402 © 2020 The Authors Journal of Hydroinformatics | 22.

2 | 2020

Knowledge-based multi-objective genetic algorithms for

the design of water distribution networks
Matthew B. Johns, Edward Keedwell and Dragan Savic

ABSTRACT

Water system design problems are complex and difficult to optimise. It has been demonstrated that Matthew B. Johns (corresponding author)
Edward Keedwell
involving engineering expertise is required to tackle real-world problems. This paper presents two Dragan Savic
College of Engineering, Mathematics and Physical
engineering inspired hybrid evolutionary algorithms (EAs) for the multi-objective design of water Sciences, Harrison Building,
University of Exeter,
distribution networks. The heuristics are developed from traditional design approaches of practicing
North Park Road, Exeter EX4 4QF,
engineers and integrated into the mutation operator of a multi-objective EA. The first engineering UK
E-mail: [email protected]
inspired heuristic is designed to identify hydraulic bottlenecks within the network and eliminate
Dragan Savic
them with a view to speeding up the algorithm’s search to the feasible solution space. The second KWR Water Research Institute,
Groningenhaven 7, 3430 BB Nieuwegein,
heuristic is based on the notion that pipe diameters smoothly transition from large, at the source, The Netherlands
to small at the extremities of the network. The performance of the engineering inspired hybrid EAs is
compared with Non-Dominated Sorting Genetic Algorithm II and assessed on three networks of varying
complexity, two benchmarks and one real-world network. The experiments presented in this paper
demonstrate that the incorporation of engineering expertise can improve EA performance, often
producing superior solutions both in terms of mathematical optimality and also engineering feasibility.
Key words | evolutionary algorithm, knowledge-based heuristic, multi-objective optimisation, water
distribution network design

INTRODUCTION

The use of evolutionary algorithms (EAs) by researchers in resilience (Prasad & Park ), reliability (Lansey ),
the field of hydroinformatics for the design and optimisation environmental impact (Marchi et al. ) and social welfare
of water systems has grown over the past two decades. With (Amit & Ramachandran ), thus making the optimisation
the emergent maturity of the field, an increased focus on of WDNs a truly multi-objective problem. It has been shown
the real-world application has also come. These real-world that the discovery of the globally optimal Pareto fronts
water distribution problems present a much greater chal- for large multi-objective water distribution network (WDN)
lenge due to their drastically increased size, complexity problems is particularly challenging (Marchi et al. ).
and number of objectives to consider. Aside from the stan- In the case of the Battle of the Water Networks II (Marchi
dard considerations, such as cost, adequate water pressure et al. ), several participant researchers utilised domain
and water quality, there are a host of additional performance knowledge and heuristic information to either decrease the
measures that have been suggested in the literature. These size of the search space or locate favourable areas of the sol-
primarily fall into the areas of risk (Murray et al. ), ution space to initialise the search. These knowledge-guided
techniques are generally aimed at achieving near-optimal sol-
This is an Open Access article distributed under the terms of the Creative
utions with the use of limited computational resources,
Commons Attribution Licence (CC BY 4.0), which permits copying,
adaptation and redistribution, provided the original work is properly cited
rather than attempting to find the globally optimal Pareto
(http://creativecommons.org/licenses/by/4.0/). front of a complex problem (Tolson & Shoemaker ;
doi: 10.2166/hydro.2019.106

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 403 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Gibbs et al. , ; Tolson et al. ; Khedr & Tolson diameters is throughout the network. Although a similar
). An important consideration when applying EAs to concept to diameter uniformity (Creaco et al. ), network
real-world problems is the large computational overhead smoothness takes into account flow direction.
incurred when solving complex hydraulic models (Maier The growing body of research in Hydroinformatics,
et al. ). It becomes apparent that there is a need for which focuses on the use of specific domain knowledge
approaches that are capable of finding near-optimal solutions and heuristic information to boost EA performance, has
within the constraints of available computational resources produced many promising results, often outperforming
and in doing so will aid in the effective application of EAs in standard methods on a range of problems. Unlike other
the practical domain (Maier et al. ). Tolson et al. () domains, however, the majority of techniques presented in
have shown that with limited computational resources, high the hydroinformatics literature tend to focus on the use of
quality solutions can be achieved if a significant amount of specific domain knowledge for the initialisation of starting
engineering judgement is used. Marchi et al. () suggest populations and less on the operators such as crossover
that there is always a trade-off between the engineering experi- and mutation. Therefore, it is interesting to explore the
ence and computational resources needed to solve complex impact that integrating engineering knowledge into the
WDN problems. However, they also claim that engineering operators of an EA would have on performance and,
judgement can never be completely avoided. This notion therefore, filling this gap in the body of research. Another
expands beyond hydroinformatics to a wider set of problem observation is that the majority of hydroinformatic knowl-
domains where domain knowledge has been shown to be edge-based EAs discussed in this section has only
an important factor when tackling real-world problems. been applied to single-objective WDN problems with the
Some examples of this can be found in the wider field of engin- exception of Keedwell & Khu () and Bi et al. ().
eering, including aeronautical (Ong & Keane ) and Therefore, exploring the impact knowledge-based operators
mechanical design (Sapuan ). has on a multi-objective EA adds to the body of knowledge.
As previously stated, there is a growing interest in the use This paper presents two hybrid multi-objective genetic
of domain-specific knowledge in the design of WDNs. algorithms (MOGAs) which employ water systems knowledge
Keedwell & Khu () developed a hybrid cellular automa- to increase computational performance and solution optimal-
ton and genetic approach which included a hydraulically ity, both from a mathematical and also a real-world feasibility
based heuristic used in the formulation of initial EA popu- standpoint. The heuristics at the heart of these algorithms
lations. The method was found to be highly effective when have been inspired by the practices of water systems engineers
tested on a set of large-scale real-world networks. The heuristic and implemented in a way as to incur minimal computation
was based on the premise that the diameter of a pipe connected overhead. Unlike the majority of methods presented in the lit-
to a demand node in pressure deficit can be expanded to erature where domain knowledge is used to produce the initial
increase pressure and the diameter of a pipe connected to a population of solutions, the algorithms presented here inte-
node in pressure excess can be decreased to improve network grate domain expertise into the mutation operators of the
cost. Zheng et al. () used knowledge of pipe network top- algorithms, guiding the search towards the feasible solution
ology and a nonlinear programming technique to identify space with a view to improving efficiency and performance.
promising areas of the solution space, subsequently seeding The performance of the algorithms is assessed on a range of
the initial population of a differential evolution (DE) algor- multi-objective WDN problems from the literature.
ithm. Another initialisation method was proposed by Kang
& Lansey () which used pipe flow velocity thresholds to
METHODS
form a set of initial solutions, and Bi et al. () then adapted
this idea and added a heuristic based on the notion that pipe Multi-objective WDN design problem
diameters generally reduce with the distance from the
source. This concept could be expressed as network smooth- There are many considerations to account for when
ness, a measure of how ‘smooth’ the transition of pipe designing a WDN. When applying new optimisation

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 404 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

methodologies to the problem, a common approach is to transition from large to small diameters from the source to
simplify the real-world nature of the problem and consider the extremities of the network. In this case, the objective
a smaller number of elements. The primary consideration is to minimise the number of pipe smoothing violations
is often the allocation of diameters to the pipes in the net- in a candidate network and is given by the following
work with the objective to minimise infrastructure cost. In expression:
addition to cost, the hydraulics of the network must be
considered to ensure that the constraints of the network X
N
f(S1 , . . . , Sn ) ¼ (Si ) (3)
are met. The most fundamental hydraulic constraint is
i¼1
ensuring the head at each demand node meets the problem
requirements. In this paper, the authors introduce a multi-
where the smoothing violations of pipe i is Si and N is the
objective formulation of the least-cost WDN design problem
number of pipes in the network. For example, in the case
(Cheung et al. ), with the addition of network smooth-
where a pipe i violates the smoothing rule Si ¼ 1; otherwise,
ness. The notion of network smoothness was introduced
if the rule is satisfied Si ¼ 0. Pipe smoothing is described in
earlier in this paper. In this formulation, the smoothness
detail in the next section.
of a network is measured by the number of smoothness vio-
To assess the hydraulic performance of a WDN solution,
lations present in the network. An example would be if the
EPANET (Rossman ) is employed. The EPANET
diameter of a given pipe is greater than the diameter of
engine enables the simulation of pressurised pipe networks
the pipe directly upstream, and this is described in more
by solving flow continuity and pipe headloss equations
detail later in the paper. This multi-objective formulation
using the gradient method (Todini & Pilati ).
enables the designer to observe the trade-off between the
hydraulic performance of the network and the infrastructure
cost with the view to making better design decisions. Water system heuristic-based genetic algorithms
The first objective is the total network cost or infrastruc-
ture cost which is given by the following equation: The genetic algorithm (GA; Holland ) has proved to be a
versatile process for solving a large variety of optimisation pro-
X
N blems spanning many fields and disciplines (Haupt & Haupt
f(D1 , . . . , Dn ) ¼ c(Di , Li ) (1) ). The strength of the approach comes from the ability
i¼1
that the GA has to traverse large search spaces, avoiding
local optima and, therefore, can be viewed as a truly global
where c(Di , Li ) is the cost of pipe i with diameter Di and
search technique (Goldberg ). The performance and versa-
length Li and N is the number of pipes in the network.
tility of the GA can be attributed partly to the independence it
This function is to be minimised during the optimisation
has over the problem being undertaken. Although seen as an
process. The second objective is to minimise the total head
asset, this problem independence can have a detrimental
deficit within the network and is given by the following
effect on performance in the case where the algorithm has
expression:
not been tuned enough to solve the problem at hand.
For the problem of WDN design, the GA relies on gen-
X
J
f(Hi , . . . , Hj ) ¼ (Hi ) (2) etic operators such as crossover and mutation to alter the
i¼1 configuration of the network (Mala-Jetmarova et al. ).
These operators, however, are blind to the direct effect any
where the head deficit in the junction i is Hi and J is the changes made to elements of the network have on the
number of junctions present in the network. The third objec- overall performance of the resultant solution. For example,
tive used in this formulation of the optimisation of least-cost from the perspective of the GA, a change in the diameter
WDNs is a measure of network smoothness. A smooth net- of a pipe has no bearing on the hydraulic behaviour of
work is achieved when pipes can be seen to ‘smoothly’ connected elements until the resultant design is evaluated

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 405 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

(e.g., by using EPANET). However, an engineer making the When solving the least-cost WDN design problem, a GA
same change would know that the head at adjacent junc- only has a direct influence over the diameters of pipes in
tions would be affected. The performance of a newly the network, as the length and roughness of the pipes are
created network, therefore, is only known following solution normally fixed parameters of the problem. Therefore, to
decoding and hydraulic simulation. Although this abstraction reduce headloss, the diameter of a pipe must be increased;
is partly why GAs can be applied to many different water however, as stated previously, this increases the cost of the
system design problems, there is definite scope for the inte- pipe and directly conflicts with the objective function
gration of problem-specific knowledge into the approach. which is trying to minimise infrastructure cost.
An important consideration when integrating problem- One of the key characteristics of a WDN is that the
specific knowledge into a GA is computational efficiency. diameters of pipes close to the source have a greater hydrau-
In most cases and particularly in large-scale real-world net- lic influence over the whole system. For example, if a pipe
works, the most computationally demanding operations are close to the source has a small diameter, large amounts of
solution evaluations. In the case of water distribution headloss can be introduced, and subsequently, the down-
design problems, this comes in the form of the hydraulic stream junctions will not receive the required hydraulic
simulations. Therefore, it is important not to incur any head; this can be referred to as a ‘bottleneck’. Figure 1
additional objective function evaluations where possible. shows two versions of a simple WDN. The first contains
Another consideration is the apparent lack of uptake a pipe (second from the source) which is introducing a
and utilisation of techniques, such as EAs, by engineers in large amount of headloss due to its small diameter, thus
the field of WDN design. One likely reason for this is the resulting in the downstream junctions not receiving
solutions produced by such methods are usually only ‘math- enough pressure and, therefore, reporting a head deficit.
ematically feasible’ and not ‘engineering feasible’ which The bottleneck is eliminated by increasing the diameter of
results in the engineer having to manually correct features the offending pipe, hence reducing headloss and increasing
of a solution network to better suit real-world application the subsequent pressure in the downstream junctions. This
and deployment. approach is often applied by water systems engineers
In this section, two separate water system heuristic when designing distribution networks to eliminate hydraulic
methods are described both of which draw upon expert bottlenecks, unlike a standard GA which cannot implement
engineering knowledge and techniques with a view of inte- this simple process as the operators do not have awareness
gration into a GA to improve search performance and of the hydraulic behaviour of the individual parts of the
solution feasibility. The heuristics presented in this paper system during the crossover and mutation stages.
have been developed and refined from earlier work
(Johns et al. , ).

Heuristic 1: targeting hydraulic deficit/surplus

One of the primary constraints of the least-cost WDN design

problem is ensuring that junction head requirements are met
throughout the network. This can be a complex task, as this
constraint is in direct conflict with the primary objective of
minimising cost through the reduction of pipe diameters.
The key issue here is headloss; as a fluid flows through a
pipe, pressure is lost due to friction along the inner surfaces
of the pipe. The Hazen–Williams equation (Williams &
Hazen ) states that headloss is directly influenced by
the length, diameter, roughness and flow rate of a pipe. Figure 1 | Bottleneck identification and elimination example.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 406 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

It is proposed that this method of bottleneck identifi- process results in junctions with a high-pressure head deficit
cation and elimination can be integrated into a GA by having a greater probability of being selected. The following
applying the heuristic directly to a modified version of the equation is used to calculate the probability of a junction
mutation operator. The aim of this operator is to direct being selected (P(i)):
the search of the algorithm to the boundary of the feasible
solution space in an efficient way using hydraulic constraint hi
P(i) ¼ PN (4)
information from prior solution evaluations. As stated j¼1 hj
previously, computational efficiency is an important con-
sideration when applying any rule-based operator into a where hi is the head deficit at junction i and N is the total
standard algorithm. Unlike some other constraint handling number of junctions in the network.
techniques such as repair algorithms, the proposed mutation Once a junction is selected, the heuristic searches
operator will not perform any additional partial or full upstream of that point until a junction with head excess
fitness evaluations. This is achieved by applying the con- is found. In the event of multiple upstream pipes, the
straint-based rule directly to the genotype without evaluating heuristic follows the path of the pipe, which has the highest
the effect this process has on network performance. head deficit of its upstream junction. The pipe immediately
During the evaluation of the solutions in the initial downstream of the identified junction is then changed to a
population, the algorithm records the flow directions of larger diameter. It has been shown that incremental pipe
each pipe. Utilising this information, the pipe and junction diameter changes during mutation are normally beneficial
directly upstream of each junction are logged, facilitating to the search of a GA. This is in contrast to large changes
the identification of pipes that are restricting junction head to network elements that can have a drastic effect on the
downstream. In some instances, an alteration to the network overall solution quality, sometimes for the worse. However,
can result in flow direction changes, the heuristic takes this it was decided that only allowing the operator to make
into account by checking the flow direction each time it single diameter increments would potentially slow the rate
encounters a pipe and updates the network model of search of the algorithm, and therefore, a weighted roulette
accordingly. wheel approach is used to select the new diameter. This is
Figure 2 shows a flow chart representation of the pro- achieved by firstly populating a list of all available pipe
cess employed by the hydraulic bottleneck elimination diameters greater than the diameter of the selected pipe
mutation operator. The modified mutation operator initially and placing them in ascending order. Each diameter is
chooses a junction using a roulette wheel procedure, which then assigned a probability of selection (P(I)) using the
allocates wheel segment sizes using head deficit information following expression:
from the previous hydraulic evaluation of the solution. This
8
> 1
< i ) (i < N)
P(I) ¼ 2 (5)
>
: 1 ) (i ¼ N)
2i1

where i is the list position of the diameter and N is the total

number of available pipe diameters present in the list. This
results in the smaller diameters in the list having the greater
probability of selection and the largest diameters having a
smaller selection probability.
In the event, where a network contains no junctions in
deficit, the modified mutation operator concentrates on
Figure 2 | Flow chart of the hydraulic bottleneck elimination algorithm. reducing network cost by targeting oversized pipes. Firstly,

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 407 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

a junction is selected using a roulette wheel where the seg-

ment size is directly proportional to the amount of head
excess at each junction. Through this process, junctions
with higher head excess have a greater probability of being
selected. The pipe directly upstream of the selected junction
is then mutated to a smaller diameter using a similar
weighted roulette wheel approach to that of the diameter
increasing method described above. In this case, the avail-
able pipe diameters smaller than the diameter of the
selected pipe are placed in a list in descending order.
The probability of a diameter being selected from the list is
dictated by Equation (5) where the smaller the diameter, Figure 3 | Smooth pipe diameter transitions example on the Hanoi network.
the greater the probability of selection.
As stated previously, the modified mutation operator increase network cost and will not add to the hydraulic capa-
requires junction head and pipe flow information to bility of the system, as it will be limited by the smaller pipe
identify potential bottlenecks in the network. Due to this upstream. One additional adverse effect of this arrangement
dependency, mutation must be applied pre-crossover to pre- is that the velocity in the larger pipe will be lower, potentially
vent the requirement to re-evaluate the hydraulic network of leading to high water age which can lead to poor water qual-
resultant solutions. Therefore, mutation precedes crossover ity. A standard GA will inevitably mutate some of these
in order to preserve the hydraulic information gained from inconsistent pipe selections from some solutions, as they
the simulation of the original solution. It should be noted have the corresponding improvement in the cost function
that the heuristic can be invoked multiple times in one with no hydraulic penalty. However, considerable exper-
mutation operation, the frequency being primarily depen- imentation has demonstrated that even in well-optimised
dant on the base mutation rate of the algorithm. solutions following hundreds of thousands of function,
evaluations of a standard GA will still contain significant
Heuristic 2: pipe smoothing numbers of incorrectly sized pipes in larger networks. This
is unsurprising given the stochastic nature of mutation and
The pipe smoothing approach described in this section aims the changing solution landscape. Given a standard mutation
to identify pipes in a network which can be mutated to rate of 2.5%, the mutation operator will only visit this pipe,
increase network smoothness (in terms of progression on average, once every 40 invocations of the mutation oper-
from one diameter to the next) using network topology ator. Once selected, the probability of the operator selecting
information and a heuristic. This is based around the prin- a ‘smooth’ diameter ranges from N  1/N in the best case,
ciple that in gravity-fed WDNs, the diameter of any pipe is where the required diameter is the second largest in the
never greater than the sum of the diameter(s) of the diameter range, to 1/N where the smallest diameter must
upstream pipes connected to the same junction. Networks be selected to adhere to the smoothness constraint. There-
that adhere to this rule can be seen to ‘smoothly’ transition fore, with 15 available diameters, a single ‘non-smooth’
from large to small diameters from the source to the extremi- pipe could be expected to be rectified, on average, once
ties of the network. Figure 3 shows an example of a ‘smooth’ every 43–600 invocations of the mutation operator. How-
solution for the Hanoi problem where the arrows indicate ever, of course, there will potentially be many of these
flow direction. within the network and as the diameters in the solution
This rule is routinely and implicitly applied by engineers change, so there is the potential to create new instances of
when selecting pipe diameters in a network, as it makes little non-smoothness which must also be rectified. Clearly, stan-
sense for a smaller diameter pipe to proceed a larger one in dard random mutation is far from an optimal method to
most situations. A larger pipe downstream will likely meet these constraints.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 408 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

The pipe smoothing mutation operator applies the The pipe smoothing mutation operator randomly
heuristic described above to the genotype without directly selects a pipe to be mutated. The maximum allowable
evaluating the impact this has on the phenotype. The heur- diameter of the current pipe is calculated by taking the
istic employed by the pipe smoothing mutation operator is sum of the diameters of the immediately upstream pipes
developed from the network topology of a specific problem, and subtracting the sum of the diameters of any pipes par-
remaining consistent throughout the algorithm’s search. It is allel to the selected pipe. This is described by the following
the aim of the heuristic to guide the algorithm’s search to expression:
the engineering feasible solution space to locate smoother
WDN designs while maintaining the performance of a ! 0 1
X
U XP
standard GA. The pipe smoothing mutation operator does Dmax
s ¼ Di @ Dj A (6)
i¼1 j¼1
not perform any additional partial or full fitness evaluations,
with pipe flow directions being established during the
evaluation of the initial population of solutions. This was where Dmax
s is the maximum allowable diameter of selected
an important consideration when developing the pipe pipe s, Di is the diameter of upstream pipe i with U being
smoothing genetic algorithm (PSGA), as additional fitness the total number of directly upstream pipes and Dj is the
evaluations would require further hydraulic simulations, diameter of parallel pipe j with P being the total number
increasing algorithm run time. of pipes parallel to the selected pipe.
Figure 4 shows two configurations of parallel pipes Similarly, to the hydraulic deficit approach, the pipe
entering and exiting a junction, the first of which (left) smoothing operator uses a skewed roulette wheel procedure
violates the pipe smoothing rule, as the sum of the down- to select the new pipe diameter. This is achieved by weight-
stream pipe diameters (A and B) is greater than the sum of ing the larger diameters within the maximum allowable size,
the diameters of the upstream pipes (C and D). It is the so that the bigger the diameter, the higher the probability of
goal of the pipe smoothing heuristic to modify the diameters use. A list is first populated of all available pipe diameters
of the downstream pipes so that the sum of the diameters is equal to and less than the maximum allowable diameter of
equal to or less than the sum of the diameters of the the selected pipe. The list is sorted into descending order
upstream pipes, resultant in a configuration which satisfies by diameter and each diameter is then assigned a probability
the pipe smoothing heuristic (right). of selection (P(I)) using the expression detailed in the
previous section (Equation (5)). This process prevents the
heuristic from selecting small diameters on every appli-
cation. With an upper-bound on possible diameters, the
repeated application of a uniform probability of selection
would result in an undersized, hydraulically infeasible net-
work. Upon a diameter being selected, the pipe being
mutated is changed to the selected diameter.
The pipe smoothing mutation operator needs each
decision pipe in the network to be ‘aware’ of the pipes
directly upstream and downstream of it. Making changes
to pipe diameters in a network can sometimes result in
flow reversal in some pipes; hence, it is necessary to swap
upstream and downstream pipes relative to the pipe in
question. Flow direction is recorded after each hydraulic
evaluation of a solution; therefore, to preserve this infor-
mation, the pipe smoothing mutation operator precedes
Figure 4 | Downstream pipe smoothing rule violation (left) and corrected downstream
diameters that satisfy the smoothing constraint (right). the crossover operator.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 409 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Non-Dominated Sorting Genetic Algorithm II employs the hydraulic deficit-based heuristic detailed
above. It was found that when the HMO used heuristic 1
The Non-Dominated Sorting Genetic Algorithm (NSGA-II) exclusively (i.e., without pipe-wise mutation) throughout
(Deb et al. ) is a multi-objective EA which utilises a the evolutionary process, the population would become
fast non-dominated sorting approach which decreases com- stagnant and prematurely converge on a suboptimal sol-
putational complexity compared with other non-dominated ution. Therefore, it was necessary to limit the amount the
sorting approaches. Although well established, NSGA-II is HMO employs the heuristic. The algorithm employs an
still considered a good benchmark algorithm, as it performs HGM presented below to control the application probability
well in a wide range of problem domains. It produces a of the heuristic:
good spread of solutions and converges close to the true
Pareto-optimal front. More recent EAs were considered gc
P(m) ¼ (7)
such as Borg (Hadka & Reed ) and NSGA-III (Deb & gi
Jain ); however, these algorithms are thought to be
not as suitable to this problem formulation. Although the where gi is the initial gradient of the hypervolume curve, gc is
Borg algorithm has shown promise on some large-scale the current gradient of the hypervolume curve and P(m) is
multi-objective WDN network problems, it was found that the probability of HMO employing heuristic 1. The gradient
the performance of NSGA-II was more consistent on a of the hypervolume curve is calculated at the end of
larger range of networks (Wang Qi et al. ). NSGA-III each generation, comparing the current hypervolume
is designed primarily for many objective (3þ) problems value with that 75 generations previous. If heuristic 1 is
and requires reference points to be supplied prior to not utilised, then random pipe mutation is used instead.
execution. NSGA-II forms the base algorithm upon which This method ensures a smooth transition between the use
the two engineering inspired heuristics are applied, but of the heuristic and random pipe mutation as the algorithm’s
they are generic and can be applied with other EAs. search progresses. This additional process ensures that the
engineering inspired heuristic is applied aggressively at the
Multi-objective adaptive locally constrained genetic start of the algorithm’s search, improving solution feasibility,
algorithm but is able to smoothly reduce the influence of the heuristic
as the search progresses and the rate of conversion slows.
The multi-objective adaptive locally constrained genetic
algorithm (MOALCO-GA) applies heuristic 1 described Multi-objective pipe smoothing genetic algorithm
earlier in the paper to target network head deficit and
surplus. The heuristic is applied to a solution through the The multi-objective pipe smoothing genetic algorithm
mutation operator where the probability of the heuristic (MOPS-GA) is based around the principle that in a WDN,
mutation operator is directly driven by the convergence the diameter of a pipe is never greater than the sum of the
rate of the population. The purpose of this operator is to diameter(s) of the pipes directly upstream (heuristic 2).
guide the algorithm’s search to the feasible solution space Networks that obey this rule can be seen to ‘smoothly’ tran-
in a fast and efficient manner utilising hydraulic data from sition from large to small diameters from the source to the
previous fitness evaluations. MOALCO-GA is essentially extremities of the network. The heuristic is applied to a sol-
NSGA-II but with some additional features; these include ution through the mutation operator where the probability
a heuristic-based mutation operator (HMO) and a hypervo- of the heuristic being applied is defined by a pre-
lume gradient monitor (HGM). The HMO is designed to set algorithm parameter, in this case, 50% probability of
guide the algorithm to feasible network designs earlier in use (random pipe mutation otherwise). It is the aim of the
the optimisation. It can be configured for use with any heuristic to direct the algorithm’s search to the engineering
appropriate objectives, but here the application to network feasible solution space to locate smoother WDN designs
hydraulic performance only is considered. MOALCO-GA while maintaining the performance of a standard MOGA.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 410 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Experimental setup velocity constraints imposed. The Modena WDN (Bragalli

et al. ) is a representation of the water supply system
This section provides details of the experimental setup of the city of Modena, Italy. The network consists of
including benchmark WDN selection, problem formulation 4 sources and 317 decision pipes with 13 available pipe
and performance evaluation. diameters to choose from resulting in a search space of
3.63 × 10352. The formulation of this problem includes both
Benchmark networks junction minimum head requirements and pipe flow velocity
constraints. Network B (Keedwell & Khu ) is a real-life
The following WDN design problems were selected from industrial WDN consisting of a single source reservoir with
the literature to assess the performance of the algorithms 1,277 decision pipes and 26 available pipe diameters with a
presented in this work. The problems range in size and search space of 1.42 × 101806. The problem has fixed mini-
complexity from a single source network with 34 decision vari- mum head requirements at all junctions in the network but
ables to a multi-reservoir, quad source network with 317 does not have any restrictions on the velocity of pipe flow.
decision variables. Also included in this set of benchmark pro-
blems is one large real-world network, Network B, with 1,277 Measuring performance
decision variables. All the following benchmark networks are
least-cost WDN design problems where the goal is to reduce To enable the comparison of MOALCO-GA, MOPS-GA and
network cost through the selection of pipe diameters while NSGA-II, the hypervolume indicator (Zitzler & Thiele ;
satisfying the hydraulic constraints set by the problem. The Bader et al. ) was employed. The hypervolume indicator
selection of a range of different network types was important allows the observation of algorithm convergence and pro-
to enable the evaluation of the hybrid algorithms. vides a measurement of population diversity. Note that the
Figure 5 shows the network layout diagrams for the hypervolume values are scaled from 0 to 1 using the theor-
WDN problems on test. The Hanoi problem (Fujiwara & etical best (utopia) and worst (nadir) points in the solution
Khang ) is a representation of a single source network space. Each of the three algorithms was run 50 times
consisting of 3 loops, 34 decision pipes and 6 available (10 times for Network B due to problem complexity and
pipe diameters with a resultant search space of 2.86 × 1026. resultant runtime). The hypervolume results were averaged
Based upon the trunk main layout for the city of Hanoi, to allow a fair performance comparison to be carried out.
Vietnam, the problem requires that a minimum fixed head In addition, the population hypervolume values produced
of 30 m is reached at all nodes in the network. In this by each algorithm were compared for statistical significance
implementation of the problem, there are no pipe flow using the Mann–Whitney U-test.

Figure 5 | Layout diagrams of the (a) Hanoi, (b) Modena and (c) Network B networks.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 411 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

RESULTS AND DISCUSSION Table 2 | Best and average hypervolume results for the Hanoi problem – NSGA-II,
MOALCO-GA and MOPS-GA comparison

Two formulations of the multi-objective WDN design Algorithm Best hypervolume Average hypervolume

problem are presented, including a novel formulation NSGA-II 0.7282 0.7201

which involves the use of a network smoothing objective. MOALCO-GA 0.7306 0.7248
To assess the performance of the engineering inspired heur- MOPS-GA 0.7441 0.7395
istics in the multi-objective domain, the newly presented
algorithms are directly compared with the standard formu- Utilising the Mann–Whitney U-test, it was found that each
lation of NSGA-II on all benchmark problems. The first algorithm, in this case, produced statistically different
experiment presented in this section is the dual-objective for- populations (p < 0.05) when compared with the other.
mulation of the WDN design problem which uses the first Figure 6 shows the average hypervolume from all
two objectives stated above, total network cost and total 50 runs for the three algorithms for the Hanoi problem.
head deficit. The final experiment in this section involves It can be seen that MOALCO-GA outperforms NSGA-II
the addition of the pipe smoothing violations objective. in the first ∼5,000 evaluations; however, at this point,
MOALCO-GA starts to converge and produces similar
Dual-objective experiments quality results to NSGA-II, while MOPS-GA goes on to sub-
stantially outperform the other two algorithms until the
This section presents the results for the dual-objective termination of the runs. It is only after 20,000 evaluations
experimentation conducted on NSGA-II, MOALCO-GA that MOALCO-GA starts to achieve better results than
and MOPS-GA. As stated previously, the two objectives NSGA-II. This behaviour is thought to be caused by the
are the minimisation of network cost and the minimisation change in heuristic application strength; increasing the
of the hydraulic deficit. To ensure a fair comparison between probability that standard mutation would be utilised instead
the three algorithms, the parameters of NSGA-II were tuned of the deficit/excess heuristic. It would seem that this shift
to each problem and the same parameter set was utilised by enabled the algorithm to explore the solution space in the
each algorithm. Table 1 gives details of these parameters for later stages of the search more effectively than NSGA-II.
each problem. Figure 7 presents the best (highest hypervolume) popu-
lations for the three algorithms. It can be observed that
Hanoi the solutions produced by both MOALCO-GA and MOPS-
GA mostly dominate the solutions found by NSGA-II,
The following set of results is from the dual-objective Hanoi especially at lower network costs. It is not surprising that
problem. Table 2 presents the best achieved hypervolume MOPS-GA achieves more dominant solutions at lower
and mean hypervolume from the 50 individual runs. These network costs as the pipe smoothing heuristic naturally
results show that both MOALCO-GA and MOPS-GA restricts the selection of larger pipe diameters; hence, the
achieve a better best hypervolume and average hypervolume algorithm promotes lower cost solutions.
than NSGA-II. It is also clear that out of the two newly
proposed algorithms, MOPS-GA produces superior results. Modena

Table 1 | Experimental parameters for problems on test The best and mean hypervolume results for the Modena pro-
blem are presented in Table 3. This shows that MOPS-GA
Tournament Pipe mutation
Problem Runs Evaluations Pop size size probability
attains a much higher hypervolume value than the other
two algorithms, which both achieve similar quality sol-
Hanoi 50 100,000 100 4 0.147
utions. In the case of these results, statistical testing
Modena 50 100,000 100 5 0.132
reveals no significant difference in the population of results
Network B 10 100,000 100 8 0.0023
between NSGA-II and MOALCO-GA; however, MOPS-GA

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 412 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 6 | Mean best hypervolume for the Hanoi problem – NSGA-II, MOALCO-GA and MOPS-GA comparison – 100,000 evaluations (left) and 5,000 evaluations (right).

Figure 7 | Pareto front for the Hanoi problem – NSGA-II, MOALCO-GA and MOPS-GA comparison – entire front (left) and zoomed front (right).

Table 3 | Best and average hypervolume results for the Modena problem – NSGA-II, The performance difference between MOPS-GA and the
MOALCO-GA and MOPS-GA comparison
other two algorithms is illustrated in Figure 8. MOPS-GA
Algorithm Best hypervolume Average hypervolume outperforms the other two algorithms significantly through-
out the entire search, ultimately achieving a much higher
NSGA-II 0.7691 0.7268
average hypervolume than NSGA-II and MOALCO-GA.
MOALCO-GA 0.7664 0.7194
MOALCO-GA does display better perform than NSGA-II
MOPS-GA 0.8414 0.8051
up until around 80,000 evaluations.
Figure 9 shows the best performing populations for the
does produce a population of results which are statistically three algorithms for the Modena problem. It is clear from
different from the other two algorithms. these results that MOPS-GA achieves much lower network

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 413 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 8 | Mean best hypervolume for the Modena problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

Figure 9 | Pareto front for the Modena problem – NSGA-II, MOALCO-GA and MOPS-GA comparison – entire front (left) and zoomed front (right).

cost solutions at zero hydraulic deficits compared with the case of a single source network, introducing a bottleneck
other competing algorithms. close to the source can have an undesirable effect on
These results suggest that the pipe smoothing heuristic hydraulic performance, while a multi-source network is
employed by MOPS-GA is very effective when applied to more resilient. Interestingly, the majority of solutions
a multi-source (reservoir) configuration such as that of the (>95%) has zero hydraulic deficit, with only a small
Modena problem. The nature of the pipe smoothing heuris- number of solutions with a hydraulic deficit. The hydraulic
tic encourages lower cost solutions, and this can sometimes requirements of the Modena problem are very easy to
result in a pipe close to the source being mutated to a small meet and are shown to have a high probability of being
diameter, introducing hydraulic deficit downstream. In the satisfied with a randomly generated solution.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 414 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Table 4 | Best and average hypervolume results for the Network B problem – NSGA-II, The best final populations generated by the three algor-
MOALCO-GA and MOPS-GA comparison
ithms for the Network B problem are shown in Figure 11. It
Algorithm Best hypervolume Average hypervolume can be observed that both MOPS-GA and NSGA-II achieve

NSGA-II 0.9121 0.9050

a relatively comparable spread of results apart from at lower

MOALCO-GA 0.9032 0.8981

network costs where the spread of solutions produced by

MOPS-GA 0.9180 0.9113

MOPS-GA is somewhat superior to NSGA-II.
It is also visible that the solutions produced by MOPS-
GA dominate those of NSGA-II and MOALCO-GA,
Network B
especially by the lower network cost solutions. MOALCO-
GA produces a large number of solutions with much
Table 4 shows the best and average hypervolume values
higher network costs than the other two algorithms, and
achieved from the 10 runs of each algorithm for the Network
in the case of this, the most complex problem from the
B problem. MOPS-GA achieves the best solution quality,
benchmarks tested, this behaviour is more pronounced.
followed by NSGA-II and finally MOALCO-GA. Each
algorithm, in this case, produced statistically different popu-
lations of hypervolume results when compared with each Tri-objective experiments
other.
The mean best hypervolume of the three algorithms The following set of experiments involves the addition of the
for the Network B problem is presented in Figure 10. pipe smoothing violations objective as stated in the Methods
MOALCO-GA displays better performance in the early section. The inclusion of pipe smoothing violations as an
stages of the search compared with the other competing additional objective is intended to aid the algorithm to
algorithms. However, at around 15,000 evaluations, the produce high quality solutions which are not only competi-
progression of the adaptive algorithm slows and the other tive but are also more feasible from the perspective of a
two algorithms overtake. It is also at this point where water systems engineer. As with the dual-objective exper-
MOPS-GA splits from NSGA-II and starts to outperform iments, NSGA-II was tuned to each problem and the same
the standard algorithm going on to achieve a better overall parameter values were used by each algorithm to ensure a
average hypervolume value. fair comparison.

Figure 10 | Mean best hypervolume for the Network B problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 415 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 11 | Pareto front for the Network B problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

Hanoi MOPS-GA obtains the highest best and average hypervo-

lume values, followed by MOALCO-GA which achieves
Table 5 presents the hypervolume results of the three algor- better results than NSGA-II. Each algorithm, in this case,
ithms for the tri-objective Hanoi problem. It can be seen that produced statistically different populations of hypervolume
results when compared with each other.
Table 5 | Best and average hypervolume results for the Hanoi problem – NSGA-II,
MOALCO-GA and MOPS-GA comparison Figure 12 displays the average hypervolume value of the
50 individual runs for each of the three algorithms. It can be
Algorithm Best hypervolume Average hypervolume
observed that both of the engineering heuristic-based algor-
NSGA-II 0.6861 0.6674 ithms display increased performance over NSGA-II in the
MOALCO-GA 0.6970 0.6734 initial stages of the search. Following the preliminary expan-
MOPS-GA 0.7108 0.6886 sion into the search space, both MOALCO-GA and NSGA-II

Figure 12 | Mean best hypervolume for the Hanoi problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 416 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 13 | Pareto front for the Hanoi problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

begin convergence at a faster rate to that of MOPS-GA. This those produced by the other algorithms at low network costs
results in MOPS-GA achieving a superior average hypervo- and high hydraulic deficit values. However, MOALCO-GA
lume value compared with that of the other two algorithms. does appear to achieve dominant solutions at low hydraulic
Figure 13 presents the best final population from each deficit values. Looking at the second plot, it is clear that the
of the algorithms for the Hanoi problem. Due to the tri- solutions produced by MOPS-GA dominate those from the
objective nature of the problem, the solutions are presented other algorithms in terms of pipe smoothing violations and
utilising four plots to increase clarity; three 2D figures network cost, and this behaviour is somewhat expected
display each side of the three-dimensional (3D) search due to the complimentary heuristic employed by the algor-
space and one 3D plot of the same data. It can be observed ithm. Interestingly, MOPS-GA produces the solutions with
that the solutions produced by MOPS-GA tend to dominate the joint highest number of pipe smoothing violations,

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 417 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

although these have a lower cost than solutions generated by objective Modena problem. It is apparent from these results
NSGA-II, which have the same number of violations. The that MOPS-GA is able to generate populations with signifi-
third plot presents the solutions in terms of hydraulic deficit cantly higher hypervolume values than the other two
and pipe smoothing violations. As previously observed, the algorithms. No statistically significant difference in results
solutions found by MOPS-GA are mainly located at low was found between NSGA-II and MOALCO-GA, although
pipe smoothing violations; however, this is at the cost of MOPS-GA produced statistically different results when com-
higher hydraulic deficit values. It can also be observed pared with the other two algorithms.
that the majority of MOPS-GA solutions with zero hydraulic The mean best hypervolume results for the three algor-
deficits has a relatively high number of smoothing viola- ithms for the Modena problem are presented in Figure 14.
tions. Interestingly, it is MOALCO-GA and NSGA-II that It is observed that MOPS-GA drastically outperforms the
achieve solutions with the lowest hydraulic deficit at zero other two algorithms throughout the entire search of the
pipe smoothing violations, mostly dominating the compet- algorithms. While MOALCO-GA does achieve better hyper-
ing solutions found by MOPS-GA. volume results in the early stages of the search compared
with NSGA-II, the difference in performance between the
Modena two algorithms diminishes in the later stages of the search.
It should be noted that it takes under 20,000 evaluations
Table 6 presents the best and average hypervolume results for MOPS-GA to achieve the highest average hypervolume
for NSGA-II, MOALCO-GA and MOPS-GA for the tri- achieved by both MOALCO-GA and NSGA-II.
Figure 15 displays the best final population of solutions
Table 6 | Best and average hypervolume results for the Modena problem – NSGA-II,
MOALCO-GA and MOPS-GA comparison generated by the three algorithms for the Modena pro-
blem. It is apparent from the first plot that MOPS-GA is
Algorithm Best hypervolume Average hypervolume
able to find the lowest cost solutions, followed by the
NSGA-II 0.6000 0.5795 other two algorithms, although this is done at the cost of
MOALCO-GA 0.6117 0.5812 an increased hydraulic deficit. The second plot shows the
MOPS-GA 0.6720 0.6463 ability of MOPS-GA to find a good number of smoother,

Figure 14 | Mean best hypervolume for the Modena problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 418 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 15 | Pareto front for the Modena problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

low cost solutions compared with the other two Network B

algorithms. It can also be observed that although
MOALCO-GA and NSGA-II do achieve solutions with Table 7 presents the best and average hypervolume results
similar network smoothness, they are mainly located at from the three algorithms for the Network B problem. It is
much higher network costs. Being a larger network, it is
more difficult to find solutions with very smooth pipe Table 7 | Best and average hypervolume results for the Network B problem – NSGA-II,
diameter transitions, hence why the lowest number of MOALCO-GA and MOPS-GA comparison

pipe smoothing violations generated by a solution is 60.

Algorithm Best hypervolume Average hypervolume
Looking at the third figure, it is apparent that although
NSGA-II 0.5847 0.5687
MOPS-GA achieves the most solutions with the least
MOALCO-GA 0.5852 0.5673
number of pipe smoothing violations, the majority of
MOPS-GA 0.5895 0.5771
these solutions has relatively high hydraulic deficit.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 419 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 16 | Mean hypervolume for the Network B problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

apparent that both NSGA-II and MOALCO-GA achieve smoothing violations tends to increase. It is also apparent
a similar average population of solutions, reaching compar- that most solutions produced by MOPS-GA dominate
able hypervolume values. MOPS-GA displays the highest those found by the other two algorithms in terms of pipe
average performance, obtaining the best hypervolume values smoothing violations and network cost. The third plot in
of all the algorithms. No statistical significance in the final the figure shows that MOPS-GA is good at finding the
population of results was found between NSGA-II and solutions with lower pipe smoothing violations at relatively
MOALCO-GA; however, MOPS-GA produced statistically sig- low deficit values, often dominating those produced by the
nificant results when compared with the other two algorithms. competing algorithms. Interestingly, it is NSGA-II that
Figure 16 shows the average hypervolume of the three finds a number of solutions with lower pipe smoothing
algorithms for the Network B problem. Interestingly, it is violations but at the cost of a high hydraulic deficit value.
MOALCO-GA that exhibits the best performance in
the early stages of the search, only being surpassed by
MOPS-GA at 20,000 and NSGA-II at the end of the CONCLUSIONS
search. NSGA-II and MOPS-GA display comparable perform-
ance during the first 10,000 evaluations; however, following MOALCO-GA and MOPS-GA have been developed and
this stage, MOPS-GA produces higher quality solutions than assessed on a number of well-known benchmarks from
the standard algorithm for the remainder of the search. the literature and one real-world network. Utilising two
Figure 17 shows the best population produced by each different heuristics, both MOALCO-GA and MOPS-GA
of the three algorithms for the Network B problem. It encode engineering knowledge into the NSGA-II with the
can be observed that the majority of solutions found by view to improving the performance of the algorithm utilising
MOPS-GA dominates those produced by the other two the mutation operator.
algorithms in terms of network cost and hydraulic deficit, MOALCO-GA has been shown to perform relatively
especially at lower network costs; although NSGA-II does well from the experiments presented in this paper when
produce some dominant solutions with large pressure defi- compared with NSGA-II. Regarding the dual-objective
cit. It is also apparent that MOALCO-GA tends to find the experiment set, MOALCO-GA performed well often achiev-
highest cost solutions, generally located at zero hydraulic ing solutions of equal or higher quality than NSGA-II. The
deficits. As network cost is decreased, the number of pipe exception to this is in the case of the large-scale problem,

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 420 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Figure 17 | Pareto front for the Network B problem – NSGA-II, MOALCO-GA and MOPS-GA comparison.

Network B, although it outperformed both NSGA-II and The introduction of a pipe smoothing component into the
MOPS-GA in the first stages of the search. In terms of the multi-objective formulation improves performance in both
tri-objective experimentations, MOALCO-GA was again dual- and tri-objective formulations. While the modified
shown to often outperform NSGA-II in a number of cases algorithm might be expected to perform well in the tri-objec-
and never produced statistically worse results than the tive case where one of the objectives reflects the heuristic, it is
standard algorithm. highly interesting that it should perform so much better on
The pipe smoothing mutation operator of MOPS-GA has the dual-objective problem. This is a key finding as it provides
been shown to outperform the standard configuration of some of the first evidence that incorporating engineering
NSGA-II on all benchmark problems tested in this expertise into an algorithm enables it to improve mathemat-
paper. For the majority of problems tested in this paper, ical optimality in multiple objectives.
MOPS-GA exhibited faster convergence than NSGA-II and The results presented in this paper have demonstrated
achieved a better set of final solutions. The results also that knowledge-guided mutation aids an EA to more
suggest that MOPS-GA performs very well when tackling efficiently solve the multi-objective formulation of the
WDN design problem that involves multiple water sources. WDN design problem. The performance gains from these

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 421 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

knowledge-guided approaches have been shown to be REFERENCES

effective over a range of network scales and complexities.
Out of the two heuristic-based mutation operators, the Amit, R. & Ramachandran, P.  Optimal Design of Water
pipe smoothing method displays the most promise, consist- Distribution Networks: A Review. Selected Works of R K Amit.
ently outperforming the bottleneck reduction process in all http://works.bepress.com/rkamit/7/.
Bader, J., Deb, K. & Zitzler, E.  Faster hypervolume-based
problems and networks. The primary cause is thought to
search using Monte Carlo sampling. In: Conference on
be the ability to apply the pipe smoothing heuristic through- Multiple Criteria Decision Making (MCDM 2008). Springer,
out the entire search, whereas the nature of the bottleneck Auckland, New Zealand, pp. 313–326.
reducing heuristic limits when it can be applied for positive Bi, W., Dandy, G. C. & Maier, H. R.  Improved genetic algorithm
optimization of water distribution system design by incorporating
effect, normally in the early stages of the search. Each
domain knowledge. Environmental Modelling & Software 69,
heuristic has its strengths, although the stage at which appli- 370–381. https://doi.org/10.1016/j.envsoft.2014.09.010.
cation occurs is a key. With this in mind, the combination Bi, W., Dandy, G. C. & Maier, H. R.  Use of domain
knowledge to increase the convergence rate of evolutionary
of these heuristics into one hybrid algorithm would be the
algorithms for optimizing the cost and resilience of water
logical next step in this research path. Another possible distribution systems. Journal of Water Resources Planning
direction for future research is employing a modified version and Management 142 (9), 04016027.
of the pipe smoothing heuristic as a post-process action Bragalli, C., D’Ambrosio, C., Lee, J., Lodi, A. & Toth, P.  On
the optimal design of water distribution networks: a practical
following optimisation. In its current form, the heuristic
MINLP approach. Optimization and Engineering 13 (2),
would most likely have a detrimental impact on the hydrau- 219–246.
lic performance of the network; however, enforcing the Cheung, P. B., Reis, L. F., Formiga, K. T., Chaudhry, F. H. & Ticona,
smoothing violation rule could be viable. If a pipe has a W. G.  Multiobjective evolutionary algorithms applied to the
rehabilitation of a water distribution system: a comparative study.
larger diameter to that of its upstream counterpart(s),
In: International Conference on Evolutionary Multi-Criterion
change the diameter to match. This should mostly sustain Optimization. Springer, Berlin, Heidelberg, pp. 662–676.
the hydraulic performance of the network while smoothing Creaco, E., Franchini, M. & Todini, E.  The combined use of
resilience and loop diameter uniformity as a good indirect
the pipe diameter transitions.
measure of network reliability. Urban Water Journal 13 (2),
In conclusion, both engineering heuristic-based multi- 167–181. https://doi.org/10.1080/1573062X.2014.949799.
objective algorithms presented in this paper were found to Deb, K. & Jain, H.  An evolutionary many-objective
outperform a tuned version of NSGA-II in the vast majority optimization algorithm using reference-point-based
nondominated sorting approach. Part I: solving problems
of cases, with MOPS-GA generally achieving the best
with box constraints. IEEE Transactions on Evolutionary
solutions out of all of the algorithms on a test. This paper Computation 18 (4), 577–601.
has gone some way in demonstrating that the incorporation Deb, K., Agrawal, S., Pratap, A. & Meyarivan, T.  A fast elitist
of water systems knowledge to an EA not only leads to non-dominated sorting genetic algorithm for multi-objective
optimization: NSGA-II. In: International Conference on
improvements in computational efficiency and mathemat-
Parallel Problem Solving from Nature. Springer, Berlin,
ical optimality but also to the generation of solutions Heidelberg, pp. 849–858.
industry engineers would find more intuitive. Fujiwara, O. & Khang, D. B.  A two-phase decomposition
method for optimal design of looped water distribution
networks. Water Resources Research 26 (4), 539–549.
Gibbs, M. S., Dandy, G. C. & Maier, H. R.  A genetic
algorithm calibration method based on convergence due to
ACKNOWLEDGEMENTS
genetic drift. Information Sciences 178 (14), 2857–2869.
Gibbs, M. S., Maier, H. R. & Dandy, G. C.  Using characteristics
The work presented in this paper was developed as part of of the optimisation problem to determine the genetic
the Human–Computer Optimisation for Water Systems algorithm population size when the number of evaluations is
limited. Environmental Modelling & Software 69, 226–239.
Planning and Management (HOWS) project funded by the
Goldberg, D. E.  Genetic Algorithms in Search, Optimization
Engineering and Physical Sciences Research Council and Machine Learning, 1st edn. Addison-Wesley Longman
(EPSRC) – grant no. EP/P009441. Publishing Co., Inc., Boston, MA, USA.

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest
 422 M. B. Johns et al. | Knowledge-based MOGAs for the design of water distribution networks Journal of Hydroinformatics | 22.2 | 2020

Hadka, D. & Reed, P.  Borg: an auto-adaptive many-objective Software. National Homeland Security Research Center, US
evolutionary computing framework. Evolutionary Environmental Protection Agency, Cincinnati, OH, USA.
Computation 21 (2), 231–259. Ong, Y. S. & Keane, A. J.  A domain knowledge based search
Haupt, R. L. & Haupt, S. E.  Practical Genetic Algorithms, advisor for design problem solving environments.
Second Edition. John Wiley & Sons, Hoboken, New Jersey. Engineering Applications of Artificial Intelligence 15 (1),
Holland, J. H.  Adaptation in Natural and Artificial Systems: An 105–116.
Introductory Analysis with Applications to Biology, Control, Prasad, T. D. & Park, N.-S.  Multiobjective genetic algorithms
and Artificial Intelligence. MIT Press, Cambridge, MA. for design of water distribution networks. Journal of Water
Johns, M. B., Keedwell, E. & Savic, D.  Pipe smoothing genetic Resources Planning and Management 130 (1), 73–82. https://
algorithm for least cost water distribution network design. In: doi.org/10.1061/(ASCE)0733-9496(2004)130:1(73).
Proceedings of the 15th Annual Conference on Genetic and Rossman, L. A.  EPANET 2: Users Manual.
Evolutionary Computation. pp. 1309–1316. https://doi.org/ Sapuan, S. M.  A knowledge-based system for materials
10.1145/2463372.2463533. selection in mechanical engineering design. Materials &
Johns, M. B., Keedwell, E. & Savic, D.  Adaptive locally Design 22 (8), 687–695.
constrained genetic algorithm for least-cost water distribution Todini, E. & Pilati, S.  A gradient method for the analysis of
network design. Journal of Hydroinformatics 16 (2), 288–301. pipe networks. Proc., Int. Conf. on Comp. Applications for
https://doi.org/10.2166/hydro.2013.218. Water Supply and Distribution. Leicester Polytechnic,
Kang, D. & Lansey, K.  Revisiting optimal water-distribution Leicester, UK.
system design: issues and a heuristic hierarchical approach. Tolson, B. A. & Shoemaker, C. A.  Dynamically dimensioned
Journal of Water Resources Planning and Management search algorithm for computationally efficient watershed model
138 (3), 208–217. calibration. Water Resources Research 43 (1), January 2007.
Keedwell, E. & Khu, S.-T.  A hybrid genetic algorithm for the Tolson, B. A., Asadzadeh, M., Maier, H. R. & Zecchin, A. 
design of water distribution networks. Engineering Hybrid discrete dynamically dimensioned search (HD-DDS)
Applications of Artificial Intelligence 18 (4), 461–472. algorithm for water distribution system design optimization.
https://doi.org/10.1016/j.engappai.2004.10.001. Water Resources Research 45 (12), W12416.
Keedwell, E. & Khu, S.-T.  A novel evolutionary meta-heuristic for Tolson, B. A., Khedr, A. & Asadzadeh, M.  The battle of the
the multi-objective optimization of real-world water distribution water networks (BWN-II): PADDS based solution approach.
networks. Engineering Optimization 38 (03), 319–333. In: WDSA 2012: 14th Water Distribution Systems Analysis
Khedr, A. & Tolson, B.  Comparing optimization techniques Conference, 24–27 September 2012. Engineers Australia,
with an engineering judgment approach to WDN design. Adelaide, South Australia, p. 291.
Journal of Water Resources Planning and Management Wang, Q., Guidolin, M., Savic, D. & Kapelan, Z.  Two-
142 (5), C4015014. objective design of benchmark problems of a water
Lansey, K.  Sustainable, robust, resilient, water distribution distribution system via MOEAs: towards the best-known
systems. In: WDSA 2012: 14th Water Distribution Systems approximation of the true Pareto front. Journal of Water
Analysis Conference, 24–27 September 2012. Engineers Resources Planning and Management 141 (3), 04014060.
Australia, Adelaide, South Australia, p. 1. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000460.
Maier, H. R., Kapelan, Z., Kasprzyk, J., Kollat, J., Matott, L. S., Williams, G. S. & Hazen, A.  Hydraulic Tables: The Elements
Cunha, M. C. & Marchi, A.  Evolutionary algorithms and of Gagings and the Friction of Water Flowing in Pipes,
other metaheuristics in water resources: current status, Aqueducts, Sewers, etc as Determined by the Hazen and
research challenges and future directions. Environmental Williams Formula and the Flow of Water Over Sharp-Edged
Modelling & Software 62, 271–299. and Irregular Weirs, and the Quantity Discharged, as
Mala-Jetmarova, H., Sultanova, N. & Savic, D.  Lost in Determined by Bazin’s Formula and Experimental
optimisation of water distribution systems? A literature review Investigations Upon Large Models. John Wiley & Sons, New
of system operation. Environmental Modelling & Software 93, York, NY.
209–254. https://doi.org/10.1016/j.envsoft.2017.02.009. Zheng, F., Simpson, A. R. & Zecchin, A. C.  A combined NLP-
Marchi, A., Salomons, E., Ostfeld, A., Kapelan, Z., Simpson, A. R., differential evolution algorithm approach for the
Zecchin, A. C. & Song, Y.  Battle of the water networks optimization of looped water distribution systems. Water
II. Journal of Water Resources Planning and Management Resources Research 47 (8), W08531.
140 (7), 04014009. Zitzler, E. & Thiele, L.  Multiobjective optimization using
Murray, R., Haxton, T., Janke, R., Hart, W. E., Berry, J. & Phillips, evolutionary algorithms – a comparative case study. In:
C.  Sensor Network Design for Drinking Water International Conference on Parallel Problem Solving from
Contamination Warning Systems: a Compendium of Nature – PPSN V, Amsterdam, The Netherlands, September
Research Results and Case Studies Using the TEVA-SPOT 27–30 1998.

First received 11 June 2019; accepted in revised form 5 November 2019. Available online 29 November 2019

Downloaded from http://iwaponline.com/jh/article-pdf/22/2/402/666274/jh0220402.pdf

by guest