Professional Documents
Culture Documents
Genetic Programming For Predicting Bridge Scour
Genetic Programming For Predicting Bridge Scour
Abstract: Bridge-pier scour is a significant problem for the safety of bridges. Extensive laboratory and field studies have been conducted
examining the effect of relevant variables. This note presents an alternative to the conventional regression-based equations 共HEC-18 and
regression equation developed by the writers兲, in the form of artificial neural networks 共ANNs兲 and genetic programming 共GP兲. There had
been 398 data sets of field measurements that were collected from published literature and were used to train the network or evolve the
program. The developed network and evolved programs were validated by using the observations that were not involved in the training.
The performance of GP was found more effective when compared to regression equations and ANNs in predicting the scour depth at
bridge piers.
DOI: 10.1061/共ASCE兲HY.1943-7900.0000133
CE Database subject headings: Bridges; Piers; Computer programming; Neural networks; Scour.
Author keywords: Bridge pier; Genetic programming; Artificial neural networks; Local scour; Radial basis function.
Introduction comparison of some bridge pier scour formulae using field and
laboratory data.
The presence of a bridge pier in a channel may cause an abrupt Alternative fitting approaches such as artificial neural net-
change in the approach flow velocity, which may result in local works 共ANNs兲 共Lee et al. 2007兲 and adaptive neurofuzzy infer-
scour at piers. The failure of Black Mount bridge in New Zealand ence system 共Bateni et al. 2007兲 have been recently shown to
yield effective estimates of ds. ANNs have been reported to pro-
was a result of undermining of its piers in a riverbed 共Melville
vide reasonably good solutions for hydraulic engineering prob-
and Coleman 2000兲. The mechanism of flow around a pier struc-
lems, in cases of highly nonlinear and complex relationship
ture is so complicated that it is difficult to establish a general
among the input-output pairs in corresponding data 共Azmathullah
empirical model to predict the scour depth ds. A reliable estima- et al. 2005, 2008兲.
tion of ds is of paramount importance in safe, economic, and The objective of this study is to predict ds using genetic pro-
technically sound bridge pier design. Most prediction formulas gramming 共GP兲. The performance of the proposed GP model is
for ds available in the literature have been developed using con- compared with a standard radial basis function 共RBF兲 neural net-
ventional regression methods. Johnson 共1995兲 reported that the work and conventional regression-based equations. The explicit
Melville and Sutherland formula tends to overpredict ds to greater formulation of the GP model is also presented.
extent than any other formulae. Recently, Mohamed et al. 共2005兲
showed that the Laursen and Toch 共1956兲 and the Colorado State
University 共CSU兲 formulas 共Mohamed et al. 2005兲 give reason- Local Scour Problem around a Pier
able estimates, while the Melville and Sutherland 共1988兲 and Jain
and Fischer 共1980兲 formulas overpredict pier scour based on the The equilibrium scour depth ds around a circular pier in a steady
flow over a bed of uniform, spherical, and noncohesive sediment
1
Senior Lecturer, River Engineering and Urban Drainage Research depends on numerous groups of variables characterizing flow,
Centre 共REDAC兲, Universiti Sains Malaysia, Engineering Campus, Ni- sediment characteristics, and pier geometry. Extensive field data
bong Tebal, Pulau Pinang 14300, Malaysia 共corresponding author兲. sets were obtained for the model developments in this study from
E-mail: redacazamath@eng.usm.my Landers and Mueller 共1999兲 共384 sets of data兲 and Mohamed et
2
Professor and Deputy Director, REDAC, Universiti Sains Malaysia,
al. 共2005兲 共14 sets of data兲. Table 1 summarizes the ranges of field
Engineering Campus, Nibong Tebal, Pulau Pinang 14300, Malaysia.
E-mail: redac02@eng.usm.my data available such as pier width 共b兲, length of pier 共L兲, flow
3
Professor and Director, REDAC, Universiti Sains Malaysia, Engi- velocity 共V兲, flow depth 共y兲, mean diameter of bed material 共d50兲,
neering Campus, Nibong Tebal, Pulau Pinang 14300, Malaysia. E-mail: and standard deviation bed grain size 共兲.
redac01@eng.usm.my
4
Assistant Professor, Dept. of Civil Engineering, Univ. of Gaziantep,
Gaziantep 27310, Turkey, E-mail: aguven@gantep.edu.tr
Neural Network Model
Note. This manuscript was submitted on June 6, 2008; approved on
July 7, 2009; published online on July 13, 2009. Discussion period open
until August 1, 2010; separate discussions must be submitted for indi- ANNs provide a random mapping between an input and an output
vidual papers. This technical note is part of the Journal of Hydraulic vector, typically consisting of three layers of neurons, namely,
Engineering, Vol. 136, No. 3, March 1, 2010. ©ASCE, ISSN 0733-9429/ input, hidden, and output, with each neuron acting as an indepen-
2010/3-165–169/$25.00. dent computational element. Neural networks derive their
strengths from the high degree of freedom associated with their Banzhaf 2001兲. An operand or an operator of an instruction is
architecture. Prior to application, the network is trained to observe changed by mutation into another symbol over the same set.
data sets. This feeds the network with input and output pairs and The fitness of a GP individual may be computed by using the
determines the values of connection weights, bias, or centers. The equation
training may require many epochs 共presentation of complete data
sets once to the network兲, being carried out until the training sum N
of squares error reaches a specified error goal. Concepts involved f= 兺 共兩X j − Y j兩兲 共1兲
behind these training schemes are outlined in the ASCE Task j=1
Committee 共2000兲. A neural network toolbox contained within the where X j = value returned by a chromosome for the fitness case j
MATLAB package was used in this study. The usual feed-forward and Y j = expected value for the fitness case j.
type of network was trained using RBFs. Out of the total of 398 In GP, the maximum size of the program is usually restricted
input-output pairs, about 75% 共300 sets兲, were selected randomly to avoid overgrowing programs without bounds 共Brameier and
and were used for training, whereas the remaining 25% 共98 sets兲 Banzhaf 2001兲. This configuration has been tested for the pro-
were employed for testing. As dictated by the use of Gaussian posed GP model and has been found sufficient. The best indi-
function, all patterns were normalized within the range of 共0.0, vidual 共program兲 of a trained GP can be converted into a
1.0兲 before their use. The RBF network 共five inputs, 36 hidden functional representation by successive replacements of variables
neurons, and one output兲 was trained by using various values of starting with the last effective instruction 共Oltean and Groşan
spread 共␣兲 between 0 and 1. The value of 0.01 was selected as it 2003兲.
yielded the best performance for the training data. To date, the application of GP in hydraulic engineering has
been limited. Davidson et al. 共1999兲 and Babovic and Keijzer
共2000兲 determined empirical relationships for the friction in tur-
GP bulent pipe flow and the additional resistance to flow induced by
flexible vegetation, respectively. Keijzer and Babovic 共2002兲 de-
GP, a branch of the genetic algorithm 共Holland 1975兲, is a method rived empirical equations using real-world hydraulic data, Gius-
for learning the most “fit” computer programs by means of arti- tolisi 共2004兲 determined Chezy resistance coefficient in
ficial evolution 共Johari et al. 2006兲. GP initializes a population corrugated channels, Kizhisseri et al. 共2005兲 explored a better
consisting of the random members known as chromosomes 共indi- correlation between the temporal pattern of flow field and sedi-
vidual兲, and the fitness of each chromosome is evaluated with ment transport by using numerical model results and field data,
respect to a target value. The principle of Darwinian natural se- and Guven and Gunal 共2008a兲 predicted local scour downstream
lection is used to select and reproduce “fitter” programs. GP cre- of grade-control structures.
ates equal or unequal length computer programs that consist of
variables 共terminal兲 and several mathematical operators 共function兲
sets as the solution. The function set of the system can be com- GP Modeling of Pier Scour
posed of arithmetic operations 共+ , − , / , ⴱ兲 and function calls 共such
as 兵ex , x , sin, cos, tan, log, sqrt, ln, power其兲. Each function implic- The equilibrium local scour depth 共ds兲 around bridge piers is in-
itly includes an assignment to a variable, which facilitates the use fluenced by the variables characterizing the flow, bed sediment,
of multiple program outputs in GP, whereas in tree-based GP and pier geometry. The following relationship describes the equi-
those side effects need to be incorporated explicitly 共Brameier librium scour depth as a function of its independent parameters:
and Banzhaf 2001兲.
The present GP utilizes a two-point string crossover. A seg-
ds = f共V,y,d50,,b,L,g兲 共2兲
ment of random position and random length is selected in both
parents and exchanged between them. If one of the resulting chil- where g = acceleration due to gravity.
dren would exceed the maximum length, crossover is abandoned From previous experience 共Azmathullah et al. 2005; Guven
and restarted by exchanging equalized segments 共Brameier and and Gunal 2008b兲, grouped nondimensional variables produced
␦=
兺 兩Y − X兩 100 共9兲
good results. The following relationship describes scour depth 兺X
normalized with flow depth 共ds / y兲 in terms of dimensionless pa-
rameters where x = 共X − X̄兲; y = 共Y − Ȳ兲; X = observed values; X̄ = mean of X;
ds/y = f共Fr,b/y,d50/y,L/y,兲 共3兲 Y = predicted value; Ȳ = mean of Y; and n = number of samples.
First, an attempt was made to assess the significance or influ-
During model developments in this study, the normalized scour ence of each input parameter on ds / y. Table 3 compares the GP
depth 共ds / y兲 is selected as output and the five dimensionless pa- models, with one of the independent parameters removed in each
rameters in Eq. 共3兲 as inputs. case, and deleting any independent parameter from the input set
In this study, four basic arithmetic operators 共+ , − , ⴱ , /兲 and yielded larger RMSE and lower R2 values. These five independent
some basic mathematical functions 共冑, x2, and power兲 were used. parameters have non-negligible influence on ds / y and so the func-
A large number of generations 共5,000兲 were tested. First, the tional relationship given in Eq. 共3兲 is used for GP modeling in this
maximum size of each program was specified as 256, starting study.
with 64 instructions for the initial program. The functional set and The GP approach resulted in highly nonlinear relationship be-
operational parameters used in GP modeling during this study are tween ds / y and the input parameters with high accuracy and rela-
listed in Table 2. tively low error. The testing performance of the proposed GP
The simplified analytic form of the proposed GP model may model revealed a high generalization capacity with R2 = 0.824 and
be expressed as
冋再冉冋 册 冊 冎
RMSE= 0.046 and AE= −13.54%.
共1 − T2兲2共L/y 兲0.5 2 2 2
ds/y = 共d50/y 兲−0.5 − 2.36 + T2 + Fr −1
共d50/y 兲2
册
Regression Model Results
0.5
− b/y − 1 共4兲
A new nonlinear regression equation was derived by using the
same data sets used earlier for training the neural network and GP
where
modeling. The least-squares fit to those observations gave the
冤 冥
共Fr − 0.224兲 2
following expression:
+ Fr − b/y − 0.739
冉 冊 冉冊 冉冊
L/y 0.042 −0.28 −0.37
T=2 + 共Fr − 0.224兲 共5兲 ds d50 b L
d50/y = 1.82 Fr0.42−0.031 59 共10兲
y y y y
Eq. 共10兲 was compared to the pier scour equation recommended
by Federal Highway Administration’s Hydraulic Engineering Cir-
Training and Testing Results of GP Modeling cular 共HEC-18兲 共Mohamed et al. 2005兲
The performance of GP in training and testing sets is validated in
terms of the common statistical measures R2 共coefficient of deter-
mination兲, root-mean-square error 共RMSE兲, average error 共AE兲,
ds
y
= 2.1
b
y
冉冊 0.65
Fr0.43 共11兲
and average absolute deviation ␦ which is intended to yield conservative estimates. Eq. 共10兲 was
冉 冊
derived from laboratory data by researchers at CSU 共Mohamed et
兺 xy
2
al. 2005兲.
2
共6兲
冑兺 x 2 兺 y 2
R =
冋 册
Analysis and Results
兺 共X − Y兲2
1/2
RMSE = 共7兲 Fig. 1 shows results of predictions of ds / y for the test data that
n
were not involved in the calibration of the models for GP, RBF,
and HEC-18-based equation, 关Eq. 共11兲兴 as well as the regression
X−Y
兺 X
100 equation developed by the writers 关Eq. 共10兲兴. Fig. 1 shows the
performance of the GP compared to RBF, HEC-18, and regression
AE = 共8兲
n equation. An excellent prediction, made by the GP, can be ob-
Table 4. Comparison of Predicted and Observed Scour Depths for the Testing Set
AE
Method Coefficient of determination R2 RMSE 共%兲 Average absolute deviation ␦
Predictions to observed ds / y for testing set
GP 0.819 0.048 ⫺13.660 26.262
RBF 0.691 0.105 ⫺59.450 43.513
Eq. 共10兲 共writers兲 0.539 0.106 ⫺17.702 38.282
Eq. 共11兲 共HEC-18兲 0.425 0.271 ⫺174.055 113.497
d50 ⫽ bed material mean diameter 共m兲; Johari, A., Habibagahi, G., and Ghahramani, A. 共2006兲. “Prediction of
Fr ⫽ Froude number 关=V / 共gy兲0.5兴; soil-water characteristic curve using genetic programming.” J. Geo-
g ⫽ acceleration due to gravity 共m / s2兲; tech. Geoenviron. Eng., 132共5兲, 661–665.
L ⫽ length of pier 共m兲; Johnson, P. A. 共1995兲. “Comparison of pier-scour equations using field
V ⫽ flow velocity 共m/s兲; data.” J. Hydraul. Eng., 121共8兲, 626–629.
Keijzer, M., and Babovic, V. 共2002兲. “Declarative and preferential bias in
y ⫽ flow depth 共m兲; and
GP-based scientific discovery.” Genet. Program. Evolvable Mach.,
⫽ standard deviation bed grain size. 3共1兲, 41–79.
Kizhisseri, A. S., Simmonds, D., Rafiq, Y., and Borthwick, M. 共2005兲.
“An evolutionary computation approach to sediment transport model-
References ing.” Proc., 5th Int. Conf. on Coastal Dynamics, ASCE, Barcelona,
Spain.
ASCE Task Committee. 共2000兲. “The ASCE Task Committee on applica- Koza, J. R. 共1992兲. “Genetic programming: On the programming of com-
tion of artificial neural networks in hydrology.” J. Hydrol. Eng., 5共2兲, puters by means of natural selection.” A Bradford book, MIT Press,
115–137. Cambridge, Mass.
Azmathullah, H. Md., Deo, M. C., and Deolalikar, P. B. 共2005兲. “Neural Landers, M. N., and Mueller, D. S. 共1999兲. “U.S. Geological Survey field
networks for estimation of scour downstream of ski-jump bucket.” J.
measurements of pier scour.” Proc., Compendium of Papers on ASCE
Hydraul. Eng., 131共10兲, 898–908.
Azmathullah, H. Md., Deo, M. C., and Deolalikar, P. B. 共2008兲. “Alter- Water Resources Engineering Confs. 1991 to 1998, ASCE, Reston,
native neural networks to estimate the scour below spillways.” Adv. Va., 585–607.
Eng. Software, 39共8兲, 689–698. Laursen, E. M., and Toch, A. 共1956兲. “Scour around bridge piers and
Babovic, V., and Keijzer, M. 共2000兲. “Genetic programming as a model abutments.” Bulletin no. 4, Iowa Road Research Board, Ames, Iowa.
induction engine.” J. Hydroinform., 2共1兲, 35–60. Lee, T. L., Jeng, D. S., Zhang, G. H., and Hong, J. H. 共2007兲. “Neural
Bateni, S. M., Borghei, S. M., and Jeng, D.-S. 共2007兲. “Neural network network modeling for estimation of scour depth around bridge piers.”
and neuro-fuzzy assessments for scour depth around bridge piers.” J. Hydrodynam., 19共3兲, 378–386.
Eng. Applic. Artif. Intell., 20, 401–414. Melville, B. W., and Coleman, S. E. 共2000兲. Bridge scour, Water Re-
Brameier, M., and Banzhaf, W. 共2001兲. “A comparison of linear genetic sources, Highlands Ranch, Colo.
programming and neural networks in medical data mining.” IEEE Melville, B. W., and Sutherland, A. J. 共1988兲. “Design method for local
Trans. Evol. Comput., 5, 17–26. scour at bridge piers.” J. Hydr. Div., 114共10兲, 1210–1226.
Davidson, J. W., Savic, D. A., and Walters, G. A. 共1999兲. “Method for Mohamed, T. H., Noor, M. J. M. M., Ghazali, A. H., and Huat, B. B. K.
identification of explicit polynomial formulae for the friction in tur- 共2005兲. “Validation of some bridge pier scour formulae using field and
bulent pipe flow.” J. Hydroinform., 1共2兲, 115–126. laboratory data.” American Journal of Environmental Science, 1共2兲,
Giustolisi, O. 共2004兲. “Using genetic programming to determine Chèzy 119–125.
resistance coefficient in corrugated channels.” J. Hydroinform., 6共3兲, Oltean, M., and Groşan, C. 共2003兲. “A comparison of several linear ge-
157–173. netic programming techniques.” Adv. Complex Syst., 14共1兲, 1–29.