Professional Documents
Culture Documents
net/publication/273369765
CITATIONS READS
62 4,228
4 authors:
63 PUBLICATIONS 881 CITATIONS
Payame Noor University
20 PUBLICATIONS 301 CITATIONS
SEE PROFILE
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Effect of Texture of Carbonate Soils in South Iran Coasts on Aggregate Crushing View project
All content following this page was uploaded by Amir A. Amini on 28 August 2015.
Abstract: The pavement condition index (PCI) is a widely used numerical index for the evaluation of the structural integrity and operational
Downloaded from ascelibrary.org by North Carolina State University on 12/10/12. Copyright ASCE. For personal use only; all rights reserved.
condition of pavements. Estimation of the PCI is based on the results of a visual inspection in which the type, severity, and quantity of
distresses are identified. The purpose of this study is to develop an alternative approach for forecasting the PCI using optimization techniques,
including artificial neural networks (ANN) and genetic programming (GP). The proposed soft computing method can reliably estimate the
PCI and can be used in a pavement management system (PMS) using simple and accessible spreadsheet softwares. A database composed of
the PCI results of more than 1,250 km of highways in Iran was used to develop the models. The results showed that the ANN- and GP-based
projected values are in good agreement with the field-measured data. In addition, the ANN-based model was more precise than the GP-based
model. For more straightforward applications, a computer program was developed based on the results obtained. DOI: 10.1061/(ASCE)TE
.1943-5436.0000454. © 2012 American Society of Civil Engineers.
CE Database subject headings: Pavements; Computer programming; Neural networks; Computation; Predictions.
Author keywords: Pavement condition index; Genetic programming; Neural networks; Soft computing.
Fig. 2. Typical deduct value curve (data from Shahin 2005) Background
where a = maximum value of the index; b = slope of deterioration Fig. 3. In this figure, alligator, transverse, and longitudinal crack-
curve; c = exponent coefficient; and t = pavement age. However, ing are predominant distress types. The PCI values for 12,487
using this approach to calculate the PCI may be dangerous when pavement segments were used to develop the model. The frequen-
the experts are actually wrong. cies of the PCI values are shown in Fig. 4. Most of the cases were
Downloaded from ascelibrary.org by North Carolina State University on 12/10/12. Copyright ASCE. For personal use only; all rights reserved.
Genetic Programming
Table 3. Specifications of Artificial Neural Network Model be simplified to a final formula including the desired input and
Network specification Value output variables that predict the output variable in a precise manner.
A typical GP flowchart is shown in Fig. 8 (i = number in the current
Number of layers 3 generation; and M = maximum number of generations).
Number of hidden layers 2
Hidden layer #1 transfer function Log-sigmoid (logsig)
Number of neurons in hidden layer #1 22 Crossover Operator
Hidden layer #2 transfer function Log-sigmoid (logsig)
Number of neurons in hidden layer #2 13 Crossover, which is used to evolve tree-based structures, is one of
Output layer transfer function Log-sigmoid (logsig) the primary GP operators. It can produce two new individuals,
Number of neurons in output layer 1 called offspring, from two other individuals as parents. Parents
Training algorithm Resilient backpropagation are selected from the previously generated population. The pro-
(trainrp) cedure consists of the following steps:
Performance measure MSE
• Selection of two individuals from the population;
• Selection of a random node from each of the selected indivi-
duals; and
prediction of foundation settlements (Rezania and Javadi 2007),
• Exchange of the subtrees of the selected nodes to generate the
modeling of the stress-strain behavior of sand under cyclic loading
new individuals.
(Shahnazari et al. 2010), and the prediction of the soil-water char-
A typical crossover operation is illustrated in Fig. 9.
acteristic curve (Johari et al. 2006) are all examples of GP appli-
cations in civil engineering. As a machine-learning approach, GP
can be used as a mathematical tool for symbolic regression and Mutation Operator
pattern recognition aims. John R. Koza introduced GP at the
Massachusetts Institute of Technology (MIT) in 1992. He used Mutation is a procedure in which a casual node in a selected tree-
the concept of Darwinian natural selection and derived GP from based individual is replaced with another random node from a
the principles of genetic algorithms (GA) (Koza 1992). related terminal or function set. The new node must have the same
In GP, the first step is to create a large random population of number of arguments as the one it replaced. That is, a functional
tree-based structured individuals with a high level of diversity. Each node is replaced by one from the function set, and the same pro-
individual is composed of functions and terminals that are selected cedure can be used for a terminal node with the terminal set
from a predefined set of functions and a set of terminals, respec- (see Fig. 10). A more detailed description of GP operators and the
tively. The functions are basic mathematical, Boolean, logical, and step-by-step explanation of the replacement process can be found
iterative operators. It is also possible to define a newly developed in Koza (1992).
function for special purposes. The terminal set is a set of constant In this study, multigene GP was used to model the PCI. This
values that serve as the arguments of the functions. The individuals method is an improved form of GP that uses a new feature called
created are compared with the desired outputs, and fitness values multigene. In this method, the model development is based on a
are calculated for each individual using a predefined fitness func- number of genes that have nonlinear behaviors but whose combi-
tion. The next stage is the creation of the next generation of the nation in a linear form shapes the final structure of the goal model.
population. The new population is created after choosing the most More details about this method can be found in Hinchliffe et al.
fit individuals and breeding them together using the GP operators, (1996) and Searson (2002, 2009). Eq. (4) shows the general form
crossover, and mutation. Then the new population replaces the of multigene GP:
old one. This procedure is repeated until a model with the desired Y ¼ c1 × G 1 þ c2 × G 2 þ c3 × G 3 þ : : : þ cn × G n þ c0 (4)
precision is obtained. The termination criterion of this iterative
process is the maximum number of generations or a special prede- where ci = coefficient of related genes; Gi = nonlinear genes;
fined fitness value. Finally, the best tree-based model should n = number of genes; c0 = bias term; and Y = output. The initial
PCI ¼ 105.64 − 0.054ðx1 þ 2x12 þ x4 Þ − 0.122ðx5 þ x10 þ x13 þ x20 Þ − 0.266ðx21 þ x6 Þ − 0.061x7 − 0.094x8 − 0.142x9 − 0.388x14
− 0.118x15 − 0.334x18 − 0.0001x23 þ 0.029ðx9 fðx2 ÞÞ þ 27.67fðx1 þ 2x2 Þ þ 1.96ðfðx9 Þ − fðx17 ÞÞ − 19.7fðx1 þ 2.78x2 Þ
− 1.066fðx15 þ x19 þ x2 Þ − 0.79fðx1 þ x5 − x7 − x8 Þ − 7.49fðx1 þ x16 þ x2 þ 1.783Þ − 1.51fðx13 − x1 − x2 þ x20 Þ
− 6.53fðx1 þ x2 þ 3.253x3 Þ − 3.84fðx11 þ 2.751x2 þ 1.864Þ − 0.737fðx2 − x15 þ 1.86x8 Þ þ 0.0166fðx2 þ x3 Þðx7 þ x8 Þ
Downloaded from ascelibrary.org by North Carolina State University on 12/10/12. Copyright ASCE. For personal use only; all rights reserved.
where f = protected logarithmic function. The other parameters illustrate the results of the parametric study for low-, medium-, and
were described previously in Table 1. The protected logarithmic high-severity alligator cracking distress, respectively.
function known as PLOG in MATLAB (2009) calculates the pro- Because all types of distresses have negative effects on the pave-
tected natural logarithm of a variable element by element. Eq. (6) ment condition, increasing the input variables decreases the PCI.
shows the definition of PLOG: In addition, the severity of the distresses has a direct impact on
the negative slopes of the curves. The results also reveal that the
if x ≠ 0 ⇒ PLOGðxÞ ¼ LnðABSðxÞÞ if x ¼ 0 ⇒ PLOGðxÞ ¼ 0 variation trends of the GP- and ANN-based models are the same
(6) and the predicted values are close to one another.
where Ln = natural logarithm; and ABSðxÞ = absolute value func-
tion that returns the absolute value of x. Fig. 11 shows the measured
PCI versus the PCI predicted by the GP-based model for training, Comparison of the Models
testing, and all of the data. As illustrated, the GP-based model The statistical performances of the ANN- and GP-based models are
had a relatively high precision and could predict the pavement summarized in Table 6. The definitions of the statistical criteria
condition index with a low estimation error. In this figure, the high used in this study are given in Table 5. Both of the models have
densities of the points are because of the large number of cases in high precision, and their R2 values show their success in modeling
the database. the PCI. The values of R2 , RMSE, and mean absolute error (MAE)
for the ANN-based models were equal to 0.9986, 0.99, and 0.49,
respectively, whereas they were equal to 0.9898, 2.63, and 1.79
ANN-Based Model for the GP-based model, considering all of the data. However,
the ANN-based evolved model had lower errors compared to the
Fig. 12 shows the performance of the final ANN-based model for
GP-based model. Fig. 16 illustrates the performance of the GP- and
training, testing, and all data. The superior performance of the
ANN-based models in comparison with the field-measured values
model demonstrates the capability of the ANN for modeling the
for some cases specified from the training subsets.
PCI. The statistical properties of the model are also illustrated
A comparison of Figs. 11(a) and 12(a) shows that the GP-based
in Fig. 12.
model had larger relative errors than the ANN-based model, which
is also illustrated by its fluctuation, as shown in Fig. 16. The figure
represents this fact in another way. It was found that cases with PCI
Parametric Study lower than 40 had relatively larger prediction errors than cases with
PCI values greater than 40. However, the errors associated with the
A parametric study of the soft computing models was conducted to
low PCI values fall within the acceptable range. The main reason
determine the effects of the input parameters on the PCI. This study
for this discrepancy is the high concentration of cases with high
was performed to further verify the proposed models. The method-
PCI values in the database (i.e., values greater than 60). In addition,
ology is based on the alteration of a single input parameter within
soft-computing approaches typically optimize the errors in the
the range of its maximum and minimum values while keeping the
ranges that capture most of the cases.
other input parameters at their mean values. Figs. 13, 14, and 15
Shahin and Kohn (1981) stated that pavements with PCI
values in the range of 40 or lower are in an unserviceable condi-
Table 5. Statistical Criteria for Evaluation of Models tion, as shown in Fig. 1. Thus, these pavements should be consid-
ered for rehabilitation/reconstruction activities by the appropriate
Performance index Mathematical definition
Pn agencies.
Coefficient of determination (R2 ) ðym −yp Þ2
R2 ¼ 1 − Pni¼1 2
ðym −y¯m Þ
i¼1
qP
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
ffi
RMSE RMSE ¼ i¼1
ðym −yp Þ2 Verification of the Models
n
Pn
jym −yp j In addition to the good performance of the models for the training
MAE MAE ¼ i¼1
n and testing subsets and the proper results of the parametric study,
Fig. 11. Measured PCI versus the predicted PCI by GP: (a) training; Fig. 12. Measured PCI versus the predicted PCI by ANN: (a) training;
(b) testing; (c) all data (b) testing; (c) all data
Fig. 13. Results of parametric study for the low alligator cracking
Fig. 16. Comparison of the measured values with the predicted values
by GP and ANN for PCI: (a) 80–70; (b) 60–40
Fig. 14. Results of parametric study for the medium alligator cracking Summary and Conclusions
Appendix
Table 8. (Continued)
Predicted Predicted
x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15 x16 x17 x18 x19 x20 x21 Measured (GP) (ANN)
18.4 0 0 0 0 0 0 0 0 0 0 13.18 0 0 0 0 0 0 3.06 0 0 72.84 74.63 72.72
6.27 7.13 0 0 0 0 0 0 0 0 0 3.9 0 0 0 0 0 0 1.68 0 0 67.75 66.19 66.84
0.79 0 0 0 0 0 0 0 0 0 0 1.16 0 0 0 0 0 0 0 0 0 95.71 96.20 95.47
3.17 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5.52 0 0 89.71 88.28 89.54
28.2 12.34 0 0 0 0 0 0 0 0 0 8.77 0 0 0 0 0 0 1.52 0 0 51.85 53.16 52.25
55.13 11.45 0 0 0 0 0 0 0 0 0 2.5 0 0 0 0 0 0 2.81 0.56 0 47.94 48.40 47.46
37.91 42.48 0 0 0 0 0 0 0 0 0 2.34 0 0 0 0 0 0 5.29 0 0 40.29 40.79 39.98
48.23 16.31 0 0 0 0 0 0 0 0 0 2.66 0 0 0 0.3 0 0 2.78 0 0 45.02 47.07 46.26
7.59 72.11 0 0 0 0 0 0 0 0 0 2.62 0 0 0 0 0 0 12.76 0 0 41.29 41.03 41.1
23.62 17.18 0 0 0 0 0 0 0 0 0 7.34 0 0 0 0 0 0 2.23 0 0 50.77 51.69 51.13
25.26 4.65 0 0 0 0 0 0 0 0 0 21.19 0 0 0 0 0 0 0.91 0 0 59.62 59.98 59.77
8.63 0 0 0 0 0 0 0 0 0 0 9.5 0 0 0 0 0 0 0 0 0 80.9 82.35 80.99
5.23 14.35 0 0 0 0 0 0 0 0 0 2.49 0 0 0 0 0 0 0.54 0 0 62.65 60.43 60.88
17.97 8.2 0 0 0 0 0 0 0 0 0 5.08 0 0 0 0 0 0 0.94 0 0 59.19 59.54 59.66
23.56 24.39 0 0 0.82 0 0 0 0 0 0 4.43 0 0 0 0 0 0 0.91 0 0 47.62 48.75 48.11
9.88 19.72 0 0 0 0 0 0 0 0 0 13.87 0 0 0 0 0 0 0 0 0 55.53 54.59 54.72
1.31 11.39 0 0 0 0 0 0 0 0 0 7.07 0 0 0 0 0 0 3.31 0 0 63.46 65.60 64.2
26.02 0 0 0 0 0 0 0 0 0 0 22.31 0 0 0 0 0 0 0.49 0 0 69.08 72.49 68.9
14.56 0.3 0 0 0 0 0 0 0 0 0 11.33 3.34 0 0 0 0 0 3.55 0 0 71.29 75.43 73.36
6.44 26.34 0 0 0 0 0 0 0 0 0 11.54 1.8 0 0.18 0.14 0 0 0 0 0 49.75 52.86 51.91
27.36 12.99 0 0 0 0 0 0 0 0 0 4.18 0 0 0 0 0 0 0.21 0 0 52.7 53.30 52.55
22.89 15.34 0 0 0 0 0 0 0 0 0 23.22 2.46 0 0 0 0 0 1.18 0 0 50.62 51.99 51.23
38.86 7.45 0 0 0 0 0 0 0 0 0 5.42 0 0 0 0 0 0 0.74 0 0 53.7 54.55 53.23
18.63 0 0 0 0.16 0 0.69 0 0 0 0 9.61 0 0 0 0 0 0 2.59 0 0 72.42 75.05 72.51
5.02 5.5 0 0 0 0 0 0 0 0 0 0.54 0 0 0 0 0 0 1.08 0 0 71.34 69.63 70.53
5.32 0 0 0 0 0 0 0 0 0 0 3.14 0 0 0 0 0 0 0 0 0 86.5 86.50 86.02