Vous êtes sur la page 1sur 15

Mathematical and Computer Modelling 44 (2006) 415429

www.elsevier.com/locate/mcm

Selection of genetic algorithm operators for urban drainage model


parameter optimisation
N.R. Siriwardene, B.J.C. Perera
School of Architectural, Civil and Mechanical Engineering, Victoria University, PO Box 14428 MCMC, Melbourne, Victoria 8001, Australia

Received 6 December 2004; accepted 11 February 2005

Abstract

Recently, genetic algorithms (GAs) have proven to be successful and efficient in identifying the optimal parameters for water
resource modelling applications. However, in order to produce efficient and robust solutions, proper selection of GA operators for
the application is necessary, before conducting the model parameter optimisation. General guidelines are available for standard GA
optimisation applications. However, there is no specific guidance available for selecting GA operators for urban drainage model
parameter optimisation. Therefore, the sensitivities of these operators were analysed through numerical experiments by repetitive
simulation considering one GA operator at a time, by integrating GA and urban drainage modelling software. The tested GA
operators in this study were the population size, the number of generations, the number of model parameter sets to be considered
from the final generation to determine the optimum set, the selection type and the crossover and mutation rates. It was found
that urban drainage models with a small number of parameters (i.e. two or less) could be optimised with any of the tested GA
operator sets. However, the proper selection of GA operators is vital to the convergence of the optimum model parameters, for
urban drainage models with a large number of parameters (i.e. five or more).
c 2005 Elsevier Ltd. All rights reserved.

Keywords: Genetic algorithms; Urban drainage; Parameter optimisation; Modelling

1. Introduction

The balance of the natural hydrological cycle is greatly disturbed by urban development, in terms of storm-water
volume and quality, as urban development prevents the absorption of rainwater to the soil due to replacement of large
pervious areas of land by impervious areas. Therefore, the management of storm-water runoff from urban catchments
has become an increasingly important environmental issue and storm-water drainage is still a major part of this overall
storm-water management.
Mathematical computer software tools are widely used to develop urban storm-water drainage system models,
and to design and analyse complex urban storm-water drainage systems. These software tools allow modelling of
hydrological (e.g. rainfall, infiltration, overland flow, evaporation) and hydraulic (e.g. pipe and open channel flow)
processes of urban catchments. Some of the urban drainage software tools widely used in Australia are SWMM [1],

Corresponding author. Tel.: +61 3 9919 4729; fax: +61 3 9919 4139.
E-mail addresses: nilmini rukma@hotmail.com (N.R. Siriwardene), Chris.Perera@vu.edu.au (B.J.C. Perera).

c 2005 Elsevier Ltd. All rights reserved.


0895-7177/$ - see front matter
doi:10.1016/j.mcm.2006.01.002
416 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

MOUSE [2], XP-UDD [3] and DRAINS [4]. Flood hydrographs and peak flow runoff can be computed by using
these software tools, which are required to design and/or upgrade the drainage systems to minimize flood damage.
However, the reliability of these models depends on the accuracy in choosing the model parameter values of the
catchments being modelled. Some of these parameter values can be physically measured, whereas other parameter
values (such as depression storage and flow roughness) are impossible or difficult to measure. However, the parameter
values that are impossible or difficult to measure physically, can be estimated through model calibration by using
good quality rainfall/runoff data, if they are available. The model calibration is done through an iterative process by
comparing model predictions with observations, until the two sets match with each other within a reasonable accuracy.
There are several methods available to calibrate mathematical models ranging from trial and error to optimisation
methods. Traditionally, model calibration was done by trial and error. With this method, the model parameters are
estimated by experienced modellers starting with educated guesses and refining these guesses by comparing model
predictions with observations. However, this method is subjective, time consuming and can also miss the optimum
parameter set. Recently, computer-based automatic optimisation methods have proven to be robust and efficient.
Automatic optimisation methods can be characterised as being either deterministic or stochastic. Deterministic
optimisation methods are designed to locate the optimum parameter set and are typically only successful when the
response surface is uni-modal (i.e. a single peak/trough). If the response surface is multi-modal, the parameter set
obtained from the deterministic method may not produce the global optimum, since the search can be trapped at a
local optimum point. Sorooshian and Gupta [5] and Hendrickson et al. [6] have shown that deterministic optimisation
techniques were not appropriate for water resource applications, since most water resource models contain a large
number of parameters (that cannot be easily optimised) and in general the parameter search space of these models
contains multiple peaks (i.e. is multi-modal).
Stochastic optimisation methods are capable of handling multi-modal functions. Some research works in water
resource applications have shown that stochastic optimisation techniques have the ability to overcome the problems
associated with deterministic optimisation techniques discussed above and are more efficient in locating the optimum
parameter set compared with deterministic methods [710]. In this research, one of the most popular stochastic
optimisation methods known as genetic algorithms (GAs) [11] was used to calibrate urban drainage models. Even
though GAs have been recognized as robust optimisation methods for estimating model parameters in many fields,
they have not been used widely for the calibration of urban drainage models.
GA operators, such as parameter representation, population size, selection methods, and crossover and mutation
rates play an important role in the convergence to the optimum model parameter set. However, there is no clear
guidance available to select appropriate GA operators for urban drainage model parameter optimisation. Schaffer
et al. [12] reported that the theory behind GAs has given little guidance for selecting proper GA operators, even
though these operators have a significance impact on GA performance. Davis [13] reported that the optimum GA
operator set varies according to the application. Franchini and Galeati [8] studied the effects of GA operators in
detail in their rainfall-runoff model calibration study and commented that a robust GA operator range was adequate,
as it did not have any significant effect on the optimum model parameter set. Ng [10] and Wardlaw and Sharif
[14] conducted comprehensive GA operator studies in their river water quality model parameter optimisation and
optimal reservoir system operation studies, respectively, and arrived at different optimum GA operators. The above
studies show that there are no clear conclusions regarding optimum GA operators to be used in model parameter
optimisation. Therefore, a detailed investigation was conducted in this study to determine the optimum GA operators
before attempting the model parameter optimisation in urban drainage modelling. The tested GA operators in this
research were the population size, the number of generations, the number of model parameter sets to be considered
from the final generation (to determine the optimum set), the selection type, and the crossover and mutation rates.

2. Genetic algorithms

Genetic algorithms (GAs) are widely used stochastic search methods originally developed by Holland [15] and
later refined by many others. GAs are theoretically and empirically proven to provide a robust search in complex non-
linear problems [11]. They use computer-based iterative procedures that employ the mechanics of natural selection
and natural genetics to select the optimum solution for a given problem. GAs search the optimum solution from one
set of possible solutions at a time, rather than one solution at a time. Specially, the notion of survival of the fittest plays
a central role in GAs.
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 417

Genetic algorithms are rooted in both natural genetics and computer science. Therefore, the GA terminology
has a mixture of natural and artificial terms. The set of possible solutions is called a population. There are several
populations in a GA run. Each of these populations is called a generation. In the GA context, the model parameter set
is defined as a chromosome, while each parameter present in the chromosome is known as a gene. At the start of the
GA optimisation of model parameters, the user has to define the GA operators (Section 2.1). The initial population
is generated according to the selected parameter representation at random or using a priori knowledge of the search
space. The initial population provides the set of all solutions for the first generation, according to the user-defined
model parameter ranges. A user-defined objective function is then used to evaluate each chromosome in the initial
population. These objective function values of the chromosomes indicate the suitability (or fitness) of the parameter
set for the given problem. After computing the objective function values for each chromosome of the initial population,
GA operators such as selection, crossover and mutation are used to generate the population in the next generation.
Numerous generations are considered in the GA process, until the user-defined termination criteria are reached.

2.1. Genetic algorithm operators

The GA operators, namely parameter representation, population size, selection type, crossover and mutation,
control the process of the GA. These operators play an important role in the efficiency and ability of GA optimisation
in reaching the optimum solution. One of the more difficult aspects of using GAs is to choose the optimum operator
set for the relevant problem.

2.1.1. Parameter representation


Parameter representation or encoding is a process of representing the model parameter values in a GA such that
the computer can interact with these values. There are two main types of parameter coding methods available, which
are bit string coding and real-value coding. In real-value coding each parameter is represented by its real value. There
are two types of bit string coding methods available, namely binary and Gray coding, which use similar concepts.
Bit string coding is the most commonly used method by GA researchers because of its simplicity. Furthermore, the
conventional GA operations and theory were developed on the basis of this fundamental structure, which has been
used in many applications [16,17].
In binary representation for GAs, each parameter in the model parameter set is encoded as a sub-string with
binary digits (i.e. 0 and 1). These sub-strings, corresponding to each parameter, are then arranged linearly to form
a string to represent the entire model parameter set. Gray coding is an ordering of binary character sets such that all
adjacent numerical numbers differ by only one bit, whereas in binary coding adjacent numbers may differ in many bit
positions. The advantage of Gray coding is that random bit flips in mutation (Section 2.1.5) are likely to induce small
changes in real parameter value and therefore smooth mapping between the real search and the encoded parameters
is achieved. Caruana and David [18] reported that Gray coding can eliminate a hidden bias in binary coding and that
the large Hamming distances (i.e. number of bit positions that differ in adjacent bit strings of equal length) in the
binary representation resulting from mutations could cause the search process to be too greatly disturbed or unable
to efficiently locate the global optimum. According to Caruana and David [18], the first suggestion of the superiority
of Gray coding was by Hollstien [19]. Gray coding was recently selected as the parameter representation method in
several water resource applications [10,14,20]. Mayer et al. [21] compared binary and real-value representations in GA
optimisation, and found that there were no differences between binary and real-value representations in converging
to the optimum parameter set in their agricultural system model studies, although the real-value representation was
expected to be superior to binary coding [22].

2.1.2. Population size


Population size is the number of chromosomes present in a population. Larger population sizes increase the amount
of variation present in the population (or population diversity), but at the expense of requiring more fitness evaluations
[11]. Furthermore, when the population size is too large, there is a tendency by the user to reduce the number of
generations in order to reduce the computing effort, since the computing effort depends on both population size and the
number of generations. Reduction in the number of generations reduces the overall solution quality. On the other hand,
a small population size can cause the GA to converge prematurely to a sub-optimal solution. Goldberg [11] reported
that population sizes ranging from 30 to 200 were the general choice of many GA researchers. Furthermore, Goldberg
418 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

pointed out that the population size was both application dependent and related to the length of the chromosome (i.e.
string length). For longer chromosomes and challenging optimisation problems, larger population sizes were needed
to maintain diversity, as they allowed better exploration.

2.1.3. Selection
The selection process determines which chromosomes participate in reproduction to generate the next population
(in the next generation) according to their fitness values in the current population. In general, this process takes
advantage of the fittest solutions by giving them greater weight when selecting the next generation and hence leads
to better solutions to the problem. There are several ways to implement selection in GA optimisation. Proportionate
selection [23], linear ranking [24] and tournament selection [25], are commonly used selection methods. Goldberg
and Deb [26] stated that no one selection method was superior to the other.

2.1.4. Crossover
The crossover operator is used to create new chromosomes for the next generation by randomly combining two
selected chromosomes from the current generation through the selection process (Section 2.1.3). However, some
algorithms use an elitist selection strategy, which ensures the fittest chromosome from one generation is propagated
into the next generation without any disturbance. The crossover rate is the user-defined probability that crossover
reproduction will be performed. For example, a crossover rate of 0.9 means that on average 90% of the population
undergoes the crossover operation. A high crossover rate encourages good mixing of the chromosomes. There are
several crossover methods available, namely single point, multi-point, uniform crossover, etc. The choice of the best
crossover method is primarily dependent on the application.

2.1.5. Mutation
Mutation introduces innovation into the population by randomly modifying the chromosomes. It prevents the
population from becoming saturated with chromosomes that are all similar and reduces the chance of premature
convergence [27]. For example, in bit string representation, mutation is done by flipping 0s to 1s and vice versa. Large
mutation rates increase the probability of destroying good chromosomes. The user defined mutation rate determines
the probability that mutation will occur. For example, if the population size is 100, the string length is 20 and the
mutation rate is 0.001, on average only two bit positions will alter in the whole population (i.e. 100 20 0.001 = 2).

3. Selection of GA operators for use in XP-UDD drainage model

3.1. XP-UDD urban drainage model

As stated in Section 1, there are several computer software packages available that are used for urban drainage
modelling. Most of these software tools use the same equations for modelling hydrological and hydraulic processes
of urban drainage catchments. XP-UDD [3] was used in this study as it is an improved version of SWMM [1], and
its input and output files are in ASCII format, which enables easy interaction with the GA software (or other external
software tools).
In XP-UDD, the urban catchment is divided into two significant sub-areas, namely impervious and pervious areas.
The impervious areas include road surfaces, roofs and other human-made hard surfaces. The pervious areas include
bare surfaces, porous pavements, grass courts and lawns. During small storm events, runoff is generally generated
only from the impervious areas after filling their depression storages, and no runoff is generated from pervious areas,
since rain falling on the pervious areas infiltrates into the soil. However, during large storm events, pervious areas
contribute to runoff, in addition to impervious areas.
Victoria University, in collaboration with ten city/shire councils in Victoria, Australia, conducted a major data
acquisition program for 26 urban catchments during 19961999, collecting data on rainfall and runoff [28]. Two of
these urban drainage catchments, namely Kew and Warringal catchments, which are typical urban drainage catchments
in the Melbourne metropolitan area, were used in this study for investigating the optimum GA operators for use in GA
optimisation of urban drainage models. The Kew catchment in the City of Boroondara, has a catchment area of 18 ha
and 34 inlet pits. The underground drainage pipe diameters range from 300 to 750 mm. The soil type is poorly graded
gravel and gravel sand mixtures, with little or no fines. The land use of the catchment is fully residential and flat
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 419

Fig. 1. Linking of XP-UDD and GENESIS.

terrain, and typical house blocks are fairly small. The Warringal catchment, in the City of Banyule, has a catchment
area of 29 ha and 71 inlet pits. The underground drainage pipe diameters range from 300 to 1200 mm. The soil type
is well or poorly graded gravel and gravel sand mixtures, with little or no fines. The land use of the catchment is fully
residential, with many single-house properties and few units.
Of all user input model parameters of XP-UDD, seven parameters were identified for calibration in this study
because of the difficulty in measuring them physically. Two of them are related to the impervious areas and they
are the percentage of the impervious area (%A) and the depression storage (DSi ). The other five are related to the
pervious areas, which are the depression storage (DS p ), the overland flow roughness (n p ), and the three Hortons soil
infiltration parameters ( f o , f c and k). Although %A can be approximately estimated using aerial photographs and/or
rainfallrunoff depth plots, it is difficult to accurately estimate as it requires the identification of individual properties
that are connected to the drainage system. Therefore, %A was selected for calibration in this study, with the initial
values for the GA optimisation obtained from aerial photographs.

3.2. GA software

There are several GA software tools available, which have been developed using Fortran, C/C++, Java and other
programming languages (e.g. Matlab). The web site http://www.aic.nrl.navy.mil/galist/src/#C provides links to some
of these public domain GA software tools. Although these software tools use the same GA theory, each software tool
is designed and implemented in a slightly different way, using various GA operator options and various compilers.
For example, as reported by Mardle and Pascoe [29], only one selection method and only one crossover type are
implemented in the Standard Genetic Algorithms SGA [11] GA software tool, whereas seven selection methods, four
crossover types and two mutation procedures are implemented in the GENEsYs [30] GA software tool.
GENESIS version 5.0 [31] was selected for this study, since it has been used successfully for various water resource
applications in the past. Liong et al. [7] used GENESIS for calibrating the SWMM model, while Ng [10] and Mulligan
and Brown [20] coupled GENESIS with river water quality models to optimise model parameters.

3.3. Linking of XP-UDD and GENESIS

A computer program was developed using the C programming language to link the operation of XP-UDD and
GENESIS, to obtain the optimum GA operators and then to perform automatic calibration of model parameters of the
selected study catchments. This computer program is shown as MODEL INTERFACE in Fig. 1. The linked overall
program is called GENESIS/XP-UDD.
GENESIS/XP-UDD prompts for the user inputs such as number of parameters to be calibrated, their parameter
ranges, population size, GA operator options, termination criteria etc. It then generates the initial population consisting
420 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

of model parameter sets, according to the given user inputs. Then MODEL INTERFACE interacts with the generated
population, XP-UDD and GENESIS to continue the generation of new populations, as shown in Fig. 1. In summary,
MODEL INTERFACE was written to perform the following tasks:
Modify the XP-UDD input file by extracting one parameter set from the GA generated population.
Run the XP-UDD model for this particular model parameter set.
Extract the resultant hydrograph ordinates from the XP-UDD output file, relevant to the above model parameter
set.
Interact with the relevant observed hydrograph ordinates (stored in a file) to compute the objective function values
for the above model parameter set.
Repeat the above steps for the entire GA population.
Feed all objective function values of the population to GENESIS.
Capture and write the results of the GA process at the end of each generation for a detailed analysis of the results
(Note: GENESIS gives only the final results, when the termination criteria are met).
Continue above steps for all generated populations, until the termination criteria are met.

3.4. Methodology, results and discussion

The GA operator study was conducted in two stages as follows:


(a) Investigation of GA operators using the Kew catchment model
(b) Validation of the above results using the Warringal catchment model.
This approach was used to reduce the computational time associated with large numbers of simulations of the
linked GENESISS/XP-UDD model. In general, it took about 7 s for an XP-UDD run of the Kew catchment model
and about 35 s for the Warringal catchment model on a 312 MB RAM, 864 MHz Pentium 4 computer. Therefore,
the Kew catchment model was used to study the optimum GA operators extensively and to determine the optimum
GA operators, and the Warringal catchment model was used only to validate the results obtained from the Kew
catchment model. The validation was conducted through investigation of optimum GA operators using the reduced
GA operator ranges (obtained from the Kew catchment) and the neighbourhood of these reduced ranges. Two studies
were conducted separately for each catchment for impervious and pervious area parameters, as the runoff generation
mechanisms are different in these two areas, as stated in Section 3.1.
The XP-UDD models of the Kew and Warringal storm-water drainage networks were assembled using information
on existing pits and pipes of the networks. All existing drainage pipes that are equal or greater than 300 mm were
assembled as links. Some of the sub-catchment input data were estimated from the catchment contour maps and aerial
photographs. These data include total sub-catchment area and slope of the sub-catchments, which were then entered
as node details when assembling the XP-UDD network. The drainage system input data, such as conduit shape, size,
length, slope, conduit invert level and ground level etc., were obtained from the drainage plans of the catchments. The
sub-catchment width was estimated by dividing the area of the sub-catchment by the average path of the overland flow,
as specified in the XP software manual [3], since there was insufficient information available to use other methods.
Two design storms (one small and the other large) were considered in optimising the GA operators in this study.
The small storm had an Annual Recurrence Interval (ARI) of 1 year and a storm duration of 30 min. This storm
produced runoff only from the impervious areas of the study catchments, as evident from the XP-UDD model output
results, and was used to study the optimum GA operators in relation to the two impervious area parameters. The large
storm, which had an ARI of 100 years and a 30 min duration, generated runoff from both impervious and pervious
areas of the catchment, as evident from the model output results, and was used to study the optimum GA operators in
relation to the remaining five pervious area parameters after fixing the two impervious area parameters obtained from
the impervious area study. Although it is beneficial to consider several storms, it is not feasible to perform an analysis
such as this, because of the high computational time associated with the repetitive GA runs.
Typical values were assumed for the model parameters (i.e. two impervious area parameters and five pervious
area parameters) to generate the two hydrographs corresponding to the above small and large storm events for each
catchment. These parameter values were considered as the actual values of the model parameters for the catchments
and the generated hydrographs above were considered as the observed hydrographs corresponding to small and large
storm events, in optimising GA operators. The objective function used in this study was the minimisation of the sum of
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 421

squared difference of computed (due to different model parameter values generated by GA) and observed hydrograph
ordinates, as it has been widely used in many previous studies (e.g. [7]) and implicitly allows for the other important
features of the hydrographs such as peak, time to peak and volume to be matched.
As detailed in the preceding paragraphs, two real catchments were used in the study. However, the model
parameters used were the typical parameter values (and not the calibrated values for these catchments, since they
were not available). The real storm data were not used, since they are not compatible with the typical model parameter
values. Instead, the design storms using the typical parameter values, generated from XP-UDD were considered as the
observed hydrographs. The real catchments, the typical values of model parameters and the design hydrographs used
in this study are compatible with each other and can be used to investigate the optimum GA operator sets effectively.
It should also be noted that in this study the optimum GA operators are determined considering both the fitness
of the objective function and the convergence to the actual parameter values (and not just the fitness of the objective
function). This approach was used since there could be several combinations of parameter values that are not reflective
of the actual parameter values, which can produce slightly better objective functions.
Although binary and Gray coding are available in GENESIS for parameter representation, Gray coding was used
in this study, as it is an enhancement of binary coding. Furthermore, the two-point crossover type was used, as it
was the only crossover method available in GENESIS. Crossover and mutation rates of 0.6 and 0.001 (respectively)
and the proportionate selection method were used in Section 3.4.2 in population size related investigations. However,
these GA operators were later optimised in Sections 3.4.3 and 3.4.4. It should be noted that in GENESIS, the default
values of crossover and mutation rates are 0.6 and 0.001 respectively, and the default selection method is proportionate
selection.
The linked GENESIS/XP-UDD model was then used to study the effects of GA operators of population size,
selection type, and crossover and mutation rates, and to determine the optimum GA operator set for urban drainage
modelling parameter optimisation. These operators were varied one at a time, keeping all other operators constant in
studying the effect of these operators. Each of these GA operator studies and their results are discussed under different
headings below. The reader is referred to Siriwardene [32] for complete study details.

3.4.1. Preliminaries
Two issues, namely the effect of seed (which is used to generate model parameters for multiple GA runs as
reflective of the GA stochastic process) on the model parameter convergence and the impact of string length on
population size (and its effect on model parameter convergence), were studied first as preliminaries.
(a) Seed. A random number generator is used in GENESIS (typical for GA software) to generate the initial population
and subsequent populations. First, a member of the initial population is specified by the user as the seed, which is
combined with random numbers to generate the remaining parameter sets. Different seeds generate different parameter
sets for the initial population, and therefore the seed may have an impact on the optimum model parameter set obtained
from the final generation of GA optimisation, and also on the optimum GA operator set.
Wang [33] and Franchini [34] considered 10 seeds in studying the effect of seed in obtaining the optimum parameter
set in their respective conceptual rainfall and runoff models. Wang [33] reported that 8 out of 10 runs had given the
same optimum parameter set with the same objective function value, while the objective function values for the
other 2 runs were marginally better, but giving almost the same optimum parameter set. Franchini [34] found that all
10 runs produced almost the same objective function value, but with significant differences in the actual parameter
values. He claimed that this difference was attributed to the errors in the data and the imperfect structure of the
model, which caused some of the parameters to be insensitive. Mohan [35], on the other hand, used 20 different
seeds and adopted the best parameter set of 20 runs as the optimum parameter set. Mohan [35] did not comment on
the differences in results due to different seeds. Ng and Perera [36] studied the effect of seed on model parameter
convergence of their river water quality modelling application and found that the seed did not have any effect on the
parameter optimisation, provided that an adequate number of GA runs was conducted by progressively reducing the
parameter search space using the results of the previous runs. As can be seen from the above studies, the effect of
the seed on the convergence to the optimum model parameter set is inconclusive, which may have an effect on the
optimum GA operators. Therefore, it was considered necessary to study the effect of the seed on the convergence to
the optimum model parameter set and to determine the optimum GA operators for GA optimisation of urban drainage
model parameter optimisation.
422 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

Fig. 2. Number of simulations vs. number of solutions with objective function values of zero for a population size of 25 for the Kew catchment
model.

Six different GA optimisation runs were conducted with six different arbitrary seeds for both impervious and
pervious area model parameter estimation studies, and the objective functions and the model parameter sets were noted
from the final GA generation. For the impervious area study, a population size of 25 with 7500 simulations (i.e. 300
generations) was considered and it was found that all model parameters converged to the actual parameter set, giving
objective function values of zero. For the pervious area study, a population size of 100 with 7500 simulations (i.e.
75 generations) was considered. The results of the final GA generation showed that the objective functions, the best
model parameter set and the mean of five best parameters corresponding to each seed run, were similar to each other
(the differences being marginal). After considering the above results of both impervious and pervious area studies, it
was concluded that the seed does not play a significant role in the urban drainage model parameter optimisation and
therefore in particular in determining the optimum GA operator set for urban drainage model calibration, provided a
sufficient number of simulations and a reasonable narrow parameter range are considered in GA runs, as was done in
this study.
(b) String length. Goldberg [37] reported that the selection of population size depends on the string length of the
model parameter set. In the bit string representation of GA, the string length of each parameter is computed based on
parameter range and required precision (i.e. number of decimal places) of the parameter value. The string length
increases with the increase in the parameter range and higher required precision (i.e. higher number of decimal
places).
Several population sizes were considered with different string lengths, to investigate the impact of string length on
parameter convergence. These population sizes were 25, 50, 75, 100 and 100, 125, 150 for the impervious and pervious
area parameter studies, respectively. Different string lengths (which implicitly account for the parameter range and
precision) were considered for both studies and they were 10, 16, 20 and 38, 48 for impervious and pervious area
parameter studies, respectively, and were computed using assumed reasonable values for parameter range and required
precision for parameters.
Fig. 2 shows an example of the results obtained for a population size of 25 for the impervious area study of
the Kew catchment model. Similar results were obtained for other population sizes in the impervious area study.
It can be seen from Fig. 2 that the longer the string length (i.e. larger parameter range and/or higher accuracy of
parameters), the longer the time required for convergence. Therefore, GA efficiency can be achieved with reduced
string lengths by limiting parameter range and the accuracy of the model parameters (i.e. number of decimal places
required) to the required level only. Similar to the impervious area results, it was found (with a population size of
100) that a larger number of simulations was required for convergence with increasing string length. It was also
noted that convergence to the actual parameter values was difficult with increased string length. Similar results were
found with other population sizes (i.e. 125, 150 and 200). Based on the above results, it can be confirmed that the
optimum population size depends on the string length of the model parameter set, as reported by Goldberg [37].
However, further investigations are required to build a relationship between population size, string length and number
of simulations.
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 423

Fig. 3. Number of simulations vs. number of solutions with objective function values of zero (as a % of population) for the Kew catchment model.

3.4.2. Population size and associated issues


In this section, two issues related to population size were studied as follows:

(a) Selecting the optimum population size and number of generations.


(b) The number of optimum model parameter sets to be considered from the final generation.

(a) Population size and number of generations. Based on the previous work of Franchini and Galeati [8], population
sizes of 75, 100, 125 and 200 were initially investigated for both impervious and pervious area studies with 7500
simulations. Using these results, further investigations were conducted for population sizes of 10, 25, 50 and 150 for
the impervious area study, and 50, 150, 300 and 500 for the pervious area study. The optimum population size and the
number of generations were then selected from these GA runs. The population size and the number of generations are
related by the total number of simulations in one GA run, and therefore these two were studied together.
It was observed that the two model parameters (i.e. %A and DSi ) converged to the actual values easily in the
impervious area study, achieving objective function values of zero. Fig. 3 shows the plot of number of simulations
against the number of solutions that had objective function values of zero in the final generation, expressed as a
percentage of the population size for population sizes of 25, 50, 75, 100, 125, 150 and 200 with 7500 simulations.
As can be seen from Fig. 3, the convergence rate decreases with an increase in population size, which is due to an
increased number of redundant solutions (with the increase in population size). The population size of 10 is not shown
in Fig. 3, since it did not converge to the actual model parameters at all. This is due to not having enough variation in
parameter values in the population.
As can be seen from Fig. 3, all parameter sets converged very quickly with a population size of 25, within 1125
simulations (i.e. 45 generations). However, the other population sizes were not able to give similar results with the
same number of generations. Similar results were obtained with the Warringal catchment model. Based on these
results, a population size of 25 and 1200 simulations were identified as the optimum population size and number of
simulations, respectively, for optimising impervious area parameters in this study. These were used in the rest of the
GA operator study.
The five pervious area parameters did not easily converge to the actual parameters giving objective function values
of zero, as in the impervious area study. Therefore, Fig. 4 was produced to illustrate the results in terms of minimum
objective function, mean of the lowest five objective functions and mean of the lowest ten objective functions in the
final generation (i.e. after 7500 simulations). Although it can be seen from Fig. 4 that population sizes of 50, 75 and
100 were equally good in terms of the objective function, a detailed analysis of the model results showed that only
a population size of 100 resulted in convergence of all five model parameters. Similar results were obtained with
the Warringal catchment model. Based on the above results, a population size of 100 and 7500 simulations (i.e. 75
generations) were identified as the optimum population size and number of simulations, respectively, for optimising
pervious area parameters. These were used in the rest of the GA operator study.
424 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

Fig. 4. Objective function value vs. population size for the Kew catchment model.

Fig. 5. Number of parameter sets vs. saturated soil infiltration rate ( f c ) for the Kew catchment model.

(b) Number of optimum parameter sets to be considered from final generation. In a typical GA run, there could be
several equally good parameter sets giving almost the best objective function in the final generation. The objective
functions of these sets may differ only by a small margin, though there could be significant differences in their
parameter values. Therefore, it is not appropriate to select a single parameter set from the final generation. Liong
et al. [7] and Wang [33] selected a single parameter set based on the best objective function. Franchini and Galeati
[8] determined the mean value of the best 20-parameter sets based on objective functions in their rainfall runoff
model. Ng [10] selected the mean value of the best 10 parameter sets based on objective function in her river water
quality modelling application. Therefore, the results obtained from the GA runs were analysed, to determine how
many parameter sets need to be considered from the final generation to determine the optimum parameter set for
urban drainage modelling.
As stated above, all parameter sets with a population size of 25 reached the actual values in the final generation
for the impervious area parameter study, and therefore need not be studied any further for this case. The values of the
pervious area parameters and their objective function values of the final generation for a population size of 100 were
studied in detail, and plots were made of these parameters to show their mean, minimum and maximum with respect
to a number of best parameter sets taken from the final generation. Fig. 5 illustrates an example of the results obtained
for f c (whose actual parameter value was 10 mm/h) for the Kew catchment model. In general, it was found that the
number of best parameter sets beyond six deviated from the actual parameter values using the results of both the Kew
and Warringal catchment models. Therefore, the mean of the best five parameter sets based on the objective function
from the final generation was considered as the value of the optimum parameter set.
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 425

Table 1
Model parameter convergence with different selection methods for the Kew catchment model

Crossover and mutation Selection Mean of 5 lowest objective np DS p fo fc k


rates method functions (l/s)2 0.03 3 (mm) 100 (mm/h) 10 (mm/h) 0.001 (1/s)
Crossover 0.6 Proportionate 4.52 0.029 3.1 99.4 10 0.001
Mutation 0.001 Linear ranking 3.14 0.028 2.93 97 12 0.0011
Crossover 0.6 Proportionate 13.79 0.032 2.72 107 11.6 0.001
Mutation 0.01 Linear ranking 10.5 0.036 2.5 107 12 0.0011
Crossover 0.9 Proportionate 4.03 0.028 3.02 103 13.4 0.0011
Mutation 0.001 Linear ranking 3.24 0.025 3.5 105 14 0.0012
Crossover 0.9 Proportionate 9.75 0.03 2.96 102 10.4 0.0011
Mutation 0.01 Linear ranking 5.99 0.028 2.92 105 12 0.0012

Table 2
Model parameter convergence with different selection methods for the Warringal catchment model

Crossover and mutation Selection Mean of 5 lowest objective np DS p fo fc k


rates method functions (l/s)2 0.03 3 (mm) 100 (mm/h) 10 (mm/h) 0.001 (1/s)
Crossover 0.6 Proportionate 10.88 0.033 2.8 98.5 10 0.001
Mutation 0.001 Linear ranking 8.55 0.020 3.9 94.0 10 0.002

3.4.3. Selection type


The options of proportionate selection and linear ranking are the only available selection methods in GENESIS.
Therefore, the effects of these two methods on the convergence to the optimum model parameter set were investigated.
Each selection method was studied with crossover rates of 0.6 and 0.9 and mutation rates of 0.001 and 0.01. These
values were the boundaries of the optimum GA operator ranges defined in the literature.
All parameter sets of population size 25 reached the actual values in the final generation (i.e. 1200 simulations)
with the proportionate and linear ranking selection methods in the impervious area parameter studies for both the Kew
and Warringal catchment models. The rate of convergence with the two selection methods was also similar.
It was found that convergence for the pervious area studies was slightly faster with the linear ranking method
compared to the proportionate selection method, as can be observed by the relatively higher objective function values
in the latter (Table 1). However, the pervious area model parameters converged to the actual values more accurately
with the proportionate selection method than the linear ranking method, as shown in Table 1 for the Kew catchment
model. The actual parameter values are shown in bold type in Table 1.
Only crossover rates of 0.6 and mutation rates of 0.001 were used to validate the above observations with the
Warringal catchment model, since the other combinations of crossover and mutation rates in Table 1 showed a similar
pattern (i.e. more accurate parameter values with the proportionate selection method compared to the linear ranking
method, but with slightly higher objective function values). The results obtained from this study are shown in Table 2
and are similar to those of the Kew catchment. Therefore, the proportionate selection method was recommended as
the selection method for urban drainage model parameter optimisation and was used for the rest of the study.

3.4.4. Crossover and mutation rate


In this part of the study, the effects of crossover rate were first investigated for the impervious and pervious area
parameter studies with the Kew and Warringal catchment models. For both studies, crossover rates ranging from 0.1
to 1, in increments of 0.1, were initially investigated, keeping the mutation rate at 0.001, which is the default value in
GENESIS. Then, these results were analysed to produce a narrow range of crossover rates, based on the convergence
of model parameters to the actual values. This narrow range was then used with different mutation rates, to produce
suitable crossover and mutation rates for urban drainage model calibration. This procedure was adopted, since the
mutation rate has less effect (or none, depending on the population size) on convergence compared to the crossover
rate. In this study, crossover and mutation rates were studied together, as they determine convergence jointly.
All ten runs with 1200 simulations converged to the actual parameter set in the impervious area study. Therefore
only a crossover rate of 0.6 (which is the default value of GENESIS) was studied with different mutation rates.
426 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

Fig. 6. Different crossover rates vs. objective function values for the Kew catchment model.

It was found that mutation rates of 0.05, 0.01, 0.005 and 0.001 were equally good in converging to the actual parameter
values, giving objective function values of zero, and therefore no further mutation rate investigations were conducted.
Based on these results, it can be concluded that crossover and mutation rates (within reasonable bounds) do not
significantly affect the convergence to the actual values of the two parameters in this study.
The pervious area study results are shown in Fig. 6. This figure shows the plot of crossover rates versus objective
function values with a mutation rate of 0.001 for a population size of 100 after 7500 simulations of the Kew catchment
model. As can be seen from Fig. 6, crossover rates of 0.2, 0.3 and 0.5 to 1 gave the best results in the pervious
area study. When the model parameters obtained from these GA runs and the actual values were compared, it was
found that they were closely matched with each other, with the crossover rates between 0.6 and 0.9. Therefore, the
conclusion was made that crossover rates between 0.6 and 0.9 need to be considered for further study with mutation
rates varying from 0.001 to 0.1.
The five model parameter values based on the mean of the five lowest objective functions obtained for these
crossover and mutation rates are shown in Table 3 for the Kew catchment model. Actual parameter values are shown
in bold type under each parameter. It can be seen from Table 3 that a crossover rate of 0.6 with a mutation rate
of 0.001 gave the best result based on convergence to the actual parameter values for this application. The other
acceptable crossover and mutation rates are shown in bold type in Table 3.
Crossover rates of 0.6 and 0.9 coupled with mutation rates of 0.01 and 0.001 were studied using the Warringal
catchment model, after reviewing the Kew catchment model results. The results are tabulated in Table 4 for a
population size of 100 after 7500 simulations. It can be seen from Table 4 that a crossover rate of 0.6 with a mutation
rate of 0.001 gave the best result based on convergence to the actual parameter values for the Warringal catchment,
similar to the Kew catchment. It should be noted that these are also the default values of GENESIS and therefore are
recommended for use in XP-UDD model parameter optimisation.

4. Summary and conclusions

Urban drainage models are widely used in urban storm-water planning and management. In order to use these
models effectively, it is necessary to calibrate them. Optimisation methods are preferred to traditional trial and error
methods for calibrating such models. Genetic algorithms (GAs) are one possible optimisation method, and are gaining
popularity in water resource applications. Before attempting to calibrate urban drainage models using GAs, it is
necessary to obtain the appropriate GA operators for the study, since there is no guidance available for GA operators to
be used in urban drainage modelling. The GA operators studied were the population size, the number of generations,
the number of model parameter sets to be considered from the final generation to determine the optimum set, the
selection type and the crossover and mutation rates.
A systematic trial and error procedure was used to investigate the optimum GA operators for use in urban drainage
model parameter optimisation. The study was conducted as two investigations to estimate impervious and pervious
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 427

Table 3
Model parameter values for the pervious area study for different crossover and mutation rates for the Kew catchment model

Crossover rate Mutation rate np DS p fo fc k Mean of 5 lowest objective


0.03 3 (mm) 100 (mm/h) 10 (mm/h) 0.001 (1/s) functions
0.6 0.001 0.029 3.1 99.4 10.0 0.001 4.52
0.002 0.034 2.68 107 11.8 0.0012 16.56
0.004 0.026 2.76 112 13.6 0.0013 8.33
0.006 0.028 3.18 99.4 13.2 0.0011 18.42
0.008 0.030 3.02 104 13.4 0.0012 24.34
0.01 0.032 2.72 107 11.6 0.001 13.79
0.05 0.023 2.7 109 12.6 0.001 40.39
0.1 0.039 2.52 106 10.2 0.0011 42.49
0.7 0.001 0.029 2.2 120 14.0 0.0014 13.9
0.002 0.024 3.1 111 12.8 0.0013 18.01
0.004 0.060 2.42 107 11.8 0.0012 75.31
0.007 0.034 2.32 115 14.0 0.0013 23.44
0.008 0.029 2.84 110 13.2 0.0013 24.59
0.01 0.029 3.02 102 13.2 0.0012 15.92
0.05 0.039 2.52 113 9.6 0.0012 53.41
0.1 0.034 2.98 95.8 11.2 0.001 22.83
0.8 0.001 0.031 3.12 97.8 10.6 0.001 13.7
0.002 0.028 3 102 12.4 0.0011 14.22
0.004 0.034 1.98 121 14.8 0.0014 10.69
0.006 0.027 2.86 107 12.6 0.0012 28.10
0.008 0.033 2.96 100 12.3 0.0011 20.42
0.01 0.035 2.78 101 13.0 0.0011 39.03
0.05 0.027 2.34 120 13.6 0.001 39.23
0.1 0.050 2.46 107 12.8 0.0013 94.77
0.9 0.001 0.028 3.02 103 13.4 0.0011 4.03
0.002 0.031 2.84 104 12.8 0.001 6.06
0.005 0.038 3.1 98.2 11.8 0.0011 29.36
0.008 0.030 3.26 99.4 12.4 0.0011 18.98
0.01 0.030 2.96 102 10.4 0.0011 9.75
0.05 0.033 3.6 92.6 10.2 0.0009 28.51
0.1 0.031 2.92 105 9.8 0.001 64.48

Table 4
Model parameter values for the pervious area sutdy for different crossover and mutation rates Warringal catchment model

Crossover rate Mutation rate np DS p fo fc k Mean of 5 lowest objective


0.03 3 (mm) 100 (mm/h) 10 (mm/h) 0.001 (1/s) functions
0.6 0.001 0.033 2.8 98.5 10 0.001 10.88
0.01 0.02 3.8 112 13 0.002 16.36
0.9 0.001 0.03 3.82 119.6 14.8 0.002 10.23
0.01 0.03 2.7 106 11.6 0.001 15.77

area model parameters, as the runoff generation mechanism is different in these two areas, which vary according to
the magnitude of the rainfall intensity. Two design storm events of duration 30 min were considered as input rainfall
in the study. The small storm, which had an Annual Recurrence Interval (ARI) of 1 year and produced runoff only
from the impervious areas, was used to study the two impervious area parameters (i.e. %A and DSi ). The large storm,
which had an ARI of 100 years and generated runoff from both impervious and pervious areas, was used to study
the remaining five pervious area parameters (i.e. n p , DS p , f c , f 0 and k). Two typical urban drainage catchments
from the Melbourne metropolitan area of Australia were used in this study. The Kew urban drainage catchment was
used extensively to study the GA operators. The Warringal urban drainage catchment was used to validate the results
obtained from the Kew catchment.
428 N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429

It was found that the GA operators were sensitive to the number of model parameters that need to be optimised
in the application. If the number of parameters to be optimised was small, as in the case of estimating impervious
area parameters (i.e. only two parameters considered), the tested GA operators did not play an important role in
converging to the optimum model parameter set, and therefore general GA operators recommended in literature
can be used. Furthermore, small population sizes (i.e. between 25 and 50) were very efficient for use in model
parameter optimisation of urban drainage models with two or fewer parameters, such as the case of optimising only the
impervious area parameters. For models with a larger number of parameters (five or more) as in the case of estimating
impervious area model parameters, the tested GA operators played an important role in achieving convergence to the
optimum parameter set. In this study, population size of 100, proportionate selection, a crossover rate of 0.6 and a
mutation rate of 0.001 gave the best results, and therefore they are recommended for model parameter optimisation
of urban drainage models with a large number of parameters (i.e. more than five), as in the case of optimising both
impervious and pervious parameters.
The above conclusions were based on limited numerical experiments, and therefore it is recommended that further
studies should be conducted using different catchments under different hydrological conditions to substantiate the
findings of this study. Nevertheless, it adds to the body of knowledge in providing guidance on selection of appropriate
GA operators for urban drainage model parameter optimisation.

Acknowledgements

The authors would like to thank the two anonymous reviewers for their helpful and constructive comments on the
paper, which certainly have improved the quality of the paper.

References

[1] USEPA, Stormwater Management Model (Version 4) Users Manual, in: US Environmental Protection Agency, Environmental Research
Laboratory, Office of Research and Development, Athens, Georgia, USA, 1992.
[2] Danish Hydraulic Institute, Modelling of Urban Sewer Systems on Microcomputers, MOUSE V3.0, in: Users Guide and Technical Reference,
Lawson and Treloar Pty Ltd, Australia, 1993.
[3] XP-Software, XP-UDD32 Getting Started Manual, ACT, Australia, 1997.
[4] G. OLoughlin, B. Stack, DRAINS Users Manual, Watercom Pty Ltd, Sydney, Australia, 1998.
[5] V. Sorooshian, K. Gupta, Model calibration, in: V.P. Singh (Ed.), Computer Models of Watershed Hydrology, Water Resources Publications,
Colorado, 1995, pp. 2668.
[6] J.D. Hendrickson, S. Sorooshian, L.E. Brazil, Comparison of Newton-type and direct search algorithms for calibration of conceptual rainfall-
runoff models, Water Resources Research 24 (1998) 691700.
[7] S.Y. Liong, W.T. Chan, J. Shreeram, Peak-flow forecasting with genetic algorithm and SWMM, ASCE Journal of Hydrologic Engineering
121 (8) (1995) 613617.
[8] M. Franchini, M. Galeati, Comparing several genetic algorithm schemes for the calibration of conceptual rainfall-runoff models, Journal of
Hydrological Sciences 42 (3) (1997) 357379.
[9] D. Savic, G.A. Walters, Genetic algorithms for least-cost design of water distribution networks, ASCE Journal of Water Resources Planning
and Management 123 (3) (1997) 6777.
[10] A.W. Ng, Parameter optimisation of river water quality models using genetic algorithms, Ph.D. Thesis, School of the Built Environment,
Victoria University, Australia, 2001.
[11] D.E. Goldberg, Genetic Algorithm in Search, Optimization and Machine Learning, Addison-Wesley Publishing Company, Inc., Reading, MA,
1989.
[12] J.D. Schaffer, R.A. Caruana, L.J. Eshelamn, R. Das, A study of control parameter affecting online performance of genetic algorithms for
function optimisation, in: Proc. Third International Conference on Genetic Algorithms, 47 October, George Mason University, California,
1989, pp. 5160.
[13] L. Davis, Handbook of Genetic Algorithm, Van Nostrand Reinhold, New York, 1991.
[14] R. Wardlaw, M. Sharif, Evaluation of genetic algorithms for optimum reservoir system operation, ASCE Journal of Water Resources Planning
and Management 125 (1) (1999) 2533.
[15] J. Holland, Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann Arbor, Michigan, 1975.
[16] K.A. De Jong, D. Fogel, H. Schwefel, A history of computation, in: T. Back et al. (Eds.), Handbook on Evolutionary Computation (Part A,
Section 2.3), Institute of Physics Publishing, Oxford University Press, New York, 1997, pp. 112.
[17] D.E. Goldberg, Sizing population for serial and parallel genetic algorithms, in: Proc. Third International Conference on Genetic Algorithms,
47 June, George Mason University, California, 1989, pp. 7079.
[18] R. Caruana, S.J. David, Gray vs. Binary Coding for Genetic Algorithm Function Optimizers, Philips Laboratories, North American Philips
Corporation, New York, 1987.
[19] R.B. Hollstien, Artificial genetic adaptation in computer control systems, Ph.D. Thesis, University of Michigan, Ann Arbor, Michigan, 1971.
N.R. Siriwardene, B.J.C. Perera / Mathematical and Computer Modelling 44 (2006) 415429 429

[20] A.E. Mulligan, L.C. Brown, Genetic algorithm for calibrating water quality models, ASCE Journal of Environmental Engineering 124 (2)
(1998) 104110.
[21] D.G. Mayer, J.A. Belward, H. Widell, K. Burrage, Survival of the fittest genetic algorithms versus evolution strategies in the optimisation
of system models, Journal of Agricultural Systems 60 (2) (1999) 113122.
[22] T. Back, U. Hammel, H.-P. Schwefel, Evolutionary computation: Comments on the history and current state, IEEE Transactions on
Evolutionary Computation 1 (1) (1997) 317.
[23] J.J. Grefenstette, Proportional selection and sampling algorithms, in: T. Back et al. (Eds.), Handbook of Evolutionary Computation (Part C,
Section 2.2), Institute of Physics Publishing, Oxford University Press, New York, 1997, pp. 16.
[24] J.E. Baker, Reducing bias and inefficiency in the selection algorithm, in: Proc. Second International Conference on Genetic Algorithm,
Massachusetts Institute of Technology, MA, 1987, pp. 1421.
[25] T. Blickle, Tournament selection, in: T. Back et al. (Eds.), Hand Book of Evolutionary Computation (Part C, Section 2.3), Institute of Physics
Publishing. Oxford University, New York, 1997, pp. 111.
[26] D.E. Goldberg, K. Deb, A Comparative analysis of selection schemes used in genetic algorithms, in: G.J. Rawlins (Ed.), Foundations of
Genetic Algorithms, Morgan Kaufman Publishers, San Mateo, California, 1991, pp. 6993.
[27] J. Hessner, R. Manner, Choosing optimal mutation rates, in: P. Schwefel, R. Manner (Eds.), Proc. First Workshop on Parallel Problem Solving
from Nature, Springer-Verlag, Berlin, 1991, pp. 2331.
[28] U.K. Maheepala, A.K. Takyi, B.J.C. Perera, Hydrological data monitoring for urban stormwater drainage systems, Journal of Hydrology 245
(2001) 3247.
[29] S. Mardle, S. Pascoe, An overview of genetic algorithms for the solution of optimisation problems, Computers in Higher Education Economics
Review 13 (1) (1999) University of Portsmouth.
[30] T. Back, A Users Guide to GENEsYs 1.0, Department of Computer Science, University of Dortmund, USA, 1992.
[31] J.J. Grefenstette, A Users Guide to GENESIS (Version 5.0), http://www.aic.nrl.navy.mil/pub/galist/src/genesis.tar.z, 1995.
[32] N.R. Siriwardene, Parameter estimation of urban drainage models, MEng Thesis, School of Architectural, Civil and Mechanical Engineering,
Victoria University, Australia, 2003.
[33] Q.J. Wang, The Genetic algorithms and its application to calibrating conceptual rainfall-runoff models, Water Resources Research 27 (9)
(1991) 24672471.
[34] M. Franchini, Use of a genetic algorithm combined with a local search method for the automatic calibration of conceptual rainfall-runoff
models, Hydrological Sciences 42 (3) (1996) 357379.
[35] S. Mohan, Parameter estimation on non-linear Muskingum models using genetic algorithm, ASCE Journal of Hydraulic Engineering 123 (8)
(1997) 137142.
[36] A.W.M. Ng, B.J.C. Perera, Selection of genetic algorithm operators for river water quality model calibration, Engineering Applications of
Artificial Intelligence 16 (2003) 529541.
[37] D.E. Goldberg, Optimal Initial Population Size for Binary-coded Genetic Algorithms, Department of Engineering Mechanics, University of
Alabama, USA, 1985.