SCIENCE CHINA

Technological Sciences
• Article • May 2016 Vol.59 No.5: 783–794
doi: 10.1007/s11431-016-6017-2

Parameters inversion of high central core rockfill dams based on

a novel genetic algorithm
ZHOU Wei1, LI ShaoLin1, MA Gang1*, CHANG XiaoLin1, MA Xing1 & ZHANG Chao2
1
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan 430072, China;
2
Changjiang Institute of Survey, Planning, Design and Research, Wuhan 430010, China

Received September 5, 2015; accepted December 30, 2015; published online February 19, 2016

Parameters identification of rockfill materials is a crucial issue for high rockfill dams. Because of the scale effect, random
sampling and sample disturbance, it is difficult to obtain the actual mechanical properties of rockfill from laboratory tests. Pa-
rameters inversion based on in situ monitoring data has been proven to be an efficient method for identifying the exact param-
eters of the rockfill. In this paper, we propose a modified genetic algorithm to solve the high-dimension multimodal and non-
linear optimal parameters inversion problem. A novel crossover operator based on the sum of differences in gene fragments
(SoDX) is proposed, inspired by the cloning of superior genes in genetic engineering. The crossover points are selected ac-
cording to the difference in the gene fragments, defining the adaptive length. The crossover operator increases the speed and
accuracy of algorithm convergence by reducing the inbreeding and enhancing the global search capability of the genetic algo-
rithm. This algorithm is compared with two existing crossover operators. The modified genetic algorithm is then used in com-
bination with radial basis function neural networks (RBFNN) to perform the parameters back analysis of a high central earth
core rockfill dam. The settlements simulated using the identified parameters show good agreement with the monitoring data,
illustrating that the back analysis is reasonable and accurate. The proposed genetic algorithm has considerable superiority for
nonlinear multimodal parameter identification problems.

rockfill dam, parameters back analysis, genetic algorithm, crossover operator, sum of differences in gene fragments

Citation: Zhou W, Li S L, Ma G, et al. Parameters inversion of high central core rockfill dams based on a novel genetic algorithm. Sci China Tech Sci, 2016,
59: 783−794, doi: 10.1007/s11431-016-6017-2

1 Introduction atory testing, engineering analogy and empirical studies.

However, due to the scale effect [2], random sampling and
sample disturbance, the in situ mechanical properties of
Deformation control is an important guarantee for the safety
rockfill may differ from those determined by laboratory
of high rockfill dams. In the design stage of a rockfill dam,
tests. The engineering analogy method and empirical meth-
various numerical methods, such as the finite element
od are too simple to capture the complex mechanical be-
method (FEM) [1], are used to predict the deformation and
havior of rockfill materials. Parameter back analysis of
stress field. Appropriate rockfill constitutive models and
rockfill materials based on the in situ settlement monitoring
mechanical parameters that match these models are the core
data has proven to be an effective way to identify these pa-
of both stress-strain analysis and the safety evaluation.
rameters.
There are three ways to obtain these parameters, i.e., labor-
The parameter back analysis of rockfill dams can be ab-
stracted as an optimization problem in mathematics. This
*Corresponding author (email: [email protected])
optimization problem is difficult because there are more

© Science China Press and Springer-Verlag Berlin Heidelberg 2016 tech.scichina.com link.springer.com
784 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5

than three different material zones in the high rockfill dams tive length coefficient strategy is utilized in the crossover to
and more than four insensitive parameters for each zone. increase the algorithm’s global search capability.
Moreover, the number of parameters needed to be identified The algorithm is used in combination with the RBFNN to
will increase dramatically when taking the creep parameters perform the parameters back analysis of a high central earth
into consideration. The parameter back analysis of rockfill core rockfill dam. Comparative studies show good agree-
dams is a strong nonlinear, high-dimension and large space ment between the measured and simulated settlements, il-
optimization problem. A variety of population-based proba- lustrating that the back analysis is reasonable and the accu-
bilistic techniques have been proposed in the literature to racy is acceptable.
solve these optimization problems, including genetic algo- This paper is organized as follows. After this introduc-
rithms (GAs) [3,4], simulated annealing (SA) [5–8], differ- tion, Section 2 presents the proposed crossover operator and
ential evolution (DE) [9–11], and particle swarm optimiza- the details of the real-coded genetic algorithm. Section 3
tion (PSO) [12–15]. Among these techniques, GAs, pro- explains the problem, experimental setup and experimental
posed by Holland [3] in 1975, have proven to be powerful results. Back analysis of the mechanical parameters for
for solving optimization problems. rockfill materials is performed in Section 4. Finally in Sec-
In a classic GA, the optimization process is driven by tion 5, the conclusions are given.
three operators, namely, crossover, mutation, and selection,
and the optimization variables are encoded into chromo-
somes in binary alphabets [16,17]. Unfortunately, binary 2 The proposed adaptive genetic crossover
GAs can be time-consuming and require large memory ca-
pacities when applied to optimization problems that seek The crossover operator is one of the most determinant oper-
high precision and have a large search space [18]. To over- ations in genetic algorithms. It affects the diversity of the
come these difficulties, real-coded GAs have been devel- population and the disproportionation between exploitation
oped. Janikow and Michalewicz [19] demonstrated that the and exploration, influences the speed of convergence and
performance of real-coded GAs surpasses that of bina- determines the global convergence of the algorithm. In clas-
ry-coded GAs in continuous optimizations. In real-coded sic genetic algorithms, the crossover points and gene frag-
GAs, the crossover operator plays a pivotal role in the algo- ments are selected randomly and blindly. It is difficult to
rithm’s ability to explore and exploit. In recent years, plenty generate new individuals when the selected gene fragments
of work has focused on constructing new crossover opera- are the same or highly similar. That is, the traditional
tors to improve the performance of real-coded GAs. Micha- crossover is not very effective for producing a new genera-
lewicz et al. [20], for instance, proposed a simple crossover tion that is different from its parent, and this crossover op-
in which two new offspring are generated by simply ex-
eration may be invalid.
changing the genes that are randomly selected from a pair of
Inspired by genetic engineering and the cloning of supe-
parents. Radcliffe [21] suggested a flat crossover from
rior genes, a novel genetic crossover operator based on the
which offspring are made using a uniform distribution. Ad-
sum of difference in gene fragments (SoDX) is proposed.
ditionally, Ono et al. [22] presented a unimodal distribution
The fragments’ crossing probabilities are first evaluated
crossover operator, utilizing multiple parents to produce
offspring that are centered around the parents. A heuristic based on the differences. The gene fragments are then se-
crossover proposed by Wright uses the fitness value to gen- lected and exchanged according to the crossing probability.
erate offspring [23]. In addition, an increasing number of This process could reduce inbreeding and the possibility of
algorithms have been suggested to improve the convergence invalid crossover operation.
speed of such GAs, such as simplex crossover [24], par- In this study, a gene fragment is a continuous period of
ent-centric crossover [25–27], and multi-parent crossover the chromosome with random length, and the crossover
[28]. However, in these crossover operators, the crossover point is the first gene of the fragment. We define the differ-
points and gene fragments are selected randomly and blind- ence in gene fragment D(j) as follows:
ly, which may lead to inbreeding or make the crossover
1 ( j + L −1)
operation invalid when the parent's gene fragments are D( j ) =
L i= j
 abs( xi(1) − xi(2) ), j = (1, 2, , n − L + 1), (1)
highly similar.
Inspired by the cloning of specific genes in genetic engi-
neering, a novel real-coded crossover operator has been where L is the length of the gene fragment; xi(1) and xi(2)
proposed based upon the sum of differences (SoD) in gene are the ith gene of the chromosome x (1) = ( x1(1) , x2(1) , ,
fragments. In the new crossover operator, the lengths of the
gene fragments are randomly determined, and the difference xn(1) ) and x (2) = ( x1(2) , x2(2) , , xn(2) ), respectively; and n is
in each gene fragment is calculated. Next, the gene frag- the length of the chromosome.
ments exchange at the cross point, which is selected based The gene fragment j is selected to intersect with proba-
on the differences in gene fragments. Furthermore, an adap- bility P(j):
 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5 785

D( j ) I is the coefficient of the gene fragment.

P( j ) = ( n − L +1)
, j = (1, 2, , n − L + 1), (2) The coefficient I controls the overall length of the gene
 D(i ) fragment, while u1 increases the diversity of the crossover
i =1 operation on the basis of the length coefficient. To evaluate
where D(i) and D(j) are the differences in gene fragments i the performance of the algorithm with different length coef-
and j, respectively. ficients, three gene fragment length strategies, i.e., the use
The binary gene fragment difference is calculated as of a fixed length coefficient, linear increasing length coeffi-
shown in Figure 1, and the differences for the first and se- cient and linear decreasing length coefficient, were studied
cond chromosomes are 9 and 21, respectively. The first with a benchmark function named Rastrigin. In the numeri-
gene fragment (I) does not markedly help produce better cal test, when I is fixed to be 0.2, the algorithm obtains the
generations, as the fragments are very similar to each other. best results using these fixed coefficient strategies (shown
Meanwhile, the second gene fragment has much higher in Table 1). The linear decreasing strategy was much more
probability of producing generations that are different from effective than the fixed or the linear increasing strategy, as
their parents due to the higher gene fragment difference. illustrated in Table 2. Additionally, the three decreasing
The greater the difference of the gene fragment, the greater strategies shown in Figure 2 were compared regarding the
the likelihood of generating new chromosomes and avoid- convergence speed of the algorithm. As shown in Table 3, it
ing invalid crossover operation. Upon this point, an innova- can be concluded that the concave strategy performed best.
tive crossover operator based on the difference in gene Thus, the concave strategy was used in this paper, and the
fragments is proposed. As mentioned before, real-coded length coefficient of the gene fragment can be determined
GAs could overcome the difficulties of binary GAs, such as as follows:
their large time and memory requirements when applied to 2
g 
optimization problems that seek high precision and have a I = I min + ( I max − I min )  − 1 , (4)
large search space [14]. In this paper, we use a real-coded  G 
chromosome, and the differences of gene fragments are where Imin and Imax are the minimum and maximum length
calculated by eq. (1). The gene fragments to intersect are coefficients, respectively; g is the current iteration; and G is
selected according to eq. (2). the maximum iteration.
In the SoDX crossover operator, the convergence is sig- Using this adaptive concave decreasing length coefficient
nificantly influenced by the length of the gene fragment, strategy, the individuals are encouraged to traverse the en-
which is defined as follows: tire search space in the early stage of the evolution, not
Lfragment = u1 Lchrom I , (3) gather around a super individual, which would lead to
premature convergence. Meanwhile, in the later period, in
where u1 (u1∈[0,1]) is a random number called the random dividuals are encouraged to search near the optimum.
length ratio, Lchrom is the total length of the chromosome and The flow of the adaptive crossover operator based on the

Figure 1 (Color online) The sum of difference (SoD) for a gene fragment.
786 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5

SoDX is defined as follows. ver (DPX) [27] and heuristic crossover (HX) [23], were
Step 1: Generate a random number u1 (u1∈[0,1]) and selected for comparison. Similarly, the other two GAs were
compute the length of the gene fragment using eqs. (3) and named DPX-NUM and HX-NUM, respectively. The fitness
(4). proportional model was used as the selection operator in all
the GAs. The pseudo code of the three GAs is shown as
Step 2: Select the parents x (1) = ( x1(1) , x2(1) , , xn(1) ) and
follows:
x (2) = ( x1(2) , x2(2) , , xn(2) ) using the roulette selection crite-
rion. Begin
Step 3: Calculate D( j )(i = 1, 2, , (n − Lfragment + 1)) m=0
initialize P(m)
and P ( j )( j = 1, 2, , (n − Lfragment + 1) . Select the cross evaluate fitness P(m)
point istart according to P ( j ) . While (criterion is not satisfied) do
m=m+1
Step 4: Generate another random number u2 (u2 ∈ [0,1]) , select P(m) from P(m–1)
intersect the selected gene fragment of parents x(1) = apply (SoDX/DPX/HX) on P(m)
( x1(1) , x2(1) , , xn(1) ) and x (2) = ( x1(2) , x2(2) , , xn(2) ) , and then apply mutation NUM operator
End do
acquire children y (1) = ( y1(1) , y2(1) , , yn(1) ) and y (2) = End
( y1(2) , y2(2) , , yn(2) ) using the follow equations:
Table 1 Results for different fixed coefficients
u2 ( xi(2) − xi(1) ) + xi(1) , i ∈ (iStart , iStart + Lfragment − 1),
yi(1) = (5) I 0.1 0.2 0.3 0.4 0.5
(1)
 xi , i ∉ (iStart , iStart + Lfragment − 1), Best fitness 3.32×10−2 4.07×10−6 4.60×10−6 3.32×10−2 3.32×10−2
I 0.6 0.7 0.8 0.9 1
yi(2) Best fitness 7.36×10−6 3.32×10−2 6.63×10−2 1.00×10−5 1.20×10−5
u2 ( xi(1) − xi(2) ) + xi(2) , i ∈ (iStart , iStart + Lfragment − 1), (6)
=
(2)
 xi , i ∉ (iStart , iStart + Lfragment − 1), Table 2 Results for different length coefficient strategies

where i is the ith gene of the chromosome and iStart is the Strategy Fixed (0.2) Linear increasing Linear decreasing
cross point. Best fitness 4.07×10−6 3.03×10−8 1.31×10−8

3 Numerical experiments and results

3.1 Benchmark functions

A set of 14 benchmark functions [14,27,29] was used to

evaluate the performance of real-coded GAs with different
crossover operators. These benchmark functions can gener-
ally be divided into three categories: unimodal functions;
multimodal functions with a few local minima, which are
easy to distinguish from the global optimal solution; and
multimodal functions with many local minima, which are
difficult to distinguish from the global optimal solution. The
benchmark functions and the functional characteristics are
detailed in Tables 4 and 5.
Figure 2 (Color online) Curves for different length coefficient decreasing
3.2 The experimental setup strategies.

A new GA code named SoDX-NUM was implemented in

MATLAB by combining the operator defined in Section 2 Table 3 Results for different length coefficient decreasing strategies
with the non-uniform mutation (NUM) [30]. To evaluate the Decreasing strategy Linear Concave Convex
performance of the proposed crossover operator SoDX, two −8 −9
Results 1.31×10 7.85×10 1.43×10−8
well-known crossover operators, the double Pareto crosso-
 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5 787

Table 4 Functional characteristics

Problem Unimodal/ Distinguishability of the local and

Function name Optimal value
name Multimodal global optimal solutions
f1 Sphere Unimodal - x* = (0, , 0), f ( x* ) = 0
f2 Rotated hyper-ellipsoid Unimodal - x* = (0, , 0), f ( x* ) = 0
Group 1
f3 Bent cigar Unimodal - x* = (0, , 0), f ( x* ) = 0
f4 Discus Unimodal - x* = (0, , 0), f ( x* ) = 0
f5 Ackley Multimodal Easy x* = (0, , 0), f ( x* ) = 0
f6 Griewank Multimodal Easy x* = (0, , 0), f ( x* ) = 0
Group 2 f7 Cosine mixture Multimodal Easy x* = (0, , 0), f ( x* ) = 0
f8 Werierstrass Multimodal Easy x* = (0, , 0), f ( x* ) = 0
f9 Ellipsoidal Multimodal Easy x* = (0, , 0), f ( x* ) = 0
f10 Rastrgin Multimodal Difficult x* = (0, , 0), f ( x* ) = 0
f11 Schwefel’s 2 Multimodal Difficult x* = (420.9687, , 420.9687), f ( x* ) = 0
Group 3 f12 Schwefel’s 22 Multimodal Difficult x* = (0, , 0), f ( x* ) = 0
f13 Levy and Montalvo Multimodal Difficult x* = (0, , 0), f ( x* ) = 0
f14 New Multimodal Difficult x* = (1, ,1), f ( x* ) = 0

Table 5 Benchmark functions

Problem number Formula Change

n 2
f1 min f ( x) =  x
i =1 i
x ∈ [ −5.12,5.12]
n i
f2 min f ( x ) =  i =1 ( j =1 x j )2 x ∈ [ −65.536, 65.536]

n
f3 min f ( x) = x12 + 106  i = 2 ( xi2 ) x ∈ [ −100,100]
n
f4 min f ( x) = 106 x12 +  i = 2 ( xi2 ) x ∈ [ −100,100]

1 n 2 1 n x ∈ [ −32.786,32.786]
f5 min f ( x ) = 20 + e − 20 exp(−0.2  xi − exp( n  i = 2 cos(2πxi ))
n i =1
1 n n x
f6 min f ( x) =  xi2 − ∏ i =1 cos( ii ) + 1
4000 i =1
x ∈ [ −600, 600]

n n
f7 min f ( x ) = 0.1 i =1 cos(5πxi ) +  i =1 xi2 x ∈ [ −1,1]
n kmax kmax
min f ( x ) =  i =1 ( k = 0  a k cos(2πb k ( xi + 0.5))  ) − n k = 0  a k cos(2πb k ⋅ 0.5) 
f8     x ∈ [ −0.5, 0.5]
Where a = 0.5 , b = 0.3 , and kmax=20
n
f9 min f ( x) =  i =1 ( xi2 − i ) x ∈ [ − n, n ]
n
f10 min f ( x ) = 10n +  ( x 2 − 10 cos(2πxi ))
i =1 i
x ∈ [ −5.12,5.12]

xi
min f ( x) = 418.9829n −  i=1 cos( x ∈ [ −500,500]
n
f11 ) +1
i
min f ( x ) =  i =1 xi − ∏ i =1 xi x ∈ [ −10,10]
n n
f12
2
f13 min f ( x) = 0.1(sin 2 (3πx1 ) +  i=1 ( xi − 1) 1 + sin 2 (3πxi+1 )  + ( xn + 1) 2 1 + sin 2 (2πxi+1 )  )
n−1
x ∈ [ −5,5]

π 2
min f ( x ) = 10sin 2 (πx1 ) +  i=1 ( xi − 1) 1 + sin 2 (3πxi+1 )  + ( xn − 1) 2 x ∈ [ −10,10]
n−1
f14
n

3.3 Experimental results DPX-NUM was set as 1; k for HX-NUM was set as 4; and
Imin and Imax for SoDX-NUM were set as 0.1 and 0.9, re-
In the three GAs, the crossover and mutation probabilities spectively. For each function, the population size for all
were set as 0.8 and 0.1, respectively. The value of β for tests was fixed at 50. Three different dimensions (30, 60 and
788 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5

90) were tested, and the corresponding maximum iteration functions. For f1 and f2, the computed optima using the
numbers for each generation were 1500, 2500 and 3500, SoDX crossover operator are the closest to the real optimum
respectively. Each algorithm ran 30 times independently value. As shown in Table 6 and Figure 3(a), the mean of the
with different initial populations. The experimental results best fitness of SoDX-NUM is significantly x∈[−5.12, 5.12]
of the 30 independent runs in terms of the means and stand-
better than that of either DPX-NUM or HX-NUM. In addi-
ard deviations of the optima for f1 to f14 are tabulated in Ta-
ble 6. The comparisons between different genetic crossover tion, SoDX-NUM has a much smaller standard deviation,
operations on to f14 are shown in Figure 3. suggesting that the solution of SoDX-NUM is much more
Functions f1–f4 are unimodal functions. Function f1 is a stable. Furthermore, in Figure 3(a), SoDX-NUM displays a
sphere model, and function f2 is a rotated hyper-ellipsoid faster convergence rate than DPX-NUM or HX- NUM due
model, which are the most widely used test benchmark to its superior searching ability. Function f3 is a bent cigar

Table 6 Means and standard deviations of benchmark function fitness values for the three GAs

Function Mean Standard deviation

Dimension Generation
name SoDX-NUM DPX-NUM HX-NUM SoDX-NUM DPX-NUM HX-NUM
30 1500 2.63×10−9 4.85×10−8 1.76×10−7 1.82×10−10 8.84×10−9 1.90×10−8
f1 60 2500 1.29×10−7 3.70×10−6 7.64×10−6 3.51×10−10 1.56×10−7 7.67×10−7
90 3500 9.76×10−7 2.86×10−5 4.43×10−5 1.22×10−8 1.66×10−6 1.10×10−6
30 1500 2.61×10−7 8.85×10−6 2.33×10−5 1.26×10−8 7.55×10−7 2.11×10−6
f2 60 2500 1.32×10−5 6.65×10−5 4.02×10−4 1.40×10−6 3.22×10−6 4.27×10−5
90 3500 1.61×10−4 3.74×10−4 3.24×10−3 4.37×10−6 4.72×10−5 7.28×10−5
30 1500 3.61×10−2 2.02 6.06 4.22×10−3 3.53×10−1 9.50×10−1
f3 60 2500 1.45 1.72×101 1.12×102 1.42×10−1 2.56 1.53×101
90 3500 6.14 7.80×101 2.22×102 7.54×10−1 3.30×101 2.00×101
30 1500 2.67×10−6 2.28×10−6 3.12×10−4 5.63×10−7 5.12×10−7 2.33×10−4
f4 60 2500 1.40×10−5 3.51×10−5 6.13×10−4 2.28×10−6 5.57×10−6 8.46×10−5
90 3500 3.86×10−5 5.56×10−5 5.75×10−4 2.99×10−6 5.74×10−6 1.50×10−4
30 1500 7.34×10−5 5.23×10−4 6.55×10−4 4.33×10−6 3.01×10−6 2.71×10−5
f5 60 2500 3.12×10−4 1.99×10−3 2.13×10−3 8.14×10−6 4.17×10−5 7.12×10−7
90 3500 5.04×10−4 3.43×10−3 3.25×10−3 4.44×10−6 1.19×10−5 2.31×10−4
30 1500 1.14×10−2 3.83×10−2 2.47×10−2 6.71×10−4 3.22×10−3 1.05×10−3
f6 60 2500 1.84×10−2 5.13×10−1 1.81×10−1 8.10×10−5 2.01×10−3 1.22×10−2
90 3500 5.16×10−2 8.59×10−1 4.03×10−1 3.08×10−3 2.19×10−3 1.00×10−2
30 1500 1.13×10−15 1.71×10−15 1.56×10−15 1.12×10−16 4.86×10−17 2.64×10−16
f7 60 2500 2.32×10−16 3.36×10−16 2.77×10−16 1.96×10−17 8.21×10−18 2.52×10−17
90 3500 9.29×10−17 1.29×10−16 1.03×10−16 1.44×10−17 1.79×10−17 1.37×10−17
30 1500 2.91×10−11 1.49×10−10 1.03×10−9 8.3×10−13 1.06×10−11 1.79×10−11
f8 60 2500 9.92×10−10 5.01×10−9 3.58×10−8 5.49×10−11 1.61×10−11 1.2×10−9
90 3500 5.83×10−9 2.29×10−8 1.57×10−7 1.52×10−11 1.61×10−9 4.49×10−9
30 1500 8.96×10−6 5.85×10−7 4.59×10−4 9.61×10−7 1.04×10−7 7.36×10−5
f9 60 2500 6.56×10−4 1.01×10−4 5.78×10−4 1.87×10−4 1.79×10−5 9.15×10−5
90 3500 2.89×10−3 7.93×10−4 6.07×10−3 4.66×10−4 2.73×10−5 7.28×10−4
30 1500 7.06×10−7 1.96×10−6 2.94×10−9 5.80×10−8 2.13×10−7 7.06×10−7
f10 60 2500 1.26 3.32×10−2 5.16×10−8 4.85×10−2 6.06×10−3 1.26
90 3500 5.18 1.29 6.06×10−3 1.46×10−1 5.45×10−2 5.18
30 1500 4.38×102 2.76×102 9.37 8.00×101 1.44×101 4.38×102
f11 60 2500 1.27×103 1.18×103 1.28×102 2.09×101 2.16×101 1.27×103
90 3500 2.42×103 2.23×103 7.28×101 5.63×101 3.89×101 2.42×103
30 1500 1.56×10−12 1.30×10−5 1.66×10−13 2.61×10−13 1.49×10−5 1.56×10−12
f12 60 2500 2.95×10−13 3.28×10−4 4.38×10−14 2.16×10−13 5.99×10−5 2.95×10−13
90 3500 1.23×10−13 1.69×10−3 1.02×10−14 1.90×10−14 3.09×10−4 1.23×10−13
30 1500 9.55×10−12 4.34×10−11 4.13×10−10 9.67×10−13 5.26×10−12 2.84×10−11
f13 60 2500 3.71×10−10 1.45×10−9 1.01×10−8 1.86×10−11 7.56×10−11 2.18×10−10
90 3500 1.98×10−9 5.29×10−9 4.4×10−8 4.36×10−11 1.96×10−10 1.61×10−9
30 1500 3.68×10−10 1.86×10−9 1.58×10−8 1.04×10−10 4.85×10−11 4.37×10−9
f14 60 2500 1.41×10−8 5.22×10−8 4.04×10−7 1.29×10−9 6.24×10−10 3.57×10−8
90 3500 6.34×10−8 2.14×10−7 1.57×10−6 2.5×10−9 2.85×10−9 5.08×10−8
 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5 789

Figure 3 (Color online) Evolution of the logarithmic average fitness for HX-NUM, DPX-NUM, and SoDX-NUM (all results are averaged over 30 runs). (a)
The Group 1; (b) the Group 2; (c) the Group 3.

model, which is smooth but of narrow ridge, and f4 is a dis- and finds the global optimum at a much faster convergence
cus model with one sensitive direction. For f3, the optimum rate (as shown in Figure 3(b)).
of the SoDX crossover is closer to the real optimum than For functions f10–f14, the mean and the standard deviation
those of the other two crossovers. For f4, the results of the obtained by the proposed SoDX-NUM are significantly
SoDX crossover are no better than those of the DPX cross- better than those obtained by DPX-NUM and HX-NUM,
over and HX crossover when the dimension is 30, but the irrespective of the dimensions. As shown Figure 3(c), the
SoDX crossover performs well when the dimensions are 60 convergence of the proposed genetic crossover SoDX is
and 90. The SoDX crossover cannot fully exert its ad- faster than that of the conventional genetic crossovers.
vantage when the number of the variables is small. For In short, the proposed genetic crossover operator SoDX
high-dimension problems, the SoDX crossover has great can effectively address both unimodal and multimodal
global search ability in the early period of the evolution and problems and outperforms two popular genetic operators.
then encourages the individuals to traverse near the opti- Both the solution accuracy and stability are satisfactory.
mum.
Table 6 and Figure 3(b) shows the mean and the standard
deviation of the best fitness for benchmark functions f5–f9. 4 Parameters inversion of a high central earth
In almost all aspects, SoDX-NUM outperforms the DPX- core rockfill dam
NUM and HX-NUM. Meanwhile, the standard deviations of
the best fitness of SoDX-NUM are significantly smaller 4.1 Constitutive model
than those of DPX-NUM and HX-NUM. For f6, all crosso-
ver operators converge to a local minima after 2000 itera- 4.1.1 Duncan & Chang E-B model
tions, while SoDX escapes from the local optimum quickly The mechanical behavior of rockfill materials is closely
790 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5

related to the gradation characteristics, initial void ratio, and b, β, mc, and nc are the creep parameters.
confining pressure, etc. [31–34]. Due to its simplicity and The limit of shear creep strain εsf is associated with the
the explicit physical meanings of its input parameters, the stress level SL, as given below:
Duncan & Chang E-B model has been widely used in civil
l
engineering [35], especially for rockfill dams in China. In  SL  c
the E-B model, the tangent modulus Et is given as follows:
ε sf = d   , (13)
 1 − SL 
n
σ  where SL is the stress level and d and lc are the creep param-
Et = Kpa  3  [1 − SL Rf ] ,
2
(7)
 pa  eters.
There are seven parameters in the creep model. α con-
where K is the modulus number, pa is the atmospheric pres- trols the creep strain rate; b, β, mc and nc are defined for the
sure, and σ3 is the minor principal stress. n and Rf are the limit of creep strain; and d and lc are defined for volume
exponent and failure ratio, respectively. The stress level SL creep. All of the parameters can be evaluated using a set of
is defined as creep triaxial tests.
(1 − sin ϕ )(σ 1 − σ 3 )
SL = , (8)
2c cos ϕ + 2σ 3 sin ϕ 4.2 Parameter sensitivity analysis

where ϕ = ϕ0 − Δϕ lg(σ 3 / pa ) is the internal friction angle For the Duncan & Chang E-B model and the Merchant
creep model of rockfill materials, fourteen parameters need
and c is the cohesive strength. σ1 and σ3 are the major and
minor principal stresses, respectively. to be identified. In addition, when considering dam material
Under unloading and reloading conditions, the elastic zoning, the number of inversion parameters will be too large
modulus Eur is expressed as to identify. Therefore, it is necessary to perform a parameter
sensitivity analysis before the parameters inversion. Using
n
σ  the modified Morris’ method, the parameter sensitivities of
Eur = K ur pa  3  . (9) the Duncan & Chang E-B model and the Merchant creep
 pa 
model are quantified to select the inversion parameters [37].
The bulk modulus Bt is given by As shown in Figure 4, the most sensitive parameters of
m
the E-B model and creep model are K, n, Kb, m and mc, nc,
σ  and d. Therefore, K, n, Kb, m, mc, nc, and d are selected to be
Bt = K b pa  3  . (10)
 pa  inverted in the parameters back analysis. The parameter α
controls the creep strain rate. There is a large difference
As illustrated above, there are nine parameters in the between the triaxial creep tests, in which the creep defor-
Duncan & Chang E-B model, i.e., c, ϕ, Δϕ, Rf, K, n, Kur, Kb, mation stabilizes within a few hours, and the in situ defor-
and m, which can be determined by a set of conventional mation monitoring data, in which the creep deformation
triaxial tests on rockfill dams.
continues for over six years. Thus, the parameter α is cho-
4.1.2 Merchant creep model of rockfill materials sen to be identified as well.
In this paper, the Merchant seven-parameter creep model is
used to describe the creep behavior of rockfill materials [36]. 4.3 Finite element model
The creep strain εt is expressed as follows:
A three-dimensional finite model is built for Pubugou Dam
ε t = ε f (1 − e −α t
), (11) (as shown in Figure 5). The FEM mesh consists of 52479
elements and 54630 nodes. Most of the elements are eight
where εf is the limit of the creep strain and α is a coefficient nodes of hexahedral elements, whereas some tetrahedron
related to the creep strain rate. The limit of creep strain εf elements are used in the boundary zone. The construction
can be divided into the limit of volume creep strain εvf and and water storage process is simulated in the FEM analysis
the limit of shear creep strain εsf. The limit of volume creep based on the real construction schedule. The layout of the in
strain εvf, determined by the minor principal stresses σ3 and situ monitoring gauges in section 0+240 is shown in Figure
the deviatoric stress, is given by the following relationship: 6.
mc n
 σ3   q c
ε vf = b   +β  , (12) 4.4 Parameters back analysis method
 pa   pa 
In parameters back analysis, one group of parameters needs
where σ3 is the confining pressure, q is the deviatoric stress, to be inverted such that the corresponding simulated dis-
 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5 791

Figure 5 The finite element model of Pubugou Dam.

of deformation at the ith measuring point, ui* is the corre-

sponding observed value and n is the number of dis-
placement measuring points used in the back analysis.
In the process of the parameters back analysis, the FEM
simulation is repetitive. To improve the computational effi-
ciency, the RBFNN is used instead of FEM to map the rela-
tionships between the mechanical parameters and the objec-
tive function fitness. As shown in Figure 7, the back analy-
sis process is as follows.
1) FEM analysis. Generate the parameter set using or-
thogonal and random algorithms to ensure the representa-
tiveness and diversity of the sample. Next, perform the non-
Figure 4 (Color online) Parameter sensitivity. (a) Duncan & Chang E-B
model; (b) Merchant creep model.
linear analysis using the FEM and calculate the fitness for
each set of parameters utilizing the FEM results and obser-
vation data.
placement approximates the observed displacements as 2) Training and testing the RBFNN. Using the sample
closely as possible. The 2-norm of the difference between parameters as the input, obtain the fitness as the output.
the simulated and observed displacements at the measuring Here, 80% of the samples are used to train the RBFNN, and
points can be taken as an objective function. 20% are used to test the RBFNN. After successful training
0.5 and testing, RBFNN is used to calculate the fitness, instead
 n  u (λ , λ ,···, λ ) − u*  2  of the FEM.
1 
f (λ1 , λ2 ,···, λn ) =    i 1 2 * n i
  , (14) 3) Evolution. Use the modified genetic algorithm
 n i =1  ui  
 (SoDX-NUM) genetic operations, such as selection, cross-
over and mutation, to find the best individual and obtain the
where λ1 , λ2 ,···, λn is the group of rockfill model parame-
inversion parameters. In the evolution, RBFNN is utilized to
ters to be inverted, ui (λ1 , λ2 ,···, λn ) is the simulated value compute the fitness.

Figure 6 Layout of the monitoring system in section 0+240 of the Pubugou CCRD.
792 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5

Figure 7 Flow chart of parameter inversion using the SoDX-NUM and RBFNN.

4.5 Back analysis results rial and its implementation are reasonable. The settlement
and horizontal displacement contours of cross-section 0+
The identified parameters of the Duncan & Chang E-B
240 based on the identified parameters are displayed in Fig-
model and creep model are presented in Table 7. The evolu-
ure 9(a) and (b). The maximum horizontal displacement is
tions of cumulative settlement with time in cross-section
107 cm, pointing to the downstream. The maximum settle-
0+240 are compared in Figure 8. The settlement simulated
using the back-analysed parameters is much closer to the ment is 432 cm, approximately 2.32% of the maximum dam
observed data than that simulated using the test parameters. height, located at the lower middle of the maximum
The magnitude and distribution of the settlements of the cross-section. As is shown in Figure 9(c), the lateral defor-
back analysis show good agreement with the observed re- mation of the dam axis-section is symmetric and the maxi-
sults, suggesting that the back analysis of the rockfill mate- mum is about 50 cm pointing to the centre of the river.

Table 7 Back analysis results

Creep model E-B model

Material zone
α (10−3) mc nc d (10−3) K n Kb m
Central core rockfill 3.495 1.123 0.790 7.636 532 0.46 264 0.32
Secondary rockfill 1.218 0.952 0.544 7.271 807 0.40 462 0.37
Main rockfill 1.333 0.310 0.433 4.45 666 0.54 261 0.26

Figure 8 (Color online) Evolution of settlements with time in cross-section 0+240. (a) At the elevation of 731 m; (b) at the elevation of 758 m.
 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5 793

Figure 9 The settlement simulated results using the indentified parameters (cm) for May 2012. (a) Settlement contours of cross-section 0+240; (b) hori-
zontal displacement contours of cross-section 0+240; (c) lateral deformation contours of the dam axis-section.

5 Conclusion genetic algorithm greatly outperforms other genetic algo-

rithms for nonlinear multimodal parameter identification
problems.
In this paper, a new adaptive genetic crossover operator
based on the sum of differences in gene fragments (SoDX)
was proposed, drawing inspiration from genetic engineering This work was supported by the National Natural Science Foundation of
and the cloning of specific genes. SoDX reduced inbreeding China (Grant Nos. 51379161 & 51509190), China Postdoctoral Science
Foundation (Grant No. 2015M572195), and the Fundamental Research
and enhanced the effectiveness and efficiency of the classic
Funds for the Central Universities.
crossovers, increasing the convergence speed and accuracy.
A set of nonlinear benchmark problems was used to test the
performance of the crossover operator. The proposed cros- 1 Wei Z, Xiaolin C, Chuangbing Z, et al. Creep analysis of high con-
crete-faced rockfill dam. Int J Numer Meth Bio, 2010, 11: 1477–1492
sover was compared with two existing crossover operators,
2 Hua J, Zhou W, Chang X, et al. Study of scale effect on stress and
DPX and HX. These results indicate that the proposed deformation of rockfill. Chin J Rock Mech Eng, 2010, 2: 328–335
crossover operator is especially effective in solving high- 3 Holland J H. Adaptation in natural and artificial system: An introduc-
dimensional nonlinear problems. By introducing SoDX into tion with application to biology, control and artificial intelligence.
a classic genetic algorithm, the ability to escape from local Dissertation of Doctor Degree. Michigan: University of Michigan,
1975
optima is enhanced, and the global search ability is
4 Holland J H. Adaptation in Natural and Artificial Systems: An Intro-
strengthened. ductory Analysis with Applications to Biology, Control, and Artifi-
Inversion analysis of the parameters for rockfill materials cial Intelligence. Cambridge: MIT Press, 1992
was performed using RBFNN and the proposed genetic al- 5 Van Laarhoven P J, Aarts E H. Simulated Annealing. Netherlands:
gorithm. Comparative studies show good agreement be- Springer Science & Business Media, 1987
6 Aarts E, Korst J. Simulated Annealing and Boltzmann Machines: A
tween the simulated settlement and observed settlement,
Stochastic Approach to Combinatorial Optimization and Neural
illustrating that the back-analysis is an effective way to de- Computing. Chichester: John Wiley and Sons Inc., 1988
termine the parameters of rockfill materials. The proposed 7 Dowsland K A, Thompson J M. Simulated Annealing. Berlin:
794 Zhou W, et al. Sci China Tech Sci May (2016) Vol.59 No.5

Springer Berlin Heidelberg, 2012. 1623–1655 23 Wright A H. Genetic algorithms for real parameter optimization.
8 Shen H, Zhou J, Peng Q, et al. Multi-objective interplanetary trajec- Found Genet Algorithm, 1991, 1: 205–218
tory optimization combining low-thrust propulsion and gravity-assist 24 Tsutsui S, Yamamura M, Higuchi T. Multi-parent recombination
maneuvers. Sci China Tech Sci, 2012, 55: 841–847 with simplex crossover in real coded genetic algorithms. In: Pro-
9 Storn R, Price K. Differential Evolution-A Simple and Efficient Adap- ceedings of the Genetic and Evolutionary Computation Conference,
tive Scheme for Global Optimization Over Continuous Spaces. Berke- Florida, 1999. 657–664
ley: ICSI, 1995 25 Deb K, Anand A, Joshi D. A computationally efficient evolutionary
10 Herrera F, Lozano M, Verdegay J L. Fuzzy connectives based cross- algorithm for real-parameter optimization. Evol Comput, 2002, 4:
over operators to model genetic algorithms population diversity. 371–395
Fuzzy Set Syst, 1997, 1: 21–30 26 Deep K, Thakur M. A new crossover operator for real coded genetic
11 Storn R, Price K. Differential evolution–a simple and efficient heu- algorithms. Appl Math Comput, 2007, 1: 895–911
ristic for global optimization over continuous spaces. J Global Optim, 27 Thakur M. A new genetic algorithm for global optimization of mul-
1997, 4: 341–359 timodal continuous functions. J Comput Sci-Neth, 2014, 2: 298–311
12 Eberhart R C, Kennedy J. A New Optimizer Using Particle Swarm 28 Elsayed S M, Sarker R A, Essam D L. A new genetic algorithm for
theory. New York: Nagoya, 1995. 39–43 solving optimization problems. Eng Appl Artif Intel, 2014, 27: 57–69
13 Kennedy J. Particle Swarm Optimization. Boston: Springer, 2010. 29 Chen Q, Liu B, Zhang Q, et al. Problem definition and evaluation
760–766 criteria for CEC 2015 special session and competition on bound con-
14 Ma G, Zhou W, Chang X. A novel particle swarm optimization algo- strained single-objective computationally expensive numerical opti-
rithm based on particle migration. Appl Math Comput, 2012, 11: mization. https://mysites.ntu.edu.sg
6620–6626 30 Michalewicz Z. Genetic Algorithms + Data Structures = Evolution
15 Zhang S, Ou J. BP-PSO-based intelligent case retrieval method for Programs. Berlin: Springer Science & Business Media, 2013
high-rise structural form selection. Sci China Tech Sci, 2013, 56: 31 Xiao Y, Liu H, Chen Y, et al. Strength and deformation of rockfill
940–944 material based on large-scale triaxial compression tests. I: Influences
16 De Jong K A. An analysis of the behavior of a class of genetic adap- of density and pressure. J Geotech Geoenviron, 2014, 12: 4014070
tive systems. Dissertation of Doctor Degree. Michigan: University of 32 Xiao Y, Liu H, Chen Y, et al. Strength and deformation of rockfill
Michigan, 1975 material based on large-scale triaxial compression tests. II: Influence
17 Goldberg D E, Holland J H. Genetic algorithms and machine learning. of particle breakage. J Geotech Geoenviron, 2014, 12: 4014071
Mach Learn, 1988, 2: 95–99 33 Xiao Y, Liu H, Chen Y, et al. Bounding surface model for rockfill
18 Goldberg D E. Real-coded genetic algorithms, virtual alphabets, and materials dependent on density and pressure under triaxial stress con-
Blocking. Urbana, 1990, 51: 61801 ditions. J Eng Mech, 2013, 4: 4014002
19 Janikow C Z, Michalewicz Z. An experimental comparison of binary 34 Xiao Y, Liu H, Chen Y, et al. Bounding surface plasticity model incor-
and floating point representations in genetic algorithms. ICGA, 1991, porating the state pressure index for rockfill materials. J Eng Mech,
31–36 2014, 11: 4014087
20 Michalewicz Z, Logan T, Swaminathan S. Evolutionary Operators 35 Duncan J, Chang C. Nonlinear analysis of stress and strain in soils. J
for Continuous Convex Parameter Spaces. River Edge, NJ: World Soil Mech Found, 1970, 5: 1629–1653
Scientific Publishing, 1994. 84–97 36 Li G, Ni Z, Fu H, et al. Experimental studies on rheological behaviors
21 Radcliffe N J. Equivalence class analysis of genetic algorithms. for rockfills in concrete faced rockfill dam. Chin J Rock Soli Mech,
Complex Syst, 1991, 2: 183–205 2004, 11: 1712–1716
22 Ono I, Kita H, Kobayashi S. A real-coded genetic algorithm using the 37 Tomlin A S. The use of global uncertainty methods for the evaluation
unimodal normal distribution crossover. Advances in evolutionary of combustion mechanisms. Reliab Eng Syst Safe, 2006, 10–11:
computing. Springer Berlin Heidelberg, 2003. 213–237 1219–1231