Académique Documents
Professionnel Documents
Culture Documents
Iztok Fister
Iztok Fister Jr. Editors
Adaptation and
Hybridization in
Computational
Intelligence
Adaptation, Learning, and Optimization
Volume 18
Series editors
Meng-Hiot Lim, Nanyang Technological University, Singapore
e-mail: emhlim@ntu.edu.sg
Yew-Soon Ong, Nanyang Technological University, Singapore
e-mail: asysong@ntu.edu.sg
About this Series
The role of adaptation, learning and optimization are becoming increasingly essen-
tial and intertwined. The capability of a system to adapt either through modification
of its physiological structure or via some revalidation process of internal mecha-
nisms that directly dictate the response or behavior is crucial in many real world
applications. Optimization lies at the heart of most machine learning approaches
while learning and optimization are two primary means to effect adaptation in var-
ious forms. They usually involve computational processes incorporated within the
system that trigger parametric updating and knowledge or model enhancement, giv-
ing rise to progressive improvement. This book series serves as a channel to con-
solidate work related to topics linked to adaptation, learning and optimization in
systems and structures. Topics covered under this series include:
complex adaptive systems including evolutionary computation, memetic com-
puting, swarm intelligence, neural networks, fuzzy systems, tabu search, sim-
ulated annealing, etc.
machine learning, data mining & mathematical programming
hybridization of techniques that span across artificial intelligence and compu-
tational intelligence for synergistic alliance of strategies for problem-solving.
aspects of adaptation in robotics
agent-based computing
autonomic/pervasive computing
dynamic optimization/learning in noisy and uncertain environment
systemic alliance of stochastic and conventional search techniques
all aspects of adaptations in man-machine systems.
This book series bridges the dichotomy of modern and conventional mathematical
and heuristic/meta-heuristics approaches to bring about effective adaptation, learn-
ing and optimization. It propels the maxim that the old and the new can come to-
gether and be combined synergistically to scale new heights in problem-solving. To
reach such a level, numerous research issues will emerge and researchers will find
the book series a convenient medium to track the progresses made.
ABC
Editors
Iztok Fister Iztok Fister Jr.
Faculty of Electrical Engineering and Faculty of Electrical Engineering and
Computer Science Computer Science
University of Maribor University of Maribor
Maribor Maribor
Slovenia Slovenia
Rapid development of digital computers has given to computer science a new momen-
tum with emergence of the computational intelligence (CI). In line with this, a several
number of algorithms has been developed to compete with each other in order to reach
an eternal human desire to make an algorithm which would be suitable of solving all
the problems that human are confronted to. Unfortunately, this desire has been reduced
to ashes because of the No-Free-Launch (NFL) theorem. In place of general problem
solvers, specific nature-inspired algorithms incorporated a domain-specific knowledge
to solve problems with sufficient accuracy in real time. The solutions that earlier cannot
be afforded because of time as well as space limitations of digital computers, nowadays
can be solved due to a rapid development of hardware components.
The most powerful algorithms today base on the following inspirations from the
nature:
human brain,
Darwinian evolution and
behavior of some species of social living insects (e.g., bees, ants, termites, etc.) and
animals (birds, dolphins, bats, etc.).
The first inspiration has led to an origin of artificial neural networks (ANNs), the
second to evolutionary algorithms (EAs) and third to swarm intelligence (SI). Unfor-
tunately, without any of the mentioned algorithms we are not near to our ideal to find
the general problem solver. For instance, despite the starting success with ANN in the
eighties of the past millennium, it has been proven that equating the capacity of hu-
man brain with the capacity of computer memory led us in dead end. Nowadays, the
computer capacity exceeds capacity of human brain, but the computers are a long way
away from the intelligence of the human brain. In line with this, the question is how
intelligent these nature-inspired algorithms are.
However, when we would like to answer this the question is what the intelligence is at
all. The answer can be found in Piagets book "The Psychology of Intelligence" arguing
the intelligence is not passive, but arises when the organism interacts with its environ-
ment. The intelligence is adaptive in its nature, where the adaptation is described as
an equilibrium between the action of the organism in the environment and vice versa.
VI Preface
Cleperde and Stern look on the intelligence as a mental adaptation to new circum-
stances. They divide the intelligence mental structures into instincts, trial-and-error and
habits. Thus, the origin of intelligence represents the most elementary empirical test-
and-error that in essence characterizes a search for a hypothesis. However, the hypoth-
esis together with a problem and control represent marks of intelligence.
The test-and-error is a method of problem solving, characterized by repeated at-
tempts until the valid solution is found. Typically, this method is significant for children
when discovering the elementary principles of the world. During this process, they ac-
quire a new knowledge based on their own experience, i.e., learned from own mistakes.
This method presents the fundamental metaphor of computational intelligence as well.
In the sense that the CI algorithms search for hypothesize by solving the problems and
thus adapt to demands of the problem to be solved, these algorithms are reasonably
referred to computational intelligence.
In fact, the adaptation is common characteristic of the nature-inspired algorithms that
are taken into account in this book. However, the adaptation is considered differently
in each of these algorithms. For instance, ANN employ learning due to minimizing
the error rate obtained as a feedback from the learning process. On the other hand, the
EAs and SI-based algorithms adapt to the dynamic demands of the problems either by
searching for the proper parameter setting or even by using the different strategies of
exploring the search space at different stages of the search process.
However, these algorithms are too general for solving the all real-world problems,
with which human are confronted to. In order to converge the solution to the global
optimum as near as possible, a domain-specific knowledge must be incorporated into
the algorithm structures. Usually, the domain-specific knowledge is conducted to the
nature-inspired algorithm via traditional construction heuristics or local search. In gen-
eral, so hybridized algorithm solves the specific problem better than the traditional. In
such so named meta-heuristic algorithms, the nature-inspired algorithm plays a role of
generating the new solution, while the incorporated heuristic solves the problem on
traditional way. This way of solving the problem is also known as generate-and-test.
In summary, the book takes a walk through recent advances in adaptation and hy-
bridization in CI domain. It consists of ten chapters that are divided into three parts.
The first part (Chapter 1) illustrates background information and provides some the-
oretical foundation tackling the CI domain, the second part (Chapters 2-6) deals with
the adaptation in CI algorithms, while the third part (Chapters 7-10) focuses on the hy-
bridization in CI. An emphasis of the book is given to nature-inspired CI algorithms,
like ANNs, EAs and SI-based. All the chapters in the second and third parts are ordered
according to the classification adopted in Chapter 1. A short description of chapters
contents are as follows.
Chapter 1 presents a background of adaptation and hybridization in CI domain. It is
focused especially on three already mentioned nature-inspired algorithms. In line with
this, biological foundations of an adaptation, as the basis for the speciation (i.e., the
formation of new species), are illustrated in the example of an adaptation of Darwins
finches. Then, foundations of nature-inspired algorithms are presented in the sense of
phenomena in the nature serving to imitate their behavior in the corresponding nature-
inspired algorithm. However, the emphasize is on the adaptation and hybridization of
Preface VII
these CI algorithms. Finally, the recent advances captured in the papers, tackling the
adaptation and hybridization are shortly surveyed.
Chapter 2 gives an overview of adaptive and self-adaptive mechanisms within the
DE algorithms. This review shows that these methods mainly base on a controlling the
mutation and crossover parameters, but less on the population size. Additionally, the
chapter proposes a new self-adaptive jDE algorithm using two strategies, selected by
the same probability during the run.
Chapter 3 takes a closer look on the self-adaptation of control parameters in evolu-
tion strategies (ES). In line with this, an analysis of classical mutation operators, like
uncorrelated mutation with one step size and uncorrelated mutation with n step sizes
is performed. Additionally, the uncorrelated mutation with n step size 4-dimensional
vectors is proposed where each 4-dimensional vector consists of a problem variable,
a mutation strength, shifting the location of normal distribution for a shift angle and
reversing the sign of the change. This means, changing a position of each problem vari-
able in a search space depends on three control parameters modified in each generation.
Chapter 4 presents a review of the most relevant adaptive techniques as proposed
in papers tackling the cooperative co-evolution (CC) in EAs. The CC divides a whole
population into sub-populations that explore the search space in different directions.
These sub-populations cooperate by exchanging information in order to evaluate the
fitness of individuals. The chapter finishes with a presentation of a new adaptive CC
firefly algorithm (FA).
Chapter 5 presents a parameter tuning of a novel SI-based optimization algorithm
inspired by krill herd. In the original krill herd (KH) algorithm, the parameter setting
is based on the real data found in biological literature. Unfortunately, the parameter
setting does not comply with the best suited parameter values needed by solving the
specific problem. Therefore, in this chapter, the best parameter setting as found during
the manual tuning process by solving the high-dimensional benchmark problems is
proposed.
Chapter 6 focuses on the SI-based algorithm inspired by the behavior of natural bats
that for an orientation in space and for hunting preys employ a physical phenomenon
called an echolocation. Using this algorithm, economic dispatch (ED) problems are
solved, in this chapter. The manual parameter tuning is proposed in order to find the
best parameter setting.
Chapter 7 deals with an automatic tuning parameters of ANNs, where the original
ANN for fire analysis of steel frame is hybridized with a real-valued meta-GA that
searches for the optimal values of ANN parameters. The meta-GA uses the genetic
operators of crossover and mutation, and evaluates the quality of parameters obtained
after applying the ANN.
Chapter 8 presents a memetic approach which is a hybridization of differential evo-
lution (DE) with a variable neighborhood search (VNS) heuristic. As identified in the
chapter, the performance of the proposed DE_VNS algorithm depends on the muta-
tion strategies, crossover operator and the standard DE control parameters. In line with
this, a multiple mutation operators can be employed to the VNS. In order to prevent a
population diversity, an inversion and injection operators are proposed in the chapter.
VIII Preface
Iztok Fister1, , Damjan Strnad1 , Xin-She Yang2 , and Iztok Fister Jr.1
1
University of Maribor, Faculty of Electrical Engineering and Computer Science
Smetanova ul. 17, 2000 Maribor, Slovenia
{iztok.fister,damjan.strnad,iztok.fister1}@um.si
2
School of Science and Technology, Middlesex University, London NW4 4BT, UK
x.yang@mdx.ac.uk
1 Introduction
The noun adaptation originates from the greek word ad aptare which means to
fit to. This word emerged primarily in biology and was later widened to other
Corresponding author.
areas as well. It designates a collective name for problems arising within dierent
areas, e.g., genetic, articial intelligence, economics, game theory, etc., encom-
passing the optimization problems of dierent diculties regarding complexity
and uncertainty [41]. Complexity means how much eort must be incorporated
in order to solve a specic problem. Uncertainty denotes the environment in
which a problem arises and typically, changes over time. In general, real-world
problems are embodied within environments which are typical dynamic, noisy
and mostly unpredictable.
An adaptive system undergoes acting operators that aects its structure. That
means, such systems adapt to the changing conditions of the environment by
modifying the structure. In fact, each system prepares itself for changes using
the so-called adaptive plan; i.e., the set of factors controlling these changes [41].
The adaptive plan determines how the structures are changed in order to best
t to the changing environment. Typically, the adaptive plans are realized by
developing the operators that determine how the changes of structures are per-
formed. There are several plans (operators) that can be used for adapting to the
environment. Which of these is the best depends on a performance measure in
which the estimation of a plan is based. Selecting the proper performance mea-
sure depends on the domain from which the specic problem arises. On the other
hand, the performance measure estimates the quality of the modied structure.
Many natural, as well as articial systems, arising within dierent domains are
adaptive in nature. Some of these systems, by their structures and performance
measures, are illustrated in Table 1.
In genetics, the structure of an adaptation is a chromosome that undergoes
the actions by the operators of crossover, mutation and inversion. The quality
of an individual is measured by its tness. The tter the individual, the more
chances it has to survive. Articial intelligence looks for a program tool that
imitates the behavior of the human brain, which should be able to learn, while
its performance is normally measured by the error function. The smaller the
value of error function, the better the program is adapted to its environment.
Production is a process of combining various material and immaterial inputs
(plans, know-how) in order to make something for consumption (the output). It
is the act of creating outputs, goods or services which have values and contribute
to the utility of individuals [35]. The higher the utility, the more the production
process is optimized.
In a game theory, a game is a mathematical model of a situation of interactive
decision making, in which every decision maker (or player) strives to attain his
the best possible outcome [42]. Indeed, each player plays a move according to
the strategy that maximize its payo. The payo matrix provides a quantitative
representation of players preference relations over the possible outcomes of the
game. The strategy for player A is the winning strategy if for every move of
player B, player A is the winner. A combination of moves must obey the game
rules by all game players.
Supramolecular chemistry may be dened as chemistry beyond the molecule,
where two molecules (i.e., receptor and substrate) are assembled into
supramolecules using intermolecular bonds [43]. Supramolecules undergo the
actions such as molecular recognition, transformation, and translocation that
may lead to the development of molecular and supramolecular species and can
provide very complex functions. These species are capable of self-organizing, self-
assembling and replicating by using molecular information. Here, the amount of
energy and information is employed as the performance measure.
In memetic computation (MC), a meme represents a building block of infor-
mation obtained by autonomic software agents obtained either by learning or
by interacting with the surrounding agents which acts within a complex dy-
namic environment [24]. Indeed, memes can represent the agents ideas and
knowledge captured as memory items and abstractions (e.g., perceptions, be-
liefs, minds) [29]. The primary memetic operator is imitation [61], which takes
place when the memes are transmitted, replicated or modied. These intelligent
agents are also confronted by selection, where the agents with the higher payos
in the previous generations have more chances for survival.
Although adaptation has emerged within dierent domains of human activi-
ties, it shares the similar characteristics, e.g., each adapted system has its struc-
ture on which operators are applied according to an adaptive plan, while the
modied structure is estimated using a suitable performance measure. The higher
the performance measure, the better the system adapts to its environment. As
a result, only the best adapted structures can continue to develop and improve
their characteristics. The less adapted ones are condemned to disappear. In this
sense, adaptation can also be viewed as an optimization process.
Obviously, most real-world problems are hard to solve. This means that prob-
lems cannot be solved exactly by an algorithm enumerating all the possible
solutions. They are too complex in terms of both the time and space necessary
for obtaining solutions [40]. Therefore, these problems are usually solved ap-
proximately by using heuristic methods that guess the solution of the problem
6 I. Fister et al.
in some (ideally smart) way. Although such a solution is not exact, it is good
enough to be used in a practice.
Nowadays, algorithm developers often try to imitate the operations of natural
processes by attempting to solve the harder, real-world problems. From the al-
gorithm development point of view, there are three types of inspiration sources
from nature:
human brain,
natural selection,
behavior of some social living insects and animals.
The rst source of inspiration has led to the emergence of the articial in-
telligence, where the algorithm tries to mimic the operations of human brains
in order to solve problems, where the main example is the articial neural net-
works (ANNs) [39]. The second source of inspiration has led to the foundations
of evolutionary algorithms (EA) [36] using the Darwinian natural selection [37],
where the ttest individual in a population can survive during the struggle for
existence. The third source of inspiration has closely related to the development
of swarm intelligence (SI) [1,173] that mimics the social behavior of some living
insects and animals [38]. Although such systems tend to obey simple rules, sim-
ple creatures such as ants are capable of performing autonomous actions, they
are still capable of doing great things, e.g., building magnicent anthills, when
acting together within a group. All three mentioned nature-inspired algorithms
can be placed under the umbrella of computational intelligence (CI). The al-
gorithms belonging to this family share the same characteristics, i.e., they are
capable of solving the problems on some sophisticated, intelligent way.
On the other hand, the behavior of an optimization algorithm is controlled by
its parameters (also strategy or control parameters). These parameters mostly
stay xed during the algorithms run. However, this is in contrast to the real-
world, where the good starting values of parameters could become bad during
the run. As a result, a need has been emerged to modify them during the run.
Here, the adaptation of control parameters can be used as well, where the values
of the control parameters are modied during the run in order to best suit the
demands of the search process.
In addition, many traditional algorithms, especially gradient-based methods,
exist that contain a lot of domain-specic knowledge within algorithm struc-
tures. Contrary, the general problem solver methods, especially nature-inspired
population-based algorithms like EAs and SI, are capable to obtain the moder-
ate results on all classes of optimization problems. In order to connect the gen-
eral problem solver methods with the traditional heuristics, the hybridization
of nature-inspired population-based algorithms with traditional heuristic algo-
rithms has been performed. Such hybridized algorithms incorporate a problem-
specic knowledge into algorithms structures and are therefore more suitable
for solving the specic problems. Using more problem-specic knowledge, these
algorithms may overcome limitation imposed by the No-Free Lunch theorem [18]
stating that two algorithms are equivalent when comparing across all classes of
problems. According to Chen et al. [24], the hybridized algorithms evolved over
Adaptation and Hybridization in Nature-Inspired Algorithms 7
simple hybrids, via adaptive hybrids to memetic automation. Simple hybrids of-
ten represent a hybridization of population-based CI algorithms with local search
heuristics. The result of connecting the adaptation with hybridization has led to
adaptive hybrids. The last step in the integration of adaptation with hybridiza-
tion forms a part of memetic computing, where, in addition to the parameters,
other algorithmic structures can also be adapted.
The remainder of this chapter is organized as follows. Section 2 deals with
optimization problems and their complexity. The origin of adaptation within
natural systems is the subject of Section 3. Section 4 analyzes the nature-inspired
algorithms. In line with this, the ANN, EA and SI-based algorithms are taken
into account. Section 5 highlights key characteristics of adaptation and diversity
in CI. Section 6 deals with a description of hybridization methods in CI. A brief
review of recent application arisen in CI is given in Section 7. Finally, some
conclusions are drawn in Section 8.
2 Classification of Problems
From a system analysis point of view, problem-solving can be seen as a system
consisted of three components: input, output, and model (Fig. 1). The model
transforms input data to output data. If the model is known, the output data
can be determined by each set of input data. The problem can also be placed
dierently, i.e., which input data produces specic output data by a known
model. Finally, knowing the input and output data, the problem is how to nd
a model that transforms the specic input data to the output data.
In line with this, three classes of problems can be dened with regard to one
of the unknown components within system analysis, as follows:
optimization: the input data that satises a criterion of optimality are
searched for by a known model and known output data,
simulation: a set of known input data are applied to the known model in
order to simulate the output data,
modeling: searching for a (mathematical) model is performed, which can
transform the known input data to the known output data, at a glance.
The optimization and simulation/modeling problems are described in the next
subsections in more detail.
8 I. Fister et al.
Fig. 2. Gal
apagos archipelago
Speciation and formation of new Galapagos nches were carried over three
phases (Fig. 2):
that have been learned by males from their parents and which are susceptible by
females of the same population. The sound is independent of the reproduction
material, although morphological characteristics of individuals are written in
genes (e.g., the size and the shape of beaks) can have an impact on the volume
and pitch of sound articulated by the bird.
Interestingly, Wrights concept of adaptive landscape [4] can be used to illus-
trate the morphological characteristics of Darwins nches according to various
food sources on the Gal apagos islands. Both dierent morphological characteris-
tics, i.e., the body size and the shape of beaks, are represented as two coordinate
axes in a 3-dimensional coordinate system, while the third axis represents se-
lective advantages or disadvantages of morphological characteristics of a specic
individual in regard to the food sources.
The adaptive landscape of morphological characteristics versus body sizes
and beak shapes can change over the longer period of time. Therefore, such
landscape is also named dynamic adaptive landscape. Similarly to the conditions
in the environment have changed over time, also the heights and positions of
hills are changed in the adaptive landscape. For instance, the height of the hill is
lowered, a valley between two hills is increased or two hills move closer to each
other or move away from each other. Various populations of Darwins nches
adapt to these changes in the environment. If, for example, two hills are moved
closer to each other because of frequent earthquakes on Galapagos archipelago,
two or more populations of Darwins nches come together, while if the hills
are moved away the groups of nches are being separated. Speciation appears
when the specic population colonizes the peak of a hill. Each hill is occupied by
exactly one nch population with the body size and the shape of beaks adapted
to the specic food source. As a result, fteen ecological niches can be discovered
on the Galapagos archipelago, on which exactly the same number of nch species
have appeared.
In computational intelligence, the adaptive landscape is known as the fitness
landscape. Furthermore, the speciation is more frequently used by solving mul-
timodal problems, where more equivalent problem solutions (i.e., more peaks
within the tness landscape) are maintained during the algorithm run. In fact,
each peak represents an ecological niche appropriate for speciation [47].
Therefore, dierent landscapes (from dierent problems) may pose dierent
challenges to dierent algorithms. It is not possible in general to adapt to all
landscapes at the same time. As a result, dierent algorithms may perform
dierently for dierent problems. In order to solve this a broad spectrum of
various problems, developers of new algorithms draw inspirations from dierent
natural systems. Nature-inspired algorithms are the most generalized terms and
we will discuss nature-inspired algorithms in greater detail in the next section.
4 Nature-Inspired Algorithms
Nature-inspired algorithms are very diverse. Loosely speaking, we can put nature-
inspired algorithms into three categories: articial neural networks, evolutionary
Adaptation and Hybridization in Nature-Inspired Algorithms 13
algorithms and swarm intelligence. It is worth pointing out that such categoriza-
tion here is not rigorous. However, it is mainly for the convenience of discussions
in this chapter.
f (x(t) )
x(t+1) = x(t) = A(x(t) ). (2)
f (x(t) )
Obviously, the convergence rate may become very slow near the optimal point
where f (x) 0. Sometimes, the true convergence rate may not be as quick as
it should be. A simple way to improve the convergence is to modify the above
formula slightly by introducing a parameter p as follows:
f (x(t) ) 1
x(t+1) = x(t) p , p= . (3)
f (x(t) ) 1 A (x )
main part of a human neuron (Fig. 3.a) is the cell body that contains a cell
nucleus [39]. The cell body branches out with a number of bers (dendrites) and
a single long ber named an axon. The neuron accepts the incoming signals from
its neighbors axons through dendrite tips at junctions called synapses, which
inhibit or amplify the signal strength. After the processing of accumulated inputs
inside the nucleus the output signal is propagated through the axon to neurons
down the communication line. The brain function is evolved through short-term
and long-term changes in the connectivity of the neurons, which is considered
as learning.
inputs threshold
weights x =-1
0
dendrite w0=q
synapse x1
dendrite w1 activation output
signal x2 w2
axon
S
axon x3 w3 v y
cell body cell body
signal
synapse
dendrite
synapse
dendrite
xn wn
transfer function
(a) Biological neuron (b) Articial neuron
y1
(1)
y1(L-1)
w1,1
(1)
1 1 (L)
w1,1
(L)
w
(L-1) y1
w1,0
(1) (L)
x1 1,0 w1,2 1
y2(1) y2(L-1) (L) (L)
2 2 ws,1 w1,0
x2 (L)
w
(L-1) ws,2
w2,0
(1)
2,0
(L)
w1,r (L)
ym
xn m
y y
(L-1)
(1)
p (L)
p r
r
ws,r wm,0
(L)
wp,n
(1)
(L-1)
w
(1)
p,0
w r,0
v= wi xi , (5)
i=0
In the second step the activation value is injected into the transfer function
to obtain the neuron output y:
y = (v). (6)
The Heaviside step function is used in place of for classication tasks, while
for regression tasks the popular choice for is the logistic function :
1
(v) = . (7)
ev/
Here, is the sigmoid slope parameter with default value 1. Fig. 5 shows the step
function on the left and the logistic function for various values of on the right-
hand side. When a signed version of the sigmoid transfer function is required,
the common choice is the hyperbolic tangent.
f(v) f(v)
1 r=10
1
r=1
0.5
0
v -0.5 0 0.5 v
Fig. 5. The step (left) and the sigmoid (right) activation function
yi =
(l) (l) (l1)
wij yj , 1 i hi ; 1 l L (8)
j=0
(0) (L)
where h0 = n, yi = xi , hL = m, and oi = yi .
Weights represent the programmable part of neural network. In order to per-
form a specic task, we need to train the MLP, i.e., adjust the weights using a set
of training samples with known input-output mappings. This is an example of
supervised learning, which is used in classication tasks with existing records of
correctly labeled patterns or regression tasks with known values of an unknown
nonlinear map in a given set of points.
The weight adaptation in neural networks is achieved by iterative training
algorithms, in which the input parts of the training samples are presented to the
network in succession. A cycle in which all of the training samples are introduced
on the network input is called an epoch. The better known supervised training
Adaptation and Hybridization in Nature-Inspired Algorithms 17
method for the MLP is the error back-propagation algorithm. For each presented
input, the computed network output o is compared with the target vector d to
obtain the prediction error. The usual error measure E in back-propagation
training is the mean squared error (MSE) of the output neurons:
1
E= (d o)T (d o) (9)
m
The weights are then updated in the direction of the negative gradient E/w
to reduce the error in the next iteration.
Training continues until the maximum number of epochs is reached or the av-
erage MSE error for the epoch falls below some prescribed tolerance . General
methods like cross-validation to prevent over-tting can also be used for prema-
ture training termination. The complete back-propagation training algorithm is
summarized in Algorithm 1.
Training continues until the maximum number of epochs is reached or the av-
erage MSE error for the epoch falls below some prescribed tolerance . General
methods like cross-validation to prevent over-tting can also be used for prema-
ture training termination. The complete back-propagation training algorithm is
summarized in Algorithm 1.
Fig. 6. Primarily, EAs dier from each other in terms of the representation of solutions.
For example, GAs operate with a population of mainly binary represented solutions,
ESs use real-valued elements of solutions, GPs represent solutions as trees implemented
in Lisp programming language, while EPs employ the solutions represented as nite
state automata.
Fig. 7. Nature-inspired SI-based algorithms - The picture presents the sources of in-
spiration from nature for developing the following SI-based algorithms that follow in
the clockwise direction: natural immune systems, particle swarm optimization, ower
pollination algorithm, bat algorithm (echolocation), cuckoo search (to lay own eggs
into other birds nests), reies (bioluminescence), bee (foraging of nectar) and ant
colonies(pheromone)
It is worth pointing out that we can only cover and discuss less than 10% of
all dierent SI-based algorithms in this brief review. However, the development
of new types of the SI-based algorithms is not nished yet. Almost every day
new SI-based algorithms have been emerging. In this way, there is no doubt that
this area will become more active in the near future.
may help to speed up the convergence of the search procedure. However, if ex-
ploitation is too strong, it can result in the quick loss of diversity in the pop-
ulation and thus may lead to the premature convergence. On the other hand,
if new search moves are not guided by local landscape information, it can typi-
cally increase the exploration capability and generate new solutions with higher
diversity. However, too much diversity and exploration may result in meandered
search paths, thus lead to the slow convergence. Therefore, adaptation of search
moves so as to balance exploration and exploitation is crucial. Consequently, to
maintain the balanced diversity in a population is also important.
Diversity in meta-heuristic algorithms can also appear in many forms. The
simplest diversity is to allow the variations of solutions in the population by
randomization. For example, solution diversity in genetic algorithms is mainly
controlled by the mutation rate and crossover mechanisms, while in simulated
annealing, diversity is achieved by random walks. In most SI-based algorithms,
new solutions are generated according to a set of deterministic equations, which
also include some random variables. Diversity is represented by the variations,
often in terms of the population variance. Once the population variance is get-
ting smaller (approaching zero), diversity also decreases, leading to converged
solution sets. However, if diversity is reduced too quickly, premature conver-
gence may occur. Therefore, a right amount of randomness and the right form
of randomization can be crucial.
In summary, adaptation and diversity in meta-heuristic algorithms can mainly
take the following forms:
balance of exploration and exploitation,
generation of new solutions,
right amount of randomness,
parameter setting, and
other subtle form.
In the remainder of this chapter, we discuss the role of adaptation and diversity
in these cases.
also local. Actually, it is for a local search. For example, hill-climbing is a method
that uses derivative information to guide the search procedure. In fact, new steps
always try to climb up the local gradient. The advantage of exploitation is that
it usually leads to very high convergence rates, but its disadvantage is that it
can get stuck in a local optimum because the nal solution point largely depends
on the starting point.
On the other hand, exploration makes it possible to explore the search space
more eciently, and it can generate solutions with enough diversity and far from
the current solutions. Therefore, the search is typically on a global scale. The
advantage of exploration is that it is less likely to get stuck in a local mode,
and the global optimality can be more accessible. However, its disadvantages are
slow convergence and waste of a lot of computational eorts because many new
solutions can be far from global optimality.
As a result, a ne balance is required so that an algorithm can achieve the
best performance. Too much exploitation and too little exploration means the
system may converge more quickly, but the probability of nding the true global
optimality may be low. On the other hand, too little exploitation and too much
exploration can cause the search path meander with very slow convergence. The
optimal balance should mean the right amount of exploration and exploitation,
which may lead to the optimal performance of an algorithm. Therefore, a proper
balance is crucially important.
However, how to achieve such a balance is still an open problem. In fact,
no algorithm can claim to have achieved such an optimal balance in the current
literature. In essence, the balance itself is a hyper-optimization problem, because
it is the optimization of an optimization algorithm. In addition, such a balance
may depend on many factors such as the working mechanism of an algorithm,
its setting of parameters, tuning and control of these parameters and even the
problem to be considered. Furthermore, such a balance may not universally
exist [18], and it may vary from problem to problem, thus requiring an adaptive
strategy.
These unresolved problems and mystery can motivate more research in this
area, and it can be expected relevant literature will increase in the near future.
Attraction and Diusion. The novel idea of attraction via light intensity as
an exploitation mechanism was rst used by Yang in the rey algorithm (FA)
in 2007 and 2008. It is simple, exible and easy to implement. This algorithm
bases on the ashing patterns and behavior of tropical reies, and can naturally
deal with nonlinear multimodal optimization problems.
The movement of rey i is attracted to another more attractive (brighter)
rey j as determined by
2
= xi + 0 erij (xj xi ) + i ,
(t+1) (t) (t) (t) (t)
xi (10)
where the second term is due to the attraction, and 0 is the attractiveness at r =
0. The third term is randomization with being the randomization parameter,
(t)
and i is a vector of random numbers drawn from a Gaussian distribution
24 I. Fister et al.
(t)
at time t. Other studies also use the randomization in terms of i that can
easily be extended to other distributions such as Levy ights. A comprehensive
review of the rey algorithm and its variants has been carried out by Fister et
al. [74,79,75].
In FA, the attractiveness (and light intensity) is intrinsically linked with the
inverse-square law of light intensity variations and the absorption coecient. As
a result, there is a novel but nonlinear term of 0 exp[r2 ] where 0 is the
attractiveness at the distance r = 0, and > 0 is the absorption coecient for
light [10].
The main function of such attraction is to enable an algorithm to converge
quickly because these multi-agent systems evolve, interact and attract, leading
to some self-organized behavior and attractors. As the swarming agents evolve,
it is possible that their attractor states will move towards to the true global
optimality.
This novel attraction mechanism in FA is the rst of its kind in the literature of
nature-inspired computation and computational intelligence. This also motivated
and inspired others to design similar or other kinds of attraction mechanisms.
Other algorithms that were developed later also used inverse-square laws, derived
from nature. For example, the charged system search (CSS) used Coulombs law,
while the gravitational search algorithm (GSA) used Newtons law of gravitation.
Whatever the attraction mechanism may be, from the meta-heuristic point of
view, the fundamental principles are the same: that is, they allow the swarming
agents to interact with one another and provide a forcing term to guide the
convergence of the population.
Attraction mainly provides the mechanisms for exploitation, but, with proper
randomization, it is also possible to carry out some degree of exploration. How-
ever, the exploration is better analyzed in the framework of random walks and
diusive randomization. From the Markov chain point of view, random walks
and diusion are both Markov chains. In fact, Brownian diusion such as the
dispersion of an ink drop in water is a random walk. Levy ights can be more
eective than standard random walks. Therefore, dierent randomization tech-
niques may lead to dierent eciency in terms of diusive moves. In fact, it is
not clear what amount of randomness is needed for a given algorithm.
Local random walks around a current solution (often the best solution),
which gives
x(t+1) = x(t) + w, (12)
where w is drawn from a Gaussian normal distribution.
Global Levy ights provide an ecient way of generating long-jump solutions
However, it is very rare for an algorithm to use only one of the above methods.
In fact, most algorithms use a combination of the above methods together with
other ways of solution generation.
where (t) is drawn from a standard normal distribution with a zero mean and
unity standard deviation. Here, the step size s determines how far a random
walker (e.g., an agent or a particle in meta-heuristics) can go for a xed number
of iterations. Obviously, if s is too large, then the new solution x(t+1) generated
will be too far away from the old solution (or more often the current best). Then,
such a move is unlikely to be accepted. If s is too small, the change is too small
to be signicant, and consequently such search is not ecient. So a proper step
size is important to maintain the search as ecient as possible. However, what
size is proper may depend on the type of the problem and can also be changed
during the iteration. Therefore, step sizes and thus the amount of randomness
may have to be adaptive.
26 I. Fister et al.
problem arises because gradient-based methods for ANN training are susceptible
to getting stuck in local optimums on complex error surfaces. For such cases,
global search methods like EAs and SI-based algorithms can provide a robust and
ecient approach for weight optimization. The second problem arises because
the optimal network structure for a specic task is rarely known in advance and
is usually determined by an expert through a tedious experimentation process.
When using EA or SI-based algorithm, the network topology can be dynamically
adapted to the problem at hand by the insertion and removal of neurons or the
connections between them.
The eld of neuro-evolution provides an unied framework for adaptive evolu-
tion and the training of neural networks. In neuro-evolution the ANN structure
and weights are adaptively developed using one of the nature-inspired optimiza-
tion methods with a problem specic tness function. We can distinguish three
groups of neuro-evolutionary methods depending on whether the network param-
eters (i.e., weights), topology or both, are evolved. Further, because the concept
of application to the training and evolution of ANN is very similar using either
EAs or SI-based methods, we regard them all under the term of neuro-evolution
in this text (Fig. 9).
initial population,
genotype-phenotype mapping,
evaluation function, and
variation and selection operators.
which generates a set of new solutions (x1 , ..., xn )(t+1) from the current pop-
ulation of n solutions. This behavior of an algorithm is largely determined by
the eigenvalues of the matrix A that are in turn controlled by the parameters
p = (p1 , . . . , pk ) and the randomness vector = (1 , ..., m ). From the Marko-
vian theory, we know that the rst largest eigenvalue is typically 1, and therefore
the convergence rate of an algorithm is mainly controlled by the second largest
eigenvalue 0 2 < 1 of A. However, it is extremely dicult to nd this eigen-
value in general. Therefore, the tuning of parameters becomes a very challenging
task.
The parameter tuning can be dened as an optimization problem that searches
for those values of the strategic parameters that optimize the performance of the
population-based CI search algorithm [17]. In fact, parameter tuning, or tuning
of parameters, is an important topic under active research [17,177]. The aim
of parameter tuning is to nd the best parameter setting so that an algorithm
can perform most eciently for a wider range of problems. At the moment,
parameter tuning is mainly carried out by detailed, extensive parametric studies,
and there is no ecient method in general. In essence, parameter tuning itself
is an optimization problem which requires higher-level optimization methods
to tackle. However, a recent study had shown that a framework for self-tuning
algorithms can be established with promising results [177].
32 I. Fister et al.
Greedy heuristics are the simplest type of construction heuristics that add new
elements to a solution according to the value of current heuristic function that
can maximize (or minimize) the current non-nal set of elements during each
construction step. When the stochastic construction heuristics [60] are used,
the results of construction may depend on some coincidence. As a result, com-
bining the population-based CI search algorithms which are stochastic in their
nature with stochastic construction heuristics form synergy suitable for solving
the hardest real-world problems.
It should be noticed that MAs represent the simplest class of so-called meme-
inspired computation (MC) that are also known as simple hybrids by Chen et al.
in [24]. Recently, MAs have merged with the eld of hybridization with adap-
tation. In line with this, several studies have been emerged that extended the
concept of adaptation of parameters also to adaptation of operators [25] that
represent the next step in evolution of MC, i.e., adaptive hybrids. In contrast to
simple hybrids in which domain knowledge is only captured and incorporated
once by a human expert during the design of MAs, adaptive hybrids incorporate
the adaptive strategies and adaptive parameters in order to better suit to solve
the problem as the search process progress [24,26]. To date, the further step of
evolution of MC represents the memetic automation already described in Sec-
tion 1 [28]. In the context of MC, all mentioned renement methods represent
the attempt to use memes as the carriers of the various kind of knowledge [27].
allow the algorithm to work on discrete spaces. Alam and Islam proposed an
interesting ABC variant called articial bee colony with self-adaptive mutation
(ABC-SAM) which tries to dynamically adapt the mutation step size with which
bees explore the problem search space. In line with this, small step sizes serve
to an exploitation component, while large mutation steps more to exploration
component of the ABC search process.
On the other hand, some interesting adaptation has also been applied to the
bat algorithm (BA). For example, Fister et al. [73] proposed a self-adaptive bat
algorithm (SABA), based on the self-adapting mechanism borrowed from the
jDE algorithm. In addition, adaptation or self-adaptation in cuckoo search (CS)
has yet to be developed.
However, there are some adaptive and self-adaptive variants of the FA. For
instance, Fister et al. extended the original FA with the self-adaptation of con-
trol parameters called also MSA-FFA and achieved better balancing between
exploration and exploitation of the search process. They tested their proposed
approach on the graph coloring and showed very promising results [63]. This
MSA-FFA was modied by Galvez and Iglesias and adopted for continuous op-
timization problems [81]. Yu et al. [113] proposed a self-adaptive step FA to
avoid falling into the local optimum and reduce the impact of the maximum
of generations. Authors core idea was to set the step of each rey varying
with the iteration according to current situation and also historical informa-
tion of reies. Roy et al. [106] developed a FA variant using self-adaptation of
the algorithm control parameter values by learning from the reies previous
experiences, which led to a more ecient algorithm.
Adaptations in improving PSO are widely spread in many papers describing
many applications. Since there are many ecient PSO variants, and readers can
refer to the following papers [99,109,115,90,114,95].
ANN for nancial forecasting [129]. The domain of stock forecasting attracted
researchers who hybridized ANNs with the ABC algorithm [150] and the sh al-
gorithm [156]. A related application of ABC to earthquake time-series prediction
is due to Shah et al. [155].
Additionally, for the most recent SI-based algorithms, adaptive hybridizations
of ANNs with the FA [142,146], the BA [148], the CS [147], and hunting algo-
rithm/harmony search combination [138] have also been carried out with good
results.
ANN training was also approached using the population-based algorithms
which are not strictly nature-inspired, such as magnetic optimization algorithm
[143], chemical reaction optimization [166], and articial photosynthesis and pho-
totropism [123].
While the majority of hybrid ANN+SI-based approaches are concerned with
ANN training, evolution of both weights and topology using the PSO was pre-
sented by Garro et al. [130] and by Ali [118]. A version of PSO called jumping
PSO was recently used by Ismail and Jeng to obtain self-evolving ANN [132].
There are many hybrid variants of EAs. Most studies in this domain are based
on hybridization with local search, and recently also on borrowing some princi-
ples from SI. In line with this, Grimaccia et al. [84] combined properties of PSO
and GA, and tested performance on the optimization of electromagnetic struc-
tures. Galinier and Hao in [80] proposed a hybrid EA (HEA) for graph coloring.
Their algorithms combined a highly specialized crossover operators with Tabu
search [83]. GA-EDA [104] is a good example of a hybrid EA which uses genetic
and estimation of distribution algorithms. Niknam [102] developed a new EA
algorithm called DPSO-HBMO, which based on the combination of honey bee
mating optimization [87] and discrete PSO [171]. Lin [98] proposed a new EA
combining DE with the real-valued GA.
ABC was also hybridized in many papers to enhance its performance and
eciency. Duan et al. [71] proposed an ABC and quantum EA, where the ABC
was adopted to increase the local search capacity and also the randomness of
populations. Data clustering was improved with hybrid ABC (HABC) [111],
where authors introduced crossover operator of genetic algorithm to ABC and
enhance information exchange between bees. Large-scale global optimization was
tackled by memetic ABC (MABC) algorithm [72], where the original ABC was
hybridized with two local search heuristics: the Nelder-Mead algorithm (NMA)
and the random walk with direction exploitation (RWDE) in order to obtain the
better balance between exploration and exploitation. Moreover, a hybrid simplex
ABC algorithm (HSABC) which combines NMA with ABC was proposed and
applied for solving the inverse analysis problems [92]. An interesting hybrid
variant of ABC was also applied to solve graph coloring problems [77].
BA has also been developed many hybrid variants, which try to enhance the
eciency, performance, quality of solutions, and faster convergence. A hybrid
BA with path relinking was proposed by Zhou et al. [116], where authors in-
tegrated the greedy randomized adaptive search procedure (GRASP) and path
relinking into the BA, and applied to capacitated vehicle routing problem. Fister
et al. [76] created a hybrid BA (HBA) in order to combine the original BA with
DE strategies as a local search instead of classic random walk. An extension of
the SABA was done by the same authors in [31] where they hybridized the SABA
(HSABA) also with ensemble DE strategies that were used as a local search for
improving current best solution directing the swarm of a solution towards the
better regions within a search space. Wand and Guo developed a novel hybrid
BA with harmony search and applied to global numerical optimization [85].
Chandrasekaran and Simon [66] proposed a hybrid CS (HCS) algorithm that
was integrated with a fuzzy system in order to cope with multi-objective unit
commitment problems. Layeb [94] developed a novel quantum inspired CS that
connects the original CS with quantum computing principles. The main advan-
tage of this hybridization was a good balance between exploration and exploita-
tion during the search process. Li and Yin [96] created a new hybrid variant of
CS called CS-based memetic algorithm and applied it for solving permutation
ow shop scheduling problems. Since the creation of CS, a diverse range of hy-
brid variants this algorithm have emerged. Therefore, readers are invited to read
the review of these algorithms in the paper [91].
FA is another example of very successful SI-based algorithm that experienced
many promising hybridizations since its birth in 2008. Although a comprehen-
sive description of this algorithm was performed in papers [74,75], let us present
some ecient and recent hybrid variants of the FA only. Kavousi-Fard et al. [93]
combined a support vector machine (SVM) and modied FA in order to get
a hybrid prediction algorithm and applied it to the short term electrical load
forecast. Guo et al. in [86] combined FA with harmony search. The result of this
hybridization was an eective algorithm for solving the global numerical opti-
mization problems. On the other hand, Fister et al [78] developed a memetic FA
(MFFA) and applied it to the graph coloring problems. Interesting approach to
40 I. Fister et al.
the distributed graph coloring problem based on the calling behavior of Japanese
tree frogs were accomplished by Hern andez and Blum in [170].
PSO underwent many hybridization suitable for continuous and combinato-
rial optimization. For instance, Lovbjerg et al. [100] created a hybrid PSO and
borrowed some concepts from EAs. A very interesting method was proposed
by Marinakis and Marinaki [101] where authors developed new approach based
on PSO, greedy randomized adaptive search procedure and expanding neigh-
borhood search. This algorithm was then tested on the probabilistic traveling
salesman problem. Zhang et al. proposed a DEPSO algorithm [172], which com-
bined PSO with DE operators, while Wang and Li [108] combined PSO with
simulated annealing (SA).
Obviously, there are other developments and applications, but the purpose of
this chapter is not to review all of they. Therefore, interested readers can refer
to more specialized literature.
8 Conclusion
Adaptation becomes the metaphor for reactions of the natural or articial system
to the conditions of the changing environment. There are a lot of renewed inter-
ests in this area. Therefore, this chapter starts from a denition of adaptive sys-
tems and identies the human domains that already deal with this phenomenon.
Adaptation has also been encountered in the domain of problem-solving. In or-
der to solve these problems, developers usually try to develop new algorithms
imitating the main characteristics of natural processes. Interestingly, the nature
does not impose questions only, but also provides the answers how to solve these.
However, these answers provides diverse sources of inspiration for scientists in
order to solve their problems.
Researchers have always been trying to nd the general problem solver suit-
able to solve all classes of the real-world problems. However, this is usually not
possible as constrained by the NFL theorem. Hybridization of nature-inspired
algorithms may partly overcome the limitations of the NFL theorem, when solv-
ing a specic problem by incorporating the problem-specic knowledge in the
algorithm structures. In line with this, some popular hybridization methods have
been presented in the chapter, with emphasize on the memetic algorithms. This
initial idea of hybridizing the population-based CI nature-inspired algorithms
with the local search has led to the emergence of the new area in CI, i.e., memetic
computation that represents the class of new general problem solvers suitable
for solving the hardest real-world problems.
Here, we have identied three main sources of inspiration that are the most
commonly used nowadays for the development of the new nature-inspired al-
gorithms, i.e., human brains, a Darwinian natural selection, and behavior or
some social living insects and animals. In line with this, three classes of nature-
inspired algorithms have been emerged, in general: ANNs, EAs and SI-based.
All the mentioned classes of algorithms placed under the umbrella of CI are de-
scribed in detail throughout this chapter. The descriptions of these algorithms
Adaptation and Hybridization in Nature-Inspired Algorithms 41
References
1. Beni, G., Wang, J.: Swarm Intelligence in Cellular Robotic Systems. In: Proceed-
ings of NATO Advanced Workshop on Robots and Biological Systems, Tuscany,
Italy, pp. 2630 (1989)
2. Turing, A.M.: Computing machinery and intelligence. Mind, 433460 (1950)
3. Grant, P.R., Grant, B.R.: Adaptive Radiation of Darwins Finches. American
Scientist 90(2), 130150 (2002)
4. Wright, S.A.: The roles of mutation, inbreeding, crossbreeding and selection in
evolution. In: Proceedings of the VI International Congress of Genetrics, vol. (1),
pp. 356366 (1932)
5. Dasgupta, D.: Information Processing in the Immune System. In: Corne, D.,
Dorigo, M., Glover, F. (eds.) New Ideas in Optimization, pp. 161167. McGraw
Hill, New York (1999)
6. Dorigo, M., Di Caro, G.: The Ant Colony Optimization Meta-heuristic. In: Corne,
D., Dorigo, M., Glover, F. (eds.) New Ideas in Optimization, pp. 1132. McGraw
Hill, London (1999)
7. Karaboga, D., Bastruk, B.: A Powerful and Ecient Algorithm for Numerical
Function Optimization: Articial Bee Colony (ABC) Algorithm. Journal of Global
Optimization 39(3), 459471 (2007)
8. Kennedy, J., Eberhart, R.: The Particle Swarm Optimization; Social Adaptation
in Information Processing. In: Corne, D., Dorigo, M., Glover, F. (eds.) New Ideas
in Optimization, pp. 379387. McGraw Hill, London (1999)
9. Yang, X.-S.: A New Metaheuristic Bat-Inspired Algorithm. In: Gonz alez, J.R.,
Pelta, D.A., Cruz, C., Terrazas, G., Krasnogor, N. (eds.) NICSO 2010. SCI,
vol. 284, pp. 6574. Springer, Heidelberg (2010)
10. Yang, X.-S.: Firey Algorithm. In: Yang, X.-S. (ed.) Nature-Inspired Metaheuris-
tic Algorithms, pp. 7990. Luniver Press, London (2008)
11. Yang, X.-S.: Flower Pollination Algorithm for Global Optimization. In: Durand-
Lose, J., Jonoska, N. (eds.) UCNC 2012. LNCS, vol. 7445, pp. 240249. Springer,
Heidelberg (2012)
12. Yang, X.-S., Deb, S.: Cuckoo Search via Levy Flights. In: World Congress &
Biologically Inspired Computing (NaBIC 2009), pp. 210214. IEEE Publication
(2009)
13. Storn, R., Price, K.: Dierential Evolution: A Simple and Ecient Heuristic
for Global Optimization over Continuous Spaces. Journal of Global Optimiza-
tion 11(4), 341359 (1997)
14. Moore, G.E.: Cramming more components onto integrated circuits. Electron-
ics 38(8), 114117 (1965)
15. Ulam, S.: Tribute to John von Neumann. Bulletin of the American Mathematical
Society 64(3), 5056 (1958)
42 I. Fister et al.
16. McCulloch, W.S., Pitts, W.: A logical calculus of the ideas immanent in nervous
activity. The Bulletin of Mathematical Biophysics 5(4), 115133 (1943)
17. Eiben, A.E., Smith, S.K.: Parameter tuning for conguring and analyzing evolu-
tionary algorithms. Swarm and Evolutionary Computation 1(1), 1931 (2011)
18. Wolpert, D., Macready, W.: No free lunch theorems for optimization. IEEE Trans-
actions on Evolutionary Computation 1(1), 6782 (1997)
19. Moscato, P.: Memetic algorithms: A short introduction. In: Corne, D., Dorigo, M.,
Glover, F. (eds.) New Ideas in Optimization, pp. 219234. McGraw Hill, London
(1999)
20. Wilfried, J.: A general cost-benet-based adaptation framework for multimeme
algorithms. Memetic Computing 2, 201218 (2010)
21. Crepin sek, M., Liu, S.-H., Mernik, M.: Exploration and exploitation in evolution-
ary algorithms: A survey. ACM Computing Surveys 45(3), 133 (2013)
22. Neri, F., Cotta, C.: Memetic algorithms and memetic computing optimization: A
literature review. Swarm and Evolutionary Computation 1(2), 114 (2011)
23. Neri, F.: Diversity Management in Memetic Algorithms. In: Neri, F., Cotta, C.,
Moscato, P. (eds.) Handbook of Memetic Algorithms, pp. 153164. Springer,
Berlin (2012)
24. Chen, X., Ong, Y.-S., Lim, M.-H., Tan, K.C.: A Multi-Facet Survey on Memetic
Computation. Trans. Evol. Comp. 15(5), 591607 (2011)
25. Ong, Y.-S., Lim, M.-H., Zhu, N., Wong, K.-W.: Classication of adaptive memetic
algorithms: a comparative study. IEEE Transactions on Systems, Man, and Cy-
bernetics, Part B: Cybernetics 36(1), 141152 (2006)
26. Garcia, S., Cano, J.R., Herrera, F.: A memetic algorithm for evolutionary proto-
type selection: A scaling up approach. Pattern Recogn. 41(8), 26932709 (2008)
27. Iacca, G., Neri, F., Mininno, E., Ong, Y.-S., Lim, M.-H.: Ockhams Razor in
memetic computing: Three stage optimal memetic exploration. Inf. Sci. 188(4),
1743 (2012)
28. Ong, Y.-S., Lim, M.H., Chen, X.: Research frontier: memetic computation-past,
present & future. Comp. Intell. Mag. 5 2(5), 2431 (2010)
29. Lynch, A.: Thought as abstract evolution. J. Ideas 2(1), 310 (1991)
30. Fister Jr., I., Yang, X.-S., Fister, I., Brest, J., Fister, D.: A brief review
of nature-inspired algorithms for optimization. Electrotehnical Review 80(3),
116122 (2013)
31. Fister, I., Fong, S., Brest, J., Fister Jr., I.: A novel hybrid self-adaptive bat algo-
rithm. The Scientic World Journal, 112 (2014)
32. Fister, I., Mernik, M., Filipic, B.: Graph 3-coloring with a hybrid self-adaptive
evolutionary algorithm. Comp. Opt. and Appl. 54(3), 741770 (2013)
33. Fister, I., Mernik, M., Filipic, B.: A hybrid self-adaptive evolutionary algorithm
for marker optimization in the clothing industry. Appl. Soft Comput. 10(2),
409422 (2010)
34. Grefenstette, J.: Optimization of control parameters for genetic algorithms. IEEE
Transactions on Systems, Man, and Cybernetics 16, 122128 (1986)
35. Kotler, P., Armstrong, G., Brown, L., Adam, S.: Marketing, 7th edn. Pearson
Education Australia/Prentice Hall, Sydney (2006)
36. Eiben, A., Smith, J.: Introduction to Evolutionary Computing. Springer, Berlin
(2003)
37. Darwin, C.: On the Origin of Species. Harvard University Press, London (1859)
38. Blum, C., Merkle, D.: Swarm Intelligence. Springer, Berlin (2008)
39. Russell, S., Norvig, P.: Articial Intelligence: A Modern Approach, 3rd edn. Pren-
tice Hall, New Jersey (2009)
Adaptation and Hybridization in Nature-Inspired Algorithms 43
40. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory
of NP-Completeness. W. H. Freeman & Co., New York (1979)
41. Holland, J.H.: Adaptation in Natural and Articial Systems: An Introductory
Analysis with Applications to Biology, Control, and Articial Intelligence. A Brad-
ford Book, Cambridge (1992)
42. Maschler, M., Solan, A., Zamir, S.: Game Theory. Cambridge University Press,
Cambridge (2013)
43. Lehn, J.M.: Supramolecular Chemistry: Concepts and Perspectives. VCH Ver-
lagsgeselschaft, Weinheim (1995)
44. Applegate, D.L., Bixby, R.E., Chv atal, V., Cook, W.: The Traveling Salesman
Problem. University Press, Princeton (2006)
45. Bondy, J.A., Murty, U.S.R.: Graph Theory. Springer, Berlin (2008)
46. Back, T.: Evolutionary Algorithms in Theory and Practice: Evolution Strategies,
Evolutionary Programming, Genetic Algorithms. Oxford University Press, Oxford
(1996)
47. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learn-
ing. Addison-Wesley Longman Publishing Co., Inc., Boston (1989)
48. Fogel, L., Owens, A., Walsh, M.: Articial Intelligence through Simulated Evolu-
tion. John Willey & Sons, Inc., New York (1966)
49. Koza, J.: Genetic Programming 2 - Automatic Discovery of Reusable Programs.
MIT Press, Cambridge (1994)
50. Searle, J.R.: The rediscovery of the mind. MIT Press, Cambridge (1992)
51. Rechenberg, I.: Evolutionsstrategie, Optimierung technischer Systeme nach
Prinzipien der biologischen Evolution. Frommann-Holzboog, Stuttgart (1973)
52. Schwefel, H.P.: Numerische Optimierung von ComputerModellen mittels der
Evolutionsstrategie. Birkh auser, Basel (1977)
53. Deb, K.: Multi-Objective Optimization Using Evolutionary Algorithms. John Wi-
ley & Sons, Inc., New York (2001)
54. Michalewicz, Z., Fogel, D.: How to solwe it: Modern heuristics. Springer (2004)
55. Michalewicz, Z.: Genetic algorithms + data structures = evolution programs.
Springer, Berlin (1992)
56. Moscato, P.: On evolution, search, optimization, genetic algorithm and martial
arts: Toward memetic algorithms. Tech. Rep. 826. California Institute of Tech-
nology, Pasadena, CA (1989)
57. Yang, X.-S.: Nature-Inspired Metaheuristic Algorithms. Luniver Press (2010)
58. Dawkins, R.: The selsh gene. Oxford University Press, Oxford (1976)
59. Aarts, E., Lenstra, J.K.: Local Search in Combinatorial Optimization. Oxford
University Press, Princeton (1997)
60. Hoos, H.H., St utzle, T.: Stochastic Local Search: Foundations and Applications.
Elsevier, Oxford (2005)
61. Blackmore, S.: The Meme Machine. Oxford University Press, New York (1999)
62. Law, A.: Simulation Modeling and Analysis with Expertt Software. McGraw-
Hill, New York (2006)
63. Fister, I., Fister Jr, I., Brest, J., Yang, X.-S.: Memetic rey algorithm for com-
binatorial optimization. In: Filipic, B., Silc, J. (eds.) Bioinspired Optimization
Methods and Their Applications: Proceedings of the Fifth International Confer-
ence on Bioinspired Optimization Methods and their Applications, BIOMA 2012,
pp. 7586. Jozef Stefan Institute, Ljubljana (2012)
64. Brest, J., Greiner, S., Boskovic, B., Mernik, M., Zumer, V.: Self-adapting control
parameters in dierential evolution: A comparative study on numerical bench-
mark problems. IEEE Transactions on Evolutionary Computation 10(6), 646657
(2006)
44 I. Fister et al.
65. Cai, Z., Peng, Z.: Cooperative coevolutionary adaptive genetic algorithm in path
planning of cooperative multi-mobile robot systems. Journal of Intelligent and
Robotic Systems 33(1), 6171 (2002)
66. Chandrasekaran, K., Simon, S.P.: Multi-objective scheduling problem: Hybrid ap-
proach using fuzzy assisted cuckoo search algorithm. Swarm and Evolutionary
Computation 5, 116 (2012)
67. Chitty, D.M., Hernandez, M.L.: A hybrid ant colony optimisation technique
for dynamic vehicle routing. In: Deb, K., Tari, Z. (eds.) GECCO 2004. LNCS,
vol. 3102, pp. 4859. Springer, Heidelberg (2004)
68. Deb, K., Beyer, H.-G.: Self-adaptive genetic algorithms with simulated binary
crossover. Evolutionary Computation 9(2), 197221 (2001)
69. Dilettoso, E., Salerno, N.: A self-adaptive niching genetic algorithm for multi-
modal optimization of electromagnetic devices. IEEE Transactions on Magnet-
ics 42(4), 12031206 (2006)
70. Duan, H., Yu, X.: Hybrid ant colony optimization using memetic algorithm for
traveling salesman problem. In: IEEE International Symposium on Approximate
Dynamic Programming and Reinforcement Learning, ADPRL 2007, pp. 9295.
IEEE (2007)
71. Duan, H.-B., Xu, C.-F., Xing, Z.-H.: A hybrid articial bee colony optimization
and quantum evolutionary algorithm for continuous optimization problems. In-
ternational Journal of Neural Systems 20(01), 3950 (2010)
72. Fister, I., Fister Jr., I., Zumer, V., Brest, J.: Memetic articial bee colony algo-
rithm for large-scale global optimization. In: 2012 IEEE Congress on Evolutionary
Computation (CEC), pp. 18. IEEE (2012)
73. Fister Jr, I., Fong, S., Brest, J., Fister, I: Towards the self-adaptation in the
bat algorithm. In: Proceedings of the 13th IASTED International Conference on
Articial Intelligence and Applications (2014)
74. Fister, I., Fister Jr., I., Yang, X.-S., Brest, J.: A comprehensive review of rey
algorithms. Swarm and Evolutionary Computation (2013)
75. Fister, I., Yang, X.-S., Fister, D., Fister Jr., I.: Firey algorithm: A brief review of
the expanding literature. In: Cuckoo Search and Firey Algorithm, pp. 347360.
Springer (2014)
76. Fister Jr., I., Fister, D., Yang, X.-S.: A hybrid bat algorithm. arXiv preprint
arXiv:1303.6310 (2013)
77. Fister Jr., I., Fister, I., Brest, J.: A hybrid articial bee colony algorithm for
graph 3-coloring. In: Rutkowski, L., Korytkowski, M., Scherer, R., Tadeusiewicz,
R., Zadeh, L.A., Zurada, J.M. (eds.) EC 2012 and SIDE 2012. LNCS, vol. 7269,
pp. 6674. Springer, Heidelberg (2012)
78. Fister Jr, I., Yang, X.-S., Fister, I., Brest, J.: Memetic rey algorithm for com-
binatorial optimization. arXiv preprint arXiv:1204.5165 (2012)
79. Fister, I., Yang, X.-S., Brest, J., Fister Jr., I.: Modied rey algorithm using
quaternion representation. Expert Syst. Appl. 40(18), 72207230 (2013)
80. Galinier, P., Hao, J.-K.: Hybrid evolutionary algorithms for graph coloring. Jour-
nal of Combinatorial Optimization 3(4), 379397 (1999)
81. Galvez, A., Iglesias, A.: New memetic self-adaptive rey algorithm for continuous
optimization. International Journal of Bio-Inspired Computation (2014)
82. Geem, Z.W., Kim, J.H., Loganathan, G.: A new heuristic optimization algorithm:
harmony search. Simulation 76(2), 6068 (2001)
83. Glover, F., Laguna, M.: Tabu search. Springer (1999)
Adaptation and Hybridization in Nature-Inspired Algorithms 45
84. Grimaccia, F., Mussetta, M., Zich, R.E.: Genetical swarm optimization: Self-
adaptive hybrid evolutionary algorithm for electromagnetics. IEEE Transactions
on Antennas and Propagation 55(3), 781785 (2007)
85. Guo, L.: A novel hybrid bat algorithm with harmony search for global numerical
optimization. Journal of Applied Mathematics 2013 (2013)
86. Guo, L., Wang, G.-G., Wang, H., Wang, D.: An eective hybrid rey algorithm
with harmony search for global numerical optimization. The Scientic World Jour-
nal 2013 (2013)
87. Haddad, O.B., Afshar, A., Marino, M.A.: Honey-bees mating optimization (hbmo)
algorithm: a new heuristic approach for water resources optimization. Water Re-
sources Management 20(5), 661680 (2006)
88. Heinonen, J., Pettersson, F.: Hybrid ant colony optimization and visibility studies
applied to a job-shop scheduling problem. Applied Mathematics and Computa-
tion 187(2), 989998 (2007)
89. Hinterding, R., Michalewicz, Z., Peachey, T.C.: Self-adaptive genetic algorithm for
numeric functions. In: Ebeling, W., Rechenberg, I., Voigt, H.-M., Schwefel, H.-P.
(eds.) PPSN 1996. LNCS, vol. 1141, pp. 420429. Springer, Heidelberg (1996)
90. Ismail, A., Engelbrecht, A.P.: The self-adaptive comprehensive learning parti-
cle swarm optimizer. In: Dorigo, M., Birattari, M., Blum, C., Christensen, A.L.,
Engelbrecht, A.P., Gro, R., St utzle, T. (eds.) ANTS 2012. LNCS, vol. 7461,
pp. 156167. Springer, Heidelberg (2012)
91. Fister Jr., I., Fister, D., Fister, I.: A comprehensive review of cuckoo search: vari-
ants and hybrids. International Journal of Mathematical Modelling and Numerical
Optimisation 4(4), 387409 (2013)
92. Kang, F., Li, J., Xu, Q.: Structural inverse analysis by hybrid simplex articial
bee colony algorithms. Computers & Structures 87(13), 861870 (2009)
93. Kavousi-Fard, A., Samet, H., Marzbani, F.: A new hybrid modied rey algo-
rithm and support vector regression model for accurate short term load forecast-
ing. Expert Systems with Applications 41(13), 60476056 (2014)
94. Layeb, A.: A novel quantum inspired cuckoo search for knapsack problems. Inter-
national Journal of Bio-Inspired Computation 3(5), 297305 (2011)
95. Li, C., Yang, S., Nguyen, T.T.: A self-learning particle swarm optimizer for global
optimization problems. IEEE Transactions on Systems, Man, and Cybernetics,
Part B: Cybernetics 42(3), 627646 (2012)
96. Li, X., Yin, M.: A hybrid cuckoo search via levy ights for the permutation ow
shop scheduling problem. International Journal of Production Research 51(16),
47324754 (2013)
97. Liao, X., Zhou, J., Zhang, R., Zhang, Y.: An adaptive articial bee colony algo-
rithm for long-term economic dispatch in cascaded hydropower systems. Interna-
tional Journal of Electrical Power & Energy Systems 43(1), 13401345 (2012)
98. Lin, W.-Y.: A gade hybrid evolutionary algorithm for path synthesis of four-bar
linkage. Mechanism and Machine Theory 45(8), 10961107 (2010)
99. Liu, S., Wang, J.: An improved self-adaptive particle swarm optimization
approach for short-term scheduling of hydro system. In: International Asia
Conference on Informatics in Control, Automation and Robotics, CAR 2009,
pp. 334338. IEEE (2009)
100. Lovbjerg, M., Rasmussen, T.K., Krink, T.: Hybrid particle swarm optimiser with
breeding and subpopulations. In: Proceedings of the Genetic and Evolutionary
Computation Conference, vol. 2001, pp. 469476. Citeseer (2001)
46 I. Fister et al.
101. Marinakis, Y., Marinaki, M.: A hybrid multi-swarm particle swarm optimization
algorithm for the probabilistic traveling salesman problem. Computers & Opera-
tions Research 37(3), 432442 (2010)
102. Niknam, T.: An ecient hybrid evolutionary algorithm based on pso and hbmo
algorithms for multi-objective distribution feeder reconguration. Energy Conver-
sion and Management 50(8), 20742082 (2009)
103. Pan, Q.-K., Fatih Tasgetiren, M., Suganthan, P.N., Chua, T.J.: A discrete arti-
cial bee colony algorithm for the lot-streaming ow shop scheduling problem.
Information Sciences 181(12), 24552468 (2011)
104. Pena, J.M., Robles, V., Larranaga, P., Herves, V., Rosales, F., Perez, M.S.: GA-
EDA: Hybrid evolutionary algorithm using genetic and estimation of distribu-
tion algorithms. In: Orchard, B., Yang, C., Ali, M. (eds.) IEA/AIE 2004. LNCS
(LNAI), vol. 3029, pp. 361371. Springer, Heidelberg (2004)
105. Qin, A.K., Suganthan, P.N.: Self-adaptive dierential evolution algorithm for nu-
merical optimization. In: The 2005 IEEE Congress on Evolutionary Computation,
vol. 2, pp. 17851791. IEEE (2005)
106. Roy, A.G., Rakshit, P., Konar, A., Bhattacharya, S., Kim, E., Nagar, A.K.: Adap-
tive rey algorithm for nonholonomic motion planning of car-like system. In:
2013 IEEE Congress on Evolutionary Computation (CEC), pp. 21622169. IEEE
(2013)
107. Wang, J., Osagie, E., Thulasiraman, P., Thulasiram, R.K.: Hopnet: A hybrid
ant colony optimization routing algorithm for mobile ad hoc network. Ad Hoc
Networks 7(4), 690705 (2009)
108. Wang, X.-H., Li, J.-J.: Hybrid particle swarm optimization with simulated an-
nealing. In: Proceedings of 2004 International Conference on Machine Learning
and Cybernetics, vol. 4, pp. 24022405. IEEE (2004)
109. Wang, Y., Li, B., Weise, T., Wang, J., Yuan, B., Tian, Q.: Self-adaptive learn-
ing based particle swarm optimization. Information Sciences 181(20), 45154538
(2011)
110. Wu, Q., Cao, Y., Wen, J.: Optimal reactive power dispatch using an adap-
tive genetic algorithm. International Journal of Electrical Power & Energy Sys-
tems 20(8), 563569 (1998)
111. Yan, X., Zhu, Y., Zou, W., Wang, L.: A new approach for data clustering using
hybrid articial bee colony algorithm. Neurocomputing 97, 241250 (2012)
112. Yang, Z., Tang, K., Yao, X.: Self-adaptive dierential evolution with neighborhood
search. In: IEEE Congress on Evolutionary Computation, CEC 2008 (IEEE World
Congress on Computational Intelligence), pp. 11101116. IEEE (2008)
113. Yu, S., Yang, S., Su, S.: Self-adaptive step rey algorithm. Journal of Applied
Mathematics 2013 (2013)
114. Zhan, Z.-H., Zhang, J., Li, Y., Chung, H.-H.: Adaptive particle swarm optimiza-
tion. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernet-
ics 39(6), 13621381 (2009)
115. Zhang, J., Ding, X.: A multi-swarm self-adaptive and cooperative particle swarm
optimization. Engineering Applications of Articial Intelligence 24(6), 958967
(2011)
116. Zhou, Y., Xie, J., Zheng, H.: A hybrid bat algorithm with path relinking for
capacitated vehicle routing problem. Mathematical Problems in Engineering 2013
(2013)
117. Alander, J.T.: An indexed bibliography of genetic algorithms and neural networks
118. Ali, Y.M.B.: Evolving multilayer feedforward neural network using adaptive par-
ticle swarm algorithm. Int. J. Hybrid Intell. Syst. 8(4), 185198 (2011)
Adaptation and Hybridization in Nature-Inspired Algorithms 47
119. Angeline, P.J., Saunders, G.M., Pollack, J.B.: An evolutionary algorithm that
constructs recurrent neural networks. IEEE Transactions on Neural Networks 5,
5465 (1994)
120. Asadi, S., Hadavandi, E., Mehmanpazir, F., Nakhostin, M.M.: Hybridization of
evolutionary levenberg-marquardt neural networks and data pre-processing for
stock market prediction. Knowl.-Based Syst. 35, 245258 (2012)
121. Caudell, T.P., Dolan, C.P.: Parametric connectivity: Training of constrained net-
works using genetic algorithms. In: David Schaer, J. (ed.) Proceedings of the
Third International Conference on Genetic Algorithms. Morgan Kaufmann Pub-
lishers (1989)
122. Cruz-Ramrez, M., Herv as-Martnez, C., Gutierrez, P.A., Perez-Ortiz, M.,
Briceno, J., de la Mata, M.: Memetic pareto dierential evolutionary neural
network used to solve an unbalanced liver transplantation problem. Soft. Com-
put. 17(2), 275284 (2013)
123. Cui, Z., Yang, C., Sanyal, S.: Training articial neural networks using appm.
IJWMC 5(2), 168174 (2012)
124. da Silva, A.J., Mineu, N.L., Ludermir, T.B.: Evolving articial neural networks
using adaptive dierential evolution. In: Kuri-Morales, A., Simari, G.R. (eds.)
IBERAMIA 2010. LNCS, vol. 6433, pp. 396405. Springer, Heidelberg (2010)
125. Delgado, M., Pegalajar, M.C., Cuellar, M.P.: Evolutionary training for dynami-
cal recurrent neural networks: an application in nantial time series prediction.
Mathware & Soft Computing 13(2), 89110 (2006)
126. Elhachmi, J., Guennoun, Z.: Evolutionary neural networks algorithm for the dy-
namic frequency assignment problem. International Journal of Computer Science
& Information Technology 3(3), 4961 (2011)
127. Fernandez, J.C., Hervas, C., Martnez-Estudillo, F.J., Gutierrez, P.A.: Memetic
pareto evolutionary articial neural networks to determine growth/no-growth in
predictive microbiology. Appl. Soft Comput. 11(1), 534550 (2011)
128. Furtuna, R., Curteanu, S., Leon, F.: Multi-objective optimization of a stacked
neural network using an evolutionary hyper-heuristic. Appl. Soft Comput. 12(1),
133144 (2012)
129. Gao, W.: Financial data forecasting by evolutionary neural network based on ant
colony algorithm. In: Deng, H., Miao, D., Lei, J., Wang, F.L. (eds.) AICI 2011,
Part III. LNCS, vol. 7004, pp. 262269. Springer, Heidelberg (2011)
130. Garro, B.A., Sossa, H., Vazquez, R.A.: Design of articial neural networks using
a modied particle swarm optimization algorithm. In: Proceedings of the 2009
International Joint Conference on Neural Networks, IJCNN 2009, pp. 23632370
(2009)
131. Igel, C.: Neuroevolution for reinforcement learning using evolution strategies. In:
Reynolds, R., Abbass, H., Tan, K.C., Mckay, B., Essam, D., Gedeon, T. (eds.)
Congress on Evolutionary Computation (CEC 2003), vol. 4, pp. 25882595. IEEE
(2003)
132. Ismail, A.Z., Jeng, D.S.: SEANN: A Self-evolving Neural Network based on PSO
and JPSO algorithms
133. Kala, R., Shukla, A., Tiwari, R.: Modular symbiotic adaptive neuro evolution for
high dimensionality classicatory problems. Intelligent Decision Technologies 5(4),
309319 (2011)
134. Kassahun, Y., Sommer, G.: Ecient reinforcement learning through evolutionary
acquisition of neural topologies. In: ESANN, pp. 259266 (2005)
135. Kaylani, A.: An Adaptive Multiobjective Evolutionary Approach to Optimize
Artmap Neural Networks. PhD thesis, Orlando, FL, USA (2008), AAI3335346
48 I. Fister et al.
136. Khan, M.M., Ahmad, A.M., Khan, G.M., Miller, J.F.: Fast learning neural net-
works using cartesian genetic programming. Neurocomputing 121, 274289 (2013)
137. Khan, M.M., Khan, G.M., Miller, J.F.: Evolution of neural networks using carte-
sian genetic programming. In: IEEE Congress on Evolutionary Computation, pp.
18. IEEE (2010)
138. Kulluk, S.: A novel hybrid algorithm combining hunting search with harmony
search algorithm for training neural networks. JORS 64(5), 748761 (2013)
139. Liao, S.-H., Hsieh, J.-G., Chang, J.-Y., Lin, C.-T.: Training neural networks via
simplied hybrid algorithm mixing neldermead and particle swarm optimization
methods. Soft Computing, 111 (2014)
140. Mandischer, M.: Representation and evolution of neural networks, pp. 643649.
Springer (1993)
141. Matteucci, M.: ELeaRNT: Evolutionary learning of rich neural network topolo-
gies. Technical report, Carnegie Mellon University (2002)
142. Lee, M.-C., Horng, M.-H., Lee, Y.-X., Liou, R.-J.: Firey Meta-Heuristic Algo-
rithm for Training the Radial Basis Function Network for Data Classication and
Disease Diagnosis. InTech (2012)
143. Mirjalili, S., Sadiq, A.S.: Magnetic optimization algorithm for training multi layer
perceptron. In: 2011 IEEE 3rd International Conference on Communication Soft-
ware and Networks (ICCSN), pp. 4246 (May 2011)
144. Mirjalili, S., Hashim, S.Z.M., Sardroudi, H.M.: Training feedforward neural net-
works using hybrid particle swarm optimization and gravitational search algo-
rithm. Applied Mathematics and Computation 218(22), 1112511137 (2012)
145. Moriarty, D., Miikkulainen, R.: Forming neural networks through ecient and
adaptive coevolution. Evolutionary Computation 5, 373399 (1998)
146. Nandy, S., Karmakar, M., Sarkar, P.P., Das, A., Abraham, A., Paul, D.: Agent
based adaptive rey back-propagation neural network training method for dy-
namic systems. In: 2012 12th International Conference on Hybrid Intelligent Sys-
tems (HIS), pp. 449454 (December 2012)
147. Nawi, N.M., Khan, A., Rehman, M.Z.: Csbprnn: A new hybridization technique
using cuckoo search to train back propagation recurrent neural network. In:
Herawan, T., Deris, M.M., Abawajy, J. (eds.) Proceedings of the First Interna-
tional Conference on Advanced Data and Information Engineering (DaEng-2013).
LNEE, vol. 285, pp. 111118. Springer, Heidelberg (2014)
148. Nawi, N.M., Rehman, M.Z., Khan, A.: A new bat based back-propagation (BAT-
BP) algorithm. In: Swiatek, J., Grzech, A., Swiatek,
P., Tomczak, J.M. (eds.)
Advances in Systems Science. AISC, vol. 240, pp. 395404. Springer, Heidelberg
(2014)
149. Neruda, R., Slusn y, S.: Parameter genetic learning of perceptron networks. In:
Proceedings of the 10th WSEAS International Conference on Systems, ICS 2006,
pp. 9297 (2006)
150. Nourani, E., Rahmani, A.-M., Navin, A.H.: Forecasting stock prices using a hybrid
articial bee colony based neural network. In: 2012 International Conference on
Innovation Management and Technology Research (ICIMTR), pp. 486490 (May
2012)
151. Oong, T.H., Isa, N.A.M.: Adaptive evolutionary articial neural networks for
pattern classication. IEEE Transactions on Neural Networks 22(11), 18231836
(2011)
152. Risi, S., Stanley, K.O.: Enhancing es-hyperneat to evolve more complex regular
neural networks. In: Proceedings of the 13th Annual Conference on Genetic and
Evolutionary Computation, GECCO 2011, pp. 15391546 (2011)
Adaptation and Hybridization in Nature-Inspired Algorithms 49
153. Sarangi, P.P., Sahu, A., Panda, M.: Article: A hybrid dierential evolution and
back-propagation algorithm for feedforward neural network training. International
Journal of Computer Applications 84(14), 19 (2013); Published by Foundation
of Computer Science, New York, USA
154. Sermpinis, G., Theolatos, K.A., Karathanasopoulos, A.S., Georgopoulos, E.F.,
Dunis, C.L.: Forecasting foreign exchange rates with adaptive neural networks
using radial-basis functions and particle swarm optimization. European Journal
of Operational Research 225(3), 528540 (2013)
155. Shah, H., Ghazali, R., Nawi, N.M.: Using articial bee colony algorithm for mlp
training on earthquake time series data prediction. CoRR, abs/1112.4628 (2011)
156. Shen, W., Guo, X., Wu, C., Wu, D.: Forecasting stock indices using radial basis
function neural networks optimized by articial sh swarm algorithm. Knowl.-
Based Syst. 24(3), 378385 (2011)
157. Slowik, A.: Application of an adaptive dierential evolution algorithm with mul-
tiple trial vectors to articial neural network training. IEEE Transactions on
Industrial Electronics 58(8), 31603167 (2011)
158. Stanley, K.O., DAmbrosio, D.B., Gauci, J.: A hypercube-based encoding for
evolving large-scale neural networks. Artif. Life 15(2), 185212 (2009)
159. Stanley, K.O., Miikkulainen, R.: Ecient reinforcement learning through evolv-
ing neural network topologies. In: Proceedings of the Genetic and Evolutionary
Computation Conference, GECCO 2002, pp. 569577 (2002)
160. Stanley, K.O., Miikkulainen, R.: Evolving neural networks through augmenting
topologies. Evol. Comput. 10(2), 99127 (2002)
161. Suchorzewski, M.: Evolving scalable and modular adaptive networks with devel-
opmental symbolic encoding. Evolutionary Intelligence 4(3), 145163 (2011)
162. Turner, A.J., Miller, J.F.: Cartesian genetic programming encoded articial neural
networks: A comparison using three benchmarks. In: Proceedings of the 15th
Annual Conference on Genetic and Evolutionary Computation, GECCO 2013,
pp. 10051012 (2013)
163. Vogl, T.P., Mangis, J.K., Rigler, A.K., Zink, W.T., Alkon, D.L.: Accelerating
the convergence of the back-propagation method. Biological Cybernetics 59(4-5),
257263 (1988)
164. Whitley, D., Starkweather, T., Bogart, C.: Genetic algorithms and neural
networks: optimizing connections and connectivity. Parallel Computing 14(3),
347361 (1990)
165. Yao, X., Liu, Y.: A new evolutionary system for evolving articial neural networks.
IEEE Transactions on Neural Networks 8, 694713 (1996)
166. Yu, J.J.Q., Lam, A.Y.S., Li, V.O.K.: Evolutionary articial neural network based
on chemical reaction optimization. In: IEEE Congress on Evolutionary Compu-
tation, pp. 20832090. IEEE (2011)
167. Zhang, J.-R., Zhang, J., Lok, T.-M., Lyu, M.R.: A hybrid particle swarm
optimization-back-propagation algorithm for feedforward neural network train-
ing. Applied Mathematics and Computation 185(2), 10261037 (2007)
168. Zhang, Y., Wu, L.: Crop classication by forward neural network with adaptive
chaotic particle swarm optimization. Sensors 11(5), 47214743 (2011)
169. Montana, D.J., Davis, L.: Training feedforward neural networks using genetic al-
gorithms. In: Proceedings of the 11th International Joint Conference on Articial
intelligence (IJCAI 1989), vol. 1, pp. 762767. Morgan Kaufmann Publishers Inc.,
San Francisco (1989)
170. Hern andez, H., Blum, C.: Distributed graph coloring: an approach based on the
calling behavior of Japanese tree frogs. Swarm Intelligence, 117150 (2012)
50 I. Fister et al.
171. Chen, W.-N., Zhang, J., Chung, H.S.H., Zhong, W.-L., Wu, W.-G., Shi, Y.-H.:
A novel set-based particle swarm optimization method for discrete optimization
problems. Trans. Evol. Comp. 14, 278300 (2010)
172. Zhang, W.-J., Xie, X.-F.: DEPSO: Hybrid Particle Swarm with Dierential Evo-
lution Operator. IEEE International Conference on Systems, Man and Cybernet-
ics 4, 38163821 (2003)
173. Yang, X.S.: Nature-Inspired Optimization Algorithms. Elsevier, London (2014)
174. Ashby, W.R.: Princinples of the self-organizing sysem. In: Von Foerster, H., Zopf
Jr., G.W. (eds.) Pricinples of Self-Organization: Transactions of the University of
Illinois Symposium, pp. 255278. Pergamon Press, London (1962)
175. Booker, L., Forrest, S., Mitchell, M., Riolo, R.: Perspectives on Adaptation in
Natural and Articial Systems. Oxford University Press, Oxford (2005)
176. Blum, C., Roli, A.: Metaheuristics in combinatorial optimisation: Overview and
conceptural comparision. ACM Comput. Surv. 35, 268308 (2003)
177. Yang, X.S., Deb, S., Loomes, M., Karamanoglu, M.: A framework for self-tuning
optimization algorithm. Neural Computing and Applications 23(7-8), 20512057
(2013)
Part II
Adaptation in Computational
Intelligence
Adaptation in the Dierential Evolution
1 Introduction
Population-based algorithms are suitable for solving continuous optimization
problems as well as for discrete optimization. Population-based algorithms in-
clude particle swarm algorithms, evolutionary algorithms, and other algorithms
inspired by nature. These algorithms usually have several control parameters
that are responsible for tuning the algorithm itself. Good values of the parame-
ters have an inuence on improving an algorithms performance during an opti-
mization process.
Globally, we distinguish between two major forms of setting parameter values:
parameter tuning and parameter control. Parameter tuning means that a user
tries to nd good values for the parameters before running the algorithm and
then running the algorithm using these values, which remain xed during the
optimization process. A particular problem may prefer some parameter values
at the early optimization stage, while other values are more suitable at the later
stage. A question can arise on how many stages an optimization process should
have. When to decide to change parameter values? If the values of the parameters
are changed during the run, we call it parameter control.
Corresponding author.
Self-adaptation allows the solver to adapt itself to any problem from a general
class of problems, to recongure itself accordingly, and to do this without any
user interaction [4]. On the other hand, when solving a particular problem using
tuning, it is possible to nd very good parameter values which are usually more
competitive than a self-adaptive mechanism.
In this chapter we conduct an overview of adaptation in Dierential Evolution
(DE). The DE algorithm is a population-based evolutionary algorithm. Since it
has only a few control parameters and is very ecient for solving real-world
problems, it has become a very popular algorithm. Adaptive and self-adaptive
DE-variants were recently proposed, in order to avoid the need for problem
specic parameter tuning and also for improving the convergence characteristics
of DE.
The chapter is structured as follows. Section 2 presents an overview of the
DE algorithm. In Section 3 a survey is given of works related to adaptation and
self-adaptation, and a case study is presented of the self-adaptive mechanism on
the jDE algorithm, as an example. Section 4 shows the experimental results of
DE algorithms with and without self-adaptive mechanisms. Section 5 concludes
this chapter.
2 Background
transformations, the algorithm stops and returns a best found solution. The
DE algorithm uses mutation, crossover, and selection operators to generate a
next population from the current population.
The DE algorithm belongs to the evolutionary algorithms but there are some
dierences between the DE and an evolutionary algorithm (EA):
Let us present the original DE algorithm [27]. It uses three operators within
an evolutionary process. The population in generation G consists of NP vectors:
(G) (G) (G) (G)
xi = (xi,1 , xi,2 , ..., xi,D ), i = 1, 2, ..., NP.
(G)
Mutation. A mutant vector vi is created by using one of the DE mutation
strategies [13,23]. Currently, there exist many mutational strategies, and the
more powerful are:
(G) (G) (G) (G)
rand/1: vi = xr1 + F (xr2 xr3 ), (1)
(G) (G) (G) (G)
best/1: vi = xbest + F (xr1 xr2 ), (2)
(G) (G) (G) (G) (G) (G)
current-to-best/1: vi = xi + F (xbest xi ) + F (xr1 xr2 ), (3)
(G) (G) (G) (G) (G) (G)
random-to-best/1: vi = xr1 + F (xbest xr1 ) + F (xr2 xr3 ), (4)
(G) (G) (G) (G) (G) (G)
best/2: vi = xbest + F (xr1 xr2 ) + F (xr3 xr4 ), (5)
(G) (G) (G) (G) (G) (G)
rand/2: vi = xr1 + F (xr2 xr3 ) + F (xr4 xr5 ), (6)
where the indexes r1 r5 represent the random and mutually dierent integers
generated within the set {1, . . . , N P } and also dierent from index i. F is a
(G)
mutation scale factor within the range [0, 2], usually less than 1. xbest denotes
the best vector in generation G.
If some components of the mutant vector are out of bounds, the proposed
solutions for repairing the mutant vector found in the literature [27,24] are as
follows: (1) they are reected onto bounds, (2) set on bounds, (3) used as they
are (out of bounds), and (4) randomly generated once again until they are out
of bounds. Which solution is the more appropriate depends on the problem we
are solving it, its characteristics, etc.
56 J. Brest, A. Zamuda, and B. Boskovic
(G)
Crossover. A crossover operator generates a trial vector ui as follows:
(G)
(G) vi,j , if rand(0, 1) CR or j = jrand ,
ui,j = (G)
xi,j , otherwise,
1: {NP ... population size, F ... scale factor, CR ... crossover parameter}
2: {xi ... i-th individual of population}
3: {MaxFEs ... maximum number of function evaluations}
4: {rand(0, 1) ... uniformly distributed random number [0, 1)}
5: Initialization()
6: {** Generate uniformly distributed random population within search space **}
7: while stopping criateria is not met do
8: for (i = 0; i < NP; i = i + 1) do
9: {*** DE/rand/1/bin ***}
10: Randomly select indexes r1 , r2 , and r3 , that are mutually dierent and
dierent also from index i.
(G) (G) (G) (G)
11: vi = xr1 + F (xr2 xr3 )
12: jrand = rand{1, . . . , D}
13: for (j = 0; j < D; j + +) do
14: if (rand(0, 1) CR or j == jrand ) then
(G) (G)
15: ui,j = vi,j
16: else
(G) (G)
17: ui,j = xi,j
18: end if
19: end for
(G) (G)
20: if (f (ui ) f (xi )) then
(G+1) (G)
21: xi = ui
22: else
(G+1) (G)
23: xi = xi
24: end if
25: end for
26: end while
Algorithm 1: DE algorithm
Adaptation in the Dierential Evolution 57
The DE [27] algorithm was proposed by Storn and Price 1997, and since then it
has been used in many practical cases. The original DE had no adaptive control
parameters, since their values were xed during the evolutionary process.
with mean value of 0.5 and standard deviation of 0.3, denoted by N (0.5, 0.3). A
set of F values are randomly sampled from such normal distribution and applied
to each target vector within the current population. SaDE gradually adjusts the
range of CR values for a given problem according to previous CR values which
have generated trial vectors for successfully entering the next generation. CR
is approximated by a normal distribution with mean value CRm and standard
deviation Std = 0.1, denoted by N (CRm, Std), where CRm is initialized as 0.5.
SaDE combined two mutation strategies DE/rand/1 and DE/current-to-best/1.
Das et al. proposed a neighborhood concept for the population members of
DE, called DEGL [7]. It is a similar idea to the communities of the Particle
Swarm Optimization (PSO) algorithms. The small neighborhoods are dened
over the index-graph of parameter vectors.
A self-adaptive dierential evolution algorithm with opposition-based mecha-
nisms is presented in [20]. This opposition-based mechanism can be used during
the initialization of a population or later during the optimization process.
Zhang and Sanderson [37] proposed self-adaptive DE (JADE) with DE/
current-to-pBest mutation strategy:
(G)
where xpBest is randomly chosen as one of the top 100p% individuals of the current
population with p (0, 1]. Fi is the scale factor associated with the ith individual
and it is updated dynamically in each generation. Instead of only adopting the best
individual in the DE/current-to-best/1 strategy, the current-to-pBest/1 strategy
utilizes information regarding other good solutions, while the best individual is
adopted in the DE/current-to-best/1 strategy. The DE/current-to-pBest/1 strat-
egy is a less greedy generalization of the DE/current-to-best/1 strategy. It also up-
dates the control parameters in an adaptive manner along with generations. The
algorithm uses an optional external archive to track the previous history of success
(G)
and failure. xr2 is in this case selected at random from a union of the current pop-
ulation and the archive. The archive size is xed and if the size exceeds a certain
threshold, then some individuals from the archive are randomly eliminated.
A new mutation strategy based on the best of a group of randomly selected
solutions from the current generation (called DE/current-to-gr best/1) was pro-
posed in [19].
Success-History based Adaptive DE (SHADE) [28] is an improved version of
JADE [37]. It uses a historical memory in order to adapt the control parameters
F and CR, current-to-pBest/1 mutation strategy, and external archive. SHADE
performed excellently on CEC2013 competition on Real-Parameter Single Ob-
jective Optimization.
Wang et al. in [34] proposed a new DE, which employs self-adapting control
parameters and generalized opposition-based learning.
Distributed DE is another very prominent optimization technique. In [12]
a distributed DE with several subpopulations and two migration selection ap-
proaches to maintaining a high diversity in the subpopulations are presented.
Adaptation in the Dierential Evolution 59
A wider overview over DE related work and its applicability within various do-
mains can be found in surveys about DE [23,13]. Recently, the adaptive DE algo-
rithm has been used for optimization within dierent domains [38,17,2,33,32,3,18],
and many others, which clearly indicates the high usability of adaptive and self-
adaptive mechanism in the DE algorithms.
The self-adaptive jDE was introduced in 2006 [7]. The self-adapting mechanism
uses rand/1/bin strategy and is applied on the control parameters F and CR.
The third control parameter NP remained unchanged. However, it seems that
NP also plays an important role among control parameters in the DE [9,5].
jDE-based algorithms have been applied to solve large-scale single objec-
tive optimization problems: CEC 2008 [11], CEC 2010 [10,6], CEC 2012 [5],
CEC2013 [6], large-scale continuous optimization problems [9], dynamic opti-
mization [8], and in real problems [36,35].
In [7] a self-adaptive control mechanism was used to change the control param-
eters F and CR during a run. Each individual in the population was extended
using the values of these two control parameters (see Figure 1). Both of them
were applied at individual level. The better values for these (encoded) control
parameters lead to better individuals which, in turn, are more likely to survive
and produce ospring and, hence, propagate these better parameter values.
(G+1) (G+1)
In jDE [7], new control parameters Fi and CR i are calculated before
the mutation operator as follows [7]:
(G+1) Fl + rand1 Fu , if rand2 < 1 ,
Fi = (G)
Fi , otherwise,
(G+1) rand3 , if rand4 < 2 ,
CR i = (G)
CR i , otherwise,
where randj , for j {1, 2, 3, 4} are uniform random values within the range [0, 1].
The jDE algorithm uses DE/rand/1/bin strategy. The presented self-adaptive
mechanism can also be used with other DE strategies [6].
In [7] parameters 1 , 2 , Fl , Fu are xed to values 0.1, 0.1, 0.1, 0.9, respectively.
Figure 1 presents the usage of one mutation strategy with two self-adaptive
control parameters in each individual. As a particular strategy might indicate
dierent performance properties during stages of the evolutionary process, it is
reasonable to apply two or more strategies in the DE algorithm. Each strategy
can have its own control parameters (see Figure 2). The simple way of perform-
ing more strategies is to use them with the same probability. More sophisticated
usages of many strategies are proposed in literature [25,9] like adaptive mecha-
nisms which usually utilize a rule that the better strategy should have a higher
probability to be chosen during the mutation.
60 J. Brest, A. Zamuda, and B. Boskovic
No. Functions fi = fi (x )
1 Sphere Function -1400
Unimodal 2 Rotated High Conditioned Elliptic Function -1300
Functions 3 Rotated Bent Cigar Function -1200
4 Rotated Discus Function -1100
5 Dierent Powers Function -1000
6 Rotated Rosenbrocks Function -900
7 Rotated Schaers F7 Function -800
8 Rotated Ackleys Function -700
9 Rotated Weierstrass Function -600
10 Rotated Griewanks Function -500
11 Rastrigins Function -400
Basic 12 Rotated Rastrigins Function -300
Multimodal 13 Non-Continuous Rotated Rastrigins Function -200
Functions 14 Schwefels Function -100
15 Rotated Schwefels Function 100
16 Rotated Katsuura Function 200
17 Lunacek Bi Rastrigin Function 300
18 Rotated Lunacek Bi Rastrigin Function 400
19 Expanded Griewanks plus Rosenbrocks Function 500
20 Expanded Scaers F6 Function 600
21 Composition Function 1 (n=5, Rotated) 700
22 Composition Function 2 (n=3, Unrotated) 800
23 Composition Function 3 (n=3, Rotated) 900
Composition 24 Composition Function 4 (n=3, Rotated) 1000
Functions 25 Composition Function 5 (n=3, Rotated) 1100
26 Composition Function 6 (n=5, Rotated) 1200
27 Composition Function 7 (n=5, Rotated) 1300
28 Composition Function 8 (n=5, Rotated) 1400
Search Range: [-100,100]D
Adaptation in the Dierential Evolution 61
...
...
...
...
xNP,1 xNP,2 ... xNP,D FNP CRNP
Fig. 1. Population and control parameters within one generation. Each individual has
its own F and CR control parameters.
strategy1 strategy2
x1,1 x1,2 ... x1,D F1 CR1 F1 CR1
...
...
...
...
...
...
xNP,1 xNP,2 ... xNP,D FNP CRNP FNP CRNP
Fig. 2. Each individual has its own two-pairs of F and CR control parameters, each
pair belongs to one DE strategy
4 Experimental Results
D = 10, and 51 runs of algorithm were executed for each function. The opti-
mal values are known a-priori for all benchmark functions, and therefore, we
can compute an error between the obtained value using our algorithm and the
optimal value. Note, error values smaller than 108 are taken as zero.
In the experiments, the parameters of the DE algorithm were set as follows:
F = 0.5, CR = 0.1,
F = 0.5, CR = 0.9,
F = 0.9, CR = 0.9, and
NP = 100.
The parameters of the jDE algorithm were set as follows:
Table 3. Experimental results of the DE algorithms with variing F and CR for di-
mension D = 10
Table 4. Experimental results jDE -2bin with dimension D = 10 (in last column, jDE
mean values are shown)
5 Conclusion
This chapter presented the adaptive and self-adaptive mechanisms of control pa-
rameters in the Dierential Evolution (DE) algorithm. More mutation strategies
can also be applied in the algorithm. These strategies can have either common
(adaptive or self-adaptive) control parameters or each strategy has its own con-
trol parameters. Using adaptation mechanisms of control parameters, more mu-
tation strategies, also dierent crossover schemes, show that the DE algorithm
can adopt to a particular problem being solved.
References
1. Ali, M.M., Torn, A.: Population Set-Based Global Optimization Algorithms: Some
Modications and Numerical Studies. Computers & Operations Research 31(10),
17031725 (2004)
2. Asafuddoula, M., Ray, T., Sarker, R.: An adaptive hybrid dierential evolution
algorithm for single objective optimization. Applied Mathematics and Computa-
tion 231, 601618 (2014)
3. Baatar, N., Jeong, K.-Y., Koh, C.-S.: Adaptive Parameter Controlling Non-
Dominated Ranking Dierential Evolution for Multi-Objective Optimization of
Electromagnetic Problems. IEEE Transactions on Magnetics 50(2) (February 2014)
4. Back, T.: Adaptive Business Intelligence Based on Evolution Strategies: Some Ap-
plication Examples of Self-Adaptive Software. Information Sciences 148, 113121
(2002)
5. Brest, J., Boskovic, B., Zamuda, A., Fister, I., Sepesy Maucec, M.: Self-Adaptive
Dierential Evolution Algorithm with a Small and Varying Population Size. In:
IEEE World Congress on Computational Intelligence (IEEE WCCI 2012), Bris-
bane, Australia, pp. 28272834 (2012)
6. Brest, J., Boskovic, B., Zamuda, A., Fister, I., Mezura-Montes, E.: Real Param-
eter Single Objective Optimization using Self-Adaptive Dierential Evolution Al-
gorithm with more Strategies. In: IEEE Congress on Evolutionary Computation
(CEC) 2013, pp. 377383 (2013)
7. Brest, J., Greiner, S., Boskovic, B., Mernik, M., Zumer, V.: Self-Adapting Control
Parameters in Dierential Evolution: A Comparative Study on Numerical Bench-
mark Problems. IEEE Transactions on Evolutionary Computation 10(6), 646657
(2006)
8. Brest, J., Korosec, P., Silc, J., Zamuda, A., Boskovic, B., Maucec, M.S.: Dier-
ential evolution and dierential ant-stigmergy on dynamic optimisation problems.
International Journal of Systems Science 44, 663679 (2013)
Adaptation in the Dierential Evolution 67
9. Brest, J., Maucec, M.S.: Self-adaptive dierential evolution algorithm using popula-
tion size reduction and three strategies. Soft Computing - A Fusion of Foundations,
Methodologies and Applications 15(11), 21572174 (2011)
10. Brest, J., Zamuda, A., Boskovic, B., Fister, I., Maucec, M.S.: Large Scale Global
Optimization using Self-adaptive Dierential Evolution Algorithm. In: IEEE World
Congress on Computational Intelligence, pp. 30973104 (2010)
11. Brest, J., Zamuda, A., Boskovic, B., Maucec, M.S., Zumer, V.: High-dimensional
Real-parameter Optimization Using Self-adaptive Dierential Evolution Algorithm
with Population Size Reduction. In: 2008 IEEE World Congress on Computational
Intelligence, pp. 20322039. IEEE Press (2008)
12. Cheng, J., Zhang, G., Neri, F.: Enhancing distributed dierential evolution
with multicultural migration for global numerical optimization. Information Sci-
ences 247, 7293 (2013)
13. Das, S., Suganthan, P.N.: Dierential evolution: A survey of the state-of-the-art.
IEEE Transactions on Evolutionary Computation 15(1), 2754 (2011)
14. Eiben, A.E., Hinterding, R., Michalewicz, Z.: Parameter Control in Evolutionary
Algorithms. IEEE Transactions on Evolutionary Computation 3(2), 124141 (1999)
15. Eiben, A.E., Smith, J.E.: Introduction to Evolutionary Computing. In: Natural
Computing. Springer, Berlin (2003)
16. Feoktistov, V.: Dierential Evolution: In Search of Solutions. Springer Optimiza-
tion and Its Applications. Springer-Verlag New York, Inc., Secaucus (2006)
17. Gong, W., Cai, Z., Yang, J., Li, X., Jian, L.: Parameter identication of an SOFC
model with an ecient, adaptive dierential evolution algorithm. International
Journal of Hydrogen Energy 39(10), 50835096 (2014)
18. Hu, Z., Xiong, S., Fang, Z., Su, Q.: A Convergent Dierential Evolution Algorithm
with Hidden Adaptation Selection for Engineering Optimization. Mathematical
Problems in Engineering (2014)
19. Islam, S.M., Das, S., Ghosh, S., Roy, S., Suganthan, P.N.: An adaptive dierential
evolution algorithm with novel mutation and crossover strategies for global numer-
ical optimization. IEEE Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics 42(2), 482500 (2012)
20. Ku, J.H., Cai, Z.H., Zheng, B., Yun, D.W.: The research of self-adaptive dierential
evolution algorithm with opposition-based mechanisms. Applied Mechanics and
Materials 543-547, 17061710 (2014)
21. Liang, J.J., Qu, B.-Y., Suganthan, P.N., Hern andez-Daz, A.G.: Problem Deni-
tions and Evaluation Criteria for the CEC 2013 Special Session and Competition
on Real-Parameter Optimization. Technical Report 201212, Computational Intelli-
gence Laboratory, Zhengzhou University, Zhengzhou China and Technical Report,
Nanyang Technological University, Singapore (2013)
22. Liu, J., Lampinen, J.: A Fuzzy Adaptive Dierential Evolution Algorithm. Soft
Computing - A Fusion of Foundations, Methodologies and Applications 9(6),
448462 (2005)
23. Neri, F., Tirronen, V.: Recent advances in dierential evolution: a survey and
experimental analysis. Articial Intelligence Review 33(1-2), 61106 (2010)
24. Price, K.V., Storn, R.M., Lampinen, J.A.: Dierential Evolution, A Practical Ap-
proach to Global Optimization. Springer (2005)
25. Qin, A.K., Huang, V.L., Suganthan, P.N.: Dierential evolution algorithm with
strategy adaptation for global numerical optimization. IEEE Transactions on Evo-
lutionary Computation 13(2), 398417 (2009)
68 J. Brest, A. Zamuda, and B. Boskovic
26. Storn, R., Price, K.: Dierential Evolution - a simple and ecient adaptive scheme
for global optimization over continuous spaces. Technical Report TR-95-012, Berke-
ley, CA (1995)
27. Storn, R., Price, K.: Dierential Evolution A Simple and Ecient Heuristic for
Global Optimization over Continuous Spaces. Journal of Global Optimization 11,
341359 (1997)
28. Tanabe, R., Fukunaga, A.: Evaluating the performance of shade on cec 2013 bench-
mark problems. In: 2013 IEEE Congress on Evolutionary Computation (CEC),
pp. 19521959 (June 2013)
29. Teo, J.: Exploring dynamic self-adaptive populations in dierential evolution. Soft
Computing - A Fusion of Foundations, Methodologies and Applications 10(8),
673686 (2006)
30. Tvrdk, J.: Competitive dierential evolution. In: MENDEL 2006, 12th Interna-
tional Conference on Soft Computing, pp. 712 (2006)
31. Tvrdk, J.: Adaptation in dierential evolution: A numerical comparison. Appl.
Soft Comput. 9(3), 11491155 (2009)
32. Vasundhara, Mandal, D., Kar, R., Ghoshal, S.P.: Digital FIR lter design using t-
ness based hybrid adaptive dierential evolution with particle swarm optimization.
Natural Computing 13(1), 5564 (2014)
33. Venske, S.M., Goncalves, R.A., Delgado, M.R.: ADEMO/D: Multiobjective op-
timization by an adaptive dierential evolution algorithm. Neurocomputing 127,
6577 (2014); 12th Brazilian Symposium on Neural Networks (SBRN) held as part
of the 1st Brazilian Conference on Intelligent Systems (BRACIS), Curitiba, Brazil,
October 20-25, 2012.
34. Wang, H., Rahnamayan, S., Wu, Z.: Parallel dierential evolution with self-
adapting control parameters and generalized opposition-based learning for solv-
ing high-dimensional optimization problems. Journal of Parallel and Distributed
Computing 73(1), 6273 (2013)
35. Zamuda, A., Brest, J.: Vectorized Procedural Models for Animated Trees Recon-
struction using Dierential Evolution. Information Sciences 278, 121 (2014)
36. Zamuda, A., Brest, J., Boskovic, B., Zumer, V.: Dierential Evolution for Param-
eterized Procedural Woody Plant Models Reconstruction. Applied Soft Comput-
ing 11, 49044912 (2011)
37. Zhang, J., Sanderson, A.C.: JADE: Adaptive Dierential Evolution with Optional
External Archive. IEEE Transactions on Evolutionary Computation 13(5), 945958
(2009)
38. Zhong, Y., Zhao, L., Zhang, L.: An Adaptive Dierential Evolution Endmember
Extraction Algorithm for Hyperspectral Remote Sensing Imagery. IEEE Geoscience
and Remote Sensing Letters 11(6), 10611065 (2014)
On the Mutation Operators
in Evolution Strategies
1 Introduction
This chapter focuses on the self-adaptation in evolution strategies (ES). Pri-
marily, the self-adaptation has been gaining popularity due to the exibility in
adaptation to dierent tness landscapes [1]. This method enables an implicit
learning of mutation strengths in the real-valued search spaces. Self-adaptation
bases on the mutation operator that modies the problem variables using the
strategy parameters to search the problem and parameter spaces simultane-
ously [2]. In line with this, the best values of problem variables and strategy
parameters survive during the evolutionary process.
Evolutionary algorithms (EAs) are intrinsically dynamic, adaptive processes [3].
Therefore, setting the strategy parameters, controlling the behavior of these algo-
rithms, xed during the run is in contrast with an idea of evolution on which bases
Corresponding author.
an evolutionary computation (EC). In line with this, the evolution of evolution has
been developed by Rechengerg [31] and Schwefel [32], where the strategy param-
eters are put into a representation of individuals and undergo operations of the
variation operators. As a result, the values of strategy parameters that modify the
problem variables, which are the most adapted to the tness landscape, as deter-
mined by the tness function of the problem to be solved during the evolutionary
search process. However, the process of the simultaneously evolution of strategy
parameters together with the problem variables is also named a self-adaptation in
EC.
Although the adaptation can tackle various elements of EAs, like representa-
tion of individuals (e.g., Schaefer in [5]) or operators (e.g., Shaer et al. in [6,7]),
it has especially been enforced by adapting the control parameters that regulate
the mutation stregths by mutation operators [8,9]. The comprehensive survey of
self-adapting in EAs can be found in [4,24,10].
In classical ES, three types of mutation operators have been applied: an uncor-
related mutation with one step size, an uncorrelated mutation with n-step sizes
and correlated mutation using a covariance matrix [28]. In the rst mutation,
the same distribution is used to mutate each problem variable, in the second,
dierent step sizes are used for dierent dimensions, while in the last mutation,
not only the magnitude in each dimension but also the direction (i.e., rotation
angle) are taken into account [33,11,12].
This paper proposes a modied mutation operator for ES, where each problem
variable in representation of individual is widen with three strategy parameters,
i.e., step size, shift angle and reverse sign determining the sign of change. In other
words, if the sign is positive the change is added to the corresponding problem
variable, otherwise the change is subtracted. Using these strategy parameters,
the magnitude, direction and sign of the change is determined. The motivation
behind the mutation operator is that each element of the solution vector explores
the search space independently of the other elements, but the way of exploring
this space depends on the strategy parameters. In fact, this mutation is uncor-
related. On the other hand, each element of solution vector is represented by a
four dimensional vector (also 4D-vector). Although mathematical structures for
description of four dimensional vector spaces already exist, like quaternions [17]
or 4-vectors [29], these structures refer to well dened algebras and are therefore
too complex for our needs. As a matter of fact, the modied mutation is referred
to an uncorrelated mutation with n 4D-vectors.
Consequently, the proposed self-adaptive evolution strategy using the uncorre-
lated mutation with n 4D-vectors (4SA-ES) was applied to test suite consisting
of ten well-known functions taken from publications. The obtained results of
4SA-ES were compared with the results of the original ESs that implement the
uncorrelated mutation with one step sizes, on the one hand and the uncorre-
lated mutation with n step size, on the other hand. Additionally, the other EA
and SI algorithms, like dierential evolution [13] (DE), self-adaptive dieren-
tial evolution [14] (jDE), and hybrid self-adaptive bat algorithms [18] (HSABA)
are also included in this comparative study. The results of the 4SA-ES showed
On the Mutation Operators in Evolution Strategies 71
that the DE, despite its maturity, still remains not completely discovered and
nevertheless oers many directions for developments also in the future. Further-
more, this contribution of self-adaptation could be applied also in contemporary
nature-inspired algorithms, i.e., the domain that usually forgets on the valuable
features of ES.
The structure in the remainder of this paper is as follows. Section 2 deals
with backround information needed by the reader. In line with this, the mutation
operator in ES were discussed in detail. Section 3 describes the proposed 4SA-ES.
In Section 4, the experiments and results are discussed. The paper concludes with
Section 5, where the possible direction for further development of this algorithm
are discussed.
for, where n denotes the number of problem variables. In order to implement the
self-adaptation of strategy parameters in ES, these parameters are added to a
representation of problem variables and become a subject of acting the variations
operators crossover and mutation.
Typically, the strategy parameters of mutation strengths are self-adapted
in the SA-ES. These parameters enable to calculate the mutation step sizes to
determine a magnitude and direction of changes applied to the corresponding
problem variables. The number of mutation strengths variables depends on the
mutation type and can vary from a single value , to n values i dedicated to
each problem variable xi , or even to the matrix of rotation angles ij , when the
correlation mutation is used. In general, the individual is represented, as [28]:
x1 , . . . , xn , 1 , . . . , n , 1 , . . . , n , (1)
2.3 Mutation
Typically, strategy parameters are used by mutation operator. This operator
bases on normal (also Gaussian) distribution that depends on two parameters,
i.e., mean value and standard deviation (t) . Mutation adds to each problem
(t)
variable xi a mutation step size xi that is randomly drawn from the normal
distribution N(, ) with corresponding probability density [36].
(t)
(t) 1 (xi )2
p(xi ) = exp . (2)
(t) 2 2 2
In practice, the mean value is set to zero ( = 0), while the candidate solution
(t) (t) (t)
x1 , . . . , xn , (t)
is mutated such that each value of variable xi is modied,
as
(t+1) (t)
xi = xi + N(0, (t+1) ), (3)
(t+1) (t+1)
where denotes the updated value of mutation strengths and N(0, )
is the random value drawn from normal distribution with mean of zero and
standard deviation (t+1) . On the other hand, a specialty of mutation in SA-ES
On the Mutation Operators in Evolution Strategies 73
is that the mutation strengths determined by (t+1) are part of selection and
undergo acting the variation operators. The slightly simplied version of the
Eq. 3 can be employed, as
(t+1) (t)
xi = xi + (t+1) N(0, 1), (4)
where is assumed that N(0, ) = N(0, 1), and N(0, 1) denotes the
(t+1) (t+1)
random value drawn from normal distribution with mean of zero and standard
deviation (t+1) .
Uncorrelated Mutation with One Step Size. In the simplest case, a single
(t)
mutation strength is applied to all problem variables xi and thus the representa-
(t) (t)
tion of individuals is reduced to x1 , . . . , xn , (t) . From the candidate solution,
(t+1) (t+1)
mutation generates the modied values of ospring x1 , . . . , xn , (t+1)
according to the following equations:
(t+1) (t)
xi = xi + (t+1) Ni (0, 1). (6)
In Eq. (5), the parameter denotes a learning rate similar to articial neural
networks (ANN) [30] set by user. Usually, this parameter is set proportionally
to the square root of the problem size, i.e., = 1/ (n). Eq. (6) shows how
(t+1)
the mutated values of program variables xi are obtained. Actually, the term
(t+1)
i = (t+1) Ni (0, 1) in the equation denotes the mutation step size of i-th
problem variable. Thus, the mutation strengths are not explicitly controlled by
the user, but are a part of solution evolving during the evolutionary process.
In order to prevent that the standard deviation comes too near to zero, the
following limitation is used:
< 0 = 0 . (7)
This mutation makes the variable step sizes possible to use. At the beginning
of the evolutionary process, when the whole search space is explored, larger step
sizes are needed. On the other hand, when the population is directed towards
the optimal solutions the step sizes can become smaller.
(t) (t)
to problem variables x1 , . . . , xn . As a result, the candidate solution is repre-
(t) (t) (t) (t)
sented as x1 , . . . , xn , 1 , . . . , n , while the mutation operation is described
using the following equations:
n
size of each solution is increased to ( 2 + 2 n) on this way, where the
The
( n2 + n) strategy parameters are self-adapted during the mutation [37]. The
mutation operator obeys the following update rules [19]:
(t+1) (t)
j = j + Nj (0, 1)), (11)
(t+1)
x(t+1) = x(t) + Ni 0, C(i , (t+1) ) , (12)
(t+1)
where Ni 0, C(i , (t+1) ) is an implementation of a normally distributed
correlated mutation vector with a zero mean vector and a covariance matrix
5
C. The parameter is xed to 5 (i.e., 180 0.0873) according to Schwe-
fel [20]. The following limitation is used in order to prevent a rotation angle j
to cross the border of the interval [, ]. The values of and are set similarly
to uncorrelated mutation with n step sizes. The rotation angles are initialized
randomly in the interval j [0, ].
In summary, the mutation in ES is an asexual operator symbolically described
as
m, , : S S, (13)
where , , are predened constants.
On the Mutation Operators in Evolution Strategies 75
2.4 Crossover
Although the rst ES has employed only a mutation operator for modifying the
candidate solution, the contemporary ES employ the crossover operator as well.
In contrast to classical EAs, crossovers in ES generate one ospring from two
parents. There are two types of ES crossovers in general, i.e., discrete and arith-
metic. The discrete crossover selects the value for ospring randomly between
values laying on the same position in the parent chromosomes. The arithmetic
crossover determines the value for ospring from values laying on the same po-
sition in the parent chromosomes according to the following equation [37]:
where the parameter can occupy values from the interval [0 . . . 1]. When
= 0.5, the corresponding crossover becomes uniform arithmetic crossover [36].
Symbolically, the ES crossover operator is presented as
c : S S S, (15)
which determines that one ospring is generated from two randomly selected
candidate solutions.
s(,) : S S , (16)
s(+) : S + S , (17)
best solutions are selected from union of parents and ospring for the next
generation. Although the ( + )-selection preserves the best solutions in the
current population, the (, )-selection is more recommended for self-adaptation
because of [36]:
in dynamic environments, the (, )-selection preserves outdated solutions
and prevents the current optimum to be moved.
the (, )-selection is suitable to forget good solutions and therefore is not
sensitive to get stuck into a local optimum.
the ( + )-selection enables the mis-adapted strategy parameters to survive
for a relatively large number of generations and thus hinders the self-adaptive
mechanism highlighting that both (i.e., problem variables as well as strategy
parameters) need to improve during the evolutionary search process.
On the other hand, the selective pressure in SA-ESs is very high because the
recommended ratio of / 1/7 indicating is much higher than [28].
76 I. Fister Jr. and I. Fister
Note that this pseudo-code summarizes all strategies to self-adapt at least one
strategy parameter. From an implementation point of view, it is important to nd
so named evolution window, where the evolutionary search can progress [36]. The
window is connected with the proper choosing the order of magnitude for strat-
egy parameters , within the reasonable performance is observed. The proper
identication of this parameter is sometimes connected to an extensive experi-
mental work.
In summary, the larger mutation step sizes in the beginning of the evolutionary
search enable exploring wide regions of the search space. When the evolution-
ary search becomes matured and the promising regions are already located, the
evolutionary search can direct itself on exploring those regions.
A motivation guiding our experimental work was to elaborate a modied
mutation operator on the basis of self-adapting the parameters using the normal
distribution. An aim of this experimentation was to develop candidate solutions
independently of each other throughout the search space. In line with this, it
is expected that the exploration power of the evolutionary search process is
increased using this operator especially by problems, where the problem variables
are uncorrelated. In these cases, the search process, armed with this operator,
can be much better adapted to the tness landscape, determined by the problem
to be solved.
The candidate solution is now represented as
x1 , x2 , . . . , xn , (18)
(t) (t) (t) (t) (t) (t)
where each step size vector is in form as xi = xi , i , i , ri , where xi
(t) (t)
denotes the problem variable, i the mutation strength, i the rotation angle
(t)
and ri the boolean variable denoting a reection, which reverse the sign of
change when the value of this variable is set to true. The meaning of the other
variables are similar to the correlated mutation.
The modied mutation mechanism is realized by the following equations:
= r (ri ),
(t+1) (t)
ri (19)
(t+1) (t)
i = i + Ni (0, 1)), (20)
where U(0, 1) is the uniformly distributed random value drawn from the interval
[0, 1] and [0, 1] some prescribed constant value determining the probability
78 I. Fister Jr. and I. Fister
of reversing the sign (also learning rate). Eects of this mutation mechanism in
two dimensions is illustrated in Fig. 1.
The uncorrelated mutation with n step sizes uses the normal distribution
N(0, ) (bell curve with location zero). In contrast, the proposed mutation sup-
ports the general form of the normal distribution N(, ) to provide changing
the location of normal distribution N(0, ) for a shift angle (bell curve with
location ). When > 0 the positive changes were preferred, while < 0 the neg-
ative. The standard deviation on the gure presents an inection point, where
the concavity changes a sign from plus to minus.
and crossover type, population model, the function evaluations to solution (FEs),
property of mutation (pm ), property of crossover (pc ), starting value of mutation
strengths (0 ) and the minimum value of mutation strengths (0 ), for each of
corresponding ES version.
From the table, it can be seen that the same values are used by almost all algo-
rithms, except the mutation type and starting value of mutation strength. Learn-
ing rates were
similar in all mentioned ES versions, as follows: = 1/ 2 n,
and = 1/ 2 n. Setting the value of learning rate is discussed later in this
paper.
The 25 independent runs were conducted by each ES algorithm, while the
minimum, maximum, average, median and standard deviation values were accu-
mulated. As a part of experimental work, the following ve tests were conducted:
In the remainder of this chapter, the test suite is described. Then, the results
of the mentioned ve tests are presented. The section concludes with discussion
of the performed work.
The benchmark suite was composed of ten well-known functions, selected from
various publications. The denitions of the benchmark functions are summarized
in Table 2, where the function name and its corresponding denition can be seen.
However, reader is invited to check a deep details about test functions in the
state-of-the art reviews [21,22,23].
Each function in the table is tagged with its sequence number from f1 to f10 .
Properties of the benchmark functions can be seen in Table 3 consisting of ve
80 I. Fister Jr. and I. Fister
Function Denition
n
x2
Griewangk f1 (x) = cos
i=1
xi
+ n i
i=1 4000 + 1
n i 2
Rastrigin f2 (x) = n 10 + i=1 (xi 10 cos(2 + xi ))
Rosenbrock f3 (x) = n1 2 2
(xi+1 xi ) + (xi 1)
i=1 100
2
n1 0.2 0.5(xi+1 +x2
2
i)
Ackley f4 (x) = i=1 20 + e20 e e0.5(cos(2xi+1 )+cos(2+xi+1 )+cos(2+xi ))
Schwefel f5 (x) == 418.9829 D D i=1 si sin( |si |)
D 2
De Jong f6 (x) = i=1 xi
D
Easom f7 (x) = (1)D ( D 2
i=1 cos (xi )) exp[ i=1 (xi ) ]
2
D 2
ixi 210
Michalewicz f8 (x) = i=1 sin(xi )[sin( )]
Yang f9 (x) = ( D |xi |) exp[ D sin(x2i )]
i=1 2 D i=1 2 4
Zakharov f10 (x) = i=1 xi + ( 2 i=1 ixi ) + ( 12 D
D 1
i=1 ixi )
f f x Domain Characteristics
f1 0 (0, 0, . . . , 0) 600, 600 Highly multi-modal
f2 0 (0, 0, . . . , 0) 15, 15 Highly multi-modal
f3 0 (1, 1, . . . , 1) 15, 15 Several local optima
f4 0 (0, 0, . . . , 0) 32.768, 32.768 Highly multi-modal
f5 0 (0, 0, . . . , 0) 500, 500 Highly multi-modal
f6 0 (0, 0, . . . , 0) 600, 600 Uni-modal, convex
f7 -1 (, , . . . ) 2, 2 Several local optima
f8 -1.80131 (2.20319, 1.57049)1 0, Several local optima
f9 0 (0, 0, . . . , 0) 2, 2 Several local optima
f10 0 (0, 0, . . . , 0) 5, 10 Uni-modal
columns: the function tag f , the value of the optimal solution f , the optimal
solution x , the parameter domains and the function characteristics.
Parameter domains limit the values of parameters into interval between their
lower and upper bounds. As a matter of fact, these determine the size of the
search space. In order to make the problems heavier to solve, the parameter
domains were selected wider than those prescribed in the standard publications.
Additionally, the problem becomes also heavier to solve when the dimensionality
of the benchmark functions are increased. As a result, benchmark functions of
more dimensions need to be optimized in the experimental work.
One of the more important characteristics of the function is the number of
local and global optima. According to this characteristics the functions are di-
vided either into uni-modal or multi-modal. The former type of functions has
only one global optimum, while the latter is able to have more local and global
optima thrown across the whole search space.
1
Valid for 2-dimensional parameter space.
On the Mutation Operators in Evolution Strategies 81
Evals Meas. f1 f2 f3 f4 f5
Best 3.01E+002 3.94E+003 7.23E+007 1.97E+001 6.49E+003
Worst 1.66E+001 4.47E+002 5.53E+005 9.31E+000 9.30E+002
1.20E+03 Mean 8.15E+000 1.74E+002 7.55E+002 2.84E+000 2.40E+002
Median 5.77E+002 5.47E+003 2.74E+008 2.07E+001 9.59E+003
StDev 4.55E+001 6.43E+002 4.75E+006 1.24E+001 2.37E+003
Best 3.02E+001 2.52E+002 1.97E+003 3.40E+000 1.23E+003
Worst 4.21E+002 4.74E+003 1.62E+008 2.03E+001 8.08E+003
6.00E+03 Mean 2.81E+001 5.31E+002 1.32E+006 1.08E+001 1.83E+003
Median 1.64E+001 2.23E+002 1.35E+003 3.24E+000 8.04E+002
StDev 4.27E+002 4.77E+003 1.55E+008 2.03E+001 8.25E+003
Best 2.24E+001 5.16E+002 9.91E+005 1.09E+001 1.88E+003
Worst 1.22E+001 2.19E+002 1.21E+003 3.28E+000 8.27E+002
3.00E+04 Mean 8.37E+001 4.48E+002 6.67E+007 3.33E-001 1.04E+003
Median 1.07E+001 5.26E+001 1.22E+006 1.08E+000 4.09E+002
StDev 8.46E+000 2.42E+001 4.10E+002 1.74E-001 3.09E+002
Evals Meas. f6 f7 f8 f9 f10
Best 1.22E+006 0.00E+000 8.79E-001 5.15E-004 4.00E+002
Worst 6.10E+004 0.00E+000 3.82E-001 3.34E-007 2.33E+002
1.20E+03 Mean 2.46E+004 0.00E+000 6.95E-003 4.63E-009 3.98E+001
Median 1.86E+006 0.00E+000 3.00E+000 5.71E-001 1.34E+004
StDev 1.39E+005 0.00E+000 1.15E+000 3.37E-005 6.13E+002
Best 7.49E+004 0.00E+000 2.88E-001 1.55E-007 8.41E+001
Worst 1.57E+006 0.00E+000 1.83E+000 1.23E-001 2.85E+003
6.00E+03 Mean 9.91E+004 0.00E+000 6.99E-001 1.71E-005 3.42E+002
Median 5.21E+004 0.00E+000 2.08E-001 4.16E-008 5.82E+001
StDev 1.62E+006 0.00E+000 1.67E+000 5.30E-003 1.07E+003
Best 9.46E+004 0.00E+000 5.27E-001 1.89E-005 2.87E+002
Worst 5.13E+004 0.00E+000 2.04E-001 3.20E-008 5.00E+001
3.00E+04 Mean 2.11E+005 0.00E+000 7.10E-001 2.07E-001 3.94E+003
Median 2.32E+004 0.00E+000 3.13E-001 1.27E-005 1.31E+002
StDev 1.61E+004 0.00E+000 8.54E-002 4.20E-008 1.61E+001
according to the observed measures were tracked at three dierent stages, i.e.,
1
at the 25 , the 15 of tness function evaluations and the end of run. The results of
the mentioned test for benchmark functions of dimensions n = 30 are presented
in Table 4, where the minimum values are presented in rows Best, maximum
in rows Worst, average in rows Mean, median in rows Median, and standard
deviation in rows StDev.
From the table, it can be seen that the ES-3 algorithm is not an elitist, because
of not preserving the best results. For instance, the mean values of the most
observed functions decrease, when the number of tness function evaluations
increases. However, this is a characteristic of the used (, ) strategy that tries
to facilitate extinction of miss-adapted solutions [36]. On the other hand, this
characteristic of the mean values indicate that the number of the tness function
evaluations was underestimated by the ES-3 algorithm during the test.
D Meas. f1 f2 f3 f4 f5
Best 1.22E+000 3.90E+001 3.39E+001 1.69E+000 9.62E+001
Worst 7.11E+000 7.35E+001 8.34E+002 2.46E+000 4.09E+002
10 Mean 4.03E+000 5.52E+001 3.21E+002 2.12E+000 2.41E+002
Median 3.85E+000 5.60E+001 2.07E+002 2.09E+000 2.37E+002
StDev 1.73E+000 1.09E+001 2.89E+002 2.47E-001 1.01E+002
Best 8.15E+000 1.74E+002 7.55E+002 2.84E+000 2.40E+002
Worst 3.02E+001 2.52E+002 1.97E+003 3.40E+000 1.23E+003
30 Mean 1.64E+001 2.23E+002 1.35E+003 3.24E+000 8.04E+002
Median 1.22E+001 2.19E+002 1.21E+003 3.28E+000 8.27E+002
StDev 8.46E+000 2.42E+001 4.10E+002 1.74E-001 3.09E+002
Best 1.50E+001 3.81E+002 1.77E+003 3.51E+000 1.53E+003
Worst 4.13E+001 4.65E+002 6.21E+003 3.82E+000 2.75E+003
50 Mean 2.46E+001 4.19E+002 3.12E+003 3.69E+000 2.10E+003
Median 2.27E+001 4.10E+002 2.57E+003 3.70E+000 1.87E+003
StDev 7.44E+000 2.67E+001 1.35E+003 1.02E-001 4.58E+002
Evals Meas. f6 f7 f8 f9 f10
Best 3.79E+003 0.00E+000 0.00E+000 1.46E-003 1.42E+000
Worst 1.94E+004 0.00E+000 4.92E-016 2.94E-003 3.00E+000
10 Mean 9.22E+003 0.00E+000 4.95E-017 2.15E-003 2.07E+000
Median 7.79E+003 0.00E+000 1.62E-025 1.87E-003 1.81E+000
StDev 4.32E+003 0.00E+000 1.55E-016 5.53E-004 5.79E-001
Best 2.46E+004 0.00E+000 6.95E-003 4.63E-009 3.98E+001
Worst 7.49E+004 0.00E+000 2.88E-001 1.55E-007 8.41E+001
30 Mean 5.21E+004 0.00E+000 2.08E-001 4.16E-008 5.82E+001
Median 5.13E+004 0.00E+000 2.04E-001 3.20E-008 5.00E+001
StDev 1.61E+004 0.00E+000 8.54E-002 4.20E-008 1.61E+001
Best 8.32E+004 0.00E+000 6.97E-001 1.55E-012 2.68E+002
Worst 1.57E+005 0.00E+000 1.50E+000 3.91E-011 4.39E+002
50 Mean 1.15E+005 0.00E+000 9.48E-001 1.19E-011 3.36E+002
Median 9.76E+004 0.00E+000 8.36E-001 4.39E-012 3.16E+002
StDev 3.11E+004 0.00E+000 2.56E-001 1.29E-011 5.68E+001
3000.0 120000000.0
Best Best
Worst Worst
2500.0 Average 100000000.0 Average
Values of fitness function
1500.0 60000000.0
1000.0 40000000.0
500.0 20000000.0
0.0 0.0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Calculation progress [%] Calculation progress [%]
12.0 800000.0
10.0 600000.0
8.0
400000.0
6.0
200000.0
4.0
2.0 0.0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Calculation progress [%] Calculation progress [%]
2.5 1000.0
2.0 800.0
1.5 600.0
1.0 400.0
0.5 200.0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Calculation progress [%] Calculation progress [%]
As can be seen from Table 6, the jDE and HSABA algorithms reached the
best results four times. The former achieved the best results by optimizing the
functions f1 , f4 , f6 , and f8 , while the latter by optimizing the functions f2 , f3 , f5
and f9 . The original DE algorithm outperformed results of the other algorithms
in test by optimizing the function f10 , while by optimizing the function f7 all
algorithms achieved similar results.
In order to evaluate these results statistically, the Friedman non-parametric
tests were also performed according to ve measurements (minimum, maximum,
mean, median and standard deviation) obtained over 25 runs for each function.
Three Friedman non-parametric tests were conducted in order to capture the
behavior of dierent algorithms, according to the dimensions of the problems.
86 I. Fister Jr. and I. Fister
The Friedman test [26,27] compares average ranks of the algorithms. A null-
hypothesis states that two algorithms are equivalent and, therefore, their ranks
should be equal. If the null-hypothesis is rejected, i.e., the performance of the
algorithms is statistically dierent, the Bonferroni-Dunn test [25] is performed
to calculate the critical dierence between the average ranks of those two algo-
rithms. When the statistical dierence is higher than the critical dierence, the
algorithms are signicantly dierent. The equation for the calculation of critical
dierence can be found in [25].
Friedman tests were performed using the signicance level 0.05. The results
of the Friedman non-parametric test are presented in Fig. 4, being divided into
three diagrams to show the normalized ranks and condence intervals (critical
dierences) for the algorithms under consideration. As closer the normalized rank
to value one, the better the algorithm. The diagrams are organized according to
the dimensions of functions. Two algorithms are signicantly dierent if their
intervals in Fig. 4 do not overlap.
The rst diagram in Fig. 4 shows that the HSABA algorithm outperforms the
other algorithms in test. Moreover, the results of HSABA are signicantly better
than the results of any other algorithm. Furthermore, the results of jDE, DE and
ES-3 are signicantly better than the results of the original ES-1. Interestingly,
On the Mutation Operators in Evolution Strategies 87
HSABA
jDE
DE
ES-3
ES-2
ES-1
0 1 2 3 4 5 0 1 2 3 4 0 1 2 3 4
Average rank (D=10) Average rank (D=30) Average rank (D=50)
the situation is dierent when taking the next two diagrams (by n = 30 and
n = 50) into consideration. In the rst case, the HSABA and jDE signicantly
improve the results of the other ES algorithms in the test, while in the second
case, the jDE and HSABA outperform the results of ES-1 and ES-2 signicantly.
Comparing the ES algorithms only, it can be seen that the ES-3 outperforms
the results of the ES-1 signicantly, and the ES-2 substantially by the observed
dimensions.
5 Conclusion
References
1. Back, T., Hammel, U., Schwefel, H.-P.: Evolutionary Computation: Comments on
the History and Current State. IEEE Trans. Evolutionary Computation 1(1), 317
(1997)
2. Beyer, H.-G., Deb, K.: On self-adaptive features in real-parameter evolutionary
algorithms. IEEE Trans. Evolutionary Computation 5(3), 250270 (2001)
3. Eiben, A.E., Hinterding, R., Michalewicz, Z.: Parameter control in evolutionary
algorithms. Trans. Evol. Comp. 3(2), 124141 (1999)
4. Hinterding, R., Michalewicz, Z., Eiben, A.E.: Adaptation in Evolutionary Com-
putation: A Survey. In: Proceedings of the Fourth International Conference on
Evolutionary Computation (ICEC 1997), pp. 6569 (1997)
5. Shaefer, C.G., Grefenstette, J.J.: The ARGOT strategy: Adaptive representation
genetic optimizer technique. In: Proc. 2nd Int. Conf. Genetic Algorithms and Their
Applications, pp. 5055 (1987)
6. Schaer, J.D., Morishima, A., Grefenstette, J.J.: An adaptive crossover distribution
mechanism for genetic algorithms. In: Proc. 2nd Int. Conf. Genetic Algorithms and
Their Applications, pp. 3640 (1987)
7. Spears, W.M., McDonnell, J.R., Reynolds, R.G., Fogel, D.B.: Adapting crossover
in evolutionary algorithms. In: Proc. 4th Annu. Conf. Evolutionary Programming,
pp. 367384. MIT Press (1995)
8. Srinivas, M., Patnaik, L.M.: Adaptive probabilities of crossover and mutation in
genetic algorithms. IEEE Trans. Syst., Man, and Cybern. 240(4), 1726 (1994)
9. Tuson, A., Ross, P., Voigt, H.-M., Ebeling, W., Rechenberg, I., Schwefel, H.-P.: Cost
based operator rate adaptation: An investigation. In: Ebeling, W., Rechenberg, I.,
Voigt, H.-M., Schwefel, H.-P. (eds.) PPSN 1996. LNCS, vol. 1141, pp. 461469.
Springer, Heidelberg (1996)
10. Zhang, J., Chen, W.-N., Zhan, Z.-H., Yu, W.-J., Li, Y.-L., Chen, N., Zhou, Q.: A
survey on algorithm adaptation in evolutionary computation. Frontiers of Electrical
and Electronic Engineering 7(1), 1631 (2012)
11. Auger, A., Hansen, N.: A restart CMA evolution strategy with increasing popu-
lation size. In: Proceedings of the 2005 Congress on Evolutionary Computation,
pp. 17691776 (2005)
12. Igel, C., Hansen, N., Roth, S.: Covariance matrix adaptation for multi-objective
optimization. Evolutionary Computation 15(1), 128 (2007)
13. Storn, R., Price, K.: Dierential Evolution: A Simple and Ecient Heuristic
for Global Optimization over Continuous Spaces. Journal of Global Optimiza-
tion 11(4), 341359 (1997)
14. Brest, J., Greiner, S., Boskovic, B., Mernik, M., Zumer, V.: Self-adapting control
parameters in dierential evolution: A comparative study on numerical benchmark
problems. IEEE Transactions on Evolutionary Computation 10(6), 646657 (2006)
15. Back, T.: An Overview of Parameter Control Methods by Self-Adaptation in Evo-
lutionary Algorithms. Fundam. Inf. 35(1-4), 5166 (1998)
On the Mutation Operators in Evolution Strategies 89
16. Fister, I., Mernik, M., Filipic, B.: Graph 3-coloring with a hybrid self-adaptive
evolutionary algorithm. Comp. Opt. and Appl. 54(3), 741770 (2013)
17. Fister, I., Yang, X.-S., Brest, J., Fister Jr., I.: Modied rey algorithm using
quaternion representation. Expert Systems with Applications 40(18), 72207230
(2013)
18. Fister, I., Fong, S., Brest, J., Fister Jr., I.: A novel hybrid self-adaptive bat algo-
rithm. The Scientic World Journal, 112 (2014)
19. Hansen, N., Ostermeier, A.: Completely derandomized self-adaptation in evolution
strategies. Evolutionary Computation 9(2), 159195 (2011)
20. Schwefel, H.-P.: Collective intelligence in evolving systems. In: Wol, W., Soeder,
C.J., Drepper, F. (eds.) Ecodynamics - Contributions to Theoretical Ecology,
pp. 95100. Springer, Berlin (1987)
21. Jamil, M., Yang, X.-S.: A literature survey of benchmark functions for global opti-
misation problems. International Journal of Mathematical Modelling and Numer-
ical Optimisation 4(2), 150194 (2013)
22. Yang, X.-S.: Appendix A: Test Problems in Optimization. Engineering Optimiza-
tion, pp. 261266. John Wiley & Sons, Inc., Hoboken (2010)
23. Yang, X.-S.: Firey algorithm, stochastic test functions and design optimisation.
International Journal of Bio-Inspired Computation 2(2), 7884 (2010)
24. Hansen, N.: The CMA Evolution Strategy: A Comparing Review. In: Lozano, J.A.,
Larranaga, P., Inza, I., Bengoetxea, E. (eds.) Towards a New Evolutionary Compu-
tation. Advances on Estimation of Distribution Algorithms, pp. 75102. Springer,
Berlin (2006)
25. Demsar, J.: Statistical Comparisons of Classiers over Multiple Data Sets. Journal
of Machine Learning Research 7, 130 (2006)
26. Friedman, M.: The use of ranks to avoid the assumption of normality implicit in the
analysis of variance. Journal of the American Statistical Association 32, 675701
(1937)
27. Friedman, M.: A comparison of alternative tests of signicance for the problem of
m rankings. The Annals of Mathematical Statistics 11, 8692 (1940)
28. Eiben, A., Smith, J.: Introduction to Evolutionary Computing. Springer, Berlin
(2003)
29. Wachter, A., Hoeber, H.: Compendium of Theoretical Physics. Springer, Berlin
(2006)
30. Russell, S., Norvig, P.: Articial Intelligence: A Modern Approach, 3rd edn. Pren-
tice Hall, New Jersey (2009)
31. Rechenberg, I.: Evolutionsstrategie, Optimierung technischer Systeme nach
Prinzipien der biologischen Evolution. Frommann-Holzboog, Stuttgart (1973)
32. Schwefel, H.P.: Numerische Optimierung von Computer-Modellen mittels der Evo-
lutionsstrategie. Birkhauser, Basel (1977)
33. Hansen, N.: The CMA evolution strategy: A tutorial. Vu le 29 (2005)
34. Baeck, T., Fogel, D.B., Michalewicz, Z.: Handbook of Evolutionary Computation.
Taylor & Francis (1997)
35. Beyer, H.-G.: The Theory of Evolution Strategies. Springer, Heidelberg (2001)
36. Back, T.: Evolutionary Algorithms in Theory and Practice: Evolution Strategies,
Evolutionary Programming, Genetic Algorithms. Oxford University Press, Oxford
(1996)
37. Deb, K.: Multi-Objective Optimization Using Evolutionary Algorithms. John Wi-
ley & Sons, Inc., New York (2001)
38. Ahsanullah, M., Kibria, B.M.G., Shakil, M.: Normal and students t distributions
and their applications. Springer, Paris (2014)
Adaptation in Cooperative Coevolutionary
Optimization
Giuseppe A. Truno
1 Introduction
The search for more eective and ecient optimization algorithms is an increas-
ingly important research topic, given the complexity of todays applications in
many elds of science and engineering. For this reason, in recent years a variety
of optimization metaheuristics have been developed, which have shown excellent
performance in many relevant real-world problems. However, most of these al-
gorithms are plagued by the so-called curse of dimensionality, which consists
of a rapid deterioration of their optimization capability as the dimensionality of
the problem increases.
Among the dierent approaches that were proposed in literature for dealing
with large-scale optimization problems, there is the Cooperative Coevolution-
ary (CC) strategy introduced in [1] and later developed by many researchers
[2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18]. In brief, the CC idea consists of de-
composing the original high-dimensional problem into a set of lower-dimensional
subproblems, which are easier to solve. Typically, to each subproblem is assigned
a subpopulation of candidate solutions, which is evolved according to the adopted
optimization metaheuristic. During the process, the only cooperation happens
in the evaluation of the tness, through an exchange of information between
subpopulations.
Corresponding author.
2 Cooperative Coevolution
The CC approach was rst applied to a Genetic Algorithm by Potter and De Jong
in [1]. Subsequently, the idea has attracted a signicant amount of research and
has been adopted in many search algorithms such as Ant Colony Optimization
[19,20], Particle Swarm Optimization (PSO) [21,5,22,15,14], Simulated Anneal-
ing [23,24], Dierential Evolution (DE) [25,7], Firey Algorithm (FA) [26,27,18]
and many others. A CC optimization is based on partitioning the d-dimensional
set of search directions G = {1, 2, . . . , d} into k sets G1 . . . Gk . Each group Gi
of directions denes a new search space S (i) in which a standard optimization
algorithm is applied. In such an approach, the whole search procedure is then
decomposed into k subcomponents associated to dierent sub-spaces whose di-
mension can be signicantly lower than d. For example, using a population-based
metaheuristic, a separate sub-population is assigned to each subcomponent gen-
erated by the groups Gi . By construction, a candidate solution in S (i) contains
only some elements of the d-dimensional vector required for computing the cor-
responding tness function f . For this reason, a common d-dimensional context
vector b is built using a representative individual (e.g. the best individual) pro-
vided by each subcomponent. Then, the candidate solutions are evaluated by
complementing them through the appropriate elements of the context vector. In
this framework, the cooperation between sub-populations emerges because the
common vector is used for the tness evaluation of all individuals.
In their original paper, Potter and De Jong [1] proposed to decompose a
d-dimensional problem into d sub-populations (i.e. Gi = {i}). The tness of
each individual was computed by evaluating the d-dimensional vector formed
by the individual itself and a selected member (e.g. the current best) from each
of the other sub-populations. The authors showed the eectiveness of the pro-
posed approach on several test functions, although the empirical evaluation was
conducted only on search spaces of up to 30 dimensions.
Later, Liu et al. [4] investigated the performances of the same cooperative
approach applied to an evolutionary programming [28] algorithm. The results
obtained on benchmark functions with 100 to 1000 dimensions were satisfactory
and the authors showed that the CC approach can signicantly improve the
scalability of the optimizer as the dimensionality of the problem increases.
Adaptation in Cooperative Coevolution 93
Subsequently, the Potter and De Jongs idea was applied to PSO [21] by
Van den Bergh and Engelbrecht in [5], where the authors introduced the
decomposition of the original d-dimensional search space into k subspaces S (i)
of the same dimension dk = d/k. In other words, in such an approach the groups
of dimensions were dened as:
Gi = {(i 1) dk + 1, . . . , i dk }
and the context vector is:
b = (b1 , . . . , bdk , b1 , . . . , bdk , . . . , b1 , . . . , bdk )T
(1) (1) (2) (2) (k) (k)
b(1) b(2) b(k)
94 G.A. Truno
Algorithm 1. CC(f , n)
1 G = {G1 , . . . , Gk } grouping(n);
2 pop initPopulation();
3 contextV ector initContextVector(pop);
4 f itnessEvaluations 0;
5 while f itnessEvaluations < M axF E do
6 foreach Gi G do
7 popi extractPopulation(pop, Gi );
8 besti optimizer(f , popi , contextV ector, Gi , maxF ESC);
9 pop storePopulation(popi , Gi );
10 f itnessEvaluations f itnessEvaluations + maxF ESC;
11 contextV ector updateContextVector(besti , Gi );
where b(i) is the dk -dimensional vector representing the contribution of the i-th
sub-population (e.g., its current best position in the subspace S (i) ):
, . . . , xdk )T
(i,j) (i,j) (i,j)
x(i,j) = (x1 , x2
, . . . , xdk , . . . , b1 , . . . , bdk )T
(1) (1) (i,j) (i,j) (k) (k)
b(i,j) = (b1 , . . . , bdk , . . . , x1
b(1) x(i,j) b(k)
In other words, the tness of x(i,j) is evaluated on the vector obtained from b
by substituting the components provided by the i-th sub-population with the
corresponding components of x(i,j) .
Except for this way of evaluating the individuals, the CC proceeds using the
standard optimizer in each subspace. Algorithm 1 outlines a possible basic CC
optimization process. First, a decomposition function creates the k groups of
directions. Then the population and the context vector are randomly initialized.
The optimization is organized in cycles. During each cycle, the optimizer is acti-
vated in a round-robin fashion for the dierent subcomponents and the context
vector is updated using the current best individual of each sub-population. A
budget of maxFESC tness evaluations is allocated to each subcomponent at
each cycle. The CC cycles terminate when the number of tness evaluations
reaches the value maxFE. Note that several variants to this scheme can be pos-
sible. For example, the context vector could be updated in a synchronous way
at the end of each cycle.
In the CC framework, many design aspects can have a signicant impact on
the optimizing performance. The main issue was early recognized in [1,3] and
Adaptation in Cooperative Coevolution 95
later conrmed by several studies (e.g. [4]): when interdependent variables are as-
signed to dierent subcomponents the search eciency can decline signicantly.
The interdependency between decision variables is a common condition in
real optimization problems and in literature is referred to as non-separability
[29] or epistasis, that is gene interaction. Basically, separability means that the
inuence of a variable on the tness value is independent of any other variables.
More formally, following [30] a function f : Rd R is separable i :
arg min f (x1 , . . . , xd ) = arg min f (x1 , . . . ), . . . , arg min f (. . . , xd ) (1)
x1 , ..., xd x1 xd
the sizes of the subspaces assigned to the dierent sub-populations. Such a co-
operative PSO outperformed some state-of-the-art evolutionary algorithms on
complex multi-modal problems. Also, in [15] the author used the same coopera-
tive approach with RG in a micro-PSO, showing that even using a small number
of individuals per sub-population (i.e. 5) the algorithm was very ecient on
high-dimensional problems. Recently, in [18] the RG approach has been applied
to a CC version of the FA [26,27].
Besides the RG strategy, further research eorts have been devoted to improve
the CC framework using adaptive approaches.
A relevant contribution in this directions was the automatic adaptation of the
subcomponent sizes, initially proposed in [6] and recently improved in [32].
Also, starting from the initial proposal of [2], many research works devised
strategies for automatically decomposing the problem into subcomponents in or-
der to account for the interdependencies between variables (e.g. [9,10,13,16,17]).
Another aspect strictly related with the problem decomposition, concerns the
possible imbalance between the contributions of the dierent subcomponents to
the improvement of the tness at each cycle. A rst proposal to cope with this
issue was presented in [12], where the authors proposed an adaptive approach
that determines the computation to spend on the subcomponents according to
their contributions to the tness. A preliminary version of a new approach is
presented in Section 3.4 of this chapter.
In the following we outlines some of the above mentioned adaptive techniques
for the enhancement of the CC approach.
3 Adaptation in CC
3.1 Random Grouping
Even if it could not be classied as an example of adaptation, the RG strategy is
part of some adaptive CC methodologies and contributed signicantly to improve
the eectiveness of CC optimization. As outlined above, RG [7] consists of a
randomized periodical re-allocation of search directions to the subcomponents
during the optimization.
Compared with the simple linear decomposition described in section 2, it
has been proved that RG increases the probability of having two interacting
variables in the same sub-population at least for some iteration of the search
algorithm [7,11]. Clearly, when a subcomponent directly operates on all the
interdependent variables, it has a better chance to adjust their values towards
the optimal direction.
More in details, in the linear decomposition proposed in [1,4,5] the i-th sub-
population operates on the group of directions Gi dened as the interval:
Gi = [(i 1) dk + 1, . . . , i dk ]
Algorithm 2. CCRG(f , n)
1 G = {G1 , . . . , Gk } randomGrouping(n, k);
2 pop initPopulation();
3 contextV ector initContextVector(pop);
4 f itnessEvaluations 0;
5 while f itnessEvaluations < M axF E do
6 foreach Gi G do
7 popi extractPopulation(pop, Gi );
8 besti optimizer(f , popi , contextV ector, Gi , maxF ESC);
9 pop storePopulation(popi , Gi );
10 f itnessEvaluations f itnessEvaluations + maxF ESC;
11 contextV ector updateContextVector(besti , Gi );
12 G = {G1 , . . . , Gk } randomGrouping(n, k);
13 //depending on the optimizer further operations may be required after
//random grouping
14 return contextV ector and f (contextV ector);
N i N i
N 1 1
Pr = 1 (2)
i=r
i k v1 k v1
98 G.A. Truno
fprev fcur
ri = (3)
|fprev |
where fprev is the best tness at the end of the previous cycle and fcur is the
best tness achieved at the end of the current cycle, in which the decomposer
dki has been used. At the beginning of each cycle, the performance indexes are
converted into probabilities using a Boltzmann soft max distribution [33]:
eri /c
pi = t (4)
rj /c
j=1 e
where c is a suitable constant. The latter should be set in such a way to associate
a high probability of being selected to the best decomposers (exploitation), still
giving some chances to all the available decomposers (exploration). The above
mechanism, allows to self-adapt the problem decomposition to the particular
objective problems and also to the evolution stages.
It is worth noting that the MLCC adaptive method can be seen in a per-
spective of a reinforcement learning (RL) approach [33], where the increase of
tness is the reinforcement signal and the actions consist in the choice of the
decomposer. However, instead of selecting actions on the basis of its long-term
utility, in the MLCC their immediate reward is used.
According to the RL literature, an alternative to the selection strategy pro-
posed in [6] could be the so-called -greedy approach: at each cycle the decom-
poser with the highest performance index is selected with probability 1 ,
otherwise random decomposer is selected. The small parameter [0, 1] should
be set in such a way to balance the exploitation of what has been already learned
and the exploration of alternative decomposers with the chance to better adapt
the performance indexes to the evolution stages.
100 G.A. Truno
In [6], the MLCC adaptation method dened by Eqs. 3 and 4 was tested, using
a RG strategy, on a suite of benchmark functions. The authors found that in
several cases the self-adaptive strategy outperformed the corresponding methods
based on the static selection of dk and on the random selection of the group sizes
at each cycle.
A simpler approach for dynamically setting the group size dk was adopted in
the cooperative PSO (CCPSO2) proposed in [14]: given a set of decomposers D,
at the end of each cycle a new decomposer is selected uniformly at random when
the tness does not improve; otherwise, the same decomposer is used for the next
cycle. A problem with this approach is that when the slope of the convergence
curve is very small, the method does not intervene and the current value of dK
is maintained.
A similar strategy was adopted in [18] for a CC implementation of the Firey
Algorithm (FA) [34,26,35]. However, besides the group size also the number of
individuals in the sub-populations was adapted accordingly, using a predened
look-up table. This further adaptation raised the problem of how to select the
individuals to be deleted from the sub-populations when it was necessary to
reduce their size. In [18] the choice was simply to eliminate the worst individuals
of each sub-population. Instead, when the size of sub-populations had to be
increased, new individuals were initialized at random.
Recently, in [32] an improvement of the MLCC adaptive approach was pre-
sented, namely the MLSoft algorithm. In particular, the authors proposed to
use a standard RL approach, replacing ri in Eq. 4 with a value function Vi . The
latter, which is an estimate of the long term utility associated to the use of a
decomposer, was dened as the arithmetic mean of all rewards ri received by
the decomposer dki during the optimization process.
In [32], the authors showed empirically that for a given fully-separable ob-
jective function f , there exists a value d of the problem dimension d to which
corresponds the best optimizers performance (i.e. dened as a sort of eciency
in using the available budget of function evaluations). In terms of CC framework,
this means that the subcomponents should be neither too small nor too large.
Thus, the objective of the adaptive MLSoft approach should be to discover
such an optimal value d < d in order to decompose in the optimal way the
original d-dimensional problem. The MLSoft algorithm was tested on eight fully-
separable functions using a rich set D of decomposers and dierent values of the
parameter c in Eq. 4. According to the results, MLSoft outperformed MLCC.
However, the MLSoft algorithm was not able to outperform the corresponding
CC framework with a xed and optimal subcomponent size. The authors argued
that there is then room for improvement in the way in which the value function
Vi is determined. Indeed, in the standard RL the value function is typically
updated on the basis of the experienced state-action pairs using the received
rewards and a learning rate [0, 1]. In addition, older values of the reward
are usually weighted through a discount factor [0, 1]. Also in the MLSoft
approach, this might be more eective than computing the value function as an
arithmetic mean of the rewards.
Adaptation in Cooperative Coevolution 101
As discussed above, having the interacting variables grouped in the same sub-
component is a crucial factor for enabling ecient CC optimizations of non-
separable problems. For this reason, a number of studies have addressed the prob-
lem of automatic and adaptive decomposition into subcomponents. In contrast
to the RG approach, such automatic procedures try to discover the underlying
structure of the problem in order to devise and adapt a suitable decomposition.
The rst attempt in this direction was carried out in [2], where the authors
proposed a technique to identify interacting variables in a CC framework. The
approach was based on the observation that if a candidate solution where two
directions have been changed achieves a better tness than the same solution
where only one of the directions was changed, then this may indicate the pres-
ence of an interdependency. The creations of groups was carried out during the
optimization process exploiting some additional tness evaluations for each indi-
viduals. The technique proved eective, although the approach was tested only
on few functions with dimensionality up to 30.
Following the idea proposed in [2], which basically consists of observing the
changes of the objective function due to a perturbation of variables, more eec-
tive methods have been developed for enhancing the CC approach. In most cases,
the decomposition stage is performed o-line, that is the groups are created be-
fore the optimization starts. Other approaches presented in the literature for
automatically grouping variables in CC are based on learning statistical models
of interdependencies [13] or on the correlation between variables [36]. However,
as noted in [10] two variables might be highly linearly correlated even when they
are completely separable. In other words, correlation coecients are not a proper
measure for separability in the CC optimization context.
A step in the development of an automatic grouping strategy for CC op-
timizations has been the Delta Grouping (DG) approach proposed in [10]. The
DG algorithm is based on the concept of improvement interval of a variable, that
is the interval in which the tness value could be improved while all the other
variables are kept constant [29,10]. It has been observed that in non-separable
functions, when a variable interacts with other variables, its improvement in-
terval tends to be smaller. Therefore, in the DG approach the identication of
interacting variables was based on measuring the amount of change (i.e. the
delta value) in each of the decision variables during the optimization process. In
particular, the DG algorithm sorts the directions according to the magnitude of
their delta values in order to group the variables with smaller delta values in the
same subcomponent.
Clearly, as noted in [10], not always a small improvement interval implies
a variable interdependency. However, when tested on both the CEC2008 [37]
and CEC2010 [38] benchmark functions, the DG method performed better
than other relevant CC methods. However, a drawback of DG is its low perfor-
mance when there is more than one non-separable subcomponent in the objective
function [10].
102 G.A. Truno
x, xi , xj :
f (x1 , . . . xi , . . . , xj , . . . , xd ) < f (x1 , . . . xi , . . . , xj , . . . , xd ) (5)
f (x1 , . . . xi , . . . , xj , . . . , xd ) > f (x1 , . . . xi , . . . , xj , . . . , xd )
The learning stage of CCVIL starts by placing each direction in a separate sub-
component, that is by separately optimizing the variables in sequence. During
this process, CCVIL tests if the currently and the previously optimized dimen-
sions interact by using Eq. 5. The latter can be applied because only two dimen-
sions changed. Before each learning cycle, the order of optimization of variables
is randomly permutated, so that each two dimensions have the same chance to
be processed in sequence. After the convergence of the learning stage in terms
of grouping, CCVIL starts the optimization stage.
In [9], the authors tested the CCVIL approach using the CEC2010 benchmark
functions [38]. According to the results, CCVIL improved the underlying CC
algorithm in most of the benchmark functions. However, a signicant issue to be
solved concerns the distribution of computational eort between learning and
optimization stages of CCVIL.
Another recent approach for adaptive grouping, named Dierential Group-
ing Algorithm (DGA), has been proposed in [17] for additively separable (AS)
functions f : Rd R, which can be expressed as the sum of k independent non-
separable functions. In this case, there exists an ideal problem decomposition
Gid composed of k groups of variables Gi such that if q Gi and r Gj , with
i = j, then q and r are independent directions. However, it is worth noting that
Gid is not necessarily the best decomposition for a CC optimization algorithm,
as can be inferred from the results presented in [32].
The DGA approach was founded on the formal proof that for AS functions,
if the forward dierences along xp :
are not equal with b = c and = 0, then xp and xq are non-separable. The
forward dierence with interval , in a point x and along the direction xp is
dened as:
fxp (x, ) = f (. . . , xp + , . . . ) f (. . . , xp , . . . )
Adaptation in Cooperative Coevolution 103
Given the ability of automatic decomposition described above, in [12] the au-
thors noted that there is often an imbalance between the contribution to the
tness of the separable and non-separable portions of an optimization problem.
In particular, in CC there are situations in which the improvements in some of
the subcomponents are not apparent simply because they are negligible in com-
parison to the tness variation caused by other subcomponents. Thus, according
to [12], in most cases devoting the same amount of computational resources to
all subcomponents (i.e. the value maxFESC in Algorithms 1 and 2) in a round-
robin fashion, can result in a waste of tness evaluations. In order to mitigate
this issue, in [12] the Contribution Based Cooperative Co-evolution (CBCC)
algorithm was proposed, where:
Clearly, the CBCC algorithm must be integrated with an eective grouping strat-
egy, which should be able to decompose the problem into independent groups
as much as possible. The CBCC has proved to be promising when tested on the
benchmark functions which have been proposed for the CEC2010 [38]. How-
ever, the experiments showed that CBCC is too much inuenced by historical
information in the early stages of evolution. For example, it may happen that
104 G.A. Truno
Table 1. Results of CCAOI and CCFA on the adopted test functions. The better
average errors are highlighted in bold when there is a signicant dierence according
to the t-test.
k k
i=1 j=1 |i j |
=
2 k 2 m
where k = |G| is the number of subcomponents. The value of [0, 1] measures
the inequality between the contributions of the subcomponents. In particular,
= 0 when all the i are the same and = 1 when the normalized contributions
are characterized by the maximum inequality. Subsequently, if the average con-
tribution is greater than zero (i.e. no stagnation) an additional global budget of
generations nAddGen is determined as nM axAddGen, where nM axAddGen
is the amount of generations corresponding to the maximum inequality between
subcomponents (line 11). This is justied by the fact that the main objective of
the adaptive procedure is to rapidly reduce the unbalance between the dierent
subcomponents. Thus, the algorithm assigns to a greater unbalance a higher
total amount of computational eort for that cycle. In the next lines 12-13, the
number of generations nAddGen is distributed to the subcomponents according
to their normalized contribution i . Each subcomponent is activated for a num-
ber geni = minGen + nAddGeni of generations. Later, such value geni is used
for normalizing the contribution of the i-th subcomponent to the global tness.
It is worth noting that the method is based on the assumption that a suitable
decomposition algorithm is available. However, a preliminary version was tested
on some functions with the imbalance characteristics taken from the CEC2010
[38] suite, for which we have manually devised an ideal grouping Gid based on the
knowledge of the functions (in the ideal grouping there is no interdependency
between any two subcomponents). Clearly, when using an automatic decompo-
sition procedure (see Section 3.3) an additional number of function evaluations
is required by the optimization.
106 G.A. Truno
A rst test of the proposed CCAOI approach was conducted using the FA
[34,26,35] as optimizer. A CC version of FA was already developed in [18], were
more details on the implementation can be found. The results for functions
f4 f13 are summarized in Table 1. Also, some average convergence plots are
shown in Fig. 3. In the experiments, the CCAOI algorithm was compared with
the corresponding CC approach in which all the subcomponents operate with
the same number of minGen generations. The adopted CCAOI parameters were
minGen = 10 and nAddGen = 50. The results were averaged over 25 inde-
pendent runs. In Table 1, for each function the best result was highlighted in
bold, according to the t-test with signicance 0.05 that was conducted. As can
be seen, the proposed approach led to improved results in 70% of the test func-
tions. In the remaining 30%, the dierences were not statistically signicant. It
is worth noting that in function f4 f8 the imbalance is relevant, while in the re-
maining functions it is less signicant (see [38] for the details). According to the
results, the proposed algorithm seems eective in addressing the issue of imbal-
ance between the tness contributions provided by the dierent subcomponents.
Adaptation in Cooperative Coevolution 107
4 Conclusions
References
11. Omidvar, M.N., Li, X., Yang, Z., Yao, X.: Cooperative co-evolution for large scale
optimization through more frequent random grouping. In: Proceedings of the IEEE
Congress on Evolutionary Computation, pp. 18. IEEE (2010)
12. Omidvar, M.N., Li, X., Yao, X.: Smart use of computational resources based on
contribution for cooperative co-evolutionary algorithms. In: Proceedings of the 13th
Annual Conference on Genetic and Evolutionary Computation, GECCO 2011, pp.
11151122. ACM, New York (2011)
13. Sun, L., Yoshida, S., Cheng, X., Liang, Y.: A cooperative particle swarm optimizer
with statistical variable interdependence learning. Information Sciences 186(1), 20
39 (2012)
14. Li, X., Yao, X.: Cooperatively coevolving particle swarms for large scale optimiza-
tion. IEEE Trans. Evolutionary Computation 16(2), 210224 (2012)
15. Parsopoulos, K.E.: Parallel cooperative micro-particle swarm optimization: A
master-slave model. Applied Soft Computing 12(11), 35523579 (2012)
16. Hasanzadeh, M., Meybodi, M., Ebadzadeh, M.: Adaptive cooperative particle
swarm optimizer. Applied Intelligence 39(2), 397420 (2013)
17. Omidvar, M.N., Li, X., Mei, Y., Yao, X.: Cooperative co-evolution with dieren-
tial grouping for large scale optimization. IEEE Trans. Evolutionary Computa-
tion 18(3), 378393 (2014)
18. Truno, G.A.: Enhancing the rey algorithm through a cooperative coevolution-
ary approach: an empirical study on benchmark optimisation problems. IJBIC 6(2),
108125 (2014)
19. Dorigo, M., Maniezzo, V., Colorni, A.: Ant system: optimization by a colony of
cooperating agents. IEEE Transactions on Systems, Man, and Cybernetics, Part
B 26(1), 2941 (1996)
20. Doerner, K., Hartl, R.F., Reimann, M.: Cooperative ant colonies for optimizing
resource allocation in transportation. In: Boers, E.J.W., Gottlieb, J., Lanzi, P.L.,
Smith, R.E., Cagnoni, S., Hart, E., Raidl, G.R., Tijink, H. (eds.) EvoWorkshop
2001. LNCS, vol. 2037, pp. 7079. Springer, Heidelberg (2001)
21. Eberhart, R., Kennedy, J.: A new optimizer using particle swarm theory. In: Pro-
ceedings of the Sixth International Symposium on Micro Machine and Human
Science, pp. 3943. IEEE (1995)
22. El-Abd, M., Kamel, M.S.: A Taxonomy of Cooperative Particle Swarm Optimizers.
International Journal of Computational Intelligence Research 4 (2008)
23. Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P.: Optimization by simulated annealing.
Science 220, 671680 (1983)
24. Sanchez-Ante, G., Ramos, F., Frausto, J.: Cooperative simulated annealing for
path planning in multi-robot systems. In: Cair o, O., Cant
u, F.J. (eds.) MICAI
2000. LNCS, vol. 1793, pp. 148157. Springer, Heidelberg (2000)
25. Storn, R., Price, K.: Dierential evolution a simple and ecient heuristic for
global optimization over continuous spaces. Journal of Global Optimization 11(4),
341359 (1997)
26. Yang, X.-S.: Firey algorithms for multimodal optimization. In: Watanabe, O.,
Zeugmann, T. (eds.) SAGA 2009. LNCS, vol. 5792, pp. 169178. Springer, Heidel-
berg (2009)
27. Fister, I., Fister Jr., I., Yang, X.S., Brest, J.: A comprehensive review of rey
algorithms. Swarm and Evolutionary Computation (2013)
28. Fogel, L., Owens, A., Walsh, M.: Articial intelligence through simulated evolution.
Wiley, Chichester (1966)
Adaptation in Cooperative Coevolution 109
29. Salomon, R.: Reevaluating genetic algorithm performance under coordinate rota-
tion of benchmark functions - a survey of some theoretical and practical aspects
of genetic algorithms. BioSystems 39, 263278 (1995)
30. Auger, A., Hansen, N., Mauny, N., Ros, R., Schoenauer, M.: Bio-inspired contin-
uous optimization: The coming of age. Invited talk at CEC 2007, Piscataway, NJ,
USA (2007)
31. Blecic, I., Cecchini, A., Truno, G.A.: Fast and accurate optimization of a GPU-
accelerated ca urban model through cooperative coevolutionary particle swarms.
Procedia Computer Science 29C, 16311643 (2014)
32. Omidvar, M.N., Mei, Y., Li, X.: Eective decomposition of large-scale separable
continuous functions for cooperative co-evolutionary algorithms. In: Proceedings
of the IEEE Congress on Evolutionary Computation. IEEE (2014)
33. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press
(1998)
34. Yang, X.S.: Nature-Inspired Metaheuristic Algorithms. Luniver Press (2008)
35. Yang, X.S.: Firey algorithm, stochastic test functions and design optimisation.
International Journal of Bio-Inspired Computation 2(2), 7884 (2010)
36. Ray, T., Yao, X.: A cooperative coevolutionary algorithm with correlation based
adaptive variable partitioning. In: Proceedings of the IEEE Congress on Evolu-
tionary Computation, pp. 983989. IEEE (2009)
37. Tang, K., Yao, X., Suganthan, P., MacNish, C., Chen, Y., Chen, C., Yang, Z.:
Benchmark functions for the CEC 2008 special session and competition on large
scale global optimization (2008)
38. Tang, K., Li, X., Suganthan, P.N., Yang, Z., Weise, T.: Benchmark functions for
the CEC 2010 special session and competition on large-scale global optimization
(2010)
39. Gini, C.: Measurement of Inequality of Incomes. The Economic Journal 31(121),
124126 (1921)
Study of Lagrangian and Evolutionary Parameters
in Krill Herd Algorithm
1 Introduction
With the development of real-life engineering techniques, the related optimization
problems are becoming more and more complex. Traditional methods seem to be
inefficient to solve such complicated problems. In order to cope with this limitation, a
variety of modern nature-inspired intelligent algorithms have been put forward and
applied to solve optimization problems. Some of them include: differential evolution
(DE) [1,2], artificial bee colony (ABC) [3-5], genetic programming (GP) [6], cuckoo
search (CS) [7-12], biogeography-based optimization (BBO) [13-16], animal migration
optimization (AMO) [17], grey wolf optimizer (GSO) [18], harmony search (HS)
[19,20], interior search algorithm (ISA) [25], particle swarm optimization (PSO)
[21-24], firefly algorithm (FA) [26-28], charged system search (CSS) [29], and bat
algorithm (BA) [30-33]. It has been proven that these methods are superior to the
traditional optimization techniques for solving several challenging problems such as
image segmentation [34], constrained optimization [35], knapsack problem [36],
*
Corresponding author.
2 KH Algorithm
Krill herd (KH) [46] is a swarm intelligence optimization method for solving
optimization problems. KH is idealized from the behavior of the krill swarms. The krill
position in two-dimensional surface is determined by three actions described as:
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 113
dX i
= N i + Fi + Di (1)
dt
where Ni, Fi and Di are three above responding actions for krill i.
where
old
and Nmax is the maximum induced speed, n is its inertia weight in [0, 1], N i is the
last motion;
local
i and target
i are local effect and target direction effect which are
provided by the neighbors and the best krill, respectively. As per the basic KH method,
we set Nmax to 0.01 (ms-1) in our study.
In KH, the effect of the neighbors is determined as:
NN
local
i
= K ij X ij (4)
j =1
X j Xi
X ij = (5)
|| X j X i || +
K Kj
K ij = i
(6)
K worst
K best
worst best
where K and K are the fitness for best and the worst krill; Ki is the fitness of
the i-th krill; Kj is the fitness of j-th (j=1, 2, , NN, NN is the number of the neighbors)
neighbor; X represents the related positions.
114 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
For choosing the neighbor, different strategies can be used. In KH method, a sensing
distance (ds) is defined around an individual.
The known target for each krill is its lowest fitness, and it can be defined by using
Eq. (8).
where, Cbest is the effective coefficient. Herein, the value of Cbest is defined as:
I
C best = i * rand * 1 + (8)
I max
where rand is a random values, i is a number between 0 and 5, I is the actual iteration
number and Imax is the maximum number of iterations.
It is influenced by the two main factors. One is the food location and the other one is the
previous experience about the food location. For the i-th krill individual, it can be
expressed as:
Fi = V f i + f Fi old (9)
where
and Vf is the foraging speed, f is its inertia weight, Fi old is the last foraging motion,
i food is the food attractive and ibest is the effect of the best i-th krill. In our study,
we set Vf to 0.02 [46].
Therefore, the food attraction for the i-th krill can be determined as:
I
C food = j * rand * 1 (12)
I max
where rand is a random value, j is a number between 0 and 5.
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 115
Di = Dmax (13)
where Dmax is the maximum diffusion speed, and is the random directional vector and
its arrays are random values in [-1, 1]. In order to accelerate the search, another term is
added to Eq. (13) that is similar to a geometrical annealing schedule:
I
Di = k * 1 (14)
I max
Therefore, Dmax can be given as
I
Dmax = k * 1 (15)
I max
where k is a number between 0 and 0.02.
In general, three above motions are able to make all the krill individuals move towards
the best position. The krill position during the t to t+t is given as:
dX i
X i ( t + t ) = X i (t ) + t (16)
dt
More details about the three main motions and KH algorithm can be found in [46].
2.5.1 Crossover
In this study, a crossover probability, Cr, determines the crossover operator that is
implemented in binomial or exponential way. By generating a random number, the
m-th component of Xi, xi,m, is defined as:
116 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
xr ,m rand i ,m < Cr
xi ,m = (17)
xi ,m else
2.5.2 Mutation
The mutation is also widely used in EAs such as ES (evolutionary strategy) and DE,
which is determined by a mutation probability (Mu). The mutation operator used in KH
method is expressed as:
xgbes ,m + ( x p ,m xq ,m ) randi , m < M u
xi ,m = (18)
xi ,m else
where p, q[1, 2, ..., i-1, i+1, ..., K} and is a number.
3 Parametric Study
The main goal of this study to find optimal values for five main parameters of the KH
algorithm (i.e., Cbest, Cfood, Dmax, Cr and Mu). In this section, a setting of these
parameters are analyzed and studied in detail by using various experiments conducted
on benchmark functions (see Table 1). Except for special notifications, all the
implementations are conducted under the same conditions as presented in [28]. More
detailed descriptions of all the benchmarks can be found in [62,13,63]. It is notable that
the dimension of functions is twenty in the current work.
n n
xi
1 2 1
0.2 cos( 2xi )
F01 Ackley n i=1 n i =1
f ( x) = 20 + e 20 e e
n
F02 Alpine f ( x ) = | xi sin( xi ) + 0.1xi |
i =1
f ( x ) = ( xi2 ) + ( xi2+1 )
n -1
xi2+1 +1 xi2 +1
F03 Brown
i =1
n
f ( x ) = ( x1 1) + i ( 2 xi2 xi 1 )
2 2
F04 Dixon & Price
i=2
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 117
Table 1. (continued)
n n
f ( x) = ( Ai Bi ) 2 , Ai = ( aij sin j + bij cos j )
i =1 j =1
F05 Fletcher-Powell n
Bi = ( aij sin x j + bij cos x j )
j =1
n
x
2
n
x
F06 Griewank f ( x) = cos i + 1
i
i =1 4000 i =1 i
n
F07 Holzman 2 function f ( x ) = ixi4
i =1
n 1
f ( x ) = sin 2 ( y1 ) + ( yi 1) (1 + 10 sin 2 ( yi + 1) )
2
i =1
F08 Levy 8
+ ( yn 1) (1 + 10 sin 2 (2 yn ) )
2
y =1+(x 1)/4
i i
n 1
f ( x) = 10sin2 ( y1 ) + ( yi 1) 1 +10sin2 ( yi +1 )
2
30 i =1
F09 Penalty #1
}
n
+ ( yn 1) + u ( xi ,10,100,4), yi = 1 + 0.25 ( xi +1)
2
i =1
n 1
f ( x ) = 0.1 sin 2 ( 3 x1 ) + ( xi 1) 1 + sin 2 ( 3 xi +1 )
2
i =1
F10 Penalty #2
}
n
+ ( xn 1) 1 + sin 2 ( 2 xn ) + u ( xi ,5,100, 4 )
2
i =1
2
n n xi k
F11 Perm #1 f ( x ) = ( i + 0.5 ) 1
k
k =1 i =1
i
n/4
f ( x) = ( x4i3 + 10x4i 2 )2 + 5( x4i1 -x4i )2
F12 Powel i =1
n
F13 Quartic with noise f ( x) = (i xi4 + U (0,1))
i =1
n
F14 Rastrigin f ( x ) = 10 n + ( xi2 10 cos(2 xi ))
i =1
118 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
Table 1. (continued)
n 1
f ( x ) = 100 ( xi +1 xi2 ) + ( xi 1)
2 2
F15 Rosenbrock
i =1
( )
D
f ( x ) = 418.9829 D xi sin xi
1/ 2
F16 Schwefel 2.26
i =1
n i
2
n
F20 Sphere f ( x) = xi2
i =1
n
F21 Step f ( x ) = 6 n + x i
i =1
n
F22 Sum function f ( x ) = ixi2
i =1
n
n
2
n
4
The performance of KH with different Cbest, Cfood, and Dmax is tested on twenty-four
optimization problems by studying the values of i, j, and k in Eqs. (8), (12) and (15),
respectively (see Fig. 1). Gandomi and Alavi [46] have proposed four different kinds of
KHs. Here, the KH I (without genetic operators) is selected to do the following
experiments.
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 119
Start
Initialization
Fitness evaluation
Three motions
Motion induced by other individuals
with different i as Eq. (8).
Is termination N
condition met?
Y
Output the best solution
End
As mentioned before, the ranges of i, j, and k are set to [0, 5], [0, 5], and [0, 0.02]
with the internal of 0.025, 0.025, and 0.002, respectively. The settings for the other
parameters used in the KH method are as given in [46].
It is well-known that the efficiency of the metaheuristic methods for finding the best
solutions depends on certain stochastic distribution. Thus, herein, 100 trials are carried
out for the KH method with certain parameter combination on each test problem in
order to get the best Cbest, Cfood, and Dmax. The obtained function values are presented in
Tables 2 and 3. In these tables, f, i, j, and k in first row represents the average function
fitness, and the values of i, j, and k.
It can be seen in Table 2 that, on average, KH method with i=3.9375, j=4.25,
k=0.0151 has the best performance among various KHs, which makes the KH
120 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
implement in the best way. Similarly, it can be observed from Table 3 that KH provides
the best results for the best function values when the i=3.8750, j=4.25, k=0.0123 has the
best performance compared to other KHs.
Referring to Tables 2-3, it can be concluded that the best performance of KH can be
achieved by setting the values of i, j, and k to 4.00, 4.25 and 0.014, respectively.
Consequently, when studying the parameters Cr and Mu, i=4.00, j=4.25, and k=0.014
are respectively adopted in Cbest, Cfood and Dmax.
f i j k
F01 5.17 2.75 4.75 0.020
F02 3.63 4.25 5.00 0.014
F03 38.00 0.00 0.00 0.000
F04 66.31 4.50 5.00 0.020
F05 3.6E5 3.25 3.50 0.018
F06 2.74 3.75 5.00 0.020
F07 15.30 5.00 4.25 0.012
F08 1.63 4.75 4.50 0.012
F09 4.62 4.75 5.00 0.008
F10 24.69 4.00 4.50 0.020
F11 7.4E29 5.00 4.75 0.018
F12 27.04 4.50 5.00 0.020
F13 0.01 5.00 4.50 0.020
F14 80.62 3.00 5.00 0.018
F15 32.20 5.00 3.75 0.020
F16 3.3E3 4.75 3.00 0.016
F17 396.37 5.00 4.25 0.020
F18 57.71 1.50 4.00 0.008
F19 7.11 3.00 3.50 0.008
F20 0.81 4.00 4.25 0.016
F21 298.00 4.75 5.00 0.014
F22 22.24 5.00 4.00 0.020
F23 135.46 2.00 3.25 0.012
F24 406.85 5.00 4.50 0.008
Average -- 3.9375 4.25 0.0151
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 121
f i j k
F01 3.65 2.75 4.75 0.020
F02 1.01 3.50 4.75 0.010
F03 11.19 4.75 4.75 0.014
F04 16.88 4.25 4.50 0.014
F05 1.1E5 4.75 2.25 0.016
F06 1.84 4.00 5.00 0.018
F07 0.99 4.25 4.75 0.010
F08 0.73 3.25 3.25 0.014
F09 1.18 4.00 4.75 0.000
F10 7.37 5.00 4.50 0.010
F11 5.0E22 0.00 4.50 0.008
F12 8.39 4.00 4.50 0.020
F13 3.55E-4 3.50 4.50 0.016
F14 41.25 5.00 4.25 0.014
F15 24.47 4.50 4.00 0.012
F16 2.4E3 5.00 4.00 0.014
F17 108.30 4.50 4.50 0.016
F18 23.62 2.25 4.75 0.014
F19 4.67 5.00 4.00 0.012
F20 0.33 4.75 4.25 0.014
F21 95.00 4.75 5.00 0.014
F22 8.59 4.00 4.25 0.016
F23 48.30 4.00 2.25 0.006
F24 277.27 1.25 4.00 0.000
Average -- 3.8750 4.25 0.0123
Start
Initialization
Fitness evaluation
Is termination N
condition met?
End
The ranges of Cr and Mu are set to [0, 0.5] and [0, 0.5] with the internal of 0.025 and
0.025, respectively. Similarly, 100 trials are carried out for KH method with certain
parameter combination on each test problem in order to get the best Cr and Mu. The
obtained function values are recorded in Tables 4 and 5. In these tables, f, Cr and Mu in
first row represents the average function fitness, the values of Cr and Mu.
As it is seen in Table 4, KH performs the best when the average Cr and Mu are
0.02375 and 0.0271, respectively. From Table 5, for the best function values, KH
performs the best when the average Cr and Mu are equal to 0.224 and 0.025,
respectively.
As per the above experimental results, we can make a conclusion that the values of
Cr and Mu are set to 0.225 and 0.025 are the optimal values for the performance
enhancement of KH.
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 123
f Cr Mu
F01 7.83 0.225 0.000
F02 4.38 0.200 0.000
F03 157.73 0.150 0.175
F04 356.31 0.250 0.000
F05 5.1E5 0.100 0.050
F06 7.24 0.275 0.000
F07 58.78 0.150 0.000
F08 3.63 0.375 0.000
F09 8.01 0.030 0.000
F10 4.3E3 0.500 0.000
F11 2.2E46 0.225 0.000
F12 85.76 0.100 0.000
F13 0.01 0.100 0.000
F14 110.92 0.175 0.000
F15 51.01 0.125 0.000
F16 4.3E3 0.275 0.425
F17 602.68 0.425 0.000
F18 35.52 0.225 0.000
F19 11.44 0.050 0.000
F20 1.69 0.450 0.000
F21 517 0.275 0.000
F22 65.61 0.250 0.000
F23 265.79 0.275 0.000
F24 560.69 0.500 0.000
Average -- 0.2375 0.0271
124 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
f Cr Mu
F01 6.38 0.350 0.000
F02 3.71 0.175 0.000
F03 23.96 0.100 0.050
F04 184.92 0.075 0.000
F05 2.6E5 0.375 0.225
F06 5.49 0.125 0.000
F07 16.48 0.150 0.000
F08 1.60 0.025 0.000
F09 3.58 0.275 0.000
F10 23.78 0.300 0.000
F11 6.7E41 0.125 0.075
F12 28.69 0.300 0.000
F13 0.00 0.075 0.000
F14 77.34 0.150 0.000
F15 35.17 0.175 0.000
F16 3.5E3 0.425 0.450
F17 195.04 0.425 0.000
F18 26.72 0.225 0.000
F19 8.49 0.375 0.000
F20 0.96 0.325 0.000
F21 400 0.275 0.000
F22 23.99 0.225 0.000
F23 106.70 0.275 0.000
F24 424.59 0.050 0.000
Average -- 0.224 0.025
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 125
4 Conclusion
The parameters used in the basic KH method are originated from experimental data
gathered from the literature. Using such information might not be always the best
choice. The main goal of this study is to derive the optimal values for the KH basic
parameters for future implementations of the algorithm. To this aim, an extensive
parametric analysis is carried out using an array of high-dimensional benchmark
problems. The performance of KH with different Cbest, Cfood, Dmax and Cr, Mu is studied
on twenty-four optimization problems. The KH algorithm without genetic operators is
selected to study the Cbest, Cfood, Dmax parameters. The parametric analysis of Cr and Mu
is done using the KH algorithm with two genetic operators: crossover and mutation
operator. The major finding is that KH would have the best performance for most
high-dimensional test functions by setting the Lagrangian parameters of i, j and k
respectively to 4.00, 4.25 and 0.014 in Cbest, Cfood and Dmax. The best genetic
parameters, Cr and Mu, are also found as 0.225, and 0.025 respectively. Undoubtedly,
this finding would eliminate the concerns regarding the optimal tuning of the KH
algorithm for its most future applications.
Acknowledgements. This work was supported by Research Fund for the Doctoral
Program of Jiangsu Normal University (No. 13XLR041).
References
1. Storn, R., Price, K.: Differential evolution-a simple and efficient heuristic for global
optimization over continuous spaces. J. Global Optim. 11(4), 341359 (1997),
doi:10.1023/A:1008202821328
2. Gandomi, A.H., Yang, X.-S., Talatahari, S., Deb, S.: Coupled eagle strategy and differential
evolution for unconstrained and constrained global optimization. Comput. Math.
Appl. 63(1), 191200 (2012), doi:10.1016/j.camwa.2011.11.010
3. Karaboga, D., Basturk, B.: A powerful and efficient algorithm for numerical function
optimization: artificial bee colony (ABC) algorithm. J. Global Optim. 39(3), 459471
(2007), doi:10.1007/s10898-007-9149-x
4. Li, X., Yin, M.: Self-adaptive constrained artificial bee colony for constrained numerical
optimization. Neural Comput. Appl. 24(3-4), 723734 (2012),
doi:10.1007/s00521-012-1285-7
5. Fister, I., Fister Jr., I., Zumer, J.B.: Memetic artificial bee colony algorithm for large-scale
global optimization. In: IEEE Congress on Evolutionary Computation (CEC 2012),
Brisbane, Australia, June 10-15, pp. 18. IEEE (2012), doi:10.1109/CEC.2012.6252938
6. Gandomi, A.H., Alavi, A.H.: Multi-stage genetic programming: A new strategy to nonlinear
system modeling. Inf. Sci. 181(23), 52275239 (2011), doi:10.1016/j.ins.2011.07.026
7. Gandomi, A.H., Yang, X.-S., Alavi, A.H.: Cuckoo search algorithm: a metaheuristic
approach to solve structural optimization problems. Eng. Comput. 29(1), 1735 (2013),
doi:10.1007/s00366-011-0241-y
8. Yang, X.S., Deb, S.: Cuckoo search via Lvy flights. In: Proceeding of World Congress on
Nature & Biologically Inspired Computing (NaBIC 2009), Coimbatore, India, pp. 210214.
IEEE Publications, USA (2009)
126 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
9. Gandomi, A.H., Talatahari, S., Yang, X.-S., Deb, S.: Design optimization of truss structures
using cuckoo search algorithm. Struct. Des. Tall Spec. 22(17), 13301349 (2013),
doi:10.1002/tal.1033
10. Li, X., Wang, J., Yin, M.: Enhancing the performance of cuckoo search algorithm using
orthogonal learning method. Neural Comput. Appl. 24(6), 12331247 (2013),
doi:10.1007/s00521-013-1354-6
11. Fister Jr, I., Yang, X.-S., Fister, D., Fister, I.: Cuckoo Search: A Brief Literature Review. In:
Yang, X.-S. (ed.) Cuckoo Search and Firefly Algorithm. SCI, vol. 516, pp. 4962. Springer,
Heidelberg (2014)
12. Fister Jr, I., Fister, D., Fister, I.: A comprehensive review of cuckoo search: variants and
hybrids. Int. J. Math. Model. Numer. Optim. 4(4), 387409 (2013)
13. Simon, D.: Biogeography-based optimization. IEEE Trans. Evolut. Comput. 12(6),
702713 (2008), doi:10.1109/TEVC.2008.919004
14. Li, X., Wang, J., Zhou, J., Yin, M.: A perturb biogeography based optimization with
mutation for global numerical optimization. Appl. Math. Comput. 218(2), 598609 (2011),
doi:10.1016/j.amc.2011.05.110
15. Li, X., Yin, M.: Multi-operator based biogeography based optimization with mutation for
global numerical optimization. Comput. Math. Appl. 64(9), 28332844 (2012),
doi:10.1016/j.camwa.2012.04.015
16. Saremi, S., Mirjalili, S., Lewis, A.: Biogeography-based optimisation with chaos. Neural
Comput. Appl. (2014), doi:10.1007/s00521-014-1597-x
17. Li, X., Zhang, J., Yin, M.: Animal migration optimization: an optimization algorithm
inspired by animal migration behavior. Neural Comput. Appl. (2013),
doi:10.1007/s00521-013-1433-8
18. Mirjalili, S., Mirjalili, S.M., Lewis, A.: Grey wolf optimizer. Adv. Eng. Softw. 69, 4661
(2014), doi:10.1016/j.advengsoft.2013.12.007
19. Geem, Z.W., Kim, J.H., Loganathan, G.V.: A new heuristic optimization algorithm:
harmony search. Simulation 76(2), 6068 (2001), doi:10.1177/003754970107600201
20. Wang, G., Guo, L., Duan, H., Wang, H., Liu, L., Shao, M.: Hybridizing harmony search
with biogeography based optimization for global numerical optimization. J. Comput. Theor.
Nanos. 10(10), 23182328 (2013), doi:10.1166/jctn.2013.3207
21. Kennedy, J., Eberhart, R.: Particle swarm optimization. Paper presented at the Proceeding of
the IEEE International Conference on Neural Networks, Perth, Australia, November
27-December 1 (1995)
22. Talatahari, S., Kheirollahi, M., Farahmandpour, C., Gandomi, A.H.: A multi-stage particle
swarm for optimum design of truss structures. Neural Comput. Appl. 23(5), 12971309
(2013), doi:10.1007/s00521-012-1072-5
23. Mirjalili, S., Lewis, A.: S-shaped versus V-shaped transfer functions for binary particle
swarm optimization. Swarm Evol. Comput. 9, 114 (2013),
doi:10.1016/j.swevo.2012.09.002
24. Mirjalili, S., Wang, G.-G., Coelho, L.S.: Binary optimization using hybrid particle swarm
optimization and gravitational search algorithm. Neural Comput. Appl. (2014),
doi:10.1007/s00521-014-1629-6
25. Gandomi, A.H.: Interior Search Algorithm (ISA): A Novel Approach for Global
Optimization. ISA Trans. (2014), doi:10.1016/j.isatra.2014.03.018
26. Yang, X.S.: Firefly algorithm, stochastic test functions and design optimisation. Int J of
Bio-Inspired Computation 2(2), 7884 (2010)
27. Fister, I., Fister Jr., I., Yang, X.-S., Brest, J.: A comprehensive review of firefly algorithms.
Swarm Evol. Comput. 13, 3446 (2013), doi:10.1016/j.swevo.2013.06.001
28. Wang, G.-G., Guo, L., Duan, H., Wang, H.: A new improved firefly algorithm for global
numerical optimization. J. Comput. Theor. Nanos. 11(2), 477485 (2014),
doi:10.1166/jctn.2014.3383
Study of Lagrangian and Evolutionary Parameters in Krill Herd Algorithm 127
29. Kaveh, A., Talatahari, S.: A novel heuristic optimization method: charged system search.
Acta Mech. 213(3-4), 267289 (2010), doi:10.1007/s00707-009-0270-4
30. Gandomi, A.H., Yang, X.-S., Alavi, A.H., Talatahari, S.: Bat algorithm for constrained
optimization tasks. Neural Comput. Appl. 22(6), 12391255 (2013),
doi:10.1007/s00521-012-1028-9
31. Yang, X.S., Gandomi, A.H.: Bat algorithm: a novel approach for global engineering
optimization. Eng. Computation 29(5), 464483 (2012), doi:10.1108/02644401211235834
32. Fister Jr., I., Fong, S., Brest, J., Fister, I.: A Novel Hybrid Self-Adaptive Bat Algorithm. Sci.
World J. 2014, 112 (2014), doi:10.1155/2014/709738
33. Mirjalili, S., Mirjalili, S.M., Yang, X.-S.: Binary bat algorithm. Neural Comput. Appl.
(2013), doi:10.1007/s00521-013-1525-5
34. Zhang, Y., Huang, D., Ji, M., Xie, F.: Image segmentation using PSO and PCM with
Mahalanobis distance. Expert Syst. Appl. 38(7), 90369040 (2011),
doi:10.1016/j.eswa.2011.01.041
35. Chen, C.-H., Yang, S.-Y.: Neural fuzzy inference systems with knowledge-based cultural
differential evolution for nonlinear system control. Inf. Sci. (2014),
doi:10.1016/j.ins.2014.02.071
36. Mukherjee, R., Patra, G.R., Kundu, R., Das, S.: Cluster-based differential evolution with
Crowding Archive for niching in dynamic environments. Inf. Sci. (2014),
doi:10.1016/j.ins.2013.11.025
37. Li, X., Yin, M.: Multiobjective binary biogeography based optimization for feature selection
using gene expression data. IEEE Trans. Nanobiosci. 12(4), 343353 (2013),
doi:10.1109/TNB.2013.2294716
38. Fister, I., Mernik, M., Filipi, B.: A hybrid self-adaptive evolutionary algorithm for marker
optimization in the clothing industry. Appl. Soft Compt. 10(2), 409422 (2010),
doi:10.1016/j.asoc.2009.08.001
39. Li, X., Yin, M.: Parameter estimation for chaotic systems by hybrid differential evolution
algorithm and artificial bee colony algorithm. Nonlinear Dynam. (2014),
doi:10.1007/s11071-014-1273-9
40. Li, X., Yin, M.: Application of Differential Evolution Algorithm on Self-Potential Data.
PLoS ONE 7(12), e51199 (2012), doi:10.1371/journal.pone.0051199
41. Mirjalili, S., Mohd Hashim, S.Z., Moradian Sardroudi, H.: Training feedforward neural
networks using hybrid particle swarm optimization and gravitational search algorithm.
Appl. Math. Comput. 218(22), 1112511137 (2012), doi:10.1016/j.amc.2012.04.069
42. Mirjalili, S., Mirjalili, S.M., Lewis, A.: Let a biogeography-based optimizer train your
Multi-Layer Perceptron. Inf. Sci. 269, 188209 (2014), doi:10.1016/j.ins.2014.01.038
43. Li, X., Yin, M.: An opposition-based differential evolution algorithm for permutation flow
shop scheduling based on diversity measure. Adv. Eng. Softw. 55, 1031 (2013),
doi:10.1016/j.advengsoft.2012.09.003
44. Yang, X.S., Gandomi, A.H., Talatahari, S., Alavi, A.H.: Metaheuristics in Water,
Geotechnical and Transport Engineering. Elsevier, Waltham (2013)
45. Gandomi, A.H., Yang, X.S., Talatahari, S., Alavi, A.H.: Metaheuristic Applications in
Structures and Infrastructures. Elsevier, Waltham (2013)
46. Gandomi, A.H., Alavi, A.H.: Krill herd: A new bio-inspired optimization algorithm.
Commun. Nonlinear Sci. Numer. Simulat. 17(12), 48314845 (2012),
doi:10.1016/j.cnsns.2012.05.010
47. Wang, G.-G., Guo, L., Gandomi, A.H., Hao, G.-S., Wang, H.: Chaotic krill herd algorithm.
Inf. Sci. 274, 1734 (2014), doi:10.1016/j.ins.2014.02.123
48. Wang, G.-G., Gandomi, A.H., Alavi, A.H.: A chaotic particle-swarm krill herd algorithm for
global numerical optimization. Kybernetes 42(6), 962978 (2013),
doi:10.1108/K-11-2012-0108
128 G.-G. Wang, A.H. Gandomi, and A.H. Alavi
49. Saremi, S., Mirjalili, S.M., Mirjalili, S.: Chaotic Krill Herd Optimization Algorithm.
Procedia Technology 12, 180185 (2014), doi:10.1016/j.protcy.2013.12.473
50. Wang, G.-G., Gandomi, A.H., Alavi, A.H.: Stud krill herd algorithm. Neurocomputing 128,
363370 (2014), doi:10.1016/j.neucom.2013.08.031
51. Wang, G., Guo, L., Gandomi, A.H., Cao, L., Alavi, A.H., Duan, H., Li, J.: Lvy-flight krill
herd algorithm. Math. Probl. Eng. 2013, 114 (2013), doi:10.1155/2013/682073
52. Guo, L., Wang, G.-G., Gandomi, A.H., Alavi, A.H., Duan, H.: A new improved krill herd
algorithm for global numerical optimization. Neurocomputing 138, 392402 (2014),
doi:10.1016/j.neucom.2014.01.023
53. Wang, G.-G., Gandomi, A.H., Alavi, A.H.: An effective krill herd algorithm with migration
operator in biogeography-based optimization. Appl. Math. Model. 38(9-10), 24542462
(2014), doi:10.1016/j.apm.2013.10.052
54. Wang, G., Guo, L., Wang, H., Duan, H., Liu, L., Li, J.: Incorporating mutation scheme into
krill herd algorithm for global numerical optimization. Neural Comput. Appl. 24(3-4),
853871 (2014), doi:10.1007/s00521-012-1304-8
55. Wang, G.-G., Gandomi, A.H., Alavi, A.H., Hao, G.-S.: Hybrid krill herd algorithm with
differential evolution for global numerical optimization. Neural Comput. Appl. 25(2),
297308 (2014), doi:10.1007/s00521-013-1485-9
56. Li, J., Tang, Y., Hua, C., Guan, X.: An improved krill herd algorithm: Krill herd with linear
decreasing step. Appl. Math. Comput. 234, 356367 (2014),
doi:10.1016/j.amc.2014.01.146
57. Wang, G.-G., Guo, L., Gandomi, A.H., Alavi, A.H., Duan, H.: Simulated annealing-based
krill herd algorithm for global optimization. Abstr. Appl. Anal. 2013, 111 (2013),
doi:10.1155/2013/213853
58. Wang, G.-G., Gandomi, A.H., Yang, X.-S., Alavi, A.H.: A new hybrid method based on
krill herd and cuckoo search for global optimization tasks. Int. J. of Bio-Inspired
Computation (2013)
59. Gandomi, A.H., Talatahari, S., Tadbiri, F., Alavi, A.H.: Krill herd algorithm for optimum
design of truss structures. Int. J. of Bio-Inspired Computation 5(5), 281288 (2013),
doi:10.1504/IJBIC.2013.057191
60. Gandomi, A.H., Alavi, A.H.: An introduction of krill herd algorithm for engineering
optimization. J. Civil Eng. Manag. (2013)
61. Gandomi, A.H., Alavi, A.H., Talatahari, S.: Structural Optimization using Krill Herd
Algorithm. In: Swarm Intelligence and Bio-Inspired Computation: Theory and
Applications, pp. 335349. Elsevier (2013)
62. Yao, X., Liu, Y., Lin, G.: Evolutionary programming made faster. IEEE Trans. Evolut.
Comput. 3(2), 82102 (1999)
63. Yang, X.-S., Cui, Z., Xiao, R., Gandomi, A.H., Karamanoglu, M.: Swarm Intelligence and
Bio-Inspired Computation. Elsevier, Waltham (2013)
Solutions of Non-smooth Economic Dispatch Problems
by Swarm Intelligence
1 Introduction
The ED problem is one of the main issues in power system operation and control. The
goal of an ED problem is to schedule the online generating units so as to satisfy the
load demand at minimum operating cost, while satisfying all the equality and inequa-
lity constraints of the units [1]. However, careful and intelligent scheduling of the
units can both decrease the operating cost significantly and assure higher reliability,
improving security with less environmental impact [2]. Therefore, the optimization of
ED problems is a challenging task and new algorithms such as meta-heuristics may be
promising for solving modern power system operations and control.
*
Corresponding author.
Traditionally, mathematical modeling of fuel costs for generating units often uses
approximate models in terms of a single quadratic cost function [3-4]. This type of prob-
lem can often be solved using several mathematical programming techniques e.g. the
lambda-iteration method, the base point and participation factors method, the interior
point method, dynamic programming, and the gradient algorithms [3, 5-8]. However,
none of these techniques can find an optimal solution satisfactorily, as they are local
search methods which can normally be trapped at a local optimum. Thus, to use the
right algorithm is very important. In addition, the effective implementation is equally
important, and even though variables are continuous. Basic ED problems consider the
power balance constraints apart from the generating capacity limits. However, a practic-
al ED model must include the prohibited operating zones, ramp rate limits, valve point
loading effects and multi-fuel options [9] so as to provide a complete formulation for the
ED problem. The resulting ED is a non-convex optimization problem, which is very
challenging to solve and cannot be resolved by the traditional approaches.
To overcome these deficiencies, evolutionary algorithms and meta-heurstics have
been utilized to solve the ED problems, and these techniques include Genetic Algo-
rithm (GA) [10], real-coded genetic algorithm (RCGA) [11], Tabu Search (TS)
[12-13], Hopfield neural network [14], Differential Evolution [15], different types of
Evolutionary Programming (EP) [16-17], biogeography-based optimization (BBO)
[18], Evolutionary Strategy (ES) [19], Particle Swarm Optimization (PSO) [2, 20-26],
an improved coordinated aggregation-based particle swarm optimization (ICA-PSO)
[27-28], Bacterial Foraging (BF) [21], and harmony search (HS) [29], Firefly Algo-
rithm (FA) [30], multiple tabu search (MTS) [31], taguchi self-adaptive real-coded
genetic algorithm (TSARGA) [32].
Although many optimization methods have been developed for the ED problems,
the complexity of the task reveals the necessity for improvement in efficient tech-
niques to accurately find the global optimum solution. Recently, a new meta-heuristic
search algorithm, called Bat Algorithm (BA), has been developed [33-35]. BA is a
new search method based on the echolocation behavior of microbats. The capability
of echolocation of microbats is fascinating, as these bats can locate their prey and
discriminate different types of insects even in complete darkness. Preliminary studies
suggest that the BA can have superior performance over genetic algorithms and par-
ticle swarm optimization [36], and it can solve real world and engineering optimiza-
tion problems [30, 37-39]. In this chapter, we will study BA further in detail and solve
ED problems. Recently two papers [40-41] have been done in this field but more de-
tailed illustration concerning theoretical and implementation feature of the proposed
algorithm is provided in the following sections. To prove the efficiency and applica-
bility of the proposed approach, several types of ED problems are studied and the
results are compared with those available in the literature.
The chapter is organized as follows: Section 2 illustrates the ED problems and its
formulation incorporating valve-loading effect multiple fuel option, prohibited operat-
ing zone (POZ) constraints and ramp rate limits. Moreover, the proposed technique
for constraints handling is described in this section. In Section 3, the Bat Algorithm is
described. In Section 4, the simulation results are presented that show the potential of
the proposed method. Finally, Section 5 concludes the paper with some discussions.
Solutions of Non-smooth Economic Dispatch Problems by Swarm Intelligence 131
2 Problem Formulations
where Fi ( Pi ) is the total generation cost for the generator unit i , which is defined
by the following equation:
Fi ( Pi ) = ai Pi 2 + bi Pi + ci (2)
P = P
i =1
i D + PL (3)
where PD is the total active power demand, PL is the network loss, Pi min is the
minimum operating limit of i-th unit, Pi max is the maximum operating limit of i-th
unit. In an ED problem PL can be approximated by a function of the unit power
outputs and the transmission loss matrix coefficients called matrix loss formula [24,
42-43]. The other important constraints are as follows.
range of all the online units for tuning the generator operation between two operating
periods [44-45]. The generation may increase or decrease with corresponding upper
and lower ramp rate limits. Therefore, units are restricted because of these ramp rate
limits as mentioned below:
If power generation increases, we have
Pi Pi 0 UR i (5)
Pi 0 Pi DRi (6)
where Pi 0 is the previous power generation of unit i . URi and DRi are the up-
ramp and down-ramp limits of the i-th generator, respectively. The inclusion of ramp
rate limits changes the generator operation constraints (5) as follows:
,
, , 2,3, , (8)
,
ai ,1 Pi 2 + bi ,1 Pi + ci ,1 , if Pi ,min Pi Pi ,1
ai , 2 Pi + bi , 2 Pi + ci , 2 , Pi ,1 Pi Pi , 2
2
if
. .
Fi ( Pi ) = (10)
. .
. .
ai ,n Pi 2 + bi ,n Pi + ci ,n , if Pi ,n 1 Pi Pi ,max
where a i , j , bi , j , and c i , j are the cost coefficients of generator i with fuel type j,
respectively; Pi ,min and Pi ,max are the minimum and maximum power generation of
unit i .
2.4 Non-smooth Cost Functions with Valve-Point Effects and Multiple Fuel
Options
To acquire a precise and practical ED solution, the realistic operation of the ED prob-
lem should deem both valve-point effects and multiple fuel options. The cost model in
this paper integrates the valve-point loadings and the fuel changes in one frame. So,
the cost function, combining (5) and (6), can be realistically written as
, , , , sin , , (11)
, , ,
3 Bat Algorithm
The bat-inspired meta-heuristic algorithm, namely the bat algorithm (BA), was recently
introduced by Xin-She Yang [33, 36], based on the echolocation of microbats. This algo-
rithm has been applied to many applications [30, 46-47]. In the real world, echolocation
usually use short impulses of a few thousandths of a second (up to about 8 to 10 ms) with
a varying frequency in the region of 25 kHz to 150 kHz, corresponding to the wave-
lengths of 2 mm to 14 mm in the air. Microbats utilize a type of sonar called echolocation
to recognize prey, escape obstacles, and locate their roosting crevices in the dark, and the
bat algorithm was inspired by this echolocation behavior. These bats transmit a very loud
134 S.S.S. Hosseini et al.
sound pulse and listen for the echo that bounces back from the surrounding objects. Their
pulses vary in properties and can be correlated with their hunting plans, depending on the
species. Most bats utilize short, frequency-modulated signals to sweep through about an
octave, while others more often utilize constant-frequency signals for echolocation. Their
signal bandwidth changes depending on the species and often increased by utilizing more
harmonics. In the standard bat algorithm, the echolocation characteristics of microbats
can be idealized as the following three rules:
i. All bats utilize echolocation to sense distance, and they know the differ-
ence between food/prey and background barriers in some magical way;
ii. Bats randomly fly with velocity vi at position xi with a fixed frequency frmin,
varying wavelength and loudness A0 to search for prey. They can automati-
cally regulate the wavelength (or frequency) of their emitted pulses and ad-
just the rate of pulse emission r [0,1], depending on the proximity of their
target;
iii. Although the loudness can change in many ways, we assume that the loud-
ness changes from a large (positive) A0 to a minimum constant value Amin
[33, 36].
For simplicity, we do not utilize ray tracing in this technique, though it can produce
an interesting feature for further extension. Generally, ray tracing can be computa-
tionally extensive, but it can be a useful feature for computational geometry and other
applications.
The basic steps of BA can be summarized as the pseudo code shown in Fig. 1.
Bat Algorithm
Objective function F(x), x = (x1, ...,xd)T
Initialize the bat population xi (i = 1,2, ...,n) and vi
Specify pulse frequency fri at xi
Initialize pulse rates r and the loudness A
while (t <Max number of iterations)
Generate new solutions by adjusting frequency,
and updating velocities and locations/solutions [equations (12) to (14)]
if (rand > r)
Select a solution among the best solutions randomly
Generate a local solution around the selected best solution by a local random
walk
end if
Produce a new solution by flying randomly
if (rand < A & F (xi) < F (x*))
Accept the new solutions
end if
Rate the bats and find the current best x*
end while
Postprocess results and visualization
Fig. 1. Pseudo code of the BA
Solutions of Non-smooth Economic Dispatch Problems by Swarm Intelligence 135
For each i-th bat, its position xit and velocity vit in a d-dimensional search space
should be defined, and upgraded during the iterations. The new solutions xti and
velocities vti at time step t can be estimated by
, , , (13)
, , , (14)
where in the range of [0,1] is a random vector drawn from a uniform distribution.
Here x* is the current global best location found yet, which is found after comparing all
the solutions amid all the n bats at the current iteration. As the product is the
velocity in the media, which is essentially fixed we can utilize either (or ) to
adjust the velocity change during fixing the other factor (or ), depending on the
type of the problem of interest. For the applications here, we will utilize frmin = 0 and
frmax = 2, depending on the domain size of the problem of interest. Initially, each bat is
randomly assigned a frequency that is drawn uniformly from [frmin, frmax] [30].
For the local search part, once a solution is selected among the current best solu-
tions, a new solution for each bat is produced locally utilizing a local random walk:
xnew = xold+ At (15)
t
where the random number is drawn from [1, 1], while A =<Ait>is the average
loudness of all the bats at this time step. In fact, this is the main updating equation of
simulated annealing. For this reason, simulated annealing could be thought as a very
special case of the BA.
Additionally, the loudness and the rate of pulse emission have to be updated
accordingly as iterations proceed. As the loudness typically decreases once a bat has
found its prey, while the rate of pulse emission grows, the loudness can be chosen as
any value of convenience. For simplicity, we can also utilize A0 = 1 and Amin = 0,
assuming Amin = 0 means that a bat has just found the prey and temporarily stop
transmitting any sound. Now we have
, 1 exp (16)
where and are constants. In fact, is analogous to the cooling factor of a cooling
schedule in the simulated annealing. For any 0 <, < 1, we have
0, as (17)
In the simplest case, we can utilize = , and in the standard BA, we can utilize
= = 0.9 to 0.975 in most cases, though have utilized = =0.9 in our simulations.
Initially, each bat should have several values of loudness and pulse emission rate, and
this can be obtained by randomization. For instance, the initial loudness can
usually be [1,2], while the initial emission rate can be around zero, or any value
0,1 .
136 S.S.S. Hosseini et al.
then the amount of generation is cut to the nearest boundary of the jth POZ as below:
j = ( Pi , j + Pi , j ) / 2
Pi ,ave LB UB
(19)
4 Numerical Results
Because of the random nature of the BA (and in fact all meta-heuristic methods), their
performance cannot be assessed by the result of a single run. Many trials with indepen-
dent population initializations should be done to obtain a useful conclusion of the per-
formance of the method. Thus, the results should be evaluated utilizing statistic measures
such as the means and standard deviations. The best, worst and mean obtained in 40 trials
are utilized to compare the performances of different EAs. To show the effectiveness of
the proposed BA, the test results are also compared with the results already reported by
recently published results utilizing most recent algorithms for solving the ED problems.
Solutions of Non-smooth Economic Dispatch Problems by Swarm Intelligence 137
The parameters of BA in our simulations are: n=20, = =0.95, and the number of total
iterations t=2,000. These are adapted to give the best results for the ED problems.
Table 1. The best, average and worst results of different ED solution methods for the 3 unit test
system
Table 2. Output power of generators in the best result of the proposed FA for the 3 unit test
system
Table 4. Output power of generators in the best result of the proposed BA for the 13 unit test
system
Table 5. The best, average and worst results of different ED solution methods for the 40 unit test system
higher nonlinearity, we have tested ED systems with variable sizes, we found that the
proposed method is potentially the best approach among all the methods we have
tested. The optimum dispatch of each generator is also recorded in order to see it in
permissible limits and shown in Table 6.
Table 6. Output power of generators in the best result of the proposed BA for the 40 unit test
system
Table 7. The best, average and worst result of different ED solution algorithms for the 15 unit
test system including POZ constraints, ramp rate limits and transmission losses
Table 8. Output power of generators and transmission losses in the best result of the proposed
BA for the 15 unit test system
the system is 2,630 MW. The best cost reported is 32,704.9$/h [61]. The problem
has a number of local optimum points, as it is more likely that most algorithms will be
trapped in a local optimum region. The prohibited operating zones embedded in the 4
units are units 2, 5, 6, and 12. This problem proves particularly challenging because
these zones result in a non-convex decision space where 192 convex subspaces can be
constituted for the dispatch problem. The remaining units have simple operational
zones. This challenging problem not only needs the proper implementation of the
constraints, but also uses an efficient search in different sub regions without wasting
too much time on the prohibited regions. Thus, a fine balance between solution quali-
ty and computational effort is needed.
The comparison of the best, average and worst solutions found by the BA and the
best solutions obtained by other methods in the most recent literature is shown in
Table 7. Again, the BA offers an improved generation cost over the other algorithms,
clearly showing the proposed approach of locating better solutions is superior to oth-
ers. Detailed results of the optimal solutions are shown in Table 8 where that all the
system constraints are met.
5 Conclusions
References
1. Park, J.B., Lee, K.S., Shin, J.R., Lee, K.Y.: A particle swarm optimization for economic
dispatch with nonsmooth cost functions. IEEE Trans. Power Syst. 20(1), 3442 (2005)
2. Abido, M.A.: Multiobjective evolutionary algorithms for electric power dispatch problem.
IEEE Trans. Evol. Comput. 10(3), 315329 (2006)
3. Wood, A.J., Wollenberg, B.F.: Power generation, operation and control. John Wiley Sons,
New York (1984)
4. Arrillaga, J., Arnold, C.P.: Computer analysis of power systems. John Wiley Sons, Great
Britain (1990)
5. IEEE Committee Report, Present practices in the economi-coperation of power systems.
IEEE Trans. Power Appar. Syst. PAS-90, 176875 (1971)
6. Chowdhury, B.H., Rahman, S.: A review of recent advances in economic dispatch. IEEE
Trans. Power Syst. 5(4), 12481259 (1990)
7. Lianf, Z.X., Glover, J.D.: A zoom feature for a dynamic programming solution to econom-
ic dispatch including transmission losses. IEEE Trans. Power Syst. 7(2), 544550 (1992)
8. Granville, S.: Optimal reactive dispatch through interior point methods. IEEE Trans. Pow-
er Syst. 9(1), 136146 (1994)
9. Lin, C.E., Viviani, G.L.: Hierarchical economic dispatch for piecewise quadratic cost func-
tions. IEEE Trans. Power Appar. Syst. PAS 103(6), 11701175 (1984)
10. Chiang, C.-L.: Improved genetic algorithm for power economic dispatch of units with
valve-point effects and multiple fuels. IEEE Trans. Power Syst. 20(4), 16901699 (2005)
11. Kumar, S., Naresh, R.: Non-convex economic load dispatch using an efficient real-coded
genetic algorithm. Appl. Soft. Comput. 9(1), 321329 (2009)
12. Pothiya, S., Ngamroo, I., Kongprawechnon, W.: Application of multiple tabu search algo-
rithm to solve dynamic economic dispatch considering generator constraints. Energy Con-
vers Manage 49(4), 506516 (2008)
13. Khamsawang, S., Jiriwibhakorn, S.: DSPSOTSA for economic dispatch problem with
nonsmooth and noncontinuous cost functions. Energ. Convers Manage 51, 365375 (2010)
14. Lee, K.Y., Sode-Yome, A., Park, J.H.: Adaptive Hopfield neural network for economic
load dispatch. IEEE Trans. Power Syst. 13(2), 519526 (1998)
15. Noman, N., Iba, H.: Differential evolution for economic load dispatch problems. Electr.
Power Syst. Res. 78(8), 13221331 (2008)
16. Park, Y.M., Won, J.R., Park, J.B.: A new approach to economic load dispatch based on
improved evolutionary programming. Eng. Intell. Syst. Electr. Eng. Commun. 6(2), 103
110 (1998)
17. Sinha, N., Chakrabati, R., Chattopadhyay, P.K.: Evolutionary programming techniques for
economic load dispatch. IEEE Trans. Evol. Comput. 7(1), 8394 (2003)
18. Bhattacharya, A., Chattopadhyay, P.K.: Solving complex economic load dispatch prob-
lems using biogeography-based optimization. Expert Syst. Appl. 37, 36053615 (2010)
19. Pereira-Neto, A., Unsihuay, C., Saavedra, O.R.: Efficient evolutionary strategy optimiza-
tion procedure to solve the nonconvex economic dispatch problem with generator con-
straints. IEE. Proc. Gen. Transm. Distrib. 152(5), 653660 (2005)
20. Gaing, Z.-L.: Particle swarm optimization to solving the economic dispatch considering
the generator constraints. IEEE Trans. Power Syst. 18(3), 11871195 (2003)
21. Panigrahi, B.K., Pandi, V.R.: Bacterial foraging optimisation: NelderMead hybrid algo-
rithm for economic load dispatch. IEE. Proc. Gen. Transm. Distrib. 2, 556565 (2008)
22. Jiejin, C., Xiaoqian, M., Lixiang, L., Haipeng, P.: Chaotic particle swarm optimization for
economic dispatch considering the generator constraints. Energy Convers Manage 48,
645653 (2007)
23. Kuo, C.-C.: A novel coding scheme for practical economic dispatch by modified particle
swarm approach. IEEE Trans. Power Syst. 23, 18251835 (2008)
Solutions of Non-smooth Economic Dispatch Problems by Swarm Intelligence 145
24. Selvakumar, A.I., Thanushkodi, K.: A new particle swarm optimization solution to non-
convex economic dispatch problems. IEEE Trans. Power Syst. 22(1), 4251 (2007)
25. Neyestani, M., Farsangi, M.M., Nezamabadi-pour, H.: Amodified particle swarm optimi-
zation for economic dispatch with non-smooth cost functions. Eng. Appl. Artif. In-
tel. 23(7), 11211126 (2010)
26. Ling, S.H., Iu, H.H.C., Chan, K.Y., Lam, H.K., Yeung, B.C.W., Leung, F.H.: Hybrid par-
ticle swarm optimization with wavelet mutation and its industrial applications. IEEE
Trans. Syst. Man. Cyb. 38(3), 743763 (2008)
27. Vlachogiannis, J.G., Lee, K.Y.: Economic Load DispatchA Comparative Study on Heu-
ristic Optimization Techniques With an Improved Coordinated Aggregation-Based PSO.
IEEE Trans. Power Syst. 24(2), 9911001 (2009)
28. Sadat Hosseini, S.S., Gandomi, A.H.: Discussion of Economic Load DispatchA Com-
parative Study on Heuristic Optimization Techniques With an Improved Coordinated Ag-
gregation-Based PSO. IEEE Trans. Power Syst. 25(1), 590 (2010)
29. Fesanghary, M., Ardehali, M.M.: A novel meta-heuristic optimization methodology for
solving various types of economic dispatch problem. Energy 34, 757766 (2009)
30. Yang, X.S., Gandomi, A.H.: Bat Algorithm: A Novel Approach for Global Engineering
Optimization. Eng. Computation 29(5), 464483 (2012)
31. Sangiamvibool, W., Pothiya, S., Ngamroo, I.: Multipletabu search algorithm for economic
dispatch problem considering valve-point effects. Electrical Power and Energy Sys-
tems 33(4), 846854 (2011)
32. Subbaraja, P., Rengaraj, R., Salivahanan, S.: Enhancement of self-adaptive real coded ge-
netic algorithm using Taguchi method for economic dispatch problem. Appl. Soft. Com-
put. 11, 8392 (2011)
33. Yang, X.S.: Bat algorithm for multi-objective optimization. Int. J. of Bio-Inspired Compu-
tation 3(5), 267274 (2011)
34. Fister, I., Fong, S., Brest, J., Fister, I.: A Novel Hybrid Self-Adaptive Bat Algorithm. The
Scientific World Journal (2014)
35. Fister, I.J., Fong, S., Brest, J., Iztok, F.: Towards the self-adaptation in the bat algorithm.
In: Proceedings of the 13th IASTED International Conference on Artificial Intelligence
and Applications (2014)
36. Yang, X.-S.: A New Metaheuristic Bat-Inspired Algorithm. In: Gonzlez, J.R., Pelta, D.A.,
Cruz, C., Terrazas, G., Krasnogor, N. (eds.) NICSO 2010. Studies in Computational Intel-
ligence, vol. 284, pp. 6574. Springer, Heidelberg (2010)
37. Gandomi, A.H., Yang, X.S., Talatahari, S., Alavi, A.H.: BatAlgorithm for Constrained Op-
timization Tasks. Neural Comput. Appl. 22(6), 12391255 (2013)
38. Gandomi, A.H., Yang, X.S.: Chaotic Bat Algorithm. J. Comput. Sci. 5(2), 224232 (2014)
39. Yang, X.S., He, X.: Bat algorithm: Literature review and applications. Int. J. of Bio-
Inspired Computation 5(3), 141149 (2013)
40. Biswal, S., Barisal, A.K., Behera, A., Prakash, T.: Optimal power dispatch using BAT al-
gorithm. In: International Conference on Energy Efficient Technologies for Sustainability
(ICEETS), pp. 10181023 (2013)
41. Sakthivel, S., Natarajan, R., Gurusamy, P.: Application of Bat Optimization Algorithm for
Economic Load Dispatch Considering Valve Point Effects. Int. J. Comput. Appl. 67(11)
(2013)
42. Panigrahi, B.K., Pandi, V.R., Das, S.: Adaptive particle swarm optimization approach for
static and dynamic economic load dispatch. Energy Convers Manage. 49(6), 14071415
(2008)
43. Wang, S.-K., Chiou, J.-P., Liu, C.-W.: Non-smooth/non-convex economic dispatch by a
novel hybrid differential evolution algorithm. IET Gen., Transm., Distrib. 1(5), 793803
(2007)
146 S.S.S. Hosseini et al.
44. Walters, D.C., Sheble, G.B.: Genetic algorithm solution of economic dispatch with valve-
point loading. IEEE Trans. Power Syst. 8(3), 13251332 (1993)
45. Ma, H., El-Keib, A.A., Smith, R.E.: A genetic algorithm-based approach to economic dis-
patch of power systems. In: IEEE Conf. (1994)
46. Fister Jr, I., Fister, D., Fister, I.: Differential evolution strategies with random forest re-
gression in the bat algorithm. In: Proceeding of the Fifteenth Annual Conference Compa-
nion on Genetic and Evolutionary Computation Conference Companion, pp. 17031706.
ACM (2013)
47. Fister Jr, I., Fister, D., Yang, X.S.: A Hybrid Bat Algorithm. Electrotechnical Re-
view 80(1-2), 17 (2013)
48. Lin, W.M., Cheng, F.S., Tsay, M.T.: An improved tabu search for economic dispatch with
multiple minima. IEEE Trans. Power Syst. 17(1), 108112 (2002)
49. Victoire, T.A.A., Jeyakumar, A.E.: Hybrid PSO-SQP for economic dispatch with valve-
point effect. Electric Power Syst. Res. 71(1), 5159 (2004)
50. Coelho, L.D.S., Mariani, V.C.: An improved harmony search algorithm for power eco-
nomic load dispatch. Energy Convers. Manage. 50, 25222526 (2009)
51. Liu, D., Cai, Y.: Taguchi method for solving the economic dispatch problem with non-
smooth cost functions. IET Gen. Transm. Distrib. 1(5), 793803 (2007)
52. Ling, S.H., Leung, F.H.F.: An Improved genetic algorithm with average-bound crossover
and wavelet mutation operation. Soft Comput. 11(1), 731 (2007)
53. Chiang, C.-L.: Genetic-based algorithm for power economic load dispatch. IEE Proc. Gen-
er. Transm. Distrib. 1(2), 261269 (2007)
54. Coelho, L.S., Mariani, V.C.: Correction to Combining of chaotic differential evolution and
quadratic programming for economic dispatch optimization with valve-point effect. IEEE
Trans. Power Syst. 21(3), 1465
55. Coelho, L.D.S., Mariani, V.C.: Combining of chaotic differential evolution and quadratic
programming for economic dispatch optimization with valve-point effect. IEEE Trans.
Power Syst. 21(2), 989996 (2006)
56. Selvakumar, A.I., Thanushkodi, K.: Anti-predatory particle swarm optimization: Solution
to nonconvex economic dispatch problems. Electric. Power Syst. Res. 78, 210 (2008)
57. Chaturvedi, K.T., Pandit, M., Srivastava, L.: Self-organizing hierarchical particle swarm
optimization for nonconvex economic dispatch. IEEE Trans. Power Syst. 23(3), 1079
1087 (2008)
58. Al-sumait, J.S., Sykulski, J.K., Al-Othman, A.K.: A hybrid GAPSSQP method to solve
power system valve-point economic dispatch problems. Appl. Energ. 87, 17731781
(2010)
59. Al-Sumait, J.S., Al-Othmam, A.K., Sykulski, J.K.: Application of pattern search method to
power system valve-point economic load dispatch. Electr. Power Energy Syst. 29(10),
720730 (2007)
60. Yang, X.S., Sadat Hosseini, S.S., Gandomi, A.H.: Firefly algorithm for solving nonconvex
economic dispatch problems with valve loading effect. Appl. Soft. Comput. 12, 1180
1186 (2012)
61. Amjady, N., Sharifzadeh, H.: Solution of non-convex economic dispatch problem consi-
dering valve loading effect by a new modified differential evolution algorithm. Int. J. Elec.
Power 32(8), 893903 (2010)
Part III
Hybridization in Computational
Intelligence
Hybrid Artificial Neural Network
for Fire Analysis of Steel Frames
1 Introduction
In the present analysis, alternative approach of modeling the mechanical be-
havior of material when exposed to high temperatures, as expected in res, is
presented. Based on series of stress-strain curves obtained experimentally for
various temperature levels [1], the articial neural network (ANN) is employed
in material modeling of steel. In general, steel at high temperatures behaves
very non-linearly. In order to avoid the ambiguities described above, we have
recently witnessed various attempts to capture the inelastic part of the material
model just by a proper set of time-independent stress-strain curves at various
temperatures involving both plastic and viscous eects. Characteristic samples
Corresponding author.
are the material models proposed by Eurocode 3 [2] and BS5950 [3] wherein the
parameters of temperature dependent bilinear diagrams with elliptic intermedi-
ate part are given. As has been stated by Huang and Tan [6], the heating rate
and the duration of elevated temperatures have considerable inuence on the
development of strains and stresses over the structure. Consequently the time-
independent material models are only suitable in the cases when the temperature
of the steel does not exceed 450 C. In real re, such temperature regime can only
be expected (i) at heat-protected structures when the exposure to the highest
temperatures does not last for too long or (ii) at very low stock of combustion
material not allowing the re to develop to full extent. In others, more realis-
tic cases, the experimental data obtained at certain temperature-time curve like
ISO 834 or constant heating rate are only of limited applicability.
A relatively large set of results of uniaxial tests on structural steel at constant
specimen heating rate of 10 C/min has been provided by Kirby and Preston [1]
for two steel qualities (grades 43A and 50B). The data are given in tabular form
as two series of stress-strain pairs at various temperature levels for the strains
up to 2% in the temperature interval from 20 C to 900C. In the temperature
interval from 250 C to 600C, the irregular wavy shapes of the stress-strain
curves do not allow to be approximated by bilinear model with either elliptic
or parabolic intermediate part. Therefore, the idea of employing the articial
neural network (ANN) was introduced in order to describe the stress-strain-
temperature relations. Some diculties appearing within the course of modeling
the properties of steel at elevated temperatures by the ANN have been solved
by combining the ANN results with linear regression in the linear elastic range
and with linear extrapolation in the hardening range.
The quality of solutions as obtained by ANN depends on setting of the ANN
parameters. However, the properly setting of parameters is not known in ad-
vance. Normally, this can be found during the extensive experimental work man-
ually. In order to avoid this time-consuming process, a genetic algorithm (GA) is
proposed for performing the tuning process. The GA at the higher level appears
as a meta-heuristic to control the parameters of ANN acting at the lower level,
i.e., solving of the problem. Thus, the optimal parameter setting is searched for.
Therefore, this GA is also named as meta-GA.
The parameters of ANN for re analysis of steel frames are coupled into
a representation of candidate solutions within the meta-GA. These solutions
undergo actions of crossover and mutation. The quality of solutions is evaluated
after applying the parameters mapped from a representation of solution in meta-
GA to ANN and obtaining the error rate from the ANN algorithm running
with the proposed parameter setting. As a matter of fact, the results of tuning
parameters with the meta-GA signicantly improved the results of the manually
tuned parameters of ANN as performed by Hozjan et al. in [16].
The structure of the chapter is as follows. Section 2 deals with background
information where the re analysis of steel frames is described and a theory of
ANN and GAs is introduced. In Section 3, the proposed meta-GA algorithm is
presented in detail. Section 4 deals with experiments and results. The chapter
Hybrid ANN for Fire Analysis of Steel Frames 151
concludes with an overview of the performed work and possible directions for
the further work are outlined.
2 Background Information
2.1 Fire Analysis of Steel Frames
In technical codes the term re resistance corresponds to the experimentally
veried endurance of individual structural elements or minor frames with re-
gard to standardized heating mode in a test furnace. Even accurate results of
experiments in a test furnace do not provide an adequate explanation of global
behavior of structure as a whole in a real re. Generally, the re resistance of
whole structures is considerably greater than that of the individual structural
elements. Since full-scale tests of structures represent high cost, but on the other
hand numerical models oer noticeably cheaper solution for determination of re
resistance of the structure. Yet they have to be ecient and accurate enough to
be able to perform such a complex analysis. Therefore, experimental data about
thermo-mechanical properties of materials and structural elements are the nec-
essary basis for development of any computational analysis and their required
reliability increases with the eciency and accuracy of available computing tools.
Current numerical models for mechanical analysis of structures exposed to re
are mostly based on nite element method with various levels of non-linearity
and are combined with non-linear material models calibrated with regard to
experimental results. However, there are several uncertainties related to the in-
uence of temperature gradients and the development of plastic and viscous
strains of steel at temperatures above 400 C.
One of the most important component in accuracy of the numerical model
for re analysis of steel frames is material model. Here, alternative approach of
constructing the material model of steel at high temperatures with help of ANN
based on available experimental data [1] is presented in short. Full description
of this application is given in [16].
This equation is illustrated in Figure 2. The activation function f(.) enables the
modeling of an arbitrary non-linear relation between input and output variables.
Dierent functions can be used as an activation function, while usual choice for
the activation function is a sigmoid function
Hybrid ANN for Fire Analysis of Steel Frames 153
output signal
nl
y1 nl (output layer)
w11 wk1nl
y1
nl-1 nl-1
y2
nl-1
yk nl-1
layer numbers
y1 ym
2 2
1
w1m
1
w2m
1
y1 y2 yn
1 1 1
0 (input layer)
x1 x2 xn
1
f (y) = , (2)
1 + ey
tanh y, or Gaussian (Figure 5). The behavior of the neural network also depends
k
on the values of the weights wij and thresholds ki which have to be determined
by the learning (training) procedure. The set of known input and output values
is termed an input-output pair. All input-output pairs are often divided into
k
two sets. With so called learning or training set the connection weights wij and
k
thresholds i are determined. Once the learning procedure ends, it means that
the neural network performs adequately for all input-output pairs in the learning
set and the neural network is assessed on the testing set of data.
Training procedure can be in some cases ill-conditioned if the input and/or
output data are not normalized (see e.g. [5]). Therefore, the values of input
and output units have to be normalized. The normalization of the values of
output units depends on the range of activation function. Usually, the linear
transformation works well, although sometimes a non-linear transformation may
help if the data are clustered.
154 T. Hozjan, G. Turk, and I. Fister
neuron on layer k
input signals
nurons on layer k-1
outut function
0 a 0 a 0 a
representation of individuals,
crossover,
mutation,
parent selection,
survivor selection.
Crossover. Crossover operators in real-coded GAs build two ospring from two
parents. They are applied to a population of solutions according to a probability
of crossover pc control parameter. Deb in [19] proposed more crossover opera-
tors that can be applied to real-coded GAs. In this chapter, we are focused on
blend (BLX) and simulated binary (SBX) crossover that are referred to in the
remainder of this chapter.
Note that generated osprings are symmetric according to the parent solutions.
Mutation. Mutation is an unary operator that uses only one parent and creates
one ospring by applying some kind of randomized change to the solution [24].
In real-coded GAs especially two mutations are employed, as follows:
random mutation,
polynomial mutation.
The mutation operator is applied to the parent solution regarding the probability
of mutation pm control parameter. In the remainder of the paper, these mutation
operators are described in detail.
(Ub) (Lb)
where xj and xj denote an upper and lower bounds of the observed element
of solution, U(0, 1) is a random variable drawn from a uniform distribution in the
interval [0, 1]. This operator is equivalent to a random initialization because of
independency of the parent solution. In order to consider the parent solution, the
random mutation can select the new value of an element in the neighborhood
of the original element using a uniform probability distribution according to
equation
(t+1) (t)
xij = xij + (U(0, 1) 0.5) j , (9)
where j is the maximal altered perturbation of the j-th problem variable.
158 T. Hozjan, G. Turk, and I. Fister
The roulette-wheel by this selection is also divided into n-sections. The posi-
tion of the rst solution is determined by a random number r, while the positions
of all other solutions are determined by moving identical distances 1/n along the
periphery of the roulette-wheel.
Survivor Selection. In general, the real-coded GAs use the generational sur-
vivor selection, where the best n solutions according to their tness values are
selected between n parents and n ospring for the next generation. Obviously,
this model is also elitist because of preserving the best parents to arise in the
next evolutionary cycle. On the other hand, the worse solutions are eliminated
from the population.
well the ANN must be learned. Some parameters can be varied and therefore
it also undergo the tuning process. For instance, the parameters, like the step
size, range of sharing weights and a learning rate ratio belong to this class.
Meaning of these parameters is as follows. The parameter step size controls the
convergence speed of back propagation networks, the parameter range of sharing
weights determines the range of numerical weights that are tuned by the ANN
learning algorithm, and the parameter learning rate ratio determines a size of
weights and bias change during learning.
The pseudo-code of real-coded meta-GA controlling the parameters of ANN
for re analysis of steel frames is illustrated in Algorithm 1.
As can be seen from Algorithm 1, two parents are selected to enter in opera-
tion of crossover in each generation. The obtained ospring undergo acting the
operator of mutation. Then, the quality of each ospring is calculated using the
evaluation function. Finally, the best solution survive and reproduce their best
characteristics in the next generation. In addition, two features are dened in
order to complete an evolutionary cycle of the meta-GA algorithm, i.e., initial-
ization and termination condition. The former initialize the individuals of the
starting population randomly, while the latter terminate the evolutionary cycle,
when a termination condition is satised.
The chromosome of the meta-GA consists of eight genes whose meanings
are represented in Table 1. In the optimization procedure up to maximum four
hidden layers of the ANN mesh were used in the analysis.
The corresponding topology in Table 1 is described with the geometry 2-
50-50-1 that denes the ANN with two hidden layers consisting of 50 neurons.
The number of output neurons dened by the user is two, while the output
neuron is only one. The evaluation function in the meta-GA is obtained by
launching the ANN algorithm with the corresponding parameters mapped from
the representation of solution and exploring the error rate obtained after nishing
the ANN. The task of meta-GA is to minimize the error rate Ep .
Hybrid ANN for Fire Analysis of Steel Frames 161
This study extends the paper of Hozjan et al. [16] that proposed the ANN for
modeling the mechanical behavior of steel frames exposed to the re. Thus,
experiments were conducted on the experimental data obtained in [1]. In the
experiments, the ANN was thought to estimate stress , while strain and
temperature T were used as input data. The calculation was carried out for steel
strength fy = 35.5 kN/cm2 .
The goal of this study was to show that the manual setting of ANN param-
eters used in [16] by eye were far to be an optimal. In contrast, the parameter
setting tuned by meta-GA can be near-optimal. Consequently, the results of
material modeling with ANN using these parameters can improve the results of
the material modeling as performed by the ANN using the manual parameter
setting. During the experiments, the allowed relative error was set to 0.05, which
is a relatively low value in an ANN training procedure.
The parameters of meta-GA were set as illustrated in Table 2. Note that
the number of tness function evaluations was F Es = 100 10 = 1, 000 in
each run of meta-GA. On the other hand, the number of independent runs was
limited to 500, in which the best parameter setting was searched for. The SBX
crossover and polynomial mutation operators were applied in the tests, because
Parameter Value
number of generations 100
population size 10
probability of crossover pc 0.9
probability of mutation pm 0.01
selection type tournament of size 2
crossover type SBX crossover
mutation type polynomial mutation
number of runs 500
162 T. Hozjan, G. Turk, and I. Fister
they produced the best ospring as shown during the extensive experimental
work. The same hold also for the other parameters in Table 1.
This section is divided into three subsections. In the rst subsection, the test
suite is described. In the second subsection searching for the optimal parameter
setting of the ANN is discussed, while in third subsection comparisons of the
material models obtained with dierent parameter setting are presented.
The total number of input-output pairs as presented in [1] was 527, of which
435 randomly selected pairs were used for learning and the remaining 92 were
used as testing pairs. The input-output data of the test and the learning set are
assembled in the way that the dispersion of the learning and testing data is as
far as possible similar (Table (3)).
Plenty of calculations with dierent topology of ANN were carried out. For
instance, one of the test result obtained by ANN with manual parameter setting
obtained coecient of correlation equals to r2 = 0.8553. Here, the ANN topology
consists of 3 hidden layers with the geometry 2-20-40-20-1. Clearly, this topology
is not good enough as can be seen in Figure 6. The better solution was obtained
when the geometry 2-30-30-1 was employed. This means, there were two hidden
layers, each of them including 30 neurons. The eciency of learning procedure
is shown in Figure 7, where actual and calculated values are compared. In this
Hybrid ANN for Fire Analysis of Steel Frames 163
150
Calculated
100
50
0
0 50 100 150 200 250 300 350 400
Actual Values
Fig. 6. Actual and calculated values of stress obtained by ANN with manual param-
eter setting
case, the coecient of correlation was higher, i.e., r2 = 0.9975, but the stress-
strain relationship constructed with this ANN topology was not good enough
and therefore further improvements were essential. From this results, it can be
seen that the procedure of searching the optimal topology for ANN by guessing
is demanding. However, the results of this procedure also depend on luck, since
the experimental data is highly non-linear.
In order to further improve the learning procedure of the ANN algorithm,
the meta-GA was introduced. As a results, the allowed relative error was also
lowered to very small value of 0.002 after using the meta-GA. Parameter values
2
350
300
250
200
150
100
50
0
0 50 100 150 200 250 300 350 400
Actual values
Fig. 7. Actual and calculated values of stress obtained by ANN with meta-GA
164 T. Hozjan, G. Turk, and I. Fister
Table 4. Manual and meta-GA suggested parameter setting for ANN training proce-
dure
manually set by an expert and the best parameter setting as proposed by the
meta-GA are presented in Table 4.
The proposed ANN topology consists of 4 hidden layers with the geometry of
2-165-103-73-79-1. Here, the coecient of correlation was almost perfect for the
suggested topology, i.e., r2 = 0.9999. (Fig. 8). This example presents the success-
ful application of the meta-GA and also shows the capability of this algorithm.
Moreover, it proves that using the meta-GA for optimizing the parameters of the
ANN is necessary due to all the non-linearities involved in this material modeling
problem.
300
250
Calculated
200
150
100
50
0
0 50 100 150 200 250 300 350 400
Actual Values
Fig. 8. Comparison between actual and calculated values of stress in case of optimized
ANN geometry
Hybrid ANN for Fire Analysis of Steel Frames 165
40
T [ C]
35 20
100
30 200
Stress [kN/cm ]
300
2
25 400
500
600
20 700
800
15 test
10
ANN and the experimental ones is very good along the entire curve for all
temperature levels (Fig. 9).
However, some diculties appear within the course of modeling the proper-
ties of steel at elevated temperatures by the ANN. Firstly, the yield points of
particular stress-strain curves are not explicitly dened by the curve shape itself.
The problem was solved by plotting the rst derivative of the experimentally ob-
tained stress-strain relations where the yield limits are much better pronounced.
Secondly, due to the regressions used in the ANN the obtained approximations
for the stress-strain relations below the yield limit exhibit certain deviations
from a linear shape. Assuming ideal linear behavior of steel the linear regression
based on actual experimental data has been used for this range. Lastly, since the
experimental data are given for the values of strains up to 2%, the ANN model
is inadequate for the range of strains higher than 2% and a constant hardening
parameter is introduced in this strain range.
According to that, the presented material model is divided into three parts
(Figure 10). The rst part is linear elastic M < Y and the stress is determined
by linear elastic law
40
35 1. part
Hooks law
3. part
30
stress s [kN/cm2]
2. part hardening
25 ANN solution
20
15
10
5
0
0 ey 0.5 1 1.5 2 2.5 3
strain e [%]
Fig. 10. Stress-strain relationship at ambient temperature
where kE,T stands for reduction factor of the elastic modulus, describing its
variation in dependency of the temperature T , referring to the elastic modulus
E20 at the room-temperature T = 20 C.
The second part stands for the plastic range where the mechanical strain
exceeds the yield strain M > Y . In this range, the stress is calculated by the
ANN, in accordance with the actual values of the mechanical strain M and
temperature T
(M , T ) = fANN (M , T ). (14)
In the range where the mechanical strain exceeds the value of M > 1.85%, a
uniform strain-hardening parameter K is considered. The value of K is deter-
mined by the slope of the stress-strain curve at the strain M = 1.85% and at
the corresponding temperature T (K = K(M = 1.85%, T )).
We should note that experimental data [1] are determined at the heating rate
10 C/min and therefore the applicability of this material model is limited. The
authors mentioned that the results for these heating rates are quite similar to the
results obtained by other researchers, whose measurements were performed at
dierent heating rates, namely 2.5 C/min, 5 C/min and 20 C/min. With this
statement the application of the presented model at dierent heating rates is
acceptable, when provided that the heating rate does not dier from 10 C/min
excessively.
4.4 Discussion
Although the meta-GA parameters were initialized with a relatively small values,
the time complexity of a meta-GA and ANN algorithms running sequentially one
Hybrid ANN for Fire Analysis of Steel Frames 167
after the other was crucial increased because of the higher time complexity of the
ANN especially, when the ANN with more hidden layers must be learned. Fortu-
nately, the meta-GA algorithm does not demand any user interaction. Therefore,
the quality of obtained results justies increasing of the performance especially
because here we deal with a real-world problem, where the results of the algo-
rithm can have a crucial impact on the behavior of material structure during
the re.
5 Conclusion
This chapter extends the results obtained by Hozjan et al. in [1] that using
the ANN searched for a material model of steel frames exposed to the high
temperatures caused by re. Although the reported error rate obtained by the
ANN learning procedure was kept within the normal range, the goal of this study
was to show that near-optimal parameter setting of the ANN can be found using
the meta-GA. In line with this, the proposed meta-GA arises as a meta-heuristic
operating at the higher level and controls the parameters of the ANN solving
the problem that operates at the lower level.
The meta-GA was applied to a benchmark problem suite as proposed in [1].
The results of ANN with parameter setting as proposed by this algorithm im-
proved the results of the ANN with parameter setting as proposed by an expert
signicantly and conrmed the fact that the nature-inspired algorithms work
better when they are hybridized. Meta-GA algorithm together with ANN pro-
vides a powerful toll which can be used in many engineering problems especially
in those where parameters behave non-linearly and irregularly.
References
1. Kirby, B.R., Preston, R.R.: High temperature properties of hot-rolled, structural
steels for use in re engineering design studies. Fire Safety Journal 13, 2737 (1988)
2. Eurocode 3: Design of steel structures, Part 1.2: General rules-Structural re de-
sign. European Committee for Standardisation, Brussels (2001)
3. British Standard Institute. BS5950-8:Structural Use of Steelwork in Building -
Part 8: Code of Practice for Fire Resistance Design. British Standards Institution,
London (2003)
4. Lippmann, R.P.: An Introduction to Computing with Neural Nets. IEEE Magazine
on Acoustics, Signal and Speech Processing 4(2), 422 (1987)
5. Sarle, W.S.: Neural Network FAQ. Periodic posting to the Usnet newsgroup
comp.ai.neural-nets (2002), ftp://ftp.sas.com/pub/neural/FAQ.html
6. Huang, Z.F., Tan, K.H.: Eects of External Bending Moments and Heating
Schemes on the Responses of Thermally-restrained Steel Columns. Engineering
Structures 26(6), 769780 (2004)
168 T. Hozjan, G. Turk, and I. Fister
7. Seed, G.M., Murphy, G.S.: The applicability of neural networks in modelling the
growth of short fatigue cracks. Fatigue & Fracture of Engineering Materials &
Structures 21, 183190 (1998)
8. Haque, M.E., Sudhakar, K.V.: ANN based prediction model for fatigue crack
growth in DP steel. Fatigue & Fracture of Engineering Materials & Structures 23,
6368 (2001)
9. Sterjovski, Z., Nolan, D., Carpenter, K.R., Dunne, D.P., Norrish, J.: Articial neu-
ral networks for modeling the mechanical properties of steels in various applica-
tions. Journal of Materials Processing Technology 170, 536544 (2005)
10. Sakla, S.S.S.: Neural network modeling of the load-carrying capacity of
eccentrically-loaded single-angle struts. Journal of Constructional Steel Re-
search 60, 965987 (2004)
11. Oreta, A., Kawashima, K.: Neural Network Modeling of Conned Compressive
Strength and Strain of Circular Concrete Columns. Journal of Structural Engi-
neering 129(4), 554561 (2003)
12. Tang, C.W., Chen, H.J., Yen, T.: Modeling Connement Eciency of Reinforced
Concrete Columns with Rectilinear Transverse Steel Using Articial Neural Net-
works. Journal of Structural Engineering 129(6), 775783 (2003)
13. Zhao, Z.: Steel columns under re - a neural network based strength model. Ad-
vances in Engineering Software 37(2), 97105 (2004)
14. Mikami, I., Tanaka, S., Hiwatashi, T.: Neural Network System for Reasoning Resid-
ual Axial Forces of High-Strength Bolts in Steel Bridges. Computer-Aided Civil and
Infrastructure Engineering 13, 237246 (1998)
15. Papadrakakis, M., Lagaros, N.D., Plevris, V.: Design optimization of steel struc-
tures considering uncertainties. Engineering Structures 27, 14081418 (2005)
16. Hozjan, T., Turk, G., Srpcic, S.: Fire analysis of steel frames with the use of articial
neural networks. Journal of Constructional Steel Research 63, 13961403 (2007)
17. Michalewicz, Z.: Genetic algorithms + data structures = evolution programs.
Springer, Berlin (1992)
18. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learn-
ing. Addison-Wesley Longman Publishing Co., Inc., Boston (1989)
19. Deb, K.: Multi-Objective Optimization Using Evolutionary Algorithms. John Wi-
ley & Sons, Inc.,, New York (2001)
20. Goldberg, D.E., Deb, K.: A comparative analysis of selection schemes used in
genetic algorithms. In: Foundations of Genetic Algorithms 1 (FOGA-1), pp. 4149
(1991)
21. Eshelman, L.J., Schaer, J.D.: Real-coded genetic algorithms and interval-
schemata. In: Foundations of Genetic Algorithms 2 (FOGA-2), pp. 187202 (1993)
22. Deb, K., Agrawal, R.B.: Simulated binary crossover for continuous search space.
Complex Systems 9(2), 115148 (1995)
23. Deb, K., Kumar, A.: Real-coded genetic algorithms with simulated binary
crossover: Studies on multi-modal and multi-objective problems. Complex Sys-
tems 9(6), 431454 (1995)
24. Eiben, A., Smith, J.: Introduction to Evolutionary Computing. Springer, Berlin
(2003)
25. Fister, I., Mernik, M., Filipic, B.: Graph 3-coloring with a hybrid self-adaptive
evolutionary algorithm. Computational Optimization and Applications 54(3), 741
770 (2013)
Hybrid ANN for Fire Analysis of Steel Frames 169
26. Fister, I., Mernik, M., Filipic, B.: A hybrid self-adaptive evolutionary algorithm
for marker optimization in the clothing industry. Applied Soft Computing 10(2),
409422 (2010)
27. Fister Jr, I., Yang, X.-S., Fister, I., Brest, J., Fister, D.: A brief review of
nature-inspired algorithms for optimization. Electrotechnical Review 80(3), 116
122 (2013)
28. Fister Jr, I., Suganthan, P.N., Strnad, D., Brest, J., Fister, I.: Articial neural
networks regression on ensemble strategies in dierential evolution. In: 20th In-
ternational Conference on Soft Computing, Mendel 2014, pp. 6570. University
of Technology, Faculty of Mechanical Engineering, Institute of Automation and
Computer Science, Brno (2014)
A Differential Evolution Algorithm with a Variable
Neighborhood Search for Constrained
Function Optimization
1 Introduction
*
Corresponding author.
In general, the equality constraints can be transformed into inequality form and can
be combined with other inequality constraints as
G (x ) + H ( x )
i i
(x ) = i =1 i = p +1
Due to the existence of several mutation strategies in the basic DE algorithms, for a
general description, we employ the DE / rand / 1 / bin variant of Storn and Price
[38,39]. In the basic DE algorithm, the initial target population is established by NP
number of individuals. A target individual in the population contains a D-dimensional
vector with parameter values. These parameter values are initially generated
uniformly between predetermined search bounds between xijmin and xijmax as follows:
( )
xijt = xijmin + x ijmax x ijmin r (6)
(
vijt = x ajt 1 + F xbjt 1 x cjt 1 ) (7)
174 M.F. Tasgetiren et al.
where a , b , and c are three randomly chosen individuals from the target population
such a way that (a b c i (1,.., NP )) and ( j = 1,2,.., D ) . F > 0 is a mutation
scale factor affecting the differential variation between two individuals.
In the next step, the trial individual can be obtained by making a uniform crossover
between the target and mutant individuals as follows:
vijt if rijt CR or j = Dj
u ijt = (8)
t 1
xij otherwise
where the D j denotes a randomly selected dimension. It guarantees that at least one
parameter of each trial individual comes from the mutant individual. CR is a cros-
sover rate within the range [0,1] , and rijt is a random number generated from U[0,1].
In case of any violation of the parameter values of trial individual during the evolu-
tion, they are restricted to:
(
u ijt = xijmin + xijmax xijmin r1 ) j = 1,2,.., D (9)
xit = t 1
( ) ( )
u it if f u it f xit 1
(10)
xi otherwise
N k ( ) = M k (v ), CR ( , v ) (13)
In other words, in the first neighborhood structure, we employ a very high muta-
tion rate F and a very high crossover rate Cr . However, in the second neighborhood
structure, we determine them uniformly in the range [0,1], randomly. With the above
definitions and temporary individuals and x * , we develop a VNS algorithm to
generate a trial individual as shown in Fig. 1.
Procedure VNS ( xi )
k max = 2
k =1
= xi
do{
x * = N k ( )
if ( )
f x * < f ( )
= x*
k =1
else
k = k +1
}while(k k max )
ui =
return ui
Endprocedure
The performance of VNS algorithms depends on what strategy is used in the first
neighborhood. As explained before, the equation (11) is used as the first neighbor-
hood whereas the equation (12) is used as the second neighborhood. Note that as long
as the first neighborhood improves the current solution, the neighborhood counter k
176 M.F. Tasgetiren et al.
will be 1 indicating that the first neighborhood will be employed. Otherwise, the
neighborhood counter k will be increased to 2 indicating that the second neighbor-
hood will be employed. If the second neighborhood improves the solution, it gets
back to the first neighborhood again until the second neighborhood fails. It should be
noted that when we compare two solutions in the VNS algorithm, we use the pena-
lized fitness values obtained by the NFT method that will be explained in Section 4.
x * = k (a + b ) x (16)
( )
xij* = kU a j , b j xij (17)
a = min (x ), b = max (x )
j ij j ij (18)
x = U (a , b ) if x < x
*
ij j j
*
ij
min
j or xij* > x max
j
After establishing and evaluating the target population, the above GOBL algorithm
is also used to obtain the opposite target individual. The better one is retained in the
target population.
for i = 1 to NP
for j = 1 to D
if (r < iP ) then
x aj = TournamentSelect ()
uij = x aj
else
uij = uij
endfor
endfor
Fig. 2. Injection Procedure
3.3 Selection
When the selection for the next generation is carried out, we employ the EC and SF
constraint handling methods that will be summarized in Section 4 as follows: For each
individual in the trial population, we check the (t ) level. If the constraint violation is
less than (t ) level, we treat the trial individual as a feasible solution. Then we em-
ploy the SF method whether or not the trial individual will survive to be in the next
generation. In addition, we simply use the SF method to update the best so far solu-
tion in the population.
3.4 Diversification
In order to further diversify the target population, we propose a diversification me-
chanism based on the inversion of the dimensional values of the target individual and
an injection procedure explained above. For a small portion of the target population,
following diversification procedure is applied to the randomly selected individuals as
shown in Fig. 3.
x aj = RandomlySelect ()
( )
x aj = invert xaj
for j = 1 to D
if (r < iP ) then
xbj = TournamentSelect ()
x aj = xbj
else
xaj = xaj
endfor
Fig. 3. Diversification Procedure
178 M.F. Tasgetiren et al.
4 Constraint Handling
Evolutionary algorithms can yield infeasible solutions. In this case, the general ten-
dency is to utilize some constraint handling approaches [7, 49]. In this paper, we used
the following constraint handling methods:
where f all is the unpenalized value of the best solution obtained so far whereas f feas
is the value of the best feasible solution yet obtained. As mentioned in [7], the adap-
tive term may result in zero-or over-penalty. Due to this reason, we only take the
dynamic part of the above penalty function with NFT threshold into account as
follows:
H j (x )
p G (x ) m
f p (x ) = f (x ) + i + NFT (21)
i =1 NFTi j = p +1 j
NFT0
The basic form of the NFT method is presented as NFT = where NFT0 is
1+ *t
the initial value of the NFT method; and t are user-defined positive value and gen-
eration counter, respectively. is severity parameter. Because of the conversion
process of the equality constraints to the inequality constraints by subtracting from
the absolute value of the constraint value and is determined beforehand, the NFT0
is chosen as 1e-4.
A Differential Evolution Algorithm with a Variable Neighborhood Search 179
The -constraint handling method was proposed in [44] in which the constraint re-
laxation is monitored by parameter. A proper control of the parameter is neces-
sary while obtaining good feasible solutions for problems with equality constraints
[44]. The level is updated according to the control generation t C . After t ex-
ceeds t C , the level is set to zero to end up with feasible solutions. The main idea
lies behind the EC method is that solutions having violations less than (t ) are consi-
dered to be feasible solutions when making selection for the next generation. The
general framework is given as follows:
(0 ) = (x ) (22)
cp
t
(0)1
(t ) = tC , 0 < t < tC . (23)
0, t tC
5 Computational Results
The DE_VNS algorithm was coded in C++ and run on an Intel P4 1.33 GHz Laptop
PC with 256MB memory. The population size is taken as NP=60. The NFT0 is fixed
at 0.0001. Injection probability is taken as 0.005 whereas the diversification probabili-
ty is taken as 0.05. For the EC constraint handling method, following parameters are
used as = 0.25 NP , tC = 0.4 * MaxGen and cp = 2 . We carried out 30 replications
for each benchmark problem and average, minimum and standard deviation of 30
replications are provided. Note that real numbers are rounded to zero after 10 digits in
the standard deviation calculations.
We compare our algorithm to the best performing algorithms from the literature
such as MDE [22], ECHT-EP2[30] and SAMO-DE [32]. The computational results
are given in Table 1. As seen from Table 1, the DE_VNS algorithm was able to find
the optimal solutions with zero standard deviations for 13 out of 22 benchmark prob-
lems. The DE_VNS algorithm was slightly better than SAMO-DE because it was able
to find 12 optimal solutions with zero standard deviations. The performance of the
ECHT-EP2 was slightly better than DE_VNS and SAMO-DE since it was able find
14 optimal solutions with zero standard deviations. The clear winner was the MDE
algorithm due to the fact that it was able to find 19 optimal solutions with zero stan-
dard deviations. However, DE_VNS, SAMO_DE and ECHT-EP2 algorithms were
run for 240000 function evaluations whereas MDE was run for 500000 function eval-
uations. In 4 benchmark functions, the standard deviation of the DE_VNS algorithm
was smaller than both SAMO_DE and ECHT-EP2, respectively. Together with all
180 M.F. Tasgetiren et al.
algorithms compared, the DE_VNS algorithm was able to find the optimal solutions
in all 30 replications. In other words, feasibility rate was 100 %. In summary, the
simple DE_VNS algorithm was competitive to the best performing algorithms from
the literature.
Table 1. (continued)
6 Conclusions
References
[1] Qin, A.K., Huang, V.L., Suganthan, P.N.: Differential Evolution Algorithm with strategy
adaptation for global numerical optimization. IEEE Trans. Evol. Comput. 13, 398417
(2009)
[2] Mohamed, A.W., Sabry, H.Z.: Constrained optimization based on modified differential
evolution algorithm. Information Sciences 194, 171208 (2012)
[3] Zamuda, A., Brest, J.: Population reduction differential evolution with multiple mutation
strategies in real world industry challenges. In: Rutkowski, L., Korytkowski, M., Scherer,
R., Tadeusiewicz, R., Zadeh, L.A., Zurada, J.M., et al. (eds.) EC 2012 and SIDE 2012.
LNCS, vol. 7269, pp. 154161. Springer, Heidelberg (2012)
[4] Babu, B.V., Onwubolu, G.C. (eds.): New Optimization Techniques in Engineering.
STUDFUZZ, vol. 141. Springer, Heidelberg (2004)
[5] Becerra, R.L., Coello, C.C.A.: Cultural Differential Evolution for Constrained Optimiza-
tion. Comput. Methods Appl. Mech. Engrg. (2005)
[6] Chiou, J.-P., Wang, F.-S.: Hybrid Method of Evolutionary Algorithms for Static and Dy-
namic Optimization Problems with Applications to a Fed-Batch fermantation Process.
Computers and Chemical Engineering 23, 12771291 (1999)
[7] Coello, C.C.A.: Theoretical and Numerical Constraint-Handling Techniques Used with
Evolutionary Algorithms: A Survey of the State of the Art. Comput. Methods Appl.
Mech. Engrg. 191(11-12), 12451287 (2002)
[8] Part Two: Differential Evolution. In: Corne, D., Dorigo, M., Glover, F. (eds.) New Ideas
in Optimization, pp. 77158. McGraw-Hill (1999)
[9] Tizhoosh, H.R.: Opposition-based learning: a new scheme for machine intelligence. In:
Proceedings of International Conference on Computational Intelligence for Modeling
Control and Automation, pp. 695701 (2005)
[10] Wang, H., Wu, Z.J., Rahnamayan, S.: Enhanced opposition-based differential evolution
for solving high-dimensional continuous optimization problems. Soft Comput. 15(11),
21272140 (2011)
[11] Wang, H., Wu, Z.J., Rahnamayan, S., Kang, L.S.: A scalability test for accelerated DE
using generalized opposition-based learning. In: Proceedings of International Conference
on Intelligent System Design and Applications, pp. 10901095 (2009)
[12] Wang, H., Wu, Z.J., Rahnamayan, S., Liu, Y., Ventresca, M.: Enhancing particle swarm
optimization using generalized opposition-based learning. Inform. Sci. 181(20),
46994714 (2011)
[13] Brest, J., Sepesy Maucec, M.: Population size reduction for the differential evolution al-
gorithm. Appl. Intell., 228247 (2008)
[14] Deb, K.: An efficient constraint handling method for genetic algorithms. Computer Me-
thods in Applied Mechanics and Engineering 186, 311338 (2000)
[15] Koziel, S., Michalewicz, Z.: Evolutionary Algorithms, Homomorphous Mappings, and
Constrained Parameter Optimization. Evol. Comput. 7(1), 1944 (1999)
[16] Lampinen, J.: Multi-Constrained Optimization by the Differential Evolution. In: Proc. of
the IASTED International Conference Artificial Intelligence Applications (AIA 2001),
pp. 177184 (2001)
[17] Lampinen, J.: Solving Problems Subject to Multiple Nonlinear Constraints by the Diffe-
rential Evolution. In: Proc. of the 7th International Conference on Soft Computing,
MENDEL 2001, pp. 5057 (2001)
A Differential Evolution Algorithm with a Variable Neighborhood Search 183
[35] Sarimveis, H., Nikolakopoulos, A.: A Line Up Evolutionary Algorithm for Solving Non-
linear Constrained Optimization Problems. Computers & Operations Research 32,
14991514 (2005)
[36] Smith, A.E., Tate, D.M.: Genetic Optimization Using a Penalty Function. In: Forrest, S.
(ed.) Proc. of the Fifth International Conference on genetic Algorithms, pp. 499503.
Morgan Kaufmann (1993)
[37] Storn, R.: System Design by Constraint Adaptation and Differential Evolution. IEEE
Transactions on Evolutionary Computation 3, 2234 (1999)
[38] Storn, R., Price, K.: Differential Evolution a Simple and Efficient Adaptive Scheme for
Global Optimization over Continuous Spaces. Technical Report TR-95-012, ICSI (1995)
[39] Storn, R., Price, K.: Differential Evolution - A Simple and Efficient Heuristic for Global
Optimization over Continuous Space. Journal of Global Optimization 11, 341359 (1997)
[40] Rahnamayan, S., Tizhoosh, H.R., Salama, M.M.A.: Opposition-based differential evolu-
tion algorithms. In: Proceedings of IEEE Congress on Evolutionary Computation,
pp. 20102017 (2006)
[41] Rahnamayan, S., Tizhoosh, H.R., Salama, M.M.A.: Opposition-based differential evolu-
tion for optimization of noisy problems. In: Proceedings of IEEE Congress on Evolutio-
nary Computation (1872)
[42] Rahnamayan, S., Tizhoosh, H.R., Salama, M.M.A.: Opposition-based differential evolu-
tion. IEEE Trans. Evol. Comput. 12(1), 6479 (2008)
[43] Das, S., Suganthan, P.N.: Differential Evolution: A Survey of the State-of-the-Art. IEEE
Trans. Evolutionary Computation 15(1), 431 (2011)
[44] Takahama, T., Sakai, S.: Constrained Optimization by the Constrained Differential Evo-
lution with Gradient-Based Mutation and Feasible Elites. In: IEEE Congress on Evolutio-
nary Computation, Sheraton Vancouver Wall Centre Hotel, Vancouver, BC, Canada,
pp. 18 (2006)
[45] Long, W., Liang, X., Huang, Y., Chen, Y.: A hybrid differential evolution augmented
Lagrangian method for constrained numerical and engineering optimization. Computer-
Aided Design 45, 15621574 (2013)
[46] Gong, W., Cai, Z., Liang, D.: Engineering Optimization by means of an improved con-
strained differential evolution. Comput. Methods Appl. Mech. Engrg. 268, 884904
(2014)
[47] Wang, Y., Cai, Z., Qingfu, Z.: Differential evolution with composite trial vector genera-
tion strategies and control parameters. IEEE Trans. Evol. Comput. 15, 5566 (2011)
[48] Jingqiao, Z., Sanderson, A.C.: JADE: adaptive differential evolution with optional exter-
nal archieve. IEEE Trans. Evol. Comput. 13, 945958 (2009)
[49] Iztok, F., Marjan, M., Bogdan, F.: Graph 3-coloring with a hybrid self-adaptive evolutio-
nary algorithm. Computational Optimization and Applications 54(3), 741770 (2013)
A Memetic Dierential Evolution Algorithm
for the Vehicle Routing Problem with Stochastic
Demands
1 Introduction
Storn and Price [38] proposed the population-based evolutionary algorithm de-
noted as Dierential Evolution (DE). Analytical presentation and surveys
can be found in [9,10,36]. Two are the basic steps of a Dierential Evolution
algorithm, but in dierent order than a classic evolutionary algorithm, the mu-
tation operator for a generation of a trial vector and the crossover operator to
produce an ospring.
The rst paper that was devoted to the solution of the Vehicle Routing Prob-
lem was published by Dantzig and Ramser [8]. For complete denition of the
problem and of its basic variants, the reader can nd more details in the follow-
ing papers [4,5,11,17,18,26,25,31,39] and in the books [19,20,34,40]. The main
dierence between the Stochastic Vehicle Routing Problems (SVRPs)
and the Classic Vehicle Routing Problem is that in the Stochastic Vehicle Rout-
ing Problems either the customers, or the customers demands or the customers
service and travel times, or all of them are not determined in the beginning of
the process as it happens in the Classic Vehicle Routing Problem but they are
Corresponding author.
ui (t), if randi (0, 1) Cr
xi (t) = (2)
xi (t), otherwise.
Initialization
Initialize the control parameters , Cr
Select the mutation operator
Select the number of generations
Generate the initial population
Calculate the initial cost function value
(tness function) of each member of the population
Main Phase
Do while maximum number of generations has not been reached
Select the parent vector xi (t)
Create the trial vector ui (t) by applying the mutation operator
Create the ospring xi (t) by applying the crossover operator
Perform VNS algorithm in each ospring
Calculate the cost function (f itness) of the ospring
if f itness(xi (t)) f itness(xi (t))
Replace the parent with the ospring for the next generation
else
MDE Algorithm for the VRPSDs 189
Algorithm VNS
Select the number of the neighbors (Nl , l = 1, , lmax )
Select an initial solution s0
l=1
Main Phase
repeat
Create a solution s in the neighbor of Nl
s = LS(s), Apply a local search phase in s
if f (s ) < f (s ) then
s = s
l=1
else
l =l+1
endif
until l lmax
Return the best solution.
A Hybrid Genetic Algorithm (HGA) is, also, used in order to compare the
results of the proposed algorithm. In the Hybrid Genetic Algorithm, an initial
population of solutions is created and each member of the population is mapped
as in the proposed MDE algorithm. Then, the tness function of each member of
the population is created. In a specic percentage of the individuals, a crossover
phase is applied using the most classic crossover operator, the 1-point crossover.
In a specic percentage of the ospring, a mutation phase is applied. After the
mutation phase, a VNS algorithm is applied in each one of the ospring. In the
next generation, the ttest individual from the whole population survives. With
the term whole population we mean the initial population and the ospring
from both mutation and crossover phases. Thus, the population is sorted based
on the tness function of the individuals and in the next generation the ttest
individuals survive. It should be mentioned that the size of the population of each
generation is equal to the initial size of the population. The stopping criterion
is the maximum number of generations.
In the following, a pseudocode of the hybrid Genetic algorithm is presented.
Initialization
Select the crossover operator
Select the mutation operator
Select the selection operator
Select the percentage of crossover operator
MDE Algorithm for the VRPSDs 191
4 Results
4.1 Parameters Selection
The whole algorithmic approach was implemented in Matlab R2009a. The se-
lected parameters are presented in Table 1. It should be noted that for the
selection of the parameters we used two criteria, the quality of the solution and
the computational time needed to achieve this solution. Thus, we used many
dierent alternative values for the parameters and we made tests with them
and the nal selected parameters were those that gave the best computational
results concerning both criteria mentioned previously. After the selection of the
nal parameters, 10 dierent runs with the selected parameters were performed
for each instance.
MDE HGA
individuals 200 150
iterations 3500 3500
lsiter 100 100
Cr 0.8 0.8
Mr - 0.2
0.5 -
algorithm, the results of the Hybrid Genetic Algorithm described in section 3.4
are, also, given. The reason that we used the Hybrid Genetic Algorithm is that
we would like to compare the proposed algorithm with another evolutionary
algorithm in order to see if the Memetic Dierential Evolution algorithm could
solve eectively a stochastic routing problem. In all tables where the notation
(i.e. quality of the solution) is presented, it means that a best known solution
(BKS) from the literature is known and the quality of the solutions measures the
eciency of the algorithm. The quality is given in terms of the relative deviation
cBKS )
from the best known solution, that is = (cM DE cBKS %, where cMDE denotes
the cost of the solution found by the proposed Memetic DE algorithm and cBKS
is the cost of the BKS solution. Similarly the quality of the solution of the GA
cBKS )
algorithm is given = (cHGA cBKS %, where cHGA denotes the cost of the
solution found by HGA algorithm.
HGA MDE
Instance p BKS cost cost
eil51 0.1 130.12 129.42 -0.54% 129.42 -0.54%
0.5 310.75 316.52 1.86% 312.52 0.57%
0.9 407.92 416.12 2.01% 414.97 1.73%
eil101 0.1 200.03 197.34 -1.35% 197.42 -1.31%
0.5 455.65 464.48 1.94% 464.06 1.85%
0.9 601.5 629.71 4.69% 623.71 3.69%
kroA100 0.1 9074.94 9175.32 1.11% 9051.77 -0.26%
0.5 16581.6 17135.1 3.34% 16569.7 -0.07%
0.9 20508.8 22590.6 10.15% 20511.3 0.01%
ch150 0.1 2510.11 2530.71 0.82% 2508.77 -0.05%
0.5 5016.85 5264.08 4.93% 5245.44 4.56%
0.9 6292.01 6617.5 5.17% 6527.73 3.75%
d198 0.1 7504.94 7770.59 3.54% 7532 0.36%
0.5 12527.6 12965.9 3.50% 12711 1.46%
0.9 15216.6 15857.9 4.21% 15568 2.31%
PTSP instances were generated starting from TSP instances and assigning
to each customer a probability p of requiring a visit. The test instances were
taken from the TSPLIB [43]. The algorithm was tested on a set of 5 Euclidean
benchmark instances with sizes ranging from 51 to 198 nodes. Each instance
is described by its TSPLIB name and size, e.g. in Table 2 the instance named
kroA100 has size equal to 100 nodes. For each PTSP instance tested, various ex-
periments were done by varying the value of the customer probability p. In Table
2, in the last four columns the results of the HGA and the proposed algorithm
(the cost and the quality of the best solution found) for three probability values
(0.1, 0.5 and 0.9) are presented. The best known solutions are taken from [30].
As it can be seen from Table 2, the proposed algorithm, compared to the best
MDE Algorithm for the VRPSDs 193
known solutions published in [30], nds new best solutions in ve out of fteen
instances. However, in one of these instances the HGA nds even better solution.
For the other 10 instances, the quality of the solutions of the proposed algorithm
varies between 0.01% and 4.56%. The HGA, as it is mentioned previously, nds
in one instance a new best solution and in another instance the same new best
solution as the MDE algorithm. In the other thirteen instances, the quality of
the solutions of the HGA varies between 0.82% and 10.15%.
HGA MDE
Instance p cost average stdev var median cost average stdev var median
eil 51 0.1 129.42 130.95 1.22 1.49 131.16 129.42 130.45 0.81 0.66 130.38
0.5 316.52 318.40 0.81 0.66 318.57 312.52 313.98 1.02 1.04 313.89
0.9 416.12 417.62 1.09 1.19 417.57 414.97 415.89 0.82 0.67 415.83
eil101 0.1 197.34 198.90 0.86 0.75 199.05 197.42 198.57 0.87 0.76 198.53
0.5 464.48 466.23 0.92 0.84 466.41 464.06 465.45 0.79 0.63 465.48
0.9 629.71 631.01 0.87 0.75 630.86 623.71 625.19 1.18 1.40 624.77
kroA100 0.1 9175.32 9177.01 0.77 0.59 9177.23 9051.77 9053.18 1.10 1.20 9053.05
0.5 17135.1 17136.27 0.63 0.40 17136.30 16569.7 16571.46 1.17 1.37 16571.93
0.9 22590.6 22591.68 0.85 0.72 22591.49 20511.3 20512.56 0.88 0.77 20512.93
ch150 0.1 2530.71 2532.27 0.92 0.84 2532.21 2508.77 2510.32 0.85 0.72 2510.41
0.5 5264.08 5265.29 1.00 1.01 5265.20 5245.44 5246.96 1.09 1.18 5247.10
0.9 6617.5 6618.47 0.91 0.82 6618.20 6527.73 6528.71 0.80 0.63 6528.61
d198 0.1 7770.59 7771.81 0.78 0.61 7771.59 7532 7533.05 1.08 1.17 7532.59
0.5 12965.9 12966.95 1.06 1.13 12966.49 12711 12711.84 0.73 0.53 12711.70
0.9 15857.9 15859.13 0.94 0.88 15858.92 15568 15569.35 0.91 0.82 15569.43
For the solution of the Vehicle Routing Problem with Stochastic Demands, there
are not commonly used benchmark instances from the researchers in the lit-
erature solving the specic problem. This issue makes the comparison of the
dierent algorithms dicult. Thus, we have tested the algorithm in two set of
benchmark instances, the one was used initially by Christiansen et al. [6]. The set
of benchmark instances contains forty instances with number of nodes between
16 and 60. The second set contains 7 out of 14 benchmark instances proposed
by Christodes [7] that have zero service time. These benchmark instances have,
initially, been proposed and used for the Capacitated Vehicle Routing Prob-
lem, but due to the fact that every variant of the Vehicle Routing Problem is
a generalization of the Capacitated Vehicle Routing Problem, these benchmark
instances have, also, been used in other variants of the Vehicle Routing Problem.
Each instance of the set contains between 51 and 200 nodes including the de-
pot. Each instance includes capacity constraints without maximum route length
restrictions and with zero service time.
Another issue concerning the Vehicle Routing Problem with Stochastic De-
mands is the fact that there are, mainly, two dierent approaches for dealing
with the route failure in this problem. Both approaches have as a goal the min-
imization of the expected cost. In the one approach [6,21], vehicles follow their
assigned routes until a route failure occurs, then, a replenishment of the capacity
is performed at the depot and, nally, a return of the vehicle at the customer
where the route failure occurred and a continuation of the service are performed.
In this approach a set of vehicles can be used. In the other approach, the one
that is, also, used in this chapter, there is a preventive restocking strategy [3,41]
and its main characteristic is that we would like to avoid the route failure. In
order to do that, a threshold value is used. If the residual load after serving a
customer is greater or equal to this value, then, it is better to move to the next
customer, otherwise a return to the depot is performed. In this case only one
vehicle is used. We use the same transformation approach as the one proposed
in [6,21] and, thus, we assume that customers demands are independent Pois-
son random variables with the mean demand for each customer equal to the
deterministic value of the demand given in the corresponding VRP problem. We
tested the proposed algorithm with both approaches used for dealing with the
route failure.
We, also, tested the algorithm using another approach concerning the demands
of the customers. In the beginning, the probability of the demand of each customer
to take a particular value is stored in a variable. This probability depends on the
value of the demands deviation. For example, if the demands deviation is r and
the real demand is R (where r R), the probability of the demand is 2r+1 1
because
the demand can take 2r+1 values (i.e., Rr, R(r1), , R, , R+(r1), R+
r) and the probability of the demand to take one of these values is the same. The
cost from the last node to the depot can be assessed directly as it does not depend
on the customer demand. In the chapter the deviation of the customers demand
takes the values: r = 0, r = 1, r = 2. The rst case (r = 0) denotes that there
MDE Algorithm for the VRPSDs 195
Table 4. Results using the preventive restocking strategy in the VRPSDs in the rst
set of benchmark instances
HGA MDE
Q cost average stdev var median cost average stdev var median
A-n32-k5 100 836.07 837.13 0.62 0.39 837.31 820.50 822.66 0.99 0.98 822.27
A-n33-k5 100 693.40 694.55 0.74 0.55 694.73 684.20 688.47 0.91 0.84 688.83
A-n33-k6 100 762.40 763.77 0.98 0.95 763.75 762.60 770.52 0.64 0.41 770.59
A-n34-k5 100 812.30 813.24 0.77 0.59 813.18 788.70 790.80 0.86 0.73 790.51
A-n36-k5 100 833.30 834.15 0.87 0.77 833.93 835.10 837.10 0.75 0.57 837.12
A-n37-k5 100 707.65 708.12 0.23 0.05 708.22 702.00 694.21 0.81 0.66 694.56
A-n37-k6 100 1018.00 1018.79 0.67 0.45 1018.80 1008.20 1000.69 0.91 0.82 1000.54
A-n38-k5 100 755.50 756.55 0.96 0.93 756.39 752.20 757.49 0.73 0.54 757.43
A-n39-k5 100 858.70 859.96 0.87 0.75 860.14 862.60 854.53 1.06 1.12 854.79
A-n39-k6 100 867.12 868.03 0.93 0.86 867.95 845.70 848.82 0.94 0.89 848.61
A-n44-k6 100 1005.90 1006.78 0.53 0.28 1006.87 980.60 980.26 1.06 1.12 980.45
A-n45-k6 100 1007.90 1009.29 1.07 1.13 1008.88 996.86 998.38 0.83 0.70 998.13
A-n45-k7 100 1239.40 1240.19 0.53 0.28 1240.03 1213.10 1177.10 0.97 0.93 1177.29
A-n46-k7 100 976.84 978.49 0.92 0.84 978.83 979.70 985.74 0.68 0.46 985.76
A-n48-k7 100 1182.30 1183.84 1.09 1.19 1183.67 1146.70 1133.71 1.13 1.27 1134.27
A-n53-k7 100 1117.80 1119.21 0.84 0.71 1119.16 1100.20 1097.43 0.65 0.42 1097.18
A-n54-k7 100 1283.90 1285.05 0.86 0.73 1284.84 1279.50 1224.33 0.81 0.66 1224.21
A-n55-k9 100 1168.10 1168.99 0.73 0.54 1169.02 1150.90 1125.56 0.78 0.61 1125.58
A-n60-k9 100 1517.25 1518.18 0.83 0.69 1518.11 1483.20 1455.12 0.60 0.37 1454.93
E-n22-k4 6000 385.12 385.97 0.86 0.75 385.59 379.16 391.85 0.93 0.86 391.59
E-n33-k4 8000 849.35 849.96 0.29 0.09 850.12 848.25 848.30 0.58 0.34 848.41
E-n51-k5 160 550.15 551.22 0.66 0.43 551.30 549.18 546.00 0.88 0.77 545.89
P-n16-k8 35 443.98 444.81 0.69 0.48 444.56 444.55 456.92 1.11 1.24 457.45
P-n19-k2 160 216.66 217.87 0.96 0.92 217.73 215.04 214.38 0.64 0.41 214.41
P-n20-k2 160 225.89 227.14 0.92 0.84 227.19 224.25 228.11 0.93 0.87 228.24
P-n21-k2 160 218.38 218.62 0.23 0.05 218.52 218.52 218.53 0.28 0.08 218.48
P-n22-k2 160 223.06 224.15 0.92 0.84 223.95 229.45 230.51 0.70 0.49 230.45
P-n22-k8 3000 587.32 588.79 0.89 0.80 588.63 589.89 591.55 0.65 0.42 591.54
P-n23-k8 40 536.07 537.12 0.68 0.46 536.99 545.26 537.32 1.00 1.01 537.14
P-n40-k5 140 471.11 471.83 0.49 0.24 471.82 472.15 472.12 0.50 0.25 472.38
P-n45-k5 150 531.29 532.36 0.91 0.83 532.07 527.90 532.13 1.15 1.33 532.58
P-n50-k10 100 755.15 756.35 1.00 0.99 756.28 724.60 741.38 1.06 1.11 741.83
P-n50-k7 150 580.34 581.20 0.65 0.42 581.13 575.92 572.40 0.96 0.93 572.60
P-n50-k8 120 658.00 658.72 0.64 0.40 658.57 664.02 660.64 0.92 0.85 660.69
P-n51-k10 80 805.80 806.71 0.81 0.65 806.45 789.04 796.41 0.88 0.78 796.17
P-n55-k10 115 742.40 743.55 0.87 0.75 743.39 730.15 738.52 0.48 0.23 738.51
P-n55-k15 70 1002.60 1003.60 1.01 1.03 1003.37 1016.40 1009.59 0.92 0.85 1009.51
P-n55-k7 170 588.34 588.45 0.07 0.00 588.44 588.47 588.37 0.28 0.08 588.55
P-n60-k10 120 803.18 803.81 0.39 0.15 803.90 790.55 773.89 0.70 0.50 773.82
P-n60-k15 80 1068.60 1069.47 0.89 0.79 1069.22 1067.60 1022.63 0.96 0.91 1022.33
is not any deviation from the actual demand R, the second case (r = 1) denotes
that the deviation is equal to 1 while the third case (r = 2) denotes that the
deviation is equal to 2. If for a customer the demand becomes negative, then,
this demand takes the value 0.
196 Y. Marinakis, M. Marinaki, and P. Spanou
Based on the tness function the vehicle either returns to the depot for replen-
ishment or it proceeds to the next customer. In all Tables, the results of the pro-
posed algorithm for the Vehicle Routing Problem with Stochastic Demands are
presented in addition with the results of the Hybrid Genetic Algorithm (HGA)
as in the previous section.
In Table 4, the results of the proposed algorithm in the rst set of benchmark
instances with the preventive restocking strategy are presented. The algorithm
was tested in forty instances, the same with the ones used in [6,21], with number
of nodes from 16 to 60. In the rst two columns of Table 4, the name of the
instance (which includes the number of nodes and the number of vehicles, for
example, the instance A-n32-k5 has 32 nodes and 5 vehicles) and the capacity of
the vehicles are presented, respectively. In columns 3 to 7 and 8 to 12 the results
of the Hybrid Genetic Algorithm (HGA) and of the proposed Memetic Dier-
ential Evolution (MDE) algorithm are presented respectively. More precisely, in
columns 3 and 8 the cost of the best out of 10 runs are presented. In columns
4 and 9 the average cost of the 10 runs, in columns 5 and 10 the standard de-
viation, in columns 6 and 11 the variance and in columns 7 and 12 the median
values for both algorithms are presented, respectively. In the forty instances, the
MDE algorithm gives better results in 27 and the HGA algorithm in 13. The
improvement in the quality of the best run for the MDE algorithm compared
to the HGA in the instances where the MDE gives better results is between
0.09% and 4.04%, while the improvement in the quality for the HGA algorithm
compared to the MDE in the instances where the HGA gives better results is
between 0.02% and 2.86%. In the average of the ten runs, the improvement in
the quality of the solutions for the MDE algorithm compared to the HGA in
the instances where the MDE gives better results is between 0.01% and 5.08%,
while the improvement in the solution for the HGA algorithm compared to the
MDE in the instances where the HGA gives better results is between 0.03% and
2.83%. Both algorithms in all runs give very good results with small dierences
between them as the variance for the MDE varies between 0.08 and 1.33 and the
standard deviation varies between 0.28 and 1.15, while for the HGA the variance
varies between 0.00 and 1.19 and the standard deviation varies between 0.07 and
1.19.
Table 5. Results using the preventive restocking strategy in the VRPSDs in the second
set of benchmark instances
HGA MDE
n Q cost average stdev var median cost average stdev var median
par1 51 160 550.15 551.82 0.88 0.78 551.82 549.18 550.27 0.92 0.84 550.30
par2 76 140 942.2357 943.75 1.03 1.06 944.07 941.54 942.52 1.01 1.01 942.22
par3 101 200 971.15 972.65 1.00 1.00 972.78 969.909 971.06 1.01 1.03 970.87
par4 151 200 1453.5 1454.92 0.90 0.80 1454.78 1418.14 1419.37 1.09 1.19 1419.23
par5 200 200 1975.37 1976.75 1.00 1.00 1976.94 1968.24 1969.35 0.89 0.78 1969.28
par11 121 200 1418.15 1419.52 1.08 1.17 1419.21 1412.11 1413.55 1.05 1.11 1413.53
par12 101 200 998.38 999.87 0.84 0.70 999.88 995.14 996.21 0.77 0.59 996.11
MDE Algorithm for the VRPSDs 197
Table 6. Results using the preventive restocking strategy and the second approach of
the demands deviation in the VRPSDs in the second set of benchmark instances
HGA MDE
Instance r Cost average stdev var median Cost average stdev var median
0 542.62 544.24 1.03 1.06 544.43 537.42 539.11 1.04 1.08 539.52
par1 1 543.8 545.05 1.13 1.28 544.73 538.35 539.97 0.96 0.93 540.23
2 540.59 541.77 1.09 1.19 541.37 536.69 537.89 1.03 1.07 537.40
0 867.86 869.12 0.86 0.74 868.96 862.39 863.98 0.94 0.89 863.83
par2 1 878.45 879.77 1.06 1.12 879.59 862.72 863.97 1.02 1.04 863.71
2 886.87 888.34 1.05 1.10 888.19 865.86 867.13 0.74 0.54 866.99
0 850.66 851.64 0.80 0.64 851.74 839.4 840.69 1.11 1.24 840.63
par3 1 854.51 855.97 0.88 0.77 856.02 866.15 867.45 0.71 0.51 867.53
2 854.18 855.74 1.01 1.02 855.68 851.52 853.11 0.99 0.98 853.21
0 1137.24 1138.84 1.12 1.27 1139.41 1095.18 1097.04 1.05 1.10 1097.42
par4 1 1128.21 1129.42 0.77 0.59 1129.30 1115.37 1116.50 1.23 1.51 1115.78
2 1141.18 1142.41 0.87 0.76 1142.45 1139 1140.27 0.71 0.51 1140.42
0 1511.21 1513.02 1.01 1.03 1513.09 1495.28 1496.52 1.09 1.19 1496.19
par5 1 1505.37 1506.80 0.77 0.59 1506.91 1499.17 1500.41 0.96 0.93 1500.61
2 1497.84 1499.59 1.10 1.22 1499.92 1489.5 1490.86 1.06 1.12 1490.91
0 1061.47 1062.92 1.22 1.49 1062.94 1055.87 1057.20 0.88 0.78 1057.16
par11 1 1085.49 1086.73 0.82 0.67 1086.78 1088.74 1089.91 0.92 0.85 1089.59
2 1117.11 1118.30 0.97 0.93 1118.29 1098.7 1100.21 1.00 1.00 1100.64
0 835.18 836.83 1.03 1.06 837.15 823.47 825.21 0.86 0.73 825.66
par12 1 862.49 863.79 1.07 1.15 863.74 859.79 861.43 1.03 1.06 861.46
2 864.46 865.73 0.87 0.75 865.57 861.28 862.52 1.07 1.15 862.45
198 Y. Marinakis, M. Marinaki, and P. Spanou
In Table 6, the results in the second set of benchmark instances using the sec-
ond approach of the demands deviation are presented. In the rst two columns
of Table 6, the name of the instance and the three dierent customers demands
deviation (r = 0, r = 1, r = 2) for each instance are presented respectively. In
columns 3 to 7 and 8 to 12 the results (cost, average, standard deviation, vari-
ance and median) of the Hybrid Genetic Algorithm (HGA) and of the proposed
Memetic Dierential Evolution (MDE) algorithm are presented, respectively. In
this case, using the three dierent deviations for the seven instances, we have
21 instances. In these instances, the MDE algorithm gives better results in 19
instances and the HGA algorithm in 2 instances. The improvement in the quality
of the best run for the MDE algorithm compared to the HGA in the instances
where the DE gives better results is between 0.19% and 3.69%, while the im-
provement in the quality for the HGA algorithm compared to the MDE in the
two instances where the HGA gives better results is 0.29% and 1.36%, respec-
tively. In the average of the ten runs, the improvement in the quality of the
solutions for the MDE algorithm compared to the HGA in the instances where
the MDE gives better results is between 0.18% and 3.66%, while the improve-
ment in the solution for the HGA algorithm compared to the MDE in the two
instances where the HGA gives better results is 0.23% and 1.34%, respectively.
Both algorithms in all runs give very good results with small dierences between
them as the variance for the MDE varies between 0.51 and 1.51 and the standard
deviation varies 0.71 and 1.23, while for the HGA the variance varies between
0.59 and 1.49 and the standard deviation varies between 0.77 and 1.22.
In Table 7, the results of the proposed algorithm and the results from the
literature without the preventing restocking strategy are presented. In the rst
column of Table 7, the name of the instance is presented. Column 2 presents
the BKS solution without the preventing restocking strategy, while columns 3
to 10 present the results of the Christiansen and Lysgaard (CL) [6] (columns 3
and 4, cost and quality of the solutions, respectively), the results of Goodson et
al. (G) [21] (columns 5 and 6, cost and quality of the solutions, respectively),
the results of the hybrid genetic algorithm (HGA) (columns 7 and 8, cost and
quality of the solutions, respectively) and the results of the proposed algorithm
(MDE) (columns 9 and 10, cost and quality of the solutions, respectively).
The reason that we tested the proposed algorithm without the preventing re-
stocking strategy is that in paper [6] the optimal solutions have been calculated
with two dierent branching strategies in some of the instances (19 out of 40
instances, denoted with bold face letters in column 2 of Table 7), and it would be
very interesting to see how the proposed algorithm performs when the optimal
values are known. As it can be seen, in the 19 instances that the optimal values
are known, the proposed algorithm succeeded to nd them in 13 and the HGA
in 8. In total, the proposed algorithm nds the best solution in 13 and the HGA
in 8 while for the other instances the quality of the solutions in the proposed al-
gorithm varies between 0.02% and 0.80% and for the HGA varies between 0.04%
and 1.26%. The average quality in all 40 instances for the proposed algorithm is
equal to 0.21% and for the HGA is equal to 0.38%.
MDE Algorithm for the VRPSDs 199
Table 7. Results without using the preventive restocking strategy in the VRPSDs in
the rst set of benchmark instances (Part A)
CL G HGA MDE
BKS cost cost cost cost
A-n32-k5 853.6 853.6 0.00 853.6 0.00 855.1 0.18 853.6 0.00
A-n33-k5 704.2 704.2 0.00 704.2 0.00 705.8 0.23 704.2 0.00
A-n33-k6 793.9 793.9 0.00 793.9 0.00 794.8 0.11 794.1 0.03
A-n34-k5 826.87 827.87 0.12 826.87 0.00 828.28 0.17 827.95 0.13
A-n36-k5 858.71 - - 858.71 0.00 861.15 0.28 860.28 0.18
A-n37-k5 708.34 708.34 0.00 708.34 0.00 709.25 0.13 709.12 0.11
A-n37-k6 1030.73 1030.75 0.00 1030.73 0.00 1031.18 0.04 1030.95 0.02
A-n38-k5 775.14 778.09 0.38 775.14 0.00 781.25 0.79 778.29 0.41
A-n39-k5 869.18 869.18 0.00 869.18 0.00 872.27 0.36 869.18 0.00
A-n39-k6 876.6 876.6 0.00 876.6 0.00 877.75 0.13 876.98 0.04
A-n44-k6 1025.48 1025.48 0.00 1025.48 0.00 1027.19 0.17 1026.85 0.13
A-n45-k6 1026.73 - - 1026.73 0.00 1029.28 0.25 1028.35 0.16
A-n45-k7 1264.83 1264.83 0.00 1264.99 0.01 1267.98 0.25 1266.15 0.10
A-n46-k7 1002.22 1002.41 0.02 1002.22 0.00 1004.58 0.24 1003.95 0.17
A-n48-k7 1187.14 - - 1187.14 0.00 1191.68 0.38 1190.57 0.29
A-n53-k7 1124.27 - - 1124.27 0.00 1127.59 0.30 1126.87 0.23
A-n54-k7 1287.07 - - 1287.07 0.00 1292.58 0.43 1290.74 0.29
A-n55-k9 1179.11 - - 1179.11 0.00 1191.24 1.03 1185.57 0.55
A-n60-k9 1529.82 - - 1529.82 0.00 1542.98 0.86 1535.24 0.35
E-n22-k4 411.57 411.57 0.00 411.57 0.00 411.57 0.00 411.57 0.00
E-n33-k4 850.27 850.27 0.00 850.27 0.00 852.14 0.22 850.27 0.00
E-n51-k5 552.26 - - 552.26 0.00 559.24 1.26 555.84 0.65
P-n16-k8 512.82 512.82 0.00 512.82 0.00 512.82 0.00 512.82 0.00
P-n19-k2 224.06 224.06 0.00 224.06 0.00 224.06 0.00 224.06 0.00
P-n20-k2 233.05 233.05 0.00 233.05 0.00 233.05 0.00 233.05 0.00
P-n21-k2 218.96 218.96 0.00 218.96 0.00 218.96 0.00 218.96 0.00
P-n22-k2 231.26 231.26 0.00 231.26 0.00 231.26 0.00 231.26 0.00
P-n22-k8 681.06 681.06 0.00 681.06 0.00 681.06 0.00 681.06 0.00
P-n23-k8 619.52 619.52 0.00 619.53 0.00 619.52 0.00 619.52 0.00
P-n40-k5 472.5 472.5 0.00 472.5 0.00 473.8 0.28 472.5 0.00
P-n45-k5 533.52 - - 533.52 0.00 535.28 0.33 534.12 0.11
P-n50-k10 760.94 - - 760.94 0.00 764.18 0.43 762.14 0.16
P-n50-k7 582.37 - - 582.37 0.00 586.47 0.70 584.15 0.31
P-n50-k8 669.81 - - 669.81 0.00 674.41 0.69 673.18 0.50
P-n51-k10 809.7 809.7 0.00 812.74 0.38 817.28 0.94 816.25 0.81
P-n55-k10 745.7 - - 745.7 0.00 751.24 0.74 749.18 0.47
P-n55-k15 1068.05 1068.05 0.00 1068.05 0.00 1077.18 0.85 1074.24 0.58
P-n55-k7 588.56 - - 588.56 0.00 595.14 1.12 593.21 0.79
P-n60-k10 804.24 - - 804.24 0.00 810.41 0.77 808.57 0.54
P-n60-k15 1085.49 1085.49 0.00 1087.41 0.18 1094.71 0.85 1092.14 0.61
200 Y. Marinakis, M. Marinaki, and P. Spanou
Table 8. Results without using the preventive restocking strategy in the VRPSDs in
the rst set of benchmark instances (Part B)
HGA MDE
cost average stdev var median cost average stdev var median
A-n32-k5 855.1 856.61 0.83 0.68 856.76 853.6 854.53 0.69 0.47 854.52
A-n33-k5 705.8 707.18 0.99 0.98 707.35 704.2 705.58 0.76 0.58 705.68
A-n33-k6 794.8 796.19 0.96 0.93 796.13 794.1 795.12 0.74 0.54 795.01
A-n34-k5 828.28 829.66 1.12 1.26 829.88 827.95 829.04 0.69 0.48 829.28
A-n36-k5 861.15 862.71 1.09 1.20 862.84 860.28 861.78 0.78 0.61 861.93
A-n37-k5 709.25 710.85 0.88 0.78 710.85 709.12 710.46 0.84 0.71 710.54
A-n37-k6 1031.18 1032.85 0.99 0.98 1033.15 1030.95 1032.07 0.97 0.93 1031.85
A-n38-k5 781.25 782.42 0.90 0.82 782.13 778.29 779.61 0.82 0.68 779.96
A-n39-k5 872.27 873.92 0.82 0.67 873.80 869.18 870.18 0.92 0.84 870.11
A-n39-k6 877.75 879.10 0.85 0.72 879.09 876.98 878.14 0.75 0.56 878.05
A-n44-k6 1027.19 1028.56 1.02 1.05 1028.37 1026.85 1028.29 0.78 0.60 1028.43
A-n45-k6 1029.28 1030.89 1.01 1.02 1030.99 1028.35 1029.69 0.89 0.79 1029.91
A-n45-k7 1267.98 1269.06 1.03 1.07 1268.67 1266.15 1266.91 0.63 0.40 1266.81
A-n46-k7 1004.58 1005.94 1.05 1.10 1005.96 1003.95 1004.98 0.75 0.56 1004.95
A-n48-k7 1191.68 1192.51 0.71 0.51 1192.45 1190.57 1191.47 0.41 0.17 1191.52
A-n53-k7 1127.59 1128.80 0.76 0.57 1128.88 1126.87 1128.05 0.87 0.75 1127.85
A-n54-k7 1292.58 1294.05 1.03 1.07 1293.96 1290.74 1292.03 0.78 0.62 1292.25
A-n55-k9 1191.24 1192.65 1.05 1.10 1192.83 1185.57 1187.01 0.73 0.53 1187.11
A-n60-k9 1542.98 1544.56 0.94 0.88 1544.60 1535.24 1536.72 0.90 0.81 1536.95
E-n22-k4 411.57 412.86 0.91 0.83 412.88 411.57 412.32 0.57 0.32 412.21
E-n33-k4 852.14 853.89 1.05 1.10 854.14 850.27 851.51 0.97 0.95 851.80
E-n51-k5 559.24 560.86 1.00 1.01 560.77 555.84 557.12 0.85 0.73 557.42
P-n16-k8 512.82 514.22 0.91 0.83 514.24 512.82 512.82 0.00 0.00 512.82
P-n19-k2 224.06 225.38 1.00 1.01 225.20 224.06 224.06 0.00 0.00 224.06
P-n20-k2 233.05 234.30 1.05 1.11 234.36 233.05 233.11 0.04 0.00 233.10
P-n21-k2 218.96 220.19 0.82 0.68 220.12 218.96 218.96 0.00 0.00 218.96
P-n22-k2 231.26 232.48 1.00 1.00 232.54 231.26 231.31 0.03 0.00 231.32
P-n22-k8 681.06 682.59 0.99 0.97 682.92 681.06 681.06 0.00 0.00 681.06
P-n23-k8 619.52 621.10 0.93 0.87 621.29 619.52 619.52 0.00 0.00 619.52
P-n40-k5 473.8 475.17 0.89 0.79 474.97 472.5 473.48 0.63 0.40 473.53
P-n45-k5 535.28 537.11 0.93 0.86 537.25 534.12 535.43 0.75 0.56 535.52
P-n50-k10 764.18 765.40 0.91 0.83 765.51 762.14 763.37 0.80 0.63 763.41
P-n50-k7 586.47 587.41 0.93 0.86 587.06 584.15 584.73 0.49 0.24 584.61
P-n50-k8 674.41 675.57 0.94 0.89 675.53 673.18 674.18 0.72 0.52 674.09
P-n51-k10 817.28 818.45 1.16 1.35 817.99 816.25 817.51 0.87 0.75 817.41
P-n55-k10 751.24 752.50 0.97 0.95 752.41 749.18 750.67 0.86 0.74 751.10
P-n55-k15 1077.18 1078.36 1.00 0.99 1078.16 1074.24 1075.62 0.77 0.60 1075.66
P-n55-k7 595.14 595.98 0.99 0.98 595.61 593.21 594.17 0.54 0.29 594.25
P-n60-k10 810.41 811.48 0.97 0.94 811.20 808.57 809.89 0.81 0.65 809.87
P-n60-k15 1094.71 1095.91 0.87 0.76 1095.87 1092.14 1093.59 0.96 0.92 1093.68
MDE Algorithm for the VRPSDs 201
we tested, initially, the algorithm using the same instances as in the previous
case using two dierent approaches for the stochastic demands and, then, we
applied the algorithm in a second set of instances, those that are used for the
Capacitated Vehicle Routing Problem. In the second strategy, we compared the
algorithm only with the hybrid genetic algorithm and we presented and analyzed
the results. Future research will be focused in two dierent directions, the one
direction will be the solution of more dicult problems as the Vehicle Routing
Problem with Stochastic Demands and Customers where both demands and cus-
tomers are stochastic variables or the Vehicle Routing Problem with Stochastic
Demands and Time Windows and the other direction will be the application of
dierent evolutionary algorithms in these problems.
References
1. Bertsimas, D.J.: Probabilistic Combinatorial Optimization Problems. Ph.D. thesis.
MIT, Cambridge, MA, USA (1988)
2. Bianchi, L.: Ant Colony Optimization and Local Search for the Probabilistic Trav-
eling Salesman Problem: A Case Study in Stochastic Combinatorial Optimization.
Ph.D. Thesis. Universite Libre de Bruxelles, Belgium (2006)
3. Bianchi, L., Birattari, M., Manfrin, M., Mastrolilli, M., Paquete, L., Rossi-Doria,
O., Schiavinotto, T.: Hybrid metaheuristics for the vehicle routing problem with
stochastic demands. Journal of Mathematical Modelling and Algorithms 5(1), 91
110 (2006)
4. Bodin, L., Golden, B.: Classication in vehicle routing and scheduling. Networks 11,
97108 (1981)
5. Bodin, L., Golden, B., Assad, A., Ball, M.: The state of the art in the routing and
scheduling of vehicles and crews. Computers and Operations Research 10, 63212
(1983)
6. Christiansen, C.H., Lysgaard, J.: A branch-and-price algorithm for the capacitated
vehicle routing problem with stochastic demands. Operations Research Letters 35,
773781 (2007)
7. Christodes, N., Mingozzi, A., Toth, P.: The vehicle routing problem. In:
Christodes, N., Mingozzi, A., Toth, P., Sandi, C. (eds.) Combinatorial Optimiza-
tion. John Wiley, Chichester (1979)
8. Dantzig, G.B., Ramser, J.H.: The truck dispatching problem. Management Sci-
ence 6(1), 8091 (1959)
9. Engelbrecht, A.P.: Computational Intelligence: An Introduction. John Wiley and
Sons, Chichester (2007)
10. Feoktistov, V.: Dierential Evolution - In Search of Solutions. Springer, NY (2006)
11. Fisher, M.L.: Vehicle routing. In: Ball, M.O., Magnanti, T.L., Momma, C.L.,
Nemhauser, G.L. (eds.) Network Routing, Handbooks in Operations Research and
Management Science, vol. 8, pp. 133. North Holland, Amsterdam (1995)
12. Fister, I., Mernik, M., Filipic, B.: A hybrid self-adaptive evolutionary algorithm
for marker optimization in the clothing industry. Applied Soft Computing 10(2),
409422 (2010)
13. Fister, I., Fister Jr., I., Brest, J., Zumer, V.: Memetic articial bee colony algo-
rithm for large-scale global optimization. In: 2012 IEEE Congress on Evolutionary
Computation (CEC). IEEE (2012)
MDE Algorithm for the VRPSDs 203
14. Fister Jr, I., Fister, D., Yang, X.S.: A Hybrid Bat Algorithm. Electrotechnical
Review 80(1-2), 17 (2013)
15. Fister, I., Mernik, M., Filipic, B.: Graph 3-coloring with a hybrid self-adaptive
evolutionary algorithm. Computational Optimization and Applications 54(3), 741
770 (2013)
16. Gendreau, M., Laport, G., Seguin, R.: Stochastic vehicle routing. European Journal
of Operational Research 88, 312 (1996)
17. Gendreau, M., Laporte, G., Potvin, J.Y.: Vehicle routing: modern heuristics. In:
Aarts, E.H.L., Lenstra, J.K. (eds.) Local Search in Combinatorial Optimization,
pp. 311336. Wiley, Chichester (1997)
18. Gendreau, M., Laporte, G., Potvin, J.Y.: Metaheuristics for the capacitated VRP.
In: Toth, P., Vigo, D. (eds.) The Vehicle Routing Problem, Monographs on Discrete
Mathematics and Applications, pp. 129154. Siam, Philadelphia (2002)
19. Golden, B.L., Assad, A.A.: Vehicle Routing: Methods and Studies. North Holland,
Amsterdam (1988)
20. Golden, B.L., Raghavan, S., Wasil, E. (eds.): The Vehicle Routing Problem: Latest
Advances and New Challenges. Springer, NY (2008)
21. Goodson, J.C., Ohlmann, J.W., Thomas, B.W.: Cyclic-order neighborhoods with
application to the vehicle routing problem with stochastic demand. European Jour-
nal of Operational Research 217, 312323 (2012)
22. Hansen, P., Mladenovic, N.: Variable neighborhood search: Principles and applica-
tions. European Journal of Operational Research 130, 449467 (2001)
23. Jaillet, P.: Probabilistic Traveling Salesman Problems. Ph.D. Thesis, MIT, Cam-
bridge, MA, USA (1985)
24. Jaillet, P.: A priori solution of a traveling salesman problem in which a random
subset of the customers are visited. Operations Research 36(6), 929936 (1988)
25. Laporte, G., Semet, F.: Classical heuristics for the capacitated VRP. In: Toth, P.,
Vigo, D. (eds.) The Vehicle Routing Problem, Monographs on Discrete Mathemat-
ics and Applications, pp. 109128. Siam, Philadelphia (2002)
26. Laporte, G., Gendreau, M., Potvin, J.Y., Semet, F.: Classical and modern heuris-
tics for the vehicle routing problem. International Transactions on Operations Re-
search 7, 285300 (2000)
27. Lawer, E.L., Lenstra, J.K., Rinnoy Kan, A.H.G., Shmoys, D.B.: The Traveling
Salesman Problem: A Guided Tour of Combinatorial Optimization. Wiley and
Sons (1985)
28. Lichtblau, D.: Discrete optimization using Mathematica. In: Callaos, N., Ebisuzaki,
T., Starr, B., Abe, J.M., Lichtblau, D. (eds.) World Multi-Conference on Systemics,
Cybernetics and Informatics (SCI 2002). International Institute of Informatics and
Systemics, vol. 16, pp. 169174 (2002)
29. Liu, Y.-H.: A hybrid scatter search for the probabilistic traveling salesman problem.
Computers and Operations Research 34(10), 29492963 (2007)
30. Marinakis, Y., Marinaki, M.: A hybrid multi-swarm particle swarm optimization
algorithm for the probabilistic traveling salesman problem. Computers and Oper-
ations Research 37, 432442 (2010)
31. Marinakis, Y., Migdalas, A.: Heuristic solutions of vehicle routing problems in
supply chain management. In: Pardalos, P.M., Migdalas, A., Burkard, R. (eds.)
Combinatorial and Global Optimization, pp. 205236. World Scientic Publishing
Co. (2002)
32. Marinakis, Y., Iordanidou, G.R., Marinaki, M.: Particle Swarm Optimization for
the Vehicle Routing Problem with Stochastic Demands. Applied Soft Comput-
ing 13, 16931704 (2013)
204 Y. Marinakis, M. Marinaki, and P. Spanou
33. Moscato, P., Cotta, C.: A gentle introduction to memetic algorithms. In: Glover,
F., Kochenberger, G.A. (eds.) Handbooks of Metaheuristics, pp. 105144. Kluwer
Academic Publishers, Dordrecht (2003)
34. Pereira, F.B., Tavares, J.: Bio-inspired algorithms for the vehicle routing problem.
SCI, vol. 161. Springer, Heidelberg (2008)
35. Powell, W.B., Jaillet, P., Odoni, A.: Stochastic and dynamic networks and routing.
In: Ball, M.O., Magnanti, T.L., Momma, C.L., Nemhauser, G.L. (eds.) Network
Routing, Handbooks in Operations Research and Management Science, vol. 8, pp.
141295. Elsevier Science B. V., Amsterdam (1995)
36. Price, K.V., Storn, R.M., Lampinen, J.A.: Dierential Evolution: A Practical Ap-
proach to Global Optimization. Springer, Berlin (2005)
37. Stewart, W.R., Golden, B.L.: Stochastic vehicle routing: A comprehensive ap-
proach. European Journal of Operational Research 14, 371385 (1983)
38. Storn, R., Price, K.: Dierential evolution - A simple and ecient heuristic for
global optimization over continuous spaces. Journal of Global Optimization 11(4),
341359 (1997)
39. Tarantilis, C.D.: Solving the vehicle routing problem with adaptive memory pro-
gramming methodology. Computers and Operations Research 32, 23092327 (2005)
40. Toth, P., Vigo, D.: The Vehicle Routing Problem. Monographs on Discrete Math-
ematics and Applications. Siam, Philadelphia (2002)
41. Yang, W.H., Mathur, K., Ballou, R.H.: Stochastic vehicle routing problem with
restocking. Transportation Science 34, 99112 (2000)
42. http://www.coin-or.org/SYMPHONY/branchandcut/VRP/data/Vrp-All.tgz
43. http://www.iwr.uni-heidelberg.de/groups/comopt/software/TSPLIB95
Modeling Nanorobot Control Using Swarm Intelligence
for Blood Vessel Repair: A Rigid-Tube Model
Abstract. The future nanorobots for diagnosis and treatment purposes in nano-
medicine may exhibit only simple behaviors and work together in their early
stage. Through exploring the existing swarm intelligence techniques, the canoni-
cal particle swarm optimization was selected to employ for adaptively controlling
the locomotion of a swarm system of early-stage nanorobots with only essential
characteristics for self-assembly into a structure in a simulations system. In this
study, we demonstrated nanorobots operating as artificial platelets for repairing
wounds in a simulated human small vessel, which may be used to treat platelet
diseases. In a rigid-tube model, we investigated how artificial platelet capabili-
ties including the perception range, maximum velocity and response speed have
impacts on wound healing effectiveness. It was found that the canonical parti-
cle swarm optimization is an efficient algorithm for controlling the early-stage
nanorobots with essential characteristics in both Newtonian and non-Newtonian
flow models. The demonstration could be beneficial as guidelines of essential
characteristics and useful locomotion control for the realization of nanorobots
for medical applications in the future.
1 Introduction
Even though the advancement of medical technology provides more effective and effi-
cient diagnosis, monitoring and treatment, there still exist some diseases that are difficult
to diagnose at their early stages and some risks/side effects from some treatments. For
example, it is truly hard to diagnose a cancer in the early stage; the traditional treatment
of radiation also damage to healthy cells near the cancer cells. Robots have been intro-
duced and used in medical applications, especially surgery which may provide fewer
risks than open and laparoscopic surgery. Although the current development of medi-
cal robots has not yet allow us to perform operation at cellular or molecular scale that
may dispel the disease at the source without harmful to healthy cells, it is anticipated
that nano-scale robots or nanorobotsa concept arising from the advance in nanotech-
nology introduced by Richard P. Feynman in 1959 [15]may be used to improve the
efficiency of medical technology. As Drexler introduced an idea of cooperative small
Corresponding author.
robots or nanorobots that could manipulate substances inside the human blood vessels
[12], these nanorobots may allow us to cure the disease by delivering drugs to specific
positions which could reduce the damage to normal cells as well as other side effects.
Although nanorobots have not yet been realized, some of nanotechnology research
and development might lead to the realization of nanorobots or nanomachines for medi-
cal and other applications in the near future. For instance, catalytic nano-wire motors [42]
could be a great way for constructing self-powered practical nanomachines and could
provide the building block to realize the future nanorobots. The artificial bacterial flag-
ella [45] that control their movement by a low-strength rotating magnetic field may have
potential to be used as a part of future nanorobots for controlling the movement of fu-
ture nanorobots. Apart from the structure and the suitable materials of the nanorobots,
the control systems of nanostructures according to the concept of nanotechnology are
also considered. Nowadays, there are many design concepts of medical nanorobots. For
example, Freitas designed medical nanorobots such as respirocytes [17] or artificial me-
chanical red blood cells, microbivores [33] or artificial mechanical white blood cells, and
clottocytes [32] or artificial platelets. In addition, the simulation of nanorobots will be
beneficial to identify the essential characteristics as well as vital functions of nanorobots
and to investigate the effectiveness of control techniques for nanorobots to achieve their
tasks; the findings can serve as guidelines for developing nanorobots in the future [21].
The examples of the simulation of nanorobots include the simulation of nanorobots trans-
porting nutrition to organ inlets developed by Cavalcanti and Freitas [5], the simulation
of DNA nanorobots for identifying the cancer cell modeled by Sanchita [34], the simu-
lation of the swarm of an early version of future nanorobots for self-assembly and self-
repair application modeled Kaewkamnerdpong and Bentley [22], and the simulation of
nanorobots traveling through blood vessel to detect the aneurysm in the brain by Cav-
alcanti et al. [4]. These studies modeled how the nanorobots would be in the future and
envisaged the way to allow the nanorobots to effectively operate. The results from these
studies could provide suggestions for the development of future nanorobots. In spite of
that, the simulation of nanorobots should be as realistic as possible so that the findings
can be truly beneficial for the realization of future nanorobots.
Based on the concept of bottom-up technology in nanotechnology, nano parts such
as molecular motor and nano swimmer have been constructed [36]. Nevertheless, the
designed nanorobots in most works are quite advanced and require complex nano parts.
With the current development in nanotechnology, complex nano parts are not yet ac-
complished, and such advanced nanorobots may not be available in recent time. Due
to the very small size of nanorobots, it is more reasonable that the early version of
nanorobots may be able to exhibit only simple behaviors. They may be able to move
around environment, interact with the individuals in the group and interact with their
environment; the external control over the nanorobots may not be available. With these
limitation, he control mechanism of nanorobots should be suitably designed according
to these potential behaviors and characteristics so that they could achieve their des-
ignated tasks under dynamic environment. Such nanorobot context, which includes
simple capabilities and no external control, is similar to social insects. The individ-
ual behaviors of social animal such as ants and termites are usually simple but their
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 207
collaborative behaviors enable them to achieve their complex tasks for survival. The col-
lective intelligence of social animal, or called swarm intelligence, is modeled and used
as problem solving techniques in various applications. Hence, the swarm intelligence
may be reasonable to control the locomotion of a swarm of nanorobots with adaptation
to effectively accomplish designated tasks even without complex characteristics. The
objective of this study was to investigate the plausibility of controlling swarm system
of nanorobots using a swarm intelligence technique for self-assembly tasks in medical
applications.
In this study, the performance of a swarm-intelligence-based system of nanorobots
was investigated as artificial platelets in blood vessel repair application. The reasons
for selecting the blood vessel repair application are that the artificial platelet model is
one of the existing designed nanorobot models, called clottocytes, in [32] and that the
primary hemostasis is an example of self-assembly in nature which is simple enough
for the early-stage nanorobots to achieve their task. In the primary hemostasis, platelets
play the most important role involving platelet adhesion and the platelet aggregation
to stop the bleeding of a wound. However, there are some patients who have abnormal
platelets that cause defects in the primary hemostasis. Additionally, the treatments of
platelet diseases also have side effects. Hence, it can be anticipated that the nanorobots
with biocompatibility that operate as artificial platelets may be used to treat patients
who have platelet diseases such as thrombocytopenias. In this study, we demonstrated
the performance of swarm-intelligence-based nanorobot system operating as artificial
platelets to repair an injured blood vessel wall in a rigid-tube model. The flow models
used in this study were obtained from those existing in the literature. The modification
for better flow models was not in the scope of this study.
This chapter is organized as follows: Section 2 describes the swarm intelligence tech-
niques and discusses on how the techniques could be appropriate to employ for control-
ling nanorobots. The simulated model for nanorobots in blood vessel repair is described
in section 3. This section includes both the model of nanotobots and their environment
in the blood vessel. Section 4 demonstrate the performance of the nanorobot system
in blood vessel repair application. The study is concluded with the analysis on control
mechanism toward the realization of future nanorobots in section 5.
optima or good enough solution. Thus, the mechanism for finding a good solution is
influenced by both their own experience and their neighbors experiences. PSO has very
few parameters to adjust; in addition, it uses less memory. Hence, PSO seems plausible
to apply in nanorobot locomotion control or other physical applications. Nevertheless,
the conventional PSO algorithm requires complex, direct communication among parti-
cles in the neighborhood to exchange detailed information on their performance which
might not be possible by real nanorobots.
There was an attempt on applying PSO-based algorithm in nanorobot locomotion
control; Kaewkamnerdpong and Bentley [22] employed the perceptive particle swarm
optimization (PPSO) [21], which is a modification of the conventional PSO, to control
the system of swarm nanorobots in surface coating application. The nanorobots must
self-assemble into a nanostructure on the desired surface. In [22], the nanorobot model
was based on an early-version of nanorobots whose characteristics and behaviors are
simple due to the limitation from their size; no direct communication with informa-
tion exchange may be available. To substitute for the lack of direct communication,
it is assumed that each particle or nanorobot has a sensory unit to sense its neighbors
and evaluate its performance in the search space for interacting with other individu-
als and its environment. PPSO takes this performance of particles at their positions in
n-dimensional space as another information in the control; hence, PPSO operates in
(n + 1)-dimensional space. By adding the additional dimension and the ability to ob-
serve the search space, it was expected that these could allow particles to perceive their
approximate performance and to reasonably perform their collaborative tasks. It was
found that although PPSO could be used to control nanorobots to achieve their task,
the additional dimension seemed unnecessary in dynamic environment. For the med-
ical nanorobots that potentially operate in an extremely dynamic environment like a
circulatory system, it may not be necessary to keep the signified fitness of their current
positions because their positions would be undetectably altered by the blood flow and,
in turn, their collected performance would be misleading the locomotion control.
In this study, we adopted the concept of early version nanorobot model but applied
the canonical PSO algorithm to control the swarm of medical nanorobots. The canonical
PSO algorithm is the constricted version of PSO proposed by Clerc and Kennedy [9].
The constriction coefficient, , is used to control the exploration versus exploitation
trade-off in velocity equation to ensure the convergence of PSO [9]. Although over the
years there are various modification of PSO proposed, the canonical PSO algorithm
was selected in this study for the reason that it is simple as analogous to early-stage
nanorobots and requires only a few parameters to be adjusted which are corresponding
to plausible characteristics of early-stage nanorobots.
platelet transfusion is well known. Platelets separated from donated blood are injected
into patients. After injection, platelets will circulate in the body for ten days in aver-
age before destroyed through bodys natural mechanism. Thus, patients usually need
platelet transfusion twice or thrice in a month. However, frequent platelet transfusion
might induce resistance to platelets from others. For patients whose platelets are de-
stroyed by their own immune systems, Immunosuppressant is used but with some side
effects as well. Another futuristic idea is using nanorobots as artificial platelets to heal
the wound from inside the body. Toward this idea, this study explored the control mech-
anism for nanorobots operating as artificial platelets to repair blood vessel.
The platelets or thrombocytes are originated from cytoplasts of megakaryocytes,
which are the largest cells in bone-marrow. The cytoplasts have many small pseudopods
that will slip off and become platelets with 2-4 m in diameter. When platelets are in
their inactivated states, they are discoid in shape but become spheroid when activated
[6]. They play an important role in hemostasis, which is the response mechanism of
human body to stop bleeding. When a blood vessel is injured, the vessel constricts to
decrease the vessel lumen in order to slow the bleeding. In hemostasis, this step can
be referred as vasoconstriction. At the opened wound, blood is exposed to the col-
lagen fibers underlying the endothelium in blood vessels. Platelets become activated
after prolonged exposure to high shear stress or when the shear stress rapidly increases
such as when vasoconstriction occurs at damaged vessel [39]. Through glycoprotein
Ia/IIa receptors on their membrane, platelets adhere to the exposed collagen and von
Willebrand factor (vWF) in the vessel wall [18]. Then, the adherent platelets release
adenosine diphosphate as well as thromboxane A2, which induces additional platelets
to become activated and adhere [1]. Prostacylin, the substance released from the normal
endothelial cells in the blood vessel adjacent to the injured area, plays a role to prevent
the aggregation of platelets along the length of a normal vessel [1]. As thromboxane A2
promotes platelet aggregation whereas prostacyclin inhibits platelet aggregation, the
balance between platelet thromboxane A2 and prostacyclin is required to perform lo-
calized platelet aggregation while preventing excessive clot and maintaining blood flow
around it [18]. The adherent platelets at the site of injury form a temporary hemostatic
plug [18]; the mechanism to form a platelet plug can be referred as primary hemostasis.
In secondary hemostasis, the hemostatic plug is, then, converted into the definitive clot.
Platelets release clotting factors to convert fibrinogena protein found in the plasma
into a dense network of fibrin stands in which blood cells and plasma are trapped [1,3].
In hemostasis, platelets take a crucial role in primary hemostasis to form a platelet
plug. The primary hemostasis is an example of natural self-assembly tasks; hence, it
is appropriate to be used for investigating the model of early-stage nanorobots. The
nanorobots acting as artificial platelets will need to contain all necessary granules to
allow usual primary hemostasis and to promote secondary hemostasis in order to com-
pletely heal the wound. While moving along blood vessels, the artificial platelets must
be able to detect whether these is any sign of a wound nearby and become activated
when the sign is discovered. Then, the artificial platelets must work together to prevent
further blood loss and accommodate to the secondary hemostasis.
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 211
carrying a folded fiber mesh which becomes sticky when the coated substance
comes in contact with plasma,
sensing its environment to detect the injury,
communicating with other clottocytes to activate them with acoustic pulses when
injury is detected, and
unleashing the mesh at the site of an injured blood vessel.
Instead of aggregating into a platelet plug, clottocytes are designed to use the artificial
nets to trap blood cells and accelerate the clotting process. According to the study of
Freitas in [32], clottocytes could stop the bleeding greatly faster than natural process.
This would be a prominent contribution for people who have problems with hemostasis
process especially people with platelet dysfunction. Nevertheless, based on the current
development of nanotechnology, this may seem too advanced for nanorobots in their
early stage.
Instead of adopting Freitas version of artificial platelets with net packets onboard, this
study regarded nanorobots as biocompatible molecules self-assembling at the wound
site to form a temporary plug to trap blood components and let the hemostasis process
continue the formation of the clot. These early version of nanorobots represent artificial
platelets that are attracted, adhere to the exposed collagen in the injured vessel wall, and
release the substances to recruit additional platelets as well as artificial platelets to the
area.
The Characteristics of Nanorobots. In order to provide guidelines that are truly ben-
eficial toward the realization of future nanorobots, this study used the nanorobot model
based on the current development of nanotechnology as well as existing characteristics
in biological systems that has potentials to be included in the early-stage nanorobots. In
[22], apart from the energy that is fundamental to power the nanorobots, the essential
characteristics of nanorobots for self-assembly and self-repair tasks have been identified
as the following:
moving around the environment,
interacting with other nanorobots as well as the environment,
carrying defined characteristics for the assembly task, and
connecting to other nanorobots at a specific surface.
Kaewkamnerdpong and Bentley [22] used these characteristics to model nanorobots
for surface coating application. These characteristics could also be used for achieving
the self-assembly task for artificial platelets in this study; with these characteristics,
the nanorobots as artificial platelets could move through the bloodstream, seek for the
wound and form into a mass at the damage site to stop the bleeding. To support that
these characteristics have the potential to be realized, Kaewkamnerdpong and Bentley
[22] explored the literature and discussed 4 features with examples of nanotechnology
212 B. Kaewkamnerdpong et al.
development and examples in nature including actuator, signal generator and sensor,
programmability, and connection as concluded in Table 1. These features are adequate
to allow swarm-intelligence-based control mechanism for nanorobots.
This study adopted the nanorobot model proposed in [22] for self-assembly. Nev-
ertheless, to function as artificial platelets the nanorobots that assemble themselves to
repair the damaged blood vessel must be able to
moving around the environment,
interacting with other nanorobots and its environment,
generating signal that attracts other nanorobots and sensing the attraction signal,
carrying defined characteristics for the assembly task
connecting to other nanorobots at a specific surface, and
operating inside human body with biocompatibility.
In this study, each artificial platelet was spherical in shape, which is similar to clottocyte
[32]. The size of artificial platelets in this study was 2 m in diameter similar to natural
platelets and clottocytes [32,40]. For practical simulation, each artificial platelet was
limited to interact with others and its environment within a defined perception range. As
this study focussed on the use of swarm intelligence to control nanorobot locomotion,
the following assumptions were made in the simulation;
The artificial platelets can move around the vessel model within a defined maximum
velocity that allow them to move in opposite direction to the blood flow to find the
wound site.
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 213
The artificial platelets can sense the changing environment within a defined percep-
tion range for measuring the concentration of chemical substance released from the
wound to locate the wound site.
The artificial platelets have connectors that can bind with vWF for adhesion to the
exposed collagen on vessel wall and can aggregate with other adhered artificial
platelets and other adhered natural platelets only at the wound area for forming the
structure.
The forming structure of the artificial platelets is stopped when the vessel releases
the endothelium-derived relaxing factor and prostacylin [40].
The artificial platelets cannot connect with other artificial platelets and other blood
cells while traveling in bloodstream.
The artificial platelets that already adhered to the wound can release the attraction
signals into the environment for inducing other artificial platelets to the wound
site and release the substances involved in blood clotting (secondary hemostasis)
after the formation of platelet plug, such as calcium, fibronectin, fibrinogen, and
coagulation factor FV and FVIII [40].
When an artificial platelet is very close to an optimal artificial platelet, the attraction
force applies; as a result, the artificial platelet is pulled to connect with that optimal
artificial platelet via the vWF, and then the new optimal artificial platelet releases
the attraction signals to the environment for inducing other artificial platelets.
Natural platelets are not included in this simulation.
After the clot dissolution, the human body will naturally dissolve the blood clot
after the wound has healed; the artificial platelets return to the bloodstream and
move along the blood vessel for repairing other wounds.
The Nanorobot Control. In term of the control mechanism for artificial platelet lo-
comotion, in this study the canonical PSO algorithm [9] was chosen to regulate the
artificial platelets to collaboratively self-assemble into a mass at the wound site to stop
the bleeding. Each artificial platelet moves in a three-dimensional model of a blood
vessel. At each time step, the artificial platelets move to new positions according to
the velocity update from the canonical PSO algorithm. With the signaling and sensing
units with limited perception range in artificial platelets for interacting with others and
the environment, only neighbors locating within the perception range can influence the
movement. The algorithm for artificial platelet locomotion can be shown in Table 2.
When the first artificial platelet find the wound site, it will adhere to the exposed colla-
gen at the wound and release attraction signal to activate others. Other artificial platelets
that can sense the attraction signal will become activated and move toward the wound
to adhere to the exposed collagen or optimal artificial platelets at the wound.
For a swarm of m artificial platelets, let xi (t) refers to the position of artificial platelet
i in [x, y, z] dimension of the search space at time t; the initial position of artificial
platelets are uniformly randomized over the search space. The velocity of artificial
platelet i in [x, y, z] dimension at time t is denoted by vi (t). For each dimension, the
velocity is initialized with uniform random between V MAX and +V MAX.
The fitness value, F(xi (t)), or the performance of each artificial platelet at its po-
sition can be the summation of the detected concentration of vWF molecules released
214 B. Kaewkamnerdpong et al.
from the wound and the detected intensity of attraction signal released from optimal
artificial platelets at the wound site. The concentration of vWF and attraction signal can
be described by Ficks second law,
In canonical PSO, each particle observes their neighbors and uses the local best
position that has the best fitness value in the neighborhood according to the network
topology to influence its movement toward a better position. In practice, the signaling
and sensing units of artificial platelets would be able to operate within a limited per-
ception range; hence, each artificial platelet can only interact to its neighbors and the
environment within its perception range. Moreover, instead of exchanging their per-
formance information with other individuals, the early-stage nanorobots may be able
to sense the presence of other individuals within its perception range only. With such
limitation, each nanorobot could not know whether any of the neighbors is in a better
position or not. Hence, the local best position, xlbest (t), in this study is determined by
randomly selecting neighboring positions within the defined perception range. The av-
erage position of all selected neighboring positions is used as the local best position.
Nevertheless, when the optimal artificial platelet is found, the local best position is set
as the optimal artificial platelet position. If there are more than one optimal artificial
platelets found, the local best position is the position of the nearest optimal artificial
platelet. In the case where both neighbors and optimal artificial platelets are not found,
the local best position is the current position.
The velocity of each nanorobot according to the canonical PSO algorithm can be
expressed as
vi (t + 1) = vi (t) + 1 x pbest,i xi (t) + 2 xlbest,i xi (t) . (3)
The blood flow along the tube is regulated by Poiseuilles law. The amount of flow,
q, for a given pressure difference, p, is
R4 p
q= , (5)
8
where is the viscosity of fluid, R is the tube radius, and is the length between two
points of the tube measured for pressure difference [44]. When blood is simplified as
Newtonian fluid, the fluid viscosity representing the internal friction in the fluid remains
constant. Hence, the average velocity of fluid, u,
is expressed as [44]
q pR2
u = = . (6)
R 2 8
The velocity of the Poiseuille flow or steady flow at a cross-sectional location of the
tube can be expressed as [44]
ks 2
us (r) = (r R2 ), (7)
4
218 B. Kaewkamnerdpong et al.
where r is the radial coordinate measured from the tube axis, and ks is the pressure
gradient driving the flow. The pressure gradient is considered to be constant and equal
to the pressure difference ps between two points of the tube divided by the length of
tube between them,
d p ps
ks = = . (8)
dx
Therefore, the maximum velocity, u, in the Poiseuille flow is at the centre of the tube;
the maximum velocity is two-folded of the average velocity or u = 2u.
As the heart pumps blood to circulate through the body, blood flows with pulsatile
motion according to the change of pressure. The driving pressure can be simulated with
a periodic function of time in term of pressure gradient [27,30],
dp
= A0 A1 (cos t), (9)
dx
where A0 is the constant component of the pressure gradient, A1 is the amplitude of
the oscillating component that gives rise to the systolic and diastolic pressure, and =
2 f p , where f p is the pulse frequency. The velocity of the pulsatile flow is a function
of radius, r, and time, t, that is [41,44]
The steady and oscillatory flow can be calculated separately. The steady flow velocity,
us (r), in a tube can be calculated by Eq. 7. The oscillatory flow velocity within the tube
can be expressed as
A1 J0 ( r)
u (r,t) = 1 ei t , (11)
i J0 ( R)
where is blood density, J0 is the Bessel function on order zero of first kind, and
= i3 / , where is kinematic viscosity.
In the blood vessels, the higher flow velocity than a critical velocity can cause tur-
bulent flow where the flow velocity become in various directions instead of the laminar
flow. Constriction of blood vessels decreases the lumen and can increase the probabil-
ity of turbulent flow [18]. Nevertheless, a blood vessel is modeled as a rigid tube for
simplicity; the turbulent flow is excluded from the model.
When nanorobots move in the blood vessels, the movement of nanorobots is influ-
enced by blood flow. The new position of artificial platelet in simulation system be-
comes
xi (t + 1) = xi (t) + vi (t + 1) + ub (xz,i (t),t) , (12)
where ub (xz,i (t),t) is the blood velocity at the z position of the artificial platelet xi and
time t. Note that the drag force that acts on the objects in fluid dynamic and affects the
fluid velocity and flow direction is ignored in this study. This is because the artificial
platelets are much smaller and less numerous than other blood cells, so the drag force
can be negligible [16].
In the blood vessels, apart from the blood flow the movement of future nanorobots
can be affected by collisions with blood cells. In this model, a collision between
nanorobots and blood cells is induced with probability p which is dependent on the
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 219
Start
Yes
Calculate fitness values at the current Update artificial platelet position in the case
position of collision with the vessel wall
No
Update new velocity according to situation Test for termination
Yes
A Stop
As the average diameter of an arteriole is 30 m [6], Iida [19] reported that when the
blood flows in the arterioles of diameter less than 0.1 mm, the velocity profiles could be
generally described by both Casson and Herchel-Bulkley (H-B) fluid models and that
the velocity profiles of blood flow in the arterioles with diameters less than 0.065 mm
could only be described by H-B fluid model. Hence, the H-B model is the appropriate
model.
In Herchel-Bulkley model, the shear rate, , can be expressed by [7]
0, if 0 r R p ;
= dV (13)
, if r > R p ;
dr
where V is the blood velocity, r is the radial coordinate position in the vessel, and R p is
plug core radiuswhich can be computed by
y
Rp = R . (14)
w
= y + m n (15)
where y is the yield stress, m is the consistency index, and n is the power-law index,
n < 0. In addition, the shear stress can be represented as a function of pressure gradient,
which is
r dP
= . (16)
2 dx
where dP/dx is the pressure gradient. The shear stress at wall can be computed by
substituting r = R in Eq. 16,
R dP
w = . (17)
2 dx
The viscosity can be expressed as
d
= (18)
d
The velocity can be computed from Eq. 13 and Eq. 15-17 following by taking inte-
gration with respect to r to obtain the steady flow velocity,
n+1 n+1
nR w 1n y n r y n
Vs (r) = 1 . (19)
n+1 m w R w
When r R p or in the plug core region, the velocity will be constant and equal to Vs at
R p . Hence, the plug core velocity is computed by substituting r = R p and using Eq. 14
in Eq. 19,
n+1
nR w 1n y n
Vp = 1 . (20)
n+1 m w
222 B. Kaewkamnerdpong et al.
The total velocity can be considered as the summation of the steady flow velocity and
oscillation flow velocity,
V (r,t) = Vs (r) + V (r,t). (22)
Fig. 3. The velocity profile of (a) Newtonian and (b) non-Newtonian fluid
4 The Demonstration
Parameter Value
Vessel thickness 20m
Vessel outer radius 35m
Vessel inner radius 15m
Endothelium layer 3m
Vessel length 500m
Pulse frequency, f p 1 Hz
Blood density, 1050 kg/m3
Dynamic viscosity, 0.00356 Pa.s
Kinematic viscosity, 3.302x106 m2 /s
Pressure gradient 20000
4000(cos t)
Hematocrit 40%
Fig. 4. Illustration of a nanorobot connecting with 8 other individuals; the nanorobot is presented
in green color whereas individuals connecting to this nanorobot are presented in black
and acceleration constants. In this study, the constriction coefficient, , was set accord-
ing to the suggestion from [9].
As such nanorobots or artificial platelets have not been realized yet, it could be ben-
eficial to study that to what extent of capability early-version nanorobots are needed to
accomplish their tasks. There are three parameters that relate to the capability and affect
the effectiveness and efficiency of artificial platelet control: the perception range, maxi-
mum velocity (V MAX) and response time. The perception range indicates the area that
an artificial platelet can interact with other individuals and its environment; it would be
determined by the operating range of signaling and sensing units in artificial platelets.
V MAX is the maximum velocity for which the artificial platelets are allowed to travel.
In real artificial platelets, this value is defined by the actuator ability. The larger the
value, the faster the artificial platelets can move. However, how fast an artificial platelet
can travel is also dependent on the response time of an artificial platelet to the control
mechanism. In this study, the effects of all three parameters were investigated in both
224 B. Kaewkamnerdpong et al.
Newtonian and non-Newtonian models. The summary of parameter settings for all three
testing parameters can be expressed in Table 4.
The performance of self-assembly in blood vessel repair can be indicated in terms
of accuracy and efficiency through observing the percentage of wound coverage and
speed to achieve the goal respectively. The resulting percentage of wound coverage is
determined by Monte Carlo simulation. The wound coverage is the ratio of the number
of testing points that have optimal artificial platelets to the total number of randomly
selected points in the wound area. The greater wound coverage indicates the better self-
assembly performance. The wound coverage can be illustrated in term of the median
value over a number of trials of the experiment. The median values are chosen to repre-
sent the result because, in the simulation, randomness is used in the control mechanism
so it is more important to investigate how reliable the control mechanism could allow
the nanorobots to achieve their task. The speed of artificial platelet system is represented
by the number of iterations used to form a mass. The lower the number of iterations is
used, the greater the speed of self-assembly.
Table 4. The setting of test parameters: the perception range (PRANGE), maximum velocity
(V MAX) and response time (NRTime)
Situations c1 c2
Both of neighbor and optimal articial platelets are not found 0.729 2.050 2.050
Neighbors are found, but no optimal articial platelet is found 0.729 2.050 2.050
Optimal articial platelets are found 0.250 0.000 0.000
Brownian motion [16,29]. The Brownian motion is the random walk of particles in
fluid. Thus, the motion of artificial platelets without control mechanism is randomized
with Gaussian distribution. Both the system with artificial platelets using canonical PSO
to control their operation and the system with artificial platelets using random move-
ment have the attraction signaling and sensing unit and other essential characteristics,
but the movement of the latter case is completely randomized even when the attraction
signal is found. However, the attraction force still applies when the artificial platelet is
very close to an optimal artificial platelet in the system with random movement.
Varying Perception Range. In the case where the perception range of artificial platelets
is varied, the results are illustrated in Fig. 5-a and 5-d. As the perception range is in-
creased, the levels of wound coverage are increased. When the perception range of the
artificial platelets is 7.5, 15 and 30 m, they reached 80 percentage of wound coverage.
226 B. Kaewkamnerdpong et al.
Fig. 5. The comparison of the results from random movement and Canonical PSO: (a) the median
wound coverage for the variation of PRANGE, (b) V MAX and (c) NRTime; (d) the mean number
of iterations used for different settings of PRANGE, (e) V MAX and (f) NRTime
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 227
When the perception range is 1.875 m, the artificial platelets cannot form the structure
or there are only a few artificial platelets adhered at the wound site. It may be because
the artificial platelets can only observe the environment close to themselves and, hence,
there is low chance to find the wound site. In addition, they may not sense the neigh-
boring artificial platelets, so the velocity modification is only influenced by their own
experience.
In term of the speed to achieve the goal, the speed of artificial platelets forming into
a structure at the wound site is increased as the perception range is increased. It may
be because when they can observe the environment in larger area, there is a greater
chance that they can find the wound site as well as the attraction signal from optimal
artificial platelets; consequently, they can move toward to the wound site and form into
a structure faster. In addition, a greater number of neighboring artificial platelets may
be observed when the perception range of artificial platelets is larger. This may lead
artificial platelets to find the wound site as the control mechanism for finding a good
solution by canonical PSO is influenced by their own experience and their neighbors
experiences.
At different perception ranges, the same level of wound covered and speed to form a
structure were obtained for the system with random movement. It is because the chang-
ing perception range has no effect on their movements. The result indicates that the
capability of signaling and sensing unit have no effect on the performance of the artifi-
cial platelet swarm system with random movement.
Varying Maximum Velocity. The results are illustrated in Fig. 5-b and 5-e. When the
V MAX of the artificial platelets is 0.9416 m, the median of wound coverage percent-
age is the lowest at 11.32%. Because the artificial platelets can move with just small
steps, they can slowly move to the wound site. With the influence from the blood flow,
they are unable to move in opposite direction to the blood flow, so they cannot move to
the wound site located behind their current position. For other cases, all systems reached
80% of wound coverage with the same levels of median wound coverage. The speed for
forming into a mass increases as the V MAX value increases. Because the greater V MAX
allows an artificial platelet to move with a larger step and can move faster and move in
opposite direction to the blood flow. Nevertheless, a larger step movement could lead
to more collision to the vessel wall or could overstep the wound. On the other hand,
the artificial platelets with lower V MAX could not move in opposite direction to the
blood flow; consequently, they could only move along the blood flow, but sometimes
the blood flow could lead a few artificial platelets to find to the damage site. In con-
clusion, the result indicates that the greater maximum velocity of the artificial platelets
brings about the better performance of the artificial platelet swarm in terms of both the
wound coverage and speed for forming into a mass.
When V MAX of the artificial platelets with random movement is 0.9416 and 1.8832
m, the median of wound coverage percentage is lower than 20% because the arti-
ficial platelets move with small step that allow they move slowly or cannot move to
target area. The same levels of wound coverage are obtained when V MAX of the artifi-
cial platelets is greater than 1.8832 m. Although the artificial platelets randomly move
228 B. Kaewkamnerdpong et al.
in blood vessel wall, the steps of movement are limited by their V MAX values. The
speed for forming into a mass increases as V MAX increases; because the greater V MAX
allows the artificial platelet to move with larger, they can move faster and move in
opposite direction to the blood flow. Moreover, the chance to find the wound increases
as well. The effects of different level of V MAX values to the performance of the artificial
platelet system with random movement are similar to the system with canonical PSO-
based control mechanism for the same reason.
Varying Response Time. In term of varying response time of artificial platelets, the
level of resulting wound coverage increases as the response time decreases (or the re-
sponse speed of artificial platelets increases). The result illustrated in Fig. 5-c and 5-f
shows the improved trend of speed for forming into a mass as the response speed of
artificial platelets increases. Artificial platelets that perform in an extremely dynamic
environment would need to quickly readjust their positions according to the changes
in the environment. With the better response ability, artificial platelets could deal with
changing environment and achieve the designated task. In addition, these could gain
more chance of artificial platelets finding the wound site and, in turn, more artificial
platelets forming into a structure at wound site. Otherwise, the artificial platelets at-
tempt to travel toward the desired position calculated by the control mechanism but
may not be able to move to the desired position due to the external influence from the
blood flow. When the response time of artificial platelets is greater than the simula-
tion time, the movement of artificial platelets depends on the blood flow more than the
control mechanism; the blood flow could lead the artificial platelet to either closer to or
away from the wound site. Hence, the greater response speed of artificial platelets allow
them to perform their task with better performance in terms of both wound coverage and
speed to form into the structure.
The same levels of wound coverage are also obtained when the artificial platelets
with random movement have different response time settings. Nevertheless, the dif-
ferent response speed of artificial platelets showed impact on the speed of forming
structure; the speed of forming into a mass increases as the response speed of artificial
platelet increases similar to the system with canonical PSO-based control mechanism.
It may be because the artificial platelets with faster response speed could exhibit more
random movement than artificial platelets with lower response speed in the same time.
Hence, the artificial platelets with the better response ability allow the greater chance to
move to the desired position faster.
Although compared to the system with random movement, the results from the system
with canonical PSO based control mechanism in Newtonian model showed better re-
sults as illustrated in Fig. 5, it could not yet assure that the artificial platelets would
effectively complete their tasks in real application. In computer simulation studies, the
realistic of simulation is the one of main concerned issues. The more the simulation
model is closer to the real environment as well as situation, the greater chance that
the result from the study could apply to the real application. In a small blood vessel,
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 229
non-Newtonian model could better simulate the blood flow than Newtonian model [7].
The region of high velocity in non-Newtonian blood is larger than in Newtonian blood.
As artificial platelets with the influence from high blood velocity might go forward
too fast and overstep the wound, it can be anticipated that artificial platelets in non-
Newtonian blood may have higher possibility to overstep the wound and need to move
against the blood flow back to the wound.
In time-independent blood, H-B model is used for simulating non-Newtonian blood
flow. The flow is assumed to be steady, so the flow rate is not changed over time. Addi-
tionally, the flow is assumed to be fully developed flow, which means that the velocity
profile is stable. Normally in human vessel, blood has Reynolds number less than 300
except in aorta, so only laminar flow is used in this simulation. At the wall, there is
no-slip condition which means that the velocity at wall is zero. For H- B model, there
are some constant parameters which are n is set as 0.5, the yield stress is 0.001 Pa [6],
the shear stress at wall can be computed from Eq. 17 with pressure gradient = 20400,
and m = 0.0019 by computing Eq. 15.
As illustrated in Fig. 3, the velocity profile of non-Newtonian fluid exhibits with
wider region of high velocity around plug core region than that of Newtonian fluid.
Hence, in this experiment, the parameter setting for the canonical PSO-based control
mechanism of artificial platelets is set to cope with non-Newtonian fluid as summa-
rized in Table 6 and 7. There are two settings for c1 and c2 used for the purpose of
comparison. Firstly, in PSO1, the value of c1 and c2 are equally set at 2.050 for giv-
ing confidence in artificial platelet own knowledge as much as social knowledge in all
situations. Secondly, in PSO2, the values of c1 and c2 are differently set to balance the
influence from artificial platelets own knowledge and social knowledge according to
the state of artificial platelets. Each experiment was run 40 times.
Fig. 6 illustrates the resulting wound coverage and number of iterations respectively
in comparison with those from random movement. The results show that the higher
PRANGE value gives the greater chance for artificial platelets to meet and get informa-
Table 6. Parameter setting of , c1 and c2 according to different situations for artificial platelets
in non-Newtonian model: PSO1
Situations c1 c2
Both of neighbor and optimal articial platelets are not found 0.729 2.050 2.050
Neighbors are found, but no optimal articial platelet is found 0.729 2.050 2.050
Optimal articial platelets are found 0.729 2.050 2.050
Table 7. Parameter setting of , c1 and c2 according to different situations for artificial platelets
in non-Newtonian model: PSO2
Situations c1 c2
Both of neighbor and optimal articial platelets are not found 0.729 2.733 1.367
Neighbors are found, but no optimal articial platelet is found 0.729 2.050 2.050
Optimal articial platelets are found 0.729 1.025 3.075
230 B. Kaewkamnerdpong et al.
tion from neighboring artificial platelets as well as from the environment. A low-valued
PRANGE artificial platelets would move pass the wound even though it gets near the
wound as it is not near enough to sense the wound in its perception range. For the max-
imum velocity, the greater value allows artificial platelets to move more freely against
the influence of blood flow. On the other hand, the greater value could also allow arti-
ficial platelets to move too fast along the blood flow and overstep the wound site; this
can be observed from the result when V MAX = 60 is worse than that when V MAX =
30. Finally, the response time denotes how fast the artificial platelets could response
to the situation in dynamic environment; usually, the smaller value should provide the
better performance. Nevertheless, the results show that it is not necessarily true as the
case with NRTime = 0.002 gives a better result than the one with NRTime = 0.001. It is
observed that the faster responding artificial platelets move slower in the environment,
so the speed to find the wound site is lower. On the other hand, when NRTime = 0.01 the
artificial platelets respond too slowly under the influence of blood flow. In most trials,
the artificial platelets could not fill the wound within 10,000 iterations.
For using PSO control mechanism, the results show that the artificial platelets us-
ing PSO1 and PSO2 can fill the wound faster than those that randomly move. In term
of the wound coverage, the system using PSO-based control performs slightly better.
Between two PSO-based control mechanisms, PSO2 performs slightly better as the me-
dian wound coverage is a little higher than that of PSO1 in many experiments; this also
indicates that PSO2 has a greater chance for artificial platelets to fill the wound. The
difference is more distinctively observed in term of self-assembly speed. PSO2 results
in faster self-assembly in several experiments. This indicates that balancing the influ-
ence from individual and social knowledge according to corresponding situations could
contribute on the self-assembly performance.
This study investigated the possibility of using a swarm intelligence technique to adap-
tively control swarm system of nanorobots for medical application. The performance of
swarm-intelligence-based nanorobot control was demonstrated in blood vessel repair
application. Through exploring different swarm intelligence techniques, the canonical
PSO algorithm was selected and used in the demonstration; the reason was that the con-
cept of control mechanism in PSO can fit to the characteristics and tasks of early-stage
nanorobots in the future better than other techniques. Moreover, the control parameters
for PSO are also corresponding to the characteristics of early-stage nanorobots.
Through demonstration, the effects of different artificial platelet capabilities includ-
ing perception range, maximum velocity and response time under the control by canon-
ical PSO were investigated in both Newtonian and non-Newtonian model of blood
flow. The results indicated that if the early-stage nanorobots comprise these capabil-
ities and use the control mechanism to adaptively control their locomotion based on
partial influences from their neighbors, they could effectively achieve their tasks as arti-
ficial platelets. From varying principal parameters of the swarm artificial platelets with
canonical PSO based control and comparing the performance with artificial platelets
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 231
Fig. 6. The comparison of experimental results between random movement, PSO1 and PSO2 in
non-Newtonian model: (a) the median wound coverage for the variation of PRANGE, (b) V MAX
and (c) NRTime; (d) the mean number of iterations used for different settings of PRANGE, (e)
V MAX and (f) NRTime.
232 B. Kaewkamnerdpong et al.
randomly moving in self-assembly application as blood vessel repair, the results from
Newtonian model can be concluded as follows:
The greater perception range increases the chance of the artificial platelets with
canonical PSO based control, which relies on their own experience and their neigh-
bors experience to influence their movement, to find the target site. However, the
different levels of perception ability have no impact on the motion of the artificial
platelets without control because their movements are random.
The greater maximum velocity that allows the artificial platelets to move with larger
step and to move in opposite direction to the blood flow increases the chance to find
the target site.
The grater response ability of artificial platelets provides a better adjustment of
their position to situation in both the system of artificial platelets with canonical
PSO and the system of artificial platelets with random movement.
From the demonstration, the results from non-Newtonian model exhibited the similar
trend, but the effect of maximum velocity and response time are distinct. The maximum
velocity was set with greater values in order to obtain the trend as in Newtonian model.
As non-Newtonian environment brings about wider region of high blood velocity in-
side the vessel, the higher setting of maximum velocity may be required to cope with
the greater blood flow influence. However, the high velocity could also induce too fast
movement so the artificial platelets overstep the wound. In term of response time, arti-
ficial platelets may not need to be so fast responding as the high blood velocity area is
in the middle of the vessel but their target wound is always at the vessel wall where the
velocity is lower.
The natural platelets form into a platelet plug through the following activities: the
non-activated platelet becomes activated when it senses vWF and, then, adheres to the
exposed collagen fiber at the wound. The activated platelets synthesize and release
chemicals for activating nearby platelets to aggregate and form into a plug. The ac-
tivated platelets can link to each other before connecting with the adhered platelets at
the wound. The natural platelets can connect with one another at all directions because
they have many receptors on their surface [40]. Nevertheless, the artificial platelets in
this study have only 8 connectors. They are designed to connect with other artificial
platelets adhered at the wound site only. Whenever the artificial platelet comes within a
very close distance to the wound, it immediately adheres to the exposed collagen fiber at
the wound and, then, releases the attraction signal which draws other artificial platelets
to get nearer and connect to form a structure to cover the wound. The examples of the
self-assembly outcome of artificial platelets into a mass at the wound site are illustrated
in Fig. 7.
Through simulation, the essential characteristics of the nanorobots with canonical
PSO based control mechanism for the self-assembly task are investigated and found
that the future nanorobots with these characteristics could plausibly operate with self-
assembly task in medical applications under swarm-intelligence-based control. The ex-
periment results suggested that the early-stage nanorobots in the future could perform
self-assemble task in blood vessel repair when they have the perception ability around
the diameter of the blood vessel and the maximum velocity of about the average blood
velocity inside the blood vessel.
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 233
Fig. 7. Examples of the output structures at the wound site for repairing the blood vessel from
artificial platelet system
Although artificial platelets in this study have no knowledge of their current loca-
tions in the environment and they cannot share their information with their neighbors,
the artificial platelets with attraction signaling and sensing unit and other essential char-
acteristics that are controlled by canonical PSO can achieve their goal. Moreover, the
movements of artificial platelets that operate in dynamic environment like a circula-
tory system are influenced by the blood flow in the vessel. In this study, the artificial
platelets with essential characteristics identified in section 3.1 cannot acknowledge the
changes in their movement caused by the blood flow. Hence, the canonical PSO al-
gorithm calculates for new movement, but the artificial platelets cannot actually travel
toward the intended location because of the external influence from the blood flow. The
information from the artificial platelets personal best position, which is calculated from
accumulating their movement from the previous personal best position, can be mislead-
ing. Nevertheless, the results indicated that PSO-based control could allow the artificial
platelets to perform better than random movement even with misleading information.
Nonetheless, if nanorobots have knowledge of their current position in the problem
space, the accurate personal best position of nanorobots will be obtained. The ability
to sense the external changes in the environment may allow the nanorobots that know
their coordinate to achieve their goal with better performance as they can get the accu-
rate personal best position and will be able to effectively move to the desired position
according to canonical PSO. This may allow nanorobots to perform the self-assembly
tasks with the higher performance level and may allow nanorobots to perform more
234 B. Kaewkamnerdpong et al.
complex task than the self-assembly task in this study. Another additional characteristic
that could be useful for nanorobots is the ability to communicate with other nanorobots
in the system. If nanorobots can communicate and share some information about their
performance with their neighbors, the nanorobots can know whether their neighbors
are in better or worse positions which will allow collective behavior in the nanorobot
swarm; this could lead to improvement of system performance and the ability to per-
form more complex tasks.
The findings in this study could serve as guidelines for the characteristics and behav-
iors required for early-stage nanorobots toward the realization of medical nanorobots in
the near future. Nevertheless, this simulation study is based on a rigid tube model due
to our limited knowledge of our circulatory system. In the future, the study on an elas-
tic tube model, which is more realistic simulation of blood vessel, would be applied.
Moreover, when there appears evidence of plausibly additional characteristics, a fur-
ther simulation study should be conducted to investigate for suitable, effective control
mechanism in more complex applications. This study on using swarm intelligence for
nanorobot locomotion control has shown an attempt to move forward toward the real-
ization of nanorobots from the computer scientists point of view. With the collaboration
in various research communities, it could be anticipated that the nanorobots would not
exist in visions but could become real with truly effective benefit in nanomedicine in
the near future.
References
1. Berne, R.M., Levy, M.N., Koeppen, B.M., Stanton, B.A.: Physiology, 5th edn. Elsevier
(2004)
2. Bonabeau, E., Dorigo, M., Theraulaz, G.: Swarm Intelligence from Natural to Artificial Sys-
tems. Oxford University Press (1999)
3. Campbell, N.A., Reece, J.B., Simon, E.J.: Essential Biology with Physiology, 2nd edn. Pear-
son Benjamin Cummings (2007)
4. Cavalcanti, A., Shirinzadeh, B., Fukuda, T., Ikeda, S.: Nanorobot for brain aneurysm. Inter-
national Journal of Robotics Research Archive 28, 558570 (2009)
5. Cavalcanti, A., Freitas Jr., R.A.: Nanorobotics control design: A collective behavior approach
for medicine. IEEE Transactions on NanoBioScience 4(2), 133140 (2005)
6. Chandran, K.B., Yoganathan, A.P., Rittgers, S.E.: Biofluid Mechanics: The Human Circula-
tion, 1st edn. Taylor & Francis (2007)
7. Chhabra, R.P., Richardson, J.F.: Non-Newtonian flow and applied rheology: engineering ap-
plications, 2nd edn. Elsevier, Amsterdam (2008)
8. Clerc, M.: The swarm and the queen: Towards a deterministic and adaptive particle swarm
optimization. In: Proceedings of the IEEE Congress on Evolutionary Computation, pp. 1951
1957 (1999)
9. Clerc, M., Kennedy, J.: The particle swarm-explosion, stability, and convergence in a mul-
tidimensional complex space. IEEE Transactions on Evolutionary Computation 6(1), 5873
(2002)
10. Colorni, A., Dorigo, M., Maniezzo, V., Elettronica, D., Milano, P.: Distributed optimization
by ant colonies. In: Proceedings of the First European Conference on Artificial Life, pp.
134142 (1992)
Modeling Nanorobot Control Using SI for Blood Vessel Repair: A Rigid-Tube Model 235
11. Dorigo, M., Gambardella, L.M.: Ant colony system: A cooperative learning approach to the
traveling salesman problem. IEEE Transactions on Evolutionary Computation 1(1), 5366
(1997)
12. Eric Drexler, K.: Engines of Creation 2.0: The Coming Era of Nanotechnology. WOWIO,
20th anniversary edition (2007)
13. Eberhart, R.C., Kennedy, J.: A new optimizer using particle swarm theory. In: Proceedings of
the Sixth International Symposium on Micro Machine and Human Science, Nagoya, Japan,
pp. 3943 (1995)
14. Engelbrecht, A.P.: Computational Intelligence: An Introduction. John Wiley & Sons (2007)
15. Feynman, R.P.: Theres plenty of room at the bottom: An invitation to enter a new field of
physics. Engineering and Science 23 (1960)
16. Fogelson, A.L., Guy, R.D.: Immersed-boundary-type models of intravascular platelet aggre-
gation. Computer Methods in Applied Mechanics and Engineering 197(25-28), 20872104
(2008)
17. Freitas Jr., R.A.: Exploratory design in medical nanotechnology: A mechanical artificial
red cell. Artificial Cells, Blood Substitutes, and Immobil. Biotech. 26, 411430 (1998),
http://www.foresight.org/Nanomedicine/Respirocytes.html
18. Ganong, W.F.: Review of Medical Physiology, 22th edn. McGraw-Hill (2005)
19. Iida, N.: Influence of plasma layer on steady blood flow in micro vessels. Japanese Journal
of Applied Physics 17, 203214 (1978)
20. Fister Jr., I., Yang, X.S., Fister, I., Brest, J., Fister, D.: A brief review of nature-inspired a
brief review of nature-inspired algorithms for optimization. Elektrotehniski Vestnik 80(3),
116122 (2013)
21. Kaewkamnerdpong, B., Bentley, P.J.: Computer science for nanotechnology: Needs and op-
portunities. In: Proceedings of the Fifth International Conference on Intelligent Processing
and Manufacturing of Materials (2005)
22. Kaewkamnerdpong, B., Bentley, P.J.: Modelling nanorobot control using swarm intelligence:
A pilot study. In: Lim, C.P., Jain, L.C., Dehuri, S. (eds.) Innovations in Swarm Intelligence.
SCI, vol. 248, pp. 175214. Springer, Heidelberg (2009)
23. Karaboga, D., Akay, B.: A comparative study of artificial bee colony algorithm. Applied
Mathematics and Computation 214, 108132 (2009)
24. Karaboga, D., Basturk, B.: A powerful and efficient algorithm for numerical function opti-
mization: Artificial bee colony (abc) algorithm. Journal of Global Optimization 39(3) (2007)
25. Karaboga, D., Basturk, B.: On the performance of artificial bee colony (abc) algorithm. Ap-
plied Soft Computing 8, 687697 (2008)
26. Levick, J.R.: An Introduction to Cardiovascular Physiology, 5th edn. Hodder Arnold (2010)
27. Massoudi, M., Phuoc, T.X.: Pulsatile flow of blood using a modified second-grade fluid
model. Computers and Mathematics with Applications 56, 199211 (2008)
28. Mody, N.A., King, M.R.: Influence of brownian motion on blood platelet flow behavior and
adhesive dynamics near a planar wall. Langmuir 23(11), 63216328 (2007)
29. Mody, N.A., King, M.R.: Platelet adhesive dynamics. part ii: High shear-induced transient
aggregation via gpiba-vwf-gpiba bridging. Biophysical 95(5), 25562574 (2008)
30. Pal, B., Misra, J.C.: A mathematical model for the study of the pulsatile flow of blood under
an externally imposed body acceleration. Mathematical and Computer Modelling 29, 89106
(1999)
31. Payton, D., Daily, M., Estowski, R., Howard, M., Lee, C.: Pheromone robotics. Autonomous
Robot 11(3), 319324 (2001)
32. Robert Jr., A.F.: Clottocytes: Artificial mechanical platelets. Foresight Update 41 (2000)
33. Robert Jr., A.F.: Microbivores: Artificial mechanical phagocytes using digest and discharge
protocol. Journal of Evolution and Technology 14, 152 (2005)
236 B. Kaewkamnerdpong et al.
34. Sanchita, P., Abhimanyu, K.S.: Identification of metastatic tumors by using dna nanorobot:
A fuzzy logic approach. International Journal of Computer Applications 1, 514 (2010)
35. Sankar, D.S., Lee, U.: Two-fluid herschel-bulkley model for blood flow in catheterized arter-
ies. Journal of Mechanical Science and Technology 22, 10081018 (2008)
36. Sharma, N.N., Mittal, R.K.: Nanorobot movement: Challenges and biologically inspired so-
lutions. International Journal on Smart Sensing and Intelligent Systems 1(1) (March 2008)
37. Slayter, H., Loscalzom, J., Bockenstedt, P., Handin, R.I.: Native conformation of human von
willebrand protein. analysis by electron microscopy and quasi-elastic light scattering. The
Journal of Biological Chemistry 260, 85598563 (1985)
38. Stanfield, C.L., Germann, W.J.: Principles of Human Physiology, 3rd edn. Pearson Benjamin
Cummings (2008)
39. Stanley, R.G., Tucker, K.L., Barrett, N.E., Gibbins, J.M.: Platelets and their role in throm-
botic and cardiovascular disease: the impact of proteomic analysis. In: Platelet Proteomics:
Principles, Analysis and Application, pp. 326. John Wiley & Sons (2011)
40. Tortora, G.J., Derrickson, B.H.: Principles of Anatomy and Physiology: Maintenance and
Continuity of the Human Body, 12th edn. John Wiley & Sons (2009)
41. Waite, L., Fine, J.: Applied Biofluid Mechanics. McGraw-Hill (2007)
42. Wang, J.: Can man-made nanomachines compete with nature biomotors? ACS Nano 3(1),
49 (2009)
43. Warkentin, T.E.: Thrombocytopenia due to platelet destruction and hypersplenism. In: Hema-
tology: Basic Principles and Practice, 5th edn., pp. 21132131. Churchill Livingstone Else-
vier (2009)
44. Zamir, M.: The Physics of Coronary Blood Flow. Springer (2005)
45. Zhang, L., Abbott, J.J., Dong, L.: Characterizing the swimming properties of artificial bacte-
rial flagella. Nano Letter 9(10), 36633667 (2009)
Author Index