Académique Documents
Professionnel Documents
Culture Documents
Automation in Construction
journal homepage: www.elsevier.com/locate/autcon
ukurova niversitesi, Mhendislik Mimarlk Fakltesi, Bilgisayar Mhendislii Blm, Balcal, Adana, Turkey
ukurova niversitesi, Mhendislik Mimarlk Fakltesi, naat Mhendislii Blm, Balcal, Adana, Turkey
ukurova niversitesi, Mhendislik Mimarlk Fakltesi, Elektrik Elektronik Mhendislii Blm, Balcal, Adana, Turkey
a r t i c l e
i n f o
Article history:
Accepted 16 September 2011
Available online 24 October 2011
Keywords:
Supervised learning
Unsupervised learning
Self Organizing Maps
Construction crew
Productivity
a b s t r a c t
Complex variability is a signicant problem in predicting construction crew productivity. Neural Networks
using supervised learning methods like Feed Forward Back Propagation (FFBP) and General Regression Neural Networks (GRNN) have proved to be more advantageous than statistical methods like multiple regression,
considering factors like the modelling ease and the prediction accuracy. Neural Networks using unsupervised
learning like Self Organizing Maps (SOM) have additionally been introduced as methods that overcome some
of the weaknesses of supervised learning methods through their clustering ability. The objective of this article
is thus to compare the performances of FFBP, GRNN and SOM in predicting construction crew productivity.
Related data has been collected from 117 plastering crews through a systematized time study and comparison of prediction performances of the three methods showed that SOM have a superior performance in predicting plastering crew productivity.
2011 Elsevier B.V. All rights reserved.
1. Introduction
Realistic project scheduling is one of the vital issues for successful
completion of construction projects and this can only be achieved if
schedules are based on realistic man-hour values. Yet, determination
of realistic man-hour values has been a complicated issue due to the
complex variability of construction labor productivity [19]. Thus,
recent researches have focused on articial neural network applications which provide a exible environment to deal with such kind
of variability. These applications have been based on supervised
learning methods, primarily Feed Forward Back Propagation (FFBP)
[12,1018], and recently General Regression Neural Network
(GRNN) [1920]. While strengths of these methods over multiple regression models, related with the modeling ease and the prediction
accuracy, have been well discussed; the weaknesses of the supervised
learning process, i.e. requiring the output vector to be known for
training, has also been pointed out [2122]. In parallel, Self Organizing Maps (SOM), based on unsupervised learning, have been introduced as applications which overcome the weaknesses of both the
statistical methods and the neural network applications based on supervised learning [2125]. However, few researchers like Hwa and
Miklas [26], Du et al. [27] and Mochnache et al. [28] used SOM for prediction purposes and these were related with heavy metal removal
performance, oil temperature of transformers and thermal aging of
transformer oil, respectively. A recent application, alternatively, has
Corresponding author. Tel.: + 90 322 338 60 84/2663 17; fax: + 90 322 3386702.
E-mail addresses: moral@cu.edu.tr (M. Oral), eoral@cu.edu.tr (E.L. Oral),
aaydin@cu.edu.tr (A. Aydn).
0926-5805/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.autcon.2011.09.002
focused on prediction of construction crew man-hour values for concrete, reinforcement and formwork crews [29] and prediction results
have been compared with the results of previous research based on
both multi regression analysis and Feed Forward Back Propagation
(FFBP) [2,30]. The objective of the current research, however, has
been to use a specic sample data and compare the prediction results
of the models developed by using SOM, FFBP and GRNN.
2. Data collection and nature of the data
Collecting realistic and consistent productivity related data is one
of the key factors in arriving at realistic man-hour estimates by using
any of the prediction methods. Various work, labor and site related
factors affect construction labor/crew productivity and these have to
be observed and analyzed systematically in order to arrive at realistic
man-hour values, and time study is a methodical process of directly
observing and measuring work [31]. Thus, time study has been undertaken with 1181 construction crews in Turkey through the use of
standard time study sheets between the years 2006 and 2008 and details related with concrete pouring, reinforcement and formwork
crews have been presented in various publications [17,29,32]. For
plastering crews, quantity and details of plastering work undertaken
by each crew were recorded together with work (location of the
site, location of the work on site, the type and the size of the material
used and the weather conditions), labor (age, education, experience,
working hours, payment method, absenteeism and crew size), and
site (site congestion, transport distances, and the availability of the;
crew, machinery, materials, equipments and site management) related factors for 31 crews initially. Man-hour values were then
272
where:
N'
A
N
Xi
n
Input
Layer
Output
Layer
Experience
Crew Size
Table 1
Distribution characteristics of productivity values for plastering crews.
Hidden
Layers
Productivity
Age
Plastering (mh/m2)
Mean productivity
Standard deviation
Coefcient of Variation
Skewness coefcient
Kurtosis coefcient
0.5
0.23
0.47
1.08
2.53
Fig. 1. The architecture a FFBP Neural Network with two hidden layers.
Input
Layer
Hidden
Layer
Experience
Pattern
Layer
273
Input Nodes
Decision
Layer
Weights
Output Nodes
Crew Size
Age
Productivity
Experience
Crew Size
Fig. 2. Schematic diagram of GRNN.
Age
the spread of RBF kernel function. The output value of a hidden neuron is passed to the two neurons in the pattern layer where one neuron is the denominator summation unit and the other is the
numerator summation unit. The denominator summation unit adds
up the weight values coming from each of the hidden neurons and
the numerator summation unit adds up the weight values multiplied
by the actual target value for each hidden neuron. The decision layer
then divides the value accumulated in the numerator summation unit
by the value in the denominator summation unit and uses the result
as the predicted target value [4950]. The prediction performance of
a GRNN strongly relies on the sigma value. It is the only parameter
that needs to be tuned by the user.
3.3. Self Organizing Maps (SOM)
Unlike FFBP and GRNN, SOM uses groups of similar instances, i.e.
input/output patterns, instead of a predened classication, i.e.
pairs consisting of an input and the desired output for training. The
training data set contains only input variables and SOM attempts to
learn the structure of the data in order to arrive at solutions. It organizes and clusters unknown data into groups of similar patterns
according to a similarity criterion (e.g. Euclidean distance) resulting
in reduction in the amount of data and organization of the data on a
low dimensional display. Thus, when the process is complete, each
node in the Output Layer has a topological position where similar
clusters position close to each other [51]. Fig. 3 shows the current
problem as an example, which adapts 4 dimensional input vector
into a 6 6 dimensional map.
Contrary to its common usage for; clustering, classication or displaying multidimensional data, SOM can be modied to make accurate predictions from which the data model is unknown. A typical
SOM has an input vector which has as many dimensions as the number of independent parameters, and no target vector. By feeding experience of the crew on the particular site, crew size, age of the
crew members and productivity as input vector, the relationships between these attributes can be displayed in two dimensional maps
produced by SOM. The maps are, in fact, visual representations of
the weight values that connect input and output nodes. The weight
values originating from crew size input node, for example, form
crew size map. As seen in Fig. 3, the output nodes do not produce
any output. Any output has to be derived from the weights. During
the training phase of SOM, similar input instances are approximated
and grouped together topologically. While some of the output nodes
correspond to the input vectors of the data set, some of them are
the estimation of the possible vector instances that lay between
these input vectors. This feature allows SOM to predict the outcome
of the instances that is not in the training data set. For the prediction
of crew productivity, FFBP and GRNN require input vectors that have
three components; experience, crew size and age, and target vectors
Productivity
274
Table 2
Validation results for FFBP for different transfer functions.
FFBP
Conguration
TSig
PLin
PLinPLin
PLinTSig
TSigPLin
TSigTSig
Nodes
MSE
MAE
MAPE
7
0.07
0.2072
49.99
9
0.0614
0.1896
46.83
7-6
0.0593
0.1882
45.97a
10-9
0.0699
0.1988
46.61
6-6
0.0699
0.201
48.31
6-5
0.0672
0.2027
47.37
Best result.
MSE
1 n
2
A Pi
n i1 i
MAE
1 n
jA Pi j
n i1 i
MAPE
CV
1 n Ai Pi
100%
n i1 Ai
100%
where;
n
Ai
Pi
neurons in the hidden layers and the transfer functions were changed as in the rst structure. The best result was obtained as 45.97,
0.0593 and 0.1882 for MAPE, MSE and MAE, respectively when
transfer functions in both layers were set to Purelinear with the
number of neurons equal to seven for the rst layer and six for the
second layer (Table 2).
The GRNN algorithm, on the other hand, needs only one parameter to be tuned up; spread of radial basis function (). parameter
was changed from 0 to 5 with 0.01 steps in order to seek the most
suitable value to achieve the best prediction performance. The best
result was obtained when was 0.3, and MAPE, MAE and MSE values
were 45.87, 0.117 and 0.023, respectively (Table 3).
SOM, however, requires many parameters to be tuned up; output
map size, initial learning rate, initial neighborhood radius, even seed
of random generator. Instead of using systematic parameter search
as in FFBP and GRNN, random twelve different congurations were
tested for SOM. The best results were obtained with output map
size of 50 by 50 and initial learning rate of 0.1. It was observed that
initial neighborhood radius has a signicant effect on the prediction
accuracy. Further experiments were carried on by changing initial
neighborhood radius from 70 to 300 with step size 1. Fig. 4 shows
the effect of initial neighborhood radius on the prediction accuracy.
The best result was obtained when the radius was set to 237. MAPE,
MAE and MSE values were obtained as 41.27, 0.193 and 0.069 respectively, with these parameters (Table 3).
Results in Table 3 show that the prediction accuracy of SOM is
superior to GRNN and FFBP with respect to MAPE values. Best and
Worst rows in Table 3 which display the best and the worst
Table 3
Validation results.
GRNN
Fold 1
Fold 2
Fold 3
Fold 4
Fold 5
Fold 6
Fold 7
Fold 8
Fold 9
Best
Worst
.
CV (%)
a
Best result.
FFBP (PLinPLin)
SOM
MSE
MAE
MAPE
MSE
MAE
MAPE
MSE
MAE
MAPE
0.022
0.013
0.019
0.020
0.055
0.036
0.019
0.011
0.012
0.011
0.055
0.023
0.014
61
0.118
0.100
0.112
0.120
0.153
0.166
0.116
0.084
0.080
0.080
0.166
0.117
0.029
25
47.85
31.15
82.12
51.07
33.59
43.88
51.74
39.91
31.56
31.15
82.12
45.87
15.78
34
0.066
0.031
0.053
0.057
0.129
0.089
0.043
0.028
0.040
0.028
0.129
0.059
0.032
54
0.175
0.155
0.199
0.204
0.234
0.264
0.170
0.134
0.158
0.134
0.264
0.1882
0.0412
22
36.87
31.02
85.50
59.52
31.63
44.71
47.50
39.74
37.29
31.02
85.50
45.97
17.23
38
0.068
0.028
0.040
0.070
0.183
0.120
0.046
0.021
0.051
0.021
0.183
0.07
0.051
74
0.199
0.123
0.168
0.212
0.298
0.278
0.158
0.120
0.182
0.120
0.298
0.193
0.062
32
42.93
21.25
66.12
49.96
39.52
40.59
43.50
29.30
38.23
21.25
66.12
41.27a
12.53
30
275
Table 4
Sensitivity analysis.
FFBP
GRNN
SOM
10-8
0.0725
0.2057
46.98
0.0586
0.1879
46.50
0.027
0.120
41.41a
5-5
0.0658
0.1929
45.82
6-5
0.0592
0.1872
44.50
0.0588
0.1857
45.67
0.0761
0.2085
45.82
9-7
0.0619
0.1867
43.56
5-6
0.0661
0.1968
46.97
10-5
0.0652
0.1923
45.46
0.0587
0.1871
46.06
0.0660
0.043
40.69a
8-8
0.0608
0.1831
42.79a
7-10
0.058
0.1833
44.52
8-10
0.0828
0.2115
47.79
9-8
0.0745
0.1953
43.20
0.0586
0.1874
46.40
0.0651
0.1860
43.96
8-8
0.0613
0.1856
43.68a
7-8
0.0681
0.1934
43.94
7-6
0.0654
0.1954
44.61
9-9
0.0622
0.1888
44.40
0.0585
0.187
46.02
0.0645
0.1939
44.56
8-8
0.0605
0.1825
42.86
9-7
0.0602
0.1863
44.71
10-5
0.0614
0.1872
44.83
6-9
0.0616
0.1892
43.51
0.0585
0.1866
45.70
0.0581
0.1858
40.54a
Conguration
PLinTSig
TSigPLin
TSigTSig
8-5
0.0791
0.2073
45.91
9-5
0.0636
0.1955
44.86
5-9
0.0605
0.1865
44.63a
PLin
TSig
PLinPLin
Best result.
Table 5
Sensitivity results of SOM.
Independent variable
Age/experience/crew size
Age/experience
Age/crew size
Experience/crew size
Age
Experience
Crew size
a
Best result.
MSE
MAE
MAPE
CV (%)
CV (%)
CV (%)
0.07
0.0761
0.0680
0.0660
0.0651
0.0645
0.0581
0.051
0.0437
0.0400
0.0304
0.0501
0.0345
0.0320
74
57.4244
58.8235
46.0606
76.9585
53.4884
55.0775
0.193
0.2085
0.1889
0.1926
0.1860
0.1939
0.1858
0.062
0.0551
0.0443
0.043
0.0772
0.0614
0.0518
32
26.4267
23.4516
22.3261
41.5054
31.6658
27.8794
41.27
45.82
43.41
40.69
43.96
44.56
40.54a
12.53
15.70
19.21
12.49
10.93
9.58
11.77
30
44.2525
30.6955
24.8635
21.4991
29.0331
30.3853
276
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
Acknowledgment
Data used in this paper has been collected during the research
project 106M055 which is supported by TBTAK (The Scientic
and Technical Research Council of Turkey). The authors would like
to thank G. Mstkoglu, E. Erdis, E.M. Ocal, and O. Paydak for their invaluable support during data collection.
[29]
[30]
[31]
[32]
[33]
References
[34]
[1] M. Radosavljevi, R.M. Horner, The evidence of complex variability in construction
labour productivity, Construction Management and Economics 20 (1) (2002) 312.
[2] R. Snmez, J.E. Rowings, Construction labor productivity modelling with neural
networks, Journl of Construction Engrg. and Mgmt. 124 (6) (1998) 498504.
[3] A. Kazaz, S. Ulubeyli, A different approach to construction labour in Turkey: comparative productivity analysis, Building and Environment 39 (2004) 93100.
[4] R. Lane, G. Goodman, Wicked Problems: Righteous Solutions Back to the Future
on Large Complex Projects, IGLC 8, Brighton, UK, 2000.
[5] S. Bertelsen, L. Koskela, Managing the Three Aspects of Production in Construction, IGLC-10, Gramado, Brazil, 2002.
[6] S. Bertelsen, ComplexityConstruction in a New Perspective, ILGC-11, Blacksburg, Virginia, 2003.
[7] S. Emmit, S. Bertelsen, A. Dam, BygLOKA Danish Experiement on Cooperation in
Construction, IGLC-12, Elsimore, Denmark, 2004.
[8] L. Koskela, G.A. Howell, The Underlying Theory of Project Management is Obsolete, Project Management Institute, 2002.
[9] G.A. Howell, G. Ballard, I.D. Tommelein, L. Koskela, Discussion of Reducing variability to improve performance as a lean construction principle, Journal of Construction Engineering Management 130 (2) (2004) 299300.
[10] L. Chao, M.J. Skibniewski, Estimating construction productivity: neural-networkbased approach, Journal of Computing in Civil Engineering 2 (1994) 234251.
[11] A.S. Ezeldin, L.M. Sharar, Neural networks for estimating the productivity of concreting
activities, Journal of Construction Engineering Management 132 (2006) 650656.
[12] J. Portas, S. AbouRizk, Neural network model for estimating construction productivity, Journal of Construction Engineering Management 123 (4) (1997) 399410.
[13] H.Y. Ersoz, Discussion of neural network model for estimating construction productivity, 125 (3), 1999, pp. 211212.
[14] S. AbouRizk, P. Knowles, U.R. Hermann, Estimating labor production rates for industrial construction activities, Journal of Construction Engineering Management
127 (6) (2001) 502511.
[15] C.O. Seung, K.S. Sunil, Construction equipment productivity estimation using articial neural network model, Construction Management and Economics 24 (2006)
10291044.
[16] C.M. Tam, T. Tong, S. Tse, Articial neural networks model for predicting excavator productivity, Engineering Construction and Architectural Management
(2002) 446452 5/6.
[17] E. Oral (Laptal), E. Erdi, G. Mstkolu, Kalp ilerinde ekip prollerinin verimlilie etkileri, 4. naat Ynetimi Kongresi, stanbul, 3031 Ekim, 2007.
[18] R. Noori, A. Khakpour, B. Omidvar, A. Farokhnia, Comparison of ANN and principal
component analysis-multivariate linear regression models for predicting the
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]