You are on page 1of 9

Neural Computing and Applications

https://doi.org/10.1007/s00521-022-08164-7 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

A hybrid artificial intelligence model for design of reinforced concrete


columns
1 •
Sinan Melih Nigdeli Melda Yücel1 • Gebrail Bekdaş1

Received: 2 February 2022 / Accepted: 6 December 2022


Ó The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
In the optimum design of structures, the optimization process is an iterative one and it may last a long time. If the structural
plan is updated, the optimization process is needed to be redone since dimensions and internal forces change. Also, the
local market prices may show differences and the proposed design may not be the optimum anymore. To skip the
optimization process, intelligence methods can be used to predict the optimum values. In the study, a model is proposed for
cost optimum results of reinforced concrete columns. A hybrid method is presented that uses harmony search as a
metaheuristic method in the optimum design and multi-layer perceptions as a type of artificial neural networks in machine
learning to generate a model. The prediction results were evaluated for several error metrics, and the model is feasible in
proposing optimum solutions.

Keywords Reinforced concrete  Machine learning  Artificial neural networks  Prediction  Columns

1 Introduction of the best solution or choice for the design and model
should be considered. This is because the method can be
Structural engineering has a wide variety of structural used to best solve the mentioned problem and converge to
designs and models. They require usability and aesthetics it. There are many types of optimization, but today, the
as well as safety and economics, both in extreme condi- most researched, perhaps wondered, preferred and used
tions and in sudden situations including winds and earth- metaheuristic methods are. The metaheuristic has the
quakes. To meet these needs together, it is important to power and ability to solve a problem in any optimal
realize the design accurately and precisely, as well as direction.
controllable and renewable solutions. Time and labor fac- In the update of the solutions, the optimization process
tors are also very important for this goal. It is known that must be redone, and this process involves numerical iter-
the use of classical analysis programs with advanced cod- ations that can take time. Intelligent techniques with
ing software has become more important in recent years. measurable performance are required, as well as achieving
Especially for optimization, coding is preferred more than the desired design criteria and solutions faster and per-
analysis programs because it is more useful and updat- forming these actions in a short time and with precision.
able and can be adjusted by any designer. In addition, if The use of these techniques provides benefits in terms of
problems in optimization are evaluated, the determination both time and effort, and in this way, early intervention can
be made against the negativities that occur in the designs.
The mentioned methods refer to machine learning
& Gebrail Bekdaş
bekdas@iuc.edu.tr techniques in artificial intelligence (AI) technology and can
find the desired results in an extremely short time for any
Sinan Melih Nigdeli
melihnig@iuc.edu.tr model with a large number of design members. Sample
studies that are carried out include pre-calculation of how
Melda Yücel
melda.yucel@yahoo.com.tr much water will be delivered to residents in any city net-
work line [1–4]; predetermination of the probability of
1
Department of Civil Engineering, Istanbul University- occurrence of vehicle traffic accidents in places [5–7];
Cerrahpaşa, 34320 Avcılar, Istanbul, Turkey

123
Neural Computing and Applications

Fig. 2 Prediction model for the determination of properties for RC


column designs

determination of the most appropriate drug dose for the


treatment of any disease [8–11]; the determination of how
much the total construction cost of a new building will be
[12–15] can be shown.
Fig. 1 3-D model illustration and loading conditions for the RC In structural engineering, several prediction studies
column using networks including prediction of optimum tuned
mass damper (TMD) parameters [16], optimum design of
carbon fiber reinforced polymer (CFRP) [17], optimum
Table 1 Optimization constants and variables
design of truss structures [18, 19] and optimum design of
Description Value reinforced concrete structures [19–21] have been con-
ducted. For reinforced concrete structures, the prediction of
Design Column length (l) Input in ANNs model
constants (Optimized for optimum results of columns was not previously investi-
Axial force (Nz)
multiple values for data gated and a prediction model for these members is
Flexural moment (My)
generation) important since these members are under the effect of both
The unit cost of concrete
per m3 (Cc) tension and compression according to the loading condi-
The unit cost of steel per
tions. In the present study, a hybrid method that predicts
ton (Cs) the optimum design of reinforced concrete (RC) columns is
Yield strength of steel (fy) 420 MPa presented. Firstly, the optimum values are found via the
Compressive strength of 30 MPa methodology using harmony search (HS). Then, the opti-
concrete (f’c) mum results for different input values including length,
The elasticity modulus of 200,000 MPa internal loads and cost of materials are used in machine
steel (Es) learning to reflect the possible change in the structural plan
The specific density of 7.86 t/m3 and market prices. A prediction of multi-layer perceptions
steel (cs)
(MLP) that is a type of artificial neural network (ANNs)
Design The breadth of column (b) 250–400 mm
was chosen.
variables The height of member (h) 300–500 mm
Steel ratio, (q) 0.01–0.06
2 Materials and methods
Table 2 Design constraints
Description Constraints 2.1 Optimization process
Maximum axial force (Nmax) g1: Nd B Nmax = 0.5fc’bh For RC members, the cost optimization problem is a
Minimum steel ratio, (qmin) g2: q C qmin = 0.01 nonlinear one because of design constraints defined
Maximum steel ratio, (qmax) g3: q B q smax = 0.06 according to design codes. Various metaheuristic methods
Flexural strength capacity (Md) g4: Md C My have been proposed for different types of members
Axial force capacity (Nd) g5: Nd C Nz [22–30].

123
Neural Computing and Applications

In this study, HS proposed by Geem et al. [31] was used


in the optimization process to collect data for machine
learning.
The RC column presented in this study is subjected to an
axial force (Nz) and a flexural moment (My). The loading is
shown in Fig. 1. The design constants and design variables
of the problem are shown in Table 1. The optimization
process was done for multiple values of the input
parameters.
The design constraints are defined according to ACI
318: Building Code Requirement for reinforced concrete
[32], and these are listed in Table 2. The optimization
objective is given as Eq. (1), and it is the total cost of RC
column. As is the area of steel reinforcements.
 X  X 
f1 ðXÞ ¼ bh  As lCc þ As lcs Cs ð1Þ

In the optimization methodology via HS, the first stage


is the definition of the design constants and ranges of
design variables. Then, an initial solution matrix is gener-
ated and it contains HMS (harmony memory size) sets of
design variables that are randomly assigned according to
the selected range. Then, the objective function is calcu-
lated and the objective is penalized with a big value (106)
when a violation of the design constraints occurs. The
candidate solution for design variables, for example, for ith
variable (Xi), is calculated via Eq. (2) for the initial values
and the global optimization stage. Xi,min and Xi,max are
lower and upper bounds of ith design variable, and a ran-
dom number between 0 and 1 is shown as rand().
 
X i ¼ X i;min þ randðÞ X i;max  X i;min ð2Þ
After the generation of initial values, new values are
generated via global or local optimization in the iterations.
The formulations of phases are shown in Eq. (3). The type
of generation is defined according to harmony memory
consideration rate (HMCR), and it is taken as 0.5. In local
optimization, a random existing solution (k) is chosen. Xi,k
is the kth existing solution of ith design variable, and PAR
is the parameter called pitch adjusting rate that is taken as
0.25. In each iteration, the objective function for new Fig. 3 Flowchart for the working principle of the hybrid optimiza-
tion–machine learning approach
values is checked according to the reduction in the objec-
tive function for the existing solutions.

Table 3 Input parameters with


Input name Symbol Ranges Increment Unit
the design ranges
Unit cost for concrete Cc 50–100 25 $
Unit cost for steel Cs 500–1500
Column length l 3000–7000 1000 cm
Value of flexural moment Mz 1.00e ? 08–5.00e ? 08 1.00e ? 08 N
Value of axial force Ny 1.00e ? 06–5.00e ? 06 1.00e ? 06 Nm

123
Neural Computing and Applications

Table 4 Obtained errors between actual and predicted values for 2.2 Prediction approach via artificial neural
outputs in training data networks (ANNs)
Error metric h b
Artificial neural networks (ANNs) have got many different
Mean absolute error (MAE) 0.3538 0.0630
kinds in terms of formation, which were developed by
Mean absolute percentage error (MAPE) 0.0428 0.0250
inspired by the natural structure and working mechanism of
Mean square error (MSE) 0.4558 0.0109 the human central neural system, and the learning process
Root mean square error (RMSE) 0.6751 0.1043 by coordinately working together with the brain. Besides, it
is a kind of machine learning algorithm. The difference
between these systems is based on the structure related to
layer number, type of feed, learning algorithm rules, etc. In
Xi;new this study, multilayer feed-forward artificial neural net-
8  
>
< HMCR [ randðÞ; Xi;min þ randðÞ Xi;max  Xi;min works or more widely known as multi-layer perceptrons
  (MLPs), which are one of the most frequently used kinds of
¼ 1 1  
>
: HMCR\randðÞ; Xi;k þ rand ; PAR Xi;max  Xi;min
2 2 ANNs, occur due to those numerous nodes acting as arti-
ficial neurons are placed on more than one layer. These
ð3Þ
layers consist of one output layer together with one or more
hidden layers, besides one input layer. In Fig. 2, the

Table 5 Optimum results of


Inputs Outputs Other values
column designs within the test
model Cc Cs L Mz Ny b h q Min. Total Cost

50 550 3250 1.00e ? 08 2.00e ? 06 250.2149 580.0830 0.01 43.9787


75 1000 5500 3.00e ? 08 1.00e ? 06 250.0714 583.7821 0.01 123.3300
60 750 3000 2.00e ? 08 2.00e ? 06 250.0000 600.3484 0.01 53.5586
75 600 6000 3.00e ? 08 3.00e ? 06 250.3849 842.0089 0.01 154.5273
75 850 6000 3.00e ? 08 3.00e ? 06 250.0899 843.0110 0.01 179.3856
85 750 3000 2.00e ? 08 2.00e ? 06 250.0000 600.3483 0.01 64.8151
75 1100 6000 3.00e ? 08 3.00e ? 06 250.0868 843.1172 0.01 204.2655
150 1000 5000 4.00e ? 08 4.00e ? 06 278.0516 1000.0000 0.01 317.8130
100 2000 5000 4.00e ? 08 4.00e ? 06 278.0512 1000.0000 0.01 357.5738
100 1500 8000 4.00e ? 08 4.00e ? 06 278.0519 1000.0000 0.01 484.7001

Table 6 Error rates between


Predicted values by ANNs Error rates
actual optimum values and
predictions for b concerning the b Error Absolute error Absolute percentage (%) Error Square error
test model
250.2627 - 0.0478 0.0478 0.0191 0.0023
249.9610 0.1104 0.1104 0.0441 0.0122
250.1071 - 0.1071 0.1071 0.0428 0.0115
250.1850 0.2000 0.2000 0.0799 0.0400
250.1866 - 0.0968 0.0968 0.0387 0.0094
250.1070 - 0.1070 0.1070 0.0428 0.0114
250.1883 - 0.1015 0.1015 0.0406 0.0103
278.0656 - 0.0140 0.0140 0.0050 0.0002
278.0602 - 0.0090 0.0090 0.0033 0.0001
278.0636 - 0.0116 0.0116 0.0042 0.0001
MAE MAPE MSE
Mean 0.0825 0.0320 0.0097
RMSE 0.0987

123
Neural Computing and Applications

Table 7 Error rates between


Predicted values by ANNs Error rates
actual optimum values and
predictions for h with respect to H Error Absolute error Absolute percentage (%) error Square error
test model
579.8253 0.2578 0.2578 0.0444 0.0665
583.8159 - 0.0338 0.0338 0.0058 0.0011
600.0803 0.2681 0.2681 0.0447 0.0719
844.6320 - 2.6231 2.6231 0.3115 6.8805
844.6264 - 1.6154 1.6154 0.1916 2.6094
600.0547 0.2935 0.2935 0.0489 0.0862
844.6208 - 1.5036 1.5036 0.1783 2.2607
999.6422 0.3578 0.3578 0.0358 0.1280
999.5261 0.4739 0.4739 0.0474 0.2246
999.5859 0.4141 0.4141 0.0414 0.1714
MAE MAPE MSE
Mean 0.7841 0.0950 1.2500
RMSE 1.1180

Table 8 Error rates between actual and calculated values with predicted outputs for minimum total costs
Calculated values by predictions Error rates
MIN. TOTAL COST Error Absolute error Absolute percentage (%) Error Square error

43.9676 0.0111 0.0111 0.0253 0.0001


123.2827 0.0473 0.0473 0.0383 0.0022
53.5576 0.0010 0.0010 0.0019 0.0000
154.8849 - 0.3576 0.3576 0.2314 0.1279
179.7988 - 0.4132 0.4132 0.2304 0.1708
64.8111 0.0040 0.0040 0.0061 0.0000
204.7128 - 0.4473 0.4473 0.2190 0.2000
317.7153 0.0977 0.0977 0.0307 0.0095
357.4160 0.1578 0.1578 0.0441 0.0249
484.5197 0.1804 0.1804 0.0372 0.0326
MAE MAPE MSE
Mean 0.1717 0.0865 0.0568
RMSE 0.2383

structure of MLPs for the current problem is illustrated artificial neural network, is one of the learning algorithms
below. [35, 36]. This algorithm is a method based on the con-
In MLPs, these nodes are connected to each other and nection weights that are rearranged by propagating the
have a specific weight value between themselves, and obtained error to back for output values in nodes, in other
layers are arranged in the way forward flowing. From this words, the previous layer, and in this way, the error is
respect, in the process of network weights determination, minimized.
namely learning, MLPs are considered a calculation tool, For the prediction process carried out by MLPs, the
which can arrange its structure by depending on the main training dataset was ensured through numerous
internal/external information flowing along with the net- optimization analyses in terms of minimization of the total
work [33, 34]. In question, learning is also known as cost for structural design. In this context, several design
training of the network, and the back-propagation method properties determined as input parameters are varied within
benefiting from the generalized delta rule, which succes- the specific value ranges to detect the optimum section
sively makes a differential calculation to train a multilayer sizes of columns (h and b) by generating multiple optimum

123
Neural Computing and Applications

Fig. 4 Differences between 280.000 0.2200


actual values–predictions and 0.2000
changing of error rates for b 275.000
values corresponding test 0.1800
designs 270.000 0.1600
265.000 0.1400

Errors
0.1200
260.000

b
0.1000
255.000 0.0800

250.000 0.0600
0.0400
245.000
0.0200
240.000 0.0000
1 2 3 4 5 6 7 8 9 10
b-actual 250.214 250.071 250.000 250.384 250.089 250.000 250.086 278.051 278.051 278.051
b-predicted 250.262 249.961 250.107 250.185 250.186 250.107 250.188 278.066 278.06 278.064
Absolute error 0.0478 0.1104 0.1071 0.2000 0.0968 0.1070 0.1015 0.0140 0.0090 0.0116
Absolute % error 0.0191 0.0441 0.0428 0.0799 0.0387 0.0428 0.0406 0.0050 0.0033 0.0042
Square error 0.0023 0.0122 0.0115 0.0400 0.0094 0.0114 0.0103 0.0002 0.0001 0.0001

Fig. 5 Differences between 1,025.000 7.2000


actual values–predictions and
changing of error rates for h 950.000 6.0000
values corresponding test
designs 875.000
4.8000
800.000

Errors
3.6000
h

725.000
2.4000
650.000

575.000 1.2000

500.000 0.0000
1 2 3 4 5 6 7 8 9 10
h-actual 580.083 583.782 600.348 842.008 843.011 600.348 843.117 1000.00 1000.00 1000.00
h-predicted 579.825 583.815 600.080 844.632 844.626 600.054 844.620 999.642 999.526 999.585
Absolute error 0.2578 0.0338 0.2681 2.6231 1.6154 0.2935 1.5036 0.3578 0.4739 0.4141
Absolute % error 0.0444 0.0058 0.0447 0.3115 0.1916 0.0489 0.1783 0.0358 0.0474 0.0414
Square error 0.0665 0.0011 0.0719 6.8805 2.6094 0.0862 2.2607 0.1280 0.2246 0.1714

design combinations. Following this process, for the optimum column designs were determined and considered
training operation of the dataset, an application called for the training of the main prediction model. Also, a
‘‘Neural Net Fitting’’ was benefited from the machine flowchart that provides to understand of the working
learning toolbox belonging MATLAB R2018a [37]. Here, principle of the proposed approach is given in Fig. 3.
the neural network structure has one hidden layer including
6 neurons. Also, the generated real optimization data were
divided into three groups for training, validation and testing 3 Numerical applications
of network model. So, while training data are including
70%, each of validation and test groups is including 15% of In this section, the success, efficiency, and usability of the
whole results. This process is also realized with k-fold hybrid model are observed via the measurement of error
cross-validation (k = 10) methodology for dividing of rates and evaluation of different test models. Also, in terms
optimal results into these three data groups. Besides, all of of different test structural models, deviation namely errors
the described ranges in terms of input parameters are are determined to occur in the generalizability of the
shown in Table 3. It can be recognized that a total of 1875 approach. The error calculations were presented for each

123
Neural Computing and Applications

Fig. 6 Differences between 500.000 0.5000


actual values–predictions and
450.000 0.4500
changing of error rates for
minimum cost values 400.000 0.4000

Minimum Total Cost


corresponding test designs
350.000 0.3500
300.000 0.3000

Errors
250.000 0.2500
200.000 0.2000
150.000 0.1500
100.000 0.1000
50.000 0.0500
0.000 0.0000
1 2 3 4 5 6 7 8 9 10
Min. Cost-actual 43.9787 123.330 53.5586 154.527 179.385 64.8151 204.265 317.813 357.573 484.700
Min. Cost-predicted 43.9676 123.282 53.5576 154.884 179.798 64.8111 204.712 317.715 357.416 484.519
Absolute error 0.0111 0.0473 0.0010 0.3576 0.4132 0.0040 0.4473 0.0977 0.1578 0.1804
Absolute % error 0.0253 0.0383 0.0019 0.2314 0.2304 0.0061 0.2190 0.0307 0.0441 0.0372
Square error 0.0001 0.0022 0.0000 0.1279 0.1708 0.0000 0.2000 0.0095 0.0249 0.0326

output parameter (h and b) concerning training data. All of According to the obtained results, the main prediction
the error values are shown in Table 4 in terms of four model can be accepted as an extremely powerful prediction
different metrics. tool thanks to the occurred error ratios for each output
In the second phase of the prediction process, a test parameter. The cause of this is that values of the detected
dataset containing several design combinations was gen- four error measurements were found at a quite low level.
erated thanks to the smallness of whole error amounts and Also, it can be seen that the MAPE metric, which expresses
thus efficiency and convergence performance of the main the percentage deviation according to the actual value of
ANNs model. In this meaning, for the generated dataset, data, is approximately 4.2% and 2.5% for h and b param-
actual optimization information of column designs and the eters for all of the training set, respectively. For this reason,
predicted values of outputs together with error rates the developed ANN model can be utilized for designing
according to actual outcomes are represented in Tables 5, 6 different column designs, which have unknown values for
and 7, respectively. output variables, through ANNs made pretty close predic-
On the other side, to validate the success and reflect the tions to data models by small deviating from actual opti-
performance of the prediction model, minimum total cost mum data.
values, which were obtained via optimization analyses, As concerns the test designs, the defined error metrics
were checked by calculating with the predicted values for were also measured in terms of b and h outputs together
outputs parameters and the other design inputs within with calculated total cost values, where the differences
Table 3 (such as C c , C s , L; etc.). The observed results are between actual and predicted values were provided at
shown in Table 8. Also, the value differences between the quite small rates for each output parameter. While all of
actual data and predictions together with error rates for the the combinations within test data were investigated, a
outputs are illustrated in Figs. 4, 5 and 6. result can be extracted related that the prediction model
is also extremely successful concerning unknown column
information (Tables 6 and 7). Also, it can be expressed
4 Conclusion that the prediction model can be accepted as a very
efficient tool due to that the total cost amounts could be
In the present study, an prediction model was created. It also directly calculated via the predicted output values
was generated for the determining of optimum section by deviating with a pretty low ratio as almost 0.24.
properties together with calculating of the total cost. In this Additionally, if these expressions must be stated, for
scope, by benefiting a machine learning methodology several test designs, which exceed the design limitations
based on ANNs, a large-scale dataset for column designs (Table 3), the ANNs model can carry out the desired
according to design constants and ranges in Table 1 was prediction process for each output parameter besides total
ensured in terms of training, and several different test cost. Also, all of the success, efficiency and convergence
models for them were controlled and evaluated via the performance of the prediction process are shown in
developed ANNs model. Figs. 4 and 5.

123
Neural Computing and Applications

As a result of the current study, both optimum values of 12. Koo C, Hong T, Hyun C (2011) The development of a con-
output parameters and objective function as the minimum struction cost prediction model with improved prediction capac-
ity using the advanced CBR approach. Expert Syst Appl
total cost can be directly and rapidly found through the 38(7):8597–8606
usage of an advanced prediction model, which was gen- 13. Magdum SK, Adamuthe AC (2017) Construction cost prediction
erated with ANNs. In that case, an iterative optimization using neural networks. ICTACT J Soft Comput 8(1):1549–1556
process was not needed to find the optimum design of 14. Shutian F, Tianyi Z, Ying Z (2017) Prediction of construction
projects’ costs based on fusion method. Eng Comput
different columns. This tool provides the detection of the 34(7):2396–2408
required values for parameters to pre-design of column 15. Chakraborty D, Elhegazy H, Elzarka H, Gutierrez L (2020) A
structures, too. For this respect, besides these advantages, novel construction cost prediction model using hybrid natural and
both cost and time efficiency statements can be made light gradient boosting. Adv Eng Inform 46:101201. https://doi.
org/10.1016/j.aei.2020.101201
possible for the defined optimization processes. 16. Yucel M, Bekdaş G, Nigdeli SM, Sevgen S (2019) Estimation of
optimum tuned mass damper parameters via machine learning.
J Build Eng 26:100847
Data availability The datasets generated during and/or analyzed 17. Yucel M, Kayabekir AE, Nigdeli SM, Bekdaş G (2020) Optimum
during the current study are available from the corresponding author design of carbon fiber-reinforced polymer (CFRP) beams for
on reasonable request.The authors declare that they have no known shear capacity via machine learning methods: optimum predic-
competing financial interests or personal relationships that could have tion methods on advance ensemble algorithms–bagging combi-
appeared to influence the work reported in this paper. nations. In: Bekdaş G, Nigdeli SM, Yücel M (eds) Artificial
intelligence and machine learning applications in civil, mechan-
ical, and industrial engineering. IGI Global, pp 85–103. https://
References doi.org/10.4018/978-1-7998-0301-0.ch005
18. Yucel M, Bekdaş G, Nigdeli SM (2021) Decision-making model
based multilayer perceptrons for estimation of optimum design
1. Lertpalangsunti N, Chan CW, Mason R, Tontiwachwuthikul P properties for truss structure. Challenge 7(4):170–179
(1999) A toolset for construction of hybrid intelligent forecasting 19. Bekdaş G, Yücel M, Nigdeli SM (2021) Estimation of optimum
systems: Application for water demand prediction. Artif Intell design of structural systems via machine learning. Front Struct
Eng 13(1):21–42 Civ Eng 15(6):1441–1452. https://doi.org/10.1007/s11709-021-
2. Msiza IS, Nelwamondo FV, Marwala T (2008) Water demand 0774-0
prediction using artificial neural networks and support vector 20. Yücel M, Nigdeli SM, Kayabekir AE, Bekdaş G (2021) Opti-
regression. J Comput 3(11):1 mization and artificial neural network models for reinforced
3. Brentan BM, Luvizotto E Jr, Herrera M, Izquierdo J, Pérez- concrete members. In: Carbas S, Toktas A, Ustun D (eds) Nature-
Garcı́a R (2017) Hybrid regression model for near real-time urban inspired metaheuristic algorithms for engineering optimization
water demand forecasting. J Comput Appl Math 309:532–541 applications. Springer Singapore, Singapore, pp 181–199. https://
4. Zubaidi SL, Ortega-Martorell S, Al-Bugharbee H, Olier I, doi.org/10.1007/978-981-33-6773-9_9
Hashim KS, Gharghan SK, Kot P, Al-Khaddar R (2020) Urban 21. Yücel M, Bekdaş G, Nigdeli SM, Kayabekir AE (2021) An
water demand prediction for a city that suffers from climate artificial intelligence-based prediction model for optimum design
change and population growth: gauteng province case study. variables of reinforced concrete retaining walls. Int J Geomech
Water 12(7):1885 21(12):04021244
5. Hu W, Xiao X, Xie D, Tan T, Maybank S (2004) Traffic accident 22. Camp CV, Assadollahi A (2015) CO2 and cost optimization of
prediction using 3-D model-based vehicle tracking. IEEE Trans reinforced concrete footings subjected to uniaxial uplift. J Build
Veh Technol 53(3):677–694 Eng 3:171–183
6. Sikka S (2014) Prediction of road accidents in Delhi using back 23. Paya-Zaforteza I, Yepes V, Hospitaler A, Gonzalez-Vidosa F
propagation neural network model. Int J Comput Sci Eng Technol (2009) CO2-optimization of reinforced concrete frames by sim-
(IJCSET) 5(08):798–804 ulated annealing. Eng Struct 31(7):1501–1508
7. Wenqi L, Dongyu L, Menghua Y (2017) A model of traffic 24. Yepes V, Gonzalez-Vidosa F, Alcala J, Villalba P (2012) CO
accident prediction based on convolutional neural network, In: 2-optimization design of reinforced concrete retaining walls
2017 2nd IEEE international conference on intelligent trans- based on a VNS-threshold acceptance strategy. J Comput Civ
portation engineering (ICITE), pp. 198–202 Eng 26(3):378–386
8. Huang HC, Chang YL, Lan TH, Chiu HJ, Liu WM, Lee TJF 25. Arama ZA, Kayabekir AE, Bekdaş G, Geem ZW (2020) CO2 and
(2008) Prediction of optimal lithium doses for Taiwanese psy- cost optimization of reinforced concrete cantilever soldier piles: a
chiatric patients. J Clin Pharm Ther 33(2):115–121 parametric study with harmony search algorithm. Sustainability
9. Grossi E, Podda GM, Pugliano M, Gabba S, Verri A, Carpani G, 12(15):5906
Buscema M, Casazza G, Cattaneo M (2014) Prediction of optimal 26. Yeo D, Potra FA (2015) Sustainable design of reinforced concrete
warfarin maintenance dose using advanced artificial neural net- structures through CO2 emission optimization. J Struct Eng
works. Pharmacogenomics 15(1):29–37 141(3):B4014002
10. McDonald JF, Mezencev R, Long TQ, Benigno B, Bonta I, Del 27. Bekdaş G, Nigdeli SM (2016) Optimum design of reinforced
Priore G (2015) Accurate prediction of optimal cancer drug concrete columns employing teaching learning-based optimiza-
therapies from molecular profiles by a machine-learning algo- tion. Chall J Struct Mech 2(4):216–219
rithm. J Clin Oncol 33(15_suppl):e22182–e22182. https://doi.org/ 28. Bekdaş G, Nigdeli SM (2015) Optimization of reinforced con-
10.1200/jco.2015.33.15_suppl.e22182 crete columns subjected to uniaxial loading. In: Lagaros ND,
11. Huang C, Mezencev R, McDonald JF, Vannberg F (2017) Open Papadrakakis M (eds) Engineering and applied sciences opti-
source machine-learning algorithms for the prediction of optimal mization. Springer International Publishing, Cham, pp 399–412.
cancer drug therapies. PLoS ONE 12(10):e0186906 https://doi.org/10.1007/978-3-319-18320-6_21

123
Neural Computing and Applications

29. Bekdas G, Nigdeli SM (2014) The optimization of slender rein- International conference, April 1994. Cary-USA, SAS Institute,
forced concrete columns. PAMM 14(1):183–184 pp. 1538–1550
30. Nigdeli SM, Bekdas G, Kim S, Geem ZW (2015) A novel har- 36. Chou JS, Lin CW, Pham AD, Shao JY (2015) Optimized artificial
mony search based optimization of reinforced concrete biaxially intelligence models for predicting project award price. Autom
loaded columns. Struct Eng Mech Int J 54(6):1097–1109 Constr 54:106–115
31. Geem ZW, Kim JH, Loganathan GV (2001) A new heuristic 37. MATLAB Mathworks Matlab (2018a) Neural net fitting, https://
optimization algorithm: harmony search. Simulation www.mathworks.com/ help/deeplearning/ref/neuralnetfitting-
76(2):60–68. https://doi.org/10.1177/003754970107600201 app.html, last accessed 2021/11/01.
32. ACI Committee. Building code requirements for structural con-
crete (ACI 318-05) and commentary (ACI 318R-05) Publisher’s Note Springer Nature remains neutral with regard to
33. Han J, Kamber M (2006) Data mining: concepts and techniques, jurisdictional claims in published maps and institutional affiliations.
2nd edn. Morgan Kaufmann, California
34. Olivas ES, Guerrero JDM, Sober MM, Benedito JRM, López AJS
Springer Nature or its licensor (e.g. a society or other partner) holds
(eds) (2009) Handbook of research on machine learning appli-
exclusive rights to this article under a publishing agreement with the
cations and trends: algorithms, methods, and techniques. IGI
author(s) or other rightsholder(s); author self-archiving of the
Global, Hershey, New York
accepted manuscript version of this article is solely governed by the
35. Sarle WS (1994) Neural networks and statistical models. In:
terms of such publishing agreement and applicable law.
Proceedings of the nineteenth annual sas users group

123

You might also like