Sie sind auf Seite 1von 7

Advances in Engineering Software 41 (2010) 886892

Contents lists available at ScienceDirect

Advances in Engineering Software


journal homepage: www.elsevier.com/locate/advengsoft

Prediction of California bearing ratio (CBR) of ne grained soils by AI methods


T. Taskiran
Department of Civil Engineering, Dicle University Engineering and Architecture Faculty, 21280 Diyarbakir, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: Advances in eld of articial intelligence (AI) offers opportunities of utilizing new algorithms and models
Received 25 September 2009 that enable researchers to solve the most complex systems. As in other engineering elds, AI methods
Received in revised form 2 December 2009 have widely been used in geotechnical engineering. Unlikely, there seems quite insufcient number of
Accepted 19 January 2010
research related to the use of AI methods for the estimation of California bearing ratio (CBR). There were
Available online 16 March 2010
actually some attempts to develop prediction models for CBR, but most of these models were essentially
statistical correlations. Nevertheless, many of these statistical correlation equations generally produce
Keywords:
unsatisfactory CBR values. However, this paper is likely one of the very rst research which aims to inves-
Articial neural network
Gene expression programming
tigate the applicability of AI methods for prediction of CBR. In this context, articial neural network
California bearing ratio (ANN) and gene expression programming (GEP) were applied for the prediction of CBR of ne grained
soils from Southeast Anatolia Region/Turkey. Using CBR test data of ne grained soils, some proper mod-
els are successfully developed. The results have shown that the both ANN and GEP are found to be able to
learn the relation between CBR and basic soil properties. Additionally, sensitivity analysis is performed
and it is found that maximum dry unit weight (cd) is the most effective parameter on CBR among the oth-
ers such as plasticity index (PI), optimum moisture content (wopt), sand content (S), clay + silt content
(C + S), liquid limit (LL) and gravel content (G) respectively.
2010 Elsevier Ltd. All rights reserved.

1. Introduction There were attempts to develop prediction models, considering


that the CBR of soils are affected by the soil index properties in
Engineering projects that involve earthwork such as road some ways. However, most of these previous models were essen-
embankment, airport runways and pavement construction often tially statistical correlations between CBR and classication data
requires construction of a proper subbase layer. These subbase lay- and/or index properties of soil. Many researchers have conducted
ers should meet certain engineering requirements such as bearing studies to show the effect of soil types and characteristics on CBR
capacity, settlement and swell properties. Therefore, a method of values. [25]. Some efforts devoted to correlate CBR with soil grain
assessment of such earthworks is necessary and takes important distribution and plasticity. Among them, Black [2] developed a cor-
part among geotechnical and road engineering applications. Cali- relation between CBR and plasticity index (PI) for cohesive soils.
fornia bearing ratio (CBR) is one of the frequently used index test Using the concept of suitably index, which varies with plasticity
to assess the stiffness modulus and shear strength of subgrade. and grading characteristics, a correlation for CBR was suggested
CBR is actually an indirect measure which indicates comparison by de Graft-Johnson and Bhatia [3]. In there, the suitability index
of the strength of subgrade material to the strength of standard was dened as below.
crushed rock mentioned in percentage values.
A
Engineers always come across with difculties in obtaining rep- Suitability Index 1
resentative CBR values for design. Due to limited budget and poor LL logPI
planning conditions, insufcient soil investigation data are ob- where A = percentage passing 2.4 mm BS sieve, LL = liquid limit,
tained in many cases. On the other hand, laboratory CBR test is PI = plasticity index.
not only laborious and time consuming but also the results some- Agarwal and Ghanekar [4] tried to develop a correlation equa-
times are not accurate due to the sample disturbance and poor tion between CBR and either liquid limit (LL), plastic limit (PL) or
quality of laboratory testing conditions. Therefore, development plasticity index (PI). However they were not able to nd any signif-
of prediction models might be useful and become a base of the icant correlation among these parameters. Instead, they found an
judgment on the validity of the CBR values [1]. improved correlation when optimum moisture content (wopt) and
LL are included. Hence, they suggested a correlation which was
only of sufcient accuracy for preliminary identication of mate-
E-mail address: taha@dicle.edu.tr rial. This correlation was presented below.

0965-9978/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advengsoft.2010.01.003
T. Taskiran / Advances in Engineering Software 41 (2010) 886892 887

CBR 2:0  16:0  logwopt 0:07  LL 2 on soft computing, enable researchers to solve the most complex
systems in different ways. The use of forecast methods not based
The effect of each parameters such as grading constant (A),
on physics equations, such as ANN and GEP methods are becoming
mean grain size (d50), size factor (SF) and the granulometric mod-
widespread in various engineering elds. Unlikely, there is, proba-
ulus (MG) on CBR were analyzed by Doshi et al., [6]. It was found
bly, very poor number of research in literature which utilizes AI
that the grading constant (A) is the best parameter to describe
methods for the estimation of CBR.
the inuence of grain size distribution on CBR. In the same study,
In the content of this study, taking the power of AI methods in
it was also reached that CBR is most dependent on maximum dry
solving complex systems and capability of processing much more
density (MDD) and least dependent on wopt. Using A, MDD and wopt
parameters, ANN and GEP methods were applied for the prediction
as independent variables, several equations for CBR were
of CBR of ne grained soils. To achieve this, 151 CBR test data of
presented.
ne grained soils were provided from 9th Regional Directorate of
Stephens [7] carried out an investigation, in which archival data
Highways Diyarbakir which was carried out for the assessment
were used, to evaluate performance of existing models for some se-
of base/subbase materials of the provincial roads of four city of
lected natal soils. He described the relationships between CBR and
Southeast Anatolia Region. The data used were belonging to plas-
various classication parameters (in both simple and multivariate
ticity properties (liquid limit, plasticity index), compaction proper-
forms) and further examined that these models were found to be
ties (optimum moisture content, maximum dry unit weight) and
generally unsatisfactory. In this study, the lack of any suitable cor-
gradation properties (clay + silt, sand and gravel fractions).
relations for universally use was discussed and a good relationship
between CBR and maximum swell was examined. Inuence of clay
fraction on CBR was reported and the interim use of the shrinkage 2. Test soil database development and laboratory testing
and grading moduli to obtain minimum CBR values for shrinking
and non-shrinking soils respectively was proposed. Another meth- In order to obtain data to establish models, 151 CBR test data
od for the estimation of CBR were presented by the British High- belonging to A-4, A-5, A-6, A-7 (AASHTO M 145) soil groups were
way Agency [8] made use of plasticity index for British soils selected among 354 total tests. The remaining of the test data were
compacted at natural moisture content where the correlations belonging to A1-A2-A3 soil groups. Tests were carried out for the
were given in the format of a table. suitability assessment of soils to be use as base and subbase mate-
National Cooperative Highway Research Program of United rial. Therefore, tested soils were collected from connecting roads of
States of America (NCHRP) through the Guide for Mechanical to be constructed for four cities (Diyarbakir, Mardin, Siirt and
Empirical Design of New and Rehabilitated Pavement Structures S
anlurfa) of Southeast Anatolia Region.
[5] suggested some correlations that describe the relationship be- All tests were conducted following the standard testing proce-
tween soil index properties and CBR. Two equations were proposed dures stipulated in the ASTM Standard, e.g., ASTM D-422 for
where one was for non-plastic coarse-grained material and the mechanical analysis, ASTM D-4318 for liquid limit, and plastic lim-
other was for soils which contain 12% nes and have some plastic- it tests. The laboratory compaction tests were conducted by using
ity. The best-tted equation proposed by NCHRP for clean, coarse- the standard proctor compaction effort in accordance with the
grained soil was given as below. standard test procedures of ASTM D-698. CBR tests were con-
ducted in accordance with AASHTO T 193. The samples were com-
CBR 28:09D60 0;358 3 pacted to optimum moisture content for determination of CBR
where D60 = diameter at 60% passing (mm). value at maximum dry density. The soil samples were soaked for
This equation is limited to D60 values varying between 0.01 mm 4 days before testing. Surcharge weights in the form of annular
and 30 mm. For D60 less than 0.01 mm, the recommended value of discs with a mass of 2 kg were placed on the soil sample to simu-
CBR is 5% whereas CBR value of 95% is recommended for D60 great- late the weight of pavement materials overlying subgrade. A cylin-
er than 30 mm. For plastic ne grained soils, the chosen soil index drical plunger is made to penetrate the soil sample at a rate
properties to correlate CBR are the percentage passing No. 200 US specied in AASHTO T 193. Tests were performed at the centre lab-
sieve or 0.075 mm size sieve and plasticity index. The suggested oratory of 9th Regional Directorate of Highways Diyarbakir.
equation by NCHRP is given below.
75 3. View of the articial neural network
CBR 4
1 0:728wPI
Due to difculties in solutions of the complex engineering sys-
where w = passing No. 200 US sieve (%). tems, in the past decade, researchers have started to study on arti-
However, when the previous researches are evaluated in gen- cial neural network (ANN) inspired by the behavior of human
eral, it is seen that satisfactory correlations could not be obtained brain and nervous system. Each ANN model can be differently or-
in many cases (even for some local soils). Many of proposed corre- ganized according to the same basic structure. There are three
lation equation do not have high enough degree of correlation coef- main layers in ANN structure; a set of input nodes, one or more lay-
cient. Therefore, it is possible to conclude that the estimation of ers of hidden nodes, and a set of output nodes. Each layer basically
CBR is difcult using the conventional statistical methods. Besides, contains a number of neurons working as an independent process-
for most studies CBR was expressed in terms of only one or two ing element and densely interconnected with each other. The neu-
soil properties group. The soil properties groups are used here- rons using the parallel computation algorithms are simply
in to refer to soil parameters belonging to plasticity, gradation and compiled with an adjustable connection weights, summation func-
compaction properties. For example PL belongs to plasticity group tion and transfer function. The methodology of ANNs is based on
while wopt is in compaction group. The reason for this unfavorable the learning procedure from the data set presented it from the in-
picture is thought to be due to inadequate representation of soil put layer and testing with other data set for the validation. A net-
properties group, existence of complex relationships among the work is trained by using a special learning function and learning
parameters and unpowerful methods of calculations. rule. In ANNs analyses, some function called learning functions is
On the other hand, advances in the eld of articial intelligence used for initialization; training, adaptation and performance func-
inuence many science topics as well as geotechnical engineering tion. During the training process, a network is continuously up-
applications. New algorithms and models, especially those based dated by a training function which repeatedly applies the input
888 T. Taskiran / Advances in Engineering Software 41 (2010) 886892

variables to a network till a desired error criterion is obtained. tor to test the accuracy of the developed network. Therefore, data
Adapt functions is employed for the simulation of a network, while group from 119 different tests was used to train the network and
the network is updated for each time step of the input vector be- data group which consist of 32 test results used for testing the
fore continuing the simulation to the next input. Performance network.
functions are used to grade the network results. In this study, Gra- The ANN toolbox of MATLAB computer aided Software [11] was
dient descent with momentum and adaptive learning rate (train- used to perform the necessary computations. In order to develop
gdx), gradient descent with momentum weight and bias learning the most appropriate ANN architecture for each model, the number
function (learngdm) and mean square error (MSE) were used for of neurons in the hidden layer and a number of multi layer net-
training function; adapt function and performance function, works with different transfer functions were tried to predict best
respectively. In the learning stage, network initially starts by ran- CBR values. They were, therefore, varied until the convergence
domly assigning the adjustable weights and threshold values for was achieved in the mean squared error. The feed-forward neural
each connection between the neurons in accordance with selected networks which consist of multilayer perceptions trained back-
ANNs model. After the weighted inputs are summed and added the propagation algorithms were employed for this study. In Fig. 1,
threshold values, they are passed through a differentiable non-lin- the architecture of the Model I formed by using feed-forward arti-
ear function dened as a transfer function. This process is contin- cial neural networks with seven inputs is shown. To evaluate the
ued, until a particular input captures to their output (i.e., target) results of the developed ANN models, the coefcient of determina-
or as far as the lowest possible error can be obtained by using an tion (R2), the mean square errors (MSE), mean (xmean) and standard
error criterion. In other words the network training is the determi- deviation (Sx) were used as statistical verication tool.
nation of the weights and the biases [9,10]. An ANN model can be The normalized data set was used to train neural network to ob-
differently composed in terms of architecture, learning rule and tain the nal weights. At the end of the analyses the network out-
self organization. The most widely used ANNs are the feed-for- puts were post processed to convert the data back into
ward, multilayer perceptions trained by back-propagation algo- unnormalized units. The connection weights and biases for Model
rithms based on gradient descent method (FFBP). This algorithm I of ANN was given in Table 2.
can provide approximating to any continuous function from one -
nite-dimensional space to another for any desired degree of accu- 5. Gene expression programming (GEP)
racy. The superiority of FFBP is that it sensitively assigns the initial
weights values and therefore it may yield closer results than the Gene expression programming, which is an algorithm based on
each other. Also this algorithm has easier application and shorter genetic algorithms (GA) and genetic programming (GP) was rst
training duration. proposed by Ferreira [12]. It has several features similar to biolog-
ical evolution to evolve a computer program encoded in linear
chromosomes of xed-length. The principal objective of GEP is to
4. Development of articial neural networks models for the
develop a mathematical function that ts a set of data presented
prediction of CBR
to GEP model. For the mathematical equation the GEP process car-
ries out the symbolic regression by means of the most of the genet-
The potential of using ANNs for the estimation of CBR were
ic operators of GA. There are some differences between GEP and
investigated by developing various proper ANN models. The vari-
GA. GA depicts the individuals as symbolic strings of xed-length
ables which have been identied in the literature review and oth-
(chromosomes), GP depicts them as non-linear entities of different
ers which appear to be potentially inuential to CBR value were
size and shapes (parse trees). Whereas in GEP the individuals is
used for ANN prediction models. Parameters belonging to three
represented as simple strings of xed-length which are subse-
groups of soil index properties which reect plasticity, gradation
quently expressed as expression trees of different size and shape
and compaction properties are utilized. Therefore, totally seven ba-
[13,14]. The primary GEP process is presented in Fig. 2. In GEP anal-
sic soil parameters such as liquid limit (LL), plasticity index (PI),
ysis, ve parameters such as the function set, terminal set, tness
dry unit weight (cd), optimum water content (wopt), No: 200 sieve
function, control parameters and stop condition are used. After
passing percent (clay + silt), sand percent (S) and gravel percent (G)
encoding of the problem for the candidate solution and the tness
were taken into consideration as input parameters for the ANN
function is specied, an initial population of viable individuals
models. The soil parameters used herein can easily be obtained
(chromosomes) is randomly created and then each chromosome
by routine laboratory tests. To obtain the best model that governs
is converted into an expression tree corresponding to a mathemat-
CBR, seven different ANN models were established by proper com-
bination of input data. The developed network models and archi-
tectures used for the training are given in Table 1. For this study
(LL)
21% of the original data was extracted at random and used for
the test stage. The remaining 79% of the data were used to train
the neural network. Testing data set were not used during develop- (PI)
ment of the neural network, and so they could form a good indica-
k
Table 1
The ANN and GEP models used for the prediction of CBR. wopt
CBR
Model Inputs Structure Transfer function
(ANN) (ANN) C+M
I (LL), (PI), ck, wopt, (C + S), S, G 7-4-1 logsiglogsig
II (LL), (PI), ck, wopt, (C + S), S, 6-5-1 tansigtansig S
III (PI), ck, wopt, (C + S), S 5-4-1 logsiglogsig
IV (PI), ck, wopt, (C + S) 4-4-1 tansigtansig
V (PI), ck, (C + M), S 4-5-1 logsiglogsig G
VI (PI), ck, (C + M) 3-4-1 logsiglogsig
VII ck, (C + M), wopt 3-5-1 logsiglogsig Fig. 1. The architecture of feed-forward articial neural networks with seven
inputs.
T. Taskiran / Advances in Engineering Software 41 (2010) 886892 889

Table 2
Connection weights and biases for ANN Model I.

Weights input layer and hidden layer


52.36109 6.84089 14.09214 65.86439 50.35212 19.68129 15.30863
67.54092 21.60191 43.18565 16.70809 74.14464 103.84416 49.24983
1.49602 2.65625 0.82929 1.72920 2.90543 1.27760 1.32442
50.01612 2.65625 0.82929 1.72920 2.90543 1.27760 1.32442
Weights hidden layer and output layer
104.27030 1.12246 5.30735 103.51451
Biases in hidden layer
47.09824 22.00911 1.66650 50.59088
Biases in output layer
2.70634

Whereas, the terminal symbols in the tail can be solely the con-
Start
stants and the independent variables of the problem like (1, a, b, c).
Initial population creation
5.2. GEP operators
Chromosome expression as ET

ET execution In GEP, the selection, mutation, transposition, and cross-over


(recombination) are the main operators. The chromosomes are
Fitness evaluation modied to obtain better tness score for the next generation by
means of these operators.
Yes The operator rates that are specied at the beginning of the
Terminate? Stop model constructions show a certain probability of a chromosome.
In general the mutation rate is recommended to be ranging from
No 0.001 to 0.1. On the other hand the transposition operator and
Chromosome selection cross-over operator are suggested to be 0.1, 0.4, respectively [16].

Reproduction
6. GEP model development
New generation creation
The main purpose for development of GEP model was to gener-
ate the mathematical functions which predict CBR value. In this
Fig. 2. GEP algorithm. study, various GEP models in which different numbers of input
variables are used were developed for the prediction of California
bearing ratio. The input parameters used for GEP models are the
ical equation. The predicted values are subsequently compared same as the ANN models and also presented in Table 1. Similar
with the actual one and the tness score for each chromosome is to ANN models, the total data used for GEP were arranged in a for-
computed. If the satisfactory tness score is obtained, the algo- mat of seven input parameters that cover liquid limit (LL), plastic-
rithm is terminated. If it is not so, some of the chromosomes are ity index (PI), dry unit weight (cd), optimum water content (wopt),
chosen via roulette-wheel sampling and then mutated to obtain No. 200 sieve passing percent (C + S), sand percent (S) and gravel
the new generations. This process is continued until satisfactory percent (G). In order to develop the best GEP model, the parame-
tness score is obtained and then non-coding of the chromosomes ters such as; the number of chromosomes were varied between
are performed for the best solution of the problem [15]. 25 and 35, number of genes between 3 and 6, and the head size be-
tween 7 and 12. During the development of GEP models, it was ob-
served that the same GEP parameters were obtained for the best
5.1. GEP elements three models and these parameters were presented in Table 3. Re-
sults have shown that Model I and Model V are two best GEP model
There are two main elements in GEP; the chromosomes and the which have better statistical performance. As can be seen from Ta-
expression trees. The chromosome is consists of one or more genes ble 1, Model I has seven and Model V has four input parameters
and it represents a mathematical equation. Two bilingual and con-
clusive languages called Karva Language (the language of the genes
and the language of the expression trees (ET)) are used for coding Table 3
any mathematical equation. This feature of the GEP is useful for the GEP parameters obtained for the best three models.

precisely inferring of the genotype. The GEP genes composed of Chromosomes 30


two parts called the head and the tail. The head of a gene that in- Fitness function error type R2
cludes some mathematical functions, variables and constants is Number of the genes 5
Head size 10
utilized for encoding of a function. But, the tail of a gene that in-
Linking function +
cludes only variables and constants are employed as supplemen- Function set +, , , /,
p
tary terminal symbols. These symbols are used, in case the Mutation rate 0.044
terminal symbols in the head are not adequate to encode a func- One-point recombination rate 0.3
tion. Many functions can be used in the head of gene such as be ba- Two-point recombination rate 0.3
Inversion rate 0.1
sic arithmetic operators, trigonometric functions or any other
Transposition rate 0.1
mathematical or user-dened functions (+, , *, /, p, sin, cos).
890 T. Taskiran / Advances in Engineering Software 41 (2010) 886892

and mathematical functions for both models are presented below. posed model gives R > 0.8, there is a strong correlation between
In order to prevent any confusion, the ne content (C + S) was de- measured and predicted values for overall the data available in
noted by F herein. the database. In addition to this, the statistical performance of
GEP Model I any model is evaluated in terms of some error criteria such as
mean squared error (MSE); mean absolute error (MAE) which is
CBR cd =PI=S F  LL  31:99  LL 11:99 signicant criteria as well as R value, since sometimes a model
31:99 sqrtLL  wopt G  PI  PI=  24:55 with high R2 value may exhibit high MSE or MAE value. Taking
these criterions into consideration, it can be expressed that the
cd cd  cd  wopt  cd =F  PI  11:30  cd
best two models of GEP are Model I and Model V while it is Model
44:14  G  F  G PI  G=18:91 I and Model VI for ANN, in terms of higher R2 and lower MSE. The
 PI=18:91 sqrtF=F 1:04  S PI  22:56 magnitudes of errors for these models appear to be in reasonable
level to use them in prediction of CBR of soils. Moreover, it can eas-
S S
ily be seen that Model I is the best model for both GEP and ANN
models in terms of all statistical indicators and the performance
GEP Model V of the model is specied in Table 5. Estimated CBR values obtained
CBR sqrtsqrtS=F  34:75 cd  PI sqrt8:28  F with Model I of ANN and GEP are graphically compared with the
measured CBR values in Figs. 3 and 4 respectively. As seen from
cd  S=cd  cd  21:49=cd cd  21:49=cd the gures, both ANN and GEP model predictions are well-agreed
2:091=cd =2:37=PI cd 2:37  PI  F S with the experimental data.
A comparison in terms of statistical performance of GEP and
S  cd 19:35  2PI  cd =23:02
ANN of Model I is believed to be useful and is also made herein.
sqrtsqrtPI=1:81  F PI  cd  1:81  cd As mentioned above, a prediction model with higher R2 and lover
 36:31 MSE indicates better agreement with actual (experimental) data.
When ANN and GEP models are compared in this respect, it can
be seen that R2 of the both models are close while MSE of GEP mod-
el is lower. This means, deviation around the regression line is
7. Results and discussion smaller for GEP model. Despite of such good aspect of GEP, it
underestimates relatively higher values of test population (approx-
The objective of this section of the paper is to examine, discuss imately for CBR values greater than 10%) as can be seen from Fig. 3.
and compare the results obtained from both ANN and GEP models. On the other hand, the statistical parameters such as; the mean
As previously mentioned, developed ANN and GEP models were (xmean), standard deviation (Sx), coefcient of variation (Cv), skew-
tested by data sets which were not employed in the training stage. ness (Csx), maximum (xmax) and minimum (xmin) reects the distri-
To evaluate how accurate the results of the developed models are, bution properties of any data set. It can be seen that, Sx, Cv, Csx and
the coefcient of determination (R2), mean square error (MSE), xmax of ANN model are better than GEP, since they are in better
mean (xmean) and standard deviation (Sx) were used as statistical agreement with the distribution parameters of experimental val-
verication tool. ues (see Table 5). Thus, if generally evaluated, while GEP of Model I
The statistical performances of both ANN and GEP models are is statistically slightly better than ANN (in terms of R2 and MSE),
summarized in Table 4. As far as Table 4 is concerned, satisfactory ANN model have better distribution properties and ANN produces
agreement between the model predictions and experimental data more reliable CBR for relatively higher values.
is observed for most of models. In Table 4, it can also be seen that Sensitivity analysis is also performed to see how much the input
the performance of each model obtained with both methods (ANN parameters are effective on output CBR parameter (see Fig. 5). As
and GEP) are similar in terms of R2 and MSE for the majority of can be seen from Fig. 5, dry unit weight are found to be the most
models. This phenomenon can be interpreted as the inputs to effective parameter on CBR and then PI, wopt, S, C + S, LL, G respec-
establish the models (given in Table 1), are selected accurately. tively. This result is in agreement with research reported by Doshi
On the other hand, in statistics, the overall error performances of et.al. [6]. However, it is known that the dry unit weight is one of
the relationship between two groups can be interpreted from coef- the signicant parameter on many engineering properties of soils
cient of correlation (R) values. According to Smith [17], if a pro- such as bearing capacity, swelling, and shrinkage characteristics.
Remembering CBR is the indirect measure of stiffness modulus
Table 4 and shear strength characteristics of soils, such impact of dry unit
Statistical performance of ANN/GEP models. weight on CBR could be considered as an expected result.
To see the performance of few correlation equations for the
Model R2 MSE xmean Sx
present soil, CBR values are calculated by means of correlation pro-
I 0.910 2.207 5.641 4.265
0.918 1.287 5.75 3.11
II 0.641 8.679 5.428 4.275
Table 5
0.882 4.36 5.231 3.394
Statistical performance of Model I.
III 0.523 29.581 6.445 7.887
0.812 4.77 5.493 3.699 ANN model GEP model

IV 0.807 4.432 5.802 4.208 Measured Predicted Measured Predicted


0.821 5.125 5.611 3.317 R 2
0.910 0.918
V 0.838 4.271 5.978 5.205 MSE 2.207 1.287
0.911 3.491 5.696 3.357 xmean 5.98 5.64 5.98 5.75
Sx 4.79 4.26 4.79 3.11
VI 0.885 2.869 5.818 4.071
Min 1.90 2.52 1.90 2.31
0.879 3.691 5.632 3.532
Max 18.00 17.71 18.00 14
VII 0.681 8.147 5.526 3.143 Csx 1.64 1.73 1.64 1.24
0.617 9.12 5.346 3.306 Cv 0.801 0.757 0.801 0.541
T. Taskiran / Advances in Engineering Software 41 (2010) 886892 891

25 3

Estimated CBRgep 2.5


20

Sensitivity
2
15
1.5

10 1

0.5
5
R2 = 0.91 0
LL PI k
G
k Wopt Clay+silt Sand Gravel
0
0 5 10 15 20 25 parameters
Actual CBR
Fig. 5. Sensitvity analysis for CBR results.

Measured GEP
20
25

Estiameted CBR values


15
CBR values

20
10
15
5
10
0
0 10 20 30 40 5 R2 = 0.74
No Test
0
Fig. 3. Comparison between estimated (by GEP) and measured values of CBR. 0 5 10 15 20 25
Actual CBR values

25 Fig. 6. Comparison between estimated by NCHRP and measured values of CBR.

20
Estimated CBR

25
15
Estimated CBR values

20
10
15

5
R2 = 0.91 10

0 5 R2 = 0.50
0 5 10 15 20 25
Actual CBR 0
0 5 10 15 20 25
Measured ANN Actual CBR values
20
Fig. 7. Comparison between estimated by Agarwal and Ghanekar [4] and measured
18 values of CBR.
16
14
CBR values

12
from the gures, both equations overestimates CBR values for the
soils used at present study. R2 value is obtained as 0.50 using Eq.
10
(2) and 0.74 by Eq. (4).
8
6
8. Summary and conclusions
4
2
The main objective of this study is to investigate the applicabil-
0
ity of two AI methods which is articial neural network (ANN) and
0 10 20 30 40
gene expression programming (GEP) for prediction of California
No Test
bearing ratio. To achieve this, CBR tests results of ne grained soils,
Fig. 4. Comparison between estimated (by ANN) and measured values of CBR. belonging to A-4, A-5, A-6, A-7 soil groups were provided from 9.
Regional Directorate of Highways Diyarbakir (Southeast Anatolia
Region). Seven ANN and GEP models which have different input
posed by the Eq. (2) (Agarwal and Ghanekar [4]) and Eq. (4) parameters were trained to establish the best interrelationship be-
(NCHRP). The results are presented in Figs. 6 and 7. As can be seen tween basic soil properties and the parameter CBR. Performances
892 T. Taskiran / Advances in Engineering Software 41 (2010) 886892

of the models were examined in terms of some statistical verica- [3] De Graft-Johnson JWS, Bhatia HS. The engineering characteristics of the
lateritic gravels of Ghana. In: Proceedings of 7th international conference on
tion criteria. The best results were produced for both GEP and ANN
soil mechanics and foundation engineering, vol. 2, Mexico, August 2829.
of Model I which have seven input parameters. Bangkok: Asian Institute of Technology; 1969. p. 1343.
Consequently, when the results are evaluated, it can be con- [4] Agarwal KB, Ghanekar KD. Prediction of CBR from plasticity characteristics of
cluded that the both ANN and GEP are found to be able to learn soil. In: Proceeding of 2nd south-east Asian conference on soil engineering,
Singapore, June 1115, 1970. Bangkok: Asian Institute of Technology; 1970. p.
the relation between the CBR and basic soil properties and can 5716.
be used for the prediction of CBR values of soils. Additionally, sen- [5] National Cooperative Highway Research Program. Guide for mechanistic and
sitivity analysis has shown that dry unit weight is the most effec- empirical design for new and rehabilitated pavement structures, nal
document. In: Appendix CC-1: correlation of CBR values with soil ndex
tive parameter on CBR and then PI, wopt, S, C + S, LL and G properties. West University Avenue Champaign, Illinois: Ara, Inc.; 2001.
respectively. Considering the CBR soil tests to be fairly difcult, [6] Doshi SN, Mesdary MS, Guirguis HR. Conference: statistical study of laboratory
time consuming and expensive, it can be emphasized that the CBR for Kuwaiti soils. In: Fourth conference of the road engineering association
of Asia and Australasia, vol. 2, Jakarta; 1983. p. 4351.
use of ANN and GEP models for estimating CBR, in terms of soil [7] Stephens DJ. Prediction of the California bearing ratio. Civ Eng South Africa
parameters, could be helpful tool to be used for preliminary iden- 1990;32(12):5237.
tication of material or/and a base of judgment for the validity of [8] The Highway Agency. Design manual for roads and bridges, vol. 7. London:
Stationery Ltd.; 1994. Section 2, Part 2 HD 25/94.
the CBR values. [9] Basma AA, Kallas N. Modeling soil collapse by articial neural networks.
Geotech Geol Eng 2004;22:42738.
[10] Moosavi M, Yazdanpanah MJ, Doostmohammadi R. Modeling the cyclic
Acknowledgement swelling pressure of mudrock using articial neural Networks. Eng Geol
2006;87:17894.
I would like to thank to M. Ali Sabaz who is the head of labora- [11] Demuth H, Beale M. Neural network toolbox for use with MATLAB. Natick,
Mass: The MathWorks Inc.; 2001. 840 pp.
tory branch of 9th Regional Directorate of Highways Diyarbakir, [12] Ferreira C. Gene expression programming: a new adaptive algorithm for
especially due to the valuable contribution and supports through- solving problems. Complex Syst 2001;13:87129.
out this article work. [13] Muoz DG. Discovering unknown equations that describe large data sets using
genetic programming techniques. Master thesis in Electronic Systems at
Linkping Institute of Technology; 2005. LITH-ISY-EX-05/3697.
References [14] Cevik A. A new formulation for longitudinally stiffened webs subjected to
patch loading. J. Constr Steel Res 2007;63:132840.
[15] Teodorescu L, Sherwood D. High energy physics event selection with gene
[1] Kin Mak Wai. California bearing ratio correlation with soil index properties
expression programming. Comput Phys Commun 2002;178:40919.
master of engineering (civilgeotechnics), Faculty of Civil Engineering,
[16] Sherrod PH. DTREG predictive modeling software; 2008. <http://www.dtreg.
University Teknology Malaysia; 2006.
com/>.
[2] Black WPM. A method of estimating the CBR of cohesive soils from plasticity
[17] Smith GN. Probability and statistics in civil engineering. London: Collins; 1986.
data. Geotechnique 1962;12:2712.

Das könnte Ihnen auch gefallen