Two-Stepped Evolutionary Algorithm and Its Application

to Stability Analysis of Slopes

C. X. Yang1; L. G. Tham2; X. T. Feng3; Y. J. Wang4; and P. K. K. Lee5

Abstract: Based on genetic algorithm and genetic programming, a new evolutionary algorithm is developed to evolve mathematical
models for predicting the behavior of complex systems. The input variables of the models are the property parameters of the systems,
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

which include the geometry, the deformation, the strength parameters, etc. On the other hand, the output variables are the system
responses, such as displacement, stress, factor of safety, etc. To improve the efficiency of the evolution process, a two-stepped approach
is adopted; the two steps are the structure evolution and parameter optimization steps. In the structure evolution step, a family of model
structures is generated by genetic programming. Each model structure is a polynomial function of the input variables. An interpreter is
then used to construct the mathematical expression for the model through simplification, regularization, and rationalization. Furthermore,
necessary internal model parameters are added to the model structures automatically. For each model structure, a genetic algorithm is then
used to search for the best values of the internal model parameters in the parameter optimization step. The two steps are repeated until the
best model is evolved. The slope stability problem is used to demonstrate that the present method can efficiently generate mathematical
models for predicting the behavior of complex engineering systems.
DOI: 10.1061/共ASCE兲0887-3801共2004兲18:2共145兲
CE Database subject headings: Algorithms; Slope stability; Stability analysis.

Introduction havior of the systems is available. Most of the methods are based
on either statistical or neural network approaches.
The enormous social and economic losses associated with hazards Various statistical and regression analysis techniques can be
such as slope failures and landslides have attracted the attention used to perform systematic data analysis and to establish relations
of both engineers and researchers. Efforts have been made to for predicting the behavior of complex systems. For example, Sah
develop analytical and numerical models for predicting their oc- et al. 共1994兲 applied the maximum likelihood method to estimate
currences and scales so that appropriate actions can be taken to the factor of safety of slopes. Though numerous methods are
minimize the losses. Because of the variability of geomaterial, available 共Weiss and Kulikowski 1991兲, the results depend on the
insufficient information regarding the site, and incomplete under- users who have to make a decision on the mathematical expres-
standing of the failure mechanism, it is almost impossible to sions of the relationships to be considered as candidates in the
model the problems exactly. Assumptions as well as simplifica- search for the best solutions. In addition, these statistical methods
tions are made in the modeling so that the solutions are tractable. are often based on the assumptions that the data are normally
However, the accuracy of the models in predicting the behavior of distributed and the variables are independent. Unfortunately, geo-
the problems depends very much on the assumptions and simpli- technical problems are usually extremely complex, and these cri-
fications 共Hoek and Bray 1981兲. teria are not necessarily satisfied.
As an alternative, data-based modeling methods can be used to A computation intelligence approach 共Mavrovouniotis 1990兲,
develop input-output relationships for the analysis of complex such as the neural network and evolutionary computation tech-
systems if a sufficiently large amount of information on the be- niques, can overcome such limitations. It has been shown math-
ematically that a neural network is capable of learning continuous
1 nonlinear input-output mapping 共Cybenko 1989兲. Feng et al.
College of Resources and Civil Engineering, Northeastern Univ.,
Shenyang 110004, People’s Republic of China. 共1996兲 proved that neural networks can adequately predict the
2
Dept. of Civil Engineering, Univ. of Hong Kong, Pokfulam Rd., stability level of slopes. Though this technique can serve as a
Hong Kong. strong predictive model for finding complex relationships, it has
3 inherent limitations in handling the uncertainties of the model
College of Resources and Civil Engineering, Northeastern Univ.,
Shenyang 110004, People’s Republic of China. inputs and deciding the size as well as topology of the required
4
College of Resources and Civil Engineering, Northeastern Univ., network. Furthermore, such an approach will not produce explicit
Shenyang 110004, People’s Republic of China. relations; therefore, the model obtained gives little insight into the
5
Dept. of Civil Engineering, Univ. of Hong Kong, Pokfulam Rd., underlying mechanisms of the problems.
Hong Kong. The evolutionary computation techniques 共Goldberg 1989;
Note. Discussion open until September 1, 2004. Separate discussions
Baäck 1995兲 provide robust alternatives. Based on the Darwinian
must be submitted for individual papers. To extend the closing date by
one month, a written request must be filed with the ASCE Managing theory of natural selection, they attempt to obtain the best solution
Editor. The manuscript for this paper was submitted for review and pos- by carrying out global optimization. They use suitable coding to
sible publication on July 18, 2002; approved on January 2, 2003. This represent possible solutions for a problem and guide the search by
paper is part of the Journal of Computing in Civil Engineering, Vol. 18, using genetic operators and the principle of ‘‘survival of the fit-
No. 2, April 1, 2004. ©ASCE, ISSN 0887-3801/2004/2-145–153/$18.00. test.’’ Of these algorithms, genetic algorithms 共Holland 1975;

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004 / 145

J. Comput. Civ. Eng. 2004.18:145-153.

Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. Stepped evolution of model structures and model parameters

Goldberg 1989兲 have established themselves as powerful search eling, ecological modeling, pattern recognition, and data mining
and optimization tools in problem solving and function optimiza- 共Sharman et al. 1995; Sen and Stoffa 1996; Shao and Yoshisada
tion 共Goldberg 1989, 1994兲, and they have been widely applied in 1996; McKay et al. 1997; Gray et al. 1998; Babovic 2000; Cao
civil and geotechnical engineering 共Feng et al. 1999, 2000; Javadi et al. 2000; Coello 2000; Yang and Soh 2000; Brameier and Ban-
et al. 1999; Nath et al. 1999; Simoes et al. 1999; Fishwick et al. zhaf 2001; Khu et al. 2001; Ross et al. 2001; Wigham and Reck-
2000; Deng and Lee 2001兲. Goh 共1999, 2000兲 used genetic algo- nagel 2001兲.
rithms to search for critical slip surfaces of slopes. However, ge- When standard genetic programming is used for model iden-
netic algorithms cannot be used to evolve complex models such tification, the models generated may include a number of internal
as equations. Therefore, genetic programming 共Koza 1992兲 has model parameters. These randomly generated parameters can
been developed to overcome this limitation. Genetic program- have significant effects on the performance of a model, because
ming can perform symbolic regression 共Koza 1992兲. It is advan-
their values affect the output of the model directly. As a poten-
tageous in that a priori assumptions and prior knowledge about
tially good model with favorable structure may be removed dur-
the structure and size of the solution are not necessary. Moreover,
ing the search, these internal model parameters have to be opti-
this technique can automatically select the system inputs to yield
a model structure that accurately fits the input-output response of mized in the model identification process. In many cases, one has
the system and provides a descriptive solution. Due to its advan- difficulties in the optimization process because the parameters
tages, genetic programming has been successfully used in the change dynamically with the evolution. Various methods have
study of engineering problems such as in civil engineering, pro- been proposed 共Sharman et al. 1995; McKay et al. 1997; Gray
cess modeling, autocontrol, signal processing, hydrological mod- et al. 1998; Cao et al. 2000兲 to overcome the difficulties by add-

146 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004

J. Comput. Civ. Eng. 2004.18:145-153.

 Fig. 2. Crossover and mutation operations in genetic algorithm
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

ing the internal model parameters explicitly. These are somewhat

artificially and not randomly generated.
A closer examination will easily reveal that evolutionary com-
putation techniques can be divided into two steps, namely, the
structure evolution step and the parameter evolution step. Of
course, one has to design an interpreter to link up the two pro-
cesses. In this paper, a two-stepped evolutionary computation
technique is developed. First, genetic programming is employed
to find and optimize the model structure in the structure evolution
step. In this step, a family of models is generated, but the model
parameters are not considered. In the next step 共the parameter
optimization step兲, the internal model parameters for each model
structure are introduced by an interpreter and such parameters of
each model structure are optimized using a genetic algorithm.
Therefore, each model structure will generate a series of evolu-
tion processes of the genetic algorithm to obtain its optimal pa-
rameters. After the two steps, a new family of models is generated
and ready for the next generation of optimization. The fitness of
each model is assessed by a target function, and the generation
will cease when the predefined terminal conditions are satisfied.

Two-Stepped Evolutionary Algorithm

Information obtained from field observations and numerical mod-

elings of an engineering system can be used to form the database
for generating a model or models so that one can make a reason- Fig. 3. Flow chart for stepped evolutionary algorithm
ably good prediction of its behavior. Mathematically, a model can
be represented as
z⫽ f 共 x 1 ,x 2 ,...x n 兲 (1)
where x 1 , x 2 , and x n ⫽input variables; and z⫽output variable. programming, the set of terminals for symbolic regression prob-
Though the exact relation is not known, one can start by gen- lems will only consist of variables, which are usually the basic
erating an initial population of relations. These relations are ran- property parameters. The internal model parameters are generated
domly generated, and the population does not necessary contain in the next step. The function set and the terminal set constitute
the optimal relation that can provide the best fit of the system the building set of the problems. For example, we can choose the
behavior. Based on various evolutionary algorithms, searches can function and terminal sets to be defined as F⫽ 兵 ⫹,⫺,⫻,/ 其 and
be carried out to find the optimal relation. In this paper, a two- T⫽ 兵 x 1 ,x 2 ,x 3 其 , respectively. Of course, the building set of the
stepped evolutionary algorithm is proposed for such search. The problem is B⫽ 兵 ⫹,⫺,⫻,/;x 1 ,x 2 ,x 3 其 . Combining the function
two steps are and terminal sets, the initial model structures 共the first generation兲
1. The structure evolution step; and can be generated by using the ramped half-and-half approach.
2. The parameter optimization step. Fig. 1 shows the tree-structure representation of four typical mod-
The details of each step are described in the subsequent sections. els:
• Model 1: ⫹x 22 ⫺x 21 ⫹x 1 x 2 /x 3 ,
• Model 2: ⫹x 21 ⫹x 1 x 3 ⫺x 2 /x 1 x 3 ,
Structure Evolution Step
• Model 3: ⫹x 21 ⫹x 2 x 3 ⫺x 2 /x 1 x 3 , and
In this step, model structures are generated by the standard ge- • Model 4: ⫹x 21 ⫹x 23 ⫺x 1 x 2 /x 3 .
netic programming technique. These model structures comprise Subsequent generations can be generated by using genetic op-
functions as well as terminals. The set of functions is composed erators, which include reproduction, crossover, and mutation. Re-
of mathematical and logical operators. Unlike standard genetic production is a direct copy of a parent into the new generation;

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004 / 147

J. Comput. Civ. Eng. 2004.18:145-153.

 Table 1. Samples of Circular Slopes 关from Sah et al. 共1994兲兴
Internal Pore
Unit Cohesion frictional Slope Slope pressure
Case weight strength angle angle height ratio Factor Failed/
number ␥ 共kN/m3兲 C 共kPa兲 ␾ 共deg兲 ␺ 共deg兲 H (m) 共r兲 of safety stable
1 18.68 26.34 15.00 35 8.23 — 1.110 Failed
2a 16.50 11.49 0.00 30 3.66 — 1.000 Failed
3 18.84 14.36 25.00 20 30.50 — 1.875 Stable
4 18.84 57.46 20.00 20 30.50 — 2.045 Stable
5 28.44 29.42 35.00 35 100.00 — 1.780 Stable
6 28.44 39.23 38.00 35 100.00 — 1.990 Stable
7 20.60 16.28 26.50 30 40.00 — 1.250 Failed
8 14.80 0.00 17.00 20 50.00 — 1.130 Failed
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

9 14.00 11.97 26.00 30 88.00 — 1.020 Failed

10 25.00 120.00 45.00 53 120.00 — 1.300 Stable
11a 26.00 150.05 45.00 50 200.00 — 1.200 Stable
12 18.50 25.00 0.00 30 6.00 — 1.090 Failed
13 18.50 12.00 0.00 30 6.00 — 0.780 Failed
14 22.40 10.00 35.00 30 10.00 — 2.000 Stable
15 21.40 10.00 30.34 30 20.00 — 1.700 Stable
16a 22.00 20.00 36.00 45 50.00 — 1.020 Failed
17 22.00 0.00 36.00 45 50.00 — 0.890 Failed
18 12.00 0.00 30.00 35 4.00 — 1.460 Stable
19 12.00 0.00 30.00 45 8.00 — 0.800 Failed
20 12.00 0.00 30.00 35 4.00 — 1.440 Stable
21 12.00 0.00 30.00 45 8.00 — 0.860 Failed
22 23.47 0.00 32.00 37 214.00 — 1.080 Stable
23 16.00 70.00 20.00 40 115.00 — 1.110 Failed
24 20.41 24.90 13.00 22 10.67 0.350 1.400 Stable
25a 19.63 11.97 20.00 22 12.19 0.405 1.350 Failed
26 21.82 8.62 32.00 28 12.80 0.490 1.030 Failed
27 20.41 33.52 11.00 16 45.72 0.200 1.280 Failed
28 18.84 15.32 30.00 25 10.67 0.380 1.630 Stable
29a 18.84 0.00 20.00 20 7.62 0.450 1.050 Failed
30 21.43 0.00 20.00 20 61.00 0.500 1.030 Failed
31 19.06 11.71 28.00 35 21.00 0.110 1.090 Failed
32 18.84 14.36 25.00 20 30.50 0.450 1.110 Failed
33 21.51 6.94 30.00 31 76.81 0.380 1.010 Failed
34 14.00 11.97 26.00 30 88.00 0.450 0.625 Failed
35 18.00 24.00 30.15 45 20.00 0.120 1.120 Failed
36 23.00 0.00 20.00 20 100.00 0.300 1.200 Failed
37 22.40 100.00 45.00 45 15.00 0.250 1.800 Stable
38 22.40 10.00 35.00 45 10.00 0.400 0.900 Failed
39 20.00 20.00 36.00 45 50.00 0.250 0.960 Failed
40 20.00 20.00 36.00 45 50.00 0.500 0.830 Failed
41 20.00 0.00 36.00 45 50.00 0.250 0.790 Failed
42 20.00 0.00 36.00 45 50.00 0.500 0.670 Failed
43 22.00 0.00 40.00 33 8.00 0.350 1.450 Stable
44a 24.00 0.00 40.00 33 8.00 0.300 1.580 Stable
45 20.00 0.00 24.50 20 8.00 0.350 1.370 Stable
46 18.00 5.00 30.00 20 8.00 0.300 2.050 Stable
a
Selected as testing cases.

therefore, no new model structure is produced. This operator is There are two mutation operators, namely, the shrink and re-
necessary, as it helps to preserve ‘‘good’’ model structures. place mutations. In shrink mutation, a subtree is deleted. By de-
On the other hand, crossover produces new model structures leting ‘‘⫹x 2 x 3 ’’ of Model 3, Model 7 is generated:
by swapping two randomly selected subtrees of two parents. For • Model 7: ⫹x 21 ⫺x 2 /x 1 x 3 .
example, by swapping the trees 共‘‘⫹x 1 x 2 /x 3 ’’ and ‘‘⫹x 1 x 3 ’’兲 be- On the other hand, replace mutation is conducted by replacing
tween Models 1 and 2, we can generate two new models 共Models a randomly selected node with a randomly selected node in the
5 and 6兲, namely building set. For example, Model 8 is created by replacing the ‘‘/’’
• Model 5: ⫹x 22 ⫺x 21 ⫹x 1 x 3 and of Model 4 with ‘‘⫻’’:
• Model 6: ⫹x 21 ⫹x 2 x 1 /x 3 ⫺x 2 /x 1 x 3 . • Model 8: ⫹x 21 ⫹x 23 ⫺x 2 x 3 /x 1 .

148 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004

J. Comput. Civ. Eng. 2004.18:145-153.

 Table 2. Control Parameters
Object Parameter
Structure evolution
Population size 2,000
Allowable maximum 100
number of generations
Reproduction probability 0.01
Crossover probability 0.85
Mutation probability 0.005
Selection method Tournament selection with tournament
size of 5 and the two best ones are
selected to reproduce
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

Termination criterion Reach the maximum number of

generations or any improvement in
the overall fitness is likely to be minimal.
The fittest structure individual is taken
as the optimum model.
Allowable maximum 10
tree depth
Parameter optimization
Population size 200
Reproduction probability 0.05
Crossover probability 0.95
Mutation probability 0.03
Selection method Roulette wheel selection 共proportionate
selection兲
Termination criterion The fitness value of the best individual
gas remained unchanged for three
generations

With the model structures generated, the next step is to introduce

and optimize the internal model parameters.

Parameter Optimization Step

For every model structure, an interpreter is used to construct
mathematical expressions that relate the input-output variables, Fig. 4. Variation of fitness values with generations; 共a兲 initial gen-
by simplifying, regularizing, and rationalizing the tree-structures eration; 共b兲 20th generation; 共c兲 40th generation; 共d兲 60th generation;
as well as introducing necessary internal model parameters. One 共e兲 80th generation
can easily note that these internal model parameters are the coef-
ficients of each term of the mathematical expressions. As these
coefficients are not yet known, one has to search for the optimal secondary ones. As we have two steps, two different sets of con-
coefficients by employing various means, and the present study trol parameters have to be selected so that the search can be
uses the genetic algorithm. In a typical genetic algorithm 共Fig. 2兲, carried out efficiently.
individuals are either represented by strings of binary digits or
with symbols. Each individual is represented by a binary string in
the present study. Genetic operators will then work on the bits of Fitness Test
the strings to find the best permutation and combination of the All generated relations have to be tested against field or numerical
bits of a string that defines the characteristic of the individual results. The results are divided into the fitness and testing groups.
solution. As in genetic programming, the operators include repro- The best relation should give the minimum error between the
duction, crossover, and mutation 共Fig. 2兲. predicted and field or numerical results; that is

Control Parameters
Parameters have to be prechosen to control the search process,
E⫽
1
n
冑兺 n

1
共 z est⫺z target兲 2 (2)

and they include the population size for each generation, the num- where E⫽fitness; z est and z target⫽estimated and target value of the
ber of generations, the probabilities of genetic operators 共repro- output variable; and n⫽total number of cases in the fitness group.
duction, crossover, and mutation兲, the selection method, the ter- As the two-stepped approach is adopted in the proposed
mination conditions, etc. The population size and number of method, the optimal set of internal model parameters for each
generations are the primary parameters, whereas the others are the model structure is determined first. Comparing the various model

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004 / 149

J. Comput. Civ. Eng. 2004.18:145-153.

 Table 3. Comparison of Factor of Safety
Factor of safety
Limit Maximum
Case equilibrium likelihood
number method estimationa Present
1 1.110 1.269 1.191
2 1.875 2.143 1.978
3 2.045 2.208 2.090
4 1.780 1.584 1.490
5 1.990 1.777 1.654
6 1.250 1.423 1.353
7 1.130 1.296 1.243
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

8 1.020 1.348 1.268

9 1.300 1.277 1.236
10 1.090 1.023 1.042
11 0.780 0.491 0.704
12 2.000 2.075 2.004
13 1.700 1.671 1.597
14 0.890 1.121 1.130
15 1.460 1.273 1.564
16 0.800 0.891 1.079
Fig. 5. Variation of best and average fitness with generation 17 1.440 1.273 1.564
18 0.860 0.891 1.079
19 1.080 1.280 1.227
structures, one can identify the model structure and corresponding 20 1.110 0.804 0.779
internal model parameters that give the best prediction of the 21 1.400 1.267 1.406
behavior of the system. 22 1.030 1.074 1.163
23 1.280 1.282 1.245
24 1.630 1.594 1.586
Implementation 25 1.030 0.772 0.816
26 1.090 1.159 1.156
Fig. 3 shows the flow chart for the proposed two-stepped algo- 27 1.110 1.254 1.258
rithm, and the pertinent points are summarized as follows: 28 1.010 0.938 0.957
1. Create a population of model structures from the building set 29 0.625 0.761 0.793
using the genetic programming technique. 30 1.120 1.003 0.984
2. Estimate the fitness of each model structure generated 31 1.200 1.080 1.057
through the following substeps: 32 1.800 2.114 1.773
a. Determine the internal model parameters of the current 33 0.900 0.792 0.935
model structure. 34 0.960 0.905 0.934
b. Generate a population of internal model parameters 35 0.830 0.625 0.707
using the genetic algorithm technique. 36 0.790 0.841 0.903
c. Use Eq. 共2兲 to estimate the fitness of each set of internal 37 0.670 0.561 0.676
model parameters for the current model structure. 38 1.450 1.296 1.430
d. If the termination conditions of parameter evolution are 39 1.370 1.256 1.366
satisfied, obtain the fitness of the best internal model 40 2.050 1.945 1.954
parameters as the fitness of the current model structure
41 1.000 0.864 1.031
and then go to Step f. Otherwise, go to Step e.
42 1.200 1.381 1.327
e. Perform genetic operations on the parameter strings to
43 1.020 1.180 1.161
generate a new population of internal model param-
44 1.350 1.131 1.226
eters. Then go to Step c.
f. If the fitness of each model structure is calculated, go to 45 1.050 0.849 1.040
Otherwise, go to Step a. 46 1.580 1.396 1.511
3. If the termination conditions are satisfied, output the best a
Sah et al. 共1994兲.
models and terminate the algorithm. Otherwise, go to Step 4.
4. Perform genetic operations on the structure trees to generate sary to perform multiple runs and obtain the best model which has
a new population of model structures. Then go to Step 2. the best permutation and combination of the major contributors.
It should be noted that the algorithm is based on a stochastic
search. Therefore, it is a probabilistic approach by nature and may Application to Stability Analysis of Circular Failure
arrive at different optimal solutions for different runs. Neverthe- Slopes
less, most of the solutions will contain some common terms.
Moreover, these common terms will often be the major contribu- The safety of the slope is expressed in terms of the factor of
tors. To eliminate the effect of this inherent variation, it is neces- safety, and the limit equilibrium approach has been the most

150 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004

J. Comput. Civ. Eng. 2004.18:145-153.

 F⫽ 兵 ⫹,⫺,⫻,/,sin,cos,tan,cot,sec,csc其 (5)
and the terminal set is
T⫽ 兵 ␥,c,␾,␺,H,r 其 (6)

One can then search for factor of safety, F̂, in terms of the build-
ing set; that is

F̂⫽g 共 ⫹,⫺,⫻,/,sin,cos,tan,cot,sec,csc兩 ␥,c,␾,␺,H,r 兲 (7)

Implementation

Twenty-three dry 共13 failed and 10 stable兲 and 23 wet 共16 failed
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

and seven stable兲 cases of circular failure were collected from Sah
et al. 共1994兲. These data are presented in Table 1. Six cases,
which are indicated by ‘‘a’’ in the table, are randomly selected to
serve as the testing cases. The others are used as the fitness cases.
The fitness function is defined as the error between the factors
of safety obtained by limit equilibrium analysis and those calcu-
lated by the present method. Mathematically, it can be expressed
as

再 冑兺 n
1
Fig. 6. Correlation between factors of safety obtained by limit 共 F LEM⫺F̂ 兲 2
fitness⫽ n i⫽1 (8)
equilibrium and present approaches
M
where F LEM⫽safety factor obtained by the limit equilibrium ap-
popular one for computing such a factor 共Hoek and Bray 1981兲. It proach; n⫽total number of fitness cases; and M⫽prescribed rela-
makes use of the plastic limit theorem of solid mechanics to ana- tively large value, and it is necessary to allow for overflow re-
lyze the stability of any mass of soil or rock under the assumption sults. Based on the parameter settings given in Table 2, we can
that mass will fail along a potential slip surface. Although differ- perform the genetic search to obtain the best expression for com-
ent assumptions have been made in limit equilibrium analyses, puting the factor of safety.
they give similar results 共Nash 1987兲. A vast database of factors
of safety for slopes can be assimilated from either open published
literature or reports. Therefore, one can make use of such infor- Results
mation and apply the present method to evolve a mathematical
expression for obtaining the factor of safety of a slope without For the fitness cases shown in Table 1, 10 runs were performed
actually carrying out limit equilibrium analyses. with different randomly generated initial models so that each run
would take a different genetic path to evolve models. Fig. 4
shows the change of fitness values with evolution of a typical run.
Problem Statement Each point represents the fitness value of a model structure with
In this section, the proposed method is applied to slope stability optimized parameters. For clarity, only models with a fitness in
analysis, and an optimal mathematical expression for computing the range of 关0.12,0.6兴 are plotted. As the model structures have
the safety factor of slopes will be generated. Though the exact been generated randomly, the distribution of the fitness values in
relation is not yet known, it is believed that there must be some the initial generation varies significantly, and only a small number
mapping relationships between the safety factor and the geotech- of models have fitness within the chosen range 关Fig. 4共a兲兴.
nical and geometry properties of the slope. The factor of safety is As evolution proceeds, more and more models have a fitness
the output, whereas these geotechnical and geometrical properties within the range 共see results of the 20th, 40th, 60th, and 80th
are the input variables. Based on the knowledge of the behavior generations兲, and the fitness values converge. This indicated that
of slopes, the factor of safety can be expressed as more and more models with better fitness were generated and the
models with poor fitness were gradually eliminated through selec-
F⫽ f 共 ␥,c,␾,␺,H,r 兲 (3) tion and contest.
where F⫽factor of safety; ␥⫽unit weight; c⫽cohesive strength; The best fitness value and the average fitness value for all
␾⫽internal friction angle; ␺⫽slope angle; H⫽slope height; and models 共Fig. 5兲 show that the performance of the models improve
r⫽pore pressure ratio. gradually. Moreover, we can note that the changes are more sig-
These property variables constitute the terminal set. As the nificant in the early generations and become more gradual in the
terminal set includes angle variables, it is also necessary, in addi- latter generations. From the 40th generation to the 80th genera-
tion to the basic arithmetical functions 共i.e., ⫹, ⫺, ⫻, /兲, to in- tion, the process almost reaches the ‘‘steady state’’; that is, the
clude trigonometric functions 共sin, cos, tan, cot, sec, and cosec兲. average fitness oscillates slightly and the best fitness almost re-
Thus, the building set of the problem can be written as mains unchanged. Therefore, one can terminate the evolution at
the 80th generation.
B⫽ 兵 ⫹,⫺,⫻,/,sin,cos,tan,cot,sec,csc;␥,c,␾,␺,H,r 其 (4)
Translating the ‘‘best’’ tree-structure into a mathematical ex-
where the function set is pression, one can show that the factor of safety is given by

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004 / 151

J. Comput. Civ. Eng. 2004.18:145-153.

 1 Cao, H. Q., Kang, L. S., and Chen, Y. P. 共2000兲. ‘‘Evolutionary modeling
F̂⫽ 共 1.406⫹0.039c csc2 ␾ 兲 ⫹1.25共 1⫺r 兲 tan ␾ cot ␺ of systems of ordinary differential equations with genetic program-
H ming.’’ Genet. Program. Evol. Mach., 1, 309–373.
Coello, C. A. C. 共2000兲. ‘‘Constraint-handling using an evolutionary mul-
1
⫹0.156 sec ␾⫹0.011 csc ␺ (9) tiobjective optimization technique.’’ Civ. Eng. Environ. Syst., 17共4兲,
␥ 319–346.
Cybenko, G. 共1989兲. ‘‘Approximation by superposition of sigmoidal
functions.’’ Math. Control, Signals, Syst., 2, 303–314.
Validation Deng, J. H., and Lee, C. F. 共2001兲. ‘‘Displacement back analysis for a
Six cases, which are indicated by ‘‘a’’ in Table 1, are chosen to steep slope at the Three Gorges Project site.’’ Int. J. Rock Mech. Min.
validate the present approach. The factors of safety for the fitness Sci., 38共2兲, 259–268.
cases and testing cases are presented in Table 3. Fig. 6 shows the Feng, X. T., Katsuyama, K., and Wang, Y. J. 共1996兲. ‘‘Self-learning based
and adaptive modeling on estimation of rock slope stability.’’ J. Min-
linear regression result between the factors of safety. It is noted
ing Mater., Process. Inst. Jpn., 112, 851– 858.
that the intercept of the straight line is almost zero, and its slope Feng, X. T., Zhang, Z. Q., and Sheng, Q. 共2000兲. ‘‘Estimating mechanical
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

is close to one. The results show that the prediction is fairly rock mass parameters relating to the Three Gorges Project permanent
accurate. As a comparison, the results obtained from maximum shiplock using an intelligent displacement back analysis method.’’ Int.
likelihood estimation are also listed in Table 3. One can note that J. Rock Mech. Min. Sci., 37共7兲, 1039–1054.
the present method can attain a better approximation. Feng, X. T., Zhang, Z. Q., and Xu, P. 共1999兲. ‘‘Adaptive and intelligent
prediction of deformation time series of high rock excavation slope.’’
Trans. Nonferrous Met. Soc. China, 9共4兲, 842– 846.
Conclusions Fishwick, R. J., Liu, X. L., and Begg, D. W. 共2000兲. ‘‘Adaptive search in
discrete limit analysis problems.’’ Comput. Methods Appl. Mech.
Eng., 189共3兲, 931–942.
For highly nonlinear and extremely complex engineering prob-
Goh, A. T. C. 共1999兲. ‘‘Genetic algorithm search for critical slip surface
lems, a two-stepped evolutionary algorithm by combining genetic
in multiple-wedge stability analysis.’’ Can. Geotech. J., 36共2兲, 382–
programming and genetic algorithms has been developed to 391.
search for the best model structure and its internal model param- Goh, A. T. C. 共2000兲. ‘‘Search for critical slip circle using genetic algo-
eters. This method has improved the search process by simplify- rithms.’’ Civ. Eng. Environ. Syst., 17共3兲, 181–211.
ing the complex problem of evolving model structure and its in- Goldberg, D. E. 共1989兲. Genetic algorithms in search, optimization, and
ternal model parameters into two steps, that is, evolution of the machine learning, Addison-Wesley, Reading, Mass.
model structure and optimization of the internal model param- Goldberg, D. E. 共1994兲. ‘‘Genetic and evolutionary algorithms come of
eters. Using this method, an empirical equation was established age.’’ Commun. ACM, 37共3兲, 113–119.
for estimating the factor of safety of slopes from the input data Gray, G. J., Murray-Smith, D. J., Li, Y., Sharman, K. C., and Weinbren-
共geotechnical properties and the geometries of slopes兲. The results ner, T. 共1998兲. ‘‘Nonlinear model structure identification using genetic
show that there is good agreement between the safety factors programming.’’ Control Eng. Pract., 6, 1341–1352.
Hoek, E., and Bray, J. 共1981兲. Rock slope engineering, Institution of
obtained by the present method and the limit equilibrium method.
Mining and Metallurgy, London.
The present method automatically constructed models directly Holland, J. H. 共1975兲. Adaptation in natural and artificial system, Uni-
from the fitness case sets. Though it is an empirical approach, the versity of Michigan, Ann Arbor, Mich.
functions obtained can still give insight into the physical nature of Javadi, A. A., Farmani, R., Toropov, V. V., and Snee, C. P. M. 共1999兲.
complex systems. The present method can offer a useful alterna- ‘‘Identification of parameters for air permeability of shotcrete tunnel
tive to other data-based modeling methodologies and can perform lining using a genetic algorithm.’’ Comput. Geotech., 25共1兲, 1–24.
model input selection, model structure selection, and parameter Khu, S. T., Liong, S. Y., Babovic, V., Madsen, H., and Muttil, N. 共2001兲.
identification. It can be widely used in engineering analysis as a ‘‘Genetic programming and its application in real-time runoff fore-
useful tool for predicting the behavior of engineering systems. casting.’’ J. Am. Water Resour. Assoc., 37共2兲, 439– 451.
Koza, J. 共1992兲. Genetic programming: On the programming of comput-
ers by natural selection, MIT Press, Cambridge, Mass.
Mavrovouniotis, M. L. 共1990兲. Artificial intelligence in process engineer-
Acknowledgments ing, Academic, New York.
McKay, B., Willis, M., and Barton, G. 共1997兲. ‘‘Steady-state modelling of
This research is supported by the Fok Ying Tung Education Foun- chemical process systems using genetic programming.’’ Comput.
dation 共Grant No. 71408兲, the Teaching and Research Award Pro- Chem. Eng., 21共9兲, 981–996.
gram for Outstanding Young Teachers in Higher Education Insti- Nash, D. 共1987兲. ‘‘A comparative review of limit equilibrium methods of
tiutions of Ministry of Education, China, and the Research Grants stability analysis.’’ Slope stability, M. G. Anderson and K. S. Rich-
Council of HKSAR. The writers would like to thank Professor ards, eds., Wiley, New York.
Tingjie Li for his generous support. Nath, S. K., Chakraborty, S., Singh, S. K., and Ganguly, N. 共1999兲. ‘‘Ve-
locity inversion in cross-hole seismic tomography by counter-
propagation neural network, genetic algorithm, and evolutionary pro-
gramming techniques.’’ Geophys. J. Int., 138共1兲, 108 –124.
References Ross, B. J., Fueten, F., and Yashkir, D. Y. 共2001兲. ‘‘Automatic mineral
identification using genetic programming.’’ Mach. Vision Appl., 13共2兲,
Baäck, T. 共1995兲. Evolutionary algorithms in theory and practice, Oxford 61– 69.
University Press, New York. Sah, N. K., Sheorey, P. R., and Upadhyaya, L. N. 共1994兲. ‘‘Maximum
Babovic, V. 共2000兲. ‘‘Data mining and knowledge discovery in sediment likelihood estimation of slope stability.’’ Int. J. Rock Mech. Min. Sci.,
transport.’’ Comput. Aided Civ. Infrastruct. Eng., 15共5兲, 383–389. 31共1兲, 47–53.
Brameier, M., and Banzhaf, W. 共2001兲. ‘‘A comparison of linear genetic Sen, M. K., and Stoffa, P. L. 共1996兲. ‘‘Bayesian inference, Gibbs’ sam-
programming and neural networks in medical data mining.’’ IEEE pler, and uncertainty estimation in geophysical inversion.’’ Geophys.
Trans. Evol. Comput., 5共1兲, 17–26. Prospect., 44共2兲, 313–350.

152 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004

J. Comput. Civ. Eng. 2004.18:145-153.

 Shao, S. W., and Yoshisada, M. 共1996兲. ‘‘Approach to failure mode analy- efficients in fractured media.’’ Geophysics, 64共4兲, 1143–1152.
sis of large structures.’’ Proc., 7th Specialty Conf. on Probabilistic Weiss, S. M., and Kulikowski, C. A. 共1991兲. Computer systems that learn,
Mechanics and Structural Reliability, D. M. Frangopol and M. D. Morgan Kaufmann, San Mateo, Calif.
Grigoriu, eds., ASCE, New York, 704 –707. Wigham, P. A., and Recknagel, F. 共2001兲. ‘‘An inductive approach to
Sharman, K. C., Alcázar, A. I. E., and Li, Y. 共1995兲. ‘‘Evolving signal- ecological time series modelling by evolutionary computation.’’ Ecol.
processing algorithms by genetic programming.’’ Genet. Algor. Eng. Modell., 146共1–3兲, 275–287.
Syst.: Innov. Appl., 12–14, 473– 480. Yang, Y. W., and Soh, C. K. 共2000兲. ‘‘Fuzzy logic integrated genetic
Simoes, I. A., Neves, F. A., Tinen, J. S., Protazio, J. S., and Costa, J. C. programming for optimization and design.’’ J. Comput. Civ. Eng.,
共1999兲. ‘‘Multiazimuthal modeling and inversion of qP reflection co- 14共4兲, 249–254.
Downloaded from ascelibrary.org by Northeastern University on 02/21/13. Copyright ASCE. For personal use only; all rights reserved.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / APRIL 2004 / 153

View publication stats J. Comput. Civ. Eng. 2004.18:145-153.