See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225718257

The approximate cost estimating model for railway bridge project in the
planning phase using CBR method

Article  in  KSCE Journal of Civil Engineering · September 2011

DOI: 10.1007/s12205-011-1342-2

CITATIONS READS

18 6,735

1 author:

Byung Soo Kim

Kyungpook National University
34 PUBLICATIONS   143 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Development of environment impact assessment model for social overhead capital View project

Nothing View project

All content following this page was uploaded by Byung Soo Kim on 10 August 2017.

The user has requested enhancement of the downloaded file.

KSCE Journal of Civil Engineering (2011) 15(7):1149-1159 Construction Management
DOI 10.1007/s12205-011-1342-2
www.springer.com/12205

The Approximate Cost Estimating Model for Railway Bridge Project

in the Planning Phase Using CBR Method
Byung Soo Kim*
Received October 10, 2010/Revised December 23, 2010/Accepted January 3, 2011

···································································································································································································································

Abstract

The budget estimated in the planning phase of construction has to be accurate to be effective until after the design phase. Therefore
the special method needs to improve the accuracy of estimating construction cost in the planning phase. Recently, to improve the
accuracy of estimating construction cost for road bridges and architectural facilities, Case-based Reasoning (CBR) which makes use
of Genetic Algorithm (GA) and Multiple Regression Analysis (MRA) and other principles are being used. This study suggests the
cost estimation model which uses CBR and makes the database reflected the character of the railroad bridge. To establish this
effective CBR model, this study examined combinations of attributes, criteria of similarities and retrieval ranks and applied GA for
an optimization of attribute weights throughout learning process. According to the results of this study, CBR should use a five-
attribute combination, retrieve six similar cases and use the similarity criteria by applying a method in which ten points are deducted
for each 10% difference to get the minimum error. According to the verification results of the CBR model, the mean absolute error
rate is 11.9% and the standard deviation is 12.7%. Therefore, the accuracy of this CBR model is 30% more effective than the criteria
of Ministry of Construction & Transportation (MOCT) and Korea Development Institute (KDI) model.
Keywords: the railroad bridge, accuracy, estimating construction cost, CBR, the planning phase
···································································································································································································································

1. Introduction retrieve rank of similar cases, and has a weak point of having
degraded accuracy in case there lacks similarity between col-
1.1 Background and Purpose lected cases and a new case.
Recently construction projects are becoming increasingly and This study intend to present a construction cost estimation
many of them are simultaneously progressing. In addition, as model with improved accuracy by reflecting characteristics of
necessity for the railroad facility which is relatively insufficient the railroad bridges to build a database and by complementing
is increasing, railroad construction projects are also steadily in- the weak point of CBR characteristics.
creasing in number. In case many projects are proceeded,
accuracy of the budget allocated at planning and design phase is 1.2 Scope and Method
an essential factor for efficient use of the budget. But, at planning For this study, the scope was set as estimation of direct
phase, accuracy must be poor due to limited information with construction cost of the railroad bridges at planning phase of
which construction cost can be estimated. Accordingly, a method which the frequency of construction is recently increasing, and a
to improve accuracy of the approximation cost estimated at CBR database was built by collecting data for 123 bridges from
planning phase is required. For this, many researchers have design companies.
studied models to improve accuracy of construction cost budget. This study is based on CBR technique which deduces similar
Genetic Algorithms (GA), Case Base Reasoning (CBR) and cases from case bases and uses them to find solutions. That is to
Multiple Regression Analysis (MRA) using correlation between say, it is used impact factors of construction cost as a method to
impact factors of construction cost are techniques valuably used find similarity with existing cases and, for calculation of weight
to estimate construction cost. Among those, CBR technique of each impact factor, it is utilized GA technique on which natural
which is recognized to be more effective than other techniques is evolution principle is grafted. Also, it is used learning process to
widely used recently. search for the cases of the smallest error rate of estimated con-
However, CBR has a characteristic of showing different levels struction cost and, in this process it is reviewed impact factor com-
of accuracy depending on the method of selecting impact factors bination, similarity evaluation method and retrieves rank of similar
and assigning attribute weight, similarity evaluation method and cases. This paper verified effectiveness and validity of approxima-

*Member, Associate Professor, Dept. of Civil Engineering, Kyungpook National University, Sang-ju 742-711, Korea (E-mail: [email protected])

− 1149 −
 Byung Soo Kim

tion cost estimation model developed through such a methodology previous circumstances similar to the present one, and by using
using actual construction costs and mean absolute error. it, seeks solutions to new problems. In other words, CBR tries to
find out a solution to a new problem by modifying the solution to
1.3 Studies on Existing Findings a previous problem. Originating from cognitive science, CBR
This paper studied findings related to estimation of construc- takes the position that events repeat themselves regularly and an
tion costs using CBR. Kim et al. (2010) said attribute weight effective countermeasure to one event can be also applied to
with minimum error rate differs depending on number of case another similar case (Schank, 1982; Gentner, 1983).
bases, criteria to decide similarity of attribute and criteria to To explain in more detail, CBR describes a new situation by
retrieve similar cases, presenting a construction cost estimation referring to a previous solution or it deduces precedent by as-
model for planning phase of PSC beam road bridges based on sessing a newly suggested solution in view of a previous case,
CBR and GA. Kim and Son et al. (2009) indicated that most of thereby making an interpretation of new a event/circumstance or
researchers are not presenting clear solution by executing proposing a balanced countermeasure to a new problem (Kim
calibration through utilization of experts in most of studies if the 2005).
construction cost error rate of the cases retrieved is big because
the similarity of the similar cases retrieved is low or there is 2.1.1 Circulation Structure of CBR
insufficient cases in the case base and, as a method to solve this Through processes of solving, understanding, learning and re-
problem, suggested a CBR calibration method using regression membering a problem, CBR provides an inference model which
model for estimation of construction cost. The object facilities integrates all these. Accordingly, the structure of CBR is consist-
were business facilities such as a post office and government ed of 4 circulation processes. That is to say, those are Retrieve
office building. Seo et al. (2010) enabled order placers to make which is a process of searching for past cases which best fit to the
decision related to design information in accordance with new problem from case bases, Reuse which is a process of
construction cost plan by checking the attribute weight rate, as a grasping and using the information required to solve the current
method to assign weight values to attributes when build a CBR problem, Revise which is the process of changing the solutions
model for public office buildings, through comparison of Feature of past cases to fit the new problem, and Retain which is a
Counting technique which applies equal weight of 1.0 to all attri- process of storing the solution obtained by revising as a case
butes and MRA technique which utilizes standardized coefficient after solving the new problem. The circulation process of CBR
(β) of Multiple Regression Analysis as attribute weight. Park et can be expressed as in Fig. 1.
al. (2010) introduced Generic Algorithms (GA) which is a
technique of solving problems through optimization to obtain
improved prediction result than the case of using standardized
regression coefficient or feature counting in order to complement
the problem in the existing deterministic method due to multiple
attribute variables and numerous solution sets in calculating
attribute weight of public apartments.
In the meantime, in foreign countries, Oberlender et al. (2001)
collected 67 cases of construction work data, which were ana-
lyzed to present 45 impact factors as well as a construction cost
prediction model utilizing Multiple Regression Analysis of the
result. Also, Willmot et al. (2003, 2006) developed construction
cost prediction model by collecting and analyzing highway in-
formation and, in his follow-up study, improved the accuracy of Fig. 1. Process of CBR
the model utilizing neural network.
When put together the existing studies, the objects were high- 2.1.2 Method to Retrieve Similar Cases
ways, road bridges, business facilities, public office buildings As typical method of CBR to retrieve similar cases, Inductive
and public apartments, and GA was utilized for optimization of Retrieval and Nearest Neighbor Retrieval are generally used
attribute weight on the basis of CBR in order to estimate con- (Park et al., 2010). Inductive retrieval is a method effective when
struction cost and regression model was used to complement the target of similar case retrieval is well defined and is a method
shortcomings of CBR. But, construction cost estimation model in which new cases are indexed by main impact factors for the
for the railroad bridges has never been attempted. result induced from past cases themselves (Lee, 1998). This
method deduces a classification method expressed in generalized
2. Background rules or formulas from the collected cases in advance and
searches for similar past cases by applying defined classification
2.1 Case-based Reasoning method when a new case is entered. Nearest Neighbor Retrieval
Case Based Reasoning (CBR) is a methodology that recalls is a method to retrieve similar cases among the stored cases by

− 1150 − KSCE Journal of Civil Engineering

 The Approximate Cost Estimating Model for Railway Bridge Project in the Planning Phase Using CBR Method

matching new cases to past cases in accordance with certain attribute weight changes depending on experts as it is based on
similarity criterion (Kim et al., 2004). subjective factors that are knowledge and experience of experts
(Ahn and Kang, 2005). Genetic Algorithm does not create
2.1.3 Similarity Evaluation problems of local minimum and collinearity as it searches for a
In order to retrieve cases similar with a new case from a case solution through trials and errors based on survival of the fittest
database, similarity evaluation should be performed. Factors for principle. Also, Genetic Algorithm can secure objectivity of the
evaluation of similarity are expressed in strings or figures, which deduced solution as it is a mathematical method (Park et al.,
are attributed, and methods of calculating similarity are differ- 2010). Moreover, it is used in diverse fields such as engineering,
ently applied in accordance with these attributes. Generally, the business and social science, etc. Table 2 shows the application
method to calculate similarity score in accordance with attribute fields of Genetic Algorithm.
is as shown in Table 1 and, for figure attribute, distance measure-
Table 2. Application Fields of Generic Algorithm (Chung, 2002)
ment principle is widely used, where, through a method as Eq.
(1), 100 score is given if cases match or have an error rate of Application
Application Examples
Fields
10% or lower while 0 score is given if cases do not match or
Optimization of numerical functions, Optimization of gas
have an error rate exceeding 10%.
pipelines, Optimization of power transmission networks,
Similarity point Optimization Optimal assignment of computer keyboard, Aircraft crew
assignment, Traveling salesman problem, Graph division
= Exist case value – New case value
--------------------------------------------------------------------------------- × 100 ≤ 10% (1) problem, and Interpretation of genetic information
New case value VLSI circuit design, aerodynamic design of aircraft
wings, Engine nozzle design, optimal design of computer
Design
Table 1. Similarity Point Criteria Each Attribute communication network, Cardiac pacemaker design, Dig-
ital filter design, and Fuzzy controller design
Property Similarity Scoring
Automatic generation of LISP program, Automatic acqui-
Character Granting a score only in the case in which the attributes of
Artificial sition of problem solving rules, Synthetic of neural net-
row the two cases being compared are completely identical
Intelligence work and learning, Pattern recognition, Natural language
Granting a score according to the ratio of being identical by processing, and Multi-agent system
Character comparing all the consecutive three characters available in
System identification, Prediction of chaotic time series,
one word System
Prediction of change in foreign exchange rate, Prediction
Granting a score according to the ratio of the number of Analysis and
and analysis problem in financial and economic fields,
Word words that the properties of the two cases being compared Prediction
and Protein structure analysis
have in common to the number of words of the new case
Control by combination of neural network and fuzzy logic
Granting a score based on the similarity determined by the Control and
Value with Genetic Algorithm, and Prediction of path of moving
distance between the two values Robotics
robots

2.1.4 Attribute Weight 2.2 Genetic Algorithm

Attribute weight is a method of assigning to an independent Genetic Algorithm (GA) has as its basis the principle of
variable a weight as big as the size of influence it has on depend- searching optimum solution by repeating the computing process
ent variables when retrieving cases similar with a new case from of Crossover, Mutation and Selection among the series of char-
a case database and uses diverse methods such as Feature acters that retain a structure similar to a chromosome (biological
Counting, Gradient Descent Method, Analytic Hierarchy Process, gene), imitating the genetic process of the natural realm. That is,
Regression Analysis and Genetic Algorithms. GA displays potential solutions as coded individual on computer,
Feature Counting is a method of assigning equal weights to all and forms a population by gathering several individuals. It then
attributes and has a weak point of not reflecting importance dif- crosses over their genetic information or produces new genetic
ference between attributes (Kim et al. 1996). Gradient Descent information over generations. Through such an imitated evolu-
Method is a method of determining the direction of search using tion process based on the principle of survival of the fittest, GA
gradient of function and of searching new points repeatedly at acts as a computing model that seeks the optimum solution to the
the present position. However, Gradient Descent Method has a given problem (Lee, 2000).
problem of unnecessarily performing search process in order to At genetic algorithm, each possible solution is deemed an
converge to a local minimum if the curvatures for different direc- individual, and a group of the individuals is called a population.
tions differ (Kim and Kang, 2004). Regression Analysis is a Normally an individual is composed of one or several chromo-
method of utilizing standardized coefficients deduced from re- somes, and the operators that transform chromosome are called
gression analysis as attribute weights. Though Regression Analysis Genetic Operators. Though difference exists to some extent
assumes linearity when estimating relation between dependent depending on the model of genetic algorithm, the overall flow of
variables and independent variables, the resulting values lack the algorithm can be illustrated as Fig. 2. Also the searching pro-
reliability as most of regression models do not fulfill the assump- cess of simple genetic algorithm for the optimum solution can be
tion (Kim, 2006). Analytic Hierarchy Process has a problem that displayed per stage like the following;

Vol. 15, No. 7 / September 2011 − 1151 −

 Byung Soo Kim

stagnate at local optima and fail to find the optimal global solu-
tion for certain starting solutions (Albright and Winston, 2007).
As a result, GA has been proven to be a versatile and effective
approach for solving nonlinear problems. It has been applied to
optimization problems such as constrained construction schedul-
ing (Leu and Yang, 1999), discrete (Koumousis and Georgiou,
1994) or mixed-discrete nonlinear optimization problems (Jenkins,
1997), site layout (Li and Love, 1998) and optimization of the
interrelationship between supply and demand points of tower
cranes (Tam et al., 2001).

3. Approximate Cost Estimating Model

The CBR process to estimate approximation cost for a railroad

Fig. 2. Simple Genetic Algorithm Flow Chart (Turban, 1995) bridge is as shown in Fig. 3. The process is largely divided into
learning step of reviewing elements required for model using
Step 1) forms the initial population P(0) that retains individuals learning cases, the step of deciding factors or methods of neces-
in M quantity. Generation g=0 sary elements representing minimum error rate through learning
Step 2) assesses fitness for all individuals included in the popula- and the step of estimating construction cost using the established
tion P(0). model. In the learning step, methods to search similar cases from
Step 3) stops implementation when an individual with a solution the case bases using learning cases are reviewed, and combina-
of the desired level is found. tion of impact factors, method of evaluating similarity and re-
Step 4) selects parent individuals from P(g) based on fitness, and trieve rank of similar cases are reviewed in order to minimize the
transforms them by genetic operators such as crossover or error rate of construction cost. Review of impact factor combi-
mutation, thereby creating new individuals in M quantity. nation will include impact factor combination obtained by solving
Step 5) assumes new population as P(g+1), and increases number the problem of collinearity.
of generation by one. g=g+1 The decision step is a step of deciding what have been
Step 6) goes back to earlier Step 2) and repeats the above process. reviewed in the learning step and impact factor combination
representing minimum error rate, similarity evaluation method
Previous approaches used experience, gradient search, fuzzy and retrieve rank are decided in this step. In this process, attribute
number, and AHP, among others. Gradient search methods may weights are optimized using Genetic Algorithm. In construction

Fig. 3. CBR Process for Approximate Cost Estimation

− 1152 − KSCE Journal of Civil Engineering

 The Approximate Cost Estimating Model for Railway Bridge Project in the Planning Phase Using CBR Method

cost estimating step, construction cost of the new case is Table 4. Example of Database
estimated using the model built in the learning and decision steps Cost
Attribute
to obtain estimated construction cost. In order to evaluate Case (million won)
effectiveness of the model, whether the accuracy is improved in A1 A2 A3 A4 A5 A6 A7 A8 Ci
comparison to other models is verified by comparing the error Br1 2 1 2 1 1 25.0 6.4 25.0 470
rate with the actual construction cost. Br2 2 1 1 1 7 25.0 7.2 175.0 1,965
Br3 2 1 2 1 2 23.0 8.5 46.0 620
Br4 2 1 1 1 3 25.0 9.0 75.0 899
3.1 Data Collection and Case Base Construction
Br5 2 1 1 1 10 25.0 10.0 250.0 4,038
The data of railroad bridges used in this study were break-
Br6 2 1 1 1 22 25.0 11.0 550.0 6,416
downs of construction cost, drawings and design reports of 123
Br7 2 1 2 1 2 25.0 12.0 50.0 1,623
railroad bridges collected from design companies. Though the
Br8 2 1 2 1 2 25.0 13.0 50.0 1,563
total number of cases investigated was 139, only 123 cases were
Br9 2 1 2 1 6 22.0 14.7 132.0 1,487
analyzed excluding 16 cases which have errors in breakdowns or
Br10 2 1 1 2 3 35.0 3.0 105.0 3,882
drawings. Among 123 cases, 7 cases were utilized as learning
Br11 3 1 2 2 1 36.0 4.0 36.0 991
cases and the data of 110 cases were utilized as a case database
Br12 2 1 2 2 5 46.8 6.4 234.0 3,467
leaving 6 cases as verification data. Table 4 shows an example of
Br13 1 1 2 4 13 25.0 13.0 325.0 5,713
database of bridge data collected and the analysis result of design
Br14 2 1 1 2 2 35.0 8.0 70.0 2,083
years is as shown in Fig. 4. Br15 2 1 2 2 1 25.0 9.0 25.0 700
Br16 3 2 2 2 5 25.0 10.5 125.0 2,844
Br17 3 2 1 2 6 25.0 10.5 150.0 1,865
Br18 2 1 3 2 9 11.7 10.5 105.0 15,357
. . . . .. .. . .. .. ..
.. .. .. .. . . .. . . .
Br110 1 1 2 46 6 40.0 14.0 240 6,441
Where, A1=Type of railroad, A2=Location, A3=Single/double-track, A4
=Superstructure type, A5=No. of spans, A6=Length of spans, A7=Height
of bridge, A8=Length of bridge.

Table 5. Numerical Status of Attribute

Type of
Attribute Numerical status
Fig. 4. Design Year of Database attribute
No. of spans, Length of
Numeric spans, Height of bridge, -
As the years when projects were executed were different, Length of bridge
construction cost coefficients were used to equalize the point in Express railroad 1
time when the construction cost is applied. As construction cost Kind of railway General railroad 2
coefficient, the construction cost index announced monthly by Urban railroad 3
Korea Institute Construction Technology (KICT) was used. Land 1
Location
Table 3 shows status of construction cost index for railroad con- Riverbed 2
struction. Single track 1
Single/double
Double track 2
Table 3. Construction Index for Railroad PSC Beam 1
Construction Construction PF Beam 2
Year Coefficient Year Coefficient
cost index cost index Steel Box 3
2010.04 140.1 1.00 2005.01 99.8 1.40 PSC Box 4
Not Simple
2010.01 136.2 1.03 2004.01 91.2 1.54 numeric Arch 5
2009.01 134.4 1.04 2003.01 85.3 1.64 Precom 6
2008.01 113.5 1.23 2002.01 76.3 1.84 Steel complex 7
Superstructure Rahmen 8
2007.01 104.8 1.34 2001.01 74.7 1.88 type
2006.01 100.9 1.39 2000.01 71.7 1.95 PSC Beam+PF Beam 12
PSC Beam+Steel Box 13
The concept of CBR is to search for the existing case which is PSC Beam+PF
123
Beam+Steel Box
most similar with the new case by evaluating similarity of Complex
PSC Beam+Steel
attributes between the new case and existing cases. In order to 134
Box+PSC Box
use such attributes, attributes should be put in order and entered PSC Beam+PF Beam
1234
in the database. +Steel Box+PSC Box

Vol. 15, No. 7 / September 2011 − 1153 −

 Byung Soo Kim

The program used for preparation of model was Microsoft Table 6. Status of Complex Bridge
Office Excel as numeric variables are more advantageous than Type of complex bridge Rate (%)
string variables in retrieving similarity of attributes. Among the PSC Beam + Steel Box 42.9
attributes which impact on construction cost, while there is no PSC Beam + PF Beam 20.4
problem with number of spans, length of spans, height of bridge PSC Beam + PF Beam + Steel Box 16.3
and length of bridge, nominal variables such as type of railroad, PSC Beam+ PSC Box 8.2
location, single/double-track and superstructure type should be
PSC Box + Precom 4.1
converted to numeric variables in a model. Table 5 shows the
PSC Beam + Steel Box + PSC Box 4.1
status of attribute variables expressed in numerical values.
PSC Beam + PF Beam + Steel Box + Arch 4.1

3.2 Characteristics of the Railroad Bridges

In order to estimate construction cost using limited information there are Investment Evaluation Guidelines Related to Public
at the initial phase, viz. planning phase, characteristics of infor- Facility Development Projects (MOCT, 2004) and Correction
mation should be well understood. In case of a road bridge, a Complementation Study on Standard Guidelines for Preliminary
result of construction cost estimation with accuracy to certain extent Feasibility Study of Road. Railroad Sector Projects (Fourth
can be achieved by building a database using factors which Edition) (KDI, 2004). The construction unit price (Investment
impact on construction cost and by running simple CBR which Evaluation Guidelines) per meter of main railroad in Table 7 is
searches for a case combination most similar with the new case presenting the unit prices by classifying railroads into single
through calculation of weight by impact factors using Genetic track, double track and double track subway and by classifying
Algorithm as in chapter 3. into general section, city center section and stream section (soft
ground) in accordance with the location of construction and is
providing composition ratio of PC beam and PF beam for pre-
paration of breakdowns by reflecting characteristics of railroad
construction which composes composite superstructure.
Standard cost (Standard Guidelines for Preliminary Feasibility
Study) for railroad construction per km in Table 8 presented the
construction cost per km by also classifying railroad into over-
bridge and city areas in addition to classifying it into single track
and double track. Here, special cities, metropolitan cities and
Fig. 5. Status Each Superstructure Type (Single bridge)
central areas of other cities are applied as city areas and remain-
ing area is applied as general area, and, special facilities (under-
In case of road bridges, when look at the models presented in ground facilities, long-span bridges) and soft ground are to be
existing studies, they are different from each other by type of separately added.
superstructure. However, in case of the railroad bridge, differ-
ently from the road bridge, there are many bridges which have Table 7. Construction Unit Cost of Major Railroads by Meter
several superstructures on top of one substructure. That is to say, (MOCT, 2004) (Unit : million won)
it can be explained as a bridge with one substructure and one Double-track
Bridge Single- Double-
superstructure versus a bridge consisted of one substructure and Unit electronic Remark
zone track track
railway
several complex superstructures. Fig. 5 shows the status of bridges
General PC beam :80%,
with a single superstructure among the railroad bridge data zone
m 10.8 19.6 19.6
PF beam:20%
collected.
Urban PC beam: 50%,
These single superstructure bridges cover 60.1% of the total. m 13.0 23.6 23.6
zone PF beam: 50%
PSC box type comprises 48.6% of single superstructure bridges River zone PC beam: 10%,
and is followed by PSC beam and PF beam showing 14.9%. m 17.4 31.7 31.7
(weak ground) PF beam: 90%
Table 6 shows the status of hybrid bridges with one substruc-
ture and several superstructures. Among hybrid bridges, bridges Table 8. Construction Unit Cost of Standard Railroad by Kilometer
consisted of PSC beam and steel box structure in combination (KDI, 2004) (Unit : 0.1billion won)
comprise the biggest portion of 42.9%. Also, there are bridges Double-track
consisted of PCS beam, PF beam and steel box structure in Bridge Single-track Double-track electronic
railway Remark
combination which cover 16.3%. zone
General Urban General Urban General Urban
3.3 Existing Construction Cost Estimation Model Over bridge 75 75 75 75 75 75
In case of domestic sector, as guidelines which can be utilized General zone 148 207 227 318 227 318
for calculation of approximation cost for the railroad bridges, Weak zone 184 258 283 396 283 396

− 1154 − KSCE Journal of Civil Engineering

 The Approximate Cost Estimating Model for Railway Bridge Project in the Planning Phase Using CBR Method

3.4 Attribute Variables and Weight Setup Table 10. Example of Error Rates by Impact Factor Combination
In order to decide combination of attribute variables and attri- and Retrieving Rank
bute weights, a combination with minimum error rate should be Number
Exception
Retrieving rank
of the
found through learning. Eight impact factors such as types of attribute
item 3 4 5 6 7 8 9 10 12 15
railroad which impact on construction cost are reviewed as attri- 8 - 16.1 14.9 10.3 6.0 8.3 9.0 11.7 14.6 24.0 28.0
bute variables to be used for CBR. Eight impact factors are type Single/dou-
20.4 17.6 11.0 9.7 9.8 10.7 12.8 16.6 16.1 21.6
of railroad, construction location, single/double track, supers- ble-Track
7
tructure type and number of spans, length of spans, heights of Superstruc-
ture Type 15.4 15.1 16.0 18.1 13.6 16.4 13.5 17.0 15.5 25.8
bridge and length of bridge. Though it is appropriate to compose
Superstruc-
combinations of attribute variables with only the factors with big ture Type
21.2 23.3 23.0 25.8 23.5 24.2 24.1 24.0 28.7 32.7
impact on construction cost, the bigger the number of the attri- Length of
bute variables, the more advantageous it is, considering the char- 6 Bridge
acteristics of CBR which searches for similar cases from case No. of spans
Length of 33.3 35.4 29.5 30.6 31.7 33.7 34.6 35.6 41.1 57.5
bases using attribute variables. In order to confirm this, this paper
Bridge
compared error rate of construction cost for each combination of
Superstruc-
attribute variable as shown in Table 10. The result was reviewed ture Type
changing the number of attribute variables to maximum 8 and Length of
Bridge 27.3 28.5 28.6 29.8 40.7 44.6 36.4 45.9 50.7 51.7
minimum 4, and, when reducing the number of attribute variables,
the items excluded were selected from items with low multiple Height of
Bridge
correlation coefficients as shown in Table 9. Also, the attribute 5
Superstruc-
variable combination obtained by solving collinearity problem in ture Type
multiple regression models was included. This combination has Height of
Bridge 39.9 39.6 27.5 28.1 24.4 21.6 25.7 28.9 27.4 69.6
6 attribute variables and is a combination from which number of
spans and length of bridge are excluded. The result was that, if Single/dou-
ble-Track
the number of attribute variables decreases, the error rate in Superstruc-
general increases as shown in Table 10. Though all the combina- ture Type
tions of attribute variables reviewed were not shown in Table 10, Length of
it shows that the case of using all 8 attribute variables has a lower Bridge
4 41.5 38.5 35.5 42.4 43.7 58.7 85.9 83.3 71.8 67.5
error rate than other cases. Height of
Bridge
Single/dou-
Table 9. Correlation of Cost Impact Factor ble-Track
Item A1 A2 A3 A4 A5 A6 A7 A8
Multiple
0.86 0.97 0.07 0.00 0.68 0.92 0.02 0.00
Regression
Where, A1=Type of railroad, A2=Location, A3=Single/double-track, A4 case base; (2) determination criteria of the similarity of attributes;
=Superstructure type, A5=No. of spans, A6=Length of spans, A7=Height
of bridge, A8=Length of bridge. and (3) retrieval criteria of similar cases with similar rankings
(Kim et al. 2010).
Accordingly, the error rate for each retrieve rank was reviewed
As the function of CBR is heavily affected by the result of cal- in Table 10 and, as shown in Table 13, attribute weight of the
culation of attribute weight, accurate calculation of attribute weight case with minimum error rate was calculated by reviewing error
is required for construction of a reliable CBR model (Dogan and rate in accordance with decision criteria for similarity of
Kartal, 2008). Accordingly, in order to calculate attribute weight, attributes. The weight for each attribute variable decided through
this paper used Generic Algorithm which has been verified to be such a process of optimization is as shown in Table 11. Attribute
effective in solving non-linear problems. This paper optimized weights decided in this way are used when similarity score is
the weight by using the error rate of the construction cost of calculated to search for cases similar to a new case.
learning cases from the database. The GA software program Table 11 presents the weight combinations of type 1 and type
used for optimization of weight was Premium Solver which is 2. Type 1 is the result of calculating weights of eight attributes
interfaced with an Excel environment. Fig. 6 is an execution and it shows that the weights of A1, A5 and A6 are very small.
screen of the captured GA program and set to minimize the error This means that three attributes have a negligible influence on
rate, and the range of weight change is between 0 and 1 including construction cost. Therefore, it should remove these three
both ends. And, the maximum iteration was set to 10,000 and the attributes and conduct the study again with only five attributes.
precision is one hundredth. The mutation is 0.25. Table 12 is the result of calculating the error rate with five and
But the weight of an attribute with minimum error changes eight attributes. Respectively, according to Table 12, the error
depends on the establishment of: (1) the number of cases in the rate of the study with five attributes is smaller than that of the

Vol. 15, No. 7 / September 2011 − 1155 −

 Byung Soo Kim

10th rank is most advantageous. The error rate for each retrieve
rank of similar cases increases from the 6th rank. This is well
shown in Table 10. That is to say, error rate increases when move
from the 6th rank to the 3rd rank and when move from the 6th
rank toward the 15th rank. Accordingly, the method of assigning
similarity score which have applied to this study is B method, 5
impact factors were applied and retrieve rank of similar cases
applied was up to the 6th rank.

Table 13. Similarity Criteria and Retrieve Rank

Error Retrieve Rank

Similarity
rate 5 impact factor 8 impact factor
criteria
(%) 3rank 6rank 10rank 3rank 6rank 10rank

Error: 0~10% : 100 Mean 27.0 55.2 68.6 21.5 12.7 39.3
A Error: more than Max. 82.1 115.8 208.7 89.4 43.2 78.9
11% : 0 Min. 0.0 15.5 0.0 6.8 0.0 9.3
Error: 0~10% : 100 Mean 16.3 5.3 15.4 16.1 5.9 14.6
Error: 11~20% : 90
B Error: 21~30% : 80 Max. 43.9 18.5 57.9 44.2 18.5 39.7
Fig. 6. GA Solver Program Error: more than
31% : 0 Min. 0.1 0.1 1.0 1.58 0.0 0.4
Table 11. Weight Each Attribute of CBR Model
Attributes A1 A2 A3 A4 A5 A6 A7 A8
The weights of similar cases retrieved in accordance with the
Type 1 0.0009 0.4299 0.1125 0.1444 0.0096 0.0002 0.0822 0.2204
result of Table 13 differ depending on similarity and the method
Type 2 - 0.5380 0.0771 0.1021 - - 0.0529 0.2299
to calculate weights for similar cases is as shown in Fig. 7 with
Table 12. Error Rate Compare of 5 and 8 Attributes
the examples shown in Table 15. Table 15 is the calculated
example of weight for the learning case number 4 bridge of
Number Retrieving rank
of the
Exception Table 14. When calculating a similarity point of case number 13
item 3 4 5 6 7 8 9 10 12 15
attribute in Table 15, the similarity point of A5 is 0.9. We obtain this result
8 - 16.1 14.9 10.3 6.0 8.3 9.0 11.7 14.6 24.0 28.0 0.18 by subtracting 11(the A5 value of case number 13 of Table
Type of 4) from 13(the A5 value of Table 14) then we divide this by 11.
Railroad This 18% error rate is 90 points according to the similarity
5
No. of
16.3 11.7 6.4 5.3 6.5 8.4 12.5 15.4 18.4 31.3 criteria B of Table 13. Therefore, when multiplying 90 points by
Spans the A5 weight of 0.0096, we get 0.9. Si is the sum of A1~A8
Length of weights and Ri presents the rank of weights.
Spans

study with eight attributes. So, based on this attribute variable

combination examination the use of five attributes is preferred.

3.5 Method to Assign Similarity Score and Retrieve Rank

Review
For evaluation of similarity, distance measurement principle is
used as introduced in section 2.1.3. That is to say, 100 score or 0
score is assign by comparing the impact factors of the new case
and existing case. However, this study minimized error rate of
the estimated construction cost by comparing two methods of
assigning similarity scores, A and B, as shown in Table 13 in
order to obtain more accurate value. Table 13 showed the
relation between the method of assigning similarity score and
error rate which Table 10 could not show. This paper found that,
as to the method of assigning similarity score, B method is
advantageous and, as to retrieve rank of similar cases, retrieving
up to the 6th rank by comparing cases from the 3rd rank to the Fig. 7. Weight Calculation Method of Similar Case

− 1156 − KSCE Journal of Civil Engineering

 The Approximate Cost Estimating Model for Railway Bridge Project in the Planning Phase Using CBR Method

Table 14. Status of Learning Case Number 4

Year correction coef- Construction
Br. Design Year A1 A2 A3 A4 A5 A6 A7 A8 Corrected Cost
ficient cost
4 2008.03 1 1 2 4 11 40 10 440 1.16 11,845,000,000 13,771,660,320

Table 15. Example of Similar Case Weight (Wi) similar case

Similarity point of attribute( Ij )
No. Si Ri RCi SSi Wi
A1 A2 A3 A4 A5 A6 A7 A8 3.7 Verification of Model
1 0.0 43.0 11.2 0.0 0.0 0.0 0.0 0.0 54.2 61 The effectiveness of CBR model was verified by comparing
2 0.0 43.0 0.0 0.0 0.0 0.0 6.6 0.0 49.6 84 the construction cost of a new case estimated in the CBR model
· · · · · · · · · · . of this study with the actual construction cost and by comparing
11 0.0 43.0 11.2 0.0 0.0 0.0 0.0 0.0 54.3 55
it with railroad construction cost estimation models of KDI and
MOCT, and then comparing the absolute error rate of each in
12 0.0 43.0 11.2 0.0 0.0 0.0 0.0 0.0 54.3 55
respect to actual construction cost.
13 0.1 43.0 11.2 14.4 0.9 0.0 0.0 17.6 87.3 5 5,713,846,554 87.3 0.140
As the result of verifying 6 new cases (Table 16 and 17) which
14 0.0 43.0 0.0 0.0 0.0 0.0 6.6 0.0 49.6 80
have been excluded earlier from the cases collected in order to
· · · · · · · · · · · verify effectiveness of the CBR model built in this study, the CBR
36 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 34 model presented by this study when compared with construction
37 0.1 43.0 11.2 14.4 0.9 0.0 0.0 19.8 89.5 2 10,422,633,510 89.5 0.144 cost in the final design statement showed an absolute error of
38 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 22 average 11.9% as shown in Table 18. It is a better error rate than
39 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 22 the result of applying unit price of the railroad bridge per meter
40 0.1 43.0 11.2 14.4 0.8 0.0 0.0 17.6 87.2 6 10,051,161,120 87.2 0.140 or km presented by Investment Evaluation Guidelines or Prelim-
41 0.1 43.0 11.2 14.4 1.0 0.0 0.0 22.0 91.8 1 15,267,998,592 91.8 0.148 inary Feasibility Study Guidelines and is within -50% and +30%
42 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 22 which is the scope of accuracy of estimated cost at planning
43 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 34 phase recommended by AACE (Association for the Advance-
44 0.1 43.0 11.2 14.4 0.8 0.0 0.0 17.6 87.2 6 18,719,010,138 87.2 0.140 ment of Engineering). It has a good standard deviation of 12.7%.
45 0.1 43.0 11.2 14.4 0.9 0.0 0.0 19.8 89.5 2 14,959,168,566 89.5 0.144 The error rates of Investment Evaluation Guidelines or Preliminary
46 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 34
Feasibility Study Guidelines are 44.7% and 41.2% respectively
and are a little out of the error rate scope recommended by AACE.
· · · · · · · · · · ·
Also, these show too big standard deviations and, especially,
50 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 22
have a weak point that the gaps between the maximum values
51 0.1 43.0 11.2 14.4 0.9 0.0 0.0 19.8 89.5 2 38,786,631,537 89.5 0.144
and minimum values are too big.
52 0.1 43.0 11.2 14.4 0.0 0.0 0.0 0.0 68.8 22
Where, A1=Type of railroad, A2=Location, A3=Single/double-track, A4 Table 16. Bridge Information of New Case for Verification
=Superstructure type, A5=No. of spans, A6=Length of spans, A7=Height
of bridge, A8=Length of Bridge, Si=Similarity points of cases, Ri=Ranks, Name of
A1 A2 A3 A4 A5 A6 A7 A8
RCi=Real coat, SSi=Similarity point of similar case, Wi=Weight of simi- bridge
lar case. Double Com-
Jungrichun General Land 21 33.7 9.0 708
track plex 1
Double
3.6 Estimation of Construction Cost for a New Case Bubgi Express Land
track
PSC Box 4 40.0 12.0 160
For estimation of construction cost for a new case, attribute
Single
variable combination (5 variables) decided through learning, Sanggae General Land PF beam 1 25.0 8.0 25
track
similarity evaluation method (B), attribute weight (Table 7) and Double
retrieve rank of similar cases (the 6th rank) are used. It will be Maryung Express Land PSC Box 1 30.0 5.0 30
track
executed by retrieving 6 cases with high similarity scores from Double Com-
Bookil General Land 43 27.9 14.0 1,198
case bases and by adding up construction cost for each weight. track plex 2
The construction cost for each weight (CWi) shall be calculated Namg- Double Com-
General Riverbed 32 30.7 7.0 983
by multiplying the weight (Wi) of retrieved similar case and real wang track plex 3
construction cost (RCi) as Eq. (2) and (3).
Table 17. Bridge Information of New Case for Verification
CWi = Wi × RCi (2) Number of
n Item Superstructure combination
combination
E cos t = ∑ CWi (3)
Complex 1 PSC Beam, PF Beam, Steel Box 3
i=1

Where, CWi: Cost for each weight, Wi : Weight of retrieved similar Complex 2 PSC Beam, PF Beam, Steel Box 3
case, RCi: Real cost, Ecost : Estimated cost, n: Retrieve ranks of Complex 3 PSC Beam, Steel Box 2

Vol. 15, No. 7 / September 2011 − 1157 −

 Byung Soo Kim

Table 18. Verification Result of Model (Unit : thousand won)

Ca CBR MOCT KDI
Real cost
se Estimated cost Mean absolute Error Estimated cost Mean absolute Error Estimated cost Mean absolute Error
1 19,976,151 19,123,260 4.3% 22,811,098 14.2% 26,418,976 32.3%
2 5,879,551 5,628,660 4.3% 5,150,685 12.4% 5,965,334 1.5%
3 906,241 910,539 0.5% 443,458 51.1% 607,702 32.9%
4 645,078 413,322 35.9% 965,753 49.7% 1,118,500 73.4%
5 55,399,434 52,797,691 4.7% 38,559,319 30.4% 44,657,987 19.4%
6 24,296,283 29,515,097 21.5% 51,180,174 110.7% 45,690,818 88.1%
Mean 11.9% 44.7% 41.2%
Standard deviation 12.7% 36.3% 33.0%
Maximum 35.9% 110.7% 88.1%
Minimum 0.5% 12.4% 1.5%

Advancement of Engineering), the CBR model presented in this

study is judged to be effective.
However, in spite of the effectiveness of the CBR model, this
paper judge that there is limit in the accuracy and reliability of
the CBR model developed as the case database has not been built
fully reflecting diverse superstructures of the railroad bridges.
Accordingly, continuous update of database for the railroad
bridges is required in the future.

Acknowledgements
Fig. 8. Comparison Result between CBR and Other Model
This research was supported by Kyungpook National Uni-
However, the 11.9% average error rate of verification cases is a versity Research Fund, 2008.
little higher than the 5.3% average error rate appeared during
learning process. It means that there are little cases similar with References
verification cases in the case basis in comparison to 7 leaning
cases or the similarity is low. Accordingly, CBR model needs to Ahn, S. H. and Kang, K. I. (2005). “A study on construction cost pre-
be built with sufficient and diverse cases in case bases. diction for pre-phase of apartment house using expert knowledge,
collection of learned papers (structural section).” Architectural
Institute of Korea, Vol. 21 No. 6, pp. 81-88.
4. Conclusions Albright, S. C. and Winston, W. L. (2007). Management science
modeling, Thomson South-Western, Mason, Ohio.
This study built a model which can predict approximation cost Bell, L. C. and Kaminsky, A. (1987). “Data base for preliminary cost
for construction of the railroad bridges using CBR model which estimating.” Journal of Transportation Engineering, Vol. 113, No. 4,
is a representative technique of predicting construction cost at pp. 341-347.
planning phase. In order to build an effective CBR model, this Carr, R. I. (1989). “Cost estimating principles.” Journal of Construction
paper reviewed the attribute variable combination, similarity Engineering and Management, Vol. 115, No. 4, pp. 545-551.
evaluation method and retrieve rank of similar case in the learning Chung, C. (2002). Optimization of sewer rehabilitation schedule and
budget by genetic algorithm, Master's Degree dissertation, Yonsei
process. Also, Genetic Algorithm was utilized to optimize attribute
University, p. 56.
weights. The result of this study showed that it is most advanta- Dogan, O. and Kartal, M. (2008). “SAR raw data simulation of targets
geous to use 5 impact factors for attribute variable combination, using three dimensional models.” EUSAR-European Conference,
and, as to similarity evaluation, to use a method of deducting 10 Vol. 7, No. 3, pp. 137-140.
scores per each 10% difference. And it showed that retrieving up Gentner, D. (1983). “Structure-mapping: A theoretical framework for
to the 6th rank is the method to obtain minimum error rate. analogy.” Cognitive Science, Vol. 7, No. 2, pp. 155-170.
The result of verifying CBR model showed average absolute Jenkins, W. M. (1997). “On the application of natural algorithm to
error rate of 11.9% and standard deviation of 12.7%, which means structural design optimization.” Engineering Structure, Vol. 194, pp.
302-308
it has about 30% effect of accuracy improvement in comparison
Kang, C. S., Lee, G. H., Kim, K. M., and Kim, K. J. (2008).
with the models of MOCT or KDI. As it is within -50% and “Development of approximate cost estimating model using case
+30% which is the accuracy scope of construction cost predicted based reasoning-centering around PSC beam bridge.” Proceeding,
at planning phase recommended by AACE (Association for the

− 1158 − KSCE Journal of Civil Engineering

 The Approximate Cost Estimating Model for Railway Bridge Project in the Planning Phase Using CBR Method

8th Conference, Korea Institute of Construction Engineering & Seoul, Korea, pp.74-84
Management. Lee, J. E. (1998). “Theoretical study and embody status of case-based
Karshenas, S. (1984). “Pre-design cost estimating method for multistory reasoning.” Sangmyung University Management Institute, No. 6, pp.
buildings.” Journal of Construction Engineering and Management, 1-31.
Vol. 110, No. 1, pp. 79-86 Leu, S. S. and Yang, C. H. (1999). “A genetic-algorithm-based resource
Kim, J. H., Kim, C. W., Kim, Y. H., and Park, J. W. (1996). “Retrieving constrained construction scheduling system.” Construction Mana-
of similar case using fuzzy concept at CBR, proceeding.” Korean gement Economic, Vol. 17, pp. 767-776.
Society for Expert System, Vol. 2, No. 1, pp. 153-160. Li, H. and Love, P. E. D. (1998). “Site-level facilities layout using genetic
Kim, K. H. (2009). A study on conceptual cost estimating using the case algorithms.” J. Comput. Civ. Eng., 124, 227-231.
based reasoning for IPC girder bridge, Master's Degree dissertation, Ministry of Construction & Transportation (2004). Investment assess-
KyungPook National University, Korea ment guideline for development project of public transport facility,
Kim, K. H. and Kang, K. I. (2004). A Study on estimating initial Korea.
construction cost for apartment house using case based reasoning, Ministry of Land, Transport & Maritime Affairs (2008). Road work
Collection of Learned Papers (Structural Section), Architectural handbook.
Institute of Korea. Oberlender, G. D. and Trost, S. M. (2001). “Predicting accuracy of early
Kim, K. H. and Kim, B. S. (2009). “An approximate cost estimation cost estimates based on estimate quality.” Journal of Construction
model for IPC girder bridge.” Journal of Research Institute of Engineering and Management, Vol. 127, No. 3, pp. 173-182.
Construction Technology, Kyungpook National University, Korea, Park, M., Seong, K., Lee, H., Ji, S., and Kim, S. (2010). “Schematic cost
Vol. 16, No. 1, pp. 117-141. estimation method using case-based reasoning : Focusing on de-
Kim, K. J. and Kim, K. M. (2010). “Preliminary cost estimation model termining attribute weight.” Korea Institute Construction Engineer-
using case-based reasoning and genetic algorithms.” ASCE Journal ing Management, Vol. 11, No. 4, pp. 22-31.
of Computing in Civil Engineering, Vol. 24, No. 6, pp. 499-505. Seo, J., Kim, S., Cho, J., and Chun, J. (2010). “Supporting model of cost
Kim, K. S. (2005). Prediction of foreign exchange crisis using case planning using case based reasoning in pre-design phase - Focused
based reasoning, Sungkyunkwan University Publishing Depart- on the case of public office building -.” Architectural Institute of
ment, Seoul, Korea. Korea, Vol. 26, No. 4, pp. 159-168.
Kim, T. G. (2006). Regression analysis for u-Can, the first edition, Schank, R. (1982). Dynamic memory: A theory of learning in computers
Human and Welfare, Seoul, pp. 53-88. and people, New York, Cambridge University Press.
Kim, Y. S., Son, M. J., Hyeon, C. T., and Hong, T. H. (2009). “A Turban, E. (1995). Decision support and expert systems management
division method using regression model in the division phase of cost support systems, Fourth Edition, Prentice Hall.
prediction model based on CBR, proceeding.” Korea Institute Tam, C. M., Tong, T. K. L., and Chan, W. K. W. (2001). “Genetic
Construction Engineering Management, No. 6, Track 2, pp. 291- algorithm for optimizing supply locations around tower crane.”
296. Journal of Construction Engineering and Management, Vol. 127,
Korea Development Institute (2004). Pre-feasibility study standard No. 4, pp. 315-321.
guideline for road·railway projects, Amendment Fourth Edition. Willmot, C. G. and Cheng, G. (2003). “Estimating future highway
Koumousis, V. K. and Georgiou, P. G. (1994). “Genetic algorithms in construction costs.” Journal of Construction Engineering and
discrete optimization of steel truss roofs.” Journal of Computing in Management, Vol. 129, No. 3, pp. 272-279.
Civil Engineering, Vol. 8, No. 3, pp. 309-325. Willmot, C. G. and Mei, B. (2005). “Neural network modeling of
Lee, C. K. (2000). Method to properly maintain bridge facility by fuzzy highway construction costs.” Journal of Construction Engineering
and genetic algorithm, PhD Dissertation, Chung-Ang University, and Management, Vol. 131, No. 7, pp. 765-771.

Vol. 15, No. 7 / September 2011 − 1159 −

View publication stats