Application of Genetic Algorithm
in Asphalt Pavement Design
Bor-Wen Tsai, Venkata Narasimham Kannekanti, and John T. Harvey
The main purpose of this study is to demonstrate the applicability of
the genetic algorithm (GA) to solve nonlinear optimization problems
encountered in asphalt pavement design. The fundamentals of the GA
are briefly discussed, and three case studies are presented. The first case
study is an example showing the backcalculation of layer moduli with
deflection data from a falling weight deflectometer and a layered-elastic
program. The second case study demonstrates how to construct the mas-
ter curve, either from a mix flexural frequency sweep test or from a
binder rheometer test, and how to apply the master curve in pavement
design. The last case shows how to apply the GA to characterize the
binder discrete relaxation spectrum with a generalized Maxwell solid
model. The results indicate that the GA is successful in resolving the
nonlinear optimization problem. The GA presents some difficulty in
terms of computing efficiency; however, several techniques have been
developed to alleviate this problem.
The genetic algorithm (GA) has long been used as an optimization
tool in resolving numerical problems, especially the problem of non-
linear optimization. However, little has been done to apply GAs to
asphalt pavement design until recently. The backcalculation of pave-
ment layer moduli, design of pavement structures, and scheduling of
pavement maintenance operations are the major applications of GAs.
This paper presents the use of GAs for backcalculation of pavement
moduli focused in the sensitivity analysis of GAs, the characterization
of mix master curve, and the binder relaxation spectrum.
In the area of backcalculation of layer moduli, Kameyama et al.
developed a method to backcalculate pavement layer moduli from
surface deflections with the GA (1), with surface deflections calcu-
lated by layered-elastic theory as the input condition. Marshall and
Meier combined the GA with artificial neural networks to search for
complex pavement models that reproduce the response of a real pave-
ment system to a suite of falling weight deflectometer (FWD) tests
(2). The computational efficiency of neural networks and GAs helped
to search for more complex solution species. The GA has replaced the
traditional calculus methods for searching for best-fit stiffness profiles
of in situ pavement systems based on nondestructive test methods (3).
GAs have been used by Liu and Wang (4) for design of flexible
pavement structure and by Hadi and Arfiadi (5) for design of rigid
structures.
B.-W. Tsai, University of California, Berkeley, Institute of Transportation Studies,
Pavement Research Center, 1353 South 46th Street, Building 480, Richmond,
CA 94804. V. N. Kannekanti and J. T. Harvey, Department of Civil and Environ-
mental Engineering, University of California, Davis, Engineering III, Room 3139,
One Shields Avenue, Davis, CA 95616.
Transportation Research Record: Journal of the Transportation Research Board,
No. 1891, TRB, National Research Council, Washington, D.C., 2004, pp. 112–120.
112
Shekharan used GAs to develop solutions to pavement deteriora-
tion models (6), and Attoh-Okine presented the application of the GA
in predicting roughness progression in flexible pavements (7). GAs
have been merged with artificial neural networks and fuzzy logic to
efficiently identify the distresses to be treated (8). The advantages of
such approaches in the design of decision support systems have been
described by Loia et al. (9).
The GA is a technique inspired by the Darwinian theory of sur-
vival of the fittest (10). It mimics the natural process of evolution to
develop an optimum solution. GAs operate on a set of randomly
generated solutions. For each of the solutions, the values of the fit-
ness function, which indicate the proximity of the solution to the
optimum solution, are evaluated. The solutions with good fitness
values are combined in an attempt to produce a better solution set.
Replacing solutions with poor fitness with new solutions completes
one iteration, or generation, of the algorithm. This process is repeated
until a sufficiently good fit is obtained.
The general procedure to conduct a GA analysis follows.
1. Define the problem. The parameter definition and fitness func-
tion associated with the problem should be clearly identified before
conducting a GA-based analysis. The rest of the GA procedure is to
find an optimum set of parameters—that is, a good gene—that min-
imize the fitness function. Intuitively, the residual sum of squares
(RSS) is a good choice for the fitness function, given that the objec-
tive is to have the measurements and predictions as close as possible
to model fitting.
2. Generate N (even number) genetic starting strings. From the
problem definition, a total of p parameters (t1, t2, t3, . . . , tp) are
selected to construct a gene string (or gene). A gene Λi consists of val-
ues {S (i) (i) (i) (i)
t1, S t2, S t3, . . . , S tp} of p parameters, which are generated by
using a uniform distribution over a specified range of each parameter.
A gene pool is defined as a set of genes, that is, {1, 2, . . . , N}:
Gene 1 1 = {St11 , St 21 , St 31 , . . . , Stp1 }
( ) ( ) ( ) ( )
Gene 2   2 = {St12 , St 22 , St 32 , . . . , Stp2 }
( ) ( ) ( ) ( )
Gene 3  N = {St1N , St 2N , St 3N , . . . , StpN }
( ) ( ) ( ) ( )
3. Compute the fitness of each gene: Π(1), Π(2), . . . , Π(N).
Fitness is the term used to measure the goodness of fit of the problem-
specified fitness function Π(i). Most of the time, it is appropriate to
define the fitness function as the RSS of the response variables.
4. Rank the genes according to fitness: *1, *2, . . . , *N.
5. Mate nearest ranked pairs (produce offspring). The offspring i
is generated as a linear combination of its parents’ genes as indicated
as follows. The reason to have an even number of genes (Step 2)
is clear.
Tsai, Kannekanti, and Harvey 113

1 = Φ 1 *1 + [1 − Φ 1 ]*2
( ) ( ) mented at various distances. According to the GA procedure outlined
previously, the following terms are defined:
 2 = Φ 2 *1 + [1 − Φ 2 ]*2
( ) ( )
• Parameters. Three-layer moduli (modulus of asphalt concrete
Eac, modulus of aggregate base Eab, and modulus of subgrade Esg).
 N −1 = Φ N −1 *N −1 + [1 − Φ N −1 ]*N
( ) ( )

• Fitness function.
 N = Φ N *N −1 + [1 − Φ N ]*N
( ) ( )

∑ (D − Dˆ i ) = min
2
RSS = i
i
where Φ(i) = Aφ (i) and 0 ≤ φ (i) ≤ 1 (a random number from uniform where Di is the FWD deflection at the ith position and D̂i is the cor-
distribution). Here, A is set at 2 . If A ≤ 1, then there is no mutation. responding deflection calculated from the layered-elastic program
6. Check design range (or constraints): S t1− ≤ S (i) + −
t1 ≤ S t1; . . . ; S tp ≤ (ELSYM5).
+ − +
S (i)
tp ≤ S tp, where S ti is the left constraint and S ti is the right constraint.
If the variable is out of range, set equal to (or preferably a little less Application of the GA procedures for this problem includes the
than) constraint. It was found that the precision of the solution was following:
affected by the range specified for each parameter. A broad param-
eter range is suggested for a novice with limited knowledge of the 1. Specify ranges of parameters.
parameter. An experienced user is able to narrow the parameter 2. Generate genes.
range to increase the probability of finding the optimum solution. 3. Calculate pavement response and evaluate the fitness function.
If an inappropriate range is specified, it could lead to a local (or 4. Rank and mate genes.
restrained) optimum. 5. Discard bad genes and replace with new genes.
7. Discard the bottom M (<N ) genes. 6. Repeat Steps 1 through 5 until the specified number of
8. Repeat with top N-M gene pool plus M new genes. generations is completed.

Several issues can be addressed regarding the GA:
1. The size of the gene pool (or gene size), the number of gener-
ations, and the number of discarded genes for each generation run are
going to affect the computational time and precision of the solution.
2. The GA could be useful in finding the initial guess for a non-
linear problem. To obtain a more accurate solution, it is better to run
the GA first and then refine the solution by using the Newton algo-
rithm, which is a gradient-based iteration method for solving a root-
finding problem (11). However, conducting the GA alone with more
generations or with a larger gene pool will also give a satisfactory
solution.
3. The GA is a simple and powerful algorithm for solving the
nonlinear optimization problem with constraints regardless of its
disadvantages: (a) computational inefficiency with respect to time
(not always the case) and (b) input of appropriate parameter range.
The generalized GA-based procedure is worth keeping in mind.
In the following sections, three case studies of GA application are
demonstrated. These cases begin with the problem definition and
follow with numeric examples to explain how the GA works.
CASE 1. BACKCALCULATION OF LAYER
MODULI USING FWD DATA
The FWD has been widely used to characterize pavement response.
A variety of approaches have been used to backcalculate elastic mod-
uli that minimize the difference between the measured peak surface
deflections from the FWD and deflections calculated by using a
static model of the FWD on a pavement characterized by layer elastic
theory—for example, MODULUS [documented by J. Uzan in Appen-
dix E: MODULUS User’s Guide of NCHRP Report 327 (12)]—or
the Odemark-Boussinesq approach—for example, ELMOD (13, 14).
An alternative is to use the GA approach to find an optimum set of
layer moduli that minimizes the RSS of the difference between mea-
sured and calculated peak surface deflections. As indicated in Fig-
ure 1a, an FWD was used on a three-layer pavement structure to
characterize the layer moduli. The surface deflection was instru-
The ranges of parameters (Eac, Eab, and Esg) were determined
according to the user’s experience and judgment. For each gene in a
gene pool, ELSYM5 was used to calculate the pavement response—
in this case, the surface deflection D̂i—and then the fitness function
was evaluated. According to the evaluated fitness, the parent genes
were ranked and then mated to generate better offspring. The mated
genes with larger fitness were discarded and replaced with the new
genes generated randomly by using the uniform distribution con-
strained by their specified ranges as in Step 2. The whole process was
iterated until the specified number of generations was completed.
This example FWD deflection basin was measured at 15.6°C
pavement surface temperature. Figure 1a schematically presents the
FWD configuration, problem definition, and pavement structure.
The properties and parameter ranges of this three-layer pavement
system are as follows:
Thickness Poisson’s Parameter range
Layer (mm) ratio (MPa)
Asphalt concrete 100 0.35 2,068 to 3,447
Aggregate base 400 0.40 345 to 1,379
Subgrade 0.45 69 to 345
Figure 1b presents the backcalculation result when a heavy FWD
load (54.6 kN) is applied to the pavement. As indicated in Figure 1b,
the calculated response is quite satisfactory compared with the mea-
sured deflections. In this case, the nonlinearity of the stress-dependent
subgrade was not considered. However, the nonlinearity can be re-
solved with a finite element program as the stress–strain engine and
following the same GA procedure as presented here.
An analysis of variance (ANOVA) was conducted to evaluate the
effects of gene size, number of generations, and FWD loading lev-
els on the parameter variation. Each factor has three levels. The
experimental design was as follows:
• Load factor H [(54.6 kN), M (39.1 kN), and L (22.3 kN)];
• Genes factor (500, 1,000, and 2,000); and
• Generations factor (50, 100, and 200).
For each combination of factor levels, 10 replicates were con-
ducted for a total of 270 GA runs. The number of discarded genes
114 Transportation Research Record 1891

Eac AC Hac
Di
Eab Aggregate Base Hab

Esg Subgrade

(a)

0
FWD Measurement
Back-Calculated Deflection
Surface Deflection (mm)

-0.1

FWD Loading = 54.55 KN

-0.2 Estimated Parameters:

Eac = 2,226.08 MPa
Eab = 534.63 MPa
Esg = 204.46 MPa
-0.3
2
RSS = 1.86131E-5 mm

-0.4
0 200 400 600 800 1,000 1,200 1,400 1,600
Distance (mm)
(b)

FIGURE 1 (a) Schematic of FWD and pavement structure (AC  asphalt

concrete) and (b) surface deflection fitting results after 50 generations with
500 genes.

was half the total number of genes. Figure 2 presents the design plots
of the mean of layer moduli and the coefficient of variation (CV) of
layer moduli.
Several findings of ANOVA can be addressed as follows:
1. The mean of Eac is sensitive to the gene size and number of
generations. On the contrary, the means of Eab and Esg are relatively
insensitive. The mean of Eac is positively proportional to load level.
However, the means of Eab and Esg are negatively proportional to
load level.
2. In general, all three parameters (Eac, Eab, and Esg) share a com-
mon tendency: the bigger the genes and the higher the number of
generations, the lower the CV of the moduli (i.e., the smaller the
modulus variation in all three parameters).
3. For the asphalt concrete layer, the lower the loading level, the
lower the CV of moduli. Conversely, for the aggregate base and sub-
grade layers, the low loading level causes a higher CV than the high
loading level.
4. On average, the magnitude of the CV of the asphalt concrete
layer is about two times that of the aggregate base layer and about
four times that of the subgrade layer.
5. One can certainly conclude from the ANOVA that the bigger
the size of genes and generations, the smaller the modulus variation.
However, a trade-off exists between the computational time and the
modulus variation. The computational time is a considerable disad-
vantage for the backcalculation of layer moduli because the layered-
elastic program needs to be called up whenever the fitness of a gene
is calculated.
With regard to computation time, the GA approach is certainly
not a good choice compared with the ELMOD and MODULUS.
However, the accurateness of the moduli of various approaches needs
to be verified through the in situ measured stress–strain data.
CASE 2. MASTER CURVE AND ITS APPLICATION
The complex modulus master curve obtained from flexural frequency
sweep tests, which are considered nondestructive tests, is a useful
tool to characterize the effects of loading frequency and temperature
on the initial stiffness of an asphalt mix. A function that describes
flexural stiffness as a function of temperature and time of loading
can be used for pavement design.
Tsai, Kannekanti, and Harvey 115

H 500 500 50
H

0.024
Mean of CvEac (MPa)
2,180.0
Mean of Eac (MPa)
50

0.022
1,000
100 100
M
2,160.0

0.020
1,000
L

0.018
200
2,000 200
2,140.0

L 2,000
Load Genes Generations Load Genes Generations
Factors Factors
(a) (b)

L 50

0.012
500
620

Mean of CvEab (MPa)

Mean of Eab (MPa)

1,000
L
600

0.011
H
100
M
580

2,000
1,000 200
100

0.010
500 50
560

M 200
H 0.009 2,000
Load Genes Generations Load Genes Generations
Factors Factors
(c) (d)
214

L 500
L
0.0050

50
100
212

Mean of CvEsg (MPa)

Mean of Esg (MPa)

M
210

1,000
0.0045
202

500 50
100
1,000
2,000 200
206

M
0.0040
204

2,000
H
202

H 200
Load Genes Generations Load Genes Generations
Factors Factors
(e) (f)

FIGURE 2 Design plots of means and CVs of three parameters (E ac , E ab , and E sg ) showing effects of load level, gene
size, and number of generations.

Flexural frequency sweep tests are conducted mostly from 15 to
0.01 Hz at three or four temperature levels. Under the assumption
that asphalt mix is a time–temperature rheologically simple mate-
rial, the curves can be shifted horizontally (or vertically if necessary)
relative to one of the test temperatures to obtain the full spectrum of
complex moduli. The ratio of relaxation times at different tempera-
tures is termed aT. The temperature shift factor ln aT is expressed in
the form ln aT = −C1(T − Tref)/(C2 + T − Tref), where C1 and C2 are
constants (15). The questions then are what should be the horizon-
tal (or vertical if necessary) distances and how should the goodness
of fit be measured?
The most intuitive measure to set up the fitness function is RSS,
while comparing the observed dependent variable data ( yi) with
the corresponding predicted values ( ŷi). With limited knowledge
about the shifted master curve—no exact mathematical function is
known—the generalized additive model with spline fitting becomes
the best candidate to calculate the predicted ŷi for fitness function (16).
In this case, the parameters and fitness function are defined as follows:
• Parameters—Horizontal shifts S1, S2, and S3 (Figure 3a).
• Fitness function—
RSS = ∑ (y − yˆi ) = min
2
i
where yi is the measured complex modulus and ŷi is the fitted complex
modulus.
116 Transportation Research Record 1891

10 RSS of 0.177 produced a quite satisfactory shifting result, as indi-

cated in Figure 4a. No function was specified for the complex mod-
9 ulus master curve. Instead, a generalized additive model with spline
5°C S1 fitting was used to calculate RSS. However, for the purpose of fur-
ther analysis, gamma fitting was suggested for the master curve as
LnE* (E*: MPa)

8
well as the time–temperature shifting relationship.
S3 The main purpose in conducting the gamma nonlinear fitting of the
7 S2
20°C data of frequency sweep tests is to find a suitable mathematical func-
6 tion that can represent the relationship of the complex modulus and
30°C reduced loading frequency at a reference temperature. The gamma
5
distribution function with shape parameter n (a positive number) and
40°C scale parameter β is expressed in the following form (18):
 − x n −1 x m
1 − exp   ∑ m
4
 x ≥ 0
 β  m = 0 β m!
-6 -4 -2 0 2 4
y = F( x ) = 
Ln(freq) (freq: Hz)

(a) 0 x < 0

10
where F (x) is the gamma distribution function and m is an index
S1 = 3.5
number.
9 S2 = -2.5 The characteristics of distribution function are as follows:
S3 = -4.0
• The values of y range from 0 to 1,
LnE* (E*: MPa)

8
RSS = 1.825 • The x values start from 0, and
• The y value is asymptotic to 1 as x increases.
7

6 10

S1 = 5.2052
5 9 S2 = -2.8206
S3 = -5.2051
LnE* (E*: MPa)

4 8
-10 -8 -6 -4 -2 0 2 4 6 8 RSS = 0.177
Tref = 20°C
Ln(freq) (freq: Hz) 7
(b)
6
FIGURE 3 (a) Parameter definition and flexural frequency sweep Gamma Fitting
test results at 5ºC, 20ºC, 30ºC, and 40ºC and (b) intermediate
fitting result (freq  frequency). 5
lnE* = ln E0* + A*{1 - exp(-lnωx/β)*[1 + lnωx/β + lnωx2/(2*β2)]}
ln E0* = 5.136, A = 4.550, β = 2.946, lnωx = lnfreq + 9.810
4
As an example, Figure 3a presents the flexural frequency sweep -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
test results for a typical California mix with asphalt from a valley Ln(freq) (freq: Hz)
source. The flexural frequency sweep test results presented here are (a)
a summary from two flexural fatigue test setups, namely, conven-
tional (without support) for temperatures of 5°C, 20°C, and 30°C 6
and modified (with support) for temperatures of 20°C, 30°C, and
40°C (17 ). “With support” means an aluminum support is placed 4 ln aT = -18.286  {1 - exp[-0.01655  (T - 20)]}
beneath the specimen to prevent the creep effect from occurring
because of specimen self-weight while it is subjected to high tem-
2
peratures. The loading frequencies used are 15, 10, 5, 2, 1, 0.5, 0.2,
Ln(aT)

0.1, 0.05, 0.02, and 0.01 Hz. The upper limit of 15 Hz is a constraint
0
imposed by the capacity of the test machine. The general principles
for conducting frequency sweep tests are “quick to slow in loading
-2
frequency and hot to cold in temperature.” The temperatures were
set at 40°C, 30°C, 20°C, and 5°C.
-4
Taking the E* master curve as an example, the parameters to be fit
are the horizontal shifts for different temperatures—that is, Λ = {S1,
-6
S2, S3}. In this case, three parameters were defined (as indicated in Fig- -20 -15 -10 -5 0 5 10 15 20 25
ure 3a), S1 shifting to the right and S2 and S3 shifting to the left accord- T - Tref(20°C)
ing to the reference temperature 20°C. The shifting is signed: positive
(b)
to the right and negative to the left. Figure 3b presents an intermedi-
ate fitting result with a set of parameters {3.5, −2.5, −4.0}, which FIGURE 4 (a) Fitting results after 50 generations with 100 genes
results in a RSS of 1.825. After 50 generations (100 genes at each gen- and gamma fitting afterward and (b) temperature-shifting
eration), an optimum parameter set {5.2052, −2.8206, −5.2051} with relationship with a gamma-type fitting (freq  frequency).
Tsai, Kannekanti, and Harvey 117

These aspects of a distribution function are used in constructing the 6

nonlinear fitting.
The procedure for nonlinear fitting of the master curve is as follows: 4 ln aT = -18.286  {1 - exp[-0.01655  (T - 20)]}

1. Take the natural logarithm for both frequency and complex 2

modulus, that is, ln(ω) and ln(E*). ln aT @ T°C

Ln(aT)
2. Shift the master curve in both x and y axes based on the low- 0
est frequency and the complex modulus at the lowest frequency so
that the frequency and complex modulus will start the value from 0, -2 ln aT = -1.452
that is, ln(ω)* = ln(ω) − ln(ω0) and ln(E*)* = ln(E*) − ln(E*0). @ 25°C
3. Conduct nonlinear fitting on the following formula: -4
T - Tref°C
  − ln(ω )*  n −1 [ln(ω )* ] 
m

ln( E* )* + Aˆ  1 − exp  ∑

-6

β  m = 0 β m! 
-20 -15 -10 -5 0 5 10 15 20 25
 
m

T - Tref(20°C)
where A is the amplitude. (a)

4. Once satisfactory fitting results are obtained, shift back to the 10

original coordinate system—that is, the fitting pairs of points are of E* (ToC; ω Hz)
the following form: 9
E* (20oC; 10 Hz)

   − ln(ω )*  nˆ −1 [ln(ω )* ] 

m
ˆ 
 ln(ω ), ln( E0 ) + A  1 − exp 
E* (25oC; 10 Hz)

LnE* (E*: MPa)

8
  ∑ ˆm
*

   βˆ  m = 0 β m! 
7

Figure 4 demonstrates the appropriate nonlinear fit by using Tref = 20°C

6 ln 10 - 1.452
gamma fitting for the master curve and the time–temperature shift-
ing relationship. The advantage of using the gamma function is to ln ω + ln aT
provide a specified continuous mathematic function that is easy to 5
10 Hz ω Hz
incorporate into programming and calculation. In a discrete sense,
using the spline function when estimating the GA parameters can do 4
-12 -10 -8 -6 -4 -2 0 2 4 6 8 10
the same job.
The major purpose of finding the master curve and the time– Ln(freq) (freq: Hz)
temperature shift factor is to correct the loading frequency and (b)
temperature effects when conducting a pavement performance
FIGURE 5 (a) Temperature shift factors at TºC and 25ºC and
prediction. (b) initial stiffness correction at TºC and 25ºC at loading frequency
As an example, Figure 5 indicates the way to correct the initial of 10 Hz.
stiffness with various loading frequencies and temperatures. The gen-
eral steps to calculate interpolatively the initial stiffness at loading
frequency ω and temperature T are as follows: CASE 3. DISCRETE RELAXATION SPECTRUM
OF ASPHALT BINDER
1. Find the temperature difference with respect to the reference
temperature. Characterization or modeling of asphalt binder properties often
2. Map the temperature difference through the time–temperature requires the relaxation time spectrum, G(t), instead of using the
relationship function to obtain the shift factor (Figure 5a). dynamic moduli, G′(ω) and G″(ω), which are easy to measure from
3. Locate the loading frequency ω in Figure 5b, add up the signed the dynamic shear rheometer test in the frequency domain. In addi-
shift factor, and then map through the master curve function to tion, the relaxation model in the time domain is easier to interpret
obtain the corrected stiffness. than in the frequency domain. The GA provides a relatively easy and
powerful way to solve the parameters of a discrete relaxation spec-
For example, suppose the temperature at the bottom of the asphalt trum gi, λi of a generalized Maxwell solid model by fitting dynamic
concrete layer is 25°C, which is 5°C greater than the reference tem- mechanical data. Instead of reinventing the wheel theoretically in
perature of 20°C; hence, from Figure 5a, the time–temperature rela- the field of asphalt rheology, the main point of this GA application
tionship, one has ln aT = −1.452. By adding this value with the in situ is to demonstrate the relatively intuitive way to find the parameters
loading frequency, for example, 10 Hz in this case, one then has of the relaxation spectrum.
the corrected initial stiffness at 25°C and 10 Hz as indicated in Fig- The relaxation modulus, G(t), of a generalized Maxwell solid
ure 5b. For a specified loading frequency, the correction moves to model can be written as a discrete set of exponential decays as
the left at higher temperature (relative to the reference temperature) follows (19):
and moves to the right at lower temperature.
N
With the GA, it is relatively easy to incorporate the parameters
G (t ) = Ge + ∑ gi exp( − t λ i )
for vertical shifts and conduct the same gamma nonlinear fitting pro- i =1
cedure to the shifted master curve. For this GA application, the com-
putational time is not an issue because no serious calculation is The formulation consists of the equilibrium modulus Ge, which is
involved as in the modulus backcalculation case. finite and greater than 0 for solids, and N relaxation modes defined
118 Transportation Research Record 1891

by their relaxation strengths gi and their relaxation times λi. The 1E+09
Storage Modulus G'
dynamic moduli, storage modulus G′, and loss modulus G″ can be 1E+08
Loss Modulus G"
expressed as follows: 1E+07 Shear Complex Modulus G*

Dynamic Modulus (Pa)

1E+06
N
(ωλ i )2
G ′(ω ) = Ge + ∑g
i =1
i
1 + (ωλ i )2
(1) 1E+05

1E+04

1E+03
N
ωλ i
G ′′(ω ) = ∑ gi
i =1 1 + (ωλ i )2
(2) 1E+02

1E+01
Reference Temperature = 25°C
1E+00
Hence, the problem definition for GA application is to find the
parameters of a discrete relaxation spectrum by simply fitting Equa- 1E-01
1E-05 1E-03 1E-01 1E+01 1E+03 1E+05 1E+07
tions 1 and 2 to G′ and G″ data. Because the ranges of parameters
Frequency (Hz)
might be over several decades, it is suggested that the ratio residual
G′(ωj)/Ĝ j′ − 1 be used instead of the normal residual G′(ωj) − Ĝ j′. As (a)
before, it is necessary to minimize RSS to obtain the optimum model.
1E+09
It should be remembered that the G′ and G″ curves need to be satis-
fied simultaneously by the same set of parameters; so, RSS is the sum 1E+08

of squares of ratio residuals of G′ and G″. For GA application, the

Dynamic Modulus (Pa)

1E+07

following are defined: 1E+06

1E+05
• Parameters—The model contains 2N + 1 parameters including 1E+04
N pairs of (gi, λi) and Ge. 1E+03
• Fitness function—
1E+02 Storage Modulus G'
1E+01 Loss Modulus G"
  G ′ (ω j ) G ′′(ω j ) 
2 2

− 1 +  − 1  = min

m
RSS = ∑  ˆ
j =1  
ˆ
1E+00 Fitted Lines (10 Replicates)
G j′   G j′′   1E-01
1E-05 1E-03 1E-01 1E+01 1E+03 1E+05 1E+07

where Ĝ j′ and Ĝ ″j are the measured data at m frequencies. Frequency (Hz)

In this case, a total of 10 GA runs were conducted with a gene size
of 1,000, 5,000 generations, and 600 discarded genes. A 12-unit gen-
eralized Maxwell solid model was adopted in this demonstration case
(i.e., a total of 25 parameters). The asphalt binder studied here was a
PG64-22. The convergence trends of RSS of these 10 GA runs pre-
sent a pattern similar to that shown in Figure 6. The RSS stabilizes
around 2,000 generations. It indicates that the result cannot be better
off if more than one run is simulated. This statement is also verified
by the fitting results of G′ and G″ data presented in Figure 7b. Be-
cause the fitting results of 10 replicates are so close, the fitted lines
overlap each other in Figure 7b.
100
80
60
RSS
(b)
FIGURE 7 (a) Plots of master curves of shear complex modulus
G*, storage modulus G, and loss modulus G with PG64-22
asphalt binder and (b) G and G fitting results using a 12-unit
generalized Maxwell solid model.
Figure 7a presents the shifted master curves (also using the GA) for
complex modulus, storage modulus, and loss modulus. Baumgaertel
and Winter (19) suggested that
• The initial number of relaxation models be chosen empirically
between 1 and 2 decades and
• The negative gi values be deleted, which is not necessary in GA
analysis because one can specify the gi range.
At the beginning, nothing was known about the magnitudes of gi
and λi; thus, a broad range (10−9 to 109) was specified for each gi,
and the range of λi was specified over 3 decades with 2 decades of

overlap. The total range of λi was 10−7 to 107. The equilibrium mod-
40 Gene Size = 1,000 ulus Ge was set in the range 0 to 108. The detail parameter range is
Generations = 5,000 presented in Table 1. Table 1 also lists the statistical summary of the
Discarded Genes = 600 10 GA runs.
20
Replicates = 10 As indicated in Figure 7b, other than the phenomenon that the loss
modulus G ″ has a wavy fitting at the high frequencies, the fitting
0 results for G′ and G″ data are quite satisfactory and consistent for
1 10 100 1,000 10,000
all 10 replicates. This waviness phenomenon also was reported by
Generations
Stastna et al. in their inspection of the Maxwell model (20). The
FIGURE 6 RSS convergence trends of 10 GA replicates against average RSS is about 0.898, the standard deviation is about 0.0504,
generations, with a 12-unit generalized Maxwell solid model. and the CV is 0.056. Figure 8a plots gi versus λi to construct a dis-
Tsai, Kannekanti, and Harvey 119

TABLE 1 Statistical Summary and Parameter Input Ranges of GA Application in Discrete Relaxation Spectrum
g i (Pa) λ i (sec)
i Parameter Range Parameter Range
µ σ CV µ σ CV
Begin End Begin End
1 1.7681E+08 1.1113E+06 0.006 1E-09 1E+09 8.2257E-07 1.1448E-08 0.014 1E-07 1E-04
2 7.7303E+07 1.0711E+05 0.001 1E-09 1E+09 3.1786E-05 1.3712E-06 0.043 1E-06 1E-03
3 2.7347E+07 9.5417E+05 0.035 1E-09 1E+09 9.6678E-04 6.4060E-05 0.066 1E-05 1E-02
4 5.5857E+06 3.6495E+05 0.065 1E-09 1E+09 2.0570E-02 3.3971E-03 0.165 1E-04 1E-01
5 8.6677E+05 1.5928E+05 0.184 1E-09 1E+09 2.6087E-01 4.7076E-02 0.180 1E-03 1E+00
6 1.2137E+05 4.1073E+04 0.338 1E-09 1E+09 3.1811E+00 2.4331E+00 0.765 1E-02 1E+01
7 3.0292E+04 3.8611E+04 1.275 1E-09 1E+09 1.5991E+01 6.1112E+00 0.382 1E-01 1E+02
8 2.3838E+03 7.7100E+02 0.323 1E-09 1E+09 1.0185E+02 1.8301E+01 0.180 1E+00 1E+03
9 1.9410E+02 3.8915E+01 0.200 1E-09 1E+09 6.5465E+02 6.6222E+01 0.101 1E+01 1E+04
10 8.3902E+00 9.2538E-01 0.110 1E-09 1E+09 5.6255E+03 3.3078E+02 0.059 1E+02 1E+05
11 1.1493E-01 4.7207E-02 0.411 1E-09 1E+09 4.0724E+05 1.6007E+05 0.393 1E+03 1E+06
12 6.9648E-02 4.0987E-02 0.588 1E-09 1E+09 5.4488E+06 1.1338E+06 0.208 1E+04 1E+07
Ge 1.0293E-01 4.8684E-02 0.473 0E+00 1E+08
RSS 8.9788E-01 5.0414E-02 0.056
σ
NOTE: µ: mean; σ: standard deviation; CV = : coefficient of variation.
µ

1E+09 crete relaxation spectrum. Note that noisy discrepancies occur at

1E+08 high relaxation times. The same disturbance is also observed in the
1E+07 plot of relaxation modulus G(t) presented in Figure 8b; in addition,
1E+06
an obvious discontinuity happens around 4,000 s.
1E+05
The fitting results indicate that using the GA approach in solving
the discrete relaxation spectrum is quite promising. The GA ap-
gi (Pa)

1E+04
proach not only can avoid the potential ill-condition problems while
1E+03
solving a system of equations but also can encourage novices in their
1E+02
Discrete Relaxation Spectrum attempts to solve the question easily and intuitively without any
1E+01 advance knowledge. The fitting results also suggest that, practically,
(10 Replicates)
1E+00 no benefit can be achieved if more than one run is simulated. In addi-
1E-01 tion, computational time is not an issue for a GA approach for this
1E-02 type of problem.
1E-07 1E-05 1E-03 1E-01 1E+01 1E+03 1E+05 1E+07
λi (sec)
(a) FINDINGS AND DISCUSSION

1E+09 From the preceding case studies, satisfactory results can be achieved
1E+08 by using the GA in solving the nonlinear optimization problems
Relaxation Modulus G(t) (Pa)

1E+07 encountered in pavement design without finalizing the solutions by

1E+06
the Newton method.
1E+05
Several findings are addressed as follows:
1E+04
1. To conduct a GA-based application, the parameters and fitness
1E+03
functions should be clearly identified and defined.
1E+02
2. Computational time is the primary problem of GAs. It is espe-
1E+01
Relaxation Modulus G(t)
Average Ge = 0.103
cially obvious when serious computation is involved, as in the case
1E+00
(10 Replicates) of backcalculation of layer moduli. This is because the number of
1E-01 times the layered-elastic program or finite element program is called
1E-02 up grows quickly as the gene size and number of generations in-
1E-07 1E-05 1E-03 1E-01 1E+01 1E+03 1E+05 1E+07 crease. If, as in Case 2 or 3, a one-time mathematical calculation is
Time (sec) all that is required, the computational time is completely tolerable.
(b) It was found that the computational time of GAs depends on how
the number of genes, the number of parameters, the number of gen-
FIGURE 8 (a) Discrete relaxation spectrum plot of relaxation
erations, and the specified range of each parameter are defined. There-
strengths g i versus relaxation times  i and (b) calculated
relaxation modulus G(t) (12-unit generalized Maxwell solid model fore, a trade-off usually exists between gene size and the number of
was used). generations.
3. One of the advantages of using the GA is that a user with exper-
tise can specify the appropriate parameter range while a novice can
120
input the parameter range with a conservative guess for the range.
Either situation can lead to the same results. However, an incorrect
specified parameter range could lead to an incorrect local optimum
solution.
4. As a general rule, the bigger the gene size and number of
generations, the smaller the variation in the predicted parameters,
computational efficiency notwithstanding.
5. With an appropriate gene size and number of generations, no
extra benefit is added if more GA replicates are conducted. To
choose the appropriate gene size and number of generations to pro-
vide satisfactory results, it is necessary for users to go through a
trial-and-error procedure.
Generally speaking, the GA is an intuitive and powerful tool to
solve the nonlinear optimization problems, although in some cases,
for example in Case 1, the computational time is barely tolerable.
To apply the GA successfully, extreme caution must be used with
regard to two issues. First, the inappropriate or incorrect parameter
input range could lead to an incorrect local optimum solution, and
thus a rather wide parameter range is required if no knowledge is
available to develop a reasonable initial guess for the parameters.
Second, the parameters and fitness function need to be defined clearly
at the beginning. In addition, judgment is needed to determine input
and to evaluate output.
Regardless of the computation time of certain GA applications,
several tasks, especially in model fitting, could be worthy of future
focus:
• Apply the GA approach developed here to extend the database
of mix master curve and binder relaxation spectrum of various mixes
and binders.
• Apply the GA approach in searching the nonlinear model of the
fatigue damage process obtained from the flexural fatigue bending test.
ACKNOWLEDGMENTS
This research was sponsored by the California Department of Trans-
portation as part of the work of the Partnered Pavement Research
Center. The authors thank Caltrans for support.
REFERENCES
1. Kameyama, S., K. Himeno, A. Kasahara, and T. Maruyama. Backcal-
culation of Pavement Layer Moduli Using Genetic Algorithms. Proc.,
8th International Conference on Asphalt Pavements, Seattle, Wash.,
Aug. 1997, pp. 1,375–1,385.
2. Marshall, C. P., and W. R. Meier. An Evolutionary Approach to Back-
Calculation Using Genetic Algorithms with Neural Networks. Inter-
Transportation Research Record 1891
national Conference on Highway Pavement Data, Analysis and Mech-
anistic Design Application, Columbus, Ohio, Sept. 2003.
3. Hunaidi, O. Evolution-Based Genetic Algorithms for Analysis of
Nondestructive Surface Wave Tests on Pavements. Nondestructive
Testing in Civil Engineering International, Vol. 31, No. 4, 1998,
pp. 273–280.
4. Liu, M.-Y., and S.-Y. Wang. Genetic Optimization Method of Asphalt
Pavement Based on Rutting and Cracking Control. Journal of Wuhan
University of Technology, Materials Science Edition, Vol. 18, No. 1,
2003, pp. 72–75.
5. Hadi, M. N. S., and Y. Arfiadi. Optimum Rigid Pavement Design by
Genetic Algorithms. Computers & Structures, Vol. 79, No. 17, 2001,
pp. 1,617–1,624.
6. Shekharan, A. R. Solution of Pavement Deterioration Equations by
Genetic Algorithms. In Transportation Research Record: Journal of the
Transportation Research Board, No. 1699, TRB, National Research
Council, Washington, D.C., 2000, pp. 101–106.
7. Attoh-Okine, N. O. Predicting Roughness Progression Models in Flexi-
ble Pavements—An Evolutionary Algorithm Approach. Intelligent Engi-
neering Systems Through Neural Networks, Vol. 8, 1998, pp. 845–853.
8. Sundin, S., and C. Braban-Ledoux. Artificial Intelligence-Based Deci-
sion Support Technologies in Pavement Management. Computer-Aided
Civil & Infrastructure Engineering, Vol. 16, No. 2, 2001, pp. 143–157.
9. Loia, V., S. Sessa, A. Staiano, and R. Tagliaferri. Merging Fuzzy Logic,
Neural Networks, and Genetic Computation in the Design of a Decision-
Support System. International Journal of Intelligent Systems, Vol. 15,
No. 7, 2000, pp. 575–594.
10. Sait, S. M., and H. Youssef. Iterative Computer Algorithms with Appli-
cation in Engineering: Solving Combinatorial and Optimization Prob-
lems. IEEE Computer Society, Los Alamitos, Calif., 1999.
11. Burden, R. L., and J. D. Faires. Numerical Analysis, 4th ed. PWS-KENT
Publishing Company, Ashburton, Devon, U.K., 1989.
12. Lytton, R. L., F. P. Germann, Y. J. Chou, and S. M. Stoffels. NCHRP
Report 327: Determining Asphaltic Concrete Pavement Structural
Properties by Nondestructive Testing. TRB, National Research Council,
Washington, D.C., 1990.
13. ELMOD Version 5.1.54. Dynatest International, Glostrup, Denmark.
14. Ullidtz, P. Modelling Flexible Pavement Response and Performance.
Polyteknisk Forlag, Oslo, Norway, 1998.
15. Ferry, J. D. Viscoelastic Properties of Polymeric Material. Wiley,
New York, 1980.
16. Hastie, T. J., and R. J. Tibshirani. Generalized Additive Models. Chapman
and Hall, New York, 1990.
17. Tsai, B.-W. High Temperature Fatigue and Fatigue Damage Process of
Aggregate-Asphalt Mixes. Ph.D. dissertation. University of California,
Berkeley, 2001.
18. Stone, C. J. A Course in Probability and Statistics. Duxbury Press,
Pacific Grove, Calif., 1996.
19. Baumgaertel, M., and H. H. Winter. Determination of Discrete Relax-
ation and Retardation Time Spectra from Dynamic Mechanical Data.
Rheologica Acta, Vol. 28, 1989, pp. 511–519.
20. Stastna, J., L. Zanzotto, and J. Berti. How Good Are Some Rheological
Models of Dynamic Material Functions of Asphalt? Journal of the Asso-
ciation of Asphalt Paving Technologists, Vol. 66, 1997, pp. 458–485.
The opinions and conclusions expressed in this paper are those of the authors
and do not necessarily represent those of Caltrans.
Publication of this paper sponsored by Characteristics of Bituminous Paving
Mixtures to Meet Structural Requirements Committee.