac-
cessed by the Case Model. The PM is also used for integration
of the component design system with other CAD applications.
SADA-II is implemented using KAPPA-PC, an object-ori-
ented programming tool and expert system shell (KAPPA
1992). The External Generic Functions are implemented using
Autolisp, the programming language of AutoCAD. SADA-II
is currently integrated with a drafting package, AutoCAD
(which is used for the preparation of production drawings),
and a spreadsheet package, MS Excel (for storing structural
analysis results).

Functionality

SADA-II contains the following modules (Fig. 1): FIG. 2. Standards Model

224 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

state DIG and the serviceability limit state DIG. DIGs are rep- However, EC2 (EC2 1992) is written in a different style
resented in classes under each member. Each Design Issue (DI) regarding the organization of its contents. EC2 Part 1 includes
class will have multiple levels of classes, which will contain sections related to design issues (torsion, buckling, etc.), rather
the requirements (implied by the provisions) for different sit- than member types. Provisions in each design issue are written
uations. generally for different member types. Nevertheless, provisions
Input and output design data items are stored in PM objects. regarding specific structural members are stated where appro-
A Design Entity (DE) is instantiated under a PM class. Its priate. But using the same organizational system discussed in
requirements can be identified by mapping it to its relevant Section 4.3.1 will be justified for the following reasons:
member class in the SM using ‘‘object-communication’’ (Fig.
2). This modeling approach for a design standard is similar to • Mapping between the Product Model objects and stan-
the context-based approach by Kiliccote et al. (1994) in terms dards requirement objects to identify relevant require-
of avoiding complexity in managing the lengthy and intercon- ments will be easier if the standard is organized in classes
nected object hierarchies found in the object-oriented model where member types are at a higher level than design
of Garrett and Hakim (1992), which is based on multiple in- issues.
heritance. The context-based approach is an extension to the • The same model can be basically used for representing
object-oriented model (Garrett and Hakim 1992), object-logic multiple design standards including traditional standards
model (Yabuki and Law 1992), and the description-logic such as BS 8110.
model (Hakim and Garrett 1993).
In the current implementation, the SM covers mainly: (1) The only difficulty in adopting the organizational system dis-
Section 3.4 of BS 8110 (Structural 1985) regarding beam de- cussed in Section 4.3.1 for EC2 is that mapping between the
sign (bending strength, shear, torsion, lateral buckling, deflec- Standards Requirements Object Hierarchy and the standard it-
tion, and flexural cracking); and (2) Section 4.3 of EC2 (EC2 self will not be as obvious as in traditional standards.
1992) regarding the ultimate limit state of bending (both uni-
axial and biaxial) combined with an axial force in column
Standards Scope Characterization
design. These are selected as examples to demonstrate the pro-
posed strategies; however, extending the implementation to in- Regarding the characterization of the standards scope, a
clude other provisions regarding other structural members is a rule-based approach is adopted. The knowledge base of the
matter of adding new branches in the SM and PM. What has scope characterization facility consists of a scope-rule base and
even further reduced the number of objects and the complexity preliminary design data repository (Badrah 2000).
of their management, in the SM of SADA-II, is the use of the Scope Rule-Base. Knowledge regarding the scope of a
object-communication approach (Badrah et al. 1998). standard is codified as a set of scope rules that can be extracted
from the standard itself and from other sources, e.g., code
Object-Communication Approach experts. These are organized in three categories and are at-
This is based on the ‘‘delegation’’ technique of object-ori- tached to appropriate SM objects (Fig. 2). The scope rule cat-
ented software technology. Delegation is defined as an imple- egories are:
mentation mechanism in which an object forwards or delegates
a request to a second object; the second object will carry out • Metascope rule sets determine if the design standard, in
the request on behalf of the first one (Wegner 1987; Gamma general, is appropriate for a context.
et al. 1995). There is no need for an inheritance relationship • Member classification rule sets determine if a context is
between the two communicated objects. Thus, delegation is classified as one of the structural members defined and
employed for reusing behavior (either partially or fully) in an covered in the standard.
object by another object representing a different context or • Design issue rule sets are implemented for each design
located in a different hierarchy. In addition to mapping be- issue, related to a member type, to determine the appli-
tween the PM objects and SM objects, object communication cability of provisions given for this issue to the design
is used for modeling standards cross-references and for the context.
integration of the Standards Model and the Case Model (Bad-
rah et al. 1998). For example, an object representing cracking An example of a scope-rule written in pseudocode, which is
requirements of a column may reference a subhierarchy of manipulated using the standards BS 8110 (1985) and other
objects representing cracking requirements of a beam and an sources (Allen 1988) and Elliott (1996), is:
object representing the shear of a channel beam in the Case
Model can reference a subhierarchy of objects representing the ‘‘If (the ‘Design Entity Depth Uniformity’ =
shear of a rectangular beam in the SM. ‘not uniform’ Or the ‘Design Entity Section
Shape’ is Not rectangular, T, L or flanged)
Organization of BS 8110 and EC2 Codes Then (Shear Applicability is OSS).’’
The modeling of a standard in classes in the previously
mentioned manner (i.e., in a hierarchy including, from top to Preliminary Design Data Repository. This holds data re-
bottom, member types, design issue groups, and design issues) lated to a design context and required for firing the relevant
will be in agreement with the way most design standards are scope rules (Fig. 2). It is implemented as object attributes
written. For example, provisions of BS 8110 (Structural 1985) where each attribute is associated with a multivalue list from
are arranged in sections where each section concerns a specific which the user can select one value according to the design
structural member, e.g., beam, column, slab, base, etc. Every context.
section covers different design issues related to the structural The outcome of SAM is the applicability of the standard
member that it handles, e.g., the beam section includes pro- regarding the general scope, member classification, and rele-
visions regarding the design issues of bending moment, shear, vant design issues and the applicable provisions. This infor-
and torsion. Other standards written in this style include AISC mation will determine the classification of the design context
LRFD, the American specification for steel construction (Load to regular and nonstandard. The standards applicability is char-
1986) and BS 5950, the British standards for steel design (BS acterized as ‘‘goals.’’ A goal is an object attribute associated
5950 1990). with a scope rule set and may take one of the following values:
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 225
 • WSS = within the scope of the standard • A graphical AutoCAD representation. This may include a
• PAG = provided as guidelines cross section, a longitudinal section, and/or a 3D model.
• OSS = outwith the explicitly defined scope of the standard
A subcase contains requirements related to a design issue
External Generic Functions within the parent case implemented as methods [in object-
oriented programming (OOP) terms].
The Standards Model of SADA-II can access External Ge-
neric Functions (Fig. 2), which are written in Autolisp/ Example Case
AutoCAD and cater for an arbitrary cross section of a struc-
tural member. Thus, they provide a method for generic As an example, ‘‘Case310’’ represents an inverted-T RC
provision representation. Generic provisions require a different beam. (See Fig. 5, in relation to Examples 1 and 2, described
form of representation and processing from other provisions later in the paper.) This case includes two groups of subcases
that can be applied only to a specific class of design entities regarding the design issues of the beam as a whole and re-
having, for example, a unified structure—e.g., shear provi- garding the design issues of its boot:
sions for rectangular RC beams and torsion provisions for box
RC beams. An example of a generic provision is section 1. Whole beam subcases are:
4.3.1.2 of EC2, Part 1: • Scase30111 for bending
• Scase30112 for shear
4.3.1.2 Design resistance to bending and longitudinal force • Scase31113 for torsion
‘‘P (1) In analysing a cross-section to determine its ultimate • Scase30114 for lateral buckling
resistance, the assumptions given below shall be used: • Scase30122 for flexural cracking
(i) Plane section remains plane. 2. Boot subcases are:
(ii) The strain. . . . ’’ • Scase31211 for boot shear
• Scase31212 for strut and tie forces
This provision can be applied to a general cross-section shape; • Scase10221 for boot flexural cracking
however, it does not give ‘‘ready-to-apply’’ requirements or
design formulas. Instead, users should use their knowledge and In general, design rules for a nonstandard component can
apply appropriate first principles along with the assumptions be based on first principles, test data, and empirical rules. They
given in such provisions to establish relevant procedures. A could also be based on provisions of the standard, since some
major task in developing an algorithm based on this provision design issues of a nonstandard component may be fully within
is to write a routine that can calculate the centroid location the scope of these provisions or can be considered as such by
and area of the concrete compression zone (an arbitrary section applying a simplifying assumption. Thus, some standards re-
shape). But this geometrical information plus reinforcement quirements (in the form of OOP methods) need to be reused.
sizes and locations can be obtained from the AutoCAD data- This requires the integration of the Case Model and the Stan-
base associated with the relevant section drawing (Badrah dards Model so that an object in CM can access other objects
2000). So the Generic Function, which is an algorithm that in the SM for reusing requirements included in them (see the
applies first principles based on assumptions given in the re- paragraph entitled ‘‘Object-Communication Approach’’).
lated provision, depends on automatic geometrical data ex-
traction. A Generic Function is called from the SM when a Description of Cases and Indexing Issues
generic provision of a relevant design context needs to be
processed. A case needs to have descriptive metadata so that its appli-
cability to a context can be determined. This has been achieved
Case Model (CM) by classifying cases in hierarchy according to their properties,
e.g., member type and section composition (Badrah 1998). At
The objectives of the Case Model are: the subcase level, every design issue has a set of indices that
defines the properties a design context should have for pos-
• To provide a structure for the requirements, implementing sibly reusing the design rules implemented in the subcase. To
the design rules related to existing nonstandard compo- establish design rules for a new design context, the latter needs
nents. to be matched with one or more cases. This is realized by
• To use the case base of nonstandard designs in the solu- comparing the specifications of the design context with the
tion of new design contexts by defining metadata describ- properties of the relevant cases/subcases. A component match-
ing case contents and the possibility of future use. ing, using metadata at the case level, should be carried out
• To be integrated into the Standards Model. first which will result in exact or partial matching. With exact
matching, design issues and their requirements can be reused.
Structure and Contents of Case Base With partial matching, users need to browse the partially
matched cases and use their own knowledge to decide what
Design rules of a component represented as a case involve design issues they need to consider and their appropriate rules.
requirements with which the design entity should comply to A design issue matching, using relevant indices, can also be
ensure safe structural behavior. For these to be represented, a achieved for better exploration of the case base.
case-subcase structure is adopted which is realized in object-
oriented modeling, where a case is a ‘‘composite’’ object hav-
Updating Case Base
ing the subcases as ‘‘aggregate parts.’’
A case encloses the following information, in addition to The case base of SADA-II has currently a limited number
the relevant subcases: of cases on beam design in the domain of reinforced concrete
structures. A full-scale system would include nonstandard
• A reference to the relevant section of the design meth- cases for different structural member types, different materials,
odology document, which provides detailed information etc. New design cases can be configured and added to the
about the principles upon which the requirements (i.e., system by those who are responsible for updating the system.
limiting conditions) are established. Instructional information in the form of a user guide or an
226 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
authoring tool are required for the purpose of updating the case Since the intention in the current implementation of SADA-
base in a real-life system. II regarding data exchange is to demonstrate the system inte-
gration capabilities, the number of integrated applications are
Integration into CAD Environment limited to the component design application (the main appli-
To develop the integration capability in SADA-II, a design cation), which also contains the Product Model, the drafting
Product Model was proposed to be incorporated to the overall package (AutoCAD), and structural analysis results in form of
model. The Product Model is intended to play the role of a an Excel spreadsheet [Fig. 3(b)]. Automatic data exchange is
central database that stores data related to different design ap- currently supported for conformance checking of regular com-
plications to facilitate information sharing [Fig. 3(a)]. Each ponents. Also, the PM in SADA-II does not conform to a
application extracts the required data from the Product Model design data exchange standard, e.g., STEP (Industrial 1995),
using the Dynamic Data Exchange (DDE) technique, which as in the CIMsteel project (Garas and Hunter 1998), but it can
enables data to be exchanged between two applications based be developed in this direction.
on the MS Windows platform. At the same time, design in-
formation that is generated in one application and can be used Data Exchange with Graphics Package
by other applications is sent to the Product Model. Adopting It is assumed that production drawings are available in a
such an approach will be appropriate for integrating a number digital format. A production drawing will be used after car-
of applications of the same discipline. For example, a struc- rying out small additions and modifications. An elevation
tural engineering integrated system may include a structural drawing needs to have the following additional data items:
analysis application, a component design application, a graph- section location lines, support center lines, and shear reinforce-
ics application, and a detailing application (i.e., reinforcement ment (i.e., link) region boundaries. These lines are recognized
detailing, or connection design in steel design packages). by assigning them to different layers. A section drawing may
include tension, compression, and other reinforcement bars in
addition to links. These are also to be assigned to different
layers. An entity already assigned to a specific layer can be
easily assigned to another layer that is required for the sake
of the data exchange subprocess. Fig. 4 shows a modified
drawing of a two-span RC beam example.
Once a section/elevation drawing is selected on the screen
(which is done by specifying top left and bottom right points
outside the drawing entities so as to surround it in a box), all
geometrical data can be accessed. Autolisp functions supported
by DDE functionality are written to manipulate this data and
to obtain component and section attributes, e.g., component
and section dimensions and rebar and link sizes and locations.
FIG. 3. SADA-II Integration Approach: (a) Future Development These attributes are exported to the relevant Product Model
Proposal; (b) Current Implementation
objects in SADA-II. Now the conformance checking process
can take place by running the Component Design Processing
Module.

ILLUSTRATIVE EXAMPLES
Example 1: Standards Applicability Checking of
Inverted-T Beam
The main purpose of this example is to illustrate standards
applicability checking using SADA-II. In this example, stan-
dards applicability is to be checked for an inverted-T precast
concrete beam. The selected standard is BS 8110 (Structural
1985). Beam B8 (Fig. 5) is a part of a three-story parking
garage made of precast concrete. To find out if the design
context is within the scope of the standard, the user should
select appropriate attribute values for data related to the design
FIG. 4. Two-Span Beam Example: (a) Elevation; (b) Cross Sec- context and needed to fire relevant scope rules. These are en-
tions tered through dialogue boxes, such as that shown in Fig. 6,

FIG. 5. B8 Drawing (Inverted-T Beam): (a) Longitudinal Section; (b) Section A-A

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 227

 FIG. 6. Attribute-Value Selection

FIG. 7. Applicability Results of B8

which is provided with a commentary to help the user select Matching Process (Design Rules Establishing Module)
the appropriate values. The applicability results would be is-
sued in a report such as that shown in Fig. 7. The applicability The matching process of the design context B8 is carried
of B8 to BS 8110 regarding general scope and member clas- out using the Design Rules Establishing Module. This involves
sification and definition was WSS. The design issues selected selecting attribute values for the design context via dialogue
for this design context are bending moment resistance, shear, boxes, as in the standards applicability checking. The matching
torsion, flexural cracking, and deflection. The applicability of process resulted in exact matching with Case310 (Fig. 8), so
BS 8110 provisions related to bending moment resistance and B8 can be processed using the Design Processing Module.
deflection was PAG. The applicability results of shear and tor-
sion were OSS. The applicability result regarding flexural Design Processing (Component Design Processing Module)
cracking was WSS. Applicable provisions are specified and the
Sizing or conformance checking of a component using the
general classification of the design context B8 was nonstan-
Component Design Processing Module involves two main ac-
dard, since no provisions were found in BS 8110 for some
tivities: data input and processing. The processing results of
design issues.
the Strut and Tie Forces (STF) issue related to the nib part of
B8 are shown in Fig. 9. This shows STF requirements defined
Example 2: Design Synthesis of Nonstandard in the subcase (Scase31212) and their firing results. For ex-
Component (Inverted; T Beam) ample, ‘‘C.6.3.2<<req1-tie-force: ten-boot -links = 292.17’’ re-
fers to the requirement for resisting horizontal tie force.
The same design context described in the previous example C.6.3.2 is the relevant section in the Design Methodology
is used to illustrate the matching and design processing of a Document (see the section entitled ‘‘Case Model’’). And ‘‘ten-
nonstandard component. boot -links = 292.17’’ states the area of the boot links in mm2.
228 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 FIG. 8. Matching Results of B8

FIG. 9. Processing Results of B8 Regarding STF Issue

Table 1 shows the processing results related to the lateral buck- lected column has a circular cross section, the same procedure
ling and deflection of B8, bending moment for section 2, shear explained in this section can be used for any section shape.
for section 1 (Fig. 5), and shear and STF for the nib part. Such The Component Design Processing Module needs to be run
results can be used to synthesize the beam reinforcement, as for the design synthesis of C3. C3 is first instantiated under a
is shown in Fig. 10. relevant PM object by entering preliminary design data such
as dimensions. Since designing the column against axial load
Example 3: Design Synthesis of Circular Column and uniaxial bending moment for this context requires exe-
Using Generic Functions cution of a generic function, the user is referred to use the
external function ‘‘UBend⫹AxialLoad’’ in AutoCAD. The lat-
The main purpose of this example is to illustrate the use of ter is developed for column design against uniaxial bending
a Generic Function for design synthesis. In this example, col- moment and axial load, based on the generic provision 4.3.1.2
umn C3, which is a part of a reinforced concrete frame and of EC2 (EC2 1992), and can be executed from the AutoCAD
subject to an axial load and uniaxial bending moment, needs extended standard toolbar (Fig. 11). The user needs to draw
to be designed according to EC2 (Fig. 11). Although the se- the column section with an initial selection of reinforcement
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 229
 TABLE 1. Processing Results for Beam B8
Section Design issue Requirements Requirement processing results
(1) (2) (3) (4)
Section 2 Bending moment Req. 1: BM reinforcement Tension reinforcement area = 2,454 mm2
Req. 2: Lever arm limit Satisfied
Req. 3: Neutral axis limit Satisfied
Section 1 Shear Req. 1: Maximum shear stress Satisfied
Req. 2: Concrete shear stress 0.86 N/mm2
Req. 3: Shear links area-upon-spacing 1.0 mm2/mm
Whole component Lateral buckling Req. 0: Slenderness limit Satisfied
Deflection Req. 0: Deflection Satisfied
Boot Shear Req. 1: Maximum shear stress Satisfied
Req. 2: Shear strength without reinforcement Satisfied
Strut and tie forces Req. 1: Tension Boot links Link reinforcement area = 292.17 mm2
Req. 2: Concrete strength under strut force Satisfied
Req. 3: Load transfer reinforcement Load transfer reinforcement area = 524.75 mm2

FIG. 10. Reinforcement Details of Section 2 in B8

FIG. 11. Design Synthesis of C3

230 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 FIG. 12. Processing Results of C3

partial matching. The case base provides another knowledge

source in addition to the Standards Model; thus, the described
system extends the use of a design standard beyond regular
components.
The scope of the standards is explicitly represented by
adopting a rule-based approach, which makes it easier to cap-
ture the standards scope. It is also more flexible to modify and
update the standards scope simply by modifying the relevant
rules or by adding new rules. Moreover, this explicitness more
readily supports developing an explanation facility that will
provide the user with more understanding of the reasoning
process behind the provision identification and context clas-
sification process. This is one potential future development
regarding this part of the system. By adopting this approach,
consistent and systematic interpretation of the standards scope
FIG. 13. Output Data of C3 can be reached by all users of the system including people
who are not familiar with a specific standard.
bars. An iterative process for sizing the reinforcement is car- The generic provision processing functions utilize the ca-
ried out as follows: pabilities of graphics packages to aid structural engineers in
sizing and conformance checking processes interactively (in
the sizing process, the designer can configure the solution on
1. Initially, ‘‘5 T 20’’ rebars arranged symmetrically around
the screen and apply generic functions to operate on these
the section with appropriate bar-covering allowance were
solutions). These functions are described as ‘‘generic’’ because
selected. However, after processing, it was found that the
they show how generic provisions in a design standard can be
moment resistance of the section is less than the ultimate
implemented to cater for arbitrary member cross sections.
design bending moment.
Extraction of all geometrical data from the drafting package
2. Next, ‘‘6 T 20’’ rebars arranged as shown in Fig. 11 were
is a complex and expensive task. In the development of the
selected. Fig. 12 shows a processing dialogue box from
data extraction algorithm in SADA-II, the principle followed
which the user can adjust interactively the ‘‘neutral axis
is to automate the process where it is effective to do so. Thus,
depth upon section depth’’ ratio until a balanced force
some data items, e.g., those regarding links (or shear reinforce-
situation is obtained. The moment resistance of the sec-
ment) area and spacing, are entered by user query through
tion is now sufficient to withstand the applied actions.
dialogue boxes. On the other hand, only 2D drawings are used
Output results can now be sent back to the design ap-
in the implementation of these algorithms for the purpose of
plication to be stored in the instantiated PM object (Fig.
structural design. However, for integration with other Com-
13).
puter-Aided Construction applications, such as a CIC for con-
crete placement, a 3D solid model could be necessary, as in
CONCLUSIONS Kunigahalli and Russell (1995) and Bernold and Salim (1992).
This paper describes a software system for the processing The contributions of this paper can be summarized as fol-
of regular and nonstandard structural designs that involve ge- lows:
ometrical and nongeometrical data. The system combines a
Standards Model and a Case Model and incorporates a facility 1. It demonstrates a Case Model (as a part of the overall
for standards scope interpretation and for generic provision model) that can be used for processing existing nonstan-
processing. It is also integrated into a graphics package dard designs and that provides support for establishing
(AutoCAD) and can be integrated into other CAD applica- design rules for future nonstandard designs.
tions. 2. It demonstrates and implements a facility for standards
The case base of nonstandard designs can be viewed as a scope characterization based on rule-based reasoning.
database incorporating methodologies, graphics, and proce- 3. It demonstrates the practicality of utilizing graphics
dures and supported with methods to perform both exact and packages capabilities for generic provision processing in
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 231
 respect to the ease of use and the maintenance of a wide Garrett, J. H., and Hakim, M. M. (1992). ‘‘Object-oriented model of en-
scope of applicability. gineering design standards.’’ J. Comp. in Civ. Engrg., ASCE, 6(3),
323–347.
4. It explores the potential for integrating the component Hakim, M. M., and Garrett Jr., J. H. (1993). ‘‘A description logic ap-
design system with other CAD applications. proach for representing engineering design standards.’’ Engrg. with
Comp., London, 9, 108–124.
APPENDIX. REFERENCES Industrial automation systems and integration: Product data represen-
Badrah, M. K. (1998). ‘‘Design cases: a supplementary knowledge source tation and exchange. (1995). ISO Central Secretariat, Geneva.
for structural component design.’’ Proc., Int. Comp. Congress held in KAPPA PC reference manual. (1992). Inillicrop, Inc., Mountain View,
Conjunction with 1998 ASCE Annual Convention & Exhibition, ASCE, Calif.
Reston, Va., 506–517. Kiliccote, H., Garrett Jr., J. H., Chmielenski, T., and Reed, K. A. (1994).
Badrah, M. K. (2000). ‘‘Information technology strategies for detailed ‘‘The context-oriented model: an improved modeling approach for rep-
structural design.’’ PhD thesis, Dept. of Civ. Engrg., University of resenting and processing design standards.’’ Proc., 1st Congress Held
Strathclyde, Glasgow, U.K. in Conjunction with A/E/C Sys.‘94, ASCE, New York, 145–152.
Badrah, M. K., MacLeod, I. A., and Kumar, B. (1998). ‘‘Using object- Kiliccote, H., Garrett, J. H., and Fenves, S. J. (1996). ‘‘A standard pro-
communication for design standards modeling.’’ J. Comp. in Civ. cessing framework.’’ Information processing in civil and structural en-
Engrg., ASCE, 12(3), 153–161. gineering design, B. Kumar, ed., Civil-Comp Press, Edinburgh, U. K.,
Bernold, L. E., and Salim, M. (1993). ‘‘Placement-oriented design and 133–140.
delivery of concrete reinforcement.’’ J. Constr. Engrg. and Mgmt., Kolodner, J. (1993). Case-based reasoning, Morgan Kaufman Publishers,
ASCE, 119(2), 323–335. San Mateo, Calif.
BS 5950, Part 1: Code of practice for design in simple and continuous Kunigahalli, R., and Russell, J. S. (1995). ‘‘Framework for development
construction, hot rolled sections. (1990). British Standards Institution, of CAD/CAC systems.’’ Automation in Constr., 3, 327–340.
London. Load and resistance factor design specifications for structural steel build-
Dym, C. L., Henchey, R. P., Delis, E. A., and Gonick, S. (1988). ‘‘Rep- ings. (1986). American Institute of Steel Construction, Chicago.
resentation and control issues in automated architectural code check- Neilson, A. (1997). ‘‘A hybrid approach to the representation and pro-
ing.’’ Comp. Aided Des., 20(3). cessing of design standards.’’ PhD thesis, Dept. of Civ. Engrg., Uni-
EC2, design of concrete structures, Part 1: General rules and rules for versity of Strathclyde, Glasgow, U.K.
buildings (together with United Kingdom national application docu- Rasdorf, W. J., and Lakmazaheri, S. (1990). ‘‘Logic-based approach for
ment). (1992). British Standards Institution, London. modeling organization of design standards.’’ J. Comp. in Civ. Engrg.,
Fenves, S. J., and Garrett Jr., J. H. (1986). ‘‘Knowledge based standards ASCE, 4(2), 102–123.
processing.’’ Artificial Intelligence, 1(1), 3–14. Structural use of concrete: BS 8110, Part 1–3. (1985). British Standards
Gamma, E., Helm, R., Johnson, R., and Vlissides, J. (1995). Design pat- Institution, London.
terns: elements of reusable object-oriented software, Addison-Wesley, Wegner, P. (1987). ‘‘Dimensions of object-based language design.’’ Proc.,
Reading, Mass. OOPSLA ‘87, N. Meyrowitz, ed., ACM Press, New York, 22(12).
Garas, F. K., and Hunter, I. (1998).‘‘CIMsteel (computer integrated man- Yabuki, N., and Law, K. H. (1992). ‘‘An integrated framework for design
ufacturing in constructional steelworks)—delivering the promise.’’ The standard processing.’’ Tech. Rep. No. 67, Stanford University, Stanford,
Struct. Engr., London, 76(3), 43–45. Calif.

232 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 PARAMETRIC COORDINATOR FOR ENGINEERING DESIGN
By Chee-Kiong Soh1 and Zhonghui Wang2

ABSTRACT: Coordination of design information is essential for the design team to provide accurate technical
documents for the actual construction. Design changes are inevitable during any ordinary building project. If
the design changes are not properly managed, design conflicts will be created, which will result in more expen-
sive design or even failure in the design-construction process. However, little research has been done to address
the issue and thereby provide tools for the management of design changes. This paper presents a novel approach
that can facilitate the coordination of design information through managing design changes with the help of a
parametric coordinator. Design changes of dimensions between different building components are specially
discussed. The proposed parametric coordinator provides each building component with the linking knowledge
that is described as a dimensional property and a reference property. A group design method is applied to check
the design consistency. Illustrative examples are presented to demonstrate the feasibility of the proposed para-
metric approach for design change management.

INTRODUCTION sented to coordinate design consistency between different ge-

ometric models. A coordinator has been developed to manage
Design can be viewed as an activity in which teams of de- design changes with the help of the knowledge-based para-
signers work together toward a final solution. The activity of metric technique and the proposed linking knowledge for ge-
designing through the interaction of the designers and the en- ometric models. This work aims at resolving the problems of
vironment is referred to as collaborative design (Saad and managing design changes in drawings, which are still used as
Maher 1996). Design issues are iterative in nature for civil the main repository of design information and principal com-
engineering problems. In an actual design environment, design munication media (Martini and Powell 1992). It also attempts
changes inevitably and continuously affect the properties of to provide a feasible tool to automate the current coordination
many building components at various stages in a building proj- process, especially the design consistency between different
ect. Such changes may occur because of the need to satisfy geometric models for collaborative design problems.
the new or modified requirements made by the owners, to The rest of the paper is organized as follows. First, an over-
change construction facility, to reduce project cost, or to rec- view of the related work is presented; then the methodology
tify existing design mistakes. If these changes are not properly and implementation of the proposed approach are introduced.
managed, design conflicts will be created and these conflicts This includes descriptions of the constraint management sys-
will result in more expensive design or even failure in the tem, the knowledge-based constraint solver, the proposed con-
design-construction process. cepts and models for building linking knowledge, and the
Coordination of design information during the design-con- group design method. Illustrative examples are then provided.
struction activity is essential for the design team to provide Finally, concluding remarks are made.
quality construction technical documents that are free of in-
compatibility errors (Pena-Mora et al. 1995; Mokhtar et al. RELATED WORK
1998). Traditionally, design data is checked manually for its
In design, management of design changes involves the col-
consistency. A common technique is to use checklists to verify
laboration of textual and geometric information. Much re-
the compatibility of various components in a project. The main
search in collaboration has attempted to provide general sys-
problem with this technique is that the design modifications
tems that are able to take into consideration various design
that are needed to be done when incompatibility errors are
and management requirements as well as coordination strate-
discovered may in turn generate other incompatibilities. Be-
gies and goals of different disciplines. For example, Hirasawa
sides, the technique cannot continuously monitor the effect of
et al. (1994) used design and construction drawing as a mul-
design changes, and manual work can be expensive and time
tilayered blackboard to manage information. Froese and Paul-
consuming.
son (1994) introduced an object model-based information sys-
Approaches on investigating feasible techniques to achieve
tem (OPIS) by establishing standard models for project
collaboration across the various disciplines involved in the de-
management and construction information. Fischer and Forese
sign-construction process have been carried out in recent
(1996) discussed the essential, general characteristics of shared
years. But few research works are available on coordination
project models. Rezgui et al. (1998) described the central mod-
of design consistencies and management of design changes
els of COMMIT (construction modeling and methodologies
(Mokhtar et al. 1998). In the writers’ view, the tools for co-
for intelligent integration of information), which can define
ordinating design consistencies between different geometric
mechanisms to handle issues relating to the management of
models are essential as geometric data contribute to the basic
information to support decision making in collaborative proj-
parts of a building project. In this paper, a constraint meth-
ects. Stouffs et al. (1998) presented an information, commu-
odology based on parametric technique (Wang 1999) is pre-
nication, and collaboration system (ICCS) to support the pro-
1
Prof., School of Civ. and Struct. Engrg., Nanyang Technological
cess of project development and administration. Approaches
Univ., Singapore 639798. in standard representation and data exchange have been con-
2
Res. Scholar, School of Civ. and Struct. Engrg., Nanyang Technolog- ducted by the International Alliance of Interoperability (IAI
ical Univ., Singapore. 1996) and International Organization for Standardization (ISO
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, 1989; Froese 1996).
2001. To extend the closing date one month, a written request must be A constraint management approach was proposed by El-
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on May 17, 1999. This
Bibany and Paulson (1994) as a method to collaborate design
paper is part of the Journal of Computing in Civil Engineering, Vol. knowledge according to the parametric representations of de-
14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/0004-0233– sign and management requirements. Lottaz et al. (1999) de-
0240/$8.00 ⫹ $.50 per page. Paper No. 21027. scribed a constraint-solving method to express large families
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 233
of acceptable solutions to facilitate and abbreviate the nego- parametric coordinator is actually a constraint-based manage-
tiation process during design and construction. The issue of ment system with the following characteristics:
acquiring the rationale in a collaborative environment was dis-
cussed by Pena-Mora et al. (1995) and de la Garza et al.
1. Built-in linking knowledge for the component models so
(1997). Pena-Mora et al. proposed the design recommenda-
that the knowledge can be easily maintained by the de-
tion-intent model (DRIM) as an ontology to represent the de-
signers
sign rational for conflict mitigation, while de la Garza et al.
2. Group patterns to check design consistency between the
used a parameter dependency network to represent design ra-
different component models
tionale for the sequent stages in order to perform their tasks
3. Knowledge-based constraint solver to render modifica-
more efficiently.
tion after a design change has been found
However, until recently the issue of managing design
changes between related components in a design project has
not received much attention (Mokhtar et al. 1998). As an at- Architecture of Constraint-Based Management
tempt to address this problem, Mokhtar et al. presented an System
information model whose core idea was to build a single re-
pository of active building components that were assigned the The constraint-based management system, CKB-Design/
task of propagating design changes. They used textual data Draft (Wang 1999), incorporates a knowledge-based constraint
rather than drawings as the main media for storing and com- solver and a group design method for managing design
municating design information. However, they did not address changes between different geometric models. The proposed
the issue of automation of the design change process. system is for the conceptual design phase, where conceptual
Effort in the geometric modeling field to manage design models of components in a construction project are needed to
constraints within a geometric part for variational design is allow project information to be represented in the computers
reflected in the so-called parametric technique (Anderl and and to build data standards to support electronic communica-
Mendgen 1996; Fowler 1996). Parametric technique allows the tion between project participants. The design changes dis-
designers to make modifications to an existing part design by cussed are that of the dimension constraints in building
changing the parameter values, then drives the design config- component models. The architecture of the proposed con-
uration to generate a new design. Typical constraint-solving straint-based management system is shown in Fig. 1.
approaches in the parametric design include the numerical The constraint-based management system is composed of
constraint solver (Light and Gossard 1982; Kondo 1991) and two parts, the building component database and the function
the knowledge-based constraint solver (Aldefeld 1988; Bouma tools, and is implemented in a client-server environment. The
et al. 1995; Brout et al. 1996; Lee and Kim 1996; Wang 1999). building component database is set on the server side and
The numerical solver translates the constraints into a system maintains all geometric component drawings that are neces-
of equations where the coordinates of characteristic points are sary for a project. Such data are used to produce technical
the variables, and then solves them using the Newton iteration construction documents suitable for erecting the building.
method. In the knowledge-based constraint solver, constraints The function tools reside on the client side to help the de-
are represented in predicate forms, and symbolic inference signers manipulate the design data through the component da-
plays a prominent role in processing the geometric constraints. tabase. They include a set of sketch tools, a constraint solver,
Currently the parametric technique is commonly used in fea- and a group of management tools. The sketch tools provide
ture-based solid modeling, where geometric parts or structures the drafting facility for the creation of component models.
are regarded as the predefined objects called features and can These concept models are assigned with constraint schemas
be combined with sweeping techniques to design 3D solids. that include explicit constraints, such as the dimensions, and
In the writers’ view, the parametric technique is quite capable implicit constraints, for example, the topological relationship
of automatically managing design changes within a geometric —i.e., parallelism and perpendicularity. The constraint solver
model. However, it can hardly manage the influence of a is applied when the constraint schema of a model needs to be
global constraint from a geometric model to its related models, modified, such that a new instance can be generated to satisfy
as the geometric model lacks the linking knowledge on how the modified constraint schema. Each new instance can con-
it is related to the other geometric models through constraints. tribute a new constraint model to the existing component da-
This paper focuses on the coordination of geometric design tabase. The management tools are used to build the model
consistency between building components and attempts to use relationships and the linking knowledge between geometric
the proposed linking knowledge as well as the group pattern components. These management tools consist of the assisting
module to extend the ability of parametric design to manage functions, to generate a related model group pattern, and the
design changes between related geometric models. editors, to specify the linking constraint knowledge between
the building components as well as the basic network facility.
CONCEPTS AND METHODOLOGY In addition, a facility for exchanging data with the commercial
Automatic knowledge integration and the creation and
maintenance of the resulting data-dependency structures re-
quire a unified methodology for knowledge representation as
well as inference (El-Bibany and Paulson 1994; Lakmazaheri
1997; Mackwork 1997). As constraints describe the knowledge
on relations for entities and models, constraint-management
methodology may create and maintain data-dependency struc-
tures that reflect the computational process.
Since graphical representations are the primary parts of a
design, this research focuses on the management of design
changes between geometric models. The proposed coordinator
is based on the knowledge-based constraint solver and at-
tempts to manage the constraint models using the built-in link-
ing knowledge to ensure consistency of design changes. The FIG. 1. Architecture of Constraint-Based Management System

234 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 FIG. 2. (a) Sketch; (b) Its Conceptual Model; (c) New Model after Dimension Modification

software AutoCAD (Autodesk 1995) via DXF file is also example, a component drawing with dimensions is regarded
available in the system. as a geometric model. However, a geometric model contains
only the geometric informations, such as coordinates and di-
Knowledge-Based Constraint Solver mensions, within itself. Such a model does not have any
The proposed knowledge-based constraint solver employs knowledge on how it is related to the other geometric models.
the rule-based graphic reasoning method (Lee and Kim 1996) Thus, additional attributes or properties with the linking
to deduce the geometric configuration, which is associated by knowledge should be attached to facilitate relation control of
the constraints. The built-in knowledge base is made up of the models.
meta-level rules, reasoning rules, and entity-connection pat- In order to clearly describe how the current model’s dimen-
terns. With the constraint solver, constraints can be treated as sions are related to the referred models, the current model
facts to be processed, and a pattern-matching mechanism is should have information on the referred models’ IDs, dimen-
used to search the constraints and fire corresponding rules to sion names, dimension values used, and the way they are ap-
infer the geometric entities. A construction plan is used to re- plied. The source model should provide the unique information
cord the inference procedure. The parametric model is then data, which can be easily accessed by the referring model.
determined by sequentially evaluating the rules in the con- Here the concepts of dimensional property and reference prop-
struction plan together with a set of dimension values. erty are proposed for this purpose. When a geometric model
Take the simple sketch shown in Fig. 2(a) as an example. is attached with a dimensional property and a reference prop-
This sketch can be interpreted by the constraint solver as a erty, such a model is regarded as a constraint model. The terms
conceptual model, as shown in Fig. 2(b), where Pi(i = 1, 6) ‘‘dimension property’’ and ‘‘reference property’’ are defined
represents the point entities, Li (i = 1, 4) the line entities, A1 below.
the arc entity, and di (i = 1, 5) the dimension constraints. The Definition 1: A dimensional property of a constraint model
explicit constraints such as dimensions, and the topological is a document object that records the dimensions’ names of
constraints such as tangency, perpendicular, etc., are recog- the model, the dimensions’ current values and formula rela-
nized by the constraint solver. Examples of representation of tions, and the dimensions’ indicators for the used reference
the constraints as predicate symbols are listed as follows: dimensions. It provides an accessible constraint data that may
be inferred to by the other constraint models.
• On P L: point P is located on line L Definition 2: A reference dimension in a constraint model is
• P㛭fixed P: the location of point P is fixed actually a dimension in another model but is referred to by the
• Tangent L C: line L is tangent to circle C current constraint model.
• Vcoordinate P1P2d: the difference of Y-coordinate be- The reference dimension is used to establish the relationship
tween Points P1 and P2 is d. between different models so as to exert the influence of a
• Vdirection L: the direction of line L is vertical global constraint. With the descriptions of the corresponding
reference dimensions in the referring models, the propagation
An illustrative rule to infer the constraint state of point P2 of design changes along all the related constraint models can
based on point P1, line L1 and dimension d1, is described be- be guaranteed.
low: Definition 3: The descriptions of a set of reference dimen-
sions used by a constraint model are recorded in its reference
P㛭fixed P1
property, which depicts the referred constraint models’ names,
Vcoordinate P1P2d1
dimensions’ names and values, and their representing symbols
On P1L1
in the current constraint model.
On P2L1
Since the dimensional property of a constraint model pro-
Vdirection L1
vides the accessible information data that can be referred to
→ P㛭fixed P2
by the other constraint models, while the reference property
When the dimensions are changed, for example, d1 and d2 are builds the link from the current model to the referred con-
changed to ‘‘72,’’ d3 to ‘‘20,’’ d4 and d5 to ‘‘60,’’ the constraint straint models, the data link can be completely set up for a
solver will start to process the geometric objects using the group of related constraint models. The dimensional property
constraints stored in the conceptual model. The successfully and the reference property are created by a dimension table
inferred rules constitute a construction plan. By evaluating the generator and reference table generator, respectively, which
construction plan with the set of changed dimensions as the will be described in the section on Illustrative Examples.
input data, a new model is generated, as shown in Fig. 2(c).
Group Design Method
Models and Properties for Coordination
In this paper, a model refers to a geometric component or Managing design changes focuses on two aspects: (1) the
part that is made up of geometric entities and constraints; for design change determination, and (2) the design change prop-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 235
agation. Such considerations are carried out in a model group
pattern.

Model Group Pattern

It is necessary for the geometric models that have related
external constraint to be grouped together. In order to clearly
describe the relationships of the constraint models within a
model group, a three-level hierarchical strategy is used to set
the position of each model in the group, according to the con-
trol parameter’s input and output. Each level in the proposed
model group pattern is defined as follows:
Definition 4: A top-level model in a model group is the
model that sends out dimension constraints to the two lower-
level constraint models, while it itself does not receive any
external dimension constraint. FIG. 3. Architecture of Group Model Manager
Definition 5: A middle-level model in a model group is the
model that sends out dimension constraints to the lower-level
constraint model, while it itself also receives dimension con- els and on the comparison between each bottom-level model
straints from the top-level models. with its referred top-level models and/or middle-level models.
Definition 6: A bottom-level model in a model group is the The general check procedure for dimension consistency of the
model that receives dimension constraints from the two higher- proposed group pattern is given below:
level models. Step 1. Obtain each list of top-level models, middle-level
Each model level is assigned a specific role in the group to models, and bottom-level models, which are specified
decide whether a corresponding model can yield controlling in the group pattern.
parameters (or constraints), and whether the model has the Step 2. Obtain an unchecked model in the middle-level model
right to receive any control parameter. Thus, the selection of list. Treat it as the current model. Search its reference
component models to be grouped together is based on certain property.
design control or functional requirements. For example, a set a. Get a referred dimension sj in the current model.
of frames can be grouped together because they need to be Get the referred model name, which should be in
assembled together and therefore share some interrelation- the list of top-level models, the original dimension
ships. Certain frames, beams, columns and/or plan-view draw- name in the referred model, and the value used
ings need to be grouped together because there are certain (which is regarded as the target value).
global constraint links between them. A group pattern editor b. Search the referred model’s dimensional property,
has been developed for this purpose and will be described in and retrieve the latest value that is treated as the
the Illustrative Examples section. The specifications of the re- source value of the referred dimension. Compare
lated models in a group pattern can be stored as a managing the target value with the source value. If the target
document in the database. The designers can negotiate to cre- value is the same as the source value, this dimen-
ate the group pattern document or to update the model infor- sion consistency is maintained. Otherwise, this di-
mation stored in the existing managing document. mension is inconsistent, and output an error mes-
sage.
c. Repeat substeps a and b until all the referred di-
Managing Design Changes
mensions have been checked.
A notification mechanism is essential to alert the designers Step 3. Repeat Step 2 until all the middle-level models have
of any model change and also to support the automation of been checked.
other information management process, such as approval. This Step 4. Get the unchecked model in the bottom-level model
includes the detection and propagation of design changes. list. Treat it as the current model. Search its reference
Propagation involves changing the dimensions of a set of tar- property:
get models because of a change introduced in a higher-level a. Get a referred dimension si in the current model.
model. In other words, propagation is the notification of ob- Get the referred model name, which should be in
jects that will then automatically update themselves. the lists of top-level models and middle-level mod-
Notification of design changes can be carried out in two els, the original dimension name in the referred
directions. In the horizontal direction—i.e., within a discipline model, and the value used (which is regarded as
—several geometric models or drawings may share some com- the target value).
mon dimensions, e.g., different level’s plan-views for a build- b. Repeat the substeps b and c in Step 2.
ing or different components needed to be assembled together. Step 5. Repeat Step 4 until all the bottom-level models have
In the vertical direction, i.e., along different disciplines, design been checked.
parameters should be transferred from the upstream models, Step 6. A check result is issued.
such as a preliminary design model, to the downstream mod-
els, such as a construction model, in a practical design project. According to the result of the check procedure, the designers
The benefit of using a group pattern is that the designers can easily locate where the design inconsistencies occur, then
can easily check the design consistency between the different activate the knowledge-based constraint solver to propagate
models, and the propagation and maintenance of design the design changes, and solve the conflicts.
changes can be automatically implemented by the knowledge-
based constraint solver. The architecture of the group model ILLUSTRATIVE EXAMPLES
manager is shown in Fig. 3. Since the top-level models them-
selves do not receive any control parameters, the check for Two examples are presented to illustrate the feasibility of
consistency is concentrated on the parameter comparison be- the proposed parametric coordinator for design changes and
tween each middle-level model with its referred top-level mod- collaboration.
236 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 FIG. 4. (a) Frame Model (Frame1); (b) Plan View Model (Plan1) Referred to by Frame1

1. Example of Relation Buildup between Two Models

An AEC design starts from the architectural design to con-

struction. Therefore a downstream component, for example a
frame, may get a reference dimension from a plan view draw-
ing. The model (Frame1) shown in Fig. 4(a) is used as an
example to build up the constraint links. Usually, the dimen-
sion d1 of Frame1 is not independent, and its parameter may
come from the plan view model’s width dimension ‘‘W.’’ Tra-
ditionally, such relation is specified in the checklist, which
does not build any active links between models—hence, any
modification of the down stream model has to be separately
modified.
The proposed method to manage such relation is reflected
in the design of a dimension interface called parameter table.
The parameter table is also responsible for generating the di-
mensional property of the current model. Fig. 5 shows the
dimension parameter table for Frame1 to which a set of initial
dimension schema has been set up. In the parameter table,
FIG. 5. Dimension Parameter Table for Frame1
items in the ‘‘Name’’ column specify the unique symbol for
each dimension. Items in the ‘‘Old Value’’ column record the
initial dimension values, and items in the ‘‘New Value’’ col-
umn hold the same contents before they are modified. How-
ever, the contents in columns ‘‘Name’’ and ‘‘New Value’’ can
be modified through the ‘‘Edit data’’ box. If a reference di-
mension is required for the current model, the designers can
use the ‘‘Create Referdim’’ tool to create a reference dimen-
sion property for the current model, and the number of refer-
ence dimensions can be specified in the ‘‘ReferDimNum’’ ed-
itor. For the current constraint model Frame1, one reference
dimension is used, and the reference property interface is
shown in Fig. 6.
The reference dimension table interface includes a reference
detail table and four controls. The reference detail table is
designed to reflect the requirements of the reference property.
Items in the ‘‘Rep㛭Name’’ column describe the representative
symbols in the current model. These symbols express the
pseudo dimension names that can be used by the current
model, and distinguish them from those generated by the pa-
rameter table. Items in the column ‘‘Ref㛭Model’’ specify the
models whose dimensions can be referred, and the items in
FIG. 6. Reference Dimension Table for Frame1
‘‘Ref㛭Dim’’ specify the dimension names in the referred mod-
els. The function of the ‘‘GetValues’’ tool is to fetch the values
of the referred dimension from the database. In this example,
a value of ‘‘75’’ has been retrieved for ‘‘W’’ and is automat- TABLE 1. Modified Dimension Schema for Frame1
ically set in the cell under the column ‘‘Values.’’ Dimension Name Old㛭Value New㛭Value Use Ref㛭Dim
Back to the dimension table, the fifth column, ‘‘Use Ref㛭 (1) (2) (3) (4) (5)
Dim,’’ is a boolean indicator whose values describe whether a 1 d1 68 s1 Yes
reference dimension is used by any of the current dimensions. 2 d2 56 60 No
When all the required dimensions have been modified, the 3 d3 11 d1 * 0.15 No
designer can start to process the constraints when he/she clicks 4 d4 9 10 No
the ‘‘Start Design’’ button. In this example, a modified dimen- 5 d5 20 20 No
sion schema to reflect the global control of the dimension in 6 d6 13.5 d1 * 0.2 No
7 d7 13.5 d6 No
Plan1 is shown in Table 1. 8 d8 11 d3 No
A new instance of Frame1 is generated after the dimension
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 237
schema has been updated, as shown in Fig. 7. Thus, it is ap- cording to the fabrication of frame works, it is required that
parent that by depending on the knowledge specified in the the dimensions of Frame1: di (i = 1,4) d2 , d2 , d4 should be
dimensional property and the reference property of the com- consistent with those of Frame2 so that the two frames can be
ponents to automatically propagate design changes, the pos- assembled together. The same dimension requirement applies
sibility of a designer forgetting to transfer any design change to Frame3 and Frame4. Additionally, d5 and d6 specified in
to the related components can be avoided. In this way, design Frame4 should be consistent with the d5 and d6 in Frame3.
consistency can be maintained. The management of design consistency for these four
frames is via a group model pattern interface, as shown in Fig.
2. Example of Group Pattern Design 9. Corresponding to the hierarchical model concept described
The construction of a building may involve thousands of in the previous section, three panel lists are designed for the
parts or components, leading to hundreds of thousands of ge- top-level model, middle-level model, and bottom-level model.
ometric constraints. It is necessary to use a hierarchical orga- In this example, Frame2 is specified in the ‘‘top model’’ list.
nization, which allows constraints to be maintained and ma- Frame 3 is specified in the ‘‘middle model’’ list because it
nipulated in comprehensible groups. In this example, four receives dimension controls from Frame2, but it also outputs
different shapes of structural frame components, as shown in a controlling dimension for Frame4. Frame1 and Frame4 are
Fig. 8, are presented as a simple example to illustrate the group specified in the ‘‘bottom model’’ because they both only re-
design concept proposed for the management of design con- ceive the controlling dimensions. Thus, at the model level, a
sistency. These four frames are supposed to be assembled to- group link has been set up for the four frames.
gether, and they share some common dimension constraints. However, the detailed constraint relations are specified in
For design consistency, it is required that Frame2 outputs the dimensional property and the reference property of each
four control dimensions, namely, d1 , d2 , d3 , d4 (d5 = d3). Ac- constraint model, which have been described in the first ex-
ample on relation buildup. These components need to be
equipped with the necessary ‘‘linking’’ constraint knowledge.
For simplification, only the used part of the dimensional prop-
erties and the reference properties are listed here. For Frame2,
part of the used dimensional property is shown in Table 2.
Tables 3–5 each describe part of the dimensional property and
the reference property of Frame1, Frame3, and Frame4. With
the specifications of each model’s dimensional property and
reference property, a set of linking relations are built between
the model’s dimension constraints. Therefore the design con-
sistency of the four frames can be guaranteed.
For illustration, suppose the value of span dimension d1 of
Frame2 is changed from ‘‘70’’ to ‘‘100.’’ Then, dimension con-
flicts would occur among the four frames if the affected frames
were not modified accordingly, because they will still be re-
ferring to the old values of d1 of Frame2. To resolve the con-
flict, the design manager can first open the frame group and
use the ‘‘Check group’’ function in the group model interface
to see how the inconsistency occurs. The result of a check of
the frame group is shown in Fig. 10.
FIG. 7. New Instance of Frame1 after Dimension Schema Is When the design manager locates the design inconsistency
Updated between the four frames and accepted the suggestion made

FIG. 8. Four Different Shapes of Frame Components

238 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 TABLE 5. Part of Dimensional Property and Reference Prop-
erty of Frame4
Dimensional Property Reference Property
Name New㛭Value Use Ref㛭Dim Rep㛭Name Ref㛭Model Ref㛭Dim
(1) (2) (3) (4) (5) (6)
d1 s1 Yes s1 Frame2 d1
d2 s2 Yes s2 Frame2 d2
d3 s3 Yes s3 Frame2 d3
d4 s4 Yes s4 Frame2 d4
d5 d5 Yes s5 Frame3 d5
d6 d5 No
d7 d1 * 0.2 No
d8 d3 No

FIG. 9. Group Model Pattern Interface

TABLE 2. Part of Dimensional Property of Frame2

Name New㛭Value Use Ref㛭Dim

(1) (2) (3)
d1 70 Yes
d2 56 No
d3 12 No
d4 10 No
d5 d3 No
FIG. 10. Check of Frame Group Pattern

TABLE 3. Part of Dimensional Property and Reference Prop-

erty of Frame1
Dimensional Property Reference Property
Name New㛭Value Use Ref㛭Dim Rep㛭Name Ref㛭Model Ref㛭Dim
(1) (2) (3) (4) (5) (6)
d1 s1 Yes s1 Frame2 d1
d2 s2 Yes s2 Frame2 d2
d3 s3 Yes s3 Frame2 d3
d4 s4 Yes s4 Frame2 d4
d5 d4 ⫹ 10 No
d6 d1 * 0.2 No
d7 d6 No
d8 d3 No

TABLE 4. Part of Dimensional Property and Reference Prop-

erty of Frame3
Dimensional Property Reference Property
Name New㛭Value Use Ref㛭Dim Rep㛭Name Ref㛭Model Ref㛭Dim FIG. 11. New Model after Design Change Has Been Propa-
(1) (2) (3) (4) (5) (6) gated (Frame4)
d1 s1 Yes s1 Frame2 d1
d2 s2 Yes s2 Frame2 d2 CONCLUSION
d3 s3 Yes s3 Frame2 d3
d4 s4 Yes s4 Frame2 d4 As the coordination of design information is a major prob-
d5 d1 * 0.15 No lem in the construction industry, management of design
d6 d5 No changes plays an important role in design collaboration. This
d7 d3 No paper describes a managing system based on the parametric
technique and the group managing method to address the is-
sues of design changes, especially for geometric information
by the check result, he/she can activate the knowledge-based and entities. The system has been developed as a proof-of-
constraint solver to display the dimensional property and then concept prototype to demonstrate and experiment with the pro-
use the reference tools to retrieve the latest value of d1 in posed method.
Frame2. Subsequently, the design manager can start the con- Compared with previous work in addressing design
straint process to obtain the modified frames. For illustration, changes, we believe that our proposed method is more flexible
Fig. 11 shows a new instance for Frame4 after the design in checking and correcting design changes for the geometric
change has been propagated from the updated Frame2 as models, as the method does not underline the different disci-
shown in Fig. 12. plines involved in a design project, but instead focuses on the
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 239
 dependency network to represent design rationale.’’ J. Comp. in Civ.
Engrg., ASCE, 11(2), 102–112.
El-Bibany, H., and Paulson, B. C. (1994). ‘‘Collaborative knowledge-
integration systems: A tool for AEC design, management and coordi-
nation.’’ J. Microcomp. in Civ. Engrg., 9(1), 29–40.
Fischer, M., and Froese, T. (1996). ‘‘Examples and characteristics of
shared project models.’’ J. Comp. in Civ. Engrg., ASCE, 10(3), 174–
182.
Fowler, J. E. (1996). ‘‘Variant design for mechanical artifacts: A state-of-
art survey.’’ Engrg. with Comp., 12, 1–15.
Froese, T., and Paulson, B. C. (1994). ‘‘OPIS: An object model-based
project information system.’’ J. Microcomp. in Civ. Engrg., 9(1), 13–
28.
Froese, T. (1996). ‘‘STEP and the building construction core model.’’
Proc., 3rd Congr. of Comp. in Civ. Engrg., ASCE, New York, 455–
451.
Helpenstein, H. J. (1993). ‘‘CAD geometry data exchange using STEP.’’
Res. Rep. for ESPRIT, Subseries PDT, Proj. 2195, CADEX, Berlin, Vol.
1.
Hirasawa, G., and Yamazki, Y. (1994). ‘‘A frame-based drawing manage-
ment system based on building object model.’’ Proc., 1st Congr. Held
in Conjunction with A/E/C Systems ’94, ASCE, New York, Vol. 1, 555–
FIG. 12. Updated Frame2 562.
IAI. (1996). ‘‘Industrial alliance for interoperability.’’ 具http://www.
interoperability.com典.
geometric models that are used. The relations represented by ISO. (1989). ‘‘International Organization for Standardization.’’ ISO/
symbolic equations are rigid and strong, and hence provide the TC184/SC4, ISO/DP 10303, Industrial Automation Systems-Product
robust links for the management of design changes made to Data Representation and Exchange, International Standards Organiza-
the related geometric models. tion, Geneva.
However, further work is needed to improve the facility of Kondo, K. (1992). ‘‘Algebraic method for manipulation of dimensional
relationships in geometric models.’’ Computer-Aided Des., 24(3), 141–
the proposed system to solve complex geometric models. The
147.
stability and efficiency of the knowledge-based constraint Lakmazaheri, S. (1997). ‘‘Constraint-based reasoning via Grobner bases.’’
solver also need more testing and improvement, since the con- J. Artificial Intell. for Engrg. Des., Anal. & Manufacturing, 11(1), 5–
straint solver is responsible for maintaining design change 15.
within a geometric model. More context properties, such as Lee, J.-Y., and Kim, K. (1996). ‘‘Geometric reasoning for knowledge-
product material and specification, will also be added. Deci- based parametric design using graph representation.’’ Computer-Aided
Des., 28(10), 831–841.
sion property for the geometric models in a group pattern is
Light, R., and Gossard, D. (1982). ‘‘Modification of geometric models
required to describe why such models’ relations should be through variational geometry.’’ Computer-Aided Des., 14(4), 209–214.
made, and can be used by the designers to negotiate with each Lottaz, C., Clement, D. E., Faltings, B. V., and Smith, I. F. C. (1999).
other to convey a better understanding of the design back- ‘‘Constraint-based support for collaboration in design and construc-
ground and rationale. This will be added in the near future. tion.’’ J. Comp. in Civ. Engrg., ASCE, 13(1), 23–35.
Besides, the STEP/EXPRESS will be adopted to describe our Mackworth, A. K. (1997). ‘‘Constraint-based design of embedded intel-
ligent system.’’ Constraints, 2(1), 83–86.
geometric models and the constraint knowledge so as to extend
Martini, K., and Powell, G. H. (1992). ‘‘Geometric modeling approaches
the scope of data exchange of the current system. It is in- for civil engineering and building design.’’ Microcomp. in Civ. Engrg.,
teresting to note that STEP researchers have also suggested 7, 3–14.
consideration of a parametric product in the future STEP/ Mokhtar, A., Bedard, C., and Fazio, P. (1998). ‘‘Information model for
EXPRESS standard (Helpenstein 1993). managing design changes in a collaborative environment.’’ J. Comp. in
Civ. Engrg., ASCE, 12(2), 82–92.
APPENDIX. REFERENCES Pena-Mora, F., Sriram, D., and Logcher, R. (1995). ‘‘Design rationale for
computer-supported conflict mitigation.’’ J. Comp. in Civ. Engrg.,
Aldefeld, B. (1988). ‘‘Variation of geometries based on geometric-rea- ASCE, 9(1), 57–72.
soning method.’’ Computer-Aided Des., 20(3), 117–126. Rezgui, Y., Cooper, G., and Brandon, P. (1998). ‘‘Information manage-
Anderl, R., and Mendgen, R. (1996). ‘‘Modeling with constraint: Theo- ment in a collaborative multiactor environment: The COMMIT ap-
retical foundation and application.’’ Computer-Aided Des., 28(3), 155– proach.’’ J. Comp. in Civ. Engrg., ASCE, 12(3), 136–144.
168. Saad, M., and Maher, M. L. (1996). ‘‘Shared understanding in computer-
Autodesk, Inc. (1995). AutoCAD user’s guide, Release 13, San Rafael, supported collaborative design.’’ Computer-Aided Des., 28(3), 183–
Calif. 192.
Borut, Z., Nikola, G., and Clapworthy, G. (1996). ‘‘Constraint-based ob- Stouffs, R., Tuncer, B., and Schmitt, G. (1998). ‘‘Support for information
ject modeling.’’ J. Engrg. Des., 7(2), 209–232. and communication in a collaborative project.’’ Artificial intelligence
Bouma, W., Fudos, I., Hoffmann, C., Cai, J., and Paige, R. (1995). ‘‘Ge- in design ’98, J. S. Gero and F. Sudweeks, eds., Kluwer Academic,
ometric constraint solver.’’ Computer-Aided Des., 27(6), 487–501. Dordrecht, The Netherlands.
Case, M. P., and Lu, S. C.-Y. (1996). ‘‘Discourse model for collaborative Wang, Z. (1999). ‘‘Knowledge-based parametric design for engineering
design.’’ Computer-Aided Des., 28(5), 333–345. and collaboration.’’ Doctoral thesis submitted to Nanyang Technolog-
de la Garza, J. M., and Alcantatara, P. T., Jr. (1997). ‘‘Using parameter ical University, Singapore.

240 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 ESTIMATING LABOR PRODUCTIVITY USING PROBABILITY INFERENCE
NEURAL NETWORK
By Ming Lu,1 S. M. AbouRizk,2 and Ulrich H. Hermann3

ABSTRACT: This paper discusses the derivation of a probabilistic neural network classification model and its
application in the construction industry. The probability inference neural network (PINN) model is based on the
same concepts as those of the learning vector quantization method combined with a probabilistic approach. The
classification and prediction networks are combined in an integrated network, which required the development
of a different training and recall algorithm. The topology and algorithm of the developed model was presented
and explained in detail. Portable computer software was developed to implement the training, testing, and recall
for PINN. The PINN was tested on real historical productivity data at a local construction company and compared
to the classic feedforward back-propagation neural network model. This showed marked improvement in per-
formance and accuracy. In addition, the effectiveness of PINN for estimating labor production rates in the context
of the application domain was validated through sensitivity analysis.

INTRODUCTION In the following section, previous NN applications in the

problem domain are first reviewed.
Problem Domain
Review of NN Applications
Estimating labor production rates (manhours per unit) is
both an art and a science. In general, the estimator develops Moselhi et al. (1990) cited the prediction of a realistic pro-
the rate for a given project by starting with a ‘‘base rate’’ and ductivity level for a certain trade as an aspect of construction
modifying it to reflect the specific conditions he/she expects that can be modeled with NNs. Factors such as job size, build-
to encounter in the project being estimated. The base rate is ing type, overtime work, and management conditions are typ-
often determined statistically from past historical data or from ically considered by an estimator and can easily be manipu-
industry standards. In the context of industrial productivity es- lated for use as NN inputs.
timating, the estimator accordingly adjusted the rate up or Karshenas and Feng (1992) analyzed earth-moving equip-
down by applying a difficulty multiplier to reflect overall fa- ment productivity with a NN application. A modular NN struc-
vorable or unfavorable conditions. In determining the diffi- ture was used to make it possible to add specifications of new
culty multiplier, consideration is only given to a few of the equipment with only a brief training session. Each module
major factors that are thought to affect job productivity, such represents a distinct type of equipment that was trained with
as the installation location of the pipe (inside a fabrication two inputs, four hidden nodes, and one output within a back-
shop or on the job site), and the material type of the welding propagation training algorithm.
(e.g., carbon steel or stainless steel). AbouRizk and Wales (1997) used NNs as a means of ap-
The challenges of this approach include the fact that it is plying the effects of environmental site conditions to the labor
not straightforward to create a conventional mathematical production rate on an activity. Daily average temperature, pre-
model to accommodate the impacts of numerous factors on the cipitation, and cumulative precipitation over the previous 7
target risk variable. The decision process relies heavily on in- days were identified as three key environmental site conditions
dividual experience, and the results are often inconsistent, re- and used as inputs into a feedforward back-propagation NN
flecting the experience and temperament of the estimator. training algorithm. The output was a productivity factor such
Artificial neural networks (NNs) have been proposed by that a value >1.0 indicates that environmental site conditions
many as an alternative for streamlining the process and reduc- produce a greater than average productivity. On the other hand,
ing the subjective nature of the work. Most models, however, a productivity factor of <1.0 indicates that the environmental
were based on point predictions of production rates with which site conditions result in below average productivity.
estimators were uncomfortable. The point prediction by NN Chao and Skibniewski (1994) performed a case study in
can be defined as a single value predicted by NN models with- which a NN was used to predict the productivity of an exca-
out any backup information on the risks of taking this value vator. They identified two main factors that affect an excava-
as correct. The new NN model presented in this paper arises tor’s productivity: job conditions and operation elements. Job
out of the need for accurate prediction in the form of a distri- conditions include the characteristics of the environment, such
as soil conditions, and specific characteristics of the excavator
bution at the output range. The estimators will be able to make
and excavation, such as the vertical position of the cutting
a decision for a future scenario based on the results recalled
edge. Operational elements, in contrast, include characteristics
by the NN model and personal preferences and experience.
not directly related to the excavating operation; for example,
1
the effect of wait time for trucks and extra tasks other than
PhD Candidate, Dept. of Civ. and Envir. Engrg., Univ. of Alberta, excavating. Two NNs were used for the purpose of this case
Edmonton, AB, Canada T6G 2G7. E-mail: [email protected]
2
Prof., Dept. of Civ. and Envir. Engrg., Univ. of Alberta, Edmonton,
study. The first was used to estimate the excavator cycle time.
AB, Canada T6G 2G7. E-mail: [email protected] Four key factors were identified as having an influence: cycle
3
Dist. Constr. Engr., PCL Industrial Constructors, Inc., Edmonton, AB, time (including swing angle), horizontal reach, vertical posi-
Canada T6E 3N7. E-mail: [email protected] tion, and soil type (job conditions). The output of the first
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, network was then incorporated into the second network, which
2001. To extend the closing date one month, a written request must be examined the effect of the operational elements on the pro-
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on April 20, 1999.
ductivity.
This paper is part of the Journal of Computing in Civil Engineering, Portas and AbouRizk (1997) proposed a feedforward back-
Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/0004-0241– propagation NN model for estimating construction production
0248/$8.00 ⫹ $.50 per page. Paper No. 20740. rates of formwork. The network outputs a single-point predic-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 241
tion along with a number of output zones, with equal likeli- median of each subrange can be used to represent the typical
hood of the production rate being in any one zone. The output value for each output zone and to derive a NN predicted value.
zones are symmetric and divided evenly across the range of The PINN uses the same strategy as the model described in
likely production rate values. During training, the output zone Knowles (1997) by incorporating Kohonen’s LVQ in NN
whose output coincides with the actual production rate is re- learning. The main difference is that the classification and pre-
warded with a primary score of 1.0, representing strong cer- diction networks are combined in an integrated network, which
tainty. A certain degree of fuzziness is considered by reward- required the development of a different training and recall al-
ing the two adjacent output zones with secondary scores of gorithm. Murtaza and Fisher (1993) utilized Kohonen’s un-
0.5, representing weak certainty. All the other output zones are supervised learning algorithm called self organizing map or
assigned a score of 0. Once the NN is trained and inputs are SOM for modular construction decision making. Kohonen’s
entered, the NN will predict a point value as well as the like- LVQ combines unsupervised and supervised learning and is
lihood of production rates being within the output zones. This recommended for statistical pattern recognition problems (Ko-
model achieved limited success, and its limitation was over- honen 1995). Three options for the LVQ algorithms (LVQ1,
come by the work discussed next. LVQ2, and LVQ3) were proposed. Kohonen’s research showed
Knowles (1997) presented a two-stage NN model in pre- that each of the three LVQ variations yields similar accuracy
dicting pipe-installation labor productivity. The input factors in most statistical pattern recognition tasks, although a differ-
are used to invoke a learning vector quantization (LVQ) clas- ent philosophy underlies each algorithm. The LVQ1 was uti-
sification process and then a predictive one. With the classi- lized in the learning process of the PINN model.
fication, the model predicts whether the output is likely in a
typical or nontypical range. The proper feedforward back- Overview of PINN Topology and Process
propagation network is then executed. The drawback of this The topology of the PINN model is given in Fig. 1. It is
method is that a buildup of errors occurs when the classifi- composed of four layers. The middle layers are a Kohonen
cation fails. For instance, if the classification accuracy is 90% classifier and a Bayesian layer. The outcome of the PINN
at the first stage of NN, and the prediction accuracy at the model at the output layer is a probability density function
second stage of NN is 85%, the prediction accuracy of the (PDF) or a distribution reflecting the likelihood of the target
whole NN is only 76.5% (90% ⫻ 85%). This problem moti- variable occurring in a given zone. The mode of the distri-
vated the development of the model described in this paper bution and its mean can serve as point predictions.
by taking a different probabilistic approach, which is more The PINN process consists of the following four stages.
direct and more meaningful in terms of giving a point predic- Stage 1. Preparation, which deals with
tion and quantifying its associated probability.
• Scaling data at the input layer, which will be discussed in
PROBABILITY INFERENCE NN (PINN) MODEL the ‘‘Data Preprocessing’’ subsection
• Setting up output zones at the Kohonen layer, which will
Introduction of PINN Model be addressed in the ‘‘Output Zone Setup’’ subsection
• Definition of processing elements (PEs) and how they
Specht (1991) revisited probabilistic NN (PNN) and general function at the Kohonen layer, which will be discussed in
regression NN (GRNN) algorithms with the objective of in- the ‘‘Processing Element at the Kohonen Layer’’ subsec-
tegrating statistics and neural training. The GRNN/PNN is a tion
memory-based feedforward NN model, where the training is
performed in one pass, thus requiring less training time. The Stage 2. Learning, which takes place between the input
GRNN/PNN is able to identify a posterior distribution over layer and the Kohonen layer using the LVQ algorithm (this
the NN weight vectors and a point-value prediction is gener- does not involve the Bayesian layer or the output layer), will
ated based on the predicted distribution. However, based on be discussed in the ‘‘NN Learning Process’’ subsection.
experimentations and observations, GRNN/PNN is not quite Once learning is achieved, the input-output patterns are
tolerant of noisy data (inaccurate or incomplete records) and coded into the weight vectors of the processing elements at
imposes a demanding standard of data quality that is hard to the Kohonen layer.
achieve in reality. The memory demand and computing time Stage 3. Investigating whether the NN has been success-
for GRNN/PNN increase very rapidly when the dimension of fully trained is accomplished through the following steps:
input vector and the quantity of training samples increase.
The PINN model uses similar topology as the GRNN/PNN 1. Feed the input vectors of the training and testing records
model but is a refinement of it. This is because PINN gener- into the input layer of the PINN model.
alizes the underlying statistical patterns within training data 2. Project the input vector of one record onto the Kohonen
and codes those patterns into a limited number of weight vec- layer by using the results of Stage 2. The euclidean dis-
tors through iterative learning. As a result, the number of tances between each processing element’s weight vector
weight vectors will not proportionally increase with the in- and the scaled input vector are calculated.
crease of dimensionality and quantity of training data. 3. An in-zone competition occurs within every output zone
The PINN model imbeds the output zone concepts described at the Kohonen layer, which is detailed in the ‘‘In-Zone
in Portas and AbouRizk (1997). In the application domain of Competition Strategy at Kohonen Layer’’ subsection.
industrial labor productivity estimating, the profile of actual The PE with the minimum euclidean distance value wins.
historical productivity data reveals a spread range. The range 4. Project the winner PE at the Kohonen layer onto the
of NN output value (i.e., the production rate multiplier) is Bayesian layer. The Bayesian layer holds a PDF approx-
evenly divided into a number of subranges, or output zones, imator. The euclidean distance values of the winner PEs
which are actually some discrete clusters with continuous are the inputs to the PDF approximator. The following
boundaries. The higher the multiplier value, the more difficult ‘‘PDF Approximator at Bayesian Layer’’ subsection dis-
and more demanding the job is, and the lower the productivity cusses the components and operations in more details.
for the job. Thus, each output zone gives an indication of the 5. The output is mapped from the Bayesian layer and pre-
relative work difficulty and productivity level, for instance, sented in the form of a PDF at the output layer. Two
output zone 0–0.7 stands for easier work and higher produc- point predictions are calculated in addition to the PDF,
tivity level when compared with output zone 0.7–1.4. The namely, the mode and the weighted average. The ‘‘Out-
242 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 FIG. 1. Topology of PINN Model

puts at Output Layer’’ subsection includes details about Output Zone Setup
the NN outputs.
As discussed in the section ‘‘Introduction of PINN Model,’’
6. Check the NN outputs against the actual outputs of the
the likely range of output values is evenly divided into a num-
training and testing records. If the results are satisfactory,
ber of output zones. The output zone boundary setup is im-
then the NN is declared to have been trained; otherwise,
portant for PINN learning and recall. Wide zones are generally
repeat Stage 1 using different parameters at each layer.
not helpful to the decision maker and hence should be avoided.
Zones that are unacceptably tight may prevent PINN from
Stage 4. Once the model calibration is done, the NN can learning. It requires some trials to obtain reasonable output
be used to recall the output value for any given input vector, zone boundaries, and the following two aspects should be con-
which is similar to Stage 3 using the final results determined sidered:
in Stages 2 and 3. A sample calculation is given in the ‘‘Sam-
ple of Recall Process’’ subsection. • Precision requirement of the user (i.e., the zone width or
subrange that suffices for the user to make decisions)
Data Preprocessing • Distribution of actual output data over the output zones;
a uniform distribution of actual output data over all zones
At the input layer of the PINN model (shown by Fig. 1),
generally yields better results
the number of input nodes corresponds to the dimension of
the input vector. The dimension of the input vector depends PEs at Kohonen Layer
on the number of input factors and the input data types. Three
input data types are used to define NN input factors (i.e., raw, At the Kohonen layer, each output zone contains an equal
rank, and binary). Raw is used simply for quantitative input number of PEs. Each PE is associated with a weight vector
factors, like general expense ratios, winter construction per- [also referred to as a codebook vector (Kohonen 1995)].
centages, or quantities of work. Rank is used to convert sub- Visually, a weight vector is a set of links that emanate from
jective factors, like crew ability ratings, into numeric format. one PE and end at each input node, as illustrated in Fig. 1.
And binary is used to group textual factors into numeric for- Thus the dimension of a weight vector is equal to that of the
mats, like material type and project definition. It should be input vector (the number of input nodes). An output zone at
noted that an input factor of the raw or rank type corresponds the Kohonen layer can be visualized as a chip containing a
to one input node at the input layer and an input factor of the number of pins (PEs). During NN training the orientation of
binary type corresponds to a number of input nodes depending those pins is gradually fine-tuned to capture the underlying
on the number of groups for the factor. For illustration, input statistical patterns within the training data (Kohonen 1995).
factors and data types for the pipe installation NN model are Our experience indicates that the number of PEs assigned
listed in Table 1. A sample record for the pipe installation NN to each class should be close to the average frequency in the
training is listed in Table 2, showing both the raw data and histogram of training data output values (i.e., the average num-
converted NN input data. The NN input data is normalized ber of training samples in one output zone).
and scaled between 0 and 1 at the input layer. These scaled
inputs will be passed forward for NN training. At the Kohonen NN Learning Process
layer all weight vectors are randomly initialized between 0 and Data of all the training records is scaled at the input layer.
1. The scaled input data is fed into the model to calibrate the
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 243
 TABLE 1. NN Input Factors and Data Type TABLE 2. NN Data Sample
NN input factor Data Options and remarks NN input factor Raw data NN input data
(1) type (3) (1) (2) (3)
(2) Project location Rural 001
Project location Binary Urban, rural, camp job Administration require- 0.235 0.235
Administration Raw General expense ment
Year of construc- Binary 89 ⬃ 92, 93 ⬃ 94, 95 ⬃ 96, 97 ⬃ 99 Year of construction 94 ⬃ 95 0100
tion Province/state Alberta 10
Province/state Binary Alberta, Saskatchewan Contract type Reimbursable 10
Contract type Binary Reimbursable, lump sum Client Shell 1.037
Client Raw Index derived from historical data Engineering firm Colt 0.682
Engineering firm Raw Index derived from historical data Project manager G. Trunn 1.036
Project manager Raw Index derived from historical data Superintendent J. Yakemchuk 0.982
Superintendent Raw Index derived from historical data Project definition Chemical 10000
Project definition Binary Chemical, cryogenic, gas, refining Work scope Confined to specific area 10
Work scope Binary Confined/scattered Project type Plant upgrade, no shutdown 0100
Project type Binary Upgrade shutdown, grass root, etc. Prefabrication/field 10%, 90%, 0% 10, 90, 0
Prefabrication/field Raw Percentages for prefabrication work
work Average crew size 25 ⬃ 50 0100
Average crew size Binary <25, 25 ⬃ 50, 50 ⬃ 100, >100 Peak crew size 50 ⬃ 100 00100
Peak crew size Binary <25, 25 ⬃ 50, 50 ⬃ 100, 100 ⬃ 150, Unionized Yes 10
>150 Equipment and mate- 9.4 9.4
Unionized Binary Yes, no rial
Equipment and ma- Raw Equipment and material cost per direct Extra work 0.661 0.661
terial manhours Change order 0.019 0.019
Extra work Raw Original project cost/final project cost Drawing and specifica- Excellent 5
Change order Raw Number of change orders per total direct tions quality
manhours Activity location clas- Inside <10-ft high 000100
Drawing and speci- Rank 1 poor, 3 average, 5 excellent sification
fications quality Total quantity (learn- 6,055 6,055
Location classifica- Binary U/G on site, fab shop, A/G on site, etc. ing)
tion Installation quantities 210, 905, 4,940 210, 905, 4,940
Total quantity Raw Total quantity in DialnFt Material type Carbon steel 00100
(learning) Method of installation Hand rigging (%), machine 10, 90
Installation quanti- Raw Quantity for size ranges, <2⬙, 2⬙ ⬃ 16⬙, rigging (%)
ties >16⬙ Pipe supports 0.45 0.45
Material type Binary Alloy, carbon steel, FRP/PVC, etc. Boltups 4.77 4.77
Method of installa- Raw Percentages of hand rigging Valves 1.59 1.59
tion Screwed joints 0 0
Pipe supports Raw Number of pipe supports per foot of Miscellaneous compo- 3.18 3.18
pipe nents
Boltups Raw Number of boltups per foot of pipe Welding impact 1.25 1.25
Valves Raw Number of valves per foot of pipe Season Winter (%), summer (%) 10, 90
Screwed joints Raw Number of screwed joints per foot of Crew ability Average 3
pipe Site working condi- Many problems 2
Miscellaneous com- Raw Installation of miscellaneous components tions
ponents (manhours per foot of pipe) Inspection, safety, and Detailed 2
Welding impact Raw Welding multiplier (miscoding on site) quality
Season Raw Percentages of winter and summer work Overall degree of diffi- Low 2
Crew ability Rank 1 very low, 3 average, 5 very high culty
Site working condi- Rank 1 extreme problems, ⬃ 5 no problem
tions
Inspection, safety, Rank 1 extremely detailed, ⬃ 5 highly tolerant of the training record; Wij = current weight vector of the
and quality global winner PEi; and W⬘ij = updated weight vector of
Overall degree of Rank 1 very low, 3 average, 5 very high the global winner PE i. The repulsion rate is initially set
difficulty between 0 and 1 and is reduced gradually until it ap-
proaches 0 at the end of learning.
weight vectors of the PEs at the Kohonen layer, using the LVQ 2. Following the global competition, an in-zone competi-
algorithm suggested by Kohonen (1995). tion among processing elements occurs only at the output
The learning process involves a number of iterations, each zone into which the actual output value of the training
of which is comprised of the following: record falls. Prior to the in-zone competition, a ‘‘con-
science’’ value is added to each PE’s euclidean distance
and adjusted over the learning iterations, so as to effec-
1. The euclidean distances between the input vector of a tively prevent one PE within a specific output zone from
training record and each PE’s weight vectors are calcu- winning all the time and activate as many PEs as possible
lated. The PE that has the smallest euclidean distance in the learning process. The formulas to calculate the
value is declared to be a global winner. If the global conscience euclidean distance can be found throughout
winner PE does not belong to the same output zone that the pertinent literature (NeuralWare 1993). Readers can
the actual output value of this training record falls into, refer to Appendix I for details. The method adopted here
the weight vector of the global winner PE is penalized is as follows: (1) The ‘‘conscience’’ euclidean distance
according to the following: between each PE’s weight vector and the input vector
W⬘ij = Wij ⫺ RR ⫻ ( X j ⫺ Wij) (1) is calculated; (2) the PE with the shortest ‘‘conscience’’
euclidean distance value is declared to be an in-zone
where RR stands for repulsion rate, which is a learning winner; and (3) only the in-zone winner PE is rewarded
rate to penalize the global winner PE i; X j = input vector using
244 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 W⬘ij = Wij ⫹ AR ⫻ ( X j ⫺ Wij) (2) TABLE 4. Updating Weight Vector in First Learning Stage
Input vector, X1 PE1, W11 PE2, W21 PE3, W ⬘31
where AR stands for attraction rate, which is a learning (1) (2) (3) (4)
rate to reward the in-zone winner PE i; X j = input vector
0.3 0.7232 0.0844 ⫺0.0927
of a training record; Wij = current weight vector of the 0.2 0.6772 0.9129 0.3073
in-zone winner PE i; and W⬘ij = updated weight vector of 0.7 0.8780 0.4422 1.1867
the in-zone winner PE i. Like the repulsion rate, the at- 0.8 0.6307 0.8774 0.8353
traction rate is initially set between 0 and 1, and is re- 0.4 0.7523 0.7261 0.3943
duced gradually until it approaches 0 at the end of learn- 0.5 0.8575 0.0439 0.1597
ing iterations. 0.5 0.2088 0.4567 0.2857
0.9 0.6310 0.4542 ⫺0.6881
0.8 0.6065 0.2413 1.0482
0.1 0.8804 0.9188 1.5346
A sample calculation of one learning iteration is presented 0.4 0.8232 0.7465 ⫺0.0004
next to illustrate the learning process. 0.6 0.0485 0.1158 0.8777
As shown in Fig. 1, the dimension of input vector is 12 and
the output range 0–4 is divided into four output zones (i.e.,
0–1, 1–2, 2–3, and 3–4). Each output zone contains three TABLE 5. Updating Weight Vector in Second Learning Stage
PEs. Note that this sample is a simple model and serves for Input vector, X1 PE1, W11 PE2, W21 PE3, W ⬘31
illustration. Problems encountered in a real situation, which (1) (2) (3) (4)
are suitable for the PINN model to solve, are mostly high
0.3 0.7232 0.0844 0.2564
dimensional; the number of input nodes may exceed 100. 0.2 0.6772 0.9129 0.2119
The input vector of a training record is scaled between 0 0.7 0.8780 0.4422 0.7541
and 1, and the weight vector of a PE at the Kohonen layer is 0.8 0.6307 0.8774 0.8039
randomly initialized between 0 and 1. Table 3 shows the input 0.4 0.7523 0.7261 0.3994
vector X 1 and the weight vectors of the three PEs in Zone 1. 0.5 0.8575 0.0439 0.4622
The euclidean distance ED is calculated between the input 0.5 0.2088 0.4567 0.4762
0.9 0.6310 0.4542 0.7235
vector X 1 and each weight vector Wi1 as ED11 = 1.4163, ED21 0.8 0.6065 0.2413 0.8276
= 1.5746, and ED31 = 1.2963. Suppose that PE3 in Zone 1 is 0.1 0.8804 0.9188 0.2594
the global winner PE among all the PEs, which gives a min- 0.4 0.8232 0.7465 0.3555
imum ED of 1.2963. If the actual output of this training record 0.6 0.0485 0.1158 0.6309
does not fall into Zone 1 (i.e., outside the subrange 0–1), then
the weight vector of PE 3 W31 is updated by (1), as shown in
Table 4. Notice that the repulsion rate is set to be 0.8 at the In-Zone Competition Strategy at Kohonen Layer
start of learning in the sample calculation. Kohonen (1995) The in-zone competition at the Kohonen layer occurring in
recommended a smaller initial value such as 0.06 for the re- the recall stage differs from that occurring in the learning
pulsion rate and attraction rate to obtain better results. stage. Once adequate training is complete, PINN is capable of
If the actual output of this training record does fall into Zone mapping the input onto the output. At the Kohonen layer, for
1 (i.e., within the subrange 0–1), then no weight vector is one output zone, the PE that has the shortest global euclidean
updated in the global competition. The learning process steps distance (no conscience value) between its weight vector and
into the second phase. the input vector is declared to be an in-zone winner PE. Only
In the first training iteration, the conscience value for every the in-zone winner PE advances to the Bayesian layer.
PE in Zone 1 is determined to be equal to 0 (Appendix I). So Unlike the GRNN/PNN, which takes the average of a PE’s
the ‘‘conscience’’ euclidean distance value is equal to the orig- euclidean distance within one output zone as the parameter to
inal euclidean distance value. The in-zone competition winner pass forward (Specht 1988), PINN takes the minimum of the
is PE3, which gives the minimum ‘‘conscience’’ ED of 1.2963, PE’s euclidean distance within one output zone as the param-
and its weight vector W31 is updated by (2), as shown in Table eter to pass forward. The reason for the difference is that in
5. GRNN/PNN each PE corresponds to one training record and
The above learning process will iterate through all the train- different numbers of PEs lie in different output zones. In the
ing records for a sufficient number of runs. Notice that during proposed PINN, the PE does not match the training record,
the learning process the repulsion rate, attraction rate, and the and an equal number of PEs dwell in each output zone and
conscience value are dynamically updated to calibrate the work together in the Kohonen layer of PINN to generalize the
weight vectors. underlying patterns within the training records by implement-
ing LVQ.

TABLE 3. Scaled Input Vector and Initial Weight Vectors PDF Approximator at Bayesian Layer
Input vector, X1 PE1, W11 PE2, W21 PE3, W31 As illustrated in Fig. 1, at the Bayesian layer each output
(1) (2) (3) (4) zone only contains the winner PE from the in-zone competi-
0.3 0.7232 0.0844 0.0819 tion at the Kohonen layer in the recall stage. The main com-
0.2 0.6772 0.9129 0.2596 ponents at the Bayesian layer are a kernel function and a soft-
0.7 0.8780 0.4422 0.9704 max activation function that are used to approximate the
0.8 0.6307 0.8774 0.8196
0.4 0.7523 0.7261 0.3969
probability density of one input vector being within each out-
0.5 0.8575 0.0439 0.3109 put zone in the following steps:
0.5 0.2088 0.4567 0.3809
0.9 0.6310 0.4542 0.0177 1. The square of the euclidean distance value of the winner
0.8 0.6065 0.2413 0.9379 PE from each zone is passed into the kernel function,
0.1 0.8804 0.9188 0.8970 which is the Gaussian function of Bayesian methods in
0.4 0.8232 0.7465 0.1775
0.6 0.0485 0.1158 0.7543
statistics as described in Specht (1988) and shown in (3).
If the number of output zones is N, then for each input
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 245
 vector X j, the kernel function is evaluated for N times TABLE 7. Recall Calculations at Bayesian Layer
and outputs one q value for each zone. Values for each zone Zone 1 Zone 2 Zone 3 Zone 4
qi = exp[⫺( Wij ⫺ Xj) ( Wij ⫺ Xj)/2␴ ]
T 2
(3) (1) (2) (3) (4) (5)
Median 0.500 1.500 2.500 3.500
where i = 1, 2, 3, . . . , N ( N = number of output zones); Winner PE’s ED 0.951 2.567 1.843 2.922
Xj = one input vector fed into PINN at the input layer; Gaussian output q 0.636 0.037 0.183 0.014
and ( Wij ⫺ Xj)T( Wij ⫺ Xj) = square of the euclidean Softmax output p 0.731 0.043 0.210 0.016
distance value between the input vector Xj and the win-
ner PE’s weight vector Wij in the output zone i(i = 1, 2,
3, . . . , N ). The value of ␴ is a smoothing factor, is the three processing elements. Table 6 lists the scaled input vector
only adjustable parameter of the Gaussian function [(3)], X1 and the weight vectors of the three processing elements in
and controls the shape of the PDF. The greater ␴, the Zone 1. The euclidean distance values ED between the weight
more dispersed the probability density graph. The value vectors and the input vector are calculated as ED11 = 0.9513,
of ␴ is critical to PINN’s predicting capability and can ED21 = 1.1670, and ED31 = 1.3249.
be determined through iterative adjustments. In regard to At output Zone 1, PE1 with a minimum ED (0.9513) is the
industrial productivity application, ␴ should fall in the in-zone competition winner and proceeds to the Bayesian
range between 0.8 and 1.2. layer. Suppose the winner PEs from the other three zones also
2. The Softmax Activation function (Sarle 1997) as shown are determined in a similar manner and proceed to the Baye-
in (4) makes the sum of the calculation results (q values) sian layer. Table 7 lists their euclidean distance values and
from (3) equal to 1, so that the final output from the outputs from the Gaussian function and the Softmax function.
Bayesian layer can be interpreted as posterior probabil- From Table 7, the probability of output being within Zone
ities ( p values). 1 is 0.731; hence, the mode output value is found to be the
median of Zone 1 (i.e., 0.5). The weighted average output
qi value is obtained by calculating the sum product of the Soft-

冘
pi = N (4) max output ( p values) and median of each output zone (i.e.,
qi 0.5 ⫻ 0.731 ⫹ 1.5 ⫻ 0.043 ⫹ 2.5 ⫻ 0.210⫹ 3.5 ⫻ 0.016 =
i=1 1.011).
where qi = output value from Gaussian function [(3)] for
output zone i (i = 1, 2, 3, . . . , N ); and N = number of IMPLEMENTATION OF PINN MODEL
output zones. Computer software based on the PINN model is developed
for learning and testing in the environment of Microsoft Ac-
Outputs at Output Layer cess 97 and Visual Basic for Applications. Historical piping
At the output layer, the probability distribution predicted by productivity data of 66 projects resulting in 119 records of a
PINN is presented in the form of a PDF graph, which portrays construction company were collected and compiled into NN
the uncertainty of the output value. In addition to the predicted input data for three labor-intensive activities (i.e., pipe instal-
distribution, PINN calculates two point prediction values: lation, pipe welding, and pipe hydro-testing). In the following
sections, pipe installation is used to illustrate the testing and
• Mode value: the median of the output zone or subrange validation of the PINN model.
that has the greatest probability The PINN model for pipe installation has a total of 81 input
• Weighted average value: the sum product of the median nodes. (The input factors and sample data are shown in Tables
and probability of each output zone—the user should 1 and 2). The output range is divided into 20 output zones,
treat this point-value prediction carefully by checking the each with an equal width of 0.72. Ten processing elements are
PDF graph. assigned to each output zone. The attraction rate and repulsion
rate are both equal to 0.06. The smoothing factor of the kernal
function is equal to 0.8.
Sample of Recall Processing
One hundred one records are used for PINN learning, and
A sample PINN recall calculation is given below for illus- 18 records are reserved to test the calibrated model. The learn-
tration. ing process takes 300 iterations.
Suppose that the NN is trained and ready to recall the output
for an input vector. As shown in Fig. 1, the dimension of the Validation of PINN Model on Testing Data
input vector is 12 and the output range 0–4 is divided into
four zones (i.e., 0–1, 1–2, 2–3, and 3–4). Each zone contains The testing of the calibrated network on the 18 unseen
records is summarized in Fig. 2. When measured against the
TABLE 6. Trained PINN Ready to Recall for Given Input Vector actual output values of the test data, for the mode value, the
average absolute error is 0.57 and the maximum absolute error
Input vector, X1 PE1, W11 PE2, W21 PE3, W31 is 2.02 and, for the weighted average value, the average ab-
(1) (2) (3) (4) solute error is 0.75 and the maximum absolute error is 2.23.
0.5 0.8668 0.9681 0.6722 Considering a wide output range of about 15, the error is rea-
0.6 0.3437 0.4966 0.1627 sonable and acceptable.
0.4 0.6202 0.0093 0.5633 To compare the PINN model with a back-propagation NN,
0.9 0.2472 0.2213 0.6362
0.4 0.3469 0.6834 0.4307
the same training records were used to train a three-layer feed-
0.7 0.7326 0.4165 0.5033 forward back-propagation NN, which has 81 input nodes at
0.8 0.6803 0.9864 0.8412 the input layer, 40 hidden nodes at the middle layer, and 1
0.1 0.1366 0.4152 0.7824 output node at the output layer. The training parameters are a
0.3 0.0517 0.5621 0.8916 learning rate equal to 0.8 and a momentum rate equal to 0.4.
0.6 0.9558 0.5190 0.7239 The transfer function is a symmetric logistic function. After
0.9 0.9264 0.5150 0.3458
0.6 0.7467 0.5772 0.0914
training was completed, the testing set of 18 records previ-
ously used to test PINN was fed into the model. The testing
246 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 FIG. 2. Comparison of PINN and Feedforward Back-Propagation NN (FF BP NN)

FIG. 3. PINN Output for Base Case Scenario

results of the back-propagation NN model compared with that Sensitivity Analysis of PINN Model
of the PINN are shown in Fig. 2. From Fig. 2, it is observed
that the PINN model outperforms the back-propagation NN A recall program based on the trained PINN model for pipe
model in terms of point prediction accuracy, by coming closer installation was developed to validate its effectiveness and ac-
to the actual output values. curacy in the context of the application domain. ‘‘What if’’
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 247
scenarios are tested on the NN model by changing some input APPENDIX I. DESINO’S METHOD TO CALCULATE
factors in order to understand the impact of such changes on CONSCIENCE VALUE
the output values. The response of the NN model is compared
against that of an experienced estimator at the participating Conscience euclidean distance is defined
construction company for the purpose of model validation. D⬘i = Di ⫹ Ci (5)
The base case scenario is taken from one testing record. The
actual difficulty multiplier for this scenario is 1.24. The mode where Di is euclidean distance; and Ci is conscience value
value predicted by the PINN model is 1.181, giving an abso-
lute error of (兩1.181 ⫺ 1.24兩 = 0.059). The weighted average Ci = cf ⫻ D ⫻ (n ⫻ wf ⫺ 1) (6)
value is 1.313, giving an absolute error of (兩1.313 ⫺ 1.24兩 = where D = maximum euclidean distance out of the global com-
0.073). Fig. 3 shows the predicted probability function or dis- petition in the previous supervised learning stage; cf = con-
tribution; the chance of output falling into Zone 2 (0.8 ⬃ 1.5, science factor, which is initially set between 0 and 1 by the
median = 1.181) is 69%. user; n = number of PEs per output zone; and wf = win fre-
In the following validation tests, the actual values remain quency. The initial estimate of the win frequency value wf0 is
unknown, so the responses of the estimator based on personal set to the reciprocal of the PE number per output zone for all
experience and common sense serve as a benchmark to mea- the PEs (i.e., 1/n).
sure the performance of the PINN model. The estimator re- With NN learning ongoing, both conscience factor cf and
sponds with a trend or direction instead of a precise number frequency estimate fe are reduced gradually until they ap-
because there are so many other input factors to take into proach 0 at the end of learning.
account. During the unsupervised learning stage, for the in-zone win-
ner PE, its wf value is updated
Scenario 1
wf ⬘ = (1 ⫺ wf0) ⫻ wf ⫹ wf0 (7)
The location of pipe installation is a major consideration
when the experienced estimator determines the pipe installa- For the in-zone loser PEs, their wf values are updated
tion productivity. The installation location for the base case
scenario is ‘‘piping within a fabrication shop.’’ What if the wf ⬘ = (1 ⫺ wf0) ⫻ wf (8)
location is changed to ‘‘operating plant installation on the
Basically, the above formulas are intended to increase the
site’’? The experienced estimator responds by increasing the
wf values for the winner PE and hence increase its conscience
difficulty multiplier to a certain extent to reflect the unfavor-
value so that the winner PE will have less chances to win again
able job conditions. Response of the PINN model is that the
than the other loser PEs in the learning iterations that follow
mode value increases to 1.903 and the weighted average value
it.
increases to 1.983; the chance of the output value falling into
Zone 3 (1.5 ⬃ 2.3, median = 1.903) increases from 13% in
the base case scenario to 78%. The PINN has taken the same ACKNOWLEDGMENTS
direction as the estimator in the decision process for this sce- The work was funded by a number of construction companies in Al-
nario. berta and the Natural Science and Engineering Research Council of Can-
ada under grant number IRC-195558/96.
Scenario 2
APPENDIX II. REFERENCES
In the base case scenario, the job is done 100% in the winter
in Alberta. What if the job is done 100% in the summer? The AbouRizk, S., and Wales, R. (1997). ‘‘Combined discrete-event/contin-
estimator anticipates a reduction for the difficulty multiplier, uous simulation for project planning.’’ J. Constr. Engrg. and Mgmt.,
which means an increase of productivity level. The response ASCE, 123(1), 11–20.
Chao, L. C., and Skibniewski, M. J. (1994). ‘‘Estimating construction
of the PINN model is that the mode value remains 1.181; productivity: Neural-network-based approach.’’ J. Comp. in Civ.
however, the weighted average value decreases to 1.154. The Engrg., ASCE, 8(2), 234–251.
chance of the output value falling into Zone 2 (0.8 ⬃ 1.5, Karshenas, S., and Feng, X. (1992). ‘‘Application of neural networks in
median = 1.181) decreases significantly from 69 to 44%, earthmoving equipment production estimating.’’ Proc., 8th Conf. Com-
whereas the chance of the output value falling into Zone 1 puting in Civ. Engrg., ASCE, New York, 841–847.
(0.2 ⬃ 0.8, median = 0.5) increases significantly, from 15% Knowles, P. (1997). ‘‘Predicting labor productivity using neural net-
works.’’ MS thesis, University of Alberta, Edmonton, Alta., Canada.
in the base case scenario to 40%. Again the PINN chooses the Kohonen, T. (1995). Self-organizing maps, Springer Ser. in Information
similar course of action as the estimator in the decision process Sci., Springer, London.
for this scenario. Moselhi, O., Hegazy, T., and Fazio, P. (1991). ‘‘Neural networks as tools
in construction.’’ J. Constr. Engrg. and Mgmt., ASCE, 117(4), 606–
CONCLUSIONS 625.
Murtaza, M. B., and Fisher, D. J. (1994). ‘‘Neuromodex—Neural network
The PINN model creates a meaningful representation of a system for modular construction decision making.’’ J. Comp. in Civ.
complex, real-life situation in the problem domain and is ef- Engrg., ASCE, 8(2), 221–233.
fective in dealing with high dimensional input-output mapping NeuralWare. (1993). NeuralWorks advanced reference guide, NeuralWare
Inc., Pittsburgh, Pa.
with multiple influential factors in a probabilistic approach. Portas, J., and AbouRizk, S. M. (1997). ‘‘Neural network model for es-
The application of the PINN model in industrial labor pro- timating construction productivity.’’ J. Constr. Engrg. and Mgmt.,
duction rate estimating helps the estimator choose a course of ASCE, 123(4), 399–410.
action by giving a better understanding of the project infor- Sarle, W. S., ed. (1997). ‘‘Introduction.’’ Neural network FAQ, part 1 of
mation available and the possible outcomes that could occur. 7, periodic posting to Usenet newsgroup 具comp.ai.neural-nets典 URL
Because the probability density of each output zone is pro- 具ftp://ftp.sas.com/pub/neural/faq.html典
Specht, D. F. (1988). ‘‘Probabilistic neural networks for classification,
vided, the predicted distribution and point-prediction values mappings, or associative memory.’’ Proc., IEEE Int. Conf. on Neural
give the estimator much more confidence in the predicted re- Networks, IEEE, Piscatawy, N.J., 1, 525–532.
sult. In combination with personal experience and preferences, Specht, D. F. (1991). ‘‘General regression neural networks.’’ IEEE Trans.
a labor production rate for a new project can be determined. on Neural Networks, 2, 525–576.

248 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 FUZZY LOGIC INTEGRATED GENETIC PROGRAMMING FOR
OPTIMIZATION AND DESIGN
By Yaowen Yang1 and Chee Kiong Soh2

ABSTRACT: A fuzzy logic integrated genetic programming (GP) based methodology is proposed to increase
the performance of the GP based approach for structural optimization and design. Fuzzy set theory is employed
to deal with the imprecise and vague information, especially the design constraints, during the structural design
process. A fuzzy logic based decision-making system incorporating expert knowledge and experience is used to
control the iteration process of genetic search. Illustrative examples have been used to demonstrate that, when
comparing the proposed fuzzy logic controlled GP approach with the pure GP method, the proposed new
approach has a higher search efficiency.

INTRODUCTION method. The benefit of integrating a fuzzy system with GP is

that, compared with crisp sets and logic, the fuzzy sets and
The realization of uncertainty, vagueness, and ambiguity in logic can provide a more loose and flexible search space. This
the world has led to the concept of fuzziness. In the last 3 benefit may improve the search efficiency of GP because va-
decades, significant progress has been made in the develop- riety is GP’s innate nature. Mathematically, constraints in en-
ment of fuzzy sets and fuzzy logic theory and their use in a gineering design problems are usually expressed in inequalities
large variety of applied topics in engineering and natural and that divide the solution space into satisfied and unsatisfied
socioeconomic sciences. The successful applications of fuzzy regions. Although a feasible solution point should fall inside
sets and fuzzy logic can be attributed to the fact that the fuzzy the limit of satisfaction, there can be a significant difference
theory reflects the true situation of the real world, where hu- between the near and far solution points from the region
man thinking is dominated by approximate reasoning logic. boundary. According to the traditional crisp set theory, some
For example, during the design stage, decision making is candidate solutions will be thrown out if they fall outside the
mainly based on conceptual understanding. Unfortunately, feasible region (i.e., violate one or more constraints) during
such information is usually vaguely defined by the experts. the early optimization evolutionary process. Consequently, the
Therefore, consideration of imprecise and vague information important information that could eventually lead to the optimal
becomes an important issue in an automated design environ- solution may also be thrown out. However, fuzzy description
ment. Moreover, structural optimization and design, in a way, of design constraints provides GP with a way of approaching
are constraint satisfaction problems; for example, when trying the optimal solution from the infeasible region to overcome
to determine the best frame design for a building, one may be this problem.
constrained by the design codes as well as by the design spec- In the applications of genetic algorithms (GAs) to the struc-
ifications such as functional, architectural, and structural be- tural optimization problems, the most commonly used method
havior requirements. Once these constraints are satisfied, the is to represent the constraints by using penalty functions
remaining problem may be to find a design that requires the (Goldberg 1989; Jenkins 1992; Hajela and Lee 1995; Shrestha
least construction material. In the traditional optimization tech- and Ghaboussi 1998). In an earlier work (Soh and Yang 2000),
niques, these constraints must be strictly satisfied. However, it a penalty method also was adopted in the GP based optimi-
is not reasonable to discard those designs that slightly violate zation approach. Experience shows that such a method has its
one or more constraints during the early design stage. These limitations in coping with imprecise and vague information,
complexities and uncertainties encountered in the optimization resulting in expensive computation during the optimization
and design of real structures provide the main motivation for process.
this work examining a fuzzy integrated system. In the last 2 decades, fuzzy theory has been applied to struc-
Genetic programming (GP) is a technique devised for the tural optimization. Several researchers (Wang and Wang 1985;
automatic generation of computer programs by means of nat- Rao 1987; Dhingra et al. 1990; Rao et al. 1990) focused on
ural selection (Koza 1992). It has been demonstrated to be optimization in a fuzzy environment. The approaches adopted
very effective at finding optimal solutions to a variety of prob- by these researchers were all based on the fuzzy mathematical
lems. This hierarchical search and optimization method has programming techniques, and the objects to be optimized were
been introduced into civil engineering problem solving by Soh only sizing optimization problems. Their results show that the
and Yang (2000). However, the characteristics that make GP use of a fuzzy set can handle the decision-making problems
robust also make it computationally intensive, requiring high with imprecise data and provide valuable information in de-
computing cost. To overcome this weakness, this study inte- ciding a promising design. Lee and Takagi (1993) proposed a
grates a fuzzy logic based decision-making system, which in- method for controlling GA using fuzzy logic techniques that
corporates expert knowledge and experience, with the pure GP were based on experience or intuition. Their results showed a
1
better performance over a simple GA in the classic inverted
Res. Scholar, Div. of Struct. and Mech., School of Civ. and Struct. pendulum control problem. Soh and Yang (1996) proposed a
Engrg., Nanyang Technol. Univ., Nanyang Ave., Singapore 639798.
2
Prof., Div. of Struct. and Mech., School of Civ. and Struct. Engrg., fuzzy-rule based GA for shape optimization of truss structures.
Nanyang Technol. Univ., Nanyang Ave., Singapore 639798. E-mail: The results obtained showed that the method costs less com-
[email protected] putation time in comparison with the pure GA approach.
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, Therefore, in this paper, based on the earlier work (Soh and
2001. To extend the closing date one month, a written request must be Yang 2000, unpublished manuscript, 2000) of using a GP
filed with the ASCE Manager of Journals. The manuscript for this paper
based approach for structural optimization and design, a fuzzy
was submitted for review and possible publication on January 21, 2000.
This paper is part of the Journal of Computing in Civil Engineering, described constraint model has been introduced into the pure
Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/0004-0249– GP to increase the GP’s performance in dealing with imprecise
0254/$8.00 ⫹ $.50 per page. Paper No. 22096. information. A fuzzy logic based decision-making system (i.e.,
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 249
a fuzzy logic controller) is integrated with the GP search point. Soh and Yang (unpublished manuscript, 2000) presented
method for structural optimization and design. To deal with a GP based methodology for this optimum design problem and
the impreciseness and vagueness in structural design and to showed that GP is more flexible and has a higher search ef-
achieve higher search efficiency, the design constraints such ficiency than GA in this situation.
as member cross-sectional area, member stress, joint displace-
ment, and member buckling are described by using fuzzy set FUZZY LOGIC CONTROLLED GP
theory. Obtained illustrative examples have shown that the
proposed fuzzy logic controlled GP is more efficient than the GP
pure GP method in solving structural optimization and design
problems. GP can be viewed as an extension of GA. It is a technique
for the automatic generation of computer programs by means
of natural selection (Koza 1992). The GP process starts by
PROBLEM DESCRIPTION OF STRUCTURAL
creating a large initial population of programs that are random
OPTIMIZATION AND OPTIMUM DESIGN
combinations of elements from the problem-specific function
The process of simultaneous sizing, geometry, and topology sets and terminal sets. The programs are usually denoted as
optimization can be stated as follows. GP parse trees. Each program in the initial population is then
Find the particular set of sizing variables A, geometric var- assessed for its fitness. This is usually accomplished by run-
iables R, and topological variable T such that structural weight ning each program on a collection of input called fitness cases
W(R, A, T) and by assigning numerical fitness values to the output of each
of these runs. The resulting values are then combined to pro-
W(R, A,T) → min (1) duce a single fitness value for the program.
The fitness values are used in producing the next generation
subject to of programs via a variety of genetic operations including re-
G Lij ⱕ Gij(R, A, T) ⱕ G Uij , i = 1, 2, . . . , nj, j = 1, 2, . . . , nc production, crossover, and mutation. The programs are ran-
(2) domly selected for participation in these operations, but the
selection function is biased toward the highly fit programs.
in which A and R = vectors of member cross-sectional areas The reproduction operator simply selects an individual and
and joint coordinates, respectively; T represents the set of ex- copies it into the next generation. The crossover operation in-
isting members; L and U = superscripts denoting the lower troduces variation by selecting two parents and by generating
and upper bounds, respectively; Gij(R, A, T) = constraints that from them two offspring; the offspring are produced by swap-
limit the relevant design variable domains; nc = total number ping random subtrees of the parents. The mutation operator
of constraint types; and nj = total number of constraints of the produces one offspring from a single parent by replacing a
jth constraint type. The constraint types may include the con- randomly selected subtree with a newly randomly generated
straints of (1) member cross-sectional areas; (2) joint coordi- subtree.
nates; (3) member allowable stresses; (4) joint allowable dis- Over many generations of fitness assessment, reproduction,
placements; (5) member buckling strength; (6) member length; crossover, and mutation, the average fitness of the population
and (7) member slenderness. may tend to improve, as may the fitness of each best-of-gen-
The optimum design problem described here refers to the eration individual. After a predetermined number of genera-
problem of finding an optimal structure within a predefined tions or after the fitness improves to a certain preset level, the
design domain, satisfying the design constraints and loading best-of-run individual is designated as the final result from the
and supporting conditions. The design domain is a specified GP system. More details on GP can be found in Koza (1992).
physical space that may have any arbitrary shape enclosing a Empirically shown to be a powerful program-induction
multiconnection area (including ‘‘void’’) (Fig. 1). The support methodology, GP has been successfully applied to a large
and the load of the structure may be at the boundary of the number of difficult problems such as automatic design, pattern
design domain or inside the design domain. The structure can recognition, robotic control, synthesis on neural architectures,
be evolved within the design domain and can acquire any to- symbolic regression, factory job scheduling, electronic circuit
pology and shape as long as it is confined within the design design, signal processing, and music and picture generation.
domain. The structure can have any number of nodes and any Recent works are described in Kinnear (1994), Angeline and
number of elements. Some nodes may be fixed to satisfy the Kinnear (1996), and Koza et al. (1996, 1997, 1998).
requirements of loading and supporting, and the others can
occupy any position within the design domain. The objective Encoding Approach
of optimum design is to minimize the weight of the structure.
The formulation of this problem also can be expressed as The key issue in the application of GP to structural design
(1) and (2). The difference between the optimization problem is to establish a mapping scheme between the kind of point-
and the optimum design problem is because there is a pre- labeled parse trees found in the world of GP and the node-
specified start point, such as ground structure and conceptual element–labeled diagrams employed in the analysis of struc-
design, for the optimization problem during the search process,
whereas the optimum design problem is free from any start

FIG. 2. Truss and Its GP Parse Tree: (a) Six-Bar Truss; (b) GP
FIG. 1. General Design Domain Parse Tree of Truss

250 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

tures. Choosing an appropriate representation of candidate W(R, A, T) → min (3)
solutions to the problem at hand is the foundation for applying
GP to solve real world problems, which conditions all the sub- subject to
sequent steps of GP. The encoding method adopted for this ⬃ ⬃ ˜U
˜ Lij ⫹ G
G ˜ Le ˜ Ue
ij ⱕ Gij ⱕ G ij ⫹ G ij (4)
study can be illustrated by the example shown in Fig. 2. Fig.
2(b) is the GP parse tree of the truss structure shown in Fig.
2(a). For more details of the encoding scheme, refer to Soh Fuzzy Logic Controller
and Yang (2000).
A fuzzy logic controller (FLC) is a rule based system that
incorporates the flexibility of human decision making by
Fuzzy Set and Logic Theory means of the use of the fuzzy set theory. The fuzzy rules of
FLC incorporate fuzzy linguistic terms (e.g., temperature is
The classical or crisp sets and mathematical logic divide the
warm) described by membership functions. These functions
world into yes and no, white and black, and true and false.
are intended to represent a human expert’s conception of the
On the other hand, fuzzy sets deal with objects that are a
linguistic terms, thus giving an approximation of the confi-
‘‘matter of degree,’’ with all possible grades of truth between
dence with which a precise numeric value is described by a
yes and no and the various shades of gray between white and
linguistic label.
black.
Fuzzy rules take the form IF (conditions) THEN (actions),
Fuzzy logic is an approximate reasoning method for coping
where conditions and actions are linguistic labels applied to
with life’s uncertainties. Occasionally, the characteristics of
input and output variables, respectively (e.g., IF temperature
various systems are difficult to describe with mathematical
is hot THEN refrigeration is high) (Kruse et al. 1994). In gen-
equations because of their complexity. In these cases, human
eral, most of these specifications in codes and the functional
experts may achieve control using knowledge or values that
requirements set by the users must be given in natural lan-
are crystallized from their life-long experiences and repre-
guage to describe the experts’ empirical knowledge of design
sented by intuitive natural language. In solving structural de-
modifications.
sign problems, loads, material behavior, and system properties
For instance, ␴ Li ⱕ ␴i ⱕ ␴ Ui in member stress constraints
may be linguistically specified. For example, the loading act-
can be replaced by ␴˜ Li ⫹ ␴˜ Le ⬃ ⬃ ˜ U ⫹ ␴˜ Ue with a relevant
i ⱕ ␴i ⱕ ␴ i i
␣ membership degree. Here, the symbol ⬃
ing on the structure may be described with linguistic variables
ⱕ means fuzzy var-
such as ‘‘severe,’’ ‘‘heavy,’’ or ‘‘light.’’ Furthermore, the mem-
iable operator and ␣ represents a series of linguistic variables
ber’s strength may be described using such qualitative terms
that mean ‘‘very small,’’ ‘‘small,’’ ‘‘medium,’’ ‘‘large,’’ ‘‘very
as ‘‘very stiff’’ or ‘‘highly flexible.’’ These linguistic terms
large,’’ and so forth. It gives a set of the intersection of the
may be defined by fuzzy sets.
transition buffer, upon which one can evaluate a constraint and
The fuzzy set theory (Zadeh 1987) has a function that ad-
control the constraint satisfaction problem. According to the
mits a degree of membership in the set from complete exclu-
interpretation of the fuzzy linguistic terms (Zadeh 1987), the
sion (0) to absolute inclusion (1). Such a function also is called
following seven fuzzy variables were used in this study: neg-
the membership function ␮A˜ ( y) of object y in the fuzzy set
˜ ␮A˜ ( y): R n → [0, 1]. The membership represents a certain ative large (NL), negative medium (NM), negative small (NS),
A:
zero (ZE), positive small (PS), positive medium (PM), and
degree of belonging of the object in the fuzzy set. The tran-
positive large (PL). They are defined by the membership func-
sition from belonging to not belonging is gradual, which gives
tion, as shown in Fig. 4. For the convenience of implemen-
one some means of handling vagueness. Fuzzy sets thus over-
tation, these seven fuzzy variables are assigned seven integer
come a major weakness of crisp sets; fuzzy sets do not have
reference numbers, that is, ⫺3, ⫺2, ⫺1, 0, 1, 2, and 3, re-
an arbitrarily established boundary to separate the members
spectively.
from the nonmembers.
A set of such fuzzy rules constitutes the fuzzy rule base of
An inequality constraint Gij in (2) for optimum design prob-
FLC. The system uses this rule base to produce precise output
lems can be defined to be a fuzzy constraint G̃ij with an ␣
membership degree, as shown in Fig. 3. Here G ˜ Lij and G
˜ Uij are,
respectively, the lower and upper bounds of the corresponding
object Gij, and G ˜ Le ˜ Ue are the extended bounds, called
ij and G ij
the ‘‘tolerances.’’ The tilde is used to denote that the con-
straints contain fuzzy information. When both G ˜ Le ˜ Ue
ij and G ij are
equal to zero, this fuzzy constraint becomes a ‘‘hard’’ con-
straint. At the same time, the value ␣ also gives us a descrip-
tion about how the constraint is satisfied. If ␣ = 1, the con-
straint Gij is fully satisfied. It is identical with the usual
constraint satisfaction. If ␣ = 0, the Gj is not satisfied. Fur- FIG. 4. Membership Function of Fuzzy Variables
thermore, its value between 0 and 1 implies the constraint is
satisfied to the relevant degree.
Based on the fuzzy set theory introduced above, (1) and (2)
can be transformed into the following optimization problem:

FIG. 3. Fuzzy Constraint G̃ij with ␣ Membership Degree FIG. 5. Framework of Fuzzy Logic Controller

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 251

values according to the input values. As shown in Fig. 5, this These heuristic rules represent the expert knowledge and
control process is divided into three stages: experience in the optimum design problems. They can be mod-
ified easily if necessary.
• Fuzzification: to calculate the fuzzy input (i.e., to evalu-
ate the input variables with respect to the corresponding FLC GP Based Structural Optimization and Design
linguistic terms in the condition side) Procedure
• Fuzzy inference: to calculate the fuzzy output (i.e., to
evaluate the activation strength of every rule base and Solutions to the optimization problems described by (3) and
combine their action sides) (4), according to the GP evolutionary process, can be sought
• Defuzzification: to calculate the actual output (i.e., to by (1) defining a coding scheme between the structures and
convert the fuzzy output into a precise numerical value) their GP parse trees; (2) producing the initial population; (3)
performing the computation with the genetic operators; and
The objective of fuzzification is, depending on the mem- (4) evaluating the fitness function with respect to the object
bership function shown in Fig. 4, to transform the input from function. Here, the highlight is on the description of the GP
a numerical value, such as the maximum of violated member based optimization search procedure. The fundamental mech-
stresses, into a set of fuzzy variables. In contrast to fuzzifi- anism of GP can be referred to in Koza (1992) and Banzhaf
cation, defuzzification performs the transformation from a et al. (1998).
fuzzy variable to a quantitative value. According to the fuzzy Fig. 6 shows the detailed procedure of the proposed GP
rules, fuzzy inference evaluates the fuzzy variables in the con- based approach with FLC for structural optimization and de-
ditions and then takes the corresponding actions. sign problems. Unlike with pure GP, the proposed fuzzy logic
The database provides the membership function of fuzzy controlled GP search approach introduces FLC for the local
sets defined for the upper and lower bounds on the constraints modification of the design variables, which is of great impor-
Gij. These functions show the degrees of membership of a set tance to its search efficiency. The role of FLC comes after
of fuzzy variables with respect to their fuzzy domains. FEM analysis is carried out and the structural performances
In the fuzzy rule base, both the input in the conditions and are evaluated. Based on the results of these evaluations, FLC
the output in the actions are described by seven fuzzy varia- is activated if necessary. If it is activated, each rule in FLC
bles. For example, as an input, the constraint Gij is usually will be sought to make a decision about whether certain design
classified into three categories: (1) active for ZE; (2) inactive variables need to be modified.
for NL, NM, and NS; and (3) violated for PS, PM, and PL. The FLC can help GP in three ways. First, it can inject
On the other hand, as an output, the modification of the mem- expert experience in the fuzzy rule base of FLC so that the
ber cross-sectional areas has the following possibilities: NL, search can reach the optimum solution regions quickly. Sec-
NM, NS, ZE, PS, PM, and PL. Two samples of the heuristic ond, by giving some direct guidance to the search process,
rules in FLC are they also can substantially reduce the computing time. Last,
due to the use of modification based on the predefined prob-
Rule 1: IF the maximum of violated member stress con- ability and fuzzy representation of the design constraints, FLC
straints is PS and All displacement constraints may reduce the risk of a premature problem solution caused
are inactive by improper rule interference and also avoid the loss of vague
THEN the change of the corresponding member information that may subsequently be useful.
cross-sectional area is PS.
Rule 2: IF all constraints are inactive and the minimum of ILLUSTRATIVE EXAMPLES
member stress constraints is NL
THEN the change of the corresponding member Two illustrative examples, including one structural optimi-
cross-sectional area is NS. zation problem and one optimum design problem, are pre-

FIG. 6. Flowchart of Fuzzy Logic Controlled GP Based Approach

252 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

sented here to demonstrate the search efficiency of the pro- The bounds on the joint coordinates are as follows: 10.16–
posed fuzzy logic controlled GP method for structural 20.32 m (400–800 in.) for x1 and 5.08–12.7 m (200–500 in.)
optimization and design problem. Both examples have been for x3, y1, and y3, respectively. The following two cases were
previously investigated using the pure GP based approach (Soh investigated: Case 1, only stress constraints, and Case 2, both
and Yang 2000, unpublished manuscript, 2000). stress and displacement constraints.
To compare the search efficiency of the proposed fuzzy
Example 1: 10-Member Planar Truss logic controlled GP with that of the pure GP, the control pa-
rameters for the GP run is set the same as Soh and Yang
Fig. 7 shows the initial shape and topology of a 10-member (2000), that is, the population size is 500, the maximum gen-
planar truss. The objective is to minimize its weight with var- eration of runs is 50, and the probabilities of reproduction,
ying shape and topology. The design data are as follows: mod- crossover, and mutation operation are 0.1, 0.8, and 0.1, re-
ulus of elasticity is 206.7 GPa (30,000 ksi), material density spectively.
is 2.768 g/cm3 (0.1 pci), the allowable tensile and compressive For both Cases 1 and 2, the final configuration of the truss
stresses are 172.25 MPa (25 ksi), and the allowable displace- is the same, as shown in Fig. 8. Comparisons of results for
ment at Joint 2 is 5.08 cm (2.0 in.). The bound on the member the two cases are presented in Table 1, where the ‘‘genera-
cross-sectional areas is 0.452–64.516 cm2 (0.07–10.0 in.2). tions’’ is the number of GP runs when the final optimal solu-
tion is obtained. It is apparent that the proposed fuzzy logic
controlled GP approach resulted in a lesser number of itera-
tions than the pure GP method; i.e., the iteration steps are
reduced by 35.5% for Case 1 and 31.6% for Case 2. Besides,
the optimum weights obtained by the fuzzy logic controlled
GP are also slightly lower.

Example 2: Optimum Design within Given Design

Domain

In this example, one will find an optimum truss structure

within the design domain given in Fig. 9, in which the loads
also are specified. All the truss members are selected from a
FIG. 7. Ten-Member Planar Truss set of 30 standard AISC sections (Manual 1989) (i.e., W14 ⫻
22 through W14 ⫻ 426). The material properties are those of
steel (E = 201 GPa, fy = 248.8 MPa, and ␳ = 7851.03 kg/m3).
The relevant AISC design specifications are followed: the al-
lowable tension stress is 0.6fy; the allowable member slender-
ness ratio is 300 for tension members and 200 for compression
members; the length of members is required to be not shorter
than 5 m and not longer than 35 m; the allowable joint dis-
placement is limited to 1/1000 of the span (i.e., 70 mm); and
the allowable compression stress ␴ bi of member i is determined
from the buckling consideration. The coordinates of the joints
FIG. 8. Final Configuration of 10-Member Truss are assumed to be discrete and may be randomly selected in
the design domain with steps of 0.5 m.
TABLE 1. Optimal Results of 10-Bar Truss Again, the control parameters for the GP run are the same
Case 1 Case 2 as those in Soh and Yang (unpublished manuscript, 2000); i.e.,
Soh and Soh and
Yang Present Yang Present
Parameter (2000) work (2000) work
(1) (2) (3) (4) (5)
(a) Member Cross-Sectional Area Ai (cm2)
A1 52.00 52.06 64.39 64.19
A2 — — — —
A3 53.29 53.29 52.26 52.58
A4 27.48 27.48 30.45 29.68
A5 — — — —
A6 — — — —
A7 36.52 36.52 36.52 36.84
A8 31.03 30.71 40.58 40.19
A9 37.68 37.74 43.68 43.87
A10 — — — — FIG. 9. Design Domain of Example 2
(b) Joint Coordinates (xj, yj) (m)
x1 — — — —
y1 — — — —
x3 9.855 9.906 9.855 9.982
y3 7.925 7.899 8.230 8.103
Weight (N) 6,932.16 6,927.57 7,870.23 7,862.5
Generations 32 21 38 26
Reduction in itera-
tion (%) — 35.5 — 31.6
FIG. 10. Optimum Truss for Example 2

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 253

 TABLE 2. Optimal Results for Example 2 CONCLUSIONS
Soh and Yang This paper presents a fuzzy logic controlled GP methodol-
Shrestha and (unpublished ogy for structural optimization and design. The imprecise and
Ghaboussi manuscript, vague information in the structural design process, especially
Parameter (1998) 2000) Present work the design constraints, is dealt with using fuzzy set theory. The
(1) (2) (3) (4) expression of constraints in fuzzy sets allows GP to access the
(a) Member Sections optimal solution from both the feasible and the infeasible re-
1-2 NA W14 ⫻ 68 W14 ⫻ 74 gion in the search space and thus increases the opportunity
2-3 NA W14 ⫻ 109 W14 ⫻ 109 and the speed in obtaining the optimal solution.
3-4 NA W14 ⫻ 132 W14 ⫻ 132 The proposed approach is able to integrate expert knowl-
4-9 NA W14 ⫻ 233 W14 ⫻ 211 edge and experience with the GP search procedure by using a
1-5 NA W14 ⫻ 132 W14 ⫻ 132 fuzzy logic controller. This artificial intelligence based simu-
5-6 NA W14 ⫻ 132 W14 ⫻ 132
6-7 NA W14 ⫻ 176 W14 ⫻ 176 lation can guide the GP search in a more intelligent way, hence
7-8 NA W14 ⫻ 193 W14 ⫻ 211 increasing its efficiency. Illustrative examples show that the
2-5 NA W14 ⫻ 61 W14 ⫻ 61 proposed method can reduce the required computational time
2-6 NA W14 ⫻ 74 W14 ⫻ 61 and enhance the search efficiency of pure GP.
3-6 NA W14 ⫻ 82 W14 ⫻ 74
4-7 NA W14 ⫻ 90 W14 ⫻ 82 APPENDIX. REFERENCES
4-8 NA W14 ⫻ 61 W14 ⫻ 61
Angeline, P. J., and Kinnear, K. E., Jr., eds. (1996). Advances in genetic
(b) Joint Cordinates (m) programming II, MIT Press, Cambridge, Mass.
x5 NA 4.0 3.5 Banzhaf, W., Nordin, P., Keller, R. E., and Francone, F. D. (1998). Ge-
y5 NA 5.5 6.0 netic programming: An introduction: On the automatic evolution of
x6 NA 13.0 12.5 computer programs and its applications, Morgan Kaufmann Publishers
y6 NA 9.0 9.0 Inc., San Francisco.
x7 NA 24.0 23.5 Dhingra, A. K., Rao, S. S., and Kumar, V. (1990). ‘‘Nonlinear member-
y7 NA 10.0 10.0 ship functions in multiobjective fuzzy optimization of mechanical and
x8 NA 35.0 35.0 structural systems.’’ AIAA J., 30(1), 251–260.
y8 NA 10.0 10.0 Goldberg, D. E. (1989). Genetic algorithms in search, optimization and
machine learning, Addison-Wesley, Reading, Mass.
Weight (kg) 60,329 45,404 45,163 Hajela, P., and Lee, E. (1995). ‘‘Genetic algorithms in truss topological
Population size 100 2,000 2,000 optimization.’’ Int. J. Solids and Struct., 32(22), 3341–3357.
Generations 9,754 83 50 Jenkins, W. M. (1992). ‘‘Plane frame optimum design environment based
Total iterations 975,400 166,000 100,000 on genetic algorithm.’’ J. Struct. Engrg., ASCE, 118(11), 3103–3112.
Note: NA denotes ‘‘not available.’’ Kinnear, K. E., Jr., ed. (1994). Advances in genetic programming, MIT
Press, Cambridge, Mass.
Koza, J. R. (1992). Genetic programming: On the programming of com-
puters by means of natural selection, MIT Press, Cambridge, Mass.
Koza, J. R., et al., eds. (1997). Proc., 2nd Annu. Conf., Genetic Pro-
gramming 1997, Morgan Kaufmann Publishers, Inc., San Francisco.
the population size is 2,000, the maximum generation of runs Koza, J. R., et al., eds. (1998). Proc., 3rd Annu. Conf., Genetic Program-
is 100, and the probabilities of reproduction, crossover, and ming 1998, Morgan Kaufmann Publishers Inc., San Francisco.
mutation are 0.1, 0.8 and 0.1, respectively. Koza, J. R., Goldberg, D. E., Fogel, D. B., and Riolo, R. L., eds. (1996).
This problem also has been previously investigated by Proc., 1st Annu. Conf., Genetic Programming 1996, MIT Press, Cam-
Shrestha and Ghaboussi (1998) using a GA based methodol- bridge, Mass.
Kruse, R., Gebhardt, J., and Klawonn, F. (1994). Foundations of fuzzy
ogy. In their investigation, the population size is set at 100 systems, Wiley, Chichester, U.K.
with relatively long chromosomes of 25,200-bit string em- Lee, M. A., and Takagi, H. (1993). ‘‘Dynamic control of genetic algo-
ployed. The best design of the truss is obtained after a run of rithms using fuzzy logic techniques.’’ Proc., 5th Int. Conf. on Genetic
9,754 generations and has an optimum weight of 60,329 kg. Algorithm, Morgan Kaufmann Publishers Inc., San Francisco, 76–83.
The final optimum truss found by the fuzzy logic controlled Manual of steel construction: Allowable stress design. (1989). 9th Ed.,
GP approach is shown in Fig. 10. Table 2 compares the results American Institute of Steel Construction, Chicago.
Rao, S. S. (1987). ‘‘Optimum design of structures in a fuzzy environ-
obtained by this study using fuzzy logic controlled GP, by ment.’’ AIAA J., 25(12), 1633–1636.
Shrestha and Ghaboussi (1998) using GA, and by Soh and Rao, S. S., Sundararaju, K., Prakash, B. G., and Balakrishna, C. (1990).
Yang (unpublished manuscript, 2000) using pure GP. The total ‘‘Fuzzy goal programming approach for structural optimization.’’ AIAA
iterations in Table 2 is defined as the product of the population J., 30(5), 1425–1432.
size and the generations of GA/GP run when the optimal so- Shrestha, S. M., and Ghaboussi, J. (1998). ‘‘Evolution of optimum shapes
using genetic algorithm.’’ J. Struct. Engrg., ASCE, 124(11), 1331–
lution is obtained. It shows that the fuzzy logic controlled GP 1338.
approach reduced the iteration steps by (166,000 ⫺ 100,000)/ Soh, C. K., and Yang, J. P. (1996). ‘‘Fuzzy controlled genetic algorithm
166,000 = 39.8%, compared with the pure GP. It is worth search for shape optimization.’’ J. Comp. in Civ. Engrg., ASCE, 10(2),
noting that the iteration steps of both the pure GP and the 143–150.
fuzzy GP are much less than that of the GA method; i.e., they Soh, C. K., and Yang, Y. W. (2000). ‘‘Genetic programming based ap-
are only 166,000/975,400 = 17% and 100,000/975,400 = proach for structural optimization.’’ J. Comp. in Civ. Engrg., ASCE,
14(1), 31–37.
10.3%, respectively, of the iteration steps of the GA method. Wang, G. Y., and Wang, W. Q. (1985). ‘‘Fuzzy optimum design of struc-
The optimal weights obtained by the GP based approach are tures.’’ Engrg. Optimization, Reading, U.K., 8(4), 291–300.
also much lighter than those obtained by the GA approach. Zadeh, L. A. (1987). Fuzzy sets and applications, Wiley, New York.

254 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

 MULTIRESERVOIR SYSTEMS OPTIMIZATION USING GENETIC
ALGORITHMS: CASE STUDY
By Mohd Sharif1 and Robin Wardlaw2

ABSTRACT: A genetic algorithm approach is presented for the optimization of multireservoir systems. The
approach is demonstrated through application to a reservoir system in Indonesia by considering the existing
development situation in the basin and two future water resource development scenarios. A generic genetic
algorithm model for the optimization of reservoir systems has been developed that is easily transportable to any
reservoir system. This generality is a distinct practical advantage of the genetic algorithm approach. A comparison
of the genetic algorithm results with those produced by discrete differential dynamic programming is also
presented. For each case considered in this study, the genetic algorithm results are very close to the optimum,
and the technique appears to be robust. Contrary to methods based on dynamic programming, discretization of
state variables is not required. Further, there is no requirement for trial state trajectories to initiate the search
using a genetic algorithm. Model sensitivity and generalizations that can be drawn from this and earlier work
by Wardlaw and Sharif are also considered.

INTRODUCTION gt (st , st⫹1) = rt

Dynamic programming (DP) (Bellman 1957; Bellman and where rt is the release in time step t such that st⫹1 = ft (st , rt);
Dreyfus 1962) has long been recognized as a powerful ap- and t is the time step.
proach in the analysis of water resource systems. The major Another approach that overcomes the dimensionality prob-
attraction of DP is that it can handle nonconvex, nonlinear, lem is the DP successive approximation technique (DPSA).
and discontinuous objective functions without difficulty. Con- Larson (1968) first proposed the technique, and Trott and Yeh
straints on both decision and state variables are easily intro- (1973) used the successive approximation technique in com-
duced. Hall and Buras (1961) was the first to propose the bination with IDP in solution of a six-reservoir problem. Suc-
application of DP to determine optimal returns from reservoir cessive approximation is one-at-a-time optimization, and its
systems. Young (1967) developed optimal operating rules for common drawback is convergence to a local optimum. Exten-
a single reservoir using DP. Extensive review of DP applica- sion to the DPSA technique has been reported by Nopmongcol
tions to reservoir systems can be found in the works of Ya- and Askew (1976), who suggested higher-level combinations,
kowitz (1982) and Yeh (1985). Wurbs (1993) describes various such as two- or three-at-a-time combinations. The technique
computational models that can be used in the analysis of water was demonstrated by application to the four-reservoir problem.
resource systems. The usefulness of DP for multireservoir sys- The requirement that the state variables be discretized is a
tems is, however, limited by the huge demand that it can in- major cause of computational burden and huge memory re-
duce on computational resources. A brief review of optimi- quirements. Murray and Yakowitz (1979) have developed dif-
zation models used in reservoir operation problems is ferential DP for multireservoir control problems. The impor-
presented here to put the work of this paper in context. tant feature of the technique is that discretization of state and
Many variants of DP have been developed to alleviate the decision space is not required. The method requires a quadratic
problem of dimensionality. Larson (1968) obtained the solu- approximation of the objective function. Application of the
tion to the four-reservoir problem using incremental dynamic method to a ten-reservoir problem by Murray and Yakowitz
programming (IDP). Hall et al. (1969) used a different version (1979) has demonstrated its effectiveness. However, the tech-
of IDP and solved a two-reservoir system. Another version, nique requires that the objective function is differentiable and
called discrete differential dynamic programming (DDDP), that the constraints are linear.
was developed by Heidari et al. (1971) and could be seen as A combined LP-DP procedure has been described by Becker
a generalization of IDP. The terms IDP and DDDP have been and Yeh (1974). The procedure has also been used by Takeuchi
used interchangeably in water resources applications. In the and Moreau (1974), Becker et al. (1976), Yeh et al. (1979),
standard IDP of Larson (1968), the number of states to be Yeh and Becker (1982), and Marino and Mohammadi (1983).
tested at each stage is limited to three per state variable. The In a combined LP-DP algorithm, the stage to stage optimiza-
computational burden and memory requirements for such an tion is carried out by LP, and DP is used for determining the
approach are functions of 3n, where n is the number of state optimal policy over a number of periods. The nonlinearities
variables. IDP does overcome the dimensionality problem to are handled by using an iterative technique, such as successive
a large extent, but requires stringent conditions to be satisfied linear programming (SLP). Grygier and Stedinger (1985) de-
for achieving an optimal solution. The IDP method requires scribe the application of SLP, an optimal control algorithm,
that there must exist a function gt (st , st⫹1) such that for every and a combined LP-DP algorithm to a multireservoir system.
pair of states st and st⫹1 The optimal control algorithm executed five times faster than
the SLP, while the LP-DP algorithm took longer to execute
1
Lect., Dept. of Civ. Engrg., Facu. of Engrg. and Technol., Jamia Millia
and produced comparatively inferior solutions. Howson and
Islamia, New Delhi, India. Sancho (1975) developed a progressive optimality algorithm
2
Sr. Lect., Dept. of Civ. and Envir. Engrg., Univ. of Edinburgh, Ed- (POA) for multistate problems. Application of the POA to an
inburgh EH9 3JN, U.K. E-mail: [email protected] example system consisting of four reservoirs in series has been
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, reported by Turgeon (1981). The most promising aspect of the
2001. To extend the closing date one month, a written request must be approach is that the state variables do not have to be discre-
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on October 5, 1999.
tized. The algorithm, however, required 823 iterations to con-
This paper is part of the Journal of Computing in Civil Engineering, verge to an optimal solution for the example considered. The
Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/0004-0255– approach has merit in that it overcomes the dimensionality
0263/$8.00 ⫹ $.50 per page. Paper No. 22076. problem.
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 255
 Optimal control theory (OCT) also provides an efficient techniques require large computer resources and derivatives of
means of overcoming the dimensionality problem. State and the objective functions. GAs work with numerical values, and
release variables need not be discretized in methods based on there is no requirement for the objective function to be a dif-
OCT. The requirement that the objective function must be dif- ferentiable function. Furthermore, the computing requirements
ferentiable limits the application of the approach to some ex- are minimal and increase only linearly with the dimensionality
tent, however. Storage constraints are not automatically satis- of the system. There are some limitations to the approach,
fied, but can be handled by use of appropriate penalty however. The convergence to an optimum cannot be guaran-
functions. Wasimi and Kitanidis (1983), Papageorgiou (1985), teed. With increasing string length, the likelihood of locating
Georgakakos and Marks (1987), Georgakakos (1989), and the optimum is reduced. Also, with increasing string length
McLaughlin and Velasco (1990) all describe the application of there are difficulties in satisfying the constraints of the prob-
OCT to reservoir optimization problems. lem. This problem of constraint satisfaction is not limited to
In recent years, genetic algorithms (GAs) have developed the GA approach. Other iterative optimization techniques, such
into a powerful optimization approach. They have been ap- as OCT approaches, also face the same problem. The dimen-
plied to a wide spectrum of problems in science and engi- sionality problem associated with DP (Bellman 1957) and the
neering and have proven to be an efficient means by identi- requirements of differentiability of objective functions for gra-
fying optimal solutions. In some cases, GAs have performed dient-based methods make GA an alternative worthy of ex-
better than traditional techniques. For example, Dandy et al. ploring.
(1996) obtained a better solution for the New York City water GAs may be set up in a number of ways, but as yet there
distribution problem using a GA than had been obtained by is no standard procedure. Wardlaw and Sharif (1999) evaluated
any other technique in the past. GAs are being increasingly alternative GA formulation procedures for multireservoir sys-
used in a number of disciplines of civil engineering to solve tems and demonstrated the robustness of the GA approach by
problems that are difficult with traditional techniques. Kou- applying it to a number of problems. The primary objective
mousis and Georgiou (1994) appled a GA for optimization of of this paper is to demonstrate the applicability of the GA
steel truss roofs. Soh and Yang (1996) used GAs in combi- approach for optimization of a real multireservoir system with
nation with fuzzy logic for structural shape optimization prob- a complex objective function. The results produced by a GA
lems. Rajeev and Krishnamoorthy (1992, 1997), Jenkins have been compared to those obtained by the DDDP tech-
(1991, 1992), and Adeli and Cheng (1993, 1994) have also nique.
reported applications of GAs to structural engineering design
problems. GENETIC ALGORITHMS
A methodology based on GAs has been developed by Li In a GA, members of a population of abstract representa-
and Love (1998) for optimizing the layout of construction site tions of candidate solutions to an optimization problem are
level facilities. Navon and McRea (1997) used the GA ap- stochastically selected, recombined, mutated, and then either
proach for selection of construction robots. Feng et al. (1997) eliminated or retained, based on their relative fitness. The GA
appled a GA to the problem of cost-time trade-offs in con- technique has been notably developed by Goldberg (1989) and
struction projects. Applications of GA to transportation engi- Michalewicz (1992), who also give an excellent introduction
neering problems have been described by Ng (1995), Fwa et to the subject.
al. (1996), Liu et al. (1997), and Cheu et al. (1998). The first step in GA formulation is to code the components
GAs have also been used in combination with other tech- of possible solutions in a chromosome. Each chromosome rep-
niques. Soh and Yang (1996) used a fuzzy controlled GA for resents a potential solution consisting of the components of
shape optimization of trusses. Franchini (1996) used a GA in the decision variables (also known as genes) that either form
combination with sequential quadratic programming for the or can be used to evaluate the objective function. The entire
automatic calibration of conceptual rainfall-runoff models. population of such chromosomes represents a generation. For
Many application of hybrid GAs have also been reported by a typical reservoir operation problem, the genes consist of res-
Davis (1991). Dasgupta and Michalewicz (1997) describe var- ervoir releases that are concatenated to form chromosomes.
ious applications of GAs to engineering optimization prob- For the purpose of exposition, suppose there is only a single
lems. reservoir in the system, and that there are 12 months in the
Goldberg and Kuo (1987) were the first to apply GAs to a optimization horizon. The chromosome representing a solution
water resource problem. They solved a pipeline optimization to such a problem consists of 12 genes representing the de-
problem using a GA. Since then, a number of applications of cision variables (releases) of the problem. The fitness of a
GAs to pipe network and other water resource problems have chromosome as a candidate solution is a function of these
been reported in the literature. A brief review of GA applica- genes and is obtained by evaluating the objective function of
tions to water resources problem can be found in the work of the problem. In application to water resource problems, infea-
Wardlaw and Sharif (1999). The literature describing the ap- sible chromosomes may be generated that fail to satisfy the
plication of GAs to water resources problems is not abundant. system constraints, such as continuity and component capaci-
The writers are aware of only a few applications of GAs in ties. The fitness of a chromosome is also a function of problem
this area. Esat and Hall (1994) applied a GA to the four-res- constraints and may be modified through the introduction of
ervoir problem. They concluded that GAs have potential in penalties when constraints are violated.
water resources optimization and that significant savings could The reproduction mechanism in a GA is composed of se-
be achieved in memory and execution times. Fahmy et al. lection, crossover, and mutation. A number of representation,
(1994) applied GAs to the optimization of reservoir systems. crossover, and mutation schemes have been described by
They compared the performance of a GA with DP and con- Goldberg (1989) and Michalewicz (1992). An evaluation of
cluded that GAs have significant potential in reservoir systems alternative GA formulations using different combinations of
optimization. Olivera and Loucks (1997) used GAs to develop these schemes in application to reservoir systems has been
operating policies for multi-reservoir systems, and concluded carried out by Wardlaw and Sharif (1999).
that GAs are a practical and robust method, and could lead to
effective operating policies. SYSTEM DESCRIPTION
The GA approach has the potential of overcoming the di- The Brantas Basin is located in the province of East Java,
mensionality problems associated with DP. Many optimization Indonesia. It has been used as a case study to demonstrate
256 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 FIG. 1. Schematic of Brantas Basin

application of the GA approach in the optimization of multi- flow sequences using the GA and the DDDP approaches.
reservoir systems. Fig. 1 is a map of the Brantas Basin. The These control curves were then applied in a simulation model
Kali Brantas is the second largest river in Java, with a total of the Brantas Basin (Wardlaw 1993) to obtain estimates of
catchment area of 12,000 km2. The primary water use in the long-term economic benefits from operation of the reservoir
basin is for irrigated agriculture. The annual diversion require- system.
ments for those irrigation schemes fed by the Brantas main With only Karangkates reservoir in operation, there are large
stream alone is about 1,600 Mm3. The only significant potable deficits in irrigation supply, even with a potable demand of 4
water demand from the K. Brantas is that of the city of Su- m3/s, although the current potable demand is probably of the
rabaya. Demand projections for Surabaya prepared by Mott order of 8 m3/s. The objective of modeling was to assess the
MacDonald (1988) indicated a potable water demand of impact of new development schemes, such as Wonorejo and
around 8 m3/s by the year 2000. Although this demand is a Beng, on the deficits in the irrigation supply while taking into
relatively small part of the total annual irrigation requirements, account the current potable demand for the city of Surabaya.
it is significant during dry periods when it competes with ir-
rigation demands for the available water resources. System Dynamics
Karangkates is the single most important multipurpose res-
ervoir in the Brantas Basin. The effective storage capacity is The dynamics of a multireservoir system can be described
253 Mm3, and although this is less than 4% of mean annual by the following equation:
basin runoff, its regulating effect on dry season flows is sig-
Si (t ⫹ 1) = Si (t) ⫹ Ii (t) ⫹ MRi (t) ⫺ Ei (t) (1)
nificant. The releases from Karangkates reservoir are used for
generating hydropower, for irrigation, and for supplying po- where Si (t) = vector of reservoir storages at time t in reservoirs
table water to the city of Surabaya. Two additional storage i = 1, n; Ii (t) = vector of reservoir inflows in time period t to
reservoirs Wonorejo and Beng (Fig. 1), have been included in reservoirs i = 1, n; Ri (t) = vector of reservoir releases in time
planning for the basin. Wonorejo is currently under construc- period t from reservoirs i = 1, n; Ei (t) = vector of reservoir
tion. evaporation in time period t from reservoirs i = 1, n; and M
= n ⫻ n matrix of indices of reservoir connections. The matrix
Database M depends upon the configuration of the system and consists
of ⫺1, 0, and 1.
The historic inflows and the other data used in this study,
The transformation from stage to stage is governed by (1).
including economic values for power and agricultural produc-
In addition to this, the releases and storages are limited by
tion, come from Mott MacDonald (1988). The DDDP and GA
physical considerations. The system is therefore subject to
models were applied to a single year of data for a critical
constraints expressed as follows:
period with a view to determining control curves for the res-
ervoirs. Critical inflow series with 90% annual exceedence Si,min ⱕ Si (t) ⱕ Si,max (2)
probability were derived. The control curves for Karangkates,
Wonorejo, and Beng reservoirs were developed for these in- Ri,min ⱕ Ri (t) ⱕ Ri,max (3)

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 257

where Si,min and Si,max are the allowable minimum and maxi- set up with real-value representation, tournament selection,
mum storages in reservoir i = 1, n, and Ri,min and Ri,max are the elitism, uniform crossover, and modified uniform mutation.
lower and upper bounds on the releases from reservoir i = 1, The chromosome length may be expressed as a product of the
n. number of stages and number of reservoirs in the system. The
decision variables are releases during each time period, R1(t),
CASE STUDY t = 1, N. In a GA, the discretization of storage and release is
not required. Each gene within the chromosome represents a
The GA approach has been applied to four cases. Case 1
release made from the reservoir and can take up any value
was the application of the approach to the situation with only
between the upper and lower bounds of releases. The aim of
Karangkates reservoir in operation, and with optimization car-
the GA approach is to find a gene sequence that yields the
ried out to maximize hydropower returns. In case 2, the op-
best chromosome, that is, maximizes the objective function
timization was carried out for hydropower and irrigation re-
value. A population size of 100 has been used and the model
turns, again with Karangkates reservoir only. Case 3
allowed to run for 1,000 generations for both cases. A cross-
considered application of the GA approach to a future water
over probability of 0.7 and a mutation probability of 0.028,
resource development scenario in the basin that includes Won-
which corresponds to one mutation per chromosome, has been
orejo reservoir in the system. Finally, case 4 considered op-
adopted.
eration of the reservoir system for another future scenario that
The evaluation function used by the GA corresponds to the
includes both Wonorejo and Beng reservoirs in the system.
objective functions given by (4) and (5) for cases EDS1 and
Both hydropower and irrigation benefits are considered for
EDS2, respectively. The constraints on storages were handled
cases 3 and 4.
by using a quadratic penalty. The GA initiates the search with
a randomly generated set of solutions that may or may not be
Case 1: Existing Development Situation, Optimizing
feasible. As the GA run progresses, the infeasible solutions are
for Hydropower Only (EDS1)
eliminated because the penalties assigned to them reduce their
The optimization was carried out for 36 10 day time steps fitness, thus making their chances slim for selection in the next
with the sole objective of maximizing the hydropower pro- generation. At the end of the run, a large number of feasible
duction while allowing deficits to occur in irrigation supplies. and near-optimal solutions are obtained. Fig. 2 shows the plot
The objective is to maximize the hydropower returns from of generation number versus best-of-generation maximum nor-
the reservoir subject to the given constraints. An additional malized fitness and average fitness of the population for EDS1
constraint is that the state of the system at the end of the and EDS2.
control horizon must be same as that at the beginning of the It can be observed from Fig. 2 that the GA starts with poor
control horizon. The objective function may be expressed as initial solutions but picks up quickly. By the end of 200 gen-

冘冘
n N

Maximize Z = Eit (S¯ it , Rit) (4)

i=1 t=1

where n is the number of reservoirs included in the model, N

is the operating horizon; Eit (S¯ it , Rit) is the economic return
from the power produced from reservoir i = 1, n, during period
t = 1, N; S̄it is the average storage in reservoir i during time
step t; and Rit is the release made from reservoir i in time step
t.

Case 2: Existing Development Situation, Optimizing

for Hydropower and Irrigation (EDS2)
For this case, the objective function comprises the hydro-
power benefits and a penalty for failing to supply the required
water for irrigation. The model attempts to minimize the def-
icits in irrigation supplies while maximizing the hydropower
production. The agricultural production was included in the
GA and DDDP models through an imposed penalty function
that was quadratic and is given by (6). The potable water de-
mand is included in the objective function through the im-
posed constraints. The objective function for this case can be
expressed as

冘冘
n N

Maximize Z = [Eit (S¯ it , Rit) ⫺ Pit ] (5)

i=1 t=1

where P1t = penalty for not meeting the irrigation demand for
the delta assumed to be of the following form:
P1t = k ⫻ [d1t ⫺ x1t ]2 (6)
where k is the penalty factor; d1t is irrigation demand at the
delta; and x1t is irrigation supply to the delta in time step t.

Solution Procedure
The solutions to the problems described above were ob- FIG. 2. Progression of Maximum and Average Fitness for Sin-
tained using the GA and the DDDP approach. The GA was gle Reservoir System

258 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

erations, a number of good solutions have been found. Once, When the optimization was carried out for both power and
the GA locates good solutions, some fine-tuning may be re- irrigation (EDS2), the target levels produced by the GA and
quired to improve those solutions. The GA was therefore al- DDDP models are lower than for the case when the optimi-
lowed to run for 1,000 generations, although the improvement zation was done for power only (EDS1). This is because the
has been found to be small after about 500 generations. models in case EDS2 attempt to satisfy the irrigation demands
The problem was also solved using DDDP so that an eval- of the delta in addition to maximizing the hydropower returns,
uation of the GA approach could be carried out. Instead of thus lowering the reservoir levels, particularly during the drier
searching over the entire state-stage domain for the optimum, periods [Fig. 3(b)]. The objective function values achieved by
as is the case for DP, DDDP analyzes only a few states at each the GA were 99.84 and 99.98% of those produced by the
stage. The procedure starts with the selection of a trial state DDDP for EDS1 and EDS2 respectively, which clearly shows
trajectory for the system and searches for improved trajectories that the GA solutions are very close to the optimum in terms
in the neighborhood of the trial trajectory. A corridor is defined of objective function values. The execution time for GA was
around the current trajectory within which the optimization is 90 s compared to 50 s for DDDP, irrespective of the objective
constrained. The traditional DP algorithm (Bellman 1957) is function used. Given the vast decision space that the GA had
then used to find the optimal trajectory within the corridor. to search for this particular problem, the results are satisfac-
The recursive equation used was tory.
A simulation model developed for the Brantas Basin (Ward-
Ft (ss) = max[Vt (st , dt) ⫹ Ft⫺1(st⫺1)] (7) law 1993) has been used to provide a more detailed evaluation
of the operating policies determined by optimization models.
where st is the state variable; dt is the decision variable; Vt (st ,
A series of economic postprocessing functions have been in-
dt) is the objective function value; Ft (st) is the cumulative re-
corporated in the model dealing with agricultural, hydropower,
turn at stage t with F0 (s0) known, and st⫹1 = g (st , dt) is the
and aquaculture production and with pumping costs. The re-
stage-to-stage transformation function. An improved trajectory
sponse of the system to the rule curves produced by the GA
is obtained after each iteration and used as a trial trajectory
and DDDP was evaluated for both cases (EDS1 and EDS2).
for the next iteration. The procedure continues until no further
From the output of the simulation model, the mean deficits in
improvement in the value of the objective function is obtained.
irrigation supply to the Brantas Delta were calculated. Fig. 4
A comparison of the trajectories produced by the GA and
shows the mean deficits in irrigation supplies to the delta for
by DDDP for EDS1 is presented in Fig. 3(a). These level tra-
EDS1 and EDS2 for the DDDP and the GA derived rule
jectories were obtained for a 10 year critical drought sequence
curves. The water resource of the basin is sufficient to permit
and give the target reservoir levels at the end of 10 day op-
continued increase in the potable water supply, but only at the
erating periods. The trajectories obtained by the GA closely
expense of irrigation water supply. This resulted in irrigation
matches that of DDDP except for a few points.
deficits in dry seasons for each scenario considered.

FIG. 4. Mean Irrigation Deficits, Existing Development Situa-

FIG. 3. Level Trajectories for Single Reservoir System tion

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 259

TABLE 1. Economic Returns in Million Rupiah, Existing De-
velopment Situation

EDS1 EDS2
Component DDDP GA DDDP GA
(1) (2) (3) (4) (5)
Irrigation production 50,681 50,905 51,626 51,591
Karangkates energy 32,981 33,058 32,741 32,789
Karangkates capacity 3,398 3,461 3,418 3,428
Total 87,060 87,424 87,785 87,808

Table 1 presents a comparison of the economic returns that

can be realized from the basin for each of the scenarios con-
sidered. The model results include irrigation production at the
Brantas Delta and energy and capacity benefits at Karangkates.
The figures in Table 1 show that the rule curves derived for
the case when optimization was carried out for both hydro-
power and irrigation (Case EDS2) produce higher total returns
than those obtained when the optimization was carried out for
hydropower only (Case EDS1).

Case 3: Future Development Scenario 1 (FDS1)

Various development schemes have been considered to im-
prove the utilization of available water resources in the basin
(Mott MacDonald 1988). Among these is the construction of
Wonorejo and Beng dams (Fig. 1). Wonorejo is a storage res-
ervoir that can be used for hydropower production and for
irrigation, and is considered as Future Development Scenario
1 (FDS1). After satisfying the local irrigation demands, the
water from Wonorejo can be used to augment the irrigation
demands at the Brantas Delta.
The combined operation of Karangkates and Wonorejo res-
ervoirs was optimized over 36 successive 10 day periods using
the GA and DDDP approaches for the objective function given
by (5). FIG. 5. Level Trajectories for Two-Reservoir System, FDS1
With the GA approach, the solution procedure is similar to
the one used for a single reservoir system. The fitness of chro-
mosomes is obtained using an evaluation function that corre-
sponds to the objective function given by (5), and where P2t
is the penalty for not satisfying local irrigation demands at
Wonorejo in time step t.
The output of the GA model includes the optimal reservoir
levels for both Karangkates and Wonorejo, which have been
plotted in Fig. 5. With GA, the objective function value
achieved was 99.07% of that produced by the DDDP. The
DDDP-derived rule curve for Karangkates is smoother than
the GA derived rule curve, but they are very similar. The ex-
ecution time for the DDDP was 210 s, while the GA took 290
s to complete 1,500 generations.
Rule curves derived using the optimization models were
used in the simulation model, as in earlier cases. Fig. 6 shows
the deficits in the delta irrigation supply under FDS1. There FIG. 6. Mean Irrigation Deficits, FDS1
is some decrease in deficits over the scenario when only Kar-
angkates was included in the model, but these are not signif- and Wonorejo are not able to meet the irrigation demands for
icant. The primary purpose of Wonorejo is to supply water to the delta, then releases would be made from Beng.
local irrigation schemes, and water is available for irrigation The evaluation function used in the GA model corresponds
in the delta only after satisfying these local demands. For this to the objective function given by (5) and is modified only to
reason, Wonorejo has little impact on deficits in the delta, de- reflect an additional reservoir in the system. The GA code used
spite its storage capacity. remains unchanged from the one used for the problems dis-
cussed earlier. The modifications are required only to the eval-
Case 4: Future Development Scenario 2 (FDS2) uation function and the data file, which may be different for
different objective functions.
The final case considers the development of operating rules A DDDP model for the three-reservoir system including
for a scenario that includes Karangkates and the Wonorejo and Karangkates, Wonorejo, and Beng was also developed. The
Beng reservoirs in the system. Beng is a potential offstream optimal trajectories produced by GA and DDDP for the three-
storage reservoir in the lower reaches of the K. Brantas (Fig. reservoir system are shown in Fig. 7. The GA trajectories for
1). During periods of high flows, water could be pumped back Karangkates and Wonorejo appear to be in very good agree-
from the main stream to Beng. If the water from Karangkates ment with the DDDP trajectories and are within 1.1% of the
260 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 for the GA-based rule is higher than that produced by the
DDDP.

ADVANTAGES OF GA APPROACH
Each of the above scenarios was run using the same GA
computer code and required only modification of the input
files. Each reservoir configuration required a different DDDP
code, however, and development of a generic DDDP code
would be difficult. The performances of the GA and DDDP
were similar, and the fact that the GA requires no discretization
of the state and decision variables is a distinct advantage. On
larger systems, the computing resources required by the GA
would be significantly lower than those required by DDDP.

SENSITIVITY TO MUTATION PROBABILITY

Proper settings of a GA’s parameters are required to achieve
the best performance from it. For this reason, a sensitivity
analysis was carried out to determine the effect of crossover
probability and mutation probability on the GA performance.
Similar analysis has been reported by Wardlaw and Sharif

FIG. 8. Mean Irrigation Deficits, FDS2

TABLE 2. Economic Returns in Million Rupiah, FDS2

Component DDDP GA
(1) (2) (3)
Net irrigation production 52,045 52,527
Wonorejo irrigation 13,304 13,068
Karangkates energy 32,657 32,821
Karangkates capacity 3,419 3,498
Wonorejo energy 1,510 1,510
Wonorejo capacity 0 0
Beng energy 1,141 1,137
Beng capacity 0 0
Total 104,076 104,561

FIG. 7. Level Trajectories for Three-Reservoir System, FDS2

optimum. There is deviation for the trajectories for Beng dur-

ing the later part of the operating horizon. The reason for dif-
ferences in trajectories for Beng could be that the same amount
of power can be produced by different combinations of re-
leases and heads. The execution time for the DDDP was 425
s, while the GA took 590 s to complete 2,000 generations.
The simulation model was run with the rule curves derived
using the optimization models. The mean deficits in irrigation
supply to the delta are shown in Fig. 8. It can be seen that,
with the inclusion of Beng in the system, there is a more sig-
nificant decrease in the mean deficits. The GA has performed
better than the DDDP in terms of delta water supply. Table 2
presents the economic returns from the operation of the res-
ervoir system under FDS2. The total return from the system FIG. 9. Sensitivity to Mutation Probability

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 261

(1999), who found that a mutation probability of one mutation • Discretization of state and decision space is not required
per chromosome produced the best results. The same result • Initial trial state trajectories are not required
has been found here for each of the scenarios considered. The • Complex reservoir systems need not be decomposed as in
results are shown in Fig. 9. Clearly, the presence of more than successive approximation-based approaches
two mutations per chromosome leads to poor GA performance. • GAs work with numerical values—highly complex ob-
jective functions do not introduce any difficulties, and
SENSITIVITY TO CROSSOVER PROBABILITY even discontinuous functions are acceptable
• Noninvertible systems can be easily handled without in-
The performance of GA is stable with respect to crossover troducing dummy state variables as in DP based ap-
probability. The variability of normalized fitness with cross- proaches
over probability is much lower than that with mutation prob- • Detailed simulation models can be used to evaluate the
ability. The trend of results was similar for all the problems fitness of solutions
considered here and for the four-reservoir problem solved by
Wardlaw and Sharif (1999). The sensitivity analysis has shown
The GA approach has some limitations. With very long
that very good results could be obtained for some particular
chromosomes, the likelihood of obtaining the optimum is re-
combination of crossover and mutation probability, but most
duced. A penalty function approach is required to satisfy the
combinations produce satisfactory results. It can be concluded
constraints of the problem, and experimentation is needed to
that the impact of crossover probability on the performance is
determine proper settings of the penalty parameters. The per-
insignificant.
formance of GAs has been found sensitive to mutation prob-
ability, and therefore proper choice of mutation probability is
EVALUATION OF ECONOMIC POSTPROCESSING vital for good GA performance. For reservoir systems, the in-
RESULTS dications are that one mutation per chromosome produces the
The response of the system to the rule curves produced by best results. It is concluded from the results of this research
the GA and DDDP was evaluated for all four cases (EDS1, that GAs can be effectively used in the optimization of control
EDS2, FDS1, and FDS2). A comparison of the total economic curves for multireservoir systems.
returns that can be realized from the basin for each of the
scenarios considered is presented in Fig. 10. From the results APPENDIX. REFERENCES
presented it can be observed that in all cases, the GA was able
to locate solutions very close to those produced by DDDP. In Adeli, H., and Cheng, N. T. (1994). ‘‘Concurrent genetic algorithms for
optimization of large structures.’’ J. Aerosp. Engrg., ASCE, 7(3), 276–
three of the four cases considered (EDS1, EDS2, and FDS2), 296.
the total economic returns for the system are higher with GA- Adeli, H., and Cheng, N. T. (1993). ‘‘Integrated genetic algorithm for
derived rules. In the remaining case, the total economic return optimization of space structures.’’ J. Aerosp. Engrg., ASCE, 6(4), 315–
produced by the GA was suboptimal by a very small amount 328.
(0.85%). The results thus demonstrate that the GAs can be Becker, L., and Yeh, W.-G. W. (1974). ‘‘Optimization of real time oper-
effectively used in identifying optimal operating policies for ation of multiple-reservoir system.’’ Water Resour. Res., 10(6), 1107–
1112.
multireservoir systems. Becker, L., Yeh, W.-G. W., Fults, D., and Sparks, D. (1976). ‘‘Operations
models for the Central Valley Project.’’ J. Water Resour. Plng. and
CONCLUSIONS Mgmt., ASCE, 102(1), 101–115.
Bellman, R. (1957). Dynamic programming, Princeton University Press,
A methodology based on the GA approach has been devel- Princeton, N.J.
oped for the optimization of reservoir systems operation. It Bellman, R., and Dreyfus, S. (1962). Applied dynamic programming,
has been demonstrated through application to the Brantas ba- Princeton University Press, Princeton, N.J.
Cheu, R. L., Jin, X., Ng, K.-C., Ng, Y.-L., and Srinivasan, D. (1998).
sin that the GA approach is easily applied to the optimization ‘‘Calibration of FREESIM for Singapore expressway using genetic al-
of multireservoir systems. In all the cases considered, the GA gorithm.’’ J. Transp. Engrg., ASCE, 124(6), 526–536.
was able to produce solutions very close to those produced by Dandy, G. C., Simpson, A. R., and Murphy, L. J. (1996). ‘‘An improved
DDDP. The approach is not limited by the complexity of the genetic algorithm for pipe network optimization.’’ Water Resour. Res.,
objective function. Unlike many optimization algorithms based 32(2), 449–458.
on nonlinear programming and optimal control theory, there Dasgupta, D., and Michalewicz, Z. (1997). Evolutionary algorithms in
engineering applications, Springer, Berlin.
is no requirement for the objective function to be differentia- Davis, L. (1991). Handbook of genetic algorithms, Van Nostrand Rein-
ble. In the context of reservoir systems optimization, specific hold, New York.
advantages are Esat, V., and Hall, M. J. (1994). ‘‘Water resources system optimisation
using genetic algorithms.’’ Proc., 1st Int. Conf. on Hydroinformatics,
Balkema, Rotterdam, The Netherlands, 225–231.
Fahmy, H. S., King, J. P., Wentzel, M. W., and Seton, J. A. (1994). ‘‘Ec-
onomic optimization of river management using genetic algorithms.’’
ASAE Int. Summer Meeting, American Society of Agricultural Engi-
neers, St. Joseph, Mich.
Feng, C.-W., Liu, L., and Burns, S. A. (1997). ‘‘Using genetic algorithms
to solve construction time-cost trade-off problem.’’ J. Comp. in Civ.
Engrg., ASCE, 11(3), 184–189.
Franchini, M. (1996). ‘‘Use of a genetic algorithm combined with a local
search method for the automatic calibration of conceptual rainfall-run-
off models,’’ J. Hydro. Sci., 41(1), 21–40.
Fwa, T. F., Chan, W. T., and Tan, C. Y. (1996). ‘‘Genetic-algorithm pro-
gramming of road maintenance and rehabilitation.’’ J. Transp. Engrg.,
ASCE, 112(3), 246–253.
Georgakakos, A. P. (1989). ‘‘Extended linear quadratic Gaussian (ELQG)
control: Further extensions.’’ Water Resour. Res., 25(2), 191–201.
Georgakakos, A. P., and Marks, D. H. (1987). ‘‘A new method for the
real time operation of reservoir systems.’’ Water Resour. Res., 23(7),
FIG. 10. Economic Returns under Different Scenarios 1376–1390.

262 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

Goldberg, D. E. (1989). Genetic algorithms in search, optimization and robot using genetic algorithm.’’ J. Comp. in Civ. Engrg., ASCE, 11(3),
machine learning, Addison-Wesley, Reading, Mass. 175–183.
Goldberg, D. E., and Kuo, C. H. (1987). ‘‘Genetic algorithms in pipeline Ng, K.-C. (1995). ‘‘Switching control systems and their design automa-
optimization.’’ J. Comp. in Civ. Engrg., ASCE, 1(2), 128–141. tion via genetic algorithms.’’ PhD thesis, University of Glasgow, Glas-
Grygier, J. C., and Stedinger, J. R. (1985). ‘‘Algorithms for optimizing gow, U.K.
hydropower system operation.’’ Water Resour. Res., 21(1), 1–10. Nopmongcol, P., and Askew, A. J. (1976). ‘‘Multilevel incremental dy-
Hall, W. A., and Buras, N. (1961). ‘‘The dynamic programming approach namic programming.’’ Water Resour. Res., 12(6), 1291–1297.
to water resources development.’’ J. Geophys. Res., 66(2), 510–520. Olivera, R., and Loucks, D. P. (1997). ‘‘Operating rules for multireservoir
Hall, W. A., Harboe, W., Yeh, W. W.-G., and Askew, A. J. (1969). ‘‘Op- systems.’’ Water Resour. Res., 33(4), 839–852.
timum firm power output from a two reservoir system by incremental Papageorgiou, M. (1985). ‘‘Optimal multireservoir network control by the
dynamic programming.’’ Water Resour. Ctr. Contrib., University of Ca- discrete maximum principle.’’ Water Resour. Res., 21(12), 1824–1830.
lif., 130, 273–282. Rajeev, S., and Krishnamoorthy, C. S. (1992). ‘‘Discrete optimization of
Heidari, M., Chow, V. T., Kokotovic, P. V., and Meredith, D. D. (1971). structures using genetic algorithms.’’ J. Struct. Engrg., ASCE, 118(5),
‘‘Discrete differential dynamic programming approach to water re- 1233–1250.
sources systems optimization.’’ Water Resour. Res., 7(2), 273–283. Rajeev, S., and Krishnamoorthy, C. S. (1997). ‘‘Genetic algorithm-based
Holland, J. H. (1975). Adaptation in natural and artificial systems, MIT methodologies for design optimization of trusses.’’ J. Struct. Engrg.,
Press, Cambridge, Mass. ASCE, 123(3), 350–358.
Howson, H. R., and Sancho, N. G. F. (1975). ‘‘A new algorithm for the Soh, C. K., and Yang, J. (1996). ‘‘Fuzzy controlled genetic algorithm
solution of multistate dynamic programming problems.’’ Math. Pro- search for shape optimization.’’ J. Comp. in Civ. Engrg., ASCE, 10(2),
gramming, 8(1), 104–116. 143–150.
Jenkins, W. M. (1991). ‘‘Structural optimisation using genetic algo- Takeuchi, K., and Moreau, D. H. (1974). ‘‘Optimal control of multiunit
rithms.’’ The Struct. Engr., London, 69(24), 418–422. interbasin water resource system.’’ Water Resour. Res., 10(3), 407–414.
Jenkins, W. M. (1992). ‘‘Plane frame optimum design environment based Trott, W. J., and Yeh, W. W.-G. (1973). ‘‘Optimization of multiple res-
on genetic algorithm.’’ J. Struct. Engrg., ASCE, 118(11), 3103–3112. ervoir system.’’ J. Hydr. Div., ASCE, 99(10), 1865–1884.
Koumousis, V. K., and Georgiou, P. G. (1994). ‘‘Genetic algorithms in Turgeon, A. (1981). ‘‘Optimal short-term hydropower scheduling from
discrete optimization of steel truss roofs.’’ J. Comp. in Civ. Engrg., the principle of progressive optimality.’’ Water Resour. Res., 17(3),
ASCE, 8(3), 309–325. 481–486.
Larson, R. E. (1968). State increment dynamic programming, Elsevier, Wardlaw, R. B. (1993). ‘‘Water resources systems modelling for the Bran-
New York. tas basin in East Java.’’ Proc., 4th Nat. Hydro. Symp., University of
Li, H., and Love, P. E. D. (1998). ‘‘Site-level facilities layout using ge- Wales, Cardiff, U.K.
netic algorithms.’’ J. Comp. in Civ. Engrg., ASCE, 12(4), 227–231. Wardlaw, R., and Sharif, M. (1999). ‘‘Evaluation of genetic algorithms
Liu, C., Hammand, A., and Itoh, Y. (1997). ‘‘Maintenance strategy opti- for optimal reservoir system operation.’’ J. Water Resour. Plng. and
mization of bridge decks using genetic algorithm.’’ J. Transp. Engrg., Mgmt., ASCE, 125(1), 25–33.
ASCE, 123(2), 91–100. Wasimi, S., and Kitanidis, P. K. (1983). ‘‘Real-time forecasting and daily
Marino, M. A., and Mohammadi, B. (1983). ‘‘Reservoir operation by operation of a multireservoir system during floods by linear quadratic
linear and dynamic programming.’’ J. Water Resour. Plng. and Mgmt., Gaussian control.’’ Water Resour. Res., 19(6), 1511–1522.
ASCE, 109(4), 303–319. Wurbs, R. A. (1993). ‘‘Reservoir-system simulation and optimization
McLaughlin, D., and Velasco, H. L. (1990). ‘‘Real-time control of a sys- models.’’ J. Water Resour. Plng. and Mgmt., ASCE, 119(4), 455–472.
tem of large hydropower reservoirs.’’ Water Resour. Res., 26(4), 623– Yakowitz, S. (1982). ‘‘Dynamic programming applications in water re-
635. sources.’’ Water Resour. Res., 18(4), 673–696.
Michalewicz, Z. (1992). Genetic algorithms ⫹ data structures = evolu- Yeh, W. W.-G. (1985). ‘‘Reservoir management and operation models: A
tion programs, Springer, New York. state-of-the-art review.’’ Water Resour. Res., 21(12), 1797–1818.
Mott MacDonald. (1988). ‘‘East Java provincial water resources master Yeh, W. W.-G., and Becker, L. (1982). ‘‘Multiobjective analysis of mul-
plan study for water supply,’’ Brantas Basin Simulation Model Final tireservoir operations.’’ Water Resour. Res., 18(5), 1326–1336.
Rep., Annexe V, Indonesia. Yeh, W. W.-G., Becker, L., and Chu, W. S. (1979). ‘‘Real time hourly
Murray, D. M., and Yakowitz, S. (1979). ‘‘Constrained differential dy- reservoir operation.’’ J. Water Resour. Plng. and Mgmt., ASCE, 105(2),
namic programming and its application to multireservoir control.’’ Wa- 187–203.
ter Resour. Res., 15(5), 1017–1027. Young, G. K. (1967). ‘‘Finding reservoir operating rules.’’ J. Hydr. Div.,
Navon, R., and McCrea, A. M. (1997). ‘‘Selection of optimal construction ASCE, 93(6), 297–321.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 263

 COMPUTER-BASED APPLICATION FOR THE PROCESSING
OF CLIENTS’ REQUIREMENTS

By J. M. Kamara,1 C. J. Anumba,2 Members, ASCE, and N. F. O. Evbuomwan3

ABSTRACT: This paper extends previous work on the use of quality function deployment for client require-
ments processing in construction. It describes a computer-based application, ClientPro, which implements a
quality-function-deployment-based model for the definition, analysis, and translation of client requirements into
solution-neutral design specifications. The methodology, development, operation, and evaluation of the prototype
are discussed, and it is concluded that ClientPro is useful in facilitating the systematic definition, analysis, and
mapping of client requirements to design specifications. It also provides for the incorporation and prioritization
of the perspectives represented by clients, the determination of the relative importance of their requirements,
and the early consideration of life-cycle issues affecting a facility. Thus, it can contribute to the ultimate satis-
faction of the requirements of construction clients.

INTRODUCTION client requirements processing, which forms the underlying

methodology on which ClientPro is based. This methodology
The use of computer-based technologies in the architectural/
and the development, operation, and evaluation by selected
engineering/construction industry is seen as a vital ingredient
industry practitioners of the prototype application are dis-
for improving the performance of the industry (e.g., in the
cussed, and the paper concludes with a brief description of the
integration of the construction process, which is otherwise
potential benefits of ClientPro and some of the ways it could
plagued with problems associated with fragmentation) (How-
be enhanced.
ard et al. 1989; Evbuomwan and Anumba 1966). Although
earlier applications of information technology (IT) in construc-
tion have focused on the more quantitative tasks (e.g., detailed REVIEW OF METHODOLOGY FOR CLIENT
design of structural members), it is now recognized that IT REQUIREMENTS PROCESSING
tools can also be used to improve the effectiveness of the brief- Client requirements processing involves the representation
ing (or architectural programming) phase of the construction of information (i.e., the needs, wishes, and expectations of the
process, which deals with the establishment of client require- person or firm responsible for commissioning and paying for
ments for a project (Benchmarking 1996). It is acknowledged the design and construction of a facility) in a format that en-
that computer-based tools can be used to record, analyze, and hances the understanding and implementation of precisely
link client requirements to design decisions, thereby support- what the client desires. It has been defined as the identification,
ing communication across disciplines regarding design con- analysis, and translation of explicit and implicit client require-
straints and trade-offs (Fiksel and Hayes-Roth 1993). This re- ments into solution-neutral design specifications (Kamara
alization has led to the development of IT tools to assist in 1999). A brief description of the methodology for client re-
the creation, documentation, communication, and management quirements processing [represented by a client requirements
of briefing information [e.g., Perkinson et al. (1994), Akin et processing model (CRPM)], which was presented in Kamara
al. (1995), Kumar (1996), Yusuf (1997)]. However, these ini- et al. (1999) now follows.
tiatives tend to be computerized (or automated) systems of
existing practices and are apparently based on the implicit as- MODEL FOR CLIENT REQUIREMENTS PROCESSING
sumption that automation is in itself desirable and that current
industry practice is likely to remain (Betts et al. 1995). The The CRPM was developed in response to the need for ef-
automation of flawed business processes can amount to the fective processing of client requirements on construction proj-
automation of existing problems and increased inefficiencies ects, which arises from the following:
(Betts et al. 1995). It is therefore necessary that computer-
based systems for briefing, which is characterized by flawed • The nature of client organizations and their motivation for
practices, should be based on re-engineered business processes initiating construction projects (Kometa and Olomolaiye
that can facilitate the effective processing of client require- 1997)
ments in construction (Kelly et al. 1992; Benchmarking 1996). • The number of ‘‘interest groups’’ [or stakeholders (e.g.,
This paper describes a computer-based application, Client- competing client departments, users, environmental pres-
Pro, that is based on an innovative approach for client require- sure groups, etc.)] that influence or are affected by the
ments processing. It extends previous work (Kamara et al. acquisition, existence, operation, and use of the facility
1999) on the use of quality function deployment (QFD) for being proposed by the project sponsor
• The need to facilitate collaborative working and trace-
1
Res. Assoc., Dept. of Civ. and Build. Engrg., Loughborough Univ., ability of client requirements through their precise defi-
Loughborough LE11 3TU, U.K. E-mail: [email protected] nition and representation in a format that is solution-neu-
2
Reader in Comp.-Integrated Constr. and Dir., Ctr. for Innovative
Constr. Engrg., Dept. of Civ. and Build. Engrg., Loughborough Univ.,
tral and that can make it easier to trace design decisions
Loughborough LE11 3TU, U.K. E-mail: [email protected] to the original intentions of the client (Perkinson et al.
3
Sr. Consultant, Whittman-Hart Inc., 6333 North Hwy. 161, Ste. 100, 1994)
Irving, TX 75038. E-mail: [email protected] • Limitations in current briefing practice in the United
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, Kingdom, which is characterized by the absence of a
2001. To extend the closing date one month, a written request must be structured procedure, poor communication, inadequate IT
filed with the ASCE Manager of Journals. The manuscript for this paper
was submitted for review and possible publication on December 20, 1999.
support, and the tendency to shift the focus from the cli-
This paper is part of the Journal of Computing in Civil Engineering, ent’s needs to the preferences of designers because the
Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/0004-0264– solution in the form of design sketches is usually used to
0271/$8.00 ⫹ $.50 per page. Paper No. 22086. clarify the problem (CIT 1996)
264 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 TABLE 1. Main Stages and Activities of Client Requirements Processing Model
Main stage Activities Required resources/tools
(1) (2) (3)
Define client requirements Establish and document basic facts about project and client. Identify and describe A multidisciplinary requirements
the people or groups (i.e., ‘‘interest groups’’) that influence, and/or are affected processing team; elicitation tech-
by acquisition, operation/use, and existence of the proposed facility. Elicit from niques (e.g., questionnaires and
client, functions and attributes of the proposed facility, information on its acqui- interview techniques)
sition, operation, future demolition, activities to be performed in the facility, and
characteristics of proposed users (i.e., ‘‘voice of the client’’).
Analyze client requirements Structure and prioritize client requirements. Restate (or decompose) client require- Requirements processing team;
ments into primary, secondary, and tertiary requirements to facilitate a clearer value tree analysis to decompose
understanding of those requirements. Determine relative importance of ‘‘interest requirements; decision-making
groups.’’ Prioritize tertiary requirements with respect to relative importance of techniques (e.g., criteria weight-
each interest group and their weighting of each tertiary requirement. ing) (ICE 1996)
Translate client requirements Generate design attributes. Determine target values for these design attributes using Requirements processing team;
information on characteristics of the project, proposed use and users of the fa- QFD ‘‘house of quality’’ matrix
cility, acquisition and operation of the facility, international standards (including
codes of practice), and target values for similar facilities. Translate tertiary client
requirements by matching them with identified design attributes to determine
which design attributes best satisfy a particular requirement. Prioritize design
attributes, which have been matched with client requirements. Prioritized design
attributes and their target values constitute solution-neutral specifications.

The CRPM is based on quality function deployment (QFD), Objectives of ClientPro

a matrix-based methodology that is used in the manufacturing
industry to translate customers’ required quality characteristics The primary objective of ClientPro is to facilitate the pro-
into appropriate product and service features and that is al- cessing of client requirements as prescribed in the CRPM. To
ready being applied in the construction industry (Serpell and achieve this, it was determined that ClientPro should allow the
Wagner 1997). It was developed using an iterative process processing of requirements for different client and project
involving academics and practitioners from the construction types and the entry, storage, viewing, and editing of infor-
industry (Kamara 1999). Other techniques from requirements mation required for processing client requirements at any stage
engineering, weapons systems acquisition, and construction in the processing activity. The system should also facilitate the
(Kamara et al. 1998) were also incorporated to enhance the calculation of relative weights for client interest groups, ter-
functionality and appropriateness of the model. The CRPM tiary requirements, and design attributes as well as the gen-
facilitates (1) the precise definition of client requirements to eration and printing of reports as required.
remove or minimize ambiguities; (2) the representation of cli-
ent requirements in a solution-neutral format that can be un- System Architecture and Information/Data Modeling
derstood by the different disciplines working on a project; (3)
the representation of client requirements in a format that To ensure that the desired features and objectives of
makes it easy to trace and correlate design decisions to the ClientPro are satisfied, an architecture for the prototype, illus-
original intentions of the client; (4) the incorporation of the trated in Fig. 1, was developed. Three modules provide the
perspectives and priorities of the client (owner, user, and other means for requirements definition, analysis, and translation.
‘‘interests’’) and maintaining focus on his/her requirements; These modules facilitate the storage and retrieval of informa-
(5) the management of the dynamic process for eliciting the tion from the data storage facility. The user interface interacts
requirements for a proposed facility; (6) the minimization of with all three modules, which are interconnected. The arrows
bias in decision making; (7) teamwork in the processing of linking the modules indicate that entry, storage, viewing, and
client requirements; and (8) the early consideration of issues editing of information can be done at any stage in the pro-
affecting the life cycle of the proposed facility. Table 1 pro- cessing activity.
vides a summary of the main stages and activities of the To implement the system architecture, an information/data
model. model that could be implemented within a computer environ-
The description of the CRPM in Kamara et al. (1999) in- ment was also developed. The underlying methodology for
cluded a case study example of the practical application of the client requirements processing (Table 1) was used to abstract
model using the actual requirements for a building project. the information/data model for ClientPro. The identified enti-
Although this demonstrated that the CRPM is of potential ben- ties and the interrelationships between them are illustrated in
efit to the construction industry, it also revealed that a paper- Fig. 2, which is an EXPRESS-G representation of the infor-
based application of the model is potentially very time con- mation model for client requirements processing. However, be-
suming. Further work to enhance the effectiveness of the
model required the development of a computer-based appli-
cation, ClientPro, for its implementation, which is the focus
of this paper.

COMPUTER IMPLEMENTATION OF CRPM

The development of ClientPro involved a number of activ-

ities, which included the establishment of the objectives,
scope, and desired features of the system; analysis of the in-
formation and process flow elements of the model; the selec-
tion of an implementation environment; and the actual devel-
opment. FIG. 1. System Architecture of ClientPro

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 265

 FIG. 2. Complete Entity-Level Diagram of CRPM Information Model

fore the information/data model in Fig. 2 can be described, the decision by a client to commission the refurbishment of
the EXPRESS-G notation needs to be explained. an existing building. The client organization determines the
business need for a project. On the other hand, the business
EXPRESS-G Notation need (e.g., need for more space) influences the type of project
(e.g., refurbishment) as well as the interest groups associated
An EXPRESS-G model is represented by three types of with the process and outcome of that project—the facility. The
graphic symbols, which together form a diagram. These are nature of the client organization and the kind of project will
the definition symbols, relationship symbols, and composition also determine how to procure, operate, and dispose of the
symbols. Definition symbols (i.e., varying forms of rectangular facility. For example, a client organization with a substantial
boxes) are used to denote simple, constructed, defined, and property portfolio can have personnel who are responsible for
entity data types as well as schema declarations. Relationship the acquisition of new property, unlike a one-off client who
symbols (lines) describe relationships that exist among the def- might require considerable assistance from outside consultants.
initions. Relationships are bidirectional, but one of the two The entity, ‘‘solution-neutral specifications,’’ shown in Fig. 2,
directions is emphasized, using an open circle in the empha- is derived from ‘‘other sources of information,’’ ‘‘facility ‘pro-
sized direction. For example, in an entity-attribute relationship, cess’,’’ ‘‘client/project characteristics,’’ ‘‘client business need,’’
the emphasized direction is toward the attribute. For inheri- and ‘‘tertiary client requirements.’’ Solution-neutral specifica-
tance relationships, the emphasized direction is toward the tions, required for the design of the facility that satisfies the
subtype. Composition symbols enable a diagram to be dis- business need of the client, consist of prioritized design attri-
played on more than one page. When there is a relationship butes and target values. The attributes for ‘‘client/project char-
between definitions on separate pages, the relationship line on acteristics,’’ ‘‘client business need,’’ ‘‘facility process,’’ and
the two pages is terminated by an oval (rounded) box that ‘‘other sources of information’’ are listed in Table 2.
contains a page number and a reference number. The page
number is the number of the page where a referenced defini- Selection of Implementation Environment
tion occurs. The reference number is used to distinguish be-
A relational database package, Microsoft (MS) Access, was
tween multiple references on a page. The composition symbol
used in the development of ClientPro. This was due to a num-
on the page where the reference originated contains the name
ber of factors, which included the vital role that database sys-
of the referenced definition (ISO 1994).
tems play in the management of requirements (Fiksel and
Information models described using EXPRESS-G are in-
Hayes-Roth 1993), the need to quickly develop the prototype
dependent of any implementation context and would therefore
at minimal expense, and the fact that MS Access had been
allow flexibility in computer implementation of the model.
used in similar research by Tsang and Jiao (1998).
Fig. 2 shows that ‘‘client requirements,’’ expressed as primary,
secondary, and tertiary requirements, describe the facility that
Development of ClientPro Using MS Access
satisfies the business need of the client. The requirements of
the client consist of information relating to the characteristics The development of ClientPro followed the general proce-
of the client and the project (i.e., ‘‘client/project characteris- dure for developing Access applications. An Access applica-
tics’’); his or her business need (i.e., ‘‘client business need’’); tion is a coordinated set of database objects—tables, forms,
and the acquisition, operation, and disposal of the facility (i.e., macros, queries, reports, and modules—that enable the user
‘‘facility ‘process’ ’’). Client requirements are influenced by to maintain the data contained in a database (McLaren 1996).
‘‘other sources of information’’ in the sense that a change in The development of an Access application involves an eval-
some standards (e.g., energy emission targets) might influence uation of the information needs of the system with particular
266 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 TABLE 2. Attributes for Defined Entities in Client Requirements Processing Model
Entity group Entity Attributes
(1) (2) (3)
Client/project characteristics Client details Name, address, business type, contact person name and address, number of employees,
average annual turnover, policy on occupancy/space use
Interest groups Name; type; relationship with client; influence in acquisition and use of facility; effect
of facility acquisition, operation, and use on group
Project details Project name, location, type, facility type, facility objectives
Client business need Use information Activity type, time of day performed, time of year performed, peak use times, required
equipment and furniture
User information User name, user type, user size, relationship with client, activity performed
Facility functions Function verb, function noun, function qualifier, function rationale (i.e., why a specific
function is required)
Facility attributes Attribute name, attribute meaning, attribute rational, function associated with attribute
Facility ‘‘process’’ Acquisition information Available budget, rationale for budget, level of client involvement (risk), rationale for
level of client involvement, client representatives, expected date of completion, ra-
tionale for completion date
Operation information Costs in use, meaning and rationale for cost in use, operation/management strategy,
rationale for operation/management strategy, level of operation/management tech-
nology, rationale for operation/management technology
Disposal information Expected life span, rationale for expected life span, etc.
Other sources of information International standards Standards for expression of user requirements, standards for air capacity for occupants
in specified building types, etc.
Benchmark/other information Operation/maintenance information for existing or similar facilities, etc

TABLE 3. Summary of How Required Features for ClientPro Were Implemented

Required feature How it was satisfied in ClientPro
(1) (2)
Allow entry, storage, viewing, and editing of Designed forms allowed display and editing of stored data. A customized menu was included on every
information at any stage in the processing form to facilitate display/editing of information at any stage of the requirements process. Development
activity of more than one form for a given table, in some cases, also facilitated display and editing of information
at appropriate stages in the processing activity.
Allow processing of requirements for differ- Predefined list of options in drop-down combo boxes guides the user on certain characteristics of project/
ent client and project types client types, but they are not restrictive as the user can enter his/her preferences. It should therefore be
possible to use ClientPro for a wide range of client and project types.
Facilitate generation of design attributes and A drop-down list of suggested design attributes facilitated the generation of design attributes as it prompts
determination of target values the user on what constitutes a design attribute. Entry and storage of target values could also be done
in ClientPro; however, to be more useful, an automated process for determining target values is needed.
Easy to use with adequate instructions on Every effort was made to ensure that forms were clear and unambiguous. Guidance was also provided on
how to carry out the processing activity every control, by written instructions in the ‘‘status bar text’’ property of those controls. This ensures
that when a button or text box receives the focus, text appears on the status bar of the form to explain
what action is required.
Facilitate generation of reports Several reports, which summarized and/or combined certain kinds of information, were designed. Appro-
priate macros and buttons were also designed to open these reports.
Allow for future extension and integration The ease with which it is possible to add other tables to ClientPro should facilitate its extension. As an
with other computer-based packages application in the MS-Office group, access can be successfully linked with other packages such as
Excel, other database systems, and AutoCAD.

TABLE 4. List of Client Requirements for Building Project (Kamara et al. 1999)
Outline requirements Supplementary requirements
(1) (2)
Prestigious building, particularly with respect to appearance Single point of access to the building and to each lab space is required to
Building that provides an open/welcoming environment promote security through staff monitoring. This will be reinforced by
Building that establishes security as a priority: limited points of entry/exit; controlled access systems (e.g., SWIPE cards or key button pads).
switch card operation (if affordable); main entrance reception for visitors Preference was expressed for entry to the building from the south where car
Building to be networked throughout and to include a mix of heavy engi- parking is to be provided. This could be integrated into a courtyard space
neering laboratories; computer labs; support for small-to-medium enter- with soft landscaping.
prises; offices; teaching space as available North elevation should address the streetscape of a major road as it is a
Engineering laboratories to be located in zones with emphasis on open plan- major frontage for the university.
ning, flexible space, and movable equipment wherever possible; light- The new facility will become the front door to the School of Science and
weight partitioning as appropriate Technology and should present an appropriate image.
Small-to-medium provision to be various and as available Existing school is spread over a number of buildings and lacks any focal
Office accommodation based normally on two persons per room; senior staff point or communal space. Opportunities exist in this project to address
in individual rooms this.

reference to the flow of data and information and the design information model (Fig. 2), which also provided the basis for
of the database objects that will solve the information prob- specifying the relationships between tables. Each table repre-
lems. sented an entity, corresponding to the defined entities in Table
The development of tables, queries, forms, reports, and 2. The entity attributes constituted the ‘‘fields’’ (or columns)
macros for ClientPro was carried out using the facilities pro- of respective tables. Queries facilitated the development of
vided within Access. The design of tables was based on the features that facilitated the analysis of requirements, the prior-
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 267
itization of client interest groups, tertiary client requirements,
and design attributes. Command buttons and customized menu
items were used in the development of the user interface. Table
3 provides a summary of how the desired features of ClientPro
were implemented in the development process.

USING CLIENTPRO TO PROCESS CLIENT

REQUIREMENTS
Running ClientPro requires the installation of MS Access
(version 7.0), although it is possible to convert it to a stand-
alone application using the MS Access Developer’s Toolkit.
When the file containing ClientPro is opened, a welcome
screen, which provides a gateway to other forms in the appli-
cation, is displayed. The use of ClientPro in the processing of
clients’ requirements, using the requirements for a building
project (listed in Table 4) is now described. However, because
of space restrictions, only a few screen shots are shown. A
detailed description of the use of ClientPro can be found in
Kamara (1999).

Defining Client Requirements FIG. 4. Structuring Client Requirements

This involves the input of information on the project; the Analyzing Client Requirements
client; interest groups represented by the client; proposed use
and users; the functions and attributes of the facility; and in- The analysis of client requirements involves the structuring
formation on the acquisition, operation, and disposal of the of requirements into primary, secondary, and tertiary require-
facility. Fig. 3 shows the forms for entering, viewing, and ed- ments; the prioritization of client interest groups; and the pri-
iting data about the functions and attributes of the facility. oritizing of tertiary requirements, using the appropriate forms
Facility functions are defined using a ‘‘verb,’’ ‘‘noun,’’ and a provided in ClientPro.
‘‘function qualifier.’’ Fig. 3 shows a function of the facility—
‘‘provide security to its users’’—that was derived from the list Structuring of Client Requirements
of requirements in Table 4. The rationale for this function is
The structuring of requirements is an iterative process,
that ‘‘the building will be used by a wide range of people;
which involves grouping requirements into a hierarchy from
adequate security is therefore required.’’ The attribute ‘‘wel-
general to specific, using the form shown on the left side of
coming to visitors’’ in the lower half of Fig. 4 is also derived
Fig. 4. On the top part of this form, previously defined func-
from Table 4. This process allows the ‘‘raw’’ requirements of
tions and attributes of the facility can be viewed. Other pre-
the client to be restated in a format that allows understanding
viously defined information on the use of the facility, interests
of what his or her needs are and the rationale behind these
groups, etc., can also be accessed by using the menu to open
needs. The entry of other information related to the ‘‘define
the relevant forms. The most general descriptions are primary
client requirements’’ function follows the same procedure as
requirements, and these are decomposed into greater levels of
for functions and attributes. Reports on the information entered
detail. To facilitate the structuring process, new statements can
on each form can be viewed by clicking on the appropriate
be introduced, or existing ones rephrased. The decomposition
button.
from general (primary) to specific (secondary or tertiary) is
based on the ability of the lower level requirement to enhance
the implementation of the higher level requirement. On the
other hand, the higher level requirement (e.g., primary require-
ments) provides justification for the lower level requirement
(secondary requirement). This means that, when a statement is
being considered to be a primary requirement (a group cate-
gory), the question ‘‘how?’’ is asked to determine how it can
be implemented, or which requirements can fall into that
group. Consistency between a secondary requirement and a
primary requirement is maintained by asking ‘‘why?’’ That is,
the primary requirement must provide a justification for the
associated secondary requirements. This process is continued
until the processing team is satisfied that they have a hierarchy
of primary, secondary, and tertiary requirements that ade-
quately represent the wishes and expectations of the client. The
right side of Fig. 4 shows the report of the combined list of
primary secondary and tertiary requirements.

Prioritization of Client Interest Groups

The prioritization of interest groups is by pairwise compar-
isons using the ‘‘client interest groups prioritization form’’
shown in Fig. 5. There are two parts to the form: the top
FIG. 3. Input Forms for Facility Functions and Attributes section, which reproduces the information on each interest
268 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 FIG. 6. Prioritization of Tertiary Client Requirements
FIG. 5. Prioritization of Client Interest Groups

Translating Client Requirements

group defined earlier for easy reference, and the bottom sec-
tion, which provides a matrix for pairwise comparison of in- Requirements translation involves generating design attri-
terest groups. [It should be mentioned that information on in- butes, determining target values, and associating tertiary re-
terests groups was derived from the client’s requirements quirements with design attributes in a matrix.
(Table 4) and other project documents.] Each group is com-
pared to each of the others, and the level of importance of one Design Attributes and Target Values
group over the other is represented by a 4, 3, 2, or 1 to indicate Design attributes for the facility are generated by selecting
major importance, medium importance, minor importance, and an appropriate attribute from a drop-down list, or by entering
equal importance, respectively. If CL1 is considered to be of a particular attribute in the space provided. A notation in the
major importance over CL2, the number 4 is entered in the form (DA01, DA02, etc.) is assigned to each attribute, whose
CL1 row under the CL2 column, and a 0 is entered in the target value is determined using previously defined and/or
opposite space on the matrix (i.e., row CL2 under column other information (e.g., user types, activity type, energy con-
CL1). The absolute weight for each interest group is the sum sumption levels, etc.) that can be displayed on the form used
of all the numbers on its corresponding row and this is auto- to generate design attributes. The first five columns of the
matically given in the ‘‘AbsWt’’ column. The relative weight lower half of Fig. 7 show the list of design attributes and their
of each interest group, calculated by rationalizing the absolute target values including unit of measurement and lower and
weights, is given in the ‘‘RelWt’’ column in the matrix. A upper limits.
report showing the prioritized list of interest groups can be
displayed by clicking on the appropriate button shown on the Translating Tertiary Requirements into Design Attributes
form. The form used to translate (or map) tertiary client require-
ments to design attributes is similar to the form displayed in
Prioritization of Tertiary Requirements
The prioritization of tertiary requirements is based on the
relative importance of interest groups and the weight that each
group places on those requirements. The ‘‘prioritization of ter-
tiary requirements’’ form (top left of Fig. 6), which displays a
matrix for entering the weightings assigned by interest groups
to each tertiary requirement, is used for this purpose. The
weighting that each interest group places on a tertiary require-
ment is entered in the appropriate place in the matrix. A num-
ber from 1 to 9 (least important to most important) is used to
indicate the level of importance a group places on each of the
requirements. After the weightings by interest groups for each
of the tertiary requirements have been entered, the calculated
absolute and relative weights are viewed by clicking on a but-
ton (i.e., ‘‘calculate weights’’) at the bottom of the form. Click-
ing this button displays a form (bottom left of Fig. 6) that
contains the list of tertiary requirements and an embedded Ex-
cel sheet that calculates the absolute and relative weights for
each tertiary requirement. The values for absolute and relative
weights on the embedded Excel sheet are then copied onto the
form containing the tertiary requirements. A report showing
the weighted list of tertiary requirements (right side of Fig. 6)
is generated and can be viewed by the user if required. FIG. 7. Translation of Tertiary Client Requirements

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 269

Fig. 6. It contains a matrix that displays the tertiary require- ClientPro does facilitate the effective processing of client re-
ments and relative weights of tertiary requirements. The pro- quirements as set out in the client requirements processing
cedure for translating tertiary requirements into design attri- model.
butes is similar to that for the calculation of tertiary
requirements described earlier. It involves comparing tertiary SUMMARY AND CONCLUSIONS
requirements against each design attribute and determining
whether there is any relationship between the two. The This paper has described a computer-based application,
strength of the relationship between requirements and design ClientPro, which extends previous work on the use of QFD
attributes is specified by using 9, 3, 1, or 0 to indicate strong for client requirements processing in construction (Kamara et
relationship, some relationship, weak relationship, and no re- al. 1999). The development operation and evaluation of the
lationship, respectively. The calculation of absolute (i.e., the prototype have also been discussed. The use of ClientPro to
sumproduct of relative weights of tertiary requirements and process client requirements offers the benefits associated with
the strength of the relationship between the two) and relative the use of computer-based applications (e.g., automated cal-
weights for design attributes follows the same procedure for culations, on-line guidance, etc.). When compared to time-con-
calculating the weights for tertiary requirements. The relative suming calculations in the paper-based application of the
weights for design attributes, together with design attributes CRPM in Kamara et al. (1999), this represents a significant
and target values (bottom half of Figure 7), make up the ‘‘so- improvement in the relative ease with which the model can be
lution-neutral specifications.’’ The generation of solution-neu- applied. A unique feature of ClientPro, however, is its under-
tral specifications concludes the processing of client require- lying methodology, which focuses on describing the proposed
ments. facility that satisfies a client’s business need, using terminol-
ogy that is familiar to the client. This description is not based
on the physical components of the facility (e.g., shape, mate-
EVALUATION OF CLIENTPRO rials, etc.), but on its functions, attributes, acquisition, opera-
To assess the effectiveness of ClientPro in facilitating the tion, and effect on people and the environment. Other features
processing of client requirements, it was evaluated by a selec- of the methodology include the following:
tion of industry practitioners who were prepared to see a dem-
onstration of the operation of ClientPro. The main aim of the • Exclusive consideration of client requirements to ensure
evaluation was to assess the effectiveness of ClientPro to pro- that adequate focus is maintained on the needs and wishes
cess client requirements. Four practitioners (an architect, a of the client
civil engineer, a project manager, and a planning and quality • The provision for clarifying, through the description of
control manager), with an average of 14 years of experience the rationale for certain statements, the real intentions of
in the construction industry, were involved in the evaluation the client and ensuring that those statements are not just
of ClientPro. Each evaluator was individually taken through ‘‘wish lists’’ but are actually based on the real (business)
the operation of ClientPro using the actual requirements for a need of the client
building project (i.e., those listed in Table 4) and was then • The incorporation of the perspectives and priorities rep-
requested to complete a questionnaire. For each question, resented by the client, which can be useful in minimizing
evaluators were asked to express their opinion on the effec- client organizational problems associated with the project
tiveness of ClientPro on a scale of 1 (poor) to 5 (excellent). • The early consideration, through operation and disposal
Comments on ways to improve the software were also solic- information, of issues affecting the life cycle of a facility,
ited. The four evaluators were generally satisfied with the ef- which can facilitate their incorporation into the design
fectiveness of ClientPro in processing client requirements, al- process
though there were suggestions for improvements to the user • The systematic mapping of client requirements to solu-
interface. Table 5 compares the average ratings of the eval- tion-neutral design specifications, which can limit a client
uators (i.e., arithmetic mean of their responses to each ques- from being overprescriptive and can therefore allow de-
tion) against the objectives for the CRPM on which ClientPro signers to be more innovative in proposing design solu-
is based. The average for all scores is 3.63 out of a possible tions (MacLeod et al. 1998)
maximum of 5. It should be pointed out that because the eval- • The utilization of a requirements processing team that is
uation was not based on actual software utilization by the separate from the design team to ensure that design con-
evaluators, but on a demonstration of its use, these ratings can siderations do not constrain the process of establishing the
only be indicative. Nevertheless, they demonstrate that client’s requirements

TABLE 5. ClientPro and Objectives for CRPM

Thus, ClientPro is useful in facilitating a better understanding
of client requirements and their implementation as well as ef-
Average fective collaborative work and design creativity. It is also use-
rating ful in minimizing uncertainties and downstream problems be-
Objective for CRPM (out of 5) cause of the early consideration of life-cycle issues and in
(1) (2) providing the basis for effective requirements management
Precise definition of requirements to remove ambiguities 3.67 throughout the life cycle of the project. Further improvements
Stated in solution-neutral format and ease of understanding (e.g., to the user interface) and extensions (e.g., integration
by different disciplines 3.85 with other packages, etc.) can enhance its effectiveness in con-
Ease of tracing and correlating requirements to original in- tributing toward the complete satisfaction of client require-
tentions of client 3.67
Reflective of perspectives and priorities of client (i.e., main- ments on construction projects.
taining focus on client; minimizing bias in decision
making) 3.58 APPENDIX. REFERENCES
Encouraging team work and early consideration of life-cycle
issues 3.50 Akin, Ö., Sen, R., Donia, M., and Zhang, Y. (1995). ‘‘SEED-PRO: Com-
Representing an improvement to the existing process of puter-assisted architectural programming in SEED.’’ J. Arch. Engrg.,
briefing 3.50 ASCE, 1(4), 153–161.
Average for all scores 3.63 Benchmarking best practice report: Briefing and design. (1996). Con-
struct IT Center of Excellence, Salford, U.K.

270 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000

Betts, M., Fischer, M. A., and Koskela, L. (1995). ‘‘The purpose and Kometa, S. T., and Olomolaiye, P. O. (1997). ‘‘Evaluation of factors in-
definition of integration.’’ Integrated construction information, P. Bran- fluencing construction clients’ decision to build.’’ J. Mgmt. in Engrg.,
don and M. Betts, eds., E & FN Spon, London. ASCE, 13(2),77–86.
Evbuomwan, N. F. O., and Anumba, C. J. (1996). ‘‘Towards an integrated Kumar, B. (1996). ‘‘A prototype design brief development assistant.’’ MS
engineering design environment.’’ Information representation and de- dissertation, University of Glasgow, Glasgow, U.K.
livery in civil and structural engineering, B. Kumar and A. Retik, eds., McLaren, B. J. (1996). Understanding and using microsoft access for
Civil-Comp Press, Edinburgh, U.K., 127–134.
Windows 95, West Publishing Co., St. Paul, Minn.
Fiksel, J., and Hayes-Roth, F. (1993). ‘‘Computer-aided requirements
management.’’ Concurrent Engrg.: Res. and Applications, 1, 83–92. MacLeod, I. A., Kumar, B., and McCullough, J. (1998). ‘‘Innovative de-
Howard, H. C., Lewitt, R. E., Paulson, B. C., Pohl, J. G., and Tatum, C. sign in the construction industry.’’ Proc., Instn. Civ. Engrs., London,
B. (1989). ‘‘Computer-integration: Reducing fragmentation in AEC in- 126, 31–38.
dustry.’’ J. Comp. in Civ. Engrg., ASCE, 3(1), 18–32. Perkinson, G. M., Sanvido, V. E., and Grobler, F. (1994). ‘‘A facility
Kamara, J. M. (1999). ‘‘Client requirements processing for concurrent programming information framework.’’ Engrg. Constr. and Arch.
life-cycle design and construction.’’ PhD thesis, University of Teesside, Mgmt., 1(1), 69–84.
Teesside, U.K. Serpell, A., and Wagner, R. (1997). ‘‘Application of quality function de-
Kamara, J. M., Anumba, C. J., and Evbuomwan, N. F. O. (1998). ‘‘Tools ployment (QFD) to the determination of the design characteristics of
for client requirements processing in concurrent life-cycle design and building apartments.’’ Lean construction, L. Alarcón, ed., Balkema,
construction.’’ Proc., 2nd Int. Symp. on Tools and Methods for Con- Rotterdam, The Netherlands, 355–363.
current Engrg., I. Horvath and A. Taleb-Bendiab, eds., Manchester
Tsang, M. M., and Jiao, J. (1998). ‘‘Computer-aided requirements man-
Metropolitan University, Manchester, U.K., 73–83.
Kamara, J. M., Anumba, C. J., and Evbuomwan, N. F. O. (1999). ‘‘Client agement for product definition: A methodology and implementation.’’
requirements processing in construction: A new approach using QFD.’’ Concurrent Engrg.: Res. and Applications, 6(2), 145–160.
J. Arch. Engrg., ASCE, 5(1), 8–14. Yusuf, F. (1997). ‘‘Information and process modeling for effective IT
Kelly, J., MacPherson, S., and Male, S. (1992). The briefing process: A implementation at the briefing stage.’’ PhD thesis, University of Sal-
review and critique, Royal Institution of Chartered Surveyors, U.K. ford, Salford, U.K.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 271

 APPLICATIONS OF MATHEMATICAL SOFTWARE PACKAGES IN
STRUCTURAL ENGINEERING EDUCATION AND PRACTICE
By Primus V. Mtenga1 and Lisa K. Spainhour,2 Members, ASCE

ABSTRACT: The use of mathematical software packages provides a number of benefits to an engineering user.
In general, the packages provide a platform that supports iterative design and parametric analysis through a
flexible, transparent interface combined with extensive computing power. This enables an engineering user to
develop design equations that are based on fundamental mechanics theories, rather than relying on the ‘‘black-
box’’ approach of most commercial design packages. As an example, a closed-form solution for obtaining
effective length factors for the design of stepped columns is presented. In the example a series of formula is
used to demonstrate the transparency of Mathcad, including the ability of using real engineering units in the
calculations, formulas as they may appear in textbooks or in codes, and ability to hide and password protect
some areas. This facilitates easier automation of the design and design checking processes. Most commercial
structural design packages can be classified as black-box packages. The analyst inputs data at one end only to
receive results at the other without fully appreciating the process the input data have gone through. This phe-
nomenon has the tendency of reducing the engineer to a technician, blindly implementing the ideas of the
software designer. The Mathcad package discussed in this paper and similar mathematical packages returns the
engineer to being in control of the design process.

INTRODUCTION the problem of stepped column design, where a closed-form

Advances in computing power have revolutionized the civil solution, made feasible by the mathematical software package,
engineering field over the past few decades. Finite-element is proposed. The advantages of using the closed-form approach
analysis programs have made it possible to solve complex sys- are clear in comparison to the current method, which involves
tems with more degrees of freedom than imagined a few years a 3D interpolation between values from several standard ta-
past. Computer-aided drafting software has virtually elimi- bles. The use of mathematical packages in the analysis and
nated hand drawing from the engineering office. Moreover, the design process allows transparency without a significant ad-
ever-decreasing cost of the microcomputer has made them af- verse effect on the speed of the process.
fordable to civil engineering firms of all sizes and specialties.
This availability, coupled with the fast-growing library of en- BENEFITS OF MATHEMATICS
gineering computer software, has greatly expanded the engi- SOFTWARE PROGRAMS
neer’s effectiveness, thus allowing the engineer to be more Mathematical software packages, including Mathcad, Math-
competitive in the marketplace. ematica, MathView, MATLAB, and Maple, provide powerful
However, with this increase in microcomputer and software computational tools for both engineering education and prac-
availability, several ethical questions concerning a possible mi- tice. With these tools, it is possible to solve very complex
crocomputer misuse have arisen and need to be addressed mathematical problems with the click of a computer mouse.
(Gifford 1987). The ‘‘black-box’’ syndrome, infinite accuracy In addition, this capacity enables engineers to base their de-
of the computer, data/input integrity, instant expertise of the signs on fundamental mechanics theories, rather than empirical
engineer, and lack of judgment on the part of the engineer are or tabular relationships. Although the capabilities and features,
identified as the microcomputer misuses of most concern. Ac- as well as the nature of the interface, vary from program to
cording to Gifford, these problems are best avoided through program, each provides an environment from which equations
education in the ethical use of microcomputers. This education representing engineering problems can be easily and quickly
should be aimed at, among other things (1) understanding the solved. Most programs of this type provide the following ca-
software’s philosophy; (2) ascertaining the software’s suitabil- pabilities:
ity and design acceptability; and (3) realizing it is the profes-
sional responsibility of the engineer to use the microcomputer • Numeric and symbolic computations
as a powerful design aid only, not allowing his judgment to • Algebraic, trigonometric, and matrix functions
be sacrificed. • Graphics capabilities
The increasing power of spreadsheet programs and mathe- • Conditional programming
matical software packages can be used as one of the tools to • Flexible, easy to use interface
address the ethical problems raised by Gifford. The benefits
of such programs are especially apparent in the classroom and Most design offices now use commercial software packages
small design office, where the ability to perform ‘‘what-if’’ for their structural analysis and design. Most commercial pack-
type calculations has proven the most useful. This paper illus- ages can be classified as black-box packages. The analyst in-
trates the application of a mathematical software package to puts data at one end only to receive results at the other without
1
Civ. Engrg. Dept., Florida A&M Univ.-Florida State Univ. Coll. of
fully appreciating the process the input data has gone through.
Engrg., 2525 Pottsdamer St., Tallahassee, FL 32310. This phenomenon has the tendency of reducing the engineer
2
Civ. Engrg. Dept., Florida A&M Univ.-Florida State Univ. Coll. of to a technician, blindly implementing the ideas of the software
Engrg., 2525 Pottsdamer St., Tallahassee, FL 32310. designer. If this trend is allowed to continue, sooner or later
Note. Editor: Sivand Lakmazaheri. Discussion open until March 1, the teaching of structural engineering will have to change to
2001. To extend the closing date one month, a written request must be reflect this development.
filed with the ASCE Manager of Journals. The manuscript for this tech-
nical note was submitted for review and possible publication on August
On the other hand, mathematical software packages provide
25, 1998. This paper is part of the Journal of Computing in Civil En- a great deal of computing power and flexibility in an inexpen-
gineering, Vol. 14, No. 4, October, 2000. 䉷ASCE, ISSN 0887-3801/00/ sive package, making them ideal for educational and small
0004-0273–0278/$8.00 ⫹ $.50 per page. Technical Note No. 19098. office environments. The biggest benefit to this group of users,
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 273
however, is the transparency, which allows the user to vary
the program inputs and see the effects at intermediate steps
throughout the solution process. This ‘‘white-box’’ or ‘‘clear-
box’’ approach can be contrasted with the black-box procedure
inherent in many software systems, where only the input and
output data are visible and all calculations are hidden to the
user.
The benefits of this approach in an academic environment
are obvious. This class of software lets undergraduate students
solve a larger number of more meaningful problems without
having to learn the details of a programming language. Be-
cause of the white-box interface, the student can see the equa-
tions involved in the solution process, which makes it easier
to find theoretical and programming errors. The student also
can vary the input and see the results immediately, helping
him or her to grasp the underlying theory, see the results of
what-if type questions, and perform parametric studies (Pavlov
and Mtenga 1997). A number of authors have described ap-
proaches to integrating mathematics software programs into
the engineering curriculum (Agelidis 1994; DeLyser 1996;
Myers 1992).
Many of these benefits carry over into the engineering de-
sign office. Mathematics software packages can assist the de- FIG. 1. Examples of Stepwise Uniform Columns
sign process by automating many routine calculations and al-
lowing the designer to evaluate different alternatives instantly. Fig. 1(a), the upper segment is designed to carry an axial load
The software is powerful enough to solve the large and cum- P1 and the associated upper column moment, whereas the
bersome equations used to characterize system architectures. lower portion is designed for P1 ⫹ P2 and the associated lower
Further, the interactive nature of the interface supports iterative column moments. This makes it necessary to determine effec-
design and allows for rapid convergence to a final design. The tive length factors for both the upper and the lower column
white-box approach also provides designers and permitting of- sections.
ficials efficient and reliable tools to check the results of more These effective length factors for the column segments are
sophisticated software, where the logic and algorithms are of- not readily available. The AISC load and resistance factor de-
ten hidden in the source code, unavailable to the engineer sign (LRFD) specifications provide no guidelines on the de-
(Magner 1995; Pavlov 1996; Pavlov and Mtenga 1997). The sign of stepped columns. However, effective length factors for
worksheetlike interface provided by some packages allows the stepped columns are tabulated in its Steel design guide series
user to add text blocks, which can be used to define variables No. 7 (Fisher 1993). These values were developed by Agrawal
and reference codes and to document the design procedure. and Stafiej (1980) based on transcendental equations devel-
oped by Anderson and Woodward (1972). The following pa-
STEPPED COLUMNS rameters were considered in developing the tabulated values:
As indicated above, the increasing power of the available (1) The end fixity types at the column ends; (2) the ratio of
mathematical software packages can eliminate the need for the bottom segment length to the total column length; (3) the
tabulated information currently used in many codes and spec- ratios of the segments’ moment of inertia; and (4) the ratio of
ifications, thus enabling the engineer to develop and utilize a the intermediate column load to the total axial load. These
complex system of equations that are grounded in basic me- tabulated values and associated design approaches are not yet
chanics theories. The problem of stepped-column design is included in the AISC specifications (AISC 1995).
used to illustrate this functionality aspect of these software In this paper, an equation-based approach to the problem of
packages. Stepwise uniform columns are a necessity in struc- stepped columns will be used in a Mathcad example, elimi-
tural applications, especially in situations where intermediate nating the need for tabulated values. The formulas-based ap-
loads are introduced into the column. A typical example of proach is easier to incorporate into mathematical software
this scenario is a column supporting a crane loading, as shown packages than the current approach based on tabulated values.
in Fig. 1(a). The lower part of the column will have to support If a designer were to use the tabulated K-factor values for a
both the crane loading and the weight of the roof structure. stepped column, advantages of automation would be mini-
As the roof loads are relatively small compared to crane load- mized. In the example MathSoft’s Mathcad software (Mathcad
ing, it may be economical to have a heavier section for the 1991) was utilized for all of the computations. However,
lower part and a lighter section for the upper portion of the Mathcad was chosen only to illustrate the capabilities of this
column. Another situation where stepped columns may be seen class of software packages. Any of the other programs men-
is in a retrofit operation where, for example, cover plates may tioned in the ‘‘Introduction’’ could have been used instead.
be added to the flanges of the column, resulting in a stepped
column as shown in Fig. 1. STEPWISE UNIFORM MEMBER
The design of stepwise uniform columns is time consuming
and complicated. This complexity has attracted a number of In this paper, the above-described approach will be applied
researchers to the study of stepped-column behavior. Examples to stepwise uniform columns. The column will have two sec-
of such studies include those reported by Lui and Sun (1995), tions, with the top referred to as the reference section and the
Bendapudi (1994), Fraser (1990), Agrawal and Stafiej (1980), bottom as the modified section. In determining the effective
and Anderson and Woodward (1972). According to AISC’s length factors, the following ratios are used:
Steel design guide series No. 7 (Fisher 1993), the best design
approach for stepped columns is to design the upper and lower • The ratio ␣ of the moment of inertia of the lower segment
portions of the column as individual segments. As shown in to the upper segment
274 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 Q4 = 12␮␤(240␤6 ⫺ 840␤5 ⫹ 756␤4 ⫹ 210␤3 ⫺ 420␤2 ⫹ 105)
(6)

Q5 = ␮␤2(⫺1,920␤5 ⫺ 7,560␤4 ⫹ 9,576␤3 ⫹ 2,940␤2 ⫹ 2,100␤

⫺ 1,260) (7)

Q6 = ␮␤3(320␤4 ⫺ 1,400␤3 ⫹ 2,247␤2 ⫺1,575␤ ⫹ 420) (8)

What is now required is to determine a factor to modify the

actual length of the column such that the buckling load will
be the same as the buckling load of (2). For the reference
section, that is, the top portion of the column, the critical buck-
ling load is

␲2S
Pcr = (9)
K 12L2

Thus, solving for K1 yields

冉冑 冊
FIG. 2. Model of Stepwise Uniform Column
S
K1 = ␲ (10)
• The ratio ␤ of the length of the bottom segment to the L2Pcr
total length of the column
• The ratio ␮ of the load at the stepped level to the load at A substitution of (2) into (10) will yield
the top of the column
␲兹[S 2(Q4 ⫹ 612) ⫹ DS(Q5 ⫺ 156) ⫹ D 2(Q6 ⫹ 12)]
K1 = (11)
Thus, the modified section will be ␤L in length and have a 兹252[S 2(Q1 ⫹ 24) ⫹ DS(Q2 ⫺ 6) ⫹ D 2(Q3 ⫹ 1)]
moment of inertia equal to ␣I. The reference section will be
(1 ⫺ ␤)L long and have a moment of inertia I, as shown in The axial load that will be acting in the modified section will
Fig. 2(b). be (1 ⫹ ␮)P. The modification factor for this section will be
At one end, a rotational restraint with spring constant kR is such that
provided. The value of kR is set equal to zero for a pinned end
and infinity for a fixed end condition. To facilitate computation ␲2␣S
(1 ⫹ ␮)Pcr = (12)
in Mathcad, a very large value, on the order of 1020 is assumed K 22L2
to be equivalent to infinity. A semirigid end condition is rep-
resented by an intermediate spring constant value. Solving for K2 yields
At the pinned end of the column (assumed to be the refer-
ence section) a compressive force P is applied. At the stepped
level of the column, a force ␮P is applied. The application of
K2 = ␲ 冉冑 冑 冊 冑
S
L2Pcr
␣
1⫹␮
= K1
␣
1⫹␮
(13)

the Rayleigh-Ritz method to the column described will yield

the critical buckling load Eqs. 21 and 23 were used to compute the effective length

冕 冕
␤L L
factors for the stepped column that were programmed in the
Mathcad example.
S( y⬙)2 dx ⫹ ␣S( y⬙)2 dx
0 ␤L
Mathcad is a Windows-based product that displays both text

冕 冕
Pcr = L ␤L
(1) and equations in a ‘‘what you see is what you get’’ (WYSI-
WYG) format. New users find the program intuitive and easy
( y⬘) dx ⫹
2
␮( y⬘) dx
2

0 0
to use. As changes are made to the files, the results are updated
automatically, making parametric studies easy to formulate.
The external work done by load P will be as a result of the Graphs are easy to create and can be frequently used to sum-
deformation of the entire length of the column. On the other marize and clarify results. Users can easily add notes, modify
hand, the external work done by the load at the stepped level equations, and add supplemental checks to verify the pro-
␮P will be due to deformation in the segment of length ␤L gram’s results.
[Fig. 2(b)]. This explains the integration limits in the denom- Mathcad and comparable mathematical software packages
inator of (1). can be used to keep engineers and students in charge of their
Using a fourth-order polynomial for y(x) in (1) will yield designs, as opposed to the black-box design software output.
Mathcad is an excellent tool that can be used to bridge the
252S [S 2(Q1 ⫹ 24) ⫹ DS(Q2 ⫺ 6) ⫹ D 2(Q3 ⫹ 1)] gap between the traditional educational approach and the au-
Pcr =
L2 [S 2(Q4 ⫹ 612) ⫹ DS(Q5 ⫺ 156) ⫹ D 2(Q6 ⫹ 12)] tomated black-box industry approach. Unlike programming
(2) languages, Mathcad does not require extensive training in
commands, syntax, or structured programming. For engineer-
where S = E ⭈ I, where E = modulus of elasticity of the material ing applications, all that is needed is knowledge of how to
and I = moment of inertia about the buckling axis; and D = solve the problem in a step-by-step flowchart manner, a clas-
kR ⭈ L, where kR = rotational stiffness of the partial restraint at sical classroom approach.
the end of the member In the classroom and in practice, students or engineers using
Q1 = (␣ ⫺ 1)␤3(144␤2 ⫺ 360␤ ⫹ 240) (3) Mathcad can solve problems, step by step, just as they would
in a pencil and paper solution. Equations, notations, and sym-
Q2 = (␣ ⫺ 1)␤ (⫺96␤ ⫹ 270␤ ⫺ 240␤ ⫹ 60)
2 3 2
(4) bols are used just as in the text and lectures. As problems
become more complex, previous models and problems can be
Q3 = (␣ ⫺ 1)␤(16␤4 ⫺ 50␤3 ⫹ 55␤2 ⫺ 25␤ ⫹ 5) (5) reused and combined to build more comprehensive problem
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 275
solutions. Equations and assumptions are visible in engineer- APPENDIX I. STEPPED COLUMN: MATHCAD
ing notation, unlike ordinary spreadsheets. Text is frequently EXAMPLE OF MTENGA AND SPAINHOUR
used to outline the design process as well as reference appli-
cable theory, standards, specifications, and literature. User Specified Units
The benefits of solving engineering problems in Mathcad N ⬅ 1 newton; KN ⬅ 1,000 N; kip = 1,000 lbf;
are numerous. Users can easily build upon their earlier work
by incorporating previous models into new projects. At the end psf ⬅ lbf/ft ; MPa ⬅ N/mm ;
2 2
ksi ⬅ 1,000 psi;
of the project, the users will have a working model of the plf ⬅ lbf/ft
entire analysis and/or design. Students not only learn how to
solve each piece of the design, they can understand the prob- Note 1: One of the most powerful characteristics of Math-
lem from an overall perspective. Showing the results both nu- cad is the ability to use units in the calculations (both built-in
merically and graphically in Mathcad reinforces the engineer- units and user-specified units). When the formulas’ noncom-
ing concepts and theories. Once the model is complete, patible units are inputted, an error will appear, thus forcing
students have a working program. Written in engineering no- the engineer or student to review the input information. Black-
tation, they can reuse the model in other classes and as a ver- box packages and even other mathematical packages permit
ification tool to check other design programs. ‘‘garbage in garbage out. Moreover, one can mix both metric
and customary units (at appropriate magnitudes, e.g., Fy = 50
ksi entered as Fy = 345 MPa) in the same worksheet and end
MATHCAD IMPLEMENTATION OF up with the same design results.
STEPPED-COLUMN PROBLEM
Input Data
Given below is a Mathcad implementation of the stepped-
column problem discussed above. In the interest of saving Note 2: With Mathcad it is possible to hide and password
space, some areas are hidden. However, unlike black-box pro- protect some regions of the worksheet. Normally one will have
grams, the contents of these hidden areas can be viewed by the ‘‘Input Data’’ area visible, which is the case in this ex-
double clicking on the name icon of the area. For security ample.
reasons some areas may be password protected; thus, to view
the contents of these areas, one may need to have the correct Material Data
password. As seen in the example (Appendix I, notes 1–6),
text can be inserted within the calculation, as is the case in a Modulus of elasticity
paper and pencil worksheet, thus creating a well-documented E := 200 KN/mm2
report. All this can be done without much impairment in the
ability and sophistication of the developed application. Data Yield stress of material (A572 grade 50 steel)
can be imported from or exported to external files and input
Fy := 345 N/mm2
and output arrays can be created, thus creating an application
capable of examining several design alternatives. Residual stress
Fr := 69 N/mm2
CONCLUSIONS
Bending strength reduction factor (per LRFD)
The use of mathematical software packages provides a num- ␾b := 0.90
ber of benefits to an engineering user. In general, the packages
provide a platform that supports iterative design and paramet- Compression strength reduction factor (per LRFD)
ric analysis through a flexible, transparent interface combined
␾c := 0.85
with extensive computing power. The following benefits need
to be emphasized
Geometry Data (Fig. 1)
Length of top section (reference section)
• Ability to incorporate the actual engineering units in the
calculation Ltop := 4 m
• Ability to see the formulas on the worksheet in the exact
form as they may appear in a textbook or a code Length of bottom section
• Ability to import and export data to external files Lbot := 8 m
Effective length of top section for buckling about weak axis
These characteristics of Mathcad allow students, teachers,
and engineers to develop sophisticated computer models that L1y := 4 m
solve engineering problems in scientific notation. Complete
Effective length of bottom section for buckling about weak
solutions to engineering problems can be developed for a axis
homework problem or to simply check the results of other
programs. Mathcad allows students and teachers to focus on L 2y := 4 m
the engineering theory and procedures while letting the soft-
ware handle the calculations along with the units of measure. Maximum length of laterally unbraced segment in top portion
As the engineering profession continues to move toward au- Lbtop := 4 m
tomated design software, engineers must find better tools that
they can use to verify engineering calculations. Engineering Maximum length of laterally unbraced segment in bottom por-
graduates who possess portfolios of working engineering mod- tion
els, produced in Mathcad as part of their college course work, Lbbot := 8 m
will be much better prepared to enter the automated paradigm
now present in the engineering profession. Rotational restraint spring constant
276 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000
 k1 := 0.0 ⭈ 1020 KN ⭈ m Q4i := 12 ⭈ ␮ ⭈ ␤ ⭈ (240 ⭈ ␤6 ⫺ 840 ⭈ ␤5 ⫹ 756 ⭈ ␤4 ⫹ 210 ⭈ ␤3
Total length of column ⫺ 420 ⭈ ␤2 ⫹ 105)
L := Ltop ⫹ Lbot Q5i := ␮ ⭈ ␤2 ⭈ ((⫺1,920 ⭈ ␤5 ⫹ 7,560 ⭈ ␤4 ⫺ 9,576 ⭈ ␤3 ⫹ 2,940 ⭈ ␤2
Fraction of column that is modified ⫹ 2,100 ⭈ ␤ ⫺ 1,260))
Lbot Q6i := ␮ ⭈ ␤3 ⭈ (320 ⭈ ␤4 ⫺ 1,400 ⭈ ␤3 ⫹ 2,247 ⭈ ␤2 ⫺ 1,575 ⭈ ␤ ⫹ 420)
␤ :=
L
The equivalent effective length factor for the top portion of
Derived variable the column [(11)]
D := k1 ⭈ L
␲ ⭈ 兹(Si)2 ⭈ (Q4i ⫹ 612) ⫹ D ⭈ Si ⭈ (Q5i ⫺ 156) ⫹ D2 ⭈ (Q6i ⫹ 12)
Kmtopi :=
Loading Data 兹252 ⭈ [(Si)2 ⭈ (Q1i ⫹ 24) ⫹ D ⭈ Si ⭈ (Q2i ⫺ 6) ⫹ D2 ⭈ (Q3i ⫹ 1)]

Factored compression force at top of column The equivalent effective length factor for the bottom portion
of the column [(13)]

冑
Putop := 80 KN
␣i
Factored compression force at the stepped level of column Kmboti := Kmtopi ⭈
(1 ⫹ ␮)
Pustp := 480 KN
Ratio of the compression forces
Note 5: The design of the column sections are in accor-
Pustp dance with the LRFD approach. Again, in the interest of short-
␮ :=
Putop ening this paper, the area is hidden.
Maximum factored bending moment in the top section
Mutop := 160 KN ⭈ m
Note 6: The combined force index is shown for the various
Maximum factored bending moment in the top section trial sections. The section is adequate only when both indices
are equal or less than 1.0.
Mubot := 220 KN ⭈ m
Moment modification factor for nonuniform moment (LRFD, Top portion of column
Eq. F1-3: top section
Cb 㛭 top := 1

Moment modification factor for nonuniform moment (LRFD,

Eq. F1-3): bottom section
Cb 㛭 bot := 1 Bottom portion of column
Note 3: One first needs to determine the effective length
factors for the top and bottom sections of the column. Before
doing that, one needs to have trial column sections. Mathcad
can be linked to section properties tables in which trial sections
can be picked. In this example three sets of trial sections were
picked and their section properties entered on the worksheet.
In the interest of space, the section is compressed (hidden). A For this example a top section of W10 ⫻ 33 and a bottom
double click on the area name in active Mathcad application section of W21 ⫻ 62 will be chosen.
will reveal the contents of the area.
APPENDIX II. REFERENCES
Agelidis, V. G. (1994). ‘‘Incorporating software tools in electrical engi-
Note 4: Using (3)–(13), the effective length factors are neering laboratory experiments—an example.’’ Proc., 1994 IEEE 1st
computed in the hidden area below. The clear formulas [(3)– Int. Conf. on Multi-Media Engrg. Educ., IEEE, Piscataway, N.J., 319–
(8), (11), and (13)] appear in the Mathcad worksheet as they 328.
would appear in a book or the code, clearly demonstrating the Agrawal, K. M., and Stafiej, A. P. (1980). ‘‘Calculation of effective
transparency of using this mathematical package. lengths of stepped columns.’’ Engrg. J., 17(4).
AISC. (1995). Manual of steel construction—load and resistance factor
design, Chicago.
Anderson, J. P., and Woodward, J. H. (1972). ‘‘Calculation of effective
lengths and effective slenderness ratios of stepped columns.’’ Engrg.
Derived values as per theory presented in this paper [(3)–(8)] J., 4th Quarter, 157–166.
Bendapudi, K. V. (1994). ‘‘Practical approaches in mill building columns
Q1i := (144 ⭈ ␤2 ⫺ 360 ⭈ ␤ ⫹ 240) ⭈ ␤3 ⭈ (␣i ⫺ 1) subjected to heavy crane loads.’’ Engrg. J., 31(4), 125–140.
Coates, R. C., Coutie, M. G., and Kong, F. K. (1988). Structural analysis,
Q2i := (⫺96 ⭈ ␤3 ⫹ 270 ⭈ ␤2 ⫺ 240 ⭈ ␤ ⫹ 60) ⭈ (␣i ⫺ 1) Van Nostrand Reinhold, New York, 331–335.
DeLyser, R. R. (1996). ‘‘Using Mathcad in electromagnetics education.’’
Q3i := (16 ⭈ ␤4 ⫺ 50 ⭈ ␤3 ⫹ 55 ⭈ ␤2 ⫺ 25 ⭈ ␤ ⫹ 5) ⭈ ␤ ⭈ (␣i ⫺ 1) IEEE Trans. on Educ., 39(2), 198–210.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000 / 277

Fisher, J. M. (1993). Industrial buildings: Roofs to column anchorage, Pavlov, A. V. (1996). ‘‘Florida Department of Transportation’s MastArm
Steel Des. Guide Ser. No. 7, AISC, Chicago. Program: Placing the engineer in control.’’ Proc., 3rd Congr. on Com-
Fraser, D. J. (1990). ‘‘The in-plane stability of a frame containing pin- puting in Civ. Engrg., ASCE, New York, 473–479.
based stepped column.’’ Engrg. J., 27(2), 49–53.
Gifford, J. B. (1987). ‘‘Microcomputers in civil engineering: Use and APPENDIX II. NOTATION
misuse.’’ J. Comp. in Civ. Engrg., ASCE, 1(1), 61–68.
Lui, E. M., and Sun, M. (1995). ‘‘Effective lengths of uniform and The following symbols are used in this paper:
stepped crane columns.’’ Engrg. J., 32(3), 98–106.
Magner, T. (1995). ‘‘Automated structural design calculations using K1 = effective length factor for upper column segment;
Mathcad.’’ Modern Steel Constr., 35(4), 42–47. K2 = effective length factor for lower column segment;
Mathcad: User’s guide. (1991). Mathsoft, Inc., Cambridge, Mass.
kR = stiffness of rotation spring (restraint) at bottom end of col-
Myers, G. E. (1992). ‘‘Equilibrium via element potentials and comput-
ers.’’ Thermodynamics and the design, analysis, and improvement of umn;
energy systems, Vol. 27, ASME, New York, 205–212. ␣ = ratio of moment of inertia of bottom section to that of top
Pavlov, A., and Mtenga, P. V. (1997). ‘‘Application of MathCAD in bridge section;
design.’’ Proc., 4th Int. Congr. on Computing in Civ. Engrg., ASCE, ␤ = ratio of bottom segment to total length of column; and
New York, 666–671. ␮ = ratio of load at stepped level to load at top.

278 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / OCTOBER 2000