This article was downloaded by: [University Library Utrecht]

On: 08 October 2013, At: 01:22

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Engineering Optimization
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/geno20

Flower pollination algorithm: A

novel approach for multiobjective
optimization
a a b
Xin-She Yang , Mehmet Karamanoglu & Xingshi He
a
School of Science and Technology, Middlesex University, The
Burroughs, London NW4 4BT, UK
b
School of Science, Xi'an Polytechnic University, Xi'an, PR China
Published online: 04 Oct 2013.

To cite this article: Xin-She Yang, Mehmet Karamanoglu & Xingshi He , Engineering Optimization
(2013): Flower pollination algorithm: A novel approach for multiobjective optimization, Engineering
Optimization, DOI: 10.1080/0305215X.2013.832237

To link to this article: http://dx.doi.org/10.1080/0305215X.2013.832237

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
 Engineering Optimization, 2013
http://dx.doi.org/10.1080/0305215X.2013.832237

Flower pollination algorithm: A novel approach for

multiobjective optimization
Xin-She Yanga∗ , Mehmet Karamanoglua and Xingshi Heb
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

a School of Science and Technology, Middlesex University, The Burroughs, London NW4 4BT, UK;
b School of Science, Xi’an Polytechnic University, Xi’an, PR China

(Received 18 April 2013; accepted 10 July 2013)

Multiobjective design optimization problems require multiobjective optimization techniques to solve, and
it is often very challenging to obtain high-quality Pareto fronts accurately. In this article, the recently
developed flower pollination algorithm (FPA) is extended to solve multiobjective optimization problems.
The proposed method is used to solve a set of multiobjective test functions and two bi-objective design
benchmarks, and a comparison of the proposed algorithm with other algorithms has been made, which
shows that the FPA is efficient with a good convergence rate. Finally, the importance for further parametric
studies and theoretical analysis is highlighted and discussed.

Keywords: algorithm; benchmark; flower pollination algorithm; optimization; metaheuristics

1. Introduction

Real-world design problems in engineering and industry are usually multiobjective or multicri-
teria, and these multiple objectives often conflict with each other, which makes it impossible
to use any single design option without compromise. Common approaches are to provide good
approximations to the true Pareto fronts of the problem of interest so that decision makers can
rank different options, depending on their preferences or their utilities (Abbass and Sarker 2002;
Babu and Gujarathi 2007; Cagnina, Esquivel, and Coello Coello 2008; Deb 1999, 2001; Deb,
Pratap, and Moitra 2000; Reyes-Sierra and Coello Coello 2006). Compared with single objective
optimization, multiobjective optimization has its own additional challenging issues such as time
complexity, inhomogeneity and dimensionality. It is usually more time consuming to obtain the
true Pareto fronts because it is usually required to produce many points on the Pareto front for
good approximations.
In addition, even if accurate solutions on a Pareto front can be obtained, there is still no guarantee
that these solution points will be distributed uniformly on the front. In fact, it is often difficult to
obtain the whole front without any part missing. For single objective optimization, the optimal
solution can often be a single point in the solution space, while for bi-objective optimization, the
Pareto front forms a curve, and for tri-objective cases, it becomes a surface. In fact, a higher-
dimensional problem can have an extremely complex hypersurface as its Pareto front (Madavan

∗ Corresponding author. Email: [email protected]

© 2013 Taylor & Francis

 2 X.-S. Yang et al.

2002; Marler and Arora 2004;Yang 2010a;Yang and Gandomi 2012). Consequently, it is typically
more challenging to solve such high-dimensional problems.
In the current literature of engineering optimization, a class of nature-inspired algorithms have
shown their promising performance and have thus become popular and widely used, and these
algorithms are mostly swarm intelligence based (Coello Coello 1999; Deb et al. 2002; Geem, Kim,
and Loganathan 2001; Geem 2009; Ray and Liew 2002;Yang 2010a, 2010b, 2011a; Gandomi and
Yang 2011; Gandomi et al. 2012). Metaheuristic algorithms such as particle swarm optimization,
harmony search and cuckoo search are among the most popular (Geem 2009; Yang 2010). For
example, harmony search, developed by Zong Woo Geem in 2001 (Geem, Kim, and Loganathan
2001; Geem 2006, 2009), has been applied in many areas such as highly challenging water
distribution networks (Geem 2006) and discrete structural optimization (Lee et al. 2005). Other
algorithms, such as the shuffled frog-leaping algorithm and particle swarm optimizers, have been
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

applied to various optimization problems (Eusuff, Lansey, and Pasha 2006; He, Prempain, and
Wu 2004; Huang 1996). There are many reasons for the popularity of metaheuristic algorithms,
and the flexibility and simplicity of these algorithms certainly contribute to their success.
The main aim of this article is to extend the flower pollination algorithm (FPA), developed
by Xin-She Yang in 2012 (Yang 2012), for single objective optimization to solve multiobjective
optimization, and thus develop a multi-objective flower pollination algorithm (MOFPA). The
rest of this article is organized as follows: Section 2 outlines the basic characteristics of flower
pollination in nature and then introduces in detail the ideas of the flower pollination algorithm.
Section 3 then presents the validation of the FPA by numerical experiments and a few selected
multiobjective benchmarks. Then, in Section 4, two real-world design benchmarks are solved to
design a welded beam and a disc brake, each with two objectives. Finally, some relevant issues
are discussed and conclusions are drawn in Section 5.

2. Flower pollination algorithm

2.1. Characteristics of flower pollination

It is estimated that there are over a quarter of a million types of flowering plants in Nature and
that about 80% of all plant species are flowering species. It still remains a mystery how flowering
plants came to dominate the landscape from the Cretaceous period (Walker 2009). Flowering
plants have been evolving for more than 125 million years and flowers have become so influential
in evolution that it is unimaginable what the plant world would look like without flowers. The
main purpose of a flower is ultimately reproduction via pollination. Flower pollination is typically
associated with the transfer of pollen, and such transfer is often linked with pollinators such as
insects, birds, bats and other animals. In fact, some flowers and insects have co-evolved into a
very specialized flower-pollinator partnership. For example, some flowers can only attract and
can only depend on a specific species of insects or birds for successful pollination.
Pollination can take two major forms: abiotic and biotic. About 90% of flowering plants depend
on biotic pollination. That is, pollen is transferred by pollinators such as insects and animals. About
10% of pollination takes the abiotic form, which does not require any pollinators. Wind and diffu-
sion help the pollination of such flowering plants, and grass is a good example of abiotic pollination
(Glover 2007). Pollinators, sometimes called pollen vectors, can be very diverse. It is estimated
there are at least 200,000 varieties of pollinators such as insects, bats and birds. Honeybees are
a good example of pollinators, and they have also developed so-called flower constancy. That
is, these pollinators tend to visit certain flower species exclusively while bypassing other flower
species. Such flower constancy may have evolutionary advantages because this will maximize the
transfer of flower pollen to the same or conspecific plants, thus maximizing the reproduction of the
 Engineering Optimization 3

same flower species. Such flower constancy may be advantageous for pollinators as well, because
they can be sure that a supply of nectar is available with their limited memory and minimum cost
of learning, switching or exploring. Rather than focusing on some unpredictable but potentially
more rewarding new flower species, flower constancy may require minimum investment cost and
a more likely guaranteed intake of nectar (Waser 1986).
Pollination can be achieved by self-pollination or cross-pollination. Cross-pollination, or
allogamy, means pollination can occur from the pollen of a flower of a different plant, while
self-pollination is the fertilization of one flower, such as peach flowers, from pollen of the same
flower or different flowers of the same plant, which often occurs when there is no reliable polli-
nator available. Biotic cross-pollination may occur over long distances, and pollinators such as
bees, bats, birds and flies can fly long distances, thus this can be considered as global pollination.
In addition, bees and birds may behave with Lévy flight behaviour, i.e. with jumping or flying
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

distance steps obeying a Lévy distribution (Pavlyukevich 2007). Furthermore, flower constancy
can be considered as an incremental stepping process using the similarity or difference of two
flowers.
From the biological evolution point of view, the objective of flower pollination is the survival
of the fittest and the optimal reproduction of plants in terms of numbers as well as the fittest. This
can be considered as a plant species optimization process. All the above factors and processes of
flower pollination interact so as to achieve the optimal reproduction of flowering plants. Therefore,
this may motivate us to design new optimization algorithms.

2.2. Flower pollination algorithm

The flower pollination algorithm (FPA) — see Figure 1 — was developed by Xin-She Yang in
2012 (Yang 2012), inspired by the flow pollination process of flowering plants. The FPA has been
extended to multi-objective optimization (Yang, Karamanoglu, and He 2013). For simplicity, the
following four rules are used.

(1) Biotic cross-pollination can be considered as a process of global pollination, and pollen-
carrying pollinators move in a way that obeys Lévy flights (Rule 1).
(2) For local pollination, abiotic pollination and self-pollination are used (Rule 2).
(3) Pollinators such as insects can develop flower constancy, which is equivalent to a reproduction
probability that is proportional to the similarity of two flowers involved (Rule 3).
(4) The interaction or switching of local pollination and global pollination can be controlled by
a switch probability p ∈ [0, 1], slightly biased towards local pollination (Rule 4).

In order to formulate the updating formulas, the above rules have to be converted into proper
updating equations. For example, in the global pollination step, flower pollen gametes are carried
by pollinators such as insects, and pollen can travel over a long distance because insects can often
fly and move over a much longer range. Therefore, Rule 1 and flower constancy (Rule 3) can be
represented mathematically as

xt+1
i = xti + γ L(λ)(g∗ − xti ), (1)

where xti is the pollen i or solution vector xi at iteration t, and g∗ is the current best solution found
among all solutions at the current generation/iteration. Here γ is a scaling factor to control the
step size.
Here L(λ) is the parameter, more specifically the Lévy-flights-based step size, that corresponds
to the strength of the pollination. Since insects may move over a long distance with various
distance steps, a Lévy flight can be used to mimic this characteristic efficiently. That is, L > 0 is
 4 X.-S. Yang et al.
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 1. Pseudo code of the proposed flower pollination algorithm (FPA).

drawn from a Lévy distribution

λ(λ) sin(π λ/2) 1

L∼ , (s  s0 > 0). (2)
π s1+λ

Here (λ) is the standard gamma function, and this distribution is valid for large steps s > 0. In
theory, it is required that |s0 |  0, but in practice s0 can be as small as 0.1. However, it is not
trivial to generate pseudo-random step sizes that correctly obey this Lévy distribution (2). There
are a few methods for drawing such random numbers, and the most efficient one from our studies
is the so-called Mantegna algorithm for drawing step size s by using two Gaussian distributions
U and V by the following transformation (Mantegna 1994):

U
s= , U ∼ N(0, σ 2 ), V ∼ N(0, 1). (3)
|V |1/λ

Here U ∼ (0, σ 2 ) means that the samples are drawn from a Gaussian normal distribution with a
zero mean and a variance of σ 2 . The variance can be calculated by
 1/λ
(1 + λ) sin(π λ/2)
σ2 = · (λ−1)/2 . (4)
λ[(1 + λ)/2] 2

This formula looks complicated, but it is just a constant for a given λ. For example, when λ = 1,
the gamma functions become (1 + λ) = 1, [(1 + λ)/2] = 1 and
 1/1
1 sin(π × 1/2)
σ =
2
· = 1. (5)
1×1 20

It has been proved mathematically that the Mantegna algorithm can produce random samples
that obey the required distribution (2) correctly (Mantegna 1994). By using this pseudo-random
number algorithm, 50 step sizes have been drawn to form a consecutive 50 steps of Lévy flights
as shown in Figure 2.
 Engineering Optimization 5
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 2. A series of 50 consecutive steps of Lévy flights.

For local pollination, both Rules 2 and 3 can be represented as

xt+1
i = xti + (xtj − xtk ), (6)

where xtj and xtk are pollen from different flowers of the same plant species. This essentially mimics
flower constancy in a limited neighbourhood. Mathematically, if xtj and xtk come from the same
species or are selected from the same population; this equivalently becomes a local random walk
if  is drawn from a uniform distribution in [0, 1].
In principle, flower pollination activities can occur at all scales, both local and global. But in
reality, adjacent flower patches or flowers in the not-so-far-away neighbourhood are more likely
to be pollinated by local flower pollen than those far away. In order to mimic this feature, a switch
probability (Rule 4) or proximity probability p can be effectively used to switch between common
global pollination to intensive local pollination. To start with, a naive value of p = 0.5 may be
used as an initially value. A preliminary parametric study showed that p = 0.8 may work better
for most applications.

2.3. Multiobjective flower pollination algorithm (MOFPA)

A multiobjective optimization problem with m objectives can be written in general as

Minimize f1 (x), f2 (x), . . . , fm (x), (7)

subject to the nonlinear equality and inequality constraints

hj (x) = 0, (j = 1, 2, . . . , J), (8)

gk (x) ≤ 0, (k = 1, 2, . . . , K). (9)

In order to use the techniques for single objective optimization or extend the methods for solving
multiobjective problems, there are different approaches to achieve this. One of the simplest ways
 6 X.-S. Yang et al.

is to use a weighted sum to combine multiple objectives into a composite single objective:


m
f = w i fi , (10)
i=1

with

m
wi = 1, wi > 0, (11)
i=1

where m is the number of objectives and wi (i = 1, . . . , m) are non-negative weights.

The fundamental idea of this weighted sum approach is that these weighting coefficients act
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

as the preferences for these multiobjectives. For a given set (w1 , w2 , . . . , wm ), the optimization
process will produce a single point of the Pareto front of the problem. For a different set of wi ,
another point on the Pareto front can be generated. With a sufficiently large number of combi-
nations of weights, a good approximation to the true Pareto front can be obtained. It is has been
proved that the solutions to the problem with the combined objective (10) are Pareto optimal if
the weights are positive for all the objectives, and these are also Pareto optimal to the original
problem (7) (Miettinen 1999; Deb 2001). In practice, a set of random numbers ui are first drawn
for a uniform distribution U(0, 1). Then, the weights wi can be calculated by normalization. That
is
ui
wi = m , (12)
i=1 ui

so that i wi = 1 can be satisfied. For example, for three objectives f1 , f2 and f3 , three random
numbers/weights can be drawn from a uniform distribution [0, 1], and they  may be u1 = 0.2915,
u2 = 0.9147 and u3 = 0.6821 in one instance  of sampling runs. Then, i = 1.8883, and w1 =
0.1544, w2 = 0.4844, w3 = 0.3612. Indeed, i wi = 1.000 is satisfied.
In order to obtain the Pareto front accurately with solutions relatively uniformly distributed on
the front, random weights wi should be used, which should be as different as possible (Miettinen
1999). From the benchmarks that have been tested, the weighted sum with random weights usually
works well as can be seen below.

3. Validation and numerical experiments

There are many different test functions for multiobjective optimization (Zitzler and Thiele 1999;
Zitzler, Deb, and Thiele 2000; Zhang et al. 2009), but a subset of some widely used functions
provides a wide range of diverse properties in terms of the Pareto front and the Pareto optimal
set. To validate the proposed MOFPA, a subset of these functions with convex, non-convex and
discontinuous Pareto fronts has been selected, including seven single objective test functions and
four multiobjective test functions, and two bi-objective design problems.

3.1. Single objective test functions

Before proceeding to solve multiobjective optimization problems, the algorithm should first be
validated by solving some well-known single objective test functions. There are at least a hundred
well-known test functions. However, there is no agreed set of test functions for validating new
algorithms, though some reviews and literature do exist (Ackley 1987; Floudas et al. 1999; Hedar
2013; Yang 2010a). Here, a subset of seven test functions with diverse properties is used.
 Engineering Optimization 7

The Ackley function can be written as

⎧  ⎫  d 
⎨ 1 1  d ⎬ 1 
f1 (x) = −20 exp − xi2 − exp cos(2π xi ) + 20 + e, (13)
⎩ 5 d ⎭ d i=1
i=1

which has a global minimum f∗ = 0 at (0, 0, . . . , 0).

The simplest of De Jong’s functions is the so-called sphere function:


n
f2 (x) = xi2 , −5.12 ≤ xi ≤ 5.12, (14)
i=1
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

whose global minimum is obviously f∗ = 0 at (0, 0, . . . , 0), and is unimodal and convex.
Easom’s function:
 d 

d 
f3 (x) = (−1) d+1
cos(xi ) exp − (xi − π ) ,
2
(15)
i=1 i=1

whose global minimum is f∗ = −1 at x∗ = (π, . . . , π ) within −100 ≤ xi ≤ 100, has many local
minima.
Griewank’s function:
 
1  2 
d d
xi
f4 (x) = xi − cos √ + 1, −600 ≤ xi ≤ 600, (16)
4000 i=1 i=1
i

whose global minimum is f∗ = 0 at x∗ = (0, 0, . . . , 0), is highly multimodal.

Rastrigin’s function:


d
f5 (x) = 10d + [xi2 − 10 cos(2π xi )], −5.12 ≤ xi ≤ 5.12, (17)
i=1

whose global minimum is f∗ = 0 at (0, 0, . . . , 0), is highly multimodal.

Rosenbrock’s function:

d−1
f6 (x) = [(xi − 1)2 + 100(xi+1 − xi2 )2 ], (18)
i=1

has its global minimum f∗ = 0 at x∗ = (1, 1, . . . , 1) in the domain −5 ≤ xi ≤ 5, where i =

1, 2, . . . , d.
Zakharov’s function:
 d 2  d 4
d
1 1
f7 (x) = xi +
2
ixi + ixi , (19)
i=1
2 i=1 2 i=1

has its global minimum f∗ = 0 at (0, 0, . . . , 0).

In order to compare the performance of the FPA with other existing algorithms, each algorithm
is first tested using the most widely used implementation and parameter settings. For genetic
algorithms (GAs), a crossover rate of pcrossover = 0.95 and a mutation rate of pmutation = 0.05
are used (Holland 1975; Goldberg 1989; Yang 2010a). For particle swarm optimization (PSO),
a version with an inertia weight θ = 0.7 is used, and its two learning parameters β1 = β2 are
 8 X.-S. Yang et al.

Table 1. Parameter values for each algorithm.

PSO n = 25, θ = 0.7, β1 = β2 = 1.5

GA n = 25, pcrossover = 0.95, pmutation = 0.05
FPA n = 25, λ = 1.5, γ = 0.1, p = 0.8
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 3. Convergence rate during iterations. The objective is plotted versus the iteration (left), and the same results
are shown on a logarithmic scale (right).

Table 2. Comparison of algorithm performance: mean values.

Functions GA PSO FPA

f1 8.29e−9 7.12e−12 5.09e−12

f2 6.61e−15 1.18e−24 2.47e−26
f3 −0.9989 −0.9998 −1.0000
f4 5.72e−9 4.69e−9 1.37e−11
f5 2.93e−6 3.44e−6 4.52e−7
f6 8.97e−6 8.21e−8 6.19e−8
f7 8.77e−4 1.58e−4 9.53e−5

set as 1.5 (Kennedy and Eberhart 1995; Yang 2010a). Also, to ensure a fair comparison, the
same population size should be used whenever possible. So n = 25 has been used for all three
algorithms.
To get some insight into the parameter settings of the FPA, a detailed parametric study has been
carried out by varying p from 0.05 to 0.95 with a step increase of 0.05, λ = 1, 1.25, 1.5, 1.75, 1.9
and n = 5, 10, 15, . . . , 50. It has been found that n = 25, p = 0.8, γ = 0.1 and λ = 1.5 work for
most cases. The parameter values used for all three algorithms are summarized in Table 1.
The convergence behaviour of genetic algorithms and PSO during iterations have been well
studied in the literature. For the FPA, various statistical measures can be obtained from a set
of runs. For example, for the Ackley function f1 , the best objective values obtained during each
iteration can be plotted in a simple graph as shown in Figure 3 where a logarithmic plot shows
that the convergence rate is almost exponential, which implies that the proposed algorithm is very
efficient.
For a fixed population size n = 25, this is equal to the total number of function evaluations,
which is 25,000. The best results obtained in terms of the means of the minimum values found
are summarized in Table 2.
 Engineering Optimization 9

3.2. Multiobjective test functions

In the rest of the article, the parameters in the MOFPA are fixed, based on a preliminary parametric
study, and p = 0.8, λ = 1.5, and a scaling factor γ = 0.1 are used. The population size n = 50
and the number of iterations is set to t = 1000. The following four functions will be tested.

• The ZDT1 function with a convex front (Zitzler and Thiele 1999; Zitzler, Deb, and Thiele
2000):

f1 (x) = x1 , f2 (x) = g(1 − f1 /g),

9 di=2 xi
g=1+ , x1 ∈ [0, 1], i = 2, . . . , 30, (20)
d−1
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

where d is the number of dimensions. Pareto optimality is reached when g = 1.

• The ZDT2 function with a non-convex front:
 2
f1
f1 (x) = x1 , f2 (x) = g 1 − ,
g

where g is the same as given in ZDT1.

• The ZDT3 function with a discontinuous front:
  
f1 f1
f1 (x) = x1 , f2 (x) = g 1 − − sin(10π f1 ) ,
g g

where g in functions ZDT2 and ZDT3 is the same as in function ZDT1. In the ZDT3 function,
f1 varies from 0 to 0.852 and f2 from −0.773 to 1.
• The LZ function (Li and Zhang 2009; Zhang and Li 2007):
  
2  jπ 2
f1 = x1 + xj − sin 6π x1 + ,
|J1 | j∈J d
1
  
√ 2  jπ 2
f2 = 1 − x1 + xj − sin 6π x1 + , (21)
|J2 | j∈J d
2

where J1 = {j|j√is odd } and J2 = {j|j is even } where 2 ≤ j ≤ d. This function has a Pareto
front f2 = 1 − f1 with a Pareto set
 
jπ
xj = sin 6π x1 + , j = 2, 3, . . . , d, x1 ∈ [0, 1]. (22)
d

After generating 100 Pareto points by the

√ MOFPA, the Pareto front generated by the MOFPA
is compared with the true front f2 = 1 − f1 of ZDT1 (see Figure 4).
Define the distance or error between the estimated Pareto front PFe to its corresponding true
front PFt as
N
Ef = PFe − PFt 2 = (PFej − PFtj )2 , (23)
j=1

where N is the number of points.

 10 X.-S. Yang et al.
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 4. (a) Pareto front of test function ZDT1, and (b) Pareto front of test function ZDT2.

Table 3. Summary of results.

Functions Errors (1000 iterations) Errors (2500 iterations)

ZDT1 1.1e−6 3.1e−19

ZDT2 2.7e−6 4.4e−10
ZDT3 1.4e−5 7.2e−12
LZ 1.2e−6 2.9e−12

Figure 5. (a) Pareto front of test function ZDT3, and (b) Pareto front of test function LZ.

The variation of convergence rates or the convergence property can be viewed by plotting out
the errors during iterations, as this measure is an absolute measure that depends on the number
of points. Sometimes, it is easier to use a relative measure in terms of the generalized distance


 N
1 
Dg = (PFej − PFtj )2 . (24)
N j=1

The results for all the functions are summarized in Table 3, and the estimated Pareto fronts and
true fronts of other functions are shown in Figures 4 and 5. In all these figures, the vertical axis
is f2 and the horizontal axis is f1 .
 Engineering Optimization 11

Table 4. Comparison of Dg for n = 50 and t = 500 iterations.

Method ZDT1 ZDT2 ZDT3 LZ

VEGA 3.79e−02 2.37e−03 3.29e−01 1.47e−03

NSGA-II 3.33e−02 7.24e−02 1.14e−01 2.77e−02
MODE 5.80e−03 5.50e−03 2.15e−02 3.19e−03
DEMO 1.08e−03 7.55e−04 1.18e−03 1.40e−03
Bees 2.40e−02 1.69e−02 1.91e−01 1.88e−02
SPEA 1.78e−03 1.34e−03 4.75e−02 1.92e−03
MOFPA 7.11e−05 1.24e−05 5.49e−04 7.92e−05

3.3. Analysis of results and comparison

Downloaded by [University Library Utrecht] at 01:22 08 October 2013

In order to compare the performance of the proposed MOFPA with other established multiobjective
algorithms, we have selected a few algorithms with available results from the literature. In cases
where results are not available, the algorithms have been implemented using well-documented
studies and then new results generated using these algorithms. In particular, other methods have
also been used for comparison, including the vector evaluated genetic algorithm (VEGA) (Schaffer
1985), NSGA-II (Deb, Pratap, and Moitra 2000), multiobjective differential evolution (MODE)
(Babu and Gujarathi 2007; Xue 2004), differential evolution for multiobjective optimization
(DEMO) (Robič and Filipič 2005), multiobjective bees algorithms (Bees) (Pham and Ghan-
barzadeh 2007), and the strength Pareto evolutionary algorithm (SPEA) (Deb et al. 2002; Madavan
2002). The performance measures in terms of generalized distance Dg are summarized in Table 4
for all the above major methods.
It is clearly seen from Table 4 that the proposed MOFPA obtained better results for almost all
four cases.

4. Structural design examples

Design optimization, especially the design of structures, has many applications in engineering and
industry. As a result, there are many different benchmarks with detailed studies in the literature
(Kim, Oh, and Lee 1997; Pham and Ghanbarzadeh 2007; Ray and Liew 2002; Rangaiah 2008). In
the rest of this article, the MOFPA will be used to solve two design case studies: design of a beam
and a disc brake (Osyczka and Kundu 1995; Ray and Liew 2002; Gong, Cai, and Zhu 2009).

4.1. Welded beam design

The multiobjective design of a welded beam is a classical benchmark that has been solved by
many researchers (Deb 1999; Ray and Liew 2002). The problem has four design variables: the
width w and length L of the welded area, the depth d and thickness h of the main beam. The
objective is to minimize both the overall fabrication cost and the end deflection δ.
The detailed formulation can be found in Deb (1999), Ray and Liew (2002) and Gong, Cai,
and Zhu (2009). Here the main problem is rewritten as
minimize f1 (x) = 1.104, 71w2 L + 0.048, 11dh(14.0 + L), minimize f2 = δ, (25)
subject to
g1 (x) = w − h ≤ 0,
g2 (x) = δ(x) − 0.25 ≤ 0,
 12 X.-S. Yang et al.
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 6. Pareto front for the bi-objective beam design where the horizontal axis corresponds to cost and the vertical
axis corresponds to deflection.

g3 (x) = τ (x) − 13, 600 ≤ 0,

g4 (x) = σ (x) − 30, 000 ≤ 0, (26)
g5 (x) = 0.104, 71w + 0.048, 11hd(14 + L) − 5.0 ≤ 0,
2

g6 (x) = 0.125 − w ≤ 0,
g7 (x) = 6000 − P(x) ≤ 0,

where
 
504, 000 L
σ (x) = , Q = 6000 14 + ,
hd 2 2
  2 
√ L (w + d)2
D = 21 L 2 + (w + d)2 , J = 2 wL + ,
6 2
65, 856 QD
δ= , β= , (27)
30, 000hd 3 J

6000 αβL
α=√ , τ (x) = + β 2,
α2 +
2wL D
 √ 
dh3 d 30/48
P = 0.614, 23 × 106 1− .
6 28
The simple limits or bounds are 0.1 ≤ L, d ≤ 10 and 0.125 ≤ w, h ≤ 2.0. This design problem has
been solved using the MOFPA. The approximate Pareto front generated by the 50 non-dominated
solutions after 1000 iterations is shown in Figure 6. This is consistent with the results obtained
by others (Ray and Liew 2002; Pham and Ghanbarzadeh 2007).

4.2. Disc brake design

The objectives are to minimize the overall mass and the braking time by choosing optimal design
variables: the inner radius r, the outer radius R of the discs, the engaging force F and the number
 Engineering Optimization 13
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 7. Pareto front of the disc brake design.

of the friction surface s. This is under design constraints such as the torque, pressure, temperature
and length of the brake (Ray and Liew 2002; Pham and Ghanbarzadeh 2007).
This bi-objective design problem can be written as

9.82 × 106 (R2 − r 2 )

Minimize f1 (x) = 4.9 × 10−5 (R2 − r 2 )(s − 1), f2 (x) = , (28)
Fs(R3 − r 3 )

subject to

g1 (x) = 20 − (R − r) ≤ 0,
g2 (x) = 2.5(s + 1) − 30 ≤ 0,
F
g3 (x) = − 0.4 ≤ 0,
3.14(R2 − r 2 )
(29)
2.22 × 10−3 F(R3 − r 3 )
g4 (x) = − 1 ≤ 0,
(R2 − r 2 )2
0.0266Fs(R3 − r 3 )
g5 (x) = 900 − ≤ 0.
(R2 − r 2 )

The simple limits are

55 ≤ r ≤ 80, 75 ≤ R ≤ 110, 1000 ≤ F ≤ 3000, 2 ≤ s ≤ 20. (30)

It is worth pointing out that s is discrete. In general, the MOFPA has to be extended in combination
with constraint handling techniques so as to deal with mixed integer problems efficiently. However,
since there is only one discrete variable, the simplest branch-and-bound method is used here.
In order to see how the proposed MOFPA performs for real-world design problems, the same
problem has also been solved using other available multiobjective algorithms. Fifty solution points
are generated using the MOFPA to form an approximatation to the true Pareto front after 1000
iterations, as shown in Figure 7.
A comparison of the convergence rates is plotted on the logarithmic scales in Figure 8. It can be
seen clearly that the convergence rate of the MOFPA is the highest in an exponentially decreasing
way. This suggests that the MOFPA provides better solutions in a more efficient way.
 14 X.-S. Yang et al.
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Figure 8. Convergence comparison for the disc brake design.

The above results for 11 test functions in total and two design examples suggest that the
MOFPA is a very efficient algorithm for multiobjective optimization. The proposed algorithm can
deal with highly nonlinear, multiobjective optimization problems with complex constraints and
diverse Pareto optimal sets.

5. Discussions and conclusions

Multiobjective optimization in engineering and industry is often very challenging to solve, neces-
sitating sophisticated techniques to tackle. Metaheuristic approaches have shown promise and
popularity in recent years.
In the present work, a new algorithm, called the flower pollination algorithm, has been for-
mulated for multiobjective optimization applications by mimicking the pollination process of
flowering plants. Numerical experiments and design benchmarks have shown that the proposed
algorithm is very efficient with an almost exponential convergence rate, based on comparison of
the FPA with other algorithms for solving multiobjective optimization problems.
It is worth pointing out that mathematical analysis is highly necessary in future work so as to
gain insight into the true working mechanisms of metaheuristic algorithms such as the MOFPA.
The FPA has advantages such as simplicity and flexibility, and in many ways it has some similarity
to cuckoo search and other algorithms with Lévy flights (Yang 2010a, 2011b); however, it is still
not clear that why the FPA works well. In terms of number of parameters, the FPA has only one
key parameter p together with a scaling factor γ , which makes the algorithm easier to implement.
However, the nonlinearity in Lévy flights make it difficult to analyse mathematically. It can be
expected that this nonlinearity in the algorithm formulations may be advantageous in enhancing
the performance of an algorithm. More research may reveal the subtlety of this feature.
For multiobjective optimization, an important issue is how to ensure the solution points can
distribute relatively uniformly on the Pareto front for test functions. However, for real-world
design problems such as the design of a disc brake and a welded beam, the solutions are not quite
uniform on the Pareto fronts, and there is still room for improvement. However, simply generating
more solution points may not solve the Pareto uniformity problem easily. In fact, how to maintain
a uniform spread on the Pareto front is still a challenging problem that requires more study. It may
be useful as a further research topic to study other approaches for multiobjective optimization,
 Engineering Optimization 15

such as the -constraint method, weighted metric methods, Benson’s method, utility methods and
evolutionary methods (Miettinen 1999; Coello Coello 1999; Deb 2001).
On the other hand, further studies could focus on more detailed parametric analysis and gain
insight into how algorithm-dependent parameters can affect the performance of an algorithm.
Furthermore, the linearity in the main updating formulas makes it possible to do some theoretical
analysis in terms of dynamic systems or Markov chain theories, while the nonlinearity in terms
of Lévy flights can make it difficult to analyse the FPA exactly. All these could form useful topics
for further research.

References
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Abbass, H. A., and R. Sarker. 2002. “The Pareto Differential Evolution Algorithm.” International Journal on Artificial
Intelligence Tools 11 (4): 531–552.
Ackley, D. H. 1987. A Connectionist Machine for Genetic Hillclimbing. Wisconsin, Madison: Kluwer Academic.
Babu, B. V., and A. M. Gujarathi. 2007. “Multi-Objective Differential Evolution (MODE) for Optimization of Supply
Chain Planning and Management.” In Proceedings of the 2007 IEEE Congress on Evolutionary Computation (CEC
2007), 25–28 September 2007, Singapore, 2732–2739. Singapore: IEEE Press. doi: 10.1109/CEC.2007.4424816.
Cagnina, L. C., S. C. Esquivel, and Carlos A. Coello Coello. 2008. “Solving Engineering Optimization Problems with the
Simple Constrained Particle Swarm Optimizer.” Informatica 32 (2): 319–326.
Coello Coello, Carlos A. 1999. “An Updated Survey of Evolutionary Multiobjective Optimization Techniques: State of
the Art and Future Trends.” In Proceedings of the 1999 IEEE Congress on Evolutionary Computation (CEC99), 6–9
July 1999, Washington, DC. Washington DC: IEEE Press. doi: 10.1109/CEC.1999.781901.
Deb, K. 1999. “Evolutionary Algorithms for Multi-Criterion Optimization in Engineering design.” In Evolutionary
Aglorithms in Engineering and Computer Science, 135–161. New York: Wiley.
Deb, K. 2001. Multi-Objective Optimization Using Evolutionary Algorithms. New York: Wiley.
Deb, K., A. Pratap, S. Agarwal, and T. Mayarivan. 2002. “A Fast and Elistist Multiobjective Algorithm: NSGA-II.” IEEE
Transactions on Evolutionary Computation 6 (2): 182–197.
Deb, K., A. Pratap, and S. Moitra. 2000.. “Mechanical Component Design for Multiple Objectives Using Elitist Non-
Dominated Sorting GA.” In Proceedings of the Parallel Problem Solving from Nature VI Conference, 16–20
September 2000, Paris, 859–868. Berlin Heidelberg: Springer.
Eusuff, M., K. Lansey, and F. Pasha. 2006. “Shuffled Frog-Leaping Algorithm: A Memetic Metaheuristic for Discrete
Optimization.” Engineering Optimization 38 (2): 129–154.
Floudas, C. A., P. M. Pardalos, C. S. Adjiman, W. R. Esposito, Z. H. Gumus, S. T. Harding, J. L. Klepeis, C. A. Meyer, and
C. A. Scheiger. 1999. Handbook of Test Problems in Local and Global Optimization. Berlin Heidelberg: Springer.
Gandomi,A. H., and X.-S.Yang. 2011. “Benchmark Problems in Structural Optimization.” In Computational Optimization,
Methods and Algorithms, edited by Slawomir Koziel and Xin-She Yang, Vol. 356 of Studies in Computational
Intelligence, 259–281. Berlin Heidelberg: Springer.
Gandomi, A. H., X.-S. Yang, S. Talatahari, and S. Deb. 2012. “Coupled Eagle Strategy and Differential Evolution for
Unconstrained and Constrained Global Optimization.” Computers & Mathematics with Applications 63 (1): 191–200.
Geem, Z. W. 2006. “Optimal Cost Design of Water Distribution Networks Using Harmony Search.” Engineering
Optimization 38 (3): 259–280.
Geem, Z. W. 2009. Music-Inspired Harmony Search Algorithm: Theory and Applications. Heidelberg: Springer.
Geem, Z. W., J. H. Kim, and G. V. Loganathan. 2001. “A New Heuristic Optimization: Harmony Search.” Simulation 76
(2): 60–68.
Glover, B. J. 2007. Understanding Flowers and Flowering: An Integrated Approach. Oxford, UK: Oxford University
Press.
Goldberg, D. E. 1989. Genetic Algorithms in Search, Optimisation and Machine Learning. Reading, MA: Addison-Wesley.
Gong, W. Y., Z. H. Cai, and L. Zhu. 2009. “An Effective Multiobjective Differential Evolution Algorithm for Engineering
Design.” Structural and Multidisciplinary Optimization 38 (2): 137–157.
He, S., E. Prempain, and Q. H. Wu. 2004. “An Improved Particle Swarm Optimizer for Mechanical Design Optimization
Problems.” Engineering Optimization 36 (5): 585–605.
Hedar, A. 2013. “Test function web pages.” Accessed June 1, 2013. http://www-optima.amp.i.kyoto-u.ac.jp /member/
student/hedar/Hedar_files/TestGO_files/Page364.htm.
Holland, J. 1975. Adaptation in Natural and Artificial Systems. Ann Arbor, MI: University of Michigan Press.
Huang, G. H. 1996. “IPWM: An Interval Parameter Water Quality Management Model.” Engineering Optimization 26
(1): 79–103.
Kennedy, J., and R. C. Eberhart. 1995. “Particle Swarm Optimization.” In Vol. 4 of Proceedings of the IEEE International
Conference on Neural Networks, 27 November–1 December 1995, Perth, Western Australia, 1942–1948. Piscataway,
NJ: IEEE Press. doi: 10.1109/ICNN.1995.488968
Kim, J. T., J. W. Oh, and I. W. Lee. 1997. “Multiobjective Optimization of Steel Box Girder Bridge.” In Proceedings of the
7th KAIST-NTU-KU Trilateral Seminar/Workshop on Civil Engineering, Kyoto, 1–3 December 1997. Piscataway,
NJ: IEEE Press.
 16 X.-S. Yang et al.

Lee, K. S., Z. W. Geem, S.-H. Lee, and K.-W. Bae. 2005. “The Harmony Search Heuristic Algorithm for Discrete Structural
Optimization.” Engineering Optimization 37 (7): 663–684.
Li, H., and Q. F. Zhang. 2009. “Multiobjective Optimization Problems with Complicated Pareto Sets, MOEA/D and
NSGA-II.” IEEE Transactions on Evolutionary Computation 13: 284–302.
Madavan, N. K. 2002. “Multiobjective Optimization Using a Pareto Differential Evolution Approach.” In Vol. 2 of
Proceedings of the Congress on Evolutionary Computation (CEC’2002), 1145–1150. Piscataway, NJ: IEEE Press.
Mantegna, R. N. 1994. “Fast, Accurate Algorithm for Numerical Simulation of Lévy Stable Stochastic Process.” Physical
Review E 49: 4677–4683.
Marler, R. T., and J. S. Arora. 2004. “Survey of Multi-Objective Optimization Methods for Engineering.” Structural and
Multidisciplinary Optimization 26: 369–395.
Miettinen, Kaisa M. 1999. Nonlinear Multiobjective Optimization. Vol. 12 of International Series in Operations Research
& Management Science. Boston, MA: Kluwer Academic.
Osyczka, Andrzej, and Sourav Kundu. 1995. “A Genetic Algorithm-Based Multicriteria Optimization Method.” In Pro-
ceedings of the 1st World Congress on Structural and Multidisciplinary Optimization, May 1995, Goslar, Germany,
edited by Niels Olhoff and George I. N. Rozvany, 909–914. Oxford: Elsevier Science.
Downloaded by [University Library Utrecht] at 01:22 08 October 2013

Pavlyukevich, I. 2007. “Lévy Flights, Non-Local Search and Simulated Annealing.” Journal of Computational Physics
226: 1830–1844.
Pham, D. T., and A. Ghanbarzadeh. 2007. “Multi-Objective Optimisation Using the Bees Algorithm.” In 3rd International
Virtual Conference on Intelligent Production Machines and Systems (IPROMS 2007), 2–13 July 2007 Whittles,
Dunbeath, Scotland.
Rangaiah, G. 2008. Multi-Objective Optimization: Techniques and Applications in Chemical Engineering. Singapore:
World Scientific. http://dc247.4shared.com/doc/yhZtUlMZ/preview.html
Ray, L., and K. M. Liew. 2002. “A Swarm Metaphor for Multiobjective Design Optimization.” Engineering Optimization
34 (2): 141–153.
Reyes-Sierra, M., and Carlos A. Coello Coello. 2006. “Multi-Objective Particle Swarm Optimizers: A Survey of the
State-of-the-Art.” International Journal of Computational Intelligence Research 2 (3): 287–308.
Robič, T., and B. Filipič. 2005. “DEMO: Differential Evolution for Multiobjective Optimization.” In Proceedings of
the Third International Conference on Evolutionary Multi-Criterion Optimization (EMO 2005), 9–11 March 2005,
Guanajuato, Mexico, edited by Carlos A. Coello Coello, Arturo Hernández Aguirre and Eckart Zitzler. In Vol. 3410
of Lecture Notes in Computer Science, 520–533. Heidelberg, Germany: Springer.
Schaffer, J. David. 1985. “Multiple Objective Optimization with Vector Evaluated Genetic Algorithms.” In Proceedings of
the 1st International Conference on Genetic Aglorithms, 24–26 July 1985, Carnegie-Mellon University, Pittsburgh,
PA, 93–100. Hillsdale, NJ: L. Erlbaum Associates.
Walker, M. 2009. “How Flowers Conquered the World.” BBC Earth News, 10 July 2009. http://news.bbc.co.uk/
earth/hi/earth_news/newsid_8143000/8143095.stm.
Waser, N. M. 1986. “Flower Constancy: Definition, Cause and Measurement.” The American Naturalist 127 (5): 596–603.
Xue, F., 2004. “Multi-Objective Differential Evolution: Theory and Applications.” PhD diss., Rensselaer Polytechnic
Institute, Troy, NY.
Yang, X.-S. 2010a. Engineering Optimization: An Introduction with Metaheuristic Applications. Hoboken, New Jersey:
Wiley.
Yang, X.-S. 2010b. “A New Metaheuristic Bat-Inspired Algorithm.” In Proceedings of Nature-Inspired Cooperative
Strategies for Optimization (NICSO 2010), edited by J. R. González et al. In Vol. 284 of Studies in Computational
Intelligence, 65–74. Berlin: Springer.
Yang, X.-S. 2011a. “Bat Algorithm for Multi-Objective Optimisation.” International Journal of Bio-Inspired Computation
3 (5): 267–274.
Yang, X.-S. 2011b. “Review of Meta-Heuristics and Generalised Evolutionary Walk Algorithm.” International Journal of
Bio-Inspired Computation 3 (2): 77–84.
Yang, X.-S. 2012. “Flower Pollination Algorithm for Global Optimization.” In Unconventional Computation and Natural
Computation, edited by J. Durand-Lose and N. Jonoska, Vol. 7445 of Lecture Notes in Computer Science, 240–249.
Berlin: Springer.
Yang, X.-S., and A. H. Gandomi. 2012. “Bat Algorithm: A Novel Approach for Global Engineering Optimization.”
Engineering Computations 29 (5): 464–483.
Yang, X.-S., M. Karamanoglu, and X. S. He. 2013. “Multi-Objective Flower Algorithm for Optimization.” Procedia
Computer Science 18: 861–868.
Zhang, Q. F., and H. Li. 2007. “MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition.” IEEE
Transactions on Evolutionary Computation 11 (6): 712–731.
Zhang, Q. F., A. M. Zhou, S. Z. Zhao, P. N. Suganthan, W. Liu, and S. Tiwari. 2009. Multiobjective Optimization Test
Instances for the CEC 2009 Special Session and Competition. Technical Report CES-487. Colchester, UK: University
of Essex.
Zitzler, E., K. Deb, and L. Thiele. 2000. “Comparison of Multiobjective Evolutionary Algorithms: Empirical Results.”
IEEE Transactions on Evolutionary Computation 8 (2): 173–195.
Zitzler, E., and L. Thiele. 1999. “Multiobjective Evolutonary Algorithms: A Comparative Case Study and the Strength
Pareto Approach.” IEEE Transactions on Evolutionary Computation 3 (4): 257–271.