1

Genetic Algorithms in Wireless Networking:Techniques, Applications, and Issues

Usama Mehboob, Junaid Qadir, Salman Ali, and Athanasios Vasilakos

Abstract—In recent times, wireless access technology is be-coming increasingly commonplace due to the ease of operationand installation of untethered wireless media. The design ofwireless networking is challenging due to the highly dynamicenvironmental condition that makes parameter optimization acomplex task. Due to the dynamic, and often unknown, operatingconditions, modern wireless networking standards increasinglyrely on machine learning and artificial intelligence algorithms.Genetic algorithms (GAs) provide a well-established frameworkfor implementing artificial intelligence tasks such as classifica-tion, learning, and optimization. GAs are well-known for theirremarkable generality and versatility, and have been applied ina wide variety of settings in wireless networks. In this paper,we provide a comprehensive survey of the applications of GAs inwireless networks. We provide both an exposition of common GAmodels and configuration and provide a broad ranging surveyof GA techniques in wireless networks. We also point out openresearch issues and define potential future work. While varioussurveys on GAs exist in literature, our paper is the first paper,to the best of our knowledge, which focuses on their applicationin wireless networks.

I. INTRODUCTION

The study of nature is a rich source of inspiration forresearchers working in the area of artificial intelligence andmachine learning. In particular, the human brain and thebiological process of evolution have helped develop and guideresearch in the neural network, and the evolutionary algorithmscommunities. A genetic algorithm (GA) is a metaheuristiccomputational method [1], inspired from biological evolution[2], that aims to imitate the robust procedures used by variousbiological organisms to adapt as part of their natural evolution.GAs have been successfully used in fields as diverse as air-craft industry, chip design, computer animation, drug design,telecommunications, software creation, and financial markets[3].

The field of genetic algorithms (GA) was established byJohn Holland who investigated the evolution of complexadaptive systems (CAS) comprising interacting genes, startingin the early 1960s, and culminating in the publication in1975 of his seminal book on this topic [4]. The idea ofembedding computers with learning inspired from evolutiongoes back even further as Alan Turing proposed a “learningmachine” which would parallel the principles of evolution in

Usama Mehboob (10beeumehboob@seecs.edu.pk), Junaid Qadir (ju-naid.qadir@seecs.edu.pk), and Salman Ali (09mscsesali@seecs.edu.pk) arewith the Electrical Engineering Department at the School of ElectricalEngineering and Computer Science (SEECS), at the National University ofSciences and Technology (NUST), Pakistan. Junaid Qadir is an Assistant Pro-fessor at SEECS, NUST. Athanasios Vasilakos (vasilako@ath.forthnet.gr) is aProfessor of Computer Engineering at the University of Western Macedonia,Greece.

his landmark 1950 paper on machine intelligence [5]. Hollandused the biological metaphor of chromosomes to refer tostrings of binary symbols encoding a candidate solution tothe given problem. Holland proposed using the computationalanalogues of the biological evolutionary processes of randommutation, crossover, and natural selection [6] to enable popu-lations of chromosomes to get increasingly better at problemsolving. The underlying premise is that given enough time, theprocess would converge towards a population that contains achromosome (or chromosomes) that solves the given problem.Apart from the mutation and crossover operators, The designof a GA framework also involves other design issues suchas genetic representation (encoding), population initialization,fitness function formulation, and a selection mechanism. Morerecently, researchers have also proposed the use of altruistictechniques in an evolutionary framework that involve cooper-ation as a fundamental primitive [7].

A. Motivations for using GA

1) Generality and Versatility: GAs apply in a wide varietyof settings and can be easily molded to particular problems.GAs constitute a very general meta-heuristic technique whichcan be thought of as the sledgehammer of the craft of algo-rithms, much like the technique of artificial neural networks(ANNs). In particular, GAs can be readily invoked in areasthat do not yield readily to standard approaches, or whenmore specialized techniques fail. GAs are capable of solvingextremely large problems that have large search spaces. GAsare very good at navigating through huge search spaces toheuristically find near optimal solutions in quick time. In ad-dition, GAs can work even when the objective function is notexactly known since GAs rely only on an objective function’sevaluation (without necessarily knowing the objective functionexplicitly). Although GAs do not guarantee optimality, GAsare generally useful in practice. Bhandar et al. [8] showed thatalthough GAs do not guarantee convergence to an optimalsolution, GAs avoid local optimas with a high probabilitythrough the use of mutation and crossover operators.

2) Adaptiveness and Online Problem Solving: To under-stand adaptivity in changing conditions, we will initiallydevelop the idea of a ‘landscape’ both as a metaphor (in whichwe visualize ourselves as climbing a landscape in pursuit ofthe highest peak) and as a mathematical object (in which thevalue of the function maps to the elevation of the landscape).Building upon the metaphor of fitness landscapes, and theinsight of GAs as stochastic search algorithms, two landscapemodels of relevance to optimization through GAs have been

arX

iv:1

411.

5323

v1 [

cs.N

I] 1

9 N

ov 2

014

2

proposed by Page [9]. Rugged landscapes are landscapes inwhich there are many peaks, valleys, and troughs (and nota single peak). Rugged landscapes assume that the fitnesslevels do not change; in evolutionary systems, however, thefitness function is also dependent on context, and on thebehavior of other competitors. This can be captured by themetaphor of dancing landscapes—which are adapted fromrugged landscapes. When a landscape ‘dances’, local peaksmay change, making a solution that was earlier optimal nolonger a peak.

Traditional techniques from optimization theory assumestatic and well-known topologies and are not suited to dynamicenvironments. In many problems of practical interest, ourfocus is more on satisficing (i.e., finding a sufficiently goodsolution that satisfies one’s purposes) rather than on optimizing(i.e., finding the best possible solution). A key benefit ofGAs is that they are well suited to the optimization task inthe dancing landscapes that characterize wireless networks (inwhich there is interaction between the various adaptive wire-less nodes and interdependency between the various node’sdecisions). GAs are well suited for building metaheuristicadaptive algorithms that can provide satisfactory performancein changing network conditions.

Another important aspect of GAs is that it is an onlineadaptive algorithm that that can operate in unknown environ-ments in an online fashion. Such an ability is crucial in variouscontrol settings in wireless networks in which decisions haveto be made automatically in run-time to cater to dynamic chan-nel parameters. In dynamic channel conditions that typicallyexist in most wireless networking configurations, optimizationis an elusive moving target since the optimal solution keepschanging as the conditions change. In online adaptive opti-mization algorithms, there is a fundamental tradeoff betweenexploration—which involves looking for potentially betterpreviously unexplored solutions—and exploitation that impliesthe use of previously known good solutions. Exploration ineffect is an attempt to find good adaptive building blocks andexploitation is the use and propagation of adaptations thatare known to perform well. We can envision mutation andrecombination of genes as analogous to exploration, whereasnatural selection can be envisioned as a form of exploitation[9]. In ever changing dynamic environments, such as wirelessnetworks, it is important to always keep on exploring. Hol-land’s original GA work used schema analysis to show that,under qualifying assumptions, GAs can achieve near-optimalexploration and exploitation tradeoff.

3) Ability to Find Good Building Blocks: By working interms of a population of candidate solutions, genetic algo-rithms can exploit the diversity of solutions to find buildingblocks—known as schemas in literature—which are substringsof chromosomes that denote high performing elements ofthe overall solution [4]. In the context of biological genet-ics, schemas correspond to constellations of genes workingtogether to help an organism adapt; evolution proceeds by se-lecting these constellations (or building blocks) through naturalselection. Using genetic operators such as recombination andcrossover, GAs can evolve better solutions through mixing of

the high performing basic building blocks.

4) Parallel Nature and Scalability: Traditional theory ofGAs presumes that GAs accomplish the culmination by dis-covering, emphasizing, and recombining the good traits ofchromosomes in a vastly parallel manner. GAs incorporatemultiple solutions together in a ‘population’ of solutions. GAsuse evolutionary techniques to test and improve the solutionsby using techniques such as mutation, crossover, selection,and recombination. The parallel nature of GAs renders it as asuitable optimization framework for scalable problem solvingin a wide range of applications.

5) Support For Multiobjective Optimization: Many prac-tical problems in wireless networks require optimizing formultiple parameters (that may potentially conflict with eachother). An important characteristic of GAs is that it can easilysupport joint optimization of multiple objectives. Support formulti-objective optimization—in which the considered objec-tives may potentially conflict with each other—is of significantpractical interest in wireless networks. Issues related to multi-objective optimization are introduced in [10].

6) Support for Global Optimization: Unlike network mod-els such as the multi-layered perceptrons (which make localchanges so as to find the local minima), GAs are suited tofinding the global optima due to a number of properties:

• They search by means of a population of individuals.• They work with an encoding of the multiple parameters.• They use a fitness function that does not require the

calculation of derivatives.• They search probabilistically.

7) Easy Implementation: GAs are computationally sim-pler compared to other complementary AI techniques suchas neural networks since they require only swapping andshifting of genes in chromosomes (unlike neural networks thatrequire adders for its multiple hidden layers). GAs are easilyimplemented and can be rapidly prototyped in digital signalprocessors (DSPs) or field programmable gate arrays (FPGAs).In addition, GAs are highly amenable to successful implemen-tation on modern parallel and cloud computing architecturesdue to its inherently parallel nature.

B. Contributions of this Paper

This paper offers a comprehensive survey of the applica-tions of GAs in wireless networking. In addition, we alsoprovide a self-contained introduction to the GAs. The aimof this paper is not to provide an exhaustive survey1, butto provide a representative sample of important works alongwith a comprehensive listing of the various applications ofGAs. We also highlight the application of GAs in wirelessnetworking configurations such as wireless mesh networks(WMNs), mobile ad-hoc networks (MANETs), cellular net-works (CNs), wireless sensor networks (WSNs), and cognitive

1The sheer breadth of GA literature precludes an exhaustive survey, theexhaustive bibliography created by Goldberg et al. in 1997—which is aexcellent reference to the pool of GA literature—amassed more than 4000publications. With the unabated interest in GAs, it will be no surprise if anupdated bibliography contains more than double the original references.

3

radio networks (CRNs). While a lot of material exists onGAs—including various textbooks [6] [11] [12], as well asgeneric survey papers [13] [14]—a focused survey article onthe applications of GAs in wireless networking is missing inliterature. This paper will be useful to networking researchersinterested in applying GAs in particular, and metaheuristicand evolutionary techniques more generally, in the context ofwireless networks. According to the best of our knowledge,this is the first survey paper that provides an extensive surveyof the applications of GAs in general wireless networks.

C. Organization of this Paper

We provide the necessary background on GA theory, work-ing and common techniques used to carry out the evolutionaryprocess in Section II. The applications of GAs in wirelessnetworks—such as routing, channel allocation, load balanc-ing, localization, and quality of service (QoS)—are listed inSection III. We highlight the various lessons learnt throughyears of GA research, in terms of common pitfalls as well asguidelines for successful implementation, in Section IV-B. InSection V, we highlight open issues and suggest directions forfuture work. Finally, this paper is concluded in Section VI.

To facilitate the reader, acronyms used in this paper arecollected in Table I as a convenient reference.

II. BACKGROUND: GENETIC ALGORITHMS

A detailed GA taxonomy is provided in Figure 1. Thistaxonomy diagram exhibits the breadth of GA’s applicationsas described in this section.

A. GA Terminology

Since the field of GA is inspired from the biological geneticevolution, the field uses a number of biological metaphorsin its terminology. The purpose of this short subsection isto introduce these biological metaphors in the context ofapplications of GAs in wireless networking. We introduce theterminology through the following list.

• Organism: It represents the entity (e.g., a radio parameter,or a wireless resource) being optimized.

• Population: It refers to the ensemble or collection oforganisms undergoing simulated genetic evolution.

• Chromosome: It encodes a particular solution to the prob-lem under study. (Biologically, a chromosome contains anorganism’s genetic makeup.)

• Fitness: It represents the utility of the current chromo-some. (In evolutionary theory, fit chromosomes are passedon through heredity while unfit chromosomes die out dueto the natural phenomenon of the ‘survival of the fittest’.)

• Gene: It is the basic building block of the chromosomedefining a particular feature of the simulated organism.(Each chromosome can contain a number of genes.)

• Allele: Each gene can take several alternative forms, eachof which is called an allele.

• Locus: It is the position on the chromosome containinga particular gene of interest.

TABLE I: Acronyms used in this paper.

Acronym Expanded Form

ACO Ant Colony OptimizationAP Access PointANN Artificial Neural NetworkBER Bit Error RateBFWA Broadband Fixed Wireless AccessBS Base StationBSP Broadcast Scheduling ProblemCAC Call Admission ControlCBDT Case Base Decision TheoryCSM Cognitive System ModuleCN Cellular NetworkCDMA Code Division Multiple AccessCR Cognitive RadioCRN Cognitive Radio NetworkDCA Dynamic Channel AssignmentDSA Dynamic Spectrum AccessDSPRP Dynamic Shortest Path Routing ProblemFA Forward AntFCA Fixed Channel AssignmentFPGA Field Programmable Gate ArrayGA Genetic AlgorithmGP Genetic ProgrammingGPRS General Packet Radio ServiceiGA Island GAHLB Hybrid Load BalancingILP Integer Linear ProgrammingLISP List Processing (Programming Language)LLF Least Loaded FirstLTE Long Term EvolutionMAC Media Access ControlMANET Mobile Ad-hoc NetworkMDP Markov Decision ProcessMOGA Multi-Objective GAMSC Mobile Switching CenterNN Neural NetworkNSGA Non-Dominated Sorted Genetic AlgorithmNSGA-II Non-Dominated Sorted Genetic Algorithm-IINSGS-IIa Adaptive NSGA-IINSGA-IIr Random NSGA-IIPGA Parallel GApNSGA-II Parallel NSGA-IIQoS Quality of ServiceRNC Radio Network ControllerRSS Radio Signal StrengthRSSI Received Signal Strength IndicatorSA Simulated AnnealingSGSN Serving GPRS NodeSINR Signal to Interference and Noise RatioSNR Signal to Noise RatioSOGA Single Objective GASP Shortest PathssNSGA-II Steady State NSGA-IISWN Software Defined Wireless NetworkUMTS Universal Mobile Telecommunications SystemWLAN Wireless Local Area NetworkWMN Wireless Mesh NetworkWSGA Wireless System Genetic Algorithm

• Mutation: It represents a random change of an allele (orthe value of a gene).

• Selection: The process through which the ‘fittest’ sim-ulated organism survives while the unfit solutions areweeded out.

B. Number-of-Objectives Based Classification

1) Single-Objective GAs: Single objective GAs (SOGAs)are popular when it comes to optimize single objective with

4

Fig. 1: Taxonomy of Genetic Algorithms

scalar fitness function. They are computationally less exhaus-tive but are required to run multiple times using differentweight vectors to optimize multiple objectives as in case ofmulti-objective GAs [15].

2) Multi-objective GAs: In a multi-objective optimization,there are multiple objectives that need to be simultaneouslyoptimized. In such a scenario, a single solution may not besuitable to multiple objectives. A solution may be best inone objective but worst in other cases. The overall systemis optimized, in all directions, by multiple solutions. Thesecategories of solutions are non-dominated solutions that lie onthe Pareto front. Each point in Pareto front is optimal, since nosingle solution exists that may improve one objective vectorwithout degrading at least one other objective. Almost all thesolutions lying on the Pareto front would be compromisesrelative to multiple objectives. There are various objectivesin wireless networking that need to be optimized such asthe signal to interference and noise ratio (SINR), the powerconsumption, and the bit error rate (BER). The Pareto front isillustrated through an example in Figure 2 with two conflictingobjectives: A is a non-dominated point with the objective ofminimizing power at the cost of BER, while the point Bcorresponds to the objective of minimizing BER at the costof power consumption. Both objectives cannot be achievedsimultaneously so a compromise solution has to be adopted.

Multi-objective GAs (MOGAs) can be employed in almostevery optimization setting provided that chromosomes encod-ing and objective function are appropriately accounted for.Details and applications of MOGAs are discussed next inSection III. In this section, we will describe one particular

instance of a MOGA framework known as Non-dominatedSorting Genetic Algorithm (NSGA).

a) Non-dominated Sorted Genetic Algorithm (NSGA):NSGA is a class of multi-objective optimization algorithmthat works at improving the adaptive fit of a population tothe Pareto front constrained by a set of objective functions.NSGA can also be seen as a case of an evolutionary algorithmfrom the domain of evolutionary computation. The populationwithin the algorithm is sorted into levels or sub-populations,which are based on the arrangement of the Pareto dominance.Major similarities between the population members of eachsub division are calculated against the Pareto front, and theresulting class, along with its similarity measure, is thenutilized to promote a solutions front that is non-dominated.Non-dominated Sorting Genetic Algorithm-II (NSGA-II) is awidely used computationally fast and elitist MOGA approach[16]. NSGA-II variants include the following:

• Steady State NSGA-II (ssNSGAII): A steady state versionof the generational genetic algorithm.

• Parallel NSGAII (pNSGA-II): A variation of the NSGA-II wherein the algorithm takes advantage of the differentcores of a computer in order to evaluate the individualsof a function in parallel.

• Random NSGAII (NSGA-IIr): NSGA-IIr is a variation ofNSGA-II wherein the three basic operations includingevolution, mutation and crossover are selected using arandom seed to generate new individuals.

• Adaptive NSGA-II (NSGA-IIa): This approach works onthe same principles as NSGA-IIr but the operators areadaptively selected through a random process.

5

C. Other Types of GAs

1) Adaptive GA: To prevent GAs from suboptimally con-verging to a local optima, a number of adaptive techniqueshave been recommended for adjusting parameters such ascrossover probability, mutation probability, and populationsize. These techniques constitute adaptive GAs. In contrast tosimple GAs that rely on fixed control parameters, adaptiveGAs utilize information about the population in each gen-eration to adaptively adjust the probability of crossover andmutation.

2) Parallel GA: Parallel GA (PGA) can be categorizedinto three subcategories: (i) master-slave GA; (ii) island GA(also referred to as distributed GA or coarse-grained GA); and(iii) cellular GA (also called fine-grained GA). Amongst thesethree configurations, the Island PGAs (island GA, or iGA, forshort) are best suited to wireless networking since they allowcontrol of the migration policy unlike the cellular PGAs—which require local communication at all agents during eachgeneration of the PGA—and the master-slave PGAs thatrequire the master and slaves to constantly communicate.

3) Distributed or Island GA: Distributed GA divide a largepopulation into several smaller subpopulations that interactweakly. Parallel GAs, on the other hand, are primed to speedup execution by implementing the sequential GA on severalcomputation engines. Amongst the distributed GA approaches,the island GA approach is important and is popularly usedin wireless networks [17] [18]. It works by separating thepopulation into sub-populations called islands. These islandscooperate with other islands through the migration of indi-viduals. In island GA, the GA runs on each node with thepossibility of good solutions migrating between nodes. Thesemigrations are performed in accordance with the migrationpolicy that controls the ‘when’ and ‘how’ of the individualmovement. In the island GAs, there is relatively infrequentmigration between island models thereby implying reducedinter-agent communication requirements compared to master-slave GA and cellular GA [19].

Fig. 2: Pareto front illustrated for two objectives

D. GA Components And Operators

A flow chart of a typical GA is provided in Figure 3. Theoperation of a GA-based framework starts by initially defininga population of candidate solutions. While it is acceptable,

Fig. 3: A flow chart of a typical genetic algorithm [20]

and indeed commonplace, to define the population randomly,the convergence process can be accelerated by starting withan appropriately defined population. Each candidate solutiondefined in the population is known as an individual, witheach individual encoded in an abstract format named as achromosome (in a metaphorical reference to the encoding roleof chromosomes in biological genetics). Various evolutionarytechniques (mutation, crossover, etc.) are then applied onthe population. In every generation, multiple individuals arestochastically selected from the current population with fitterindividuals being more likely selections (due to the geneticprinciple of ‘survival of the fittest’). The selected individualsare then genetically modified (mutated or recombined) toform the next generation of the population. The usage ofgenetic operators and stochastic selection allow a gradualimprovement in the ‘fitness’ of the solution and allow GAsto keep away from local optima.

In the following subsections, we will provide more detailsabout the various steps involved in problem solving throughGAs.

1) Encoding into Chromosomes: Genetic algorithms oper-ate by initially defining a population of candidate solutions(each of which is called an individual). Individuals are encodedin an abstract representation known as a chromosome (whichmay be problem specific although representation in strings of1s and 0s is common; in the related field of ‘genetic pro-gramming’, the encoding is in terms of snippets of computercode). The population size can be a significant factor in GAperformance and efficiency. In particular, the performance ofGAs can poor when the population size is very small [22].

The encoding of candidate solutions in the form of chro-mosomes can also be of variable length. Such variable lengthencoding is well suited to problems such as shortest-pathrouting problem that compares paths of different lengths. Asan example of variable length chromosome encoding, referto the work of Ahn et al. [23] in which the routing path isencoded in the chromosome as a sequence of positive integersrepresenting node IDs of the nodes enroute the end-to-endrouting path. Each locus of the chromosome corresponds to

6

(a) Binary Encoding. Most common encoding type in which eachchromosome is represented using a binary string.

(b) Permutation Encoding. Useful for ordering problems in which eachchromosome represents position in a sequence.

(c) Value Encoding. Chromosome are strings of somevalues (which can be form numbers, real numbers, chars,or some other complicated object).

(d) Tree Encoding. Every chromosome is a tree of someobjects (such as functions or commands in programminglanguage).

Fig. 4: Illustration of basic types of encoding schemes usedby GA (Adapted from [21]).

the order of the node in the end-to-end routing path with thesource node always represented by the gene of first locus. Thelength of the chromosome is variable but is upper bounded byN (the total number of network nodes).

Several encoding schemes for GA exist including:

a) Binary Encoding: Binary encoding is the most commonapproach, mainly because initial GA-based works usedthis encoding. In binary encoding every chromosome isa string of bits, i.e. 0 or 1. Binary encoding providesmany possible chromosomes even for a small amountof alleles. On the other hand, this encoding is in somecases not natural for many problems and in other casescorrections need to be made after mutation or crossover.

b) Octal Encoding: The chromosome is represented in octalnumbers (0-7).

c) Hexadecimal Encoding: The chromosomes are repre-sented using hexadecimal numbers. (0-9, A-F).

d) Permutation Encoding: Permutation encoding can beused in GA for ordering problems, such as task orderingproblems, or the ‘traveling salesman’ problem. In per-mutation encoding, each chromosome is a representedas a string of numbers, which is actually a number in asequence. In some problems, for some types of mutationand crossover, corrections must be inculcated to leavethe chromosome in a consistent form, viz. it should becorrected to ensure that it has a real sequence in it.

e) Value Encoding: Direct value encoding is useful inproblems that deal with complicated values (such as areal number). Use of binary encoding method in thiscase would be very difficult. In value encoding, everychromosome is a string of some values. Values can bein any arbitrary form including numbers, characters, orany other complicated object. While value encoding is agood approach for some problems, it often requires newcrossover and mutation operators defined specifically forthe problem.

f) Tree Encoding: Tree encoding is used for expressions orevolving programs, specifically for genetic programming(GP). In tree encoding, every chromosome or gene is ina tree form of some objects, like commands in program-ming language or a function. Programming language likeLISP is mostly used for this, because genetic programsin it are encoded in this form and can be easily parsedas a tree structure, thereby facilitating mutation andcrossover to be performed relatively easily.

For illustrative purposes, an example binary encoding,permutation encoding, value encoding, and tree encoding ispresented in Figure 4.

Encoding example: Consider an example encoding of radioparameters in chromosomes and genes for GAs. The inherentdifficulty in such a representation is the large relative differ-ence between radio parameters. For example, frequency rangeof 1 MHz to 6 GHz with step size of 1 Hz would requiremore than 30 bit genes, while modulation has only handfulof techniques so it would require far less bits. Chromosomescan, therefore, utilize large number of bits for frequencywhile utilizing small number of bits for the modulation gene.To make GAs independent of which radio it is optimizing,the variable bit chromosomes representation is suggested byScaperoth et al. [24] in the context of software defined radios(SDR).

2) Fitness Function: The fitness function, sometimes calledthe objective function, offers a mechanism for the evaluationof the individual chromosomes. Fitness function can be singleobjective, or multi-objective. The fitness function is devisedin a way that the individual chromosome is its variableinput parameter. Fitness function ensures that the chromo-somes passed on to the next generation are not violatingany constraints. It assigns a fitness score to each individualdepending on the quality of the individual in terms of theoptimization goals and objectives. In particular, for the fieldsof both genetic algorithms and genetic programming, eachdesign solution is represented as a string of numbers referredto as the chromosome. In fitness evaluation, after each roundof completed simulation or testing, N worst design solutionsare deleted, and N new ones from the best design solutionsare created. Each design solution thus needs to be awarded afigure of merit, which indicates how close it came to meet theoverall specification. This is achieved by applying the fitnessfunction to the simulation or test results obtained from thatsolution.

Genetic algorithms require much effort in designing a work-

7

able fitness function. Though a computer might generate thefinal form of the algorithm, it takes a human expert to definethe fitness function since an inefficient function can result inconvergence to an inappropriate solution, or no convergence atall. Speed of execution of the fitness function is very important,since a typical GA needs to be iterated many times in orderto produce a usable result for a non-trivial problem. Hence,fitness function requirements include:

• Minimum fitness computation time of candidate solution.• Precise model definition for fitness computation.• Removal of noise from the fitness function values.Two main categories of fitness functions exist: one where

the fitness function does not change over time (as in thecase of optimizing a fixed function, or testing with a fixedset of test cases) and the other where the fitness function ismutable (as in the case of co-evolving the set of test casesor in niche differentiation). Another way of defining fitnessfunctions is in terms of a fitness landscape that depicts thefitness level for each possible chromosome. Definition of thefitness function may not be straightforward in many cases. Insome scenarios, it becomes very hard or impossible to come upeven with a guess of what fitness function definition might be.This difficulty is solved by ‘interactive GA’, which outsourcesevaluation to external agents including humans. The functionis calculated iteratively if the fittest solutions produced by GAare not in the desired range.

There are various strategies for assignment of fitness valuesto the subject population. The ‘aggregation-based’ strategysums the objectives into a single parameterized objectivevalue, e.g., an aggregation of weighted-sum. The ‘criterion-based’ method transitions between the objectives during theselection phase of the algorithm. Every time an individual ischosen for reproduction, a potentially different objective willdictate which member of the population will be copied intothe mating pool. In the ‘Pareto dominance based’ methods,different approaches like dominance depth, dominance count,and dominance rank are used. The dominance rank dictatesthe number of individuals by which certain individual isdominated. In the dominance depth, the population is dividedinto several fronts and the dominance depth reflects the frontto which an individual belongs. The dominance count isthe number of individuals that are dominated by a certainindividual.

Fitness function example: The evaluation of the solution isat the heart of GAs since it is on the basis of fitness valuesthat GAs evolve solutions and remove poor individuals. It canbe thought of as the performance metric for the optimizationalgorithm. As an example, for the routing problem, the fitnesscan be equated with high average throughput (fi = ti:where fi is the fitness of ith individual and ti is its averagethroughput), low delay (fi = −di) where di is delay perceivedby the ith individual. For classification problems, the fitnessfunction can be some measure of the error. Realistic fitnessfunctions are typically much more complex than this toyexample, and can depend on many factors.

3) Mutation: The mutation operator is used for explorationwith randomly altering genes in chromosomes to discover

new horizons and new traits. Mutation helps to diversify thepopulation, thus helping GAs to avoid local optima and pavethe way towards global optima. Several types of mutationtechniques exist including:

• Bit-string Mutation: The mutation of bit strings occurs,through bit flips at random positions of chromosomes (asillustrated in Figure 5).

• Flip Bit: This mutation operator takes the chosen genomechromosome and inverts the bits, i.e. if the genome bit is0, it is changed to 1 and vice versa.

• Boundary: This mutation operator replaces the chromo-some with either upper or lower bound in a random man-ner. This can be used for float and integer representationof chromosomes.

• Non-Uniform: The probability that the extent of mutationwill go down to 0 with the next generation of chromo-some is increased by using non-uniform mutation. Thiskeeps the population from stagnating at early stages ofthe evolution process. It also alters solution in later stagesof evolution. The mutation operator can only be used forfloat and integer chromosome genes.

• Uniform: This operator replaces the score of the chosenchromosome with a uniform random value selected in therange of the user-specified upper and lower bounds. Thismutation operator can only be used for integer and floatgenes.

• Gaussian: This operator adds a unit Gaussian distributedrandom value to the selected chromosome. If it fallsoutside of the boundary of user-specified upper or lowerbounds for that chromosome, the new value is clipped.This mutation operator is used for integer and floatstrings.

Fig. 5: Illustration of mutation and crossover operation.

4) Crossover: The crossover process is essentially an at-tempt to exploit the best traits of the current chromosomes andto mix them in a bid to improve their fitness. This operatorrandomly chooses a locus and exchanges the sub-sequencesbefore and after that locus between two parent chromosomesto create a pair of offspring. A crossover point is randomlychosen. There can be multiple crossover points to aid moreexploration. In biological systems, crossover is much morerampantly observed than mutation, often as much as a milliontimes more frequent [25]. Crossover along with mutationprovides the necessary ‘evolutionary mix of small steps andoccasional wild gambles’ to facilitate robust search in complexsolution spaces [26].

In GAs, crossover is viewed as synthesis of best practicesand mutation is viewed as a method of spontaneous inspirationand creativity. The evolution can start from a populationof completely random individuals and can evolve to better

8

solutions through survival of the fittest after application of ge-netic operators. The selected individuals (or chromosomes) aregenetically modified (mutated or recombined) to form the nextgeneration of the population. The usage of genetic operatorsand stochastic selection allow a gradual improvement in thefitness of the solution and allow GAs to keep away from localoptima.

Fig. 6: Illustration of examples of one point, two points, anduniform crossover methods (Adapted from [27])

Several methods exist for crossover of chromosomes whichuse different data structures [6].

• One-Point Crossover: A single crossover point on bothparent’s chromosome strings is selected. All data beyondthat point, in either organism chromosome, is swappedbetween the two parent strings. The resulting strings arethe new children chromosome.

• Two-Point Crossover: Two-point crossover requires twopoints to be selected on the parent chromosome strings.Everything between the two points is swapped betweenthe parent chromosome strings, producing two childchromosomes.

• Uniform Crossover and Half Uniform Crossover: Theuniform crossover approach uses a fixed mixing ratiobetween two parents strings. Unlike one-point and two-point crossover, this approach enables the parent chro-mosomes to contribute at the whole chromosome genelevel rather than at the segment level. For example ifthe mixing ratio is 0.5, the offspring chromosome hasapproximately half of the genes from first parent and theother half from other parent, although cross over pointscan also be randomly chosen instead. In the half uniformcrossover method, exactly half of the non-matching bitsare swapped between the chromosomes. First the numberof differing bits, or the Hamming distance, is calculated.This number is then divided by two; the resulting numberdepicts the number of bits that do not match between thetwo parents, and will thus be swapped. An illustrativeexample comparing one point, two points, and uniformcrossover is presented in Figure 6.

• Cut and Splice: In the crossover variant called the ‘cut

and splice’ approach, a change in length of the childrenstrings occurs. The reason for this difference in length isthat each parent string has a separate choice of decidingcrossover point.

• Three Parent Crossover: In this method, the child chro-mosome is derived from three parents which can berandomly chosen. Each bit of first parent is comparedwith bit of second parent to see if they are the same. Ifthe bits are same then that corresponding bit is taken forthe offspring otherwise the bit from the third parent istaken for the offspring chromosome.

• Ordered Chromosome Crossover: In some cases, a directswap may not be possible. One such case is whenthe chromosomes are in an ordered list. The N-pointcrossover can be used for ordered chromosomes, butalways requires a corresponding repair process. Some ofthe approaches for ordered pair crossover include cyclecrossover, position based crossover, alternating positioncrossover, edge recombination crossover, sequential con-structive crossover and voting recombination crossover.

5) Selection: Selection in GA allows individual genomesto be chosen from a population for later recombination orcrossover (breeding). This operator performs the selection ofchromosomes from population for reproduction. The fitter thechromosomes, the more likely they are to be passed on to thenext generation. A number of selection techniques has beenemployed in GAs to carry out the survival of the fittest froma pool of individuals [28].

A common selection procedure may be implemented in thefollowing way:

• The fitness function is evaluated for each individual andis then normalized for algorithm use. The fitness value ofeach individual chromosome is typically normalized bythe aggregation of all fitness values to ensure that all theresulting fitness values sums up to 1.

• The population is then sorted in descending order offitness values.

• Accumulated normalized fitness scores are calculated.Accumulated fitness value of an individual equals the sumof fitness scores of all the previous individuals and itsown fitness score. Accumulated fitness of the previousindividual must always be 1—Any other value can beused as indication of a wrong normalization procedure.

• A random number R is chosen in the range 0 to 1. Thefirst individual whose summed up normalized value isgreater than the number R is selected.

This selection method is called fitness proportionate selec-tion or roulette-wheel selection, since the procedure is repeateduntil there are enough selected individuals. If instead of using asingle pointer that is spun a number of times, multiple equallyspaced pointers are used on a wheel that is spun once, the se-lection process will be called ‘stochastic universal sampling’.The selection process will be a tournament selection processonce there is repeated selection of the best-fitted individualfrom a randomly chosen subset pool. This will be furthercalled truncation selection once we take the best half, thirdor some other proportion of the individuals.

9

There are also other selection mechanisms that do not takeinto account all individuals for the selection process, but onlythose that have a fitness value, greater than a given constant.Other algorithms select the individual from a restricted poolwhere only a constrained percentage of the individuals areallowed, based on their fitness scores. Retaining the bestindividual chromosomes in a generation unaltered, in the nextset of generation, is called elitist selection or elitism. It isconsidered a successful variant of the general procedure ofconstructing a new population pool.

The main selection techniques used by GAs are summarizedin the following list.

• Relative Tournament Evaluation: In relative tournament,two parents are compared randomly and the one with highfitness value is selected to be the parent of next gener-ation. Winning chromosomes are awarded by adjustingtheir weight corresponding to the objective function. Thisweighting increases the importance of that chromosomesrelative to others.

• Roulette Wheel Selection: Roulette wheel selection is away of choosing mother and father chromosomes fromthe population in a way that is proportional to theirfitness. The chromosomes with higher fitness have highchances of being selected using this technique.

• Relative Pooling Tournament Evaluation: In this tech-nique, each parent is thrown in a competition with theindividual chromosomes in population, independently.The individual that is the winner of maximum number oftournaments gets a chance to be passed into next genera-tion through reproduction. This technique is comparableto the proposed method in Horn et al. [29] of fightingtwo individual against various subsets of population. Thistechnique is computationally intensive, and thereby canstress GA deployment due to the computational overhead.

• Elitism: Elitism is a method to store and use previouslyfound best-suited solutions in subsequent population gen-erations of GA. In a GA approach that uses elitism, thestatistic related to the population’s best solutions doesnot degrade with generation. Maintenance of archivesfor non-dominated solutions is an important and criticalissue. The final contents of archive usually represent theresult returned by applying the optimization process. It ishighly effective and common to employ the archive as apool, from which guidance is taken for the generation ofnew solutions. Some algorithms use solutions availablein the archive exclusively for such purpose, while otherstend to rely on the archive to varying degrees accord-ing to settings of related parameters. The computationalcomplexity of maintaining the archive which involveschecking for non-dominance of newly generated solutionssuggests relatively modest size of bounded archive [30].

III. APPLICATIONS OF GAS IN WIRELESS NETWORKING

After providing background on GAs in the last section,we are now in a position to survey the applications of GAsin wireless networking. GAs have been applied in wireless

TABLE II: Applications of genetic algorithms in wirelessnetworking (covered in Section III-A)

Application Discussed in

Routing Section III-A1 and Table IIIQuality of Service (QoS) Section III-A2 and Table IVLoad balancing Section III-A3 and Table VBandwidth allocation Section III-A4 and Table VI.Channel allocation & assignment Section III-A5 and Table VII.Localization Section III-A6 and Table VIIIWireless AP placement Section III-A7 and Table IXGeneral Applications Section III-A8 and Table X

networking in a wide variety of contexts. We will firstlydescribe the various applications of GA in wireless networkingin Section III-A. We will then provide a network-type basedclassification of GA application in Section III-B. Finally, wewill discuss ‘hybridization’ proposals that combine GAs withother techniques in Section III-C.

A. Application Based ClassificationIn this section, we will discuss the applications of GAs in

wireless networking. In particular, we will discuss routing,QoS, load balancing, channel allocation and assignment, local-ization, WLAN AP placement, and other general applications,in this particular order. The organization of this section issummarized in the Table II.

1) Routing: The problem of shortest path routing problemis a well-studied problem in literature. GAs are well suitedto the routing problem allowing convenient mechanisms fordeveloping adaptive routing solutions. GAs have been used fordeveloping routing solutions in a variety of wireless settingsas summarized in Table III. Ahn et al. [23] have tackled thisproblem through GAs. The routing table for each source-destination pair needs to be constructed first. Obviously, thenumber of possible routes between two nodes heavily dependson the network topology. If the network is densely connectedor the size of the network is large, the number of possibleroutes of a source-destination pair becomes huge. Hence, it isimpossible to list all the possible paths in the routing table. Apredefined number of routes are randomly generated and fed topopulation pool, and chromosomes are then optimized. In thesimulation, crossover and mutation are run on variable lengthchromosomes. Infeasible solutions are also kept in check. Thepaper discussed the issues of path-oriented encoding, and path-based crossover and mutation, which are relevant to the routingproblem.

Yang and Wang [34] studied how GAs can be used fordynamic shortest path routing problem (DSPRP). Algorithmslike Bellman-Ford and Dijkstra perform well in fixed infras-tructure but on dynamic scale, high computational complexityrenders them unacceptable. A DSPRP model is built up inthis paper using GAs. A specialized GA is designed for theshortest path problem in MANETs. Several immigrants andmemory schemes that have been developed for GAs for gen-eral dynamic optimization problems are adapted and integratedinto the specialized GA to solve the DSPRP in MANETs.The experimental results indicate that both immigrants and

10

TABLE III: Using genetic algorithms for routing in wireless networks

Project &reference

Network No. ofobjectives

Technique used Brief summary

Ahn and Ramakr-ishna [23]

MANET SOGA Integer coded, variable lengthchromosomes

Unlike Dijkstra and Depth-First Search (DFS) algorithms, multiplesolutions can be found for shortest path routing that have same costusing GAs.

Cheng et al. [31] MANET MOGA Distributed multipopulation GA A distributed GA approach incorporating multi-population with randomimmigrant schemes solves the shortest path routing in MANET.

Gen et al. [32] LANs SOGA Integer coded GA for randomdistributions

A priority encoding scheme is proposed using GAs to represent andoptimize all possible paths in network graph.

Davies et al. [33] WAN SOGA Simple GA with unbiased cod-ing scheme

A GA approach continuously re-routes and assigns weight to trafficpaths to implement adaptive control over dynamic routes and networktopology.

Yang et al. [34] MANET SOGA GA with immigrants and mem-ory scheme

A GA proposed for investigating effectiveness and efficiency of GAswith immigrants and memory schemes in solving the dynamic shortestpath routing problem in the real-world networks.

Badia et al. [35] WMNs SOGA ILP formulation with GAs GA addressed the routing and link scheduling problem when ILP cannotcope with computational complexity of large WMNs.

Al-Karaki et al.[36]

WSN MOGA Centralized GA and ILP A metaheuristic algorithm that solves the joint problem of optimalrouting with data aggregation to optimize the network lifetime whileretaining the energy and latency efficiency.

Lorenzo and Gal-isic. [37]

CN MOGA Sequential GA A sequential GA solves the NP-hard problem by optimizing relaytopology while eschewing the inter-cell interference.

Apetroaei et al.[38]

WSN SOGA Tree coded GA on spanning treetopology

GA inspired spanning tree topology for WSNs that adapts (accordingto the nodes, residual energy) to maximize the usage of the network.

Yen et al. [39] MANET SOGA Prüfer number encoding A GA strategy for energy efficient routing by maintaining the energybalanced nodes in entire MANET distribution.

Chiang et al. [40] MANET SOGA Tree sequenced coding A GA based on sequence and topology encoding employed topologycrossovers to solve the multicast routing constraints efficiently.

memory schemes enhance the performance of GAs for theDSPRP in MANETs.

In MANETs, all the nodes communicate over wirelesslinks without any established infrastructure. MANETs are wellsuited to environments in which there is no fixed infrastructureor when the infrastructure is not trusted. In such networks, acommon strategy is to form clusters where each node is linkedto a clusterhead for efficient routing with other nodes that arenot in its direct range. GAs have been used in such cluster-based routing schemes for MANETs. Al Gazal et al. [41]have proposed a GA-based protocol named ‘cluster gatewayswitch routing protocol’ (CGSRP) to select the clusterheadto carry out communication between nodes. Clusterhead musthave enough resources, power, and bandwidth to avoid dangersof bottlenecks. This scheme works by encoding each node’sunique ID in the chromosomes. The chromosomes have infor-mation about cluster-head, members, number of links in eachcluster-head. The encoded chromosomes are then evaluatedagainst certain criteria as defined by the fitness function(which may incorporate features such as load balancing andbandwidth conservation). Each chromosome’s fitness is thenevaluated. The process of the survival of the fittest leads toan optimum selection of nodes as cluster-heads that optimallyutilize resources.

GAs have also been proposed for use in multicast andbroadcast routing in wireless networks. A study on multicastroutes optimization using GAs is done by Banerjee and Das[42]. An on demand routing scheme with GAs is discussedto meet bandwidth, delay constraints while maintaining theQoS in wireless networks. Chromosomes from different poolsare randomly concatenated to obtain a multicast solution. Indoing so, all possible routes between source and destination

nodes are encoded and fed to GA in the form of populationchromosomes. GA then uses evolutionary techniques to selectthe best chromosomes by evaluating them against a partic-ular fitness function, which can be in the form of differentconstraints related to bandwidth, QoS, and channel allocation,etc. In another work, Salcedo-Sanz et al. [43] proposed mixedneural-genetic algorithm for the ‘broadcast scheduling prob-lem’ (BSP). The algorithm solves this optimization problem intwo stages: firstly, it finds a feasible solution, and secondly itobtains maximal transmission throughput by employing fittestchromosomes through GAs. A track of interference and framematrix are kept and the channel slots are assigned accordingly.Nodes that are 1-hop away are not assigned the same slotto avoid mutual interference. After finding feasible solutions,these solutions are encoded in chromosomes. Multiple GAoperators—such as mutation, crossovers—are then appliedleading to maximized throughput.

The use of hybrid GA algorithms have been proposed forrouting in wireless networks. Cauvery et al. [44] proposedusing ant algorithms, combined with GAs, for route discovery.It was observed that the use of ant algorithm leads to reductionin the size of the routing table. We note here that ant-colony-optimization (ACO) algorithms belong to the broaderclass of AI known as swarm intelligence. While GAs cannotuse global information of the network, the use of ACO andGA techniques in tandem leads to independent explorationof the network helping in effective computation of routesbetween any pair of nodes. The proposed algorithm creates aninitial population, determines the forward-ants, generates newpopulations using genetic operators and fitness evaluation, andfinally updates routing table. The hybrid algorithm performswell in not only generating routes between pair of nodes but

11

also for achieving efficient load balancing.

2) Quality of Service (QoS): With increasing prevalenceof multimedia applications (such as video conferencing andstreaming), and the development of high-speed wireless com-munication technologies, efficient and reliable QoS provision-ing has become increasingly important. Modern multimediaapplications place rigorous QoS demands for different param-eters like bandwidth, delay, and jitter [71]. This motivates thedevelopment of effective resource allocation schemes that canensure QoS while also satisfying these strict QoS constraints.In this section, we will describe the various GA-based QoS so-lutions that have been proposed in literature. This informationis presented in summarized form in Table IV.

Various analytic approaches have been proposed in literatureincluding those that rely on Semi-Markov decision process(SMDP) theory for optimal call admission control (CAC)[72]. These methods unfortunately fail to scale to large-scalenetworks where the decision and action space is large. GAscan provide near optimal solutions in large-scale complexnetworks. Xiao et al. [73] presented an effective near optimalGA-based CAC approach for large-scale wireless/ mobilenetworks. The chromosomes encoded are binary strings where1 and 0 represent the acceptance and rejection of calls by CAC.These chromosomes are randomly generated and then evalu-ated against fitness function of channel utility and constraints.Evolution of best traits leads to near optimal solutions forQoS. This scheme offers adaptive QoS to multi-class userswhile being computationally less intensive compared to theSMDP approach.

A multi-objective GA (MOGA) approach on multicast QoSrouting is proposed by Roy et al. [49]. To deliver on-demandQoS, different solutions are kept in an optimized pool. Multi-ple solutions are offered depending upon the service demands.A MOGA approach optimizes the prioritized objectives asstated in the user QoS requirements. A similar approach forQoS aware web-based service under strict constraints andoptimizing the targeted QoS parameters is done by Canforaet al. [51]. It is shown in this study that the widely adoptedoptimization approach of linear integer programming is unableto offer non-linear QoS functions. GAs can be employed towork with non-linear QoS attributes. In yet another MOGAwork, Roy et al. have proposed QM2RP as a QoS-basedmulticast routing protocol framework for mobile wirelesscellular networks [56]. The proposed protocol determines near-optimal routes on demand while working with potentiallyinaccurate network state information. The protocol considersa number of QoS optimization metrics including end-to-enddelay, bandwidth, and residual capacity.

With the increase in number of service candidates, thetraditional chromosome model for QoS optimization becomescumbersome in web services. To address this, Gao et al.[74] studied a ‘tree based GA’ for optimizing QoS. Thismodel is faster than traditional GA since it offers impulsivechromosomes coding and decoding and present run timeintegration of QoS resources. Chromosomes are representedin series of nodes and all the leaf nodes execute the process.A tree characterizes the structured model for different QoS

composition, and requires less computation for large networks.Sub-tree crossover exchanges the task nodes and updatesthe QoS vector from the leaf nodes to the tree’s root. Astandardized and comprehensive fitness function verifies thewhole QoS process plan. Results have showed that the treebased models have lesser overhead than the traditional GAswith serial chromosomes while offering excellent replanningand service composition.

3) Load Balancing: GAs have been proposed for use invarious load balancing solutions for various wireless network-ing configurations in literature (as shown in Table V).

There are three main types of load balancing categories:(i) strongest ‘received signal strength indicator’ (RSSI); (ii)least loaded first (LLF); (iii) hybrid load balancing (HLB)incorporating RSSI & LLB. These techniques fall under thecategory of network-centric load balancing, and are applied ona network scale to achieve global optima, unlike user-centeredtechniques that are more likely to achieve local optima only.Chromosomes consist of randomly assigned individuals toan access point. Crossover and mutation are performed onindividuals and fitness evaluation is done. Both load balancingand congestion control problems have been addressed. QoShas also been addressed by assigning different weight to users,depending on its service requirement like streaming and mediaetc. Experimental results and analysis showed that microGAoutperforms macroGA at first but in the long run macroGAdoes better load balancing by pushing more congested userstowards peripheral APs.

Ozugar et al. [82] proposed the multi-objective hierarchicalmobility management optimization for UMTS networks todistribute the load among different parts of network suchas serving GPRS nodes (SGSN), radio network controllers(RNC), and mobile switching centers (MSC) in addition toperforming the role of minimizing cost for location update inthese intricate networks.

Selecting resource-rich nodes for trivial tasks creates animbalance in energy distribution among nodes resulting in thedemise of multicast service. To mitigate such a possibility,Yen et al. [39] have proposed an energy-efficient GA routingin MANETs. The algorithms proposed in [39] continuouslyobserve the condition at various nodes and improve the en-ergy balance by replacing the weaker nodes with energy-richnodes. An extended Prüfer encoding scheme is proposed forenergy efficient exchange of leaf nodes through mutation andcrossover. In this way a substantial reduction in convergencetime is achieved, thus prolonging the service duration effi-ciently.

GAs have also been proposed for load balancing at APs inIEEE 802.11 WLANs for efficient distribution of bandwidthamong users. Scully and Brown [70] have discussed MicroGAsand MacroGAs to address the problem of load balancing.A comparison of MicroGAs and MacroGAs has been doneby Kumar et al. [83]. MicroGAs has a population of fiveindividuals leading to brisk evolution.

4) Bandwidth Allocation: The problem of bandwidth allo-cation under QoS constraints in wireless networks is typicallya complex task. A proper bandwidth allocation scheme must be

12

TABLE IV: Using genetic algorithms for QoS optimization in wireless networks

Reference Goal No. ofobjectives

Technique used Brief summary

Resources Scheduling for QoSSherif et al.[45]

Resources allocation withcall admission control

SOGA GA with reallocation ofunused bandwidth

Accelerate allocation through crossover and mutation. Reducescomputation of NP-hard problem. Dynamic assignment of net-work resources and bandwidth

Gozupek andAlagoz [46]

Scheduling under interfer-ence constraints

MOGA Binary encoding withrandom seeding

GA based sub-optimal schedulers for maximizing the throughputand improving the latency in CRNs.

Cheng andZhuang [47]

Packet level resource allo-cation

SOGA Karush-Kuhn Tucker(KKT) and GA drivenapproach

An effectual hybrid GA optimizes the QoS provisioning andresource allocation to keep a balance between performance andcomputation.

Lopez et al.[48]

Transmit power controlfor maximum utility

MOGA GA with call admissioncontrol

MOGA approach adopted to maintain the transmit power of CRs,while maintaining the QoS and optimizing the spectrum access.

Roy et al. [49] Energy efficient resourcescheduling

MOGA Depth first search withNSGA

On demand QoS routing scheme has been proposed based on enduser needs. A solution meeting the constraints is offered from anoptimized pool of chromosomes.

QoS for Multimedia and Web-ServicesJiang et al. [50] Multi-Constrained QoS

for web-servicesSOGA Variable length chromo-

some GAA scalable QoS aware web service composition is introducedusing GA for diversified functionality through cut and spliceoperation of mutation and crossover.

Canfora et al.[51]

QoS service composition SOGA GA with ILP and elitism GA approach for web based QoS services with non-linear aggre-gate functions while meeting a set of constraints.

Li et al. [52] Video On Demand MOGA GA with interferencemodel

GA is employed to leverage multi-source, multipath solution todesign a concurrent video streaming environment in WMNs, alongwith route discovery.

Energy-Efficient QoS Multicast RoutingEkabtaniFard etal. [53]

Energy Efficient QoSrouting in clusters

MOGA NSGA-II A hierarchical energy efficient routing protocol, for two tier WSNnetworks, under tight QoS constraints, is developed using MOGAoptimization with NSGA-II. [54].

GAMAN Topological Multicasting SOGA Tree coded GA withmulti-population

A GA for MANETs (GAMAN) is proposed by Barolli et al. [55].The algorithm works locally on smaller population and offerspromising QoS for ad-hoc networks.

QM2RP [56] QoS-based multicast rout-ing

MOGA Tournament selectionand niched Paretooptimization

A QoS-based mobile multicast routing protocol using multi-objective genetic algorithm.

Cui et al. [57] Optimizing conflictingQoS parameters

MOGA Depth-first search algo-rithm and GA

A MOGA approach is used to optimize multiple multicast routeslying on Pareto front.

Lu et al. [58] Energy efficient QoS mul-ticast routing

MOGA Tree structured coding Proposed GA solves QoS by constructing a multicast tree with aminimum energy cost and bounded end-to-end delay.

Yen et al. [59] Multicast routing throughflooding

MOGA Tree coded flooded GA A GA based method is proposed for multi-constrained QoSmulticast routing, while incorporating elitism, to optimize thecomputation time.

Huang and Liu[60]

QoS Aware MulticastRouting

MOGA Greedy and family com-petition with GA

A hybrid GA along with greedy algorithm mitigate prematureconvergence problems to avoid local optimas.

Sun et al. [61] Multi-constrainedoptimization

MOGA Spanning tree encoding The evaluated GA can optimize the cost of the multicast tree,maximum link utilization, the average delay selection of the long-life path.

Abdullah [62] Modified QoS RoutingProtocol

MOGA Variable length encoding A GA approach look for the best QoS route, using route mutation,to optimize the design of MANET routing protocol by improvingpacket delay and throughput.

Route DiscoveryGelenbe et al.[63]

Autonomous route discov-ery with GA

SOGA GA with distributedCPN protocol

Proposed autonomous route discovery in cognitive packet net-works (CPNs) with GAs. Paths discovered by CPN are provided toGA which then produces new routes using evolutionary techniques

Cauvery andVisvanatha [44]

Route discovery throughmobile agents

SOGA GA with Ant algorithm A hybrid algorithm (combining ant algorithms and GA) is pro-posed to reduce the size of routing table and help in routing amongautonomous peers.

Queue Management at Routers (AQM)Fatta et al. [64] Designing FPI controller

for AQMSOGA GA with fuzzy logic Fuzzy proportional integer (FPI) controller is purposed and FPI

parameters are controlled by GAs dynamically to optimize activequeue management (AQM) policies.

Selamat et al.[65]

Query Retrieval for activeAQM

SOGA Integer coded distributedGA

A GA with mobile agents search system (MaSS) is proposed tominimize the query retrieval cost while maintaining a reasonablepath delay.

13

TABLE V: Using genetic algorithms for load balancing in wireless networks

Reference No. ofobjectives

Technique used Brief summary

Wireless Mesh NetworksHsu et al. [66] SOGA GA with Dijkstra algorithm A hybrid GA coupled with Dijkstra’s algorithm look for the best available paths with

bounded delays and low cost network configuration.Zeng et al. [67] SOGA GA with greedy algorithm and bi-

nary encodingLoad balanced disjoint clustering is done by a GA coupled with greedy algorithmfor efficient distribution of network resources.

Mobile Ad-hoc NetworksYen et al. [39] MOGA GA with Prüfer number coding In time-critical load balancing scenario, GAs inspired strategy is applied in multicast

tree to balance the energy consumption, leading to prolonged network lifetime.Cheng et al.[68]

SOGA Distributed GA with multi-population and immigrant schemes

A dynamic GA evaluates the fitness of each feasible clustering structure against loadbalancing metric.

Wireless Sensor NetworksHe et al. [69] SOGA 1-to-1 mapping of nodes to genes

in chromosomesThe application of GA to find a load-balanced connected dominating set (CDS) tobe used as a virtual backbone of WSNs.

Wireless Local Area NetworksScully andBrown [70]

SOGA MicroGA & MacroGA A comparison of MicroGA and MacroGA is done against QoS attributes likethroughput and load balancing. MicroGA performs better in brisk load balancingunder time constraints.

TABLE VI: Using genetic algorithms for bandwidth allocation in wireless networks

Reference No. ofobjectives

Techniques used Brief summary

Lin et al. [75] MOGA GA with adaptive inheri-tance probability

A prioritized bandwidth allocation scheme for transmitting medical patients data isapplied for multimedia sub-streams.

Kandavanam et al.[76]

MOGA GA with variable neigh-borhood search approach

A hybrid GA optimizes the residual bandwidth allocation to routing links for multiclasstraffic in all networks.

Karabudak et al. [77] SOGA GA with Markov DecisionModel and Elitism

The proposed GA scheme minimizes handoff latency by using elitism technique toallocate the maximum network resources to admitted calls.

Vendanthan et al. [78] SOGA GA with greedy algo-rithms

A GA approach to efficiently utilize the bandwidth for different incoming call streamsto maximize the revenue generation.

Riedl [79] SOGA GA with local searchheuristics

GA approach with bandwidth and delay metrics optimizes the IP traffic engineering inwireless networks.

Kobayashi et al. [80] MOGA Distributed GA A distributed GA works by utilizing both local and global optimization in parallel onnetworks nodes to utilize the network resources and bandwidth.

Qahtani et al. [81] SOGA GA for asynchronoustransfer mode (ATM)

A genetic routing algorithm (GRA) is designed to allocate bandwidth to virtual pathsubnetworks in a wireless network.

adopted to overcome the problem of limited wireless resourcesand utilize the channel allocation efficiently. Various GA-basedapproached have been proposed in literature as solutions tothe problem of bandwidth allocation in wireless networks. Wewill describe a sample of these works (a tabulated summaryis presented in Table VI.

An accelerated GA approach with dynamic adjusting ofmutation and inheritance probability is proposed in [75] as asolution to the bandwidth allocation problem in WLANs. Theprobability of mutation and crossover is made dependent onthe individual chromosomes fitness. This approach is shown toquickly converge to global optima while substantially reducingthe number of generations as well as the computation time. Inanother work, Sherif et al. [45] proposed a GA-based approachto analyze the network bandwidth allocation using a predeter-mined range of QoS levels (high, medium, low) declared byeach multimedia substream. The proposed algorithm assignsthe best available QoS level to each substream dependingupon the availability of bandwidth resources in network.In the case of network congestion, the algorithm tries topreserve resources by degrading the quality of some admittedcalls. Using this technique, the algorithm attempts maximum

resource utilization and fair distribution of bandwidth in allthe networks while keeping in view the network constraints.

5) Channel Allocation and Assignment: With the rapidproliferation of wireless and cellular networks, and the prob-lem of limited frequency spectrum, the problem of channelassignment and reuse has become very important. The recentadvancement in wireless technology has enabled many high-speed data services leading to an upsurge in number of wire-less users. Since the frequency spectrum is a limited resource,there is a great emphasis on the reuse and sharing of frequencyamongst the ever-increasing user base. The goal of channelallocation is to maximize frequency reuse by minimizing thenumber of channels required to cover all the network cells.

There has been a lot of GA-based work done in the areaof channel allocation and assignment in wireless networks (asummary of which is provided in Table VII). We discuss someimportant GA-based channel allocation/ assignment worksbelow.

Broadly speaking, there are three main constraints that chan-nel assignment algorithms in wireless networks must consider:(i) co-channel interference constraints—which requires that

14

TABLE VII: Using genetic algorithms for channel allocation/ assignment in wireless networks

Reference No. ofobjectives

Techniques used Brief summary

Channel allocation in CRNsZhao et al. [84] MOGA Quantum computation

with GA [85]A hybrid GA implemented to decrease the search space and ease the mapping processbetween the channel matrix and qubit chromosomes.

Bhattacharjee et al.[86]

MOGA Comparison of GA withinterference graphs

Channel assignment optimization for single cell cognitive network using GA.

Friend et al. [17] SOGA Distributed Island GA An Island GA, solves the channel allocation through dynamic spectrum access incognitive radio networks.

ElNainay et al. [18] SOGA Distributed Island GA A localized island GA cooperate with other cognitive network nodes to utilize theavailable spectrum opportunistically and efficiently through channel allocation.

Ali et al. [87] MOGA Luby transform codes andGAs

A GA assisted resource management scheme is presented for reliable multimediadelivery.

Channel allocation in cellular networksZhenhua et al. [88] MOGA Immuned GA with elitism A modified immune GA, integrating immune operators, and minimum separation

encoding to obtain the conflict free channel assignment in cellular radio networks.Pinagapany andKulkarve [89]

SOGA Parallel GA with decimalencoding

A GA based scheme used to solve both fixed and dynamic channel assignment problemsin multi-cell cellular radio network.

Jose et al. [90] SOGA GA with Elitism and fit-ness entropy

Diversity guided Micro GA in cellular radios to optimize frequency/channel reuse infixed channel assignment problem.

Lima et al. [91] MOGA GA with immigrants andmemory schemes

Two adaptive GAs for locking and switching channel, solve the dynamic spectrumaccess problem through distribution of channels among cell in cellular networks.

Channel allocation for broadband wireless networksWong and Wassel[92]

SOGA GA based dynamic chan-nel allocation

GA based technique incorporated channel segregation to maximize the throughput inbroadband fixed wireless access networks.

Fang et al. [93] SOGA GAs with video codingscheme

Evolutionary technique alongwith Reed-Solomon coding and SNR matrix allocatechannel rates for scalable video transmission.

the same frequency or channel cannot be assigned to two dif-ferent radios simultaneously; (ii) adjacent channel interferenceconstraints—which requires that adjacent frequencies not beassigned to adjacent channels simultaneously; and, (iii) co-site interference constraints—that requires minimum spacingbetween a pair of frequencies assigned to a channel must haveminimum spacing.

Kassotakis et al. have proposed an approach based on hybridgenetic approach—that combines genetic algorithms with localimprovement operator—for channel reuse in multiple accesstelecommunication networks [94]. The proposed scheme aimsto establish the maximal number of connections while mini-mizing the number of isochronous channels. Wong and Wassel[92] investigated the channel allocation for a ‘broadbandfixed wireless access’ (BFWA) networks. A channel allocationmatrix keeps track of channel parameters for conflict freeassignments. It is observed that as the network grows, cen-tralized scheme would offer unacceptable signaling overhead.GA is used for distributed channel where each access point(AP) transmits the interference power of the entire availablechannel to a control unit. Control unit employs GAs and ranksthe chromosomes of channel priorities according to fitnessvalue of interference. Crossover and mutation rate are also setto promote the best channels and explore the new channels.Experimental Statistics have revealed that better SNR gain(more than 14.5 dB) and throughput is achieved with GAs incomparison with the other two dynamic channel assignmentmethods: (i) least interference [95], and (ii) channel segrega-tion [96].

A study on GAs in channel assignment in cellular networksis performed by Kim et al. [97]. This paper reports that neuralnetworks are more likely to get trapped on a local optima

since a neural network’s output critically depends on the theinitial values. In the case of GAs, they possess crossover andmutation as their key strength to escape local optimas. GAs,therefore, have far better chances to overcome the channelallocation problems such as (i) co-channel interference (ii)adjacent channel interference and, (iii) co-site interference.GAs are applied on clusters in cellular networks and it isshowed that if the number of cells in one cluster are increased,the convergence rate is decreased leading to an increase in thenumber of generations due to the reuse of the same frequencyamong different cells. GA comes handy in keeping the clustersize optimum for efficient frequency distribution among cells.

Ngo and Li [98] presented a modified GA to contribute inconflict free channel allocation and spectrum efficient assign-ment in cellular radio networks. There are generally two kindsof channel assignments: (i) fixed channel assignment (FCA)—where channel slots are pre-allocated to cells, and (ii) dynamicchannel assignment (DCA)—where channels are assignedupon request. FCA generally outperforms DCA under heavytraffic load. This algorithm, called the genetic-fix algorithm,generates and manipulates individuals with fixed size, andhence greatly reduces the search space. Furthermore, using theminimum-separation encoding scheme, the required number ofbits for representing solutions is substantially reduced.

6) Localization: The task of determining the location of awireless node is known as localization. Localization is impor-tant in many different settings (such as military monitoringservices and disaster management) for providing location-aware services. GAs have been used to assist the localizationprocess in various wireless settings. In this section, we willprovide an overview of such work. A tabulated summary of

15

TABLE VIII: Using genetic algorithms for localization in wireless networks

Reference Goal No. ofobjectives

Technique used Brief Summary

Zhang et al.[99]

Node localization withrandom distributions

MOGA GA with simulatedannealing

A GA approach for centralized architecture to optimize thelocalization in WSN with least mean localization error.

Yun et al. [100] Centroid localization SOGA GA with fuzzy logic A hybrid GA with fuzzy membership is developed for edgeweights to localize the nodes using received signal strengthindicator (RSSI) information.

Tam et al. [101] Localization with ad-hoc positioning system(APS)

SOGAs MicroGA as post-optimizer with APS

A single objective MicroGA, using smaller population worked outbest to improve the localization accuracy in APS for both isotropicand an-isotropic topologies.

Nan et al. [102] Sensor nodes localiza-tion

SOGA GA with elitism andpartial map crossover(PMX)

A real-coded version of GA utilizes roulette wheel selection withInteger coding, thus, helping in achieving precise node localizationin sensor networks.

Yun et al. [103] Localization with RSSI SOGA GA and fuzzy logicsystem

Calculation of edge weight with reference of anchor nodes, fol-lowed by modeling through fuzzy logic functions and optimizationby GA in the end.

TABLE IX: Using genetic algorithms for wireless AP placement optimization

Reference No. of Objec-tives

Technique used Brief Summary

Wireless Mesh NetworksSmadi et al.[104]

Bi-ObjectiveGA

GA with ILP A Joint clustering and gateway problem is addressed by GAs for placement ofhybrid free space opticals (FSO) to supplement the gateway capacity in WMNs

Xhafa et al.[105]

Bi-objectiveGA

GA on grid area with rectangularmutation and crossover

Proposed GAs to optimize the connectivity and user coverage area of meshnetwork through optimal placement of nodes in the area.

Mobile Ad-hoc NetworksKusyk et al.[106]

MOGA Distributed forced GA with evolu-tionary game theory

A GA alongwith traditional game theory, running at each node in autonomousnetwork, helps nodes deciding their mobility and organization.

Sahin et al.[107]

SOGA Distributed forced GA A Distributed GA adjust speed, location, and direction of mobile nodes over ageographical terrain for military applications.

Hoffman et al.[108]

SOGA GA with TDMA table and time slotexchanges.

A GA approach to gateway allocation in aeronautical networks and TDMAresource scheduling to minimize the latency.

Wireless Sensor NetworksYoussef andYounis [109]

SOGA GAs on pre-calculation genericrouting table

A GA places the WSN nodes for disjoint clustering and minimizes the numberof hops between sensor and nodes to have efficient intra-cluster communication.

Tripathi et al.[110]

Bi-objective GA and Genetic Programming(GP)

GP improves the deployment, while GA optimizes the node placement formaximum coverage in WSN.

Pandey et al.[111]

MOGA GA with greedy algorithm and ‘bi-nary integer linear programming’(BILP)

A hybrid GAs along with greedy algorithm optimizes the node placements fortraffic balancing, in two tier hierarchical heterogeneous WSNs.

Xu and Yao[112]

SOGA GA with integer coding and ordermapped crossover (OMX)

A GA approach worked on optimal placement of WSN nodes in grid area full ofhindrances and obstacles to match in real-time environment.

some sample works are presented in Table VIII.

A simple way to localize the nodes is to deploy GPS, butinstalling GPS in thousands of closely placed nodes is im-practical. A better approach is to employ GPS in a few nodes,called anchor nodes, and apply techniques to locate all theother nodes relatively, with reference to multiple anchor nodes.Tam et al. [101] proposed localization technique in whicheach node finds its position relative to three anchor nodes,thus, enabling each node to locate itself through triangulation.This information is then broadcasted to base station wherea centralized GA works as a post-optimizer, thus, leadingto precise network graph with each node’s location known.Zhang et al. presented the centralized GA with simulatedannealing [99] to minimize the error in position estimationof nodes in WSNs. This approach is also based on anchornodes localization scheme. Simulated annealing is local searchtechnique and it avoids pre-mature convergence, by mixing

inferior and sub-optimal solutions with the optimal ones. Inthe approach of centroid localization [100], proposed in [100],GAs and fuzzy logics are used by each node to performself-localization by estimating the edge weights based on thereceived signal strength (RSS).

7) WLAN AP placement: Since the arrival of IEEE 802.11,WLAN services are widely adopted with the predominantWLAN setting being infrastructure mode (in which the topolo-gies around APs). Optimizing the placement of WLAN APsplays a substantial role in maximizing the throughput and datarate of WLANs. In this section, we will summarize the variousproposals for using GAs in WLAN AP placement. A tabulatedsummary is presented in Table IX.

A multi-objective GA (MOGA) approach for WLAN APplacement (with the objectives of minimizing the number ofAPs and maximizing SNR) has been described by researchersin [130]. GA-provided solutions have lesser variance implying

16

TABLE X: Using genetic algorithms for general applications in wireless networks

Reference NetworkType

No. ofobjectives

Technique used Brief summary

Area CoverageXhafa et al. [105] WMNs Bi-objective

GARectangular mutation &crossover

A GA approach is used to maximize the ‘giant component’ andcoverage area in WMNs while using multiple distribution strategiesof client nodes.

Alba et al. [113] MANET MOGA Broadcasting updates from basestation running GA

A hybrid broadcasting protocol using evolutionary algorithms. AMOGA approach is used to optimize metropolitan-scope MANETcoverage and usage.

Hu et al. [114] WSN SOGA GAs with schedule transition op-erations

Hybrid GA with scheduled transition and forward encoding schemefinds the maximum number of disjoint cover sets to enhance thedisjoint covered sets in areas.

Sahin et al. [115] MANET MOGA A Distributed forced GA A forced GA is proposed letting the nodes decide their movement,direction and speed through molecular force equilibrium to get maxi-mum area coverage

Lai et al. [116] WSN SOGA Integer coding GA Based scheme with one-to-one mapping of genes and sensors tofind maximum disjoint sensor covers to prolong the network life span.

Quintao et al.[117]

WSN SOGA GAs coupled with mathematicalmodeling and ILP

GA based heuristic approach with integer linear programming (ILP)and mathematical formulation to solve dynamic network coverageproblem.

Network Planning and designBhondekar et al.[118]

WSN MOGA GA model for grid deployment A GA decides the transmission ranges and the active/ sleeping nodesin WSN topology while accounting for connectivity, energy, and otherapplication specific parameters.

Jourdan et al.[119]

WSN MOGA GA for non-linear objectives A MOGA approach design the Pareto optimal layouts for WSN nodesdepending upon the sensing and communication range.

Pries et al. [120] WMNs SOGA Tree coding GA operators like subtree crossover and cell crossover, respectively,are employed in solving small and large-scale WMN gateway opti-mization problems.

Ayyadurai et al.[121]

CN SOGA Single cell optimization An addaptive GA for dynamic system reconfiguration optimizes theresource allocation in cellular networks.

Ghosh et al.[122]

WMN SOGA Binary encoding of nodes A dynamic mesh topology design, using GAs, while maintaining lownetwork cost and path redundancy at base station for reliability.

Sheen et al. [123] CN MOGA Joint optimization A GA for maximizing joint multi-cell CN system efficiency, throughoptimizing relay stations (RS) placement, frequency reuse pattern, andmaximizing throughput to enhance the end user experience.

Chiaraviglio etal. [124]

CN MOGA Hybrid GA with centralizedsorting of base stations

GA based switch off strategies for enabling sleep modes in basestations to design energy efficient cellular networks.

Cluster OptimizationZeng et al. [67] WMN MOGA Greedy algorithm with GAs A hybrid GA meets the QoS and load balancing constraints separately

by constructing the graph and dividing the WMNs in disjoint clusters.Hussain et al.[125]

WMN SOGA Cluster optimization throughdistributed broadcasting

GAs augmented the network life when compared with LEACH [126]by employing energy efficient data dissemination and optimizing thecluster distances.

Zhang et al. [127] WSN SOGA Simulated annealing (SA) [128] A GA approach provided the energy efficient WSN topology byimproving the data aggregation rate and minimizing the intra-clusterdistances.

Huruiala et al.[129]

WSN SOGA Centralized GA A GA running on BS coordinate with nodes to minimize latency andpower consumption for routing under tight constraints.

that a larger number of area points within the network wouldhave average SNR. Statistical analysis reveals that MOGAapproach generate solutions with a high SNR data profile.

Optimizing the radio coverage inside buildings is as im-portant as outdoors radio optimization. WLANs are mostlyused inside homes and offices to provide mobile and Internetcoverage to subscribers. Nagy and Farkas [131] have studiedthe problem of optimizing the placement of indoor BaseStations (BS) while taking in account the effects of floors,ceiling, and walls on signal propagation and path losses. GAused the microwave propagation model to account for theeffect of influence of brick, concrete, wood, furniture, andglass obstacles. Different weighting coefficients were assignedto each obstacle and parallel heuristic approach of GA is usedto optimize the coverage results within indoor environment.

Another important application of GAs is in the optimizationof AP placement in cellular and mobile networks. Sincefinding the set of coverage areas whose sum is the maximumcovered area is an NP-hard problem, GAs come handy. Lieskaet al. [132] have explored in their study the combination ofsimulation and evolutionary techniques. Radio coverage areasare obtained by the series of BS with received power greaterthan -60dbm. Algorithm works by selecting the chromosomesbased on required BS. The chromosomes are than evaluatedagainst constraints of power and SNR; afterwards, the healthychromosomes are allowed to reproduce leading to globalconvergence of the entire network. This fitness function canbe modified to include other constraints based on networkrequirement. A major advantage of a GA, therefore, is itsability to adapt to multiple situations and problems.

17

8) General Applications: A survey for solving computa-tionally difficult problem, through GAs, in wireless networksis presented by Das et al. [133] in his book “Handbook ofBio-Inspired Algorithm and Applications”. It has been shownthat GAs have the potential to solve any optimization problemprovided that proper modeling and chromosomes encodinghas been performed. In this section, we will cover someof the other miscellaneous applications of GAs in wirelessnetworking (that we have not covered in previous sections). Atabulated summary of these applications are presented in TableX. We describe some of these applications here; some other ap-plications are discussed when we cover network-configurationbased categorization of GA applications in Section III-B).

GAs have been proposed for energy efficient planning andmanagement of cellular networks by enabling sleep modes inbase stations [124]. Johnson and Samii [134] have proposeda GA based approach to find the best node distributionwith optimized antenna pattern and SNR. The authors haveemployed simple binary coding scheme for a pool of randomlygenerated chromosomes for different nodes distribution andevaluated their fitness against the fitness function (in which Nis the total number of nodes):

Fitness = max− 1

N

∑i

Path_lengthi (1)

The winner chromosomes (having minimum cost of pathlength) are more probable to be inherited in the next gener-ation. This approach is also applied for random encoding ofantenna pattern and transmission power.

Churn prediction and forecasting have application in almostall the real world phenomenon including medical, finance,products sale, telecommunications and network. Forecastingthe future sale and growth in turbulent markets like telecom-munications is always a big issue for product managers. Evenapproximate insights can yield large profit return in the longrun. Venkatesan and Kumar studied GAs for forecasting [135]the magnitude of future sales, time period to peak sales andthe value of sales during peak time. Considering the significantadvantages of accurate forecast of product lifecycle, GAs areused to estimate the diffusion model based on previous datapoints. GAs are fed with several critical factors like price,competitive intensity and network effects in obtaining betterpredictions during growth phase. Since GAs are inherentlyparallel, so, different scenarios can be taken care of whilemaking predictions. Study shows that predictions by GAsare robust and accurate in predicting the future growth of amarket. Pendharkar [136] presented the hybrid GA with neuralnetworks to forecast the churn in cellular network services.The hybrid algorithms optimize the objectives after learningand assigning weights to certain parameters. Results obtainedfrom GA method are shown to be more accurate than statisticalmodels but are computationally expensive.

B. Network Type Based ClassificationIn this section, we will describe the applications of GAs in

wireless networking categorized according to the type of wire-less network. In particular, we will present GA applications

in the networking configurations of wireless sensor networks(WSNs), wireless mesh networks (WMNs), cognitive radionetworks (CRNs), cellular networks (CNs), and software-defined wireless networks (SWNs).

1) Wireless Sensor Networks (WSN): WSNs have emergedas an exciting technology that can be used to monitor the phys-ical world. WSN is composed of self-organizing, miniaturizedsensors nodes, running on batteries, and communicating overwireless links. The major challenge in WSNs is to providearea coverage with minimum WSN nodes while preservingthe resources and energy consumption.

Jia et al. [137] addressed this problem through MOGA forenergy efficient coverage control in WSNs. The algorithmprovides a mechanism to cover the maximum deployed areawith minimum number of WSN nodes. Since connectivity andcoverage are both dependent on each other so this problemis addressed by MOGA approach by achieving the Paretooptimal front. This is a non-dominated problem so it comprisesof multiple solution individuals that are evaluated againsttwo fitness function coverage area and sensor used rate. TheMOGA approach applied performs well in maximizing thearea covered with least computation. Another advantage isthat it avoids partial solutions in whole search space and resetsensor nodes once in whole network simulation.

Frentinos and Tsiligiridis [138] studied an adaptive WSNdesign based on MOGA. Individual with random placementof nodes are utilized as population and GAs decides whichsensors would be active and which would act as clusterhead.Random designs are encoded and fed to GAs and the winnerchromosomes are going to decide the network infrastructureincluding clusterheads and active or passive nodes alongwith signal strengths. From evolution of characteristics, itwas shown that generally, it is more favorable to have largenumber of sensor nodes with low energy rather than havinglesser number of sensors with high energy consumption. Thisapplication of adaptive GA in WSN can lead to increase innetwork life span while maintaining network properties closeto optimal value as well as it helps in adapting sophisticateddecisions about operating modes (clusterheads, active nodeswith optimal signal range).

Extending the life span of any WSN network is activearea of research. Different schemes have been proposed likecentralized monitoring GA for keeping active and passivenodes in network. A similar approach has been discussed byJin et al. [139] to cut the energy consumption in WSNs. AGA is used for independent cluster formation to reduce thecommunication distances thus bringing the route cost down.Experimental results showed that by keeping the number ofcluster to 10% of total nodes, results in 80% reduction inpath cost and communication distance as compared to randomdistribution of WSNs. In another work, Sengupta et al. [140]have proposed an evolutionary multiobjective sleep-schedulingscheme for differentiated coverage in WSNs.

Zhang et al. [141] used GA to extend the work of Niculescuand Nath on network localization in WSNs [142]. The centralidea is to employ GPS enabled nodes, called anchor nodes, inthe network and to have all the other network nodes estimate

18

their position through 1-hop distances to the anchor nodes. Thelocations are then flooded by each node to the clusterheadwhich uses the location information to construct the graphof the entire network. By using standard GA operators likemutation and crossover, the proposed GA based approachwas able to better estimate locations resulting in lesser meanposition error than other algorithms in literature such as the‘semi-definite programming with gradient search localization’(SDPL) [143] and the ‘simulated annealing based localization’(SAL) [144].

In another work, Hussain et al. [125] have studied and sim-ulated GAs for centralized hierarchical WSNs. The proposedscheme comprises a hierarchy of layers with the BS beingat the top layer and the clusterheads and the sensor nodesoccupying the bottom layer. GAs are used to construct clusterswith a clusterheads and other nodes. The BS layer keeps trackof the entire cluster through their clusterheads. A GA approachminimizes the average cluster distances while simultaneouslyincreasing the total number of transmissions. After configuringsensor nodes in optimized formation, the data transmissionphase follows.

2) Wireless Mesh Networks (WMNs): WMNs provide apopular mechanism for the provisioning of cost-effectivewireless broadband connectivity. WMNs are composed of(potentially mobile) mesh clients and (typically fixed) meshrouters that connect over a relatively static multihop wirelessnetwork [145]. To address the issue of efficient placement ofmesh routers, Xhafa et al. [105] employed GAs to optimizethe user coverage area and connectivity of mesh networkthrough optimal placement of nodes in an area. Chromosomesare designed in rectangular grids having information of bothclients and mesh router nodes. Crossover is implementedto preserve the chromosomes traits by randomly exchangingparent and child router nodes from one cell to another in gridarea, thereby promoting high scoring chromosomes. Mutationis then utilized to randomly migrate a router to another gridand recalculating the fitness. Various models can be adoptedfor distributing client nodes such as the normal, uniform, andexponential distribution. GAs are excellent in computing themesh router placements and always succeed in achieving theconnectivity, irrespective of client distribution, however, areacoverage is subjected to client nodes in network area.

3) Cognitive Radio Networks (CRNs): Cognitive radio net-works (CRNs) offer a potential solution for improving thespectrum utilization in the world of limited spectrum. Indynamic spectrum access (DSA) based CRNs, the secondaryusers of a network aim to utilize the spectrum opportunisticallyto offer QoS for secondary users without causing any interfer-ence to licensed primary users. Such DSA based CRNs utilizecognitive radios (CRs) that incorporate various AI techniquesfor intelligent decision making [20]. The main benefits ofapplying GA in CRNs are that: (i) it is conceptually easy tounderstand; (ii) it is inherently amenable to parallel solutionsand therefore can be easily distributed; and that (iii) GAs lendsupport to multi-objective optimization and performs well innoisy unknown environments.

An early application of GA techniques to CRNs is doc-

umented in a paper authored by Rondeau et al. [146]. Thispaper presented the adaptation mechanism of a cognitiveengine, implemented by the authors, that used GAs to evolvea radio’s parameters to make them best for the user’s cur-rent needs. A ‘wireless system genetic algorithm’ (WSGA)is also proposed to realize adaptive waveform control andcross-layer optimization [146]. GAs have also been used forbuilding distributed parallel solutions, using the technique ofisland genetic algorithms, to mitigate the channel assignmentproblem in CRNs [17]. An island GA divides the overallpopulation into subpopulations known as islands, which theninteract by migrating individuals to the other island. Moredetails of parallel genetic algorithms, and of general parallelmeta-heuristics, can be found in the survey [19] and [147],respectively.

a) MOGAs for optimization of CRNs: Multi-objectiveoptimization aims at estimating a true Pareto front opti-mization in wireless medium depending upon the externalenvironment parameters. These radio environment parametersserves as genes and are encoded in binary strings, whichform chromosomes. These chromosomes are evaluated againstfitness function regularly in successive generations. Afterrealizing sufficient fitness or completing predefined number ofgeneration, these set of transmission parameters are sent forupdating new radio configuration. Due to variation in radiochannel characteristics and spectrum availability, a cognitiveradio has to maintain time varying QoS while optimizingdifferent parameters such as BER, throughput, power con-sumption, and interference.

MOGAs have been used extensively for optimization ofcognitive radio networks (CRNs) [28]. The foremost aim ofcognitive engine (CE) is optimizing the radio parameters forbest QoS and secondary goal is to observe and learn for adap-tation. To achieve these goals cognitive engines have (i) cog-nitive system module (CSM)—for modeling and learning (ii)Wireless System Genetic Algorithm (WSGA)—for actions (iii)Wireless Channel Genetic Algorithm (WCGA)—for hardwareimplementation of radio parameters. These parameter and ra-dio traits called genes represent chromosomes, are modulation,frequency, spreading, filtering, and FEC coding, etc. Differentobjectives are optimized based on relative weights assigned tothem using multi-objective approach. These objectives includeoptimizing BER, power consumption, interference and MOGAapproach is applied on population and chromosomes, whichperforms best in most of the objective, are chosen as new setof radio parameters.

Case Base Decision Theory (CBDT) for Initialization ofGAs: Since the secondary objective of CR is to adapt andevolve with changing environmental parameters by updatingradio configurations. So a CR system must be provided witha memory database that continuously monitors and learns GAoutcomes in real-time optimization. As the system matures,with the passage of time, each new optimization would takeless time to achieve global optima. More thoroughly speaking,CE keeps track of previous objectives and the winner chromo-somes. So whenever a new objective or case is encountered,it is compared to previous cases in memory and chromosomes

19

of the case/objective having maximum similarity and utilityare used to partially initialize the GA population for currentobjective. Rest of the population is randomly generated to notonly maintain diversity but also to help in achieving optima inlesser number of iterations. This adaptation scheme is carriedout through algebraic manipulations with the independentlyconstructed channel availability matrix, the interference ma-trix, and the channel reward matrix (which are each kept upto date about the external radio environment conditions) [148].

Rieser [149] presented a biologically inspired cognitiveradio model, using modern AI techniques. This model isbased on adaptive cognitive engine that learns the radioenvironment, optimizes the radio parameters, and presentsthe best possible experience to secondary user. He built hisfoundation of GA based technique on original Mitola [150]proposed technique of software radios. He gave insights onall the popular AI Algorithms and how ineffective they are inevolving as compared to GAs, when faced with unanticipatedradio channel environment. He applied GA to hidden Markovmodel for error modeling in uncertainties in radio channelenvironment. Wireless channel GA has its foundation on theseradio environmental modeling.

The biologically inspired cognitive engine serves andevolves well in unanticipated radio channel environment,which make it usable for military purpose and emergencysituation. This whole cognitive architecture works in multiplesteps. Firstly, the engine gathers the channel informationwhich is then subjected to machine learning techniques suchas learning classifiers. After learning and modeling, theseinformation are taken up to alter the radio configurations fortweaking appropriate settings [28]. Wireless system GA treatsthe radio as the analogue of a biological creature, and makesoptimized discoveries to keep the appropriate balance of radioparameters. The cognitive engine architecture, with WCGA,CSM, and WSGA, is displayed in Figure 7. We will onlybriefly explain below the main infrastructural components ofthe cognitive engine developed by Rieser.

1) Wireless Channel Genetic Algorithm (WCGA): WCGAtakes the information about radio channel through phys-ical and MAC layer and constructs a statistical modelbased on external channel environment. This informationis then forwarded to CSM.

2) Cognitive System monitor (CSM): CSM incorporates thecapability of learning and adaptation. CSM after learningand generating appropriate action goals instructs WSGA[28] for actions leading to different radio configurations.

3) Wireless System Genetic Algorithm (WSGA): WSGAreceives the channel model by CSM with fitness functionand initial chromosomes. WSGA performs all the GAcomputation through mutation, selection and crossover,thereby helping to select the best chromosomes. Thesechromosomes are, in fact, different radio configurations,which are then used for implementing new radio set-tings.

b) Cognitive radio (CR) testbeds: Rieser [149] devel-oped a biologically inspired cognitive radio CR testbed. Anumber of cognitive theories are developed and evaluated in

Fig. 7: Cognitive Engine Architecture with WCGA, CSM, andWSGA [28]

parallel through the use of GA operators. This developedtestbed consists of multilayered cognitive engine inspired byhuman brain evolution. The architecture is shown in figure. Itprovides multilayered scaled functionality with sensing radioparameters at both MAC and physical layers. It uses threealgorithms that can be co-located in same radio or distributedto multiple radios. This cross-layered cognitive engine archi-tecture is shown in Figure 7. Another GA-based softwaretestbed for cognitive engine (CE) is proposed by Kim et al.[151]. A two-step GA is used to avoid problem of optimizingthe biased dominant parameters. To implement this, GA basediteration is run for channel selection and transmitter parametersselection separately. Simulation of this testbed demonstratesthat bandwidth and transmit power converge to optimal valueswith in few iterations of GAs.

4) Cellular Networks (CNs): Next generation wireless net-works have to cope with the massive increase in multimediaservice along with cellular data. This motivates the devel-opment of self-consolidating networks that are capable ofmaking intelligent decision to offer the best user quality ofexperience (QoE). Ayyadurai et al. [121] presented an adap-tive GA to optimize the cellular networks spectral efficiencythrough dynamic allocation of resources, between the relaystations and the base stations. Population adjustment camealong the way to improve the search capabilities leading toquick optimum. Chiaraviglio et al. [124] presented the energyefficient deployment of cellular networks by leveraging GA tominimize the power consumption and number of transmitterin cells. This involves employing sleep modes, where, a basestation decides which nodes should be powered on, dependingon number of active users in given cell.

A conflict free channel assignment is an NP-hard problem.A powerful GA incorporating immune operators includingimmune clone, immune vaccination, and immune selectionhas been proposed by Zhenhua et al. [88] to mitigate the

20

interference constraints and obtain a conflict free channelassignment. Vaccination involves changing some bits of agene to have an improved fitness value. Elitism keeps thepopulation diversity and minimum separation encoding helpsin keeping a frequency distance among individual channels.A similar approach has been studied by Pinagapany andKulkarve [89] for solving both dynamic and fixed channelassignments in cellular networks. Channel compatibility ma-trix is constructed first to record the violation of individuals.After that, cell and bandwidth information are encoded inindividual chromosomes, which are then subjected to crossoverand mutation. Jose et al. [90] offered low computational microGA with elitism strategy to solve the fixed channel assignment(FCA) problem. The proposed algorithm spotlighted how tokeep diversity by varying mutation probability to avoid localoptima.

In cellular networks, it is desirable to minimize the calldrop and call blocking probability. Mobile cellular networksare often evaluated against the criteria of dropping probabilityof current calls and the blocking probability of new calls. Limaet al. [91] studied two adaptive GAs for locking and switchingchannel to solve the dynamic spectrum access problem incellular networks. The proposed fitness function evaluates thecapacity of each channel for new calls and also keep tracksof existing calls. The algorithm takes care of randomness byallowing random immigrants and parameter adaptation to keepexploring new option for each channel.

5) Software Defined Wireless Networks (SWNs): In recenttimes, software defined networking (SDN) has been proposedas a new architecture for conveniently operating and managingdiverse wireless networks. SDN is generally understood toentail (i) the separation of the control and data plane; (ii)the control of multiple data planes by a single controller; andthe iii) development of new sophisticated control abstractions[152]. Wireless networks using SDN principles are referred toas software defined wireless networks (SWNs). Although thisnetworking configuration is relatively new, the use of GA hasbeen proposed in SWNs. For example, Bojic et al. have pro-posed a small-cell mobile backhaul resource manager, basedon GA, that interworks with software defined management[153].

C. Hybridization: GA With Other Techniques

GAs is a useful framework that can be profitably used in awide variety of settings. It is also possible to use GAs withother adaptive algorithmic approaches such as game theory,reinforcement learning (RL), graph theory, fuzzy logic basedand quantum theory based systems. Such a symbioses is not atall antithetical from the theme underlying the GA frameworkwhich encourages pluralism in preference to relying on anyone solution; it is often the case that we can perform betterby combining the best parts of different solutions as buildingblocks. In this section, we will describe hybrid approaches thatcombine various techniques with the principles of GAs.

1) Combining GAs with Game Theory: Game theory isa mathematical decision framework through which we can

study and analyze competitive interaction between two ormore self-interested rational agents. The field of evolutionarygame theory utilizes the concepts of evolutionary computingand genetic algorithms in the backdrop of a game-theoreticbackground [154]. Evolutionary games have been used ina wide variety of contexts in networks including coopera-tive sensing [155], multiple-access-control and power control[156], routing games [157], network selection games [158],congestion control [159], adaptive TCP [160], and for dynamicbandwidth allocation [161]. The interested reader is referred toa chapter on this topic in a recently published book on gametheory in wireless and communication networks [162] for adetailed treatment.

2) Combining GAs with Reinforcement Learning: Rein-forcement learning, unlike supervised learning, is a tech-nique where agent has no prior knowledge of transitionsand learns along the way through experience and interactionwith the environment. We note that evolutionary algorithmslike GAs are related to reinforcement learning as they relieson exploration and exploitation [163]. More specifically, GAutilizes the genetic operators of mutation and recombination asprocedures for exploration along with selection procedures toexploit known good solutions. Reinforcement learning can becarried out both by searching through the policy space or thevalue-function space. Genetic algorithms can be thought of asreinforcement learning scheme of the policy space searchingvariety. Policies are encoded in chromosomes and determinis-tic procedure is used to evaluate the cumulative fitness of theagent the follows a given policy. According to evolutionaryprinciple of survival of the fittest, only good policies areadopted while failed or weaker policies are relinquished [164].

3) Combining GAs with Graph Theory: In order to producerobust dynamic graph-theoretic solutions, graph theory canbe combined with GAs in particular, and with evolutionaryalgorithms more generally. The combination of graph theorywith evolutionary algorithms is known as evolutionary graph(EG) theory in literature. Ferreira et al. [165] have appliedEG theory in MANETs with known connectivity pattern fordeveloping efficient practical routing protocols. The dynamicbehavior of network is captured through evolving graphs.There are various metrics that can be used for determiningoptimal routes in EG-based models with known connectivitypatterns, including the three metrics of shortest, fastest, andforemost proposed in [166]. These metrics determine respec-tively the minimum number of hops, the minimum time span,and the earliest arrival data. Ferreira et al. [165] utilizedthe foremost journey metric along with the EG theory. Theproposed algorithm can estimate the congested nodes andsuggest the parallel paths. In another work, a bio-inspiredgraph theoretic solution for ensuring energy efficiency incooperative communication networks has been proposed in[167]. The presented algorithm combine the graph theorywith evolutionary techniques to the let the nodes operate indecentralized fashion for intelligent decision making, and doforwarding by selecting energy conserving nodes.

21

4) Combining GAs with Fuzzy Logic: Fuzzy logic is alogical system that can entertain logical variables that takecontinuous values between 0 and 1 instead of discrete valuesof either 0 or 1 (as in classical digital logic). Fuzzy logicis as an effective method for knowledge representation, con-trol implementation, and cross-layer optimization in wirelessnetworks [168]. Fuzzy logic is well suited for knowledgerepresentation in wireless networks since network conditionsin wireless networks are often incompletely and impreciselyknown. Fuzzy systems that utilize fuzzy logic are highlysuited to problems with imprecise, noisy, and incomplete inputinformation. GA can be used for both optimizing the fuzzymembership function as well as the fuzzy rules that drive thefuzzy control system. The intersection of hybrid technologiesof GA, fuzzy logic, and other AI techniques (most prominentlyneural networks) is studied in the research theme of ‘softcomputing’ [169]. The trend of genetic fuzzy systems for softcomputing applications is less developed than the trend ofneuro-fuzzy systems, especially in the context of industrialcontrol applications. Nonetheless the use of genetic fuzzytechniques is popular for both communication systems [170]and wireless networks [103].

5) Combining GAs with Quantum Theory: Quantum theoryderives its root from quantum mechanics, i.e., mechanicsof atoms. A comparative study has been performed in [85]between quantum GA and classical GA. Quantum GA provesto be faster and more robust then classical GA for travelingsales person problem. A quantum GA is proposed in [171]to solve the NP-Complete QoS metrics problem in WSNs.Quantum GA operate on qubit as a smallest unit of infor-mation, that serves as a linear superposition of two states,i.e., 1 and 0. A rotation gate is applied to keep the diversityof population in quantum GA. Quantum GA behaves moreeffectively than traditional GA for larger input set owing toits linear superposition states.

6) Combining GAs with Local Search: A very commonform of a hybrid GA, sometimes referred to as Memetic algo-rithms [172] in literature, is created by combining GAs withlocal search techniques and to incorporate domain-specificknowledge in the search process [27]. This can help producestronger results but at the cost of higher computational costs.The local search operator is applied to each member of thepopulation after each generation. Memetic algorithms havebeen used in the context of wireless networks for varioustasks such as energy-aware topology control [173] and cellassignment to switches [174].

7) Combining GAs with Other Metaheuristics: A populararea of research is combining GAs with heuristic methods likeant colony optimization (ACO), neural networks (NN), andfuzzy logics to solve the dynamic constrained multi-objectiveoptimization. In particular, there has a been a lot of work inapplying GAs for adjusting the connection weights and theconnectivity itself in NNs [175]. Notwithstanding the amountof work done, there is a lot of scope for further optimizationand improvement.

IV. PRACTICAL ISSUES AND LESSONS LEARNT

In this section, we will summarize the main insights andlessons learnt through decades of GA research that can guidesuccessful problem solving through GAs. We first describecommon pitfalls in Section IV-A, and then describe someguidelines and insights for successful GA implementation inSection IV-B.

A. Common Pitfalls

While GAs provide a useful framework for problem-solvingin a wide variety of domains, it is not a panacea. In this section,we will highlight some deficiencies of GAs and how it maybe incorrectly applied.

1) Genetic Drift or Bias: In GAs, there is a risk ofslow convergence and settling for local minimas, especiallywith inappropriate choice of model parameters. While GAscan often converge to the global optima in the majority ofapplications, there are no mathematical guarantees to thiseffect. In particular, a GA can get get stuck in the vicinityof a suboptimal point if it lacks population diversity, or if hasan inappropriately small population. Such a phenomenon isknown as ‘genetic drift’ or ’genetic bias’. Certain problemscannot be solved by means of simple genetic operations. Thisis mainly due to the choice of poor fitness functions that gen-erate poor chromosome blocks. There cannot be an absoluteassurance that a GA will be able to find a global optimum—this can especially be a problem when the population has alot of subjects.

2) Using GAs for Real-Time Applications: Like most otherAI techniques, GAs are not well suited to real-time applicationthat require guaranteed response times due to its stochasticand heuristic nature. In addition, the response time of GAcan have significant variance depending on the various controlparameters and population initialization. This potential pitfallshould be kept in mind if GAs are used for online learningand optimization in wireless networks. In particular, QoSrequirements in wireless networks may put constraints onachieving the optimum decision in limited time forcing GAsto terminate after predefined number of iterations.

3) GA Deception: It has been reported in literature thatcertain objective functions, referred to as GA-deceptive func-tions, can be very difficult to optimize. With such deceptivefunctions, combination of good building blocks may result invery bad chromosomes, causing the building block theory tofail. A potential solution to this problem is using messy geneticalgorithms [3].

B. Implementation Insights

1) Setting Population Size: The size of the population in aGA can be a major factor in determining its quality and con-vergence speed. Generally speaking, larger populations resultin better solutions. A few studies have addressed the problemof determining exactly how large a population should be toensure a certain quality [176] [177] [178]. For a moderatelycomplex problem, a population of 50 to 100 chromosomes

22

can serve as a good default value [179]. The precise choice ofthe population size depends on the problem complexity andthe number of search variables. While it is typical for thepopulation size to remain fixed from one generation to thenext, it can also be made to grow or contract depending onthe rate of convergence, and other factors such as the currentgenetic diversity.

2) Designing Solution Encoding and Fitness Function:The efficiency, processing power, and the convergence ofGAs depend critically on how the solution is encoded inthe form of chromosomes and on the formulation of anappropriate fitness function. This is the first necessary steptowards an efficient implementation. There is no one-size-fits-all solution here, and the correct model choice depends onwhat is most useful for the problem at hand. The encoding ofchromosomes should be carefully considered. While a majorityof implementations utilize bit representation, due to the easewith which crossover and mutation can be implemented, otherrepresentations, utilizing data structures, may be appropriatefor the considered problem and should also be considered [27].Since there can be numerous encoding and formulations, thisstep involves considerable ingenuity. Some guidelines on theseissues are provided in [22] [180].

3) Setting of Mutation and Crossover Rates: The optimiza-tion of the control variables of genetic algorithms (such as thesetting of mutation and crossover rates) plays an importantrole in the performance of a GA based framework. A GAframework can avoid prematurely settling in a suboptimal rutby emphasizing exploration using mutation. Various solutionshave been proposed to the problem of genetic drift/ biasincluding fitness sharing [181], crowding and pre-selection[182]. Bhandari et al. [8] showed that the probability offast convergence increases by seeding the initial populationwith high scoring chromosomes. As a general rule-of-thumb,crossover and mutation rates of 0.6 and 0.001 have been pro-posed for large population sizes (of ∼100), with correspondingvalues of 0.9 and 0.01 proposed for small population sizes (of∼30) [22] [180]. Sastry et al. [27] similarly recommend amutation probability of .05 and a crossover rate of 0.6 with apopulation size of ∼50. A tabulated summary of mutation andcrossover rates generally recommended by the GA communityand accepted as de facto ‘standards’ are presented in Table XI.

4) Computation of Fitness Values: In many problems,the computation of the fitness function can be extremelycomputationally intensive, thus motivating research in fitnessapproximation [184]. An important practical issue is to de-termine ways of modeling large complex systems so that thefitness value can be computed without executing the real-worldsystem itself (which can be impractical in many situations).

5) Setting of Termination Criteria: Assuming the pop-ulation, and the various GA control variables, have beenset up properly, the GA will improve towards the optimalsolution over successive generations. In light of potentiallyslow convergence, it is very important to for timely problemsolving with GAs to devise a suitable termination criterionthat defines a ‘good enough’ solution. Defining these criteria is

non-trivial and context-dependent. While this criteria dependson the number of objective functions, and the number ofvariables, a default value of 2.5 times the population size canbe used as a reasonable maximum generation count estimate[179].

V. OPEN ISSUES AND FUTURE WORK

GAs have become a powerful tool for solving almost anykind of an optimization problem–subject to proper modelingand encoding of chromosome. GAs are contributing in mili-tary, disaster management, spectrum utilization, and networkoptimization. There are numerous challenges in devising anappropriate GA based solution including suitable definitionsof population size and evaluation function (which is typicallynon-trivial to define). In this section, we provide the recom-mendation on scope of future research on GAs in the field ofnetworking.

A. Theoretical Results for GAs

A lot of effort has focused on providing theoretical un-derstanding of why and how GAs work but rigorous resultshave been slow in coming. Some of the results, such as “NoFree Lunch”, have sobering implications. Simply stated, nofree lunch implies that all search algorithms perform equallywhen averaged over all problems. This further entails that ifwe are comparing GAs with other search heuristics (such ashill climbing or simulation annealing), even if GA performsbetter for some class of problems, the other algorithms willnecessarily perform better on some problems outside this class.GAs are also population-based heuristics; even as the entirepopulation is improved by GAs, the same cannot be guaranteedfor an individual existing within that population. There is aneed for enhanced theoretical understanding of various facetsof GAs to lay a strong foundation upon which the GA user’strust can be safely placed.

B. How to Exploit Problem Specific Knowledge?

Techniques like GA and neural networks are often resortedto as black-box solutions to complex problems where themechanisms to be used for problem solving are not wellunderstood. Even without problem specific knowledge, GAscan allow insights to develop thorough identification of sim-ilarities between highly fit individuals [6]. GAs exploit thisinformation by combining highly fit similarities to explorebeyond the currently known best solutions. When the subjectknowledge is available, in the form of good understanding ofthe underlying problem and plausible candidate solutions, itis desirable to exploit this information to expedite GA’s pro-gression towards a solution. One way of embedding problem-specific information in the GA system is by seeing GAs withgood structures and populations thereby accelerating searchand learning progress. However, the use of problem specificknowledge to generate specialized heuristics or operators canentail some loss of generality of the GA system [6]. Thequestion of how to embed wireless-specific subject knowledgeinto the GA frameworks is very much an open research issuerequiring further exploration.

23

TABLE XI: Summary of GA paramater settings guidelines generally accepted by the GA community

De Jong and Spears [180] Grefenstette [22] Sastry et al. [27] Carrol [183] (Micro GA)

Population Size 50 30 50 5Number of Generations 1000 Not Specified Not Specified 100Mutation Rates 0.001 0.01 0.05 0.02; 0.04Mutation Types bit flip bit flip Not Specified jump and creepCrossover Rate 0.6 0.9 0.6 0.5Crossover Type 1 or 2 point (typically, 2 point) 1 or 2 point (typically, 2 point) Not Specified Uniform

C. Efficient Multi-Objective and Distributed GAsThe development of practical algorithms for wireless net-

works often requires the use of multi-objective optimizationperformed in a distributed fashion. Currently, the most popularapproaches in wireless networks depend on centralized GA[116]: e.g., in mobile cellular networks, the GA runs on theBS and helps decide the cluster size and nodes deployment.In wireless networks, there is often a need to distribute thecomputational load to various network nodes. This can bedone by a locally run GA on each node combined with localsearch techniques for robustness. There is an ongoing researchon dynamic cluster design based on multiple metrics. We canalso use distributed GA to provide maximum disjoint coversto enhance the coverage area and network life. There is roomfor improving the throughput by extending the GAs fromsingle cell to multi-cell optimization, with inter-channel noisecancellation [121].

D. GAs, Complexity Theory, and Wireless NetworksIt has been stated in [9] that a system can be called a

‘complex system’ if the agents that constitute this systemincorporate: (i) diversity, (ii) connectivity, (iii) interdepen-dence, and (iv) adaptation. Modern wireless networks, bythese definitions, define a complex system, and are thusamenable to analysis, modeling, and optimization techniquesfrom complexity theory. Already, there has been work inapplying complexity science to networks [185] [186], but moreresearch is needed, particularly for the specific case of wirelessnetworks.

E. Using Big Data Techniques For Scaling GAsBig data techniques have the ability to scalably and reliably

process massive amounts of data using a cluster of commod-ity servers [187]. Big data processing techniques, such asGoogle’s MapReduce and other frameworks of its ilk, holda lot of promise for processing large problems in quick time.This is done by horizontal scaling, or ‘scaling out’, throughwhich more and more computers are employed to attack agiven problem in parallel. There is potentially a lot of scope inemploying these big data processing techniques to acceleratethe convergence of GAs. A steady trickle of research hasstarted emerging that is proposing the use of big data tools(such as the open-source Apache Hadoop platform which isbuilt on Google’s MapReduce paradigm) to scale up GAs[188] [189]. More research needs to be done to determine howto best utilize big data tools and techniques for improving GAperformance.

VI. CONCLUSIONS

We have provided a survey of the applications of geneticalgorithms (GAs) in wireless networking. In addition to pro-viding a self-contained introduction to common models andconfigurations of GAs, we have also provided a detailedsurvey of applications of GAs in wireless networking. Wehave categorized the applications of GAs in wireless network-ing both according to the wireless networking configuration,and according to the different kinds of GA techniques. Wehave also highlighted pitfalls and challenges in successfullyimplementing GAs in wireless networks. Finally, we havehighlighted a number of open issues and have identifiedpotential directions for future work.

REFERENCES

[1] F. Hillier and G. Lieberman, “Introduction to operations research,”McGraw Hill, New York, 2001.

[2] M. Ridley, Evolution. 3rd Edition. Blackwell, 2004.[3] M. Mitchell, An introduction to genetic algorithms. MIT press, 1998.[4] J. Holland, “Genetic algorithms, computer programs that evolve in

ways that even their creators do not fully understand,” ScientificAmerican, pp. 66–72, 1975.

[5] A. M. Turing, “Computing machinery and intelligence,” Mind, pp. 433–460, 1950.

[6] D. E. Goldberg and J. H. Holland, “Genetic algorithms and machinelearning,” Machine learning, vol. 3, no. 2, pp. 95–99, 1988.

[7] R. Highfield and M. Nowak, SuperCooperators: Evolution, Altruismand Human Behaviour (or why we need each other to succeed). TextPublishing, 2011.

[8] D. Bhandari, C. Murthy, and S. K. Pal, “Genetic algorithm withelitist model and its convergence,” International Journal of PatternRecognition and Artificial Intelligence, vol. 10, no. 06, pp. 731–747,1996.

[9] S. E. Page, Diversity and complexity. Princeton University Press, 2010.[10] A. Konak, D. W. Coit, and A. E. Smith, “Multi-objective optimization

using genetic algorithms: A tutorial,” Reliability Engineering & SystemSafety, vol. 91, no. 9, pp. 992–1007, 2006.

[11] L. Davis et al., Handbook of genetic algorithms, vol. 115. Van NostrandReinhold New York, 1991.

[12] W. Pedrycz and A. Vasilakos, Computational intelligence in telecom-munications networks. CRC press, 2010.

[13] M. Srinivas and L. M. Patnaik, “Genetic algorithms: A survey,”Computer, vol. 27, no. 6, pp. 17–26, 1994.

[14] K.-F. Man, K.-S. Tang, and S. Kwong, “Genetic algorithms: conceptsand applications,” IEEE Transactions on Industrial Electronics, vol. 43,no. 5, pp. 519–534, 1996.

[15] H. Ishibuchi, Y. Nojima, et al., “Comparison between single-objectiveand multi-objective genetic algorithms: Performance comparison andperformance measures,” in Evolutionary Computation, 2006. CEC2006. IEEE Congress on, pp. 1143–1150, IEEE, 2006.

[16] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitistmultiobjective genetic algorithm: Nsga-ii,” Evolutionary Computation,IEEE Transactions on, vol. 6, no. 2, pp. 182–197, 2002.

[17] D. H. Friend, M. EINainay, Y. Shi, and A. B. MacKenzie, “Architec-ture and performance of an island genetic algorithm-based cognitivenetwork,” in Consumer communications and networking conference,2008. CCNC 2008. 5th IEEE, pp. 993–997, IEEE, 2008.

24

[18] M. Y. ElNainay, F. Ge, Y. Wang, A. E. Hilal, Y. Shi, A. B. MacKen-zie, and C. W. Bostian, “Channel allocation for dynamic spectrumaccess cognitive networks using localized island genetic algorithm,”in Testbeds and Research Infrastructures for the Development ofNetworks & Communities and Workshops, 2009. TridentCom 2009. 5thInternational Conference on, pp. 1–3, IEEE, 2009.

[19] E. Alba and J. M. Troya, “A survey of parallel distributed geneticalgorithms,” Complexity, vol. 4, no. 4, pp. 31–52, 1999.

[20] J. Qadir, “Artificial intelligence based cognitive routing for cognitiveradio networks,” arXiv preprint arXiv:1309.0085, 2013.

[21] S. Sivanandam and S. Deepa, “Introduction to genetic algorithms.2008.”

[22] J. J. Grefenstette, “Optimization of control parameters for geneticalgorithms,” Systems, Man and Cybernetics, IEEE Transactions on,vol. 16, no. 1, pp. 122–128, 1986.

[23] C. W. Ahn and R. S. Ramakrishna, “A genetic algorithm for shortestpath routing problem and the sizing of populations,” EvolutionaryComputation, IEEE Transactions on, vol. 6, no. 6, pp. 566–579, 2002.

[24] D. Scaperoth, B. Le, T. Rondeau, D. Maldonado, C. W. Bostian, andS. Harrison, “Cognitive radio platform development for interoperabil-ity,” in Military Communications Conference, 2006. MILCOM 2006.IEEE, pp. 1–6, IEEE, 2006.

[25] J. H. Holland, Hidden order: How adaptation builds complexity. BasicBooks, 1995.

[26] T. Harford, Adapt: Why success always starts with failure. Macmillan,2011.

[27] K. Sastry, D. Goldberg, and G. Kendall, “Genetic algorithms,” in Searchmethodologies, pp. 97–125, Springer, 2005.

[28] B. A. Fette, Cognitive radio technology. Access Online via Elsevier,2009.

[29] J. Horn, N. Nafpliotis, and D. E. Goldberg, “A niched Pareto geneticalgorithm for multiobjective optimization,” in Evolutionary Compu-tation, 1994. IEEE World Congress on Computational Intelligence.,Proceedings of the First IEEE Conference on, pp. 82–87, Ieee, 1994.

[30] J. Knowles and D. Corne, “Bounded Pareto archiving: Theory andpractice,” in Metaheuristics for multiobjective optimisation, pp. 39–64,Springer, 2004.

[31] H. Cheng and S. Yang, “Multi-population genetic algorithms withimmigrants scheme for dynamic shortest path routing problems inmobile ad hoc networks,” in Applications of Evolutionary Computation,pp. 562–571, Springer, 2010.

[32] M. Gen, R. Cheng, and Q. Wang, “Genetic algorithms for solvingshortest path problems,” in Evolutionary Computation, 1997., IEEEInternational Conference on, pp. 401–406, IEEE, 1997.

[33] C. Davies and P. Lingras, “Genetic algorithms for rerouting shortestpaths in dynamic and stochastic networks,” European Journal ofOperational Research, vol. 144, no. 1, pp. 27–38, 2003.

[34] S. Yang, H. Cheng, and F. Wang, “Genetic algorithms with immigrantsand memory schemes for dynamic shortest path routing problems inmobile ad hoc networks,” Systems, Man, and Cybernetics, Part C:Applications and Reviews, IEEE Transactions on, vol. 40, no. 1, pp. 52–63, 2010.

[35] L. Badia, A. Botta, and L. Lenzini, “A genetic approach to joint routingand link scheduling for wireless mesh networks,” Ad Hoc Networks,vol. 7, no. 4, pp. 654–664, 2009.

[36] J. N. Al-Karaki, R. Ul-Mustafa, and A. E. Kamal, “Data aggregationand routing in wireless sensor networks: Optimal and heuristic algo-rithms,” Computer networks, vol. 53, no. 7, pp. 945–960, 2009.

[37] B. Lorenzo and S. Glisic, “Optimal routing and traffic scheduling formultihop cellular networks using genetic algorithm,” Mobile Comput-ing, IEEE Transactions on, vol. 12, no. 11, pp. 2274–2288, 2013.

[38] I. Apetroaei, I.-A. Oprea, B.-E. Proca, and L. Gheorghe, “Geneticalgorithms applied in routing protocols for wireless sensor networks,”in Roedunet International Conference (RoEduNet), 2011 10th, pp. 1–6,IEEE, 2011.

[39] Y.-S. Yen, Y.-K. Chan, H.-C. Chao, and J. H. Park, “A genetic algorithmfor energy-efficient based multicast routing on manets,” ComputerCommunications, vol. 31, no. 4, pp. 858–869, 2008.

[40] T.-C. Chiang, C.-H. Liu, and Y.-M. Huang, “A near-optimal multicastscheme for mobile ad hoc networks using a hybrid genetic algorithm,”Expert Systems with Applications, vol. 33, no. 3, pp. 734–742, 2007.

[41] M. Al-Ghazal, A. El-Sayed, and H. Kelash, “Routing optimlzationusing genetic algorithm in ad hoc networks,” in Signal Processingand Information Technology, 2007 IEEE International Symposium on,pp. 497–503, IEEE, 2007.

[42] N. Banerjee and S. K. Das, “Fast determination of qos-based multicastroutes in wireless networks using genetic algorithm,” in Communi-cations, 2001. ICC 2001. IEEE International Conference on, vol. 8,pp. 2588–2592, IEEE, 2001.

[43] S. Salcedo-Sanz, C. Bousoño-Calzón, and A. R. Figueiras-Vidal, “Amixed neural-genetic algorithm for the broadcast scheduling prob-lem,” Wireless Communications, IEEE Transactions on, vol. 2, no. 2,pp. 277–283, 2003.

[44] F. C. NK and S. D. K. Viswanatha, “Routing algorithm using mobileagents and genetic algorithm,”

[45] M. R. Sherif, I. W. Habib, M. Nagshineh, and P. Kermani, “Adaptiveallocation of resources and call admission control for wireless atmusing genetic algorithms,” Selected Areas in Communications, IEEEJournal on, vol. 18, no. 2, pp. 268–282, 2000.

[46] D. Gözüpek and F. Alagöz, “Genetic algorithm-based scheduling incognitive radio networks under interference temperature constraints,”International Journal of Communication Systems, vol. 24, no. 2,pp. 239–257, 2011.

[47] H. T. Cheng and W. Zhuang, “Novel packet-level resource allocationwith effective QoS provisioning for wireless mesh networks,” WirelessCommunications, IEEE Transactions on, vol. 8, no. 2, pp. 694–700,2009.

[48] R. B. Lopez, S. M. Sanchez, E. M. Fernandez, R. D. Souza, andH. Alves, “Genetic algorithm aided transmit power control in cognitiveradio networks,” in Cognitive Radio Oriented Wireless Networks andCommunications (CROWNCOM), 2014 9th International Conferenceon, pp. 61–66, IEEE, 2014.

[49] A. Roy, N. Banerjee, and S. K. Das, “An efficient multi-objective qos-routing algorithm for wireless multicasting,” in Vehicular TechnologyConference, 2002. VTC Spring 2002. IEEE 55th, vol. 3, pp. 1160–1164,IEEE, 2002.

[50] H. Jiang, X. Yang, K. Yin, S. Zhang, and J. A. Cristoforo, “Multi-pathqos-aware web service composition using variable length chromosomegenetic algorithm,” Information Technology Journal, vol. 10, no. 1,pp. 113–119, 2011.

[51] G. Canfora, M. Di Penta, R. Esposito, and M. L. Villani, “An approachfor qos-aware service composition based on genetic algorithms,” inProceedings of the 7th annual conference on Genetic and evolutionarycomputation, pp. 1069–1075, ACM, 2005.

[52] D. Li, Q. Zhang, C.-N. Chuah, and S. Ben Yoo, “Multi-source multi-path video streaming over wireless mesh networks,” in Circuits andSystems, 2006. ISCAS 2006. Proceedings. 2006 IEEE InternationalSymposium on, pp. 4–pp, IEEE, 2006.

[53] G. H. EkbataniFard, R. Monsefi, M.-R. Akbarzadeh-T, M. Yaghmaee,et al., “A multi-objective genetic algorithm based approach for en-ergy efficient qos-routing in two-tiered wireless sensor networks,” inWireless Pervasive Computing (ISWPC), 2010 5th IEEE InternationalSymposium on, pp. 80–85, IEEE, 2010.

[54] K. Deb, S. Agrawal, A. Pratap, and T. Meyarivan, “A fast elitist non-dominated sorting genetic algorithm for multi-objective optimization:Nsga-ii,” Lecture notes in computer science, vol. 1917, pp. 849–858,2000.

[55] L. Barolli, A. Koyama, and N. Shiratori, “A QoS routing method forad-hoc networks based on genetic algorithm,” in Database and ExpertSystems Applications, 2003. Proceedings. 14th International Workshopon, pp. 175–179, IEEE, 2003.

[56] A. Roy and S. K. Das, “Qm2rp: a qos-based mobile multicast routingprotocol using multi-objective genetic algorithm,” Wireless Networks,vol. 10, no. 3, pp. 271–286, 2004.

[57] X. Cui, C. Lin, and Y. Wei, “A multiobjective model for QoS multicastrouting based on genetic algorithm,” in Computer Networks and MobileComputing, 2003. ICCNMC 2003. 2003 International Conference on,pp. 49–53, IEEE, 2003.

[58] T. Lu and J. Zhu, “Genetic algorithm for energy-efficient QoS multicastrouting,” Communications Letters, IEEE, vol. 17, no. 1, pp. 31–34,2013.

[59] Y.-S. Yen, H.-C. Chao, R.-S. Chang, and A. Vasilakos, “Flooding-limited and multi-constrained QoS multicast routing based on thegenetic algorithm for manets,” Mathematical and Computer Modelling,vol. 53, no. 11, pp. 2238–2250, 2011.

[60] J. Huang and Y. Liu, “Moeaq: A qos-aware multicast routing algorithmfor manet,” Expert Systems with Applications, vol. 37, no. 2, pp. 1391–1399, 2010.

[61] B. Sun, S. Pi, C. Gui, Y. Zeng, B. Yan, W. Wang, and Q. Qin, “Multipleconstraints QoS multicast routing optimization algorithm in manetbased on ga,” Progress in Natural Science, vol. 18, no. 3, pp. 331–336, 2008.

25

[62] J. Abdullah, “Multiobjectives ga-based QoS routing protocol for mo-bile ad hoc network,” International Journal of Grid and DistributedComputing, vol. 3, no. 4, pp. 57–68, 2010.

[63] E. Gelenbe, P. Liu, and J. Laine, “Genetic algorithms for autonomicroute discovery,” in Distributed Intelligent Systems: Collective Intel-ligence and Its Applications, 2006. DIS 2006. IEEE Workshop on,pp. 371–376, IEEE, 2006.

[64] G. Di Fatta, F. Hoffmann, G. Lo Re, and A. Urso, “A genetic algorithmfor the design of a fuzzy controller for active queue management,”Systems, Man, and Cybernetics, Part C: Applications and Reviews,IEEE Transactions on, vol. 33, no. 3, pp. 313–324, 2003.

[65] A. Selamat, M. H. Selamat, et al., “Routing algorithm of mobile agentsfor query retrieval using genetic algorithm,” Malaysian Journal ofComputer Science, vol. 17, no. 2, pp. 1–10, 2004.

[66] C.-Y. Hsu, J.-L. C. Wu, S.-T. Wang, and C.-Y. Hong, “Survivable anddelay-guaranteed backbone wireless mesh network design,” Journal ofParallel and Distributed Computing, vol. 68, no. 3, pp. 306–320, 2008.

[67] F. Zeng and Z. Chen, “Load balancing placement of gateways inwireless mesh networks with QoS constraints,” in Young ComputerScientists, 2008. ICYCS 2008. The 9th International Conference for,pp. 445–450, IEEE, 2008.

[68] H. Cheng, S. Yang, and J. Cao, “Dynamic genetic algorithms for thedynamic load balanced clustering problem in mobile ad hoc networks,”Expert Systems with Applications, vol. 40, no. 4, pp. 1381–1392, 2013.

[69] J. He, S. Ji, M. Yan, Y. Pan, and Y. Li, “Load–balanced CDS construc-tion in wireless sensor networks via genetic algorithm,” InternationalJournal of Sensor Networks, vol. 11, no. 3, pp. 166–178, 2012.

[70] T. Scully and K. N. Brown, “Wireless LAN load balancing with geneticalgorithms,” Knowledge-Based Systems, vol. 22, no. 7, pp. 529–534,2009.

[71] A. Malik, J. Qadir, B. Ahmad, K.-L. A. Yau, and U. Ullah, “Qos inieee 802.11-based wireless networks: A contemporary survey,” arXivpreprint arXiv:1411.2852, 2014.

[72] J. Choi, T. Kwon, Y. Choi, and M. Naghshineh, “Call admission controlfor multimedia services in mobile cellular networks: a markov decisionapproach,” in Computers and Communications, 2000. Proceedings.ISCC 2000. Fifth IEEE Symposium on, pp. 594–599, IEEE, 2000.

[73] Y. Xiao, C. P. Chen, and Y. Wang, “A near optimal call admission con-trol with genetic algorithm for multimedia services in wireless/mobilenetworks,” in National Aerospace and Electronics Conference, 2000.NAECON 2000. Proceedings of the IEEE 2000, pp. 787–792, IEEE,2000.

[74] C. Gao, M. Cai, and H. Chen, “Qos-aware service composition basedon tree-coded genetic algorithm,” in Computer Software and Applica-tions Conference, 2007. COMPSAC 2007. 31st Annual International,vol. 1, pp. 361–367, IEEE, 2007.

[75] D. Lin and F. Labeau, “Accelerated genetic algorithm for bandwidthallocation in view of emi for wireless healthcare,” in Wireless Commu-nications and Networking Conference (WCNC), 2012 IEEE, pp. 3312–3317, IEEE, 2012.

[76] G. Kandavanam, D. Botvich, S. Balasubramaniam, and B. Jennings,“A hybrid genetic algorithm/variable neighborhood search approach tomaximizing residual bandwidth of links for route planning,” in ArtificalEvolution, pp. 49–60, Springer, 2010.

[77] D. Karabudak, C.-C. Hung, and B. Bing, “A call admission controlscheme using genetic algorithms,” in Proceedings of the 2004 ACMsymposium on Applied computing, pp. 1151–1158, ACM, 2004.

[78] S. Vedantham and S. S. Iyengar, “The bandwidth allocation problemin the atm network model is np-complete,” Information ProcessingLetters, vol. 65, no. 4, pp. 179–182, 1998.

[79] A. Riedl, “A hybrid genetic algorithm for routing optimization in ipnetworks utilizing bandwidth and delay metrics,” in IP Operations andManagement, 2002 IEEE Workshop on, pp. 166–170, IEEE, 2002.

[80] H. Kobayashi, M. Munetomo, K. Akama, and Y. Sato, “Designing a dis-tributed algorithm for bandwidth allocation with a genetic algorithm,”Systems and Computers in Japan, vol. 35, no. 3, pp. 37–45, 2004.

[81] T. A. Al-Qahtani, M. J. Abedin, and S. I. Ahson, “Dynamic routing inhomogenous atm networks using genetic algorithms,” in EvolutionaryComputation Proceedings, 1998. IEEE World Congress on Compu-tational Intelligence., The 1998 IEEE International Conference on,pp. 114–119, IEEE, 1998.

[82] T. Ozugur, A. Bellary, and F. Sarkar, “Multiobjective hierarchical2G/3G mobility management optimization: niched Pareto genetic al-gorithm,” in Global Telecommunications Conference, 2001. GLOBE-COM’01. IEEE, vol. 6, pp. 3681–3685, IEEE, 2001.

[83] K. Krishnakumar, “Micro-genetic algorithms for stationary and non-stationary function optimization,” in 1989 Advances in Intelligent

Robotics Systems Conference, pp. 289–296, International Society forOptics and Photonics, 1990.

[84] Z. Zhao, Z. Peng, S. Zheng, and J. Shang, “Cognitive radio spectrumallocation using evolutionary algorithms,” Wireless Communications,IEEE Transactions on, vol. 8, no. 9, pp. 4421–4425, 2009.

[85] A. Narayanan and M. Moore, “Quantum-inspired genetic algorithms,”in Evolutionary Computation, 1996., Proceedings of IEEE Interna-tional Conference on, pp. 61–66, IEEE, 1996.

[86] S. Bhattacharjee, A. Konar, and A. K. Nagar, “Channel allocationfor a single cell cognitive radio network using genetic algorithm,”in Innovative mobile and internet services in ubiquitous computing(IMIS), 2011 fifth international conference on, pp. 258–264, IEEE,2011.

[87] S. Ali, A. Munir, S. B. Qaisar, and J. Qadir, “A genetic algorithmassisted resource management scheme for reliable multimedia de-livery over cognitive networks,” in Computational Science and ItsApplications–ICCSA 2012, pp. 352–367, Springer, 2012.

[88] Y. Zhenhua, Y. Guangwen, L. Shanwei, and Z. Qishan, “A modifiedimmune genetic algorithm for channel assignment problems in cellularradio networks,” in Intelligent System Design and Engineering Applica-tion (ISDEA), 2010 International Conference on, vol. 2, pp. 823–826,IEEE, 2010.

[89] S. Pinagapany and A. Kulkarni, “Solving channel allocation problemin cellular radio networks using genetic algorithm,” in CommunicationSystems Software and Middleware and Workshops, 2008. COMSWARE2008. 3rd International Conference on, pp. 239–244, IEEE, 2008.

[90] S. José-Revuelta et al., “A new adaptive genetic algorithm for fixedchannel assignment,” Information Sciences, vol. 177, no. 13, pp. 2655–2678, 2007.

[91] M. A. Lima, A. F. Araujo, and A. C. Cesar, “Adaptive geneticalgorithms for dynamic channel assignment in mobile cellular com-munication systems,” Vehicular Technology, IEEE Transactions on,vol. 56, no. 5, pp. 2685–2696, 2007.

[92] S. H. Wong and I. Wassell, “Dynamic channel allocation using agenetic algorithm for a tdd broadband fixed wireless access network,”in IASTED International Conference in Wireless and Optical Commu-nications, Banff, Canada, pp. 521–526, 2002.

[93] T. Fang and L.-P. Chau, “Gop-based channel rate allocation usinggenetic algorithm for scalable video streaming over error-prone net-works,” Image Processing, IEEE Transactions on, vol. 15, no. 6,pp. 1323–1330, 2006.

[94] I. E. Kassotakis, M. E. Markaki, and A. V. Vasilakos, “A hybrid geneticapproach for channel reuse in multiple access telecommunication net-works,” Selected Areas in Communications, IEEE Journal on, vol. 18,no. 2, pp. 234–243, 2000.

[95] M. Cheng and L. F. Chang, “Wireless dynamic channel assignmentperformance under packet data traffic,” Selected Areas in Communica-tions, IEEE Journal on, vol. 17, no. 7, pp. 1257–1269, 1999.

[96] Y. Akaiwa and H. Andoh, “Channel segregation-a self-organized dy-namic channel allocation method: application to tdma/fdma microcel-lular system,” Selected Areas in Communications, IEEE Journal on,vol. 11, no. 6, pp. 949–954, 1993.

[97] J.-S. Kim, S. Park, P. Dowd, and N. Nasrabadi, “Channel assignmentin cellular radio using genetic algorithms,” Wireless Personal Commu-nications, vol. 3, no. 3, pp. 273–286, 1996.

[98] C. Y. Ngo and V. O. Li, “Fixed channel assignment in cellular radionetworks using a modified genetic algorithm,” Vehicular Technology,IEEE Transactions on, vol. 47, no. 1, pp. 163–172, 1998.

[99] Q. Zhang, J. Wang, C. Jin, and Q. Zeng, “Localization algorithmfor wireless sensor network based on genetic simulated annealingalgorithm,” in Wireless Communications, Networking and Mobile Com-puting, 2008. WiCOM’08. 4th International Conference on, pp. 1–5,IEEE, 2008.

[100] S. Yun, J. Lee, W. Chung, and E. Kim, “Centroid localization methodin wireless sensor networks using tsk fuzzy modeling,” in Internationalsymposium on advanced intelligent systems, pp. 971–974, 2008.

[101] V. Tam, K.-Y. Cheng, and K.-S. Lui, “Using micro-genetic algorithmsto improve localization in wireless sensor networks,” Journal of Com-munications, vol. 1, no. 4, pp. 1–10, 2006.

[102] G.-F. Nan, M.-Q. Li, and J. Li, “Estimation of node localization witha real-coded genetic algorithm in wsns,” in Machine Learning andCybernetics, 2007 International Conference on, vol. 2, pp. 873–878,IEEE, 2007.

[103] S. Yun, J. Lee, W. Chung, E. Kim, and S. Kim, “A soft computingapproach to localization in wireless sensor networks,” Expert Systemswith Applications, vol. 36, no. 4, pp. 7552–7561, 2009.

26

[104] M. N. Smadi, S. C. Ghosh, A. A. Farid, T. D. Todd, and S. Hranilovic,“Free-space optical gateway placement in hybrid wireless mesh net-works,” Journal of Lightwave Technology, vol. 27, no. 14, pp. 2688–2697, 2009.

[105] F. Xhafa, C. Sánchez, and L. Barolli, “Genetic algorithms for efficientplacement of router nodes in wireless mesh networks,” in AdvancedInformation Networking and Applications (AINA), 2010 24th IEEEInternational Conference on, pp. 465–472, IEEE, 2010.

[106] J. Kusyk, C. S. Sahin, M. Umit Uyar, E. Urrea, and S. Gundry, “Self-organization of nodes in mobile ad hoc networks using evolutionarygames and genetic algorithms,” Journal of Advanced Research, vol. 2,no. 3, pp. 253–264, 2011.

[107] C. S. Sahin, E. Urrea, M. U. Uyar, M. Conner, I. Hokelek, G. Bertoli,and C. Pizzo, “Uniform distribution of mobile agents using geneticalgorithms for military applications in manets,” in Military Communi-cations Conference, 2008. MILCOM 2008. IEEE, pp. 1–7, IEEE, 2008.

[108] F. Hoffmann, D. Medina, and A. Wolisz, “Optimization of routingand gateway allocation in aeronautical ad hoc networks using geneticalgorithms,” in Wireless Communications and Mobile Computing Con-ference (IWCMC), 2011 7th International, pp. 1391–1396, IEEE, 2011.

[109] W. Youssef and M. Younis, “Intelligent gateways placement for reduceddata latency in wireless sensor networks,” in Communications, 2007.ICC’07. IEEE International Conference on, pp. 3805–3810, IEEE,2007.

[110] A. Tripathi, P. Gupta, A. Trivedi, and R. Kala, “Wireless sensor nodeplacement using hybrid genetic programming and genetic algorithms,”International Journal of Intelligent Information Technologies (IJIIT),vol. 7, no. 2, pp. 63–83, 2011.

[111] S. Pandey, S. Dong, P. Agrawal, and K. Sivalingam, “A hybrid approachto optimize node placements in hierarchical heterogeneous networks,”in Wireless Communications and Networking Conference, 2007. WCNC2007. IEEE, pp. 3918–3923, IEEE, 2007.

[112] Y. Xu and X. Yao, “A GA approach to the optimal placement ofsensors in wireless sensor networks with obstacles and preferences,” inConsumer Communications and Networking Conference, 2006. CCNC2006. 3rd IEEE, vol. 1, pp. 127–131, IEEE, 2006.

[113] E. Alba, B. Dorronsoro, F. Luna, A. J. Nebro, P. Bouvry, and L. Hogie,“A cellular multi-objective genetic algorithm for optimal broadcastingstrategy in metropolitan manets,” Computer Communications, vol. 30,no. 4, pp. 685–697, 2007.

[114] X.-M. Hu, J. Zhang, Y. Yu, H.-H. Chung, Y.-L. Li, Y.-H. Shi, and X.-N. Luo, “Hybrid genetic algorithm using a forward encoding schemefor lifetime maximization of wireless sensor networks,” EvolutionaryComputation, IEEE Transactions on, vol. 14, no. 5, pp. 766–781, 2010.

[115] C. S. Sahin, E. Urrea, M. U. Uyar, M. Conner, I. Hokelek, M. Conner,G. Bertoli, and C. Pizzo, “Genetic algorithms for self-spreading nodesin manets,” in Proceedings of the 10th Annual Conference on Geneticand Evolutionary Computation, GECCO ’08, (New York, NY, USA),pp. 1141–1142, ACM, 2008.

[116] C.-C. Lai, C.-K. Ting, and R.-S. Ko, “An effective genetic algorithm toimprove wireless sensor network lifetime for large-scale surveillanceapplications,” in Evolutionary Computation, 2007. CEC 2007. IEEECongress on, pp. 3531–3538, IEEE, 2007.

[117] F. P. Quintao, F. G. Nakamura, and G. R. Mateus, “Evolutionaryalgorithm for the dynamic coverage problem applied to wireless sensornetworks design,” in Evolutionary Computation, 2005. The 2005 IEEECongress on, vol. 2, pp. 1589–1596, IEEE, 2005.

[118] A. P. Bhondekar, R. Vig, M. L. Singla, C. Ghanshyam, and P. Kapur,“Genetic algorithm based node placement methodology for wirelesssensor networks,” in Proceedings of the international multiconferenceof engineers and computer scientists, vol. 1, pp. 18–20, Citeseer, 2009.

[119] D. Jourdan and O. L. de Weck, “Layout optimization for a wirelesssensor network using a multi-objective genetic algorithm,” in VehicularTechnology Conference, 2004. VTC 2004-Spring. 2004 IEEE 59th,vol. 5, pp. 2466–2470, IEEE, 2004.

[120] R. Pries, D. Staehle, M. Stoykova, B. Staehle, and P. Tran-Gia,“Wireless mesh network planning and optimization through geneticalgorithms,” in Advances in Mesh Networks, 2009. MESH 2009. SecondInternational Conference on, pp. 55–61, IEEE, 2009.

[121] V. Ayyadurai, K. Moessner, and R. Tafazolli, “Multihop cellularnetwork optimization using genetic algorithms,” in Proceedings of the7th International Conference on Network and Services Management,pp. 399–403, International Federation for Information Processing,2011.

[122] S. Ghosh, P. Ghosh, K. Basu, and S. K. Das, “GaMa: an evolution-ary algorithmic approach for the design of mesh-based radio access

networks,” in Local Computer Networks, 2005. 30th Anniversary. TheIEEE Conference on, pp. 8–pp, IEEE, 2005.

[123] W.-H. Sheen, S.-J. Lin, and C.-C. Huang, “Downlink optimization andperformance of relay-assisted cellular networks in multicell environ-ments,” Vehicular Technology, IEEE Transactions on, vol. 59, no. 5,pp. 2529–2542, 2010.

[124] L. Chiaraviglio, D. Ciullo, G. Koutitas, M. Meo, and L. Tassiulas,“Energy-efficient planning and management of cellular networks,” inWireless On-demand Network Systems and Services (WONS), 2012 9thAnnual Conference on, pp. 159–166, IEEE, 2012.

[125] S. Hussain, A. W. Matin, and O. Islam, “Genetic algorithm forhierarchical wireless sensor networks,” Journal of Networks, vol. 2,no. 5, pp. 87–97, 2007.

[126] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-efficient communication protocol for wireless microsensor networks,”in System Sciences, 2000. Proceedings of the 33rd Annual HawaiiInternational Conference on, pp. 10–pp, IEEE, 2000.

[127] J. Zhang, Y. Lin, C. Zhou, and J. Ouyang, “Optimal model for energy-efficient clustering in wireless sensor networks using global simulatedannealing genetic algorithm,” in Intelligent Information TechnologyApplication Workshops, 2008. IITAW’08. International Symposium on,pp. 656–660, IEEE, 2008.

[128] J. D. Stumpf, X. Feng, and R. W. Kelnhofer, “An enhanced operator-oriented genetic search algorithm,” in Evolutionary Computation, 1994.IEEE World Congress on Computational Intelligence., Proceedings ofthe First IEEE Conference on, pp. 235–238, IEEE, 1994.

[129] P.-C. Huruiala, A. Urzica, and L. Gheorghe, “Hierarchical routing pro-tocol based on evolutionary algorithms for wireless sensor networks,”in Roedunet International Conference (RoEduNet), 2010 9th, pp. 387–392, IEEE, 2010.

[130] K. Maksuriwong, V. Varavithya, and N. Chaiyaratana, “Wireless LANaccess point placement using a multi-objective genetic algorithm,” inSystems, Man and Cybernetics, 2003. IEEE International Conferenceon, vol. 2, pp. 1944–1949, IEEE, 2003.

[131] L. Nagy and L. Farkas, “Indoor base station location optimizationusing genetic algorithms,” in Personal, Indoor and Mobile RadioCommunications, 2000. PIMRC 2000. The 11th IEEE InternationalSymposium on, vol. 2, pp. 843–846, IEEE, 2000.

[132] K. Lieska, E. Laitinen, and J. Lahteenmaki, “Radio coverage optimiza-tion with genetic algorithms,” in Personal, Indoor and Mobile RadioCommunications, 1998. The Ninth IEEE International Symposium on,vol. 1, pp. 318–322, IEEE, 1998.

[133] S. K. Das, N. Banerjee, and A. Roy, “Solving optimization problems inwireless networks using genetic algorithms,” Handbook of BioinspiredAlgorithms and Applications, p. 219, 2006.

[134] J. M. Johnson and Y. Rahmat-Samii, “Genetic algorithm optimizationof wireless communication networks,” in Antennas and PropagationSociety International Symposium, 1995. AP-S. Digest, vol. 4, pp. 1964–1967, IEEE, 1995.

[135] R. Venkatesan and V. Kumar, “A genetic algorithms approach to growthphase forecasting of wireless subscribers,” International Journal ofForecasting, vol. 18, no. 4, pp. 625–646, 2002.

[136] P. C. Pendharkar, “Genetic algorithm based neural network approachesfor predicting churn in cellular wireless network services,” ExpertSystems with Applications, vol. 36, no. 3, pp. 6714–6720, 2009.

[137] J. Jia, J. Chen, G. Chang, and Z. Tan, “Energy efficient coveragecontrol in wireless sensor networks based on multi-objective geneticalgorithm,” Computers & Mathematics with Applications, vol. 57,no. 11, pp. 1756–1766, 2009.

[138] K. P. Ferentinos and T. A. Tsiligiridis, “Adaptive design optimizationof wireless sensor networks using genetic algorithms,” ComputerNetworks, vol. 51, no. 4, pp. 1031–1051, 2007.

[139] S. Jin, M. Zhou, and A. S. Wu, “Sensor network optimization usinga genetic algorithm,” in Proceedings of the 7th World Multiconferenceon Systemics, Cybernetics and Informatics, pp. 109–116, 2003.

[140] S. Sengupta, S. Das, M. Nasir, A. V. Vasilakos, and W. Pedrycz, “Anevolutionary multiobjective sleep-scheduling scheme for differentiatedcoverage in wireless sensor networks,” Systems, Man, and Cybernetics,Part C: Applications and Reviews, IEEE Transactions on, vol. 42, no. 6,pp. 1093–1102, 2012.

[141] Q. Zhang, J. Wang, C. Jin, J. Ye, C. Ma, and W. Zhang, “Geneticalgorithm based wireless sensor network localization,” in NaturalComputation, 2008. ICNC’08. Fourth International Conference on,vol. 1, pp. 608–613, IEEE, 2008.

[142] D. Niculescu and B. Nath, “Ad hoc positioning system (aps) usingaoa,” in INFOCOM 2003. Twenty-Second Annual Joint Conference

27

of the IEEE Computer and Communications. IEEE Societies, vol. 3,pp. 1734–1743, IEEE, 2003.

[143] T.-C. Liang, T.-C. Wang, and Y. Ye, “A gradient search method to roundthe semidefinite programming relaxation solution for ad hoc wirelesssensor network localization,” Sanford University, formal report, vol. 5,2004.

[144] A. A. Kannan, G. Mao, and B. Vucetic, “Simulated annealing basedlocalization in wireless sensor network,” in Local Computer Networks,2005. 30th Anniversary. The IEEE Conference on, pp. 2–pp, IEEE,2005.

[145] C. T. Chou, A. Misra, and J. Qadir, “Low-latency broadcast in multiratewireless mesh networks,” Selected Areas in Communications, IEEEJournal on, vol. 24, no. 11, pp. 2081–2091, 2006.

[146] T. W. Rondeau, B. Le, C. J. Rieser, and C. W. Bostian, “Cognitiveradios with genetic algorithms: Intelligent control of software definedradios,” in Software defined radio forum technical conference, pp. C3–C8, Citeseer, 2004.

[147] E. Alba, Parallel metaheuristics: a new class of algorithms, vol. 47.Wiley. com, 2005.

[148] F. Ye, R. Yang, and Y. Li, “Genetic algorithm based spectrum assign-ment model in cognitive radio networks,” in Information Engineeringand Computer Science (ICIECS), 2010 2nd International Conferenceon, pp. 1–4, IEEE, 2010.

[149] C. J. Rieser, Biologically inspired cognitive radio engine model utilizingdistributed genetic algorithms for secure and robust wireless commu-nications and networking. PhD thesis, Virginia Polytechnic Instituteand State University, 2004.

[150] J. Mitola III, Cognitive Radio Architecture: The Engineering Founda-tions of Radio XML. John Wiley and Sons, 2006.

[151] J. M. Kim, S. H. Sohn, N. Han, G. Zheng, Y. M. Kim, and J. K.Lee, “Cognitive radio software testbed using dual optimization ingenetic algorithm,” in Cognitive Radio Oriented Wireless Networks andCommunications, 2008. CrownCom 2008. 3rd International Conferenceon, pp. 1–6, IEEE, 2008.

[152] J. Qadir, N. Ahmed, and N. Ahad, “Building programmable wirelessnetworks: an architectural survey,” EURASIP Journal on WirelessCommunications and Networking, vol. 2014, no. 1, p. 172, 2014.

[153] D. Bojic, E. Sasaki, N. Cvijetic, T. Wang, J. Kuno, J. Lessmann,S. Schmid, H. Ishii, and S. Nakamura, “Advanced wireless and opticaltechnologies for small-cell mobile backhaul with dynamic software-defined management,” IEEE Commun. Mag, vol. 51, no. 9, pp. 86–93,2013.

[154] J. W. Weibull, Evolutionary game theory. MIT press, 1997.[155] B. Wang, K. R. Liu, and T. C. Clancy, “Evolutionary cooperative

spectrum sensing game: how to collaborate?,” Communications, IEEETransactions on, vol. 58, no. 3, pp. 890–900, 2010.

[156] H. Tembine, E. Altman, R. El-Azouzi, and Y. Hayel, “Evolutionarygames in wireless networks,” Systems, Man, and Cybernetics, Part B:Cybernetics, IEEE Transactions on, vol. 40, no. 3, pp. 634–646, 2010.

[157] E. Altman, Y. Hayel, and H. Kameda, “Evolutionary dynamics andpotential games in non-cooperative routing,” in Modeling and Opti-mization in Mobile, Ad Hoc and Wireless Networks and Workshops,2007. WiOpt 2007. 5th International Symposium on, pp. 1–5, IEEE,2007.

[158] D. Niyato and E. Hossain, “Dynamics of network selection in hetero-geneous wireless networks: an evolutionary game approach,” VehicularTechnology, IEEE Transactions on, vol. 58, no. 4, pp. 2008–2017, 2009.

[159] E. Altman, R. Elazouzi, Y. Hayel, and H. Tembine, “An evolutionarygame approach for the design of congestion control protocols inwireless networks,” in Modeling and Optimization in Mobile, AdHoc, and Wireless Networks and Workshops, 2008. WiOPT 2008. 6thInternational Symposium on, pp. 547–552, IEEE, 2008.

[160] M. P. Anastasopoulos, D. K. Petraki, R. Kannan, and A. V. Vasilakos,“TCP throughput adaptation in WiMax networks using replicatordynamics,” Systems, Man, and Cybernetics, Part B: Cybernetics, IEEETransactions on, vol. 40, no. 3, pp. 647–655, 2010.

[161] K. Zhu, D. Niyato, and P. Wang, “Optimal bandwidth allocationwith dynamic service selection in heterogeneous wireless networks,”in Global Telecommunications Conference (GLOBECOM 2010), 2010IEEE, pp. 1–5, IEEE, 2010.

[162] Z. Han, Chapter 6, Game theory in wireless and communicationnetworks: theory, models, and applications. Cambridge UniversityPress, 2012.

[163] M. Crepinšek, S.-H. Liu, and M. Mernik, “Exploration and exploita-tion in evolutionary algorithms: a survey,” ACM Computing Surveys(CSUR), vol. 45, no. 3, p. 35, 2013.

[164] D. E. Moriarty, A. C. Schultz, and J. J. Grefenstette, “Evolutionaryalgorithms for reinforcement learning,” Journal of Artificial IntelligenceResearch, vol. 11, pp. 241–276, 1999.

[165] A. Ferreira, A. Goldman, and J. Monteiro, “Performance evaluation ofrouting protocols for MANETs with known connectivity patterns usingevolving graphs,” Wireless Networks, vol. 16, no. 3, pp. 627–640, 2010.

[166] B. B. Xuan, A. Ferreira, and A. Jarry, “Computing shortest, fastest,and foremost journeys in dynamic networks,” International Journal ofFoundations of Computer Science, vol. 14, no. 02, pp. 267–285, 2003.

[167] A. Gajduk, Z. Utkovski, L. Basnarkov, and L. Kocarev, “Energy-efficiency in decentralized wireless networks: A game-theoretic ap-proach inspired by evolutionary biology,” CoRR, vol. abs/1405.3491,2014.

[168] N. Baldo and M. Zorzi, “Fuzzy logic for cross-layer optimization incognitive radio networks,” IEEE Communications Magazine, vol. 46,no. 4, p. 64, 2008.

[169] L. A. Zadeh, “Fuzzy logic, neural networks, and soft computing,”Communications of the ACM, vol. 37, no. 3, pp. 77–84, 1994.

[170] L. Wang, Soft Computing in Communications, vol. 136. Springer, 2003.[171] W. Luo, “A quantum genetic algorithm based QoS routing protocol

for wireless sensor networks,” in Software Engineering and ServiceSciences (ICSESS), 2010 IEEE International Conference on, pp. 37–40, IEEE, 2010.

[172] P. Moscato and C. Cotta, “A modern introduction to memetic algo-rithms,” in Handbook of Metaheuristics, pp. 141–183, Springer, 2010.

[173] A. Konstantinidis, K. Yang, H.-H. Chen, and Q. Zhang, “Energy-aware topology control for wireless sensor networks using memeticalgorithms,” Computer Communications, vol. 30, no. 14, pp. 2753–2764, 2007.

[174] A. Quintero and S. Pierre, “A memetic algorithm for assigning cells toswitches in cellular mobile networks,” Communications Letters, IEEE,vol. 6, no. 11, pp. 484–486, 2002.

[175] D. Whitley, T. Starkweather, and C. Bogart, “Genetic algorithms andneural networks: Optimizing connections and connectivity,” Parallelcomputing, vol. 14, no. 3, pp. 347–361, 1990.

[176] D. E. Goldberg, “Sizing populations for serial and parallel geneticalgorithms,” in Proceedings of the 3rd International Conference onGenetic Algorithms, pp. 70–79, Morgan Kaufmann Publishers Inc.,1989.

[177] D. E. Goldberg, K. Deb, and J. H. Clark, “Genetic algorithms, noise,and the sizing of populations,” Complex systems, vol. 6, pp. 333–362,1991.

[178] G. Harik, E. Cantú-Paz, D. E. Goldberg, and B. L. Miller, “The gam-bler’s ruin problem, genetic algorithms, and the sizing of populations,”Evolutionary Computation, vol. 7, no. 3, pp. 231–253, 1999.

[179] E. Cox, Fuzzy modeling and genetic algorithms for data mining andexploration. Academic Press, 2005.

[180] K. A. De Jong and W. M. Spears, “An analysis of the interacting rolesof population size and crossover in genetic algorithms,” in Parallelproblem solving from nature, pp. 38–47, Springer, 1991.

[181] C. M. Fonseca and P. J. Fleming, “Multiobjective genetic algorithmsmade easy: selection sharing and mating restriction,” 1995.

[182] S. W. Mahfoud, “Crowding and preselection revisited,” Urbana,vol. 51, p. 61801, 1992.

[183] D. L. Carroll, “Chemical laser modeling with genetic algorithms,” AIAAjournal, vol. 34, no. 2, pp. 338–346, 1996.

[184] Y. Jin, “A comprehensive survey of fitness approximation in evolution-ary computation,” Soft computing, vol. 9, no. 1, pp. 3–12, 2005.

[185] M. Mitchell, “Complex systems: Network thinking,” Artificial Intelli-gence, vol. 170, no. 18, pp. 1194–1212, 2006.

[186] S. H. Strogatz, “Exploring complex networks,” Nature, vol. 410,no. 6825, pp. 268–276, 2001.

[187] J. Qadir, N. Ahad, E. Mushtaq, and M. Bilal, “SDN, clouds, and bigdata: New opportunities,” 12th International Conference on Frontiersof Information (FIT), 2014.

[188] A. Verma, X. Llora, D. E. Goldberg, and R. H. Campbell, “Scalinggenetic algorithms using mapreduce,” in Intelligent Systems Designand Applications, 2009. ISDA’09. Ninth International Conference on,pp. 13–18, IEEE, 2009.

[189] X. Llora, A. Verma, R. H. Campbell, and D. E. Goldberg, “When hugeis routine: scaling genetic algorithms and estimation of distributionalgorithms via data-intensive computing,” in Parallel and DistributedComputational Intelligence, pp. 11–41, Springer, 2010.