Hindawi

Advances in Civil Engineering

Volume 2019, Article ID 5972620, 11 pages
https://doi.org/10.1155/2019/5972620

Research Article
Application of Artificial Neural Network(s) in Predicting
Formwork Labour Productivity

Sasan Golnaraghi, Zahra Zangenehmadar , Osama Moselhi, and Sabah Alkass

Department of Building, Civil and Environmental Engineering, Concordia University, Montréal, Canada

Correspondence should be addressed to Zahra Zangenehmadar; z_zange@encs.concordia.ca

Received 2 August 2018; Revised 5 October 2018; Accepted 19 November 2018; Published 2 January 2019

Guest Editor: Ali R. Vosoughi

Copyright © 2019 Sasan Golnaraghi et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Productivity is described as the quantitative measure between the number of resources used and the output produced, generally
referred to man-hours required to produce the final product in comparison to planned man-hours. Productivity is a key element
in determining the success and failure of any construction project. Construction as a labour-driven industry is a major contributor
to the gross domestic product of an economy and variations in labour productivity have a significant impact on the economy.
Attaining a holistic view of labour productivity is not an easy task because productivity is a function of manageable and un-
manageable factors. Compound irregularity is a significant issue in modeling construction labour productivity. Artificial Neural
Network (ANN) techniques that use supervised learning algorithms have proved to be more useful than statistical regression
techniques considering factors like modeling ease and prediction accuracy. In this study, the expected productivity considering
environmental and operational variables was modeled. Various ANN techniques were used including General Regression Neural
Network (GRNN), Backpropagation Neural Network (BNN), Radial Base Function Neural Network (RBFNN), and Adaptive
Neuro-Fuzzy Inference System (ANFIS) to compare their respective results in order to choose the best method for estimating
expected productivity. Results show that BNN outperforms other techniques for modeling construction labour productivity.

1. Introduction as a measure of output achieved by a combination of inputs.

Considering this perspective, two concepts for measuring
Artificial Intelligence (AI) has been a powerful tool in the productivity described in the literature are total factor
construction industry over the past decade. Several AI productivity and circulating capital [2]. Total factor pro-
modeling techniques have been employed in the con- ductivity is the most common construction productivity
struction industry such as expert systems (ES) and Artificial measurement technique where the output is measured
Neural Network (ANN). Modeling labour productivity is against all inputs. Partial factor productivity is referred to as
challenging as it requires the quantification of substantial single-factor productivity, in which the output is measured
factors that affect labour productivity and consideration of against a single input or selected inputs. Partial factor
influential factor interdependencies. Productivity is a deli- productivity is a cost-effective model and very advantageous
cate aspect of any construction project. Unquestionably, for developing strategy and assessing the state of the
arriving at a definition of construction productivity can economy; however, it is not beneficial for contractors [2].
cause confusion because of the various different ways at Circulating capital is any kind of capital that will be depleted
defining it. Strictly speaking, productivity is a component of during the course of a project, such as material and operating
cost and is not a method for estimating the cost of resources; expenses, whereas fixed capital refers to any kind of capital
rather, it is a quantitative assessment of the correlation that is not exhausted during the course of a project.
among the number of resources used and the amount of Productivity modeling has been a topic of interest for
output made [1]. Productivity in construction is considered many researchers and the various models being developed
2 Advances in Civil Engineering

today can be classified into two major groups: statistical and applied ANN for estimating the bricklayer (builder) pro-
AI. Regression analysis is the most common statistical ductivity and modeled 13 productivity-influencing factors.
method for modeling labour productivity. The main ad- El-Gohary et al. [16] proposed a framework to document,
vantage to regression analysis is that a productivity model control, and predict contractor labour productivity using
can be developed to reach anticipated clarification or ANN and hyperbolic tan function. They considered factors
forecasting levels with as few predictor variables as possible. at micro and macro/microlevels and applied the models to
However, for regression methods, the degree of relationship construction crafts, carpentry, and fixing reinforcing steel
(linear and nonlinear) needs to be selected prior to model bars.
development. In AI modeling, ANN models are the most A considerable issue in the construction industry is that a
common for developing labour construction productivity. lot of problems such as last-minute bids, design under
Unlike regression methods, the degree of relationship is not pressure, and so on are analogy-based in form. Thus, ANN
a concern in ANN modeling. Those studies that have applied techniques as compared to other conventional practices are
various ANN methods to predict different types of pro- more appropriate in modeling construction industry
ductivity are discussed in the next section. problems that demand analogy-based resolutions [17]. There
Lu et al. [3] estimated construction labour productivity are four major steps for modeling analogy-based problems
using real historical data from local construction companies. using ANN: (i) gathering historical data, (ii) building and
They applied a Probability Inference Neural Network configuring relevant network, (iii) initializing weights and
(PINN) model and compared it to a feed-forward back- biases, and (iv) training and validation step. In this research,
propagation neural network model. AbouRizk et al. [4] four types of ANN were applied for modeling labour pro-
developed a two-stage ANN model for predicting labour ductivity. These were Backpropagation Neural Network
productivity rates. They stated that understanding input (BNN), Radial Basis Network (RBF), Generalized Regression
factors and having a sufficient historical database are the Neural Network (GRNN), and Adaptive Neuro-Fuzzy In-
most important parts in productivity prediction. Later, [5] ference System (ANFIS). A detailed explanation of all the
introduced a neural network model for defining the impact applied modeling techniques will be discussed along with the
of a change order on labour productivity and found that the data collection procedure in the following sections.
ANN model shows better performance in comparison to
other techniques. Ezledin and Sharara [6] established an 2. ANN Model Development
ANN model for productivity prediction in formwork ac-
tivity, steel fixing, and concrete-pouring activities. Ok and Construction projects are highly dynamic with many
Sinha [7] applied ANN to estimate the daily productivity of challenges in the areas of costs, delays and disruptions,
earthmoving equipment. Song and AbouRizk [8] presented a impaired productivity, quality issues, safety aspects, mate-
productivity model for steel drafting and fabrication pro- rials unavailability, and escalation among others. These
ductivity through ANN and discrete-event simulation using challenges are highly complicated in nature and information
actual data. Oral and Oral [9] utilized a Self Organizing Map related to these challenges is vague. Therefore, construction
(SOM) to analyze the relationship between construction projects are within the purview of ANN in which the given
crew productivity and different factors. They also predicted ambiguous information can be effectively interpreted in
productivity in given situations for ready-mixed concrete, order to arrive at meaningful conclusions. In other words,
formwork, and reinforcement crew. Data were collected because of ANN’s capability to draw the relationships be-
randomly from a construction site in Turkey. They con- tween input and output provided via a training dataset, ANN
cluded that SOM can predict productivity better than re- is suitable for nonlinear problems where vague information,
gression methods due to its complexity. Muqeem et al. [10] subjective judgment, experience, and surrounding condi-
predicted production rate values for installation of beam tions are key features, and traditional approaches are in-
formwork using ANN. Meanwhile, Mohammed et al. [11] sufficient to calculate the complex input-output relationship
predicted the productivity of ceramic wall construction necessary for predicting construction labour productivity.
using data from general contractor companies. They applied This paper reviews the application of various ANNs in
ANN since analysis required performing complex mapping predicting construction labour productivity for formwork
of environment and management factors to productivity. assembly with a limited given dataset. Figure 1 shows the
AL-Zwainy et al. [12] developed a model for estimating overall flowchart of the research. The first step in the model
construction labour productivity in marble finishing works. development process was the choice of inputs from the
They used multilayer perceptron training through a BNN available data and appropriate model outputs. Then, the data
algorithm. Moselhi and Khan [13] ranked labour were processed through normalization and for handling
productivity-influencing parameters in construction using missing data. The data were divided into training and testing,
Fuzzy Subtractive Clustering, Neural Network Modeling, and the BNN, RBF, ANFIS, and GRNN AI models were
and Stepwise Variable Selection. They determined that work applied to the datasets. The models were calibrated via a
type, floor level, and temperature were the parameters with a performance evaluation and were compared using a de-
larger effect on productivity. Heravi and Eslamdoost [14] termination coefficient (R-squared), Mean Squared Error
considered 15 important factors in the motivation of labour, (MSE), and Mean Absolute Error (MAE). Ultimately, the
supervision sufficiency, and competency and suggested a best AI model was selected. Development of each model and
model for labour productivity rate estimation. Aswed [15] their results is presented in the following sections.
Advances in Civil Engineering 3

ANN productivity Data collection Formwork activity for

modeling (221 data points) 2 high-rise buildings

Normalization
Choice of inputs/output Data processing
Handling missing data

Data division Calibrated model

Training Testing Model Model

(80%) (20%) structure calibration

Performance
ANN models development evaluation

RBF BNN
Comparing developed models:
R-squared, MSE, MAE, correlation
coefficient

GRNN ANFIS
Best ANN model

Figure 1: Overall flowchart of the study.

2.1. Data Collection. The dataset used for modeling labour Table 1: Labour productivity factors.
productivity in this research was gathered by Khan [1] and
collected from field observations and data collection from Weather Crew Project
two high-rise buildings located in downtown Montreal over Temperature (T) Gang size (GS) Work type (WT)
a period of eighteen months. The buildings were 17 and 16 Humidity (H) Labour percentage (LP) Floor level (FL)
Wind speed (WS) Work method (WM)
floors. The first building is a concrete, mainly flat slab
Precipitation (P)
structure with Roller Compacted Concrete (RCC) con-
Temperature: based on an average of eight working hours a day (°C).
struction and several typical levels with a surface of Humidity: average humidity of eight working hours a day in percentage
68,000 m2. The project was constructed over three years. The form. Precipitation: includes four numerical value terms assigned as: no
second building also has a similar structure system and is a precipitation � 0, light rain � 1, snow � 2, and rain � 3. Wind speed: average
flat slab building. Two hundred and twenty-one data points wind speed for eight working hours a day and included in the calculation in
were collected from both projects for formwork activity. The terms of km/h. Height: related to the floor being worked on and included in
terms of the number of floors. Work type: three different types of formwork
collected data was classified into three groups of weather, installation included as slabs � 1, walls � 2, and columns � 3. Work method:
crew, and project. Data related to temperature, humidity, covers two techniques built in place (BIP) and flying forms (FF). BIP was
wind speed, and precipitation were classified into weather coded as 1 and FF was coded as 2. Gang size: number of persons in a crew.
data while gang size and labour percentage as crew. Floor Labour percentage: the ratio of labour to gang size obtained from the
level, work type, and method were the parameters used in following equation: (labour% � labour size/gang size) × 100.
the project category. These variables were selected because Table 2: Collected data descriptive statistics.
they cause variations in productivity on a daily basis [1].
Data of nine factors classified into three major categories SE Std.
Variable Mean Min. Median Max.
were available, as shown in Table 1, for performing the task. mean dev
These variables were chosen since they can cause dif- Temperature 4.08 0.81 12.03 -26 3 25
ferences in productivity on a daily basis or in the short-term. Humidity 66.34 1.05 15.67 18 67 97
Short-term influence means that factors change value every Precipitation 0.28 0.04 0.6 0 0 3
day and do not have a cumulated or ripple effect impact on Wind speed 15.42 0.57 8.46 3 14 43
Gang size 16.03 0.34 5.07 8 18 24
other activities. Thus, it is worthwhile to consider the
Labour
abovementioned labour productivity factors for modeling 35.49 0.26 3.79 29 36 47
percentage
labour productivity. Table 2 shows the descriptive statistics Work type 1.43 0.03 0.51 1 1 3
of the collected data which can provide a summary of the Floor level 11.38 0.25 3.75 1 12 17
dataset. Work method 1.44 0.03 0.5 1 1 2
The Pearson correlation coefficient is used to examine Productivity 1.57 0.02 0.35 0.82 1.51 2.53
the strength and direction of a linear relationship between
two variables in a database. A correlation coefficient ranges Pearson correlation, an absolute value of 1 specifies a perfect
between −1 and +1. A larger absolute coefficient value results linear relationship and a value of 0 indicates nonlinear re-
in a stronger relationship between variables. In the case of a lationship between variables. Table 3 shows that the
4 Advances in Civil Engineering

Table 3: Pearson correlation matrix for input and output parameters.

Variables T H P WS GS LP WT FL WM Productivity
T 1 0.151 −0.093 −0.122 0.390 −0.120 −0.122 0.358 0.078 0.589
H 0.151 1 0.338 0.026 −0.167 0.219 −0.110 0.048 0.176 0.090
P −0.093 0.338 1 0.076 0.065 −0.063 −0.037 −0.288 −0.057 −0.175
WS −0.122 0.026 0.076 1 −0.415 0.051 0.065 0.235 0.030 −0.202
GS 0.390 −0.167 0.065 −0.415 1 −0.310 −0.175 −0.352 −0.036 0.183
LP −0.120 0.219 −0.063 0.051 −0.310 1 −0.135 0.142 0.177 −0.053
WT −0.122 −0.110 −0.037 0.065 −0.175 −0.135 1 −0.052 −0.761 −0.353
FL 0.358 0.048 −0.288 0.235 −0.352 0.142 −0.052 1 0.225 0.301
WM 0.078 0.176 −0.057 0.030 −0.036 0.177 −0.761 0.225 1 0.328
Productivity 0.589 0.090 −0.175 −0.202 0.183 −0.053 −0.353 0.301 0.328 1

correlation between parameters most of the time is an ap- and evaluates the percentage of total differences between
proximate near 0. Consequently, none of the correlates very estimated and target values with respect to the average.
much and there is no over-estimation phenomenon. The Several BNN models with different numbers of neurons and
Pearson p-values and R-squared prove the same the inputs hidden layers were developed to find the best model for
and output behaviour. identifying labour productivity. The number of hidden
layers varied between one and two and the models were
2.2. Backpropagation Productivity Modeling. In this section, trained by five, ten, . . ., 100 neurons. Considering the dif-
BNN was applied to model labour productivity. BNN is ferences in the number of neurons and the hidden layer, 32
mostly used for unknown function approximation. As de- different models were developed and their results compared.
scribed in the literature review, a key BNN feature is its Figures 2 and 3 display the effect the number of neurons
learning ability. It can be trained by historical datasets to find has on the R-squared for one and two hidden layers. As can
the accurate relation between inputs and outputs as well as be seen from Figure 2, R-squared values are mostly between
predict the output(s) for new inputs. In this research, BNN 90 and 100% for the training phase and 70–90% for the
models were developed, trained, validated, and tested in testing phase. The model with one hidden layer and 50
MATLAB 2017a with 221 data points. The dataset was neurons shows maximum accuracy. For the two hidden
randomly divided into 80% and 20% groups used for layers’ models, the model with 20 neurons in each layer
training and testing results, respectively. Several BNN showed the best performance in predicting labour
models were developed which were different in three aspects: productivity.
number of neurons in hidden layer varied between five and Increasing the number of hidden layers resulted in better
100, random groups of datasets, and the number of hidden performance; however, this approach takes more compu-
layers of one and two. tational time and does not change model accuracy in any
Bayesian Regularization (BR) algorithm was used for significant way. In this research, the maximum number of
data training. BR is commonly used in noisy and small neurons that could be considered in the ANN model was set
problems. The algorithms attempted to minimize the sum of at 60 due to the extreme computational time needed for what
the squared errors by updating the network’s bias and is an insignificant improvement to model accuracy. Figures 4
weight. Training sets were used to adjust the network and 5 show that the MSE value was the smallest in the
structure based on the associated errors until the best models with two hidden layers and 20 neurons and one
structure was reached. Validation sets were utilized to hidden layer and 50 neurons.
measure network generalization capabilities and to pause It should be mentioned that no performance index was
training when generalization stopped improvement. After available during the validation phase while the given datasets
training, testing sets provided an independent network were trained by the BR algorithm because the algorithm does
performance index. For each BNN model, trials were per- not validate data and the datasets were randomly divided
formed to reach lowest error. Model performance was into trained and tested datasets only.
assessed based on R-squared and MSE developed through Model performances were assessed based on R2 and
MATLAB coding according to the following equations, MSE, as summarized in Table 4 for one hidden layer and
where “ti” is the target value while “oi ” is the output value: Table 5 for two hidden layers. As can be seen, the final model
2 was the one with two hidden layers and 20 neurons, which
􏽐i ti − oi 􏼁 showed the highest accuracy for identifying labour pro-
R2 � 1 − 2, (1)
􏽐i ti −(1/n)􏽐i ti 􏼁 ductivity. The model had MSE and R2 performance value
indices of 0.0054 and 0.949, respectively. Therefore, this
1 2 model was considered for comparison with the ANIFS
MSE � 􏽘 t −o 􏼁 . (2)
n i i i model in the next sections.
The error histogram of the final model with two hidden
R-squared is often used in statistical analysis since it is layers and 20 neurons demonstrates that most of the errors
easy to calculate and understand. It fluctuates between [0, 1] oscillate between −0.55 and 0.75 in all training and testing
Advances in Civil Engineering 5

100% 0.18
90% 0.16
80% 0.14
0.12
70%
R2

0.1

MSE
60%
0.08
50% 0.06
40% 0.04
0 20 40 60 80 100
0.02
Neurons
0
R2 training 0 20 40 60 80 100
R2 testing Neurons
R2 MSE
Figure 2: R2 values of ANN models using BP algorithm with 1 HL. MSE training
MSE testing
Figure 5: MSE values of ANN models in 2HL for different phases
100%
of ANN.
90%
80%
70% for the selected model as shown in Figure 6. The model
shows that temperature and floor level are the most im-
R2

60%
50% portant factors in labour productivity.
40%
30%
0 10 20 30 40 50 60 2.3. RBF Productivity Modeling. RBF is a three-layer forward
Neurons network applied for modeling science and engineering
problems fast and precisely. RBF is a branch of ANN first
R2 training
R2 testing
introduced in the late eighties. RBFNN architecture is simple
R2 and includes one hidden layer and output. The RBF was
selected to model labour productivity because of its feed-
Figure 3: R2 values of ANN models using BP algorithm with 2 HL. forward training done on a layer-by-layer basis in default of
having input signals going through convoluted and time
consuming multihidden layer developments. Thus, as
0.160
compared to other ANN techniques, RBFNN is faster and
0.140
has application flexibility [18].
0.120 Because of the abovementioned advantages, an effort
0.100 was made in this research to model the nine predictor
MSE

0.080 variables’ convoluted relations and target based on actual

0.060 datasets gathered from two high-rise buildings using
0.040 RBFNN. The RBFNN model was trained using a BP al-
0.020 gorithm to minimize MSE with selected predictor variables.
0.000 The RBF neural network included three layers: input,
0 10 20 30 40 50 60 hidden, and summation.
Neurons In this model, there is one neuron in the input layer for
MSE each predictor variable, where N is the number of categories
MSE training and N-1 the neurons used. Input neurons normalize a range
MSE testing of the values by subtracting the median and dividing it by an
interquartile range. The input neurons feed the values to
Figure 4: MSE values of ANN models in 1HL for distinct phases of
each of the neurons in the hidden layer. The hidden layer has
ANN.
a changing number of neurons (the optimal number is
determined by the training process). Each neuron contains a
phases. The concentration of errors was 0.003, which is a radial basis function centered on a point with the di-
small error for prediction. The R-squared in the selected mensions equal to predictor variables. The radius of a RBF
model shows the fitted line for all data as output  0.97 × function is different for each dimension. Here, the training
target + 0.099 and an R2 value of 97.68% in the training process determined the centers and spreads. A hidden
dataset, demonstrating that the outputs are very close to neuron measured the Euclidean distance of the test case
target values. The R2 value for testing was 83.27%, proof that from the neuron’s center point and then applied a RBF
the model is able to predict 83% of future outcomes accu- kernel function to this distance using the spread values when
rately. The applied method is able to rank predictor variables presented with the x vector of input values from the input
6 Advances in Civil Engineering

Table 4: Performance indices for models with one hidden layer.

Neurons 5 10 15 20 25 30 35 40 45 50 60
MSE 0.014 0.014 0.017 0.012 0.015 0.014 0.017 0.015 0.030 0.008 0.011
MSE train 0.010 0.004 0.002 0.002 0.001 0.002 0.002 0.002 0.001 0.005 0.003
MSE test 0.034 0.061 0.086 0.060 0.082 0.075 0.085 0.080 0.069 0.025 0.050
R2 0.94 0.94 0.93 0.95 0.94 0.94 0.94 0.94 0.91 0.97 0.96
R2 train 0.96 0.98 0.99 0.99 1 0.99 0.99 0.99 1 0.98 0.99
R2 test 0.88 0.77 0.69 0.72 0.68 0.79 0.5 0.77 0.75 0.89 0.77

Table 5: Performance indices for models with two hidden layers.

Neurons 5 10 15 20 25 30 35 40 45 50 60
MSE 0.0162 0.0096 0.1264 0.0215 0.0545 0.0384 0.0212 0.1264 0.0183 0.0184 0.0096
MSE train 0.0055 0.0022 0.1242 0.0200 0.0008 0.0000 0.0000 0.1267 0.0005 0.0008 0.0007
MSE test 0.0665 0.0446 0.1369 0.0230 0.1071 0.1193 0.1213 0.1251 0.1024 0.1010 0.0518
R2 0.935 0.942 0.499 0.949 0.813 0.872 0.929 0.512 0.927 0.932 0.962
R2 train 0.979 0.981 0.522 0.976 0.997 1 1 0.537 0.998 0.997 0.997
R2 test 0.693 0.693 0.403 0.832 0.335 0.482 0.733 0.395 0.714 0.694 0.785

100 variables, increasing the population size is recommended.

90 Increasing the population size helps to prevent local minima
80 and find the optimal global solution. In addition, DTREG
70 lets the user modify the network and neuron parameters as
Importance

60 well as the testing and validation percentage, select how to

50 handle missing predictor variable values, and select one of
40 four options for target categories’ prior probabilities. An-
30
other interesting option available to the user is that the
20
software can compute predictor variables’ importance [19].
10
0
To train the DTREG algorithm, sequential orthogonal
T FL H GS WS WT WM LP P training developed by [20] was used. This algorithm uses an
evolutionary approach to determine the optimal center
Figure 6: BNN model relative variable importance.
points and spreads for each neuron. It also determines when
to stop adding neurons to the network by monitoring the
layer. The resulting value was handed to the summation estimated Leave-One-Out (LOO) error and terminating
layer. The coming out value of a neuron in the hidden layer when the LOO error beings to increase due to overfitting.
was multiplied by a weight associated with the neuron (W1, Optimal weight computation between the neurons in the
W2, . . ., Wn) and passed to the summation which added up hidden layer and the summation layer was done using ridge
the weighted values and presented this sum as the output of regression. An iterative procedure was used to compute the
the network. A bias value of 1.0 was multiplied by a weight optimal regularization Lambda parameter that minimizes
(W0) and fed into the summation layer. For classification generalized cross-validation (GCV) error [20]. During
reasons, there was one output along with a separate set of training, it was found that model errors can be reduced
weights and summation unit for each target category. The sufficiently to a lower level after incorporating 11 neurons
output value of a category equaled to the probability that the and the model reached a steady state with 47 neurons. Thus,
evaluated case has that category. 47 neurons were used to model labour productivity. The RBF
In order to develop a reliable model, the dataset was network with 47 neurons developed in this research has R-
randomly divided into two separate training and testing squared values of 0.91 and 0.67 for training and testing,
subsets. Eighty percent of the dataset was considered for respectively. Table 6 shows the developed model perfor-
network training and 20% of data was used for network mance indicators. The RBF network algorithm ranked hu-
reliability and to avoiding overfitting. It should be noted that midity, floor level, and temperature as the most important
RBF was developed using DTREG predictive modeling variables, as shown in Figure 7.
software. One of the key advantages of using DTREG for
RBF development is that DTREG uses an evolutionary
method called Repeating Weighted Boosting Search (RWBS) 2.4. GRNN Productivity Modeling. GRNN, proposed by
for building neural networks. In DTREG, a population of Donald F. Specht in 1990, is often used for nonlinear
candidate neurons is first built with random centers and function approximation. It has a special linear and radial
spreads which is limited by the minimum and maximum basis layer which makes it different from radial basis net-
specified radius. The population size parameter controls how works. GRNN is a neural network model that mimics
many candidate neurons are created. If there are many nonlinear relations between a target variable and a set of
Advances in Civil Engineering 7

Table 6: Performances indices for RBF. minimizing error, minimizing the number of neurons, and
Performance index R2
MSE RMSE MAE
limiting the number of neurons to a certain number.
The developed GRNN model’s accuracy was compared
Training 91.30% 0.0103 0.1026 0.0756
with other models’ using the same dataset. Therefore, 80% of
Testing 85.84% 0.0471 0.1531 0.1421
the dataset was selected randomly as the training set, which
corresponded to 177 input-output pairs. Twenty percent of
100
the data were kept unused for testing, which corresponds to
90
44 input-output pairs. Note that all techniques used for
80 modeling labour productivity in this study used the same
70 training and testing datasets for a proper comparison
approach.
Importance

60
50 In addition, a Gaussian kernel function type was selected
40 as it is the best function among other kernel functions, and a
30 single sigma for the whole model was selected to reduce
20 computational time. Using a trial and error approach to
10 select the best model, three models were developed based on
0 three options provided by the software: remove unnecessary
FL T H WM LP GS WS WT P
neurons, minimize error, and the constant number of
Figure 7: Relative importance of variables in the RBFNN model. neuron. Table 7 summarizes various statistical indices for the
developed models.
The models with 34 and 10 neurons had overfitting in
predictor variables. GRNN falls into the class of probabilistic their training process since there was a large difference
neural networks and requires less training samples in between the R2 in the training and testing phases.
comparison to a BNN. A GRNN’s main advantage is that Therefore, the best GRNN model was found to have 107
since available datasets for developing neural networks are nodes with the R2 value of 87.87%, which is higher than the
not usually sufficient, probabilistic neural networks are more two other approaches. Like the RBF neural network,
attractive for modeling. In other words, GRNN can solve any GRNN is able to rank predictor variables for the selected
function approximation problem in case sufficient data is model as shown in Figure 8. Here, temperature and floor
available in abbreviated time. levels were the significant factors found for modeling
In GRNN, the target value of the predictor is achieved by productivity.
considering the weighted average of the values of its
neighbouring points. Target neighbour variable distance
plays a key role in predicting target value. Neighbouring 2.5. ANFIS Productivity Modeling. ANFIS is used in various
points close to target points have a greater impact on target engineering fields such as environmental, civil, electrical, etc.
value; distant points, on the other hand, are not influential as [23–25]. ANFIS utilizes a hybrid learning algorithm which
much as close neighbouring points. A radius base function is can model the relationship between predictor variables and
used for calculating the neighbouring point influence level. respond variables based on expert knowledge by using
As mentioned, GRNN is able to build a model with a rel- neural network capabilities. It represents expert knowledge
atively small dataset and has the capability to handle outliers in the form of fuzzy “if-then” rules with an approximation of
[21]. There are two main disadvantages associated with membership functions from given predictors and response
GRNN; it needs considerable calculations to evaluate new datasets. Fuzzy logic handles the vagueness and uncertainty
points and is not able to ignore unrelated inputs without associated with the system being modeled, whereas the
assistance and needs major algorithm modifications. Con- neural network provides model adaptability. By combining
sequently, this method is not a choice for problems with a the learning abilities of a neural work with the reasoning
substantial number of predictor variables. A GRNN algo- capacities of fuzzy logic into a unified platform, ANFIS can
rithm can be enhanced by advancing GRNN in two ways: be considered an enhanced prediction tool in comparison
using clustering versions of GRNN and applying parallel with a single methodology one. ANFIS can adjust mem-
calculations to take advantage of GRNN structure charac- bership function (MF) parameters and linguistic rules di-
teristics [22]. rectly from neural network training capabilities with respect
In addition to the abovementioned drawbacks, GRNN to refining model performance. ANFIS is able to capture
models can be large due to having one neuron for each expert knowledge regarding a nonlinear system and its
training row. In the developed GRNN model, DTREG was behaviour in a qualitative model without quantitative de-
utilized; thus, after the model was constructed, DTREG scriptions of the system. Fuzzy interference system (FIS) is a
provided an option for facilitating the removal of un- knowledge interpretation technique based on the concept of
necessary neurons from the model. By removing un- fuzzy set theory, fuzzy “if-then” rules, and fuzzy reasoning,
necessary neurons, computational time was reduced and it where each fuzzy rule characterizes a state of the system.
became possible to improve model accuracy. DTREG was ANFIS uses a Sugeno FIS for a structured approach to
utilized in order to select the best possible model. Three generate fuzzy rules by using a given dataset [26]. Training
criteria available for guiding the removal of neurons are and testing data are matrices with ten columns where the
8 Advances in Civil Engineering

Table 7: Performances indices for GRNN.

Number of R2 train R2 test MSE MSE RMSE RMSE MAE MAE
Performance index
neuron (%) (%) train test train test train test
Minimize error 107 87.87 75.32 0.0103 0.0475 0.1108 0.2219 0.0651 0.1421
Minimize number of neurons 34 85.84 48.70 0.0172 0.0716 0.1311 0.2676 0.0984 0.1785
Number of neurons 10 64.82 41.07 0.0427 0.0823 0.2067 0.2868 0.1471 0.2164

100 Table 8: ANFIS parameters for modeling labour productivity.

90
ANFIS parameters
80
70 Number of input variables 9
Training data points 177
Importance

60
50 Testing data points 44
40 Number of layers 5
30
Operator Subtractive
20
Number of membership functions (MF) 63
MF type Bell shape
10
Transfer function of output layer Linear
0
T FL WT WS GS P H LP WM Training algorithm Hybrid
Error tolerance 0
Figure 8: Relative importance of variables in the selected GRNN Number of epochs 1000
model.

first nine columns contain data for each FIS predictor To recognize which algorithm is the best method to be
variable and the last column contains the response data. It utilized in modeling construction labour productivity,
should be noted that the same 177 data points were used for analysis of variance (ANOVA) was applied to demonstrate
training and the other 44 for testing purposes. Then, the FIS the significant superiority of the BNN algorithm over other
model structure was generated by choosing the subtractive algorithms, which had the lowest F-Value.
clustering technique. Subtractive clustering technique is Furthermore, the results of the model were compared
faster than grid partitioning and with a satisfactory result for with the SOM model developed by [9] and results show that
justifying use. Based on the selected parameters of sub- for formwork productivity prediction, BNN performs better
tractive clustering, 63 clusters were detected as the most in comparison to SOM. The correlation of coefficient and the
suitable MF number. Here, the number of clusters and MFs MSE for the available database were 94.9% and 0.0215 for
were equal. The hybrid method was selected for training the backpropagation method while it was 89.25% and 0.07 for
MF because it generates better results rather than BP. The SOM.
hybrid method includes BP and least squares for MF pa- Both regression and AI techniques have merits and
rameter estimation, BP for estimating input MF parameters, demerits. Three statistical methods, namely, best subset,
and least square for MF parameters. ANFIS parameters were stepwise, and Evolutionary Polynomial Regression (EPR),
selected to reach higher accuracy with less computational were applied to the available database, and the coefficient of
time, as shown in Table 8. correlation and MSE were calculated and are presented in
Figure 9 shows the predicted daily productivity against Table 11. EPR predicted the data better than best subset and
corresponding actual values, and Table 9 summarizes the stepwise. However, BNN outperformed and achieved a better
dataset various statistical indices using the developed ANFIS fit and forecast with the given dataset than the regression
model. models due to the nonlinearity of the dataset in modeling
ANFIS is also able to rank predictor variables for the labour productivity for formwork. The statistical perfor-
selected model as shown in Figure 10. Temperature and floor mance of those models is far behind BNN. The analysis of
level are the important factors in modeling productivity. variance for different techniques is presented in Table 12.
Table 10 summarizes the performance indices of R2 Train, R2
Test, MSE Train, and MSE Test for the four models. BNN 3. Conclusions
shows the highest R-squared in the training phase followed
by RBF, ANFIS, and GRNN. Moreover, BNN shows the One of the major strategic components in determining the
highest R-squared in testing phases followed by RBF, ANFIS, success or failure of a construction project is the productivity
and GRNN. GRNN is the least accurate technique among all rate, which has a relationship with different factors. This
the techniques for the given dataset in both training and paper attempts to demonstrate a way to use AI models to
testing phases. BNN has the lowest MSE of the techniques in predict labour productivity. These are effective tools for
the training phase followed by RBF. ANFIS and BNN show quantifying loss of productivity and can be used as a support
the lowest MSE in the testing phase followed by RBF and method for actual loss of productivity calculations. GRNN,
GRNN. BNN, RBFNN, and ANFIS were tested against Khan’s
Advances in Civil Engineering 9

1.00
0.90
0.80
0.70

Norm. productvity
0.60
0.50
0.40
0.30
0.20
0.10
0.00
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
Data points
Actual
Predicted
Figure 9: Actual vs predicted value using ANFIS.

Table 9: Statistical indicators of the developed ANFIS model.

Performance index R (%) R2 train (%) R2 test (%) MSE train MSE test RMSE train RMSE test MAE train MAE test
Value 81.1 89.3 66.2 0.01 0.02 0.114 0.146 0.017 0.097

100 Table 12: Analysis of variance for different techniques.

90
Source DF Adj. SS Adj. MS F-value P-value
80
70 Actual 37 2.14235 0.05790 4.42 0.034
Importance

60 BNN Error 6 0.07864 0.01311

50 Total 42 2.22099
40 Actual 36 5.69202 0.15811 15.53 0.001
30 ANFIS Error 6 0.06108 0.01018
20 Total 42 5.75310
10 Actual 34 3.4887 0.10261 6.67 0.004
0 RBF Error 8 0.1231 0.01539
T FL H GS WS WT WM LP P
Total 42 3.6118
Figure 10: Relative importance of variables in ANFIS Model. Actual 37 5.59931 0.151333 31.89 0.002
GRNN Error 4 0.01898 0.004745
Table 10: Performance results comparison. Total 42 5.61829

R2 train R2 test MSE train MSE test

BNN 0.98 0.83 0.0023 0.020
(i.e., decreasing the number of nodes), early training phase
RBF 0.91 0.85 0.0096 0.018
GRNN 0.88 0.75 0.0103 0.047
stopping, or weight decay use. Furthermore, the dataset used
ANFIS 0.89 0.66 0.0100 0.020 to develop the aforementioned models was a raw and un-
balanced dataset. Studying the behaviour of the given
datasets prior to feeding it to any AI techniques is required in
Table 11: Statistical performance of the regression methods. order to have a robust model for modeling labour
productivity.
Technique R2 MSE Time (sec) Number of variables
Researchers have proved that supervised BNN models
Best subset 46.8 0.259 ∼2 8
are more successful in predicting construction crew pro-
Stepwise 48.61 0.259 ∼2 7
ductivity in comparison to statistical methods like re-
EPR 52.69 0.057 1140 8
gression. Furthermore, if the causal relationship between
input and output has a complex variability in areas other
datasets of two high-rise buildings related to formwork than construction, in most cases, the learning task is easier
operation. From the comparisons, BNN showed the best with unsupervised learning. Therefore, this study focuses on
performance among the techniques. However, BNN can be the application of supervised methods in formwork crew
considered a black box approach and is prone to overfitting, productivity data to compare the predicted results. This
which can be the result of network architecture study focused on formwork installation operations since
10 Advances in Civil Engineering

they constitute a substantial part of the overall labour [3] M. Lu, S. M. AbouRizk, and U. H. Hermann, “Estimating
component of concrete framing in building construction. labor productivity using probability inference neural net-
Results reveal that BNN shows superior performance in work,” Journal of Computing in Civil Engineering, vol. 14,
comparison to RFBNN, ANFIS, and GRNN for formwork no. 4, pp. 241–248, 2000.
productivity prediction in the following ways: [4] S. AbouRizk, P. Knowles, and U. R. Hermann, “Estimating
labor production rates for industrial construction activities,”
(1) Selected input variables are those that cause varia- Journal of Construction Engineering and Management,
tions in productivity in the short-term or daily basis. vol. 127, no. 6, pp. 502–511, 2001.
The developed models were compared based on [5] O. Moselhi, I. Assem, and K. El-Rayes, “Change orders impact
statistical performances and BNN outperformed the on labor productivity,” Journal of Construction Engineering
techniques of RBF, ANFIS, and GRNN. The de- and Management, vol. 131, no. 3, pp. 354–359, 2005.
[6] A. S. Ezeldin and L. M. Sharara, “Neural networks for esti-
veloped model of this research can be utilized in
mating the productivity of concreting activities,” Journal of
different ways.
Construction Engineering and Management, vol. 132, no. 6,
(2) The model can help to estimate formwork pro- pp. 650–656, 2006.
ductivity by considering variables such as temper- [7] S. C. Ok and S. K. Sinha, “Construction equipment pro-
ature, humidity, gang size, labour percentages, work ductivity estimation using artificial neural network model,”
type, etc. In addition, this study also found that Construction Management and Economics, vol. 24, no. 10,
productivity is not significantly correlated with pp. 1029–1044, 2006.
precipitation, labour percentage, work method, and [8] L. Song and S. M. AbouRizk, “Measuring and modeling labor
humidity. Within the scope of the conducted study, productivity using historical data,” Journal of Construction
Engineering and Management, vol. 134, no. 10, pp. 786–794,
the number of parameters observed, and the range of
2008.
their values, this study found temperature to have the [9] E. L. Oral and M. Oral, “Predicting construction crew pro-
most significant impact on productivity followed by ductivity by using self organizing maps,” Automation in
floor level. Construction, vol. 19, pp. 791–797, 2010.
(3) The model can also be useful for quantifying loss of [10] S. Muqeem, M. Idrus, and F. Khamidi, “Construction labor
productivity by considering the output of the de- production rates modeling using artificial neural network,”
veloped model as the value for unimpacted pro- Journal of Information Technology in Construction, vol. 16,
ductivity period since the identifying unimpacted pp. 713–725, 2011.
[11] S. Mohammed and A. Tofan, “Neural networks for estimating
period for quantifying loss of productivity is im-
the ceramic productivity of walls,” Journal of Engineering,
possible to calculate sometimes. Therefore, BNN can vol. 17, no. 2, pp. 200–217, 2011.
help save time and cost associated with quantifying [12] F. M. S. AL-Zwainy, H. A. Rasheed, and H. F. Ibraheem,
loss of productivity. “Development of the construction productivity Estimation
Ultimately, this study shows that the backpropagation model using artificial neural network for finishing works for
model can be an alternative tool to supervised learning- floors with marble,” ARPN Journal of Engineering and Applied
Sciences, vol. 7, no. 6, pp. 714–722, 2012.
based tools and can be used in various prediction applica-
[13] O. Moselhi and Z. Khan, “Significance ranking of parameters
tions. One limitation of this study is that the findings are impacting construction labour productivity,” Construction
limited to the collected data range and parameters con- Innovation, vol. 12, no. 3, pp. 272–296, 2012.
sidered in the study. It should be noted that the developed [14] G. Heravi and E. Eslamdoost, “Applying artificial neural
model does not involve any parameters that directly account networks for measuring and predicting construction-labor
for management strategies and skills or any project-specific productivity,” Journal of Construction Engineering and
conditions. Management, vol. 141, no. 10, article 04015032, 2015.
[15] G. Aswed, “Productivity estimation model for bricklayer in
construction projects using neural network,” Al-Qadisiyah
Data Availability Journal for Engineering Sciences, vol. 9, no. 2, pp. 183–199,
The data used to support the findings of this study are 2016.
[16] K. M. El-Gohary, R. F. Aziz, and H. A. Abdel-Khalek, “En-
available from the corresponding author upon request.
gineering approach using ANN to improve and predict
construction labor productivity under different influences,”
Conflicts of Interest Journal of Construction Engineering and Management,
vol. 143, no. 8, article 04017045, 2017.
The authors declare that they have no conflicts of interest. [17] O. Moselhi, T. Hegazy, and P. Fazio, “Neural networks as tools
in construction,” Journal of Construction Engineering and
References Management, vol. 117, no. 4, pp. 606–625, 1991.
[18] M. T. Musavi, W. Ahmed, K. H. Chan, K. B. Faris, and
[1] Z. U. Khan, “Modeling and parameter ranking of construction D. M. Hummels, “On the training of radial basis function
labor productivity,” Doctoral Dissertation, Concordia Uni- classifiers,” Neural Networks, vol. 5, no. 4, pp. 595–603, 1992.
versity, Montreal, Canada, 2005. [19] P. Sherrod, “DTREG predictive modeling software,” 2018,
[2] W. Yi and A. P. C. Chan, “Critical review of labor productivity http://www.dtreg.com.
research in construction journals,” Journal of Management in [20] S. Chen, X. Hong, and C. J. Harris, Orthogonal Forward
Engineering, vol. 30, no. 20, pp. 214–225, 2014. Selection for Constructing the Radial Basis Function Network
Advances in Civil Engineering 11

with Tunable Nodes, in Lecture Notes in Computer Science,

Vol. 3644, 2005.
[21] H.-l. Yip, H. Fan, and Y.-h. Chiang, “Predicting the main-
tenance cost of construction equipment: comparison between
general regression neural network and Box-Jenkins time se-
ries models,” Automation in Construction, vol. 38, pp. 30–38,
2014.
[22] D. F. Specht, “Probabilistic neural networks,” Neural Net-
works, vol. 3, no. 1, pp. 109–118, 1990.
[23] H. Alasha’ary, B. Moghtaderi, A. Page, and H. Sugo, “A
neuro–fuzzy model for prediction of the indoor temperature
in typical Australian residential buildings,” Energy and
Buildings, vol. 41, no. 7, pp. 703–710, 2009.
[24] A. Subasi, A. S. Yilmaz, and H. Binici, “Prediction of early heat
of hydration of plain and blended cements using neuro-fuzzy
modelling techniques,” Expert Systems with Applications,
vol. 36, no. 3, pp. 4940–4950, 2009.
[25] L.-C. Ying and M.-C. Pan, “Using adaptive network based
fuzzy inference system to forecast regional electricity loads,”
Energy Conversion and Management, vol. 49, no. 2,
pp. 205–211, 2008.
[26] M. Negnevitsky, ANFIS: Adaptive Neruo-Fuzzy Inference
System, Artificial Intelligence-A Guide to Intelligent Systems,
Pearson Education Limited, Essex, UK, 2nd edition, 2005.
 International Journal of

Rotating Advances in
Machinery Multimedia

The Scientific
Engineering
Journal of
Journal of

Hindawi
World Journal
Hindawi Publishing Corporation Hindawi
Sensors
Hindawi Hindawi
www.hindawi.com Volume 2018 http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi
Engineering
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

VLSI Design
Advances in
OptoElectronics
International Journal of

International Journal of
Modelling &
Simulation
Aerospace
Hindawi Volume 2018
Navigation and
Observation
Hindawi
www.hindawi.com Volume 2018
in Engineering
Hindawi
www.hindawi.com Volume 2018
Engineering
Hindawi
www.hindawi.com Volume 2018
Hindawi
www.hindawi.com www.hindawi.com Volume 2018

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Hindawi Hindawi Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018