Neurocomputing is widely implemented in time series area, however the nearness of exceptions that for the most part happen in information time arrangement might be hurtful to the information organize preparing. This is on the grounds that the capacity to consequently discover any examples without earlier suppositions and loss of all-inclusive statement. In principle, the most well-known preparing calculation for Backpropagation calculations inclines toward lessening ordinary least squares estimator (OLS) or all the more particularly, the mean squared error (MSE). In any case, this calculation is not completely hearty when exceptions exist in preparing information, and it will prompt false estimate future esteem. Along these lines, in this paper, we show another calculation that control calculations firefly on slightest middle squares estimator (FFA-LMedS) for BFGS quasi-newton backpropagation neural network nonlinear autoregressive moving (BPNN-NARMA) model to lessen the effect of exceptions in time arrangement information. In the in the mean time, the monthly data of Malaysian Roof Materials cost index from January 1980 to December 2012 (base year 1980=100) with various level of exceptions issue is adjusted in this examination. Toward the finish of this paper, it was found that the upgraded BPNN-NARMA models utilizing FFA-LMedS performed extremely well with RMSE values just about zero errors. It is expected that the finding would help the specialists in Malaysian development activities to handle cost indices data accordingly.

1. Indonesian Journal of Electrical Engineering and Computer Science
Vol. 8, No. 1, October 2017, pp. 101 ~ 106
DOI: 10.11591/ijeecs.v8.i1.pp101-106  101
Received June 1, 2017; Revised August 5, 2017; Accepted September 1, 2017
Enhanced BFGS Quasi-Newton Backpropagation
Models on MCCI Data
Nor Azura Md. Ghani*
1
, Saadi Ahmad Kamaruddin
2
, Norazan Mohammed Ramli
3
,
Ismail Musirin
4
, Hishamuddin Hashim
5
1,3
Center for Statistical Studies and Decision Sciences, Faculty of Computer and Mathematical Sciences,
Universiti Teknologi MARA, Selangor Darul Ehsan, Malaysia
2
Department of Computational and Theoretical Sciences, Kulliyyah of Science, International Islamic
University Malaysia, Pahang Darul Makmur, Malaysia
4,5
Center for Electrical Power Engineering Studies, Faculty of Electrical Engineering,
Universiti Teknologi MARA, Selangor Darul Ehsan, Malaysia
Pilihsepakat Sdn. Bhd., Lot 5431, Tingkat 1(B), Jalan J9, Fasa 6, Taman Melawati, Kuala Lumpur,
Malaysia
*Corresponding author, e-mail: azura@tmsk.uitm.edu.my
1
, saadiak@iium.edu.my
2
,
norazan@tmsk.uitm.edu.my
3
, ismailbm@salam.uitm.edu.my
4
, pssbf6@gmail.com
5
Abstract
Neurocomputing is widely implemented in time series area, however the nearness of exceptions
that for the most part happen in information time arrangement might be hurtful to the information organize
preparing. This is on the grounds that the capacity to consequently discover any examples without earlier
suppositions and loss of all-inclusive statement. In principle, the most well-known preparing calculation for
Backpropagation calculations inclines toward lessening ordinary least squares estimator (OLS) or all the
more particularly, the mean squared error (MSE). In any case, this calculation is not completely hearty
when exceptions exist in preparing information, and it will prompt false estimate future esteem. Along
these lines, in this paper, we show another calculation that control calculations firefly on slightest middle
squares estimator (FFA-LMedS) for BFGS quasi-newton backpropagation neural network nonlinear
autoregressive moving (BPNN-NARMA) model to lessen the effect of exceptions in time arrangement
information. In the in the mean time, the monthly data of Malaysian Roof Materials cost index from January
1980 to December 2012 (base year 1980=100) with various level of exceptions issue is adjusted in this
examination. Toward the finish of this paper, it was found that the upgraded BPNN-NARMA models
utilizing FFA-LMedS performed extremely well with RMSE values just about zero errors. It is expected that
the finding would help the specialists in Malaysian development activities to handle cost indices data
accordingly.
Keywords: Outliers, CCI data, trainbfg, robust estimators, evolutionary algorithms
Copyright © 2017 Institute of Advanced Engineering and Science. All rights reserved.
1. Introduction
Monetary Activities (PFI) is these days a pattern in Malaysia, as it is general with the
administration's promoting more prominent non-open segment's inclusion inside abetment the
notoriety over system administrations. The most basic donor with regards to PFI is worth since
money (VFM), where ideal trademark over advancement activities incorporating regard as per
customer's pleasure yet ventures are in this manner achieved effectively. It is key to tally in
regard to texture charges, so much are acquired at some phase in the PFI developments as per
determine as overspending wish not happen. Since the building occasion then circumstance
conveyance is the administration motivation inside the Malaysian PFI, endeavors bear been
done in similarity with anticipate the present list concerning development texture worth lists
inside Malaysia. It was appropriately introduced that concrete's controlled esteem has been
wrecked by means of the Malaysian government, beginning concerning 5 June 2008 [1]. From
that point forward, so was a sparkling intensify in regard to the concrete goodness in June 2008
which is with the guide of 23.3% among Peninsular Malaysia, while 6.5% was accounted for
among Sabah and 5.2% among Sarawak [1].
In the meanwhile, Malaysian government had actualized Merchandise and Ventures
Expense (GST) all through the country since first April 2015. Merchandise at that point

2.  ISSN: 2502-4752
IJEECS Vol. 8, No. 1, October 2017 : 101 – 106
102
Administrations Duty (GST) is a multi-organize intricacy over household utilization. GST is
charged on whole assessable segments over frill and applications in Malaysia other than these
especially exempted. GST is likewise charged in regard to importation over products yet
purposes inside Malaysia [2]. Because of the usage of this new approach in Malaysia,
engineers are mostly hit by the cost of crude materials [3]. The more terrible effect is, industry
players and specialists expect the costs of private properties to rise 2% to 4% post-GST
notwithstanding the way that such properties are not subject to the GST. Accordingly, with the
execution with GST, combined with the harder working condition, property designers are
probably going to methodologies to cradle any negative effect.
The esteem bring is additionally pertinent up in agreement with the rest of the
improvement materials-steel, prepared blend concrete yet various others [4]. As building
material expenses into Malaysia have been met including vulnerability, the prevalent technique
has been examined in impersonation of depend estimation of the development fabric costs in
agreement as per the mean area on Malaysia. Next, the legacy on actualities constant among
that direction is depicted briefly inside part II. In share III, the technique diagram is moreover
provided, yet the strategy old as per break down the information clarified. Next, the finished
results and discourse about the promising gauging technique in light of the fact that assessing
the material charge files concurring in impersonation of Malaysian regions are presented in
divide IV. At last, divide V conveys the conclusion on the investigation, summation a proposal
on account of future works.
2. Foundation of Data
The information had been formed from 3 unique sources especially Unit Kerjasama
Awam Swasta (UKAS) on Head administrator's Area of expertise, Development Industry
Advancement Board (CIDB) at that point Malaysian Insights Office as had suggested the
improvement costs lists in light of the fact that the medium locale on the Peninsular Malaysia
which comprises of three states Kuala Lumpur Government Domain, Selangor, Negeri Sembilan
then Melaka. The genuine modern month-to-month measurements over Malaysian Total esteem
records past January 1980 in impersonation of December 2012 (base a year 1980=100) were
adjusted, with anomalies 6.1 percent of the general informational index. The variable of intrigue
experiences anomalies issue as can be found in Figure 1.
Figure 1. The dispersion of Malaysian Rooftop Materials costs files information
3. Methodology
The flowchart over the query perform keep considered in Figure 2. Here, the current
effective estimators of backpropagation neural system have been actualized. To answer the
most vital objective over the investigation, the conceivable intense estimator’s crossover inside
nonlinear autoregressive (NAR) yet nonlinear autoregressive moving average (NARMA) over
neural system day accumulation had been played out the utilization of MATLAB R2012a. At that

3. IJEECS ISSN: 2502-4752 
Enhanced BFGS Quasi-Newton Backpropagation Models on MCCI Data (Nor Azura Md. Ghani)
103
progression, MATLAB contents then coding have been composed parallel as per the scientific
equation taken before. From that point onward, the execution concerning the proposed
robustified neural system designs have been interestingly the utilization of genuine certainties
the use of the root means square error (RMSE). The unrivaled near impacts have been
extended appropriate here where the valuable model was picked. The fundamental BPNN-NAR
detailing can be represented as:
𝐻(𝑥) = 𝑝𝑝𝑝𝑝𝑝𝑝𝑝 �� 𝑤𝑗𝑗 �tanh �� 𝑤𝑖𝑖
𝑙
𝑖=1
�𝑥(𝑡 − 1), 𝑥(𝑡 − 2), … , 𝑥�𝑡 − 𝑛 𝑦�� + 𝜀(𝑡)��
𝑚
𝑗=1
� (1)
The finished BPNN-NARMA definition can be represented as:
𝐻(𝑥) = 𝑝𝑝𝑝𝑝𝑙𝑖𝑖 �� 𝑤𝑗𝑗 �tanh �� 𝑤𝑖𝑖 �
�𝑥(𝑡 − 1), 𝑥(𝑡 − 2), … , 𝑥�𝑡 − 𝑛 𝑦�,
𝜀(𝑡 − 1), 𝜀(𝑡 − 2), … , 𝜀(𝑡 − 𝑛 𝜀)] + 𝜀(𝑡)
�
𝑙
𝑖=1
��
𝑚
𝑗=1
� (2)
where H(x) is the NARMA model, x(t-1), x(t-2),…, x(t-ny) are input lags, ε(t-1), ε(t-2),…ε(t-
nε) are residual lags terms. Hence, ε(t) are the white noise. l is the input neurons with index i, m
is the hidden neurons with index j, and n is the output neurons with index k.
The nearby important segment over the investigation is the scientific plan enchancment
offer of backpropagation neural system calculation utilizing factual solid estimators. To make
hearty the normal backpropagation calculation based absolutely in regard to the M-estimators
thinking for bringing down anomaly impact, the squared residuals strength 𝜀𝑖
2
among the system
cataclysm with the guide of another component in regard to the residuals.
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅(𝐸) =
1
𝑁
� 𝜀𝑖
2
𝑁
𝑖
, (3)
which comply with,
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅(𝐸) =
1
𝑁
� 𝜌(𝜀𝑖)
𝑁
𝑖
, (4)
where N is the sum assortment with respect to tests reachable in light of the fact that system
preparing. We are inferring the refreshing on the system weights based absolutely finished the
inclination better than average learning calculation. To keep the hardship in regard to all-
inclusive statement, a feedforward neural system together with one dark level delight lie did
between this examination. The weights past the concealed neurons in impersonation of yield
neurons, 𝑊𝑗,𝑖 are communicated so.
∆𝑊𝑗,𝑖 = −𝛼 �
𝜕𝜕
𝜕𝑊𝑗,𝑖
� = −
𝛼
𝑁
��
𝜕𝜕(𝜀𝑖)
𝜕𝑊𝑗,𝑖
𝑁
𝑖
�
= −
𝛼
𝑁
�� 𝜑(𝑟𝑖).
𝜕𝑓𝑗
𝜕𝑛𝑛𝑛𝑗
. 𝑂𝑖
𝑁
𝑖
�,
(5)
the place 𝛼 is a constant,𝑂ℎ is the yield of the 𝑖 𝑡ℎ
neuron, 𝑂𝑗 = 𝑓𝑗�𝑛𝑛𝑛𝑗� is produced over the 𝑗 𝑡ℎ
output neuron, 𝑛𝑛𝑛𝑗 = ∑ 𝑊𝑗𝑗 𝑂𝑖𝑖 is produced at the input on the activation function associated with
the output neuron (𝑗), and 𝑓𝑗 is the activation function of the neurons in the output layer. A

4.  ISSN: 2502-4752
IJEECS Vol. 8, No. 1, October 2017 : 101 – 106
104
linear (purelin) is used in the output layer’s neurons. The weights from the input to hidden
neurons 𝑊𝑗,𝑖 are updated as.
∆𝑊𝑗,𝑖 = −𝛼 �
𝜕𝜕
𝜕𝑊𝑗,𝑖
� = −
𝛼
𝑁
��
𝜕𝜕(𝜀𝑖)
𝜕𝑊𝑗,𝑖
𝑁
𝑖
�
= −
𝛼
𝑁
�� � 𝜑(𝑟𝑖).
𝜕𝑓𝑗
𝜕𝑛𝑛𝑛𝑗
. 𝑊𝑗,𝑖.
𝜕𝑓𝑖
𝜕𝑛𝑛𝑛𝑖
𝑗
𝑁
𝑖
. 𝐼𝑖�,
(6)
Figure 2. Flow of enhanced BFGS quasi-newton backpropagation

5. IJEECS ISSN: 2502-4752 
Enhanced BFGS Quasi-Newton Backpropagation Models on MCCI Data (Nor Azura Md. Ghani)
105
where Ii is the enter after the ith enter neuron, 𝑛𝑛𝑛𝑗 = ∑ 𝑊𝑗𝑗 𝑂𝑖𝑖 is realized local subject advanced
at the contribution over the enactment work connected together with the hidden neuron (𝑖), and
𝑓𝑗 is the initiation work about the neurons between the mystery layer. We have the goal as per
uses the tan-sigmoid trademark to be specific the initiation includes on the grounds that the dark
layer's neurons in light of the fact that on its adaptability. The leastmedian squares (LMedS)
estimator appraises the parameters by methods for tackling the nonlinear minimization issue.
𝑚𝑚𝑚[𝑚𝑚𝑚𝑖(𝜀𝑖
2
)] (7)
That is, the estimator should create the littlest excellence for the middle over squared residuals
processed for the entire data set. It appears to be so much this approach is completely strong in
congruity with false fits then also in impersonation of anomalies staying in impersonation of
deteriorative confinement [6]. Not sort of the M-estimators, nonetheless, the LMedS
inconvenience can't stand diminished by a weighted minimum squares issue. It is maybe now
not reasonable in impersonation of shadow underneath a simple equation in view of the result
on LMedS estimator. Thus, deterministic calculations may also not be skilled as per trademark
in impersonation of limit that estimator. The Monte-Carlo technique [7] has been worked on as
indicated by clear up it inconvenience among halfway non-neural applications. Life span
Stochastic calculations are also recognized as the streamlining calculations as utilization freely
hunt to harvest an answer. Stochastic calculations are thus genuinely moderate, however in that
place is likelihood that want find the worldwide least. One really well-known enhancement
calculation used by diminish a LMedS-based system perplexity is manufactured drink (SA)
calculation. SA is a glorious calculation since it is especially broad, yet such has the inclining not
after be brought gotten of either the incomplete least then greatest [6]. Nonetheless, [5] finds so
iterated LMedS tends as indicated by outsail the SA-LMedS.
Table 1. Stopping Criteria
Values NN Terms
1000 Maximum number of epochs to train
0 Performance goal
1e^-7 Minimum performance gradient
4. Results
From the examination on MCCI datasets, it can be assumed that the proposed
backpropagation neural framework time arrangement models performed well when the
information involves exceptions. Since the fact of the matter is to find the best fitted deciding
models for MCCI datasets, this examination can unravel the disclosures as in Table 2. The most
noticeably awful design for Rooftop Materials information is 5-5-20 of BPNN-NARMA show with
RMSE = 0.833. The best design is BPNN-NARMA show with setup 10-10-10 where the RMSE
= 0.414. This is trailed by BPNN-NAR display with design 10-10-10 and 15-15-15, where the
RMSE = 0.598.
Table 2. Yield of Ordinary BPNN-NAR and BPNN-NARMA Models on Malaysian Roof
Materials Cost Indices Data based on Different Lags
Input Error Hidden RMSE
Lags Lags Nodes BPNN-NAR BPNN- NARMA
5 5 20 0.689 0.833
10 10 10 0.598 0.414
15 15 15 0.598 0.649
20 20 20 0.624 0.684
25 25 25 0.649 0.717
30 30 30 0.671 0.749
35 35 35 0.678 0.764
40 40 45 0.709 0.807

6.  ISSN: 2502-4752
IJEECS Vol. 8, No. 1, October 2017 : 101 – 106
106
5. Discussions
This implies if the system is given sufficient number of info slacks and mistake slacks,
consolidated with satisfactory number of concealed hubs, the system can have the capacity to
perform ideally. For this situation, the best model is BPNN-NARMA demonstrate.
6. Conclusion
The reasonable model suitable in light of the fact that Malaysian Rooftop Materials
charge files measurements is BPNN-NARMA together with setup 10-10-10. In the following
exertion, FFA-LMedS might stand tested concerning certifiable information who relate on 30%
to half distant information. The proposed Colossal calculations since training neural systems
may keep achievable after stay customized a scope of fields of counterfeit consciousness,
arrangement recognizable proof, specimen acknowledgment, machine learning, quality control
and streamlining yet logical figuring. The proposed algorithm can be further implemented in many
processing activities, such as image processing [8], water treatment plant [9] and power plant [10].
Acknowledgement
We want to express our gratefulness and appreciation to Unit Kerjasama Awam Swasta
(UKAS) over Prime Minister's Office, Construction Industry Development Board (CIDB) then
Malaysian Statistics Department. Appreciation to Universiti Teknologi MARA and Malaysian
Service over Advanced education (MOHE) as a result of supporting it investigation underneath
the Research Grant No. 600-RMI/DANA 5/3/PSI (197/2013) or No. 600-RMI/FRGS 5/3
(137/2014).
References
[1] HM Foad, A Mulup. Harga siling simen dimansuh 5 Jun. Utusan, Putrajaya. 2008.
[2] J Goh. Developers strategizing to buffer GST impact. The Edge Malaysia, MSN News. 2015: 27.
[3] Royal Malaysian Customs. Good and Sevices Tax: Guide on Construction Industry. 2014: 1-27.
[4] SBA Kamaruddin, NAM Ghani, NM Ramli. Best Forecasting Models for Private Financial Initiative
Unitary Charges Data of East Coast and Southern Regions in Peninsular Malaysia. International
Journal of Economics and Statistics. 2014: 2.
[5] A Rusiecki. Robust Learning Algorithm Based on Iterative Least Median of Squares. Neural Process
Lett. 2012; 36: 145-160.
[6] MT El-Melegy, MH Essai, AA Ali. Robust training of artificial feedforward neural networks. Found
Comput Intell 1, Springer, Berlin, Heidelberg. 2009; 1: 217–242.
[7] Z Zhang. Parameter estimation techniques: A tutorial with application to conic fitting. Image and vision
Computing. 1997; 15(1): 59-76.
[8] B Saichandana, K Srinivas, R KiranKumar. Image fusion in hyperspectral image classification using
genetic algorithm. Indonesian Journal of Electrical Engineering and Computer Science. 2017: 2(3):
703-711.
[9] MS Gaya, LA Yusuf, M Mustapha, B Muhammad, A Sani, ATA Tijjani, MTM Khairi. Estimation of
Turbidity in Water Treatment Plant using Hammerstein-Wiener and Neural Network
Technique. Indonesian Journal of Electrical Engineering and Computer Science. 2017; 5(3): 666-672.
[10] S Chakraborty, PK Sadhu, N Pal. A New Approach towards Ideal Location Selection for PV Power
Plant in India. Indonesian Journal of Electrical Engineering and Computer Science. 2014: 12(11):
7681-7689.