1

Unveiling Swarm Intelligence with Network

Science—the Metaphor Explained
Marcos Oliveira, Diego Pinheiro, Mariana Macedo, Carmelo Bastos-Filho, Senior Member, IEEE, and Ronaldo
Menezes, Member, IEEE

Abstract—Self-organization is a natural phenomenon that researchers often examine the techniques from the perspective
emerges in systems with a large number of interacting compo- of their natural inspirations. For instance, in some flocking
nents. Self-organized systems show robustness, scalability, and models that mimic bird flocks, the velocities of individuals
flexibility, which are essential properties when handling real-
world problems. Swarm intelligence seeks to design nature- are usually used to understand the system behavior [11]. In
inspired algorithms with a high degree of self-organization. Yet, these systems, the lack of spatial coordination or the excess
we do not know why swarm-based algorithms work well and of coordination among individuals generally leads to poor
neither we can compare the different approaches in the literature. performance in solving problems. In the case of foraging-
The lack of a common framework capable of characterizing based models inspired by ant colonies, many studies attempt
these several swarm-based algorithms, transcending their par-
ticularities, has led to a stream of publications inspired by to understand the performance of these models by examining
different aspects of nature without much regard as to whether the pheromone that agents deposit on the environment [12].
they are similar to already existing approaches. We address this This usual approach of analyzing models via their inspiration
gap by introducing a network-based framework—the interaction has helped to improve algorithms by building new proce-
network—to examine computational swarm-based systems via dures [13]–[16].
the optics of social dynamics. We discuss the social dimension of
several swarm classes and provide a case study of the Particle These analyses, however, are confined to the specific niches
Swarm Optimization. The interaction network enables a better that have their metaphor (e.g., ants following pheromone, birds
understanding of the plethora of approaches currently available searching for food, fireflies trying to synchronize) and jargon
by looking at them from a general perspective focusing on the (e.g., pheromone, velocity, fish weight). The broad variety of
structure of the social interactions.
natural inspirations makes it challenging to find interchange-
Index Terms—self-organization; complex systems; network able concepts between swarm-intelligent techniques [17]. The
science; swarm intelligence; particle swarm optimization absence of niche-free analyses restricts the findings of a
model to its own narrowed sub-field. Such myopia leads
I. I NTRODUCTION us to miss the underlying mechanisms driving a system to
WARM INTELLIGENCE refers to the global order that
S emerges from simple social individuals interacting among
themselves [1]–[6]. In the past three decades, swarm intelli-
the imbalanced states that new techniques (or procedures)
endlessly try to avoid. In this scenario, we need agnostic
quantitative approaches to analyze swarm intelligence in a
gence has inspired many computational models, allowing us general manner and thus provide the means to understand and
to understand social phenomena and to solve real-world prob- improve algorithms in whatever niche.
lems [6]. The field of computational intelligence has witnessed The field lacks general methodologies to analyze swarms
the development of various swarm-based techniques that share because of the absence of a generic framework to examine
the principle of social interactions while having different their main similarity: the social interactions (see Fig. 1).
natural inspirations such as ants [7], fishes [8], fireflies [9], Indeed, the concept of social interaction is fundamental in
birds [10], to name a few. Though researchers have studied swarm intelligence; it refers to the exchange of information
such techniques in detail, the lack of general approaches to through diverse mechanisms [3], [5]. In this definition, social
assess these systems prevents us from uncovering what makes interactions are not only the mere exchange of information
them intelligent and understanding the differences between
techniques beyond their inspirations.
Much research has been devoted to understand and improve
these bio-inspired algorithms [5], [6], [11]. In the literature,
Marcos Oliveira is with the department of Computational Social Science,
GESIS–Leibniz Institute for the Social Sciences, Cologne, Germany.
Diego Pinheiro is with the Department of Internal Medicine, University of
California, Davis, Sacramento, USA.
Mariana Macedo and Ronaldo Menezes are with the Department of Com- Fig. 1. The social interaction at the mezzo level is still overlooked by
puter Science, University of Exeter, Devon, UK. researches who often devote considerable efforts to understand how changing
Carmelo Bastos-Filho is with the Polytechnic School of Pernambuco, the simple rules at the micro level (e.g., procedures, equations) directly affects
University of Pernambuco, Recife, Brazil. the collective behavior of the system at the macro level. In fact, these micro-
This work has been submitted to the IEEE for possible publication. level rules create the conditions to social interaction at the mezzo level which
Copyright may be transferred without notice, after which this version may in turn enables the necessary swarm dynamics to solve complex problems at
no longer be accessible. the macro level.
 2

between peers but also have the potential to change indi- way a swarm handles information defines its underlying self-
viduals [5]. The sophisticated behavior emerging from social organization mechanism. We can describe a system using three
interactions enables the system to adjust itself to solve prob- aspects of information: (i) the definition of information, (ii)
lems [5]. In swarm-intelligent techniques, individuals process how individuals use information, and (iii) how information
information and locally interact among themselves, spreading flows within the system (see Fig. 2). This approach classifies
knowledge within the swarm which results in the emergent swarm systems but fails to examine them quantitatively.
system ability. In this sense, examining the social mechanisms In fact, the literature has various approaches to classify
is fundamental to understand intelligence in these systems. swarm systems [28]–[30] and metaheuristics in general [31]–
This general perspective also helps us to assess swarms [33]. These efforts are essential to organize the field. They
with different natural inspirations. Instead of relying on the are necessary initial steps to understand current and new
complete understanding of the micro-level properties (e.g., algorithms. Still, the absence of quantitative approaches pre-
velocity, pheromone, weight), we can assess the swarm via vents us from characterizing the particularities of methods and
the structure and dynamics of the social interactions [3]. quantifying their differences.
Notably, the field of Network Science has shown that every In some cases, researchers measure the swarm diversity
complex system has an underlying network encoding the to understand the technique [30]. This diversity is often the
interactions between the components of the system and that the diversity of the candidate solutions when solving a given prob-
understanding of the structure of this network is sine qua non lem [11], [12], [34]–[37]. With such approach, however, we
for learning the behavior of the system itself [18]. Network focus on the final outcome of the swarm dynamics, neglecting
Science advocates that the complex systems comprehension the underlying mechanism leading to these dynamics. We
can be reached by observing the structure and dynamics lack a framework enabling us to examine the system from
of their underlying networks [19]–[21]. Though the idea of an intermediate perspective.
using networks as frameworks for understanding complex
phenomena dates back to Moreno’s use of sociograms in the A. The Social Interactions in Swarm Systems
1940s [22], it has popularized after two seminal papers from The dynamics of swarm-based systems depend on social
Watts and Strogatz [23], and Barabási and Albert [24] in interaction. The system lacks coordination without enough
the late 1990s. Recent works in the field have demonstrated interaction among the individuals and looses adaptability with
that even small variations in fundamental structural properties, the excess [38]. The local rules in such systems promote or
such as the degree distribution, can significantly influence the undermine the level of interaction within the swarm (Fig. 1). In
behavior of the system described by the network [25], [26]. this sense, the social interactions are halfway between the mi-
Here we propose a network-based framework—the interac- cro and macro behavior of the system. The network emerging
tion network—for the general assessment and comparison of from these complex interactions is a natural universal mezzo-
swarm intelligence techniques. In the following sections, we level perspective of swarms.
start by describing the importance of understanding swarm- Previous research has used the network paradigm to ex-
based algorithms and explaining the definition of the in- amine the emergent behavior in social animals and its un-
teraction network. We show how the interaction network
can be defined for other swarm-based metaphors using a
system categorization proposed by Mamei et al. [27]. Then,
we demonstrate a complete case study using the concept of
flocking and show the relationship between the interaction
network and the swarm behavior.

II. U NDERSTANDING S WARM S YSTEMS

In the field of Swarm Intelligence, scholars often analyze
algorithms via their performance on given problems [17].
In many cases, innovation means better results on a set
of benchmark functions. These improvements, however, tend
to arise without much explanation. Researchers often use
jargons to justify their approaches and to describe the improve-
Fig. 2. Three dimensions define the self-organized mechanism in a swarm
ments [17]. This black-box approach sidetracks us from the system: how information flows, how information is used, and the definition
underlying mechanisms in these systems. We lack a general of information. The diffusion flow occurs when individuals passively receive
interpretability of results. The case occurs because of the lack information that other individuals spread in the environment whereas serendip-
itous flow occurs when individuals need to search information left in the
of a comprehensive view of swarm-based algorithms. Though environment by other individuals. When using information in a trigger-based
some efforts have been made to understand swarm systems system, individuals act in the environment by performing specific, mostly
from a general perspective, they tend to be qualitative in one-off, actions, while in a follow-through, they are guided by what they find,
and the action can be more long-lasting. The marker-based information is
nature. explicitly defined for interaction purposes (e.g., pheromone), while individuals
Mamei et al. have proposed to look at the swarm as a implicitly share sematectonic information as the current state of the population
system processing information [27]. From this perspective, the (figure adapted from [27]).
 3

Fig. 3. Different viewpoints of swarm systems. We highlight that the interaction space is a general way of viewing these systems. (A) Models of swarm
intelligence are commonly used to solve continuous (top) and discrete (bottom) problems that can be represented in the problem space (left). Each model
incorporates metaphor-specific aspects such as velocities (first row) and pheromone (second row) in PSO and ACO, respectively, that can be represented in
the metaheuristic space (middle). However, regardless the problem and the metaphor, the actions of individuals are interdependent and depend on the social
interactions. The use of the interaction network allows us to represent these social interactions in the interaction space (right), enabling a unified assessment
of metaheuristics. (B) A general illustration of an interaction network for swarm systems where each node is an individual in the population and each link
represents the direction and the extent of the individuals’ influence. Each color depicts a distinct sub-network to which members are highly integrated and
tightly connected when compared to nodes outside the sub-network. For instance, individuals 1, 4 and 5 belong to the leftmost sub-network depicted in red
whereas individuals 10, 11 and 12 belong to the rightmost sub-network depicted in blue, while individual 9 act as a bridge between the sub-networks.

derlying mechanisms [39]–[42]. Some developments in the and so the swarm, we use I(t) to describe the influence that
computational intelligence field have also taken advantage of individuals exert on each other at time t.
networks [43]–[49]. They have been used to understand swarm The interaction network is a representation of the swarm and
systems [50], [51] and their respective collective behaviors the result of the rules that define the swarm system. However,
such as flocking [52]–[59] and foraging [49], [60]. In this the network I belongs to the interaction space I (see Fig. 3A).
regard, Oliveira et al. proposed one of the first approaches to This is an agnostic space exempt from the particularities of the
examine interactions within the swarm in the Particle Swarm swarm algorithm or problems being solved by the algorithm.
Optimization [52]. Yet, these preliminary efforts have focused Note that both the algorithm (i.e., rules) and problem modify
on specific techniques, missing the fact that social interaction the social dynamics within the system and have an impact on I.
is the common feature driving swarm intelligence. Yet, when we look at algorithms from this general framework,
In this work, we argue that the social dynamics in swarm- we have the potential to assess different algorithms that are,
based algorithms should be more analyzed and explored to at their surface, completely distinct (i.e., inspired by distinct
provide insights into the dynamic network behind the rules and natural phenomena).
inspirations, which may lead to a possible meta-classification The analyses of the network structure—at both global- and
of the systems but from the dynamics of social interactions individual-levels—enable us to assess different aspects of the
rather than the natural inspiration for the social system. In swarm, and aspects across different swarm approaches. For
Section III, we define the interaction network and use the instance, Fig. 3B depicts a conceptual interaction network for
categorization proposed by Mamei et al. to elaborate on swarm systems. At the individual-level, the network positions
the plausibility of describing and employing the interaction occupied by individuals indicate the types of interdependencies
network to assess models of swarm intelligence inspired by that were created by the swarm and the influence individuals
different mechanisms of self-organization. Though different may exert on one another. The individuals with a high degree
mechanisms, an interaction network can be built to character- centrality (e.g., individuals 1 and 12) typically exert higher
ize the system over a common space: the interaction space. influence when compared to other individuals. Similarly, in-
In the interaction space, the interaction network becomes a dividuals that connect different groups (e.g., individual 9)
general framework that allows for the unified assessment of act as bridges between subgroups of individuals and control
swarm-intelligent models with distinct inspirations. the cascade of influence between sub-networks. Thus, some
individuals in a swarm system can develop important roles
as bridges and hubs. Lastly, at a global-level, the interaction
III. T HE N ETWORK OF S OCIAL I NTERACTION
network indicates the extent of local and global exploration
We propose to examine the social interactions within a by providing the relationship between natural niches formed
swarm as a way to assess the behavior of swarm-intelligent by individuals (e.g., green and blue sub-networks).
systems. Here we develop the concept of interaction network In order to analyze a swarm using the interaction network,
to represent the interdependencies of the actions of the indi- we need to understand the rules and mechanisms that allow
viduals. For a given swarm system, the interaction network I individuals to influence the action of each other within each
consists of nodes that represent its individuals and edges Iij swarm system. For this, we employ the dimensions described
that indicate the extent to which individual i influences the in Fig. 2 to guide our understanding of algorithms and thus
action of the individual j. As social interactions are dynamic to define its network. For a given algorithm, we have to
 4

TABLE I
S OCIAL INTERACTIONS OF THE EIGHT SWARM - INTELLIGENT CATEGORIES AS ORGANIZED BY M AMEI ET AL . [27].

Name Inspiration Heuristic Communication Social interaction Example

Quorum bacteria decision-making Direct Triggered based on the passively perceived BSO [61]
current arrangement of their local environment
Embryogenesis cells [62] self-assembling Direct Triggered based on the passively perceived GSO [63]
markers of social interaction diffused in the
environment
Molding slime mold swarm coordinated Direct Passively perceived markers of social Graduate
amoebas [64] aggregation interaction they diffuse in the environment Routing [65]
Flocking birds [5] swarm coordinated Direct Continuously self-propel based on the current PSO [10]
motion arrangement of their passively perceived
environment
Brood sorting worker ants [3] sorting of items Indirect Triggered based on the actively found current SMOA [66]
arrangement of their local environment
Nest Building termites [67] coordinated construction Indirect Triggered based on the actively found markers MBO [68]
of complex structures of social interaction left in the environment
Foraging ants [3] coordinated exploration Indirect Continuously self-propel based on the actively ACO [7]
of the environment found markers of social reinforcement left in
the environment
Web Weaving spiders [69] coordinated construction Indirect Continuously self-propel based on the actively RIO [70]
of complex structures found current arrangement of their local envi-
ronment

identify how each individual uses information after it received literature often focuses on the spatial diversity [34]–[37]. From
information from other individuals. It is clear that first we need this viewpoint, researchers are interested on the outcomes
to characterize what information is in the system then describe of social interactions such as the positions (i.e., solutions)
how information exchange influences individuals. Table I or velocities of the particles in the search space. Indeed, a
describe how the interaction network can be constructed for poor trade-off between exploration and exploitation affects
the approaches proposed by Mamei et al. [27]. In Section IV, the diversity of solutions found by the swarm. For instance,
we provide a detailed case study of a well-known flocking researchers have developed different metrics that quantify
technique, the Particle Swarm Optimization, for which we use swarm diversity as the degree of dispersion of particles around
a definition of I to analyze the swarm behavior. a given spatial centroid [35]–[37], [73]. These approaches
have succeed in developing novel mechanisms (e.g., adaptive
swarm [74]) to improve the performance of the algorithm. Yet,
IV. T HE S OCIAL I NTERACTION IN THE PARTICLE S WARM
with these approaches, we fail to understand the underlying
O PTIMIZATION
social interactions driving the swarms to particular imbalanced
The Particle Swarm Optimization (PSO) is a population- states that new mechanisms try to avoid.
based optimization method that relies on the interactions Still, a few works have attempted to analyze the particles’
of individuals sharing the best positions they found during interactions in order to examine the swarm behavior. Some
the search process [10]. The method—inspired by the social of these efforts analyzed the impact of the infrastructure of
behavior in flocks of birds—consists of a population of simple the swarm communication on the swarm performance [59],
reactive agents (particles) that explore the search space by [75], [76]. Though these studies neglected the actual inter-
locally perceiving the environment and interacting among actions between particles, they showed that bounding social
themselves to improve their solutions. interactions influences the swarm behavior. Oliveira et al.
In the particle-swarm perspective, exploration and exploita- were the first ones to examine the actual interactions among
tion refer to the ability of individuals to broadly explore particles in order to assess the swarm [52]. They proposed the
the whole search space or specifically focus on a particular analysis of the swarm using a network in which the nodes
area [5]. Similar to other optimization techniques [71], the (particles) are connected if they share information in a given
PSO suffers from premature convergence [72]: the swarm iteration, and later extended the concept to capture historical
reaches a local optimum solution because of an early equi- information [53], [54], [56]. Later on, Pluhacek et al. provided
librium state. This undesired situation often results from visualizations of the interactions in the swarm [58].
an exploration–exploitation imbalance that leads to a poorly In the next sections, we briefly describe the PSO technique
explored search space. and define the interaction network I to assess the swarm
In order to understand the behavior of a swarm (as in the using the methods developed by Oliveira et al. [55]. With the
PSO) and properly balance the aforementioned two properties, definition of I, we are able to uncover relationships between
researchers tend to look at diversity within the swarm. The the swarm dynamics, the swarm performance, and the social
 5

interactions. The understanding of such relationships demon- approach [77]. The topology influences the social interaction
strates that one can understand and assess swarm approaches within the swarm and has been shown to impact the swarm
from the meta-level of social interactions. performance [75]–[77], [81].
Clerc proposed a somewhat different definition of swarm
A. Particle Swarm Optimization topology—the so-called graph of influence—which explic-
itly includes the social information and presents directed
In the standard definition of the PSO, each particle i consists
edges [82]. Still, regardless of definition, the swarm topology
of four vectors in a d-dimensional search space: its current
only refers to the structure for a potential exchange of infor-
position x~i (t), its best position found so far p~i (t), its velocity
mation and neglects effective interaction among particles.
~vi (t), and the best position found by its neighbors ~ni (t) [77].
The position of each particle represents a candidate solution
to a d-dimensional continuous optimization problem and the B. A Network for the Particle Swarm Optimization
swarm moves through the problem search space seeking for To examine a swarm system from the perspective of its
better solutions. To enable this capability, all particles change social interactions, we need to build the interaction network
their positions, at each iteration t, according to their velocities to capture the structure and dynamics of the social influence
~vi (t) which are updated based on the personal best position exerted among individuals. In the case of the PSO, a social
p~i (t) and the social best position ~ni (t). interaction occurs when a particle i updates its position based
In the first definition of the algorithm, the swarm could be in on the position of a particle j. This happens when j is the
the so-called explosion state in which the particles indefinitely best neighbor of i at a given iteration; that is, ni (t) = j.
increase their velocities [78]. Some approaches have been Here we use a simple (yet powerful) definition of interaction
proposed to prevent this state [78]–[80]. Clerc and Kennedy network I(t) in which the weight of an edge (i, j) is the
developed the constricted PSO in which the velocities are number of times the particle i was the best neighbor of the
constricted by a constriction factor that avoids the explosion particle j or vice-versa until the iteration t [53]. This network
state. This factor, χ, is defined as follows: represents the social interactions that systematically affect the
2 swarm. We can use a time window tw to control the recency
χ=   , where ϕ = c1 + c2 , (1)
of the analysis, thus the interaction network at iteration t with
p
2 − ϕ − ϕ2 − 4ϕ
window tw is defined as the following:
which adjusts the influence of the previous particle velocities
t  
during the optimization process. The constricted PSO employs X
the following update equation: Iij (t) = δi,nj (t0 ) + δj,ni (t0 ) , (4)
   t0 =t−tw +1

v~i (t + 1) = χ · v~i (t) + ~r1 c1 · p~i (t) − x~i (t) with t ≥ tw ≥ 1 and where δi,j is Kronecker delta. In this
  (2) definition, nodes (i.e., particles) are connected by an edge
+~r2 c2 · ~ni (t) − x~i (t) . with weight equals to the number of times two particles
shared information in at most tw iterations before the iteration
x~i (t + 1) = x~i (t) + v~i (t + 1), (3) t [53]. The time window tw tunes the frequency–recency
balance in the analysis. High tw makes the network dominated
where ~r1 and ~r2 are random vectors generated from a uniform by most frequent interactions; low tw only includes most
probability density distribution in the interval [0,1] for each recent interactions and when tw = 1 we have instantaneous
particle; while c1 and c2 are the cognitive and the social accel- interactions.
eration constants that weigh the contribution of the cognitive Note that the definition of an interaction network for a
and social components [5]. swarm system depends on the rules that promote social in-
The particles in the swarm only interact with a subset of the teraction in the system. Here we pinpointed that, in PSO, a
swarm. The swarm topology defines the infrastructure through social interaction between i and j occurs when the particle
which particles communicate and thus enables the particles to i updates its velocity v~i using the position of a particle
retrieve information from other particles (i.e., their neighbors). j. Yet, this definition of I is a simple one that includes
At each iteration t, each particle i seeks for its best neighbor only the occurrence of social interaction between particles.
ni (t) in its neighborhood (i.e., the one with the best solution More complex definitions may include edge direction or other
so far). Note that each particle i uses information only from aspects of the algorithm such as the social constant c2 or the
ni (t) at each iteration t. realizations of r~2 . Nevertheless, with this simple definition,
In the original PSO paper, the implicit communication we can already better understand the swarm [52]–[56]. Other
scheme is defined over a global topology in which all particles swarm systems, however, have different rules and distinct
of the swarm have the same neighborhood [10]. The infor- forms of social interactions, as we pointed out in Section II-A
mation shared among particles is the same for each particle. and showed in Table I.
In local topologies, however, particles have different subsets
of neighbors, implying different social information within
the swarm [77]. For instance, the ring topology—the most C. Examining the Social Interactions with I
popular local topology—has a structure in which particles The formation of structures in the interaction network arises
communicate with only two other particles using a labeling from the way information flow within the swarm which, in
 6

Fig. 4. The pace at which components emerge while edges are gradually removed from I is associated with the search mode of the swarm. An exploration
mode is characterized by a slow increase in the number of components due to the different information flows present in the swarm. The network, however, is
rapidly destroyed in a swarm that depends only on a small set of individuals, a behavior related to an exploitation search mode. In the PSO, (A) the weighted
interaction network of a run with the swarm using a von Neumann topology has edges removed based on their weight: below 20% of the highest possible
weight, 25% and 30%. The colors represent components with more than one node. In this process, edges with lowest weights are removed first. (B) The
impact of the edges removal on the growth of the number of components depends on the structure of the swarm topology. The different colors/markers in
the plot represent the time window tw . The normalized weight is the weight value divided by 2tw which is the highest possible weight in the network. The
rapidly increasing in the number of components of the global topology leads to a behavior related to the exploitation search mode. In the ring topology, the
number of components increases slowly, indicating the existence of sub-swarms searching more independently through the search space [53]. In all cases, the
swarm consists of 100 individuals. (C) Each topology leads to distinct interaction diversity that can be described by the number of components emerging (color
intensity) as edges are removed (y-axis) of the interaction network with different time windows (x-axis).

turn, alters the dynamics of the swarm. The existence of well- the swarm search when using different structures: the com-
connected nodes in I indicates frequent information flows in munication topology affects the diversity of the interactions
the swarm. The constant interaction among certain individuals in the swarm, generating different interaction networks.
leads to their respective nodes in the interaction network to be To consider the swarm ability to maintain different frequent
clustered. To capture these clusters, we can gradually remove information flows, we can analyze the network destruction
the edges of I according to their weight; the components while varying tw to include frequency and recency in the anal-
that emerge during this network destruction represent the ysis of the flows. Fig. 4C depicts the number of components
information flows within the swarm (see Fig. 4A). The dis- that emerge when edges are removed from I with increasing
connectivity emerges from the percolation threshold surpassed. time windows. The interaction network of a particle swarm
Note that the pace at which these components appear relates with global topology seems to be destroyed at the same pace in
to the swarm dynamics. A slow increase suggests an explo- both perspectives of frequency (i.e., high tw ) and recency (i.e.,
ration search mode in which individuals share information low tw ). The interactions of the particles within this topology
among distinct groups and thus create social interactions promote a lack of diversity in the information flows in short
with various levels of tie strength. A rapid increase suggests, and long terms.
however, an exploitation search mode in which individuals This diversity regards to the ability of the swarm to have
interact with few same sources and thus create a center of a diverse flow of information—a perspective different from
information with similar levels of tie strength. spatial diversity in which d-dimensional properties of particles
With the definition in Eq. (4), we can now examine the are compared to particular definitions of swarm center [35]–
search mode in the PSO. For instance, we analyze I of [37], [83]. Note that the lack of diversity in the information
swarms using different topology parameters—that are known flow can decrease the spatial diversity in a swarm. The
to lead the swarm to behave differently—while solving the absence of multiple information flows leads particles retrieving
same problem. As shown in Fig. 4B, with the global topology, information from few sources and drives particles to move
the particle swarm presents exploitation behavior, whereas the towards the same region of the search space; lack of interaction
ring topology leads the system to explore different information diversity pushes individuals to the same direction.
sources. Note that this analysis differs from the typical analysis To quantify interaction diversity, we measure the destruction
on the relationship between fitness and topology structure [75], pace of interaction networks with different time windows. For
[76], [81]. Here we focus on the way particles interact during a given time window tw , the area under the destruction curve
 7

Atw can be seen as a measure of diversity in the information to fast information flow which decreases the diversity [77].
flow. High values of Atw indicate fast destruction, whereas Our results revealed that the interaction diversity in the swarm
low values imply in a slower destruction. Hence, we can define depends on the problem; the same topology leads to distinct
the interaction diversity ID (previously called communication levels of diversity when optimizing different functions (see
diversity [55]) as the mean diversity over a set of time windows also Table II). Though the topology bounds the interactions
T , as the following: among particles, the swarm organize the way information flow
1 X to optimize a function.
ID(t) = 1 − Atw =t0w (t), (5) Indeed, swarm-intelligent systems have the capability to
|S||T | 0
tw ∈T self-organize during the optimization process; they can adapt
where |S| is the number of particles in the swarm. Thus, their behavior towards an optimal behavior to solve a given
swarms exhibiting high ID (i.e., low values for Atw ) have the problem. Hence, to assess the relationship between swarm
ability to have diverse information flows, while low values for search and interaction diversity, we examine the pace f∆ at
ID imply in swarms with only few information flows (i.e., which a swarm improves and the interaction diversity at each
high value for Atw ). An ideal set T would be one taking into iteration. We found that ID exhibits a non-trivial relationship
account all time windows (i.e., interactions from tw = 1 until with f∆ , as seen in Fig. 6(A) for the function F2 . The average
tw = t). This procedure, however, can be computationally f∆ increases with the average ID until reaches a maximum
expensive given the vast number of possible time windows, pace after which f∆ decreases with ID. The increase of
and a more reasonable approach is to have a sample set of diversity in the social interactions of the swarm leads to faster
time windows. swarm pace only until a certain level of diversity; then the
1) Experimental Design: To investigate the extent to which swarm starts to slow down—swarm dynamics that impacts
the interaction diversity assess the swarm at each iteration, we the overall swarm performance, as seen in Fig. 6(B). We also
systematically examined the swarm using different topologies found a non-trivial association between k-regular topologies
that lead the swarm to behave differently. We employ different and the best fitness found at the end of the optimization
connected k-regular graphs (i.e., graphs that nodes have k process. From global to 30-regular topologies, the fitness
links) as the swarm communication topology with k equal decreases from 8.06 × 103 and improves down to 6.77 × 103 ,
to 2, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 60, 70, 80, 90, then deteriorates up to 1.01 × 104 .
and 100 (special cases are: k = 2, ring topology; k = 4, 3) Discussion: Our results demonstrate the capability of
von Neumann topology; and k = 100, global topology). Here interaction diversity ID to explain the behavior of the swarm
we consider a distinct group of four benchmark functions during the optimization process in the Particle Swarm Op-
F2 , F6 , F14 , and F19 from the CEC’2010 set which possess timization technique. ID enables us to identify changes in
varied characteristics and cover major aspects of optimization the way information flow within the swarm depending on the
problems such as multi-modality and non-separability [84]. In type of problem. The leverage capability of proposed analysis
all experiments, we set the number of dimensions to 1000 and, procedure brings on the possibility to identify imbalance
when applicable, the degree of non-separability m to 50; and during the search process and further understand the flow of
define the swarm with 100 particles that are updated according information within the swarm. More than using this to select
to Eq. (2) with c1 = 2.05, c2 = 2.05 [78]. which is the best topology for a particular problem, one can
We analyze the relationship between ID and fitness im-
provement over time; thus we define fitness improvement
f∆ (t) at iteration t as the speed at which the fitness fg (t)
of the swarm changes between the two immediate iterations
f (t)−fg (t−1)
t and t − 1 as follows: f∆ (t) = g fg (t−1) , where fg (t)
is the global best fitness of the swarm at iteration t. In
the simulations, we set as the stopping criterion whether the
maximum number of iterations, tmax = 10000, is reached or
the swarm converged at iteration ts . We define that a swarm
converged at iteration ts if the global best fitness does not
improve, that is, if f∆ (t) < 10−5 , until iteration ts + δ with
δ = 500. For each considered swarm topology, we run a PSO
implementation 30 times while measuring ID and f∆ at each
iteration in each execution.
2) Results: First, we analyze the impact of the infrastruc-
ture of communication (i.e., swarm topology) on the diversity
Fig. 5. The characteristic interaction diversity of the swarm regarding
of the information flows within a swarm. We found that k- different topologies of communication and benchmark functions. Each box-
regular topologies promote higher diversity as k decreases plot represents 30 repeated simulations. The interaction diversity is typically
when solving the same problem (see Fig. 5 and Table II). With higher for less connected topologies (i.e., 2-regular and 10-regular) when
compared to highly connected topologies (i.e., 30-regular and 40-regular).
less connected topologies, swarms exhibit greater interaction Similarly, some functions appear to consistently present higher interaction
diversity than with more connected ones. Given previous stud- diversity when compared to other regardless of the underlying communication
ies, this is an expected result: short topological distances lead topology.
 8

TABLE II
I NTERACTION DIVERSITY (95% C ONFIDENCE INTERVALS ) AND THE FINAL FITNESSES FOR DIFFERENT k- REGULAR TOPOLOGIES .

k-regular topologies
Ring Neumann 5 6 7 8 9 10
F2 (0.350, 0.359) (0.325, 0.329) (0.345, 0.356) (0.328, 0.332) (0.329, 0.333) (0.283, 0.287) (0.285, 0.288) (0.255, 0.258)
1.0184 × 104 8.9214 × 103 1.0297 × 104 9.0308 × 103 9.0245 × 103 8.3778 × 103 8.4711 × 103 8.2141 × 103
F6 (0.359, 0.370) (0.317, 0.321) (0.360, 0.369) (0.316, 0.320) (0.318, 0.322) (0.268, 0.274) (0.268, 0.275) (0.238, 0.249)
1.7889 × 106 1.4784 × 106 1.5871 × 106 1.5463 × 106 1.4926 × 106 1.3893 × 106 1.4189 × 106 1.0583 × 106
F14 (0.397, 0.404) (0.363, 0.366) (0.399, 0.405) (0.360, 0.363) (0.361, 0.364) (0.320, 0.323) (0.321, 0.325) (0.295, 0.298)
1.2450 × 1010 1.0920 × 1010 1.2628 × 1010 1.0686 × 1010 1.0823 × 1010 1.0007 × 1010 9.8929 × 109 9.2693 × 109
F19 (0.381, 0.387) (0.387, 0.391) (0.373, 0.381) (0.382, 0.384) (0.383, 0.386) (0.353, 0.355) (0.352, 0.356) (0.333, 0.336)
8.6326 × 106 7.3515 × 106 8.5287 × 106 7.4901 × 106 7.5940 × 106 7.0125 × 106 6.8728 × 106 6.7804 × 106
20 30 40 60 70 80 90 Global
F2 (0.195, 0.200) (0.162, 0.169) (0.127, 0.129) (0.098, 0.100) (0.090, 0.091) (0.082, 0.083) (0.073, 0.074) (0.057, 0.058)
7.1551 × 103 6.7687 × 103 6.9666 × 103 7.2088 × 103 7.2479 × 103 7.2293 × 103 7.4163 × 103 8.0666 × 103
F6 (0.165, 0.175) (0.132, 0.140) (0.108, 0.112) (0.083, 0.087) (0.076, 0.079) (0.069, 0.071) (0.062, 0.065) (0.055, 0.058)
1.2550 × 106 1.6004 × 106 1.6119 × 106 1.7215 × 106 2.1324 × 106 2.0674 × 106 2.6710 × 106 2.6931 × 106
F14 (0.241, 0.246) (0.215, 0.222) (0.183, 0.188) (0.153, 0.159) (0.142, 0.147) (0.129, 0.134) (0.119, 0.124) (0.096, 0.099)
7.9751 × 109 7.0154 × 109 7.9144 × 109 9.0796 × 109 1.0001 × 1010 1.0116 × 1010 1.0436 × 1010 1.1396 × 1010
F19 (0.294, 0.300) (0.281, 0.288) (0.252, 0.257) (0.211, 0.214) (0.197, 0.200) (0.183, 0.186) (0.166, 0.169) (0.127, 0.128)
6.0334 × 106 5.7086 × 106 5.6747 × 106 5.3514 × 106 5.4846 × 106 5.5568 × 106 5.4663 × 106 5.3245 × 106

even deploy this to propose adaptive mechanisms to avoid swarm-based systems because it does not consider peculiar-
imbalance during search processes [81]. ities associated with the swarm metaphor. The approach is
The diversity of interaction is a general measure to assess defined over the structure of the network—the interaction
space—which is entirely based on the social interactions.
For instance, the results regarding the associations between
topology, interaction diversity, and fitness, clearly indicate that
interaction diversity can be used to understand the complex
behavior exhibited in the PSO (our case study) with different
communication topologies. In fact, this approach can help
researchers to perform parametric analyses; due to the lack of
analytical tools, previous parametric studies tend to consider
simplified version of the algorithm [78].

V. D ISCUSSION
Bees, ants, birds, and many other animals have inspired
several swarm-based algorithms, but the literature fails to
explain their differences and their complex behavior, losing
their full potential. In the field, we often describe the dif-
ferences between the techniques or their versions via the
performance achieved when solving distinct problems. This
black-box approach has enabled the area to grow over the years
and to develop excellent general-use tools. This approach,
however, lacks interpretability. How to interpret, for instance,
that including a diversity procedure improves the performance
of a swarm algorithm? Is this modification the same as using
a different algorithm? With this opaque approach, we miss the
opportunity to understand swarm intelligence.
The main barrier to understanding the swarm complex be-
havior is the discontinuity between the micro-level actions of
individuals and the macro-level behavior of the swarm. In our
Fig. 6. The interaction diversity, fitness improvement, final fitness and k-
topologies are associated in a non-trivial manner. In the results for F6 function,
work, we argue that the interaction network is the necessary
(A) although the correlation of −0.79 indicates a strongly negative linear mezzo level to explain and understand these systems. With
relationship between the average interaction diversity and the mean fitness this approach, we can examine a system via an intermediary
improvement, one can easily see that they are associated in a non-monotonic
way. (B) Similarly, the final quality of the fitness found by the swarm
structure that emerges from the social interactions within
also presents a non-monotonic behavior regarding k-regular topologies and the swarm. We can now analyze the patterns of these self-
consequently interaction diversity. organized interactions. The interaction network also grants an
 9

agnostic representation of swarm systems in the interaction [12] M. Dorigo and T. Stützle, Ant Colony Optimization. The MIT Press,
space which provides us with a more general perspective of 2004.
[13] J. Sun, S.-W. Xiong, and F.-M. Guo, “A new pheromone updating strat-
swarm-based algorithms. egy in ant colony optimization,” in Proceedings of 2004 International
To verify the plausibility of this network-based approach, Conference on Machine Learning and Cybernetics (IEEE), 2004, pp.
we considered the self-organization mechanism of flocking 620–625.
[14] W. Dong and M. Zhou, “A supervised learning and control method to
as a case study and investigated its most popular optimiza- improve particle swarm optimization algorithms,” IEEE Transactions on
tion technique, namely, the Particle Swarm Optimization. We Systems, Man, and Cybernetics: Systems, vol. 47, no. 7, pp. 1135–1148,
also discussed the social interaction in other self-organization 2017.
[15] A. Babalık, H. İşcan, İ. Babaoğlu, and M. Gndz, “An improvement
mechanisms to guide definitions of their interaction network. in fruit fly optimization algorithm by using sign parameters,” Soft
In the analysis of the PSO, we found that the interaction Computing, vol. 22, no. 22, pp. 7587–7603, 2017.
network helps us to disentangle complex features of swarm [16] Y. Cao, Y. Lu, X. Pan, and N. Sun, “An improved global best guided
artificial bee colony algorithm for continuous optimization problems,”
systems. We analyzed its interplay with the quality and im- Cluster Computing, pp. 1–9, 2018.
provement of fitness, and we found that some characteristics [17] K. Sörensen, “Metaheuristics-the metaphor exposed,” International
of the interaction network can be used to explain parametric Transactions in Operational Research, vol. 22, no. 1, pp. 3–18, 2013.
[18] A.-L. Barabási, Linked: The new science of networks. Perseus Books
settings in the algorithm. Specifically, we studied the diversity Group, 2003.
in the network (i.e., the Interaction Diversity). Our results re- [19] S. H. Strogatz, “Exploring complex networks,” Nature, vol. 410, no.
vealed that different communication topology leads the swarm 6825, pp. 268–276, 2001.
[20] G. Palla, I. Derényi, I. Farkas, and T. Vicsek, “Uncovering the overlap-
to distinct search mode that also depends on the problem ping community structure of complex networks in nature and society,”
landscape. Nature, vol. 435, no. 7043, pp. 814–818, 2005.
The network-based perspective of swarms unfolds a path- [21] A.-L. Barabási, “The network takeover,” Nature Physics, vol. 8, no. 1,
pp. 14–16, 2012.
way to researchers to study these systems comprehensively. [22] J. L. Moreno, “Sociogram and sociomatrix,” Sociometry, vol. 9, pp.
This perspective creates opportunities on two fronts. First, it 348–349, 1946.
brings the required general viewpoint to build an objective [23] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’
networks,” Nature, vol. 393, no. 6684, pp. 440–442, 1998.
classification of swarm-based algorithms. This classification [24] A.-L. Barabási and R. Albert, “Emergence of scaling in random net-
guides the algorithm selection for problem-solving and the works,” Science, vol. 286, no. 5439, pp. 509–512, 1999.
development of novel or hybrid methods. Second, the network [25] M. Santolini and A.-L. Barabási, “Predicting perturbation patterns from
the topology of biological networks,” Proceedings of the National
empowers scholars to examine swarms from an intermediate Academy of Sciences, vol. 115, no. 27, pp. E6375–E6383, 2018.
level that is crucial to understand the complex behavior of [26] E. Muller and R. Peres, “The effect of social networks structure on inno-
these systems. At this mezzo level, we expose the effects of vation performance: A review and directions for research,” International
Journal of Research in Marketing, 2018.
the swarm rules which are hidden in the swarm behavior. [27] M. Mamei, R. Menezes, R. Tolksdorf, and F. Zambonelli, “Case studies
In this study, though we limited our numerical analyses to for self-organization in computer science,” Journal of Systems Architec-
the PSO, we proposed a general approach that makes possible ture, vol. 52, no. 8-9, pp. 443–460, 2006.
[28] R. Parpinelli and H. Lopes, “New inspirations in swarm intelligence:
to perform parametric analyses, quantify differences between a survey,” International Journal of Bio-Inspired Computation, vol. 3,
methods, balance techniques with hybrid or adaptive versions, no. 1, p. 1, 2011.
and build mezzo-level mechanisms. These are directions for [29] H. Duan and Q. Luo, “New progresses in swarm intelligence-based
computation,” International Journal of Bio-Inspired Computation, vol. 7,
future research. no. 1, p. 26, 2015.
[30] X. Chu, T. Wu, J. D. Weir, Y. Shi, B. Niu, and L. Li, “Learn-
ing–interaction–diversification framework for swarm intelligence opti-
R EFERENCES mizers: a unified perspective,” Neural Computing and Applications,
[1] S. Kaufmann, The origins of order. Oxford University Press, 1993, 2018.
vol. 209. [31] M. Gendreau and J.-Y. Potvin, “Metaheuristics in combinatorial opti-
[2] P. Bak, How Nature Works, 1st ed. Springer New York, 1996. mization,” Annals of Operations Research, vol. 140, no. 1, pp. 189–213,
[3] E. Bonabeau, M. Dorigo, and G. Theraulaz, Swarm intelligence: from 2005.
natural to artificial systems. Oxford university press, 1999, no. 1. [32] J. L. Fernandez-Marquez, G. D. M. Serugendo, S. Montagna, M. Viroli,
[4] T. Vicsek, “A question of scale,” Nature, vol. 411, no. 6836, pp. 421– and J. L. Arcos, “Description and composition of bio-inspired design
421, 2001. patterns: a complete overview,” Natural Computing, vol. 12, no. 1, pp.
[5] J. Kennedy and R. C. Eberhart, Swarm Intelligence. Morgan Kaufmann 43–67, 2012.
Publishers Inc., 2001. [33] S. Fong, X. Wang, Q. Xu, R. Wong, J. Fiaidhi, and S. Mohammed,
[6] A. P. Engelbrecht, Fundamentals of Computational Swarm Intelligence. “Recent advances in metaheuristic algorithms: Does the makara dragon
John Wiley & Sons, 2006. exist?” The Journal of Supercomputing, vol. 72, no. 10, pp. 3764–3786,
[7] M. Dorigo and G. D. Caro, “Ant colony optimization: a new meta- 2015.
heuristic,” in Proceedings of the 1999 Congress on Evolutionary [34] T. Krink, J. Vesterstrom, and J. Riget, “Particle swarm optimisation
Computation-CEC99, 1999, pp. 1470–1477. with spatial particle extension,” in Proceedings of the 2002 Congress
[8] C. J. Bastos-Filho, F. B. de Lima Neto, A. J. Lins, A. I. Nascimento, and on Evolutionary Computation. CEC, 2002, pp. 1474–1479.
M. P. Lima, “A novel search algorithm based on fish school behavior,” in [35] Y. Shi and R. C. Eberhart, “Population diversity of particle swarms,”
2008 IEEE International Conference on Systems, Man and Cybernetics. in 2008 IEEE Congress on Evolutionary Computation (IEEE World
IEEE, 2008, pp. 2646–2651. Congress on Computational Intelligence), 2008, pp. 1063–1067.
[9] X.-S. Yang, “Firefly algorithm, lévy flights and global optimization,” [36] O. Olorunda and A. P. Engelbrecht, “Measuring exploration/exploitation
in Research and Development in Intelligent Systems XXVI, 2009, pp. in particle swarms using swarm diversity,” in 2008 IEEE Congress on
209–218. Evolutionary Computation (IEEE World Congress on Computational
[10] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proceed- Intelligence), 2008, pp. 1128–1134.
ings of ICNN 95 - International Conference on Neural Networks, 1995, [37] Y. Shi and R. Eberhart, “Monitoring of particle swarm optimization,”
pp. 1942–1948. Frontiers of Computer Science in China, vol. 3, no. 1, pp. 31–37, 2009.
[11] A. P. Engelbrecht, Computational Intelligence: An Introduction, 2nd ed. [38] H. Chaté and M. A. Muñoz, “Viewpoint: insect swarms go critical,”
Wiley Publishing, 2007. Physics, vol. 7, p. 120, 2014.
 10

[39] J. H. Fewell, “Social insect networks,” Science, vol. 301, no. 5641, pp. [61] E. Ali and S. Abd-Elazim, “Coordinated design of PSSs and TCSC
1867–1870, 2003. via bacterial swarm optimization algorithm in a multimachine power
[40] D. Lusseau, “The emergent properties of a dolphin social network,” system,” International Journal of Electrical Power & Energy Systems,
Proceedings of the Royal Society B: Biological Sciences, vol. 270, pp. vol. 36, no. 1, pp. 84–92, 2012.
S186–S188, 2003. [62] P. Lawrence, “The making of a fly: the genetics of animal design,”
[41] A. Strandburg-Peshkin, C. R. Twomey, N. W. Bode, A. B. Kao, Y. Katz, Trends in Genetics, no. 6, pp. 220–221, 1992.
C. C. Ioannou, S. B. Rosenthal, C. J. Torney, H. S. Wu, S. A. Levin, [63] K. Krishnanand and D. Ghose, “Detection of multiple source locations
and I. D. Couzin, “Visual sensory networks and effective information using a glowworm metaphor with applications to collective robotics,”
transfer in animal groups,” Current Biology, vol. 23, no. 17, pp. R709– in Proceedings 2005 IEEE Swarm Intelligence Symposium, 2005. SIS.
R711, 2013. IEEE, 2005, pp. 84–91.
[42] S. B. Rosenthal, C. R. Twomey, A. T. Hartnett, H. S. Wu, and I. D. [64] T. Schmickl and K. Crailsheim, “A navigation algorithm for swarm
Couzin, “Revealing the hidden networks of interaction in mobile animal robotics inspired by slime mold aggregation,” in Swarm Robotics, 2007,
groups allows prediction of complex behavioral contagion,” Proceedings pp. 1–13.
of the National Academy of Sciences, vol. 112, no. 15, pp. 4690–4695, [65] R. D. Poor, “Embedded networks: Pervasive, low-power, wireless con-
2015. nectivity,” Ph.D. dissertation, Massachusetts Institute of Technology,
[43] M. Giacobini, M. Preuss, and M. Tomassini, “Effects of Scale-Free 2001.
and Small-World Topologies on Binary Coded Self-adaptive CEA,” in [66] D. R. Monismith and B. E. Mayfield, “Slime mold as a model for
Lecture Notes in Computer Science (including subseries Lecture Notes numerical optimization,” in 2008 IEEE Swarm Intelligence Symposium.
in Artificial Intelligence and Lecture Notes in Bioinformatics), 2006, vol. IEEE, 2008, pp. 1–8.
3906 LNCS, pp. 86–98. [67] Dynamic Concepts. (2014) Cooperative multi-robot control architecture.
[44] J. L. Payne and M. J. Eppstein, “Evolutionary dynamics on scale-free http://www.dynamic-concepts.com/.
interaction networks,” IEEE Transactions on Evolutionary Computation, [68] H. A. Abbass, “A monogenous mbo approach to satisfiability,” in Pro-
vol. 13, no. 4, pp. 895–912, 2009. ceeding of the International Conference on Computational Intelligence
[45] A. Godoy and F. J. von Zuben, “A complex neighborhood based for Modelling, 2001.
particle swarm optimization,” in 2009 IEEE Congress on Evolutionary [69] C. Bourjot, V. Chevrier, and V. Thomas, “A new swarm mechanism
Computation, 2009, pp. 720–727. based on social spiders colonies: From web weaving to region detection,”
[46] B. Dorronsoro and P. Bouvry, “Study of different small-world topology Web Intelligence and Agent Systems: An international journal, vol. 1,
generation mechanisms for genetic algorithms,” in 2012 IEEE Congress no. 1, pp. 47–64, 2003.
on Evolutionary Computation, 2012, pp. 10–15. [70] T. C. Havens, C. J. Spain, N. G. Salmon, and J. M. Keller, “Roach
[47] M. Oliveira, C. J. A. Bastos Filho, and R. Menezes, “Using network infestation optimization,” in 2008 IEEE Swarm Intelligence Symposium,
science to define a dynamic communication topology for particle swarm 2008, pp. 1–7.
optimizers,” in Complex Networks, 2013, pp. 39–47. [71] H. M. Pandey, A. Chaudhary, and D. Mehrotra, “A comparative review
[48] C. Liu, W.-B. Du, and W.-X. Wang, “Particle swarm optimization with of approaches to prevent premature convergence in GA,” Applied Soft
scale-free interactions,” PLoS ONE, vol. 9, no. 5, p. e97822, 2014. Computing, vol. 24, pp. 1047–1077, 2014.
[49] M. Metlicka and D. Davendra, “Ensemble centralities based adaptive [72] P. J. Angeline, “Evolutionary optimization versus particle swarm op-
artificial bee algorithm,” in 2015 IEEE Congress on Evolutionary timization: Philosophy and performance differences,” in International
Computation (CEC), 2015, pp. 3370–3376. Conference on Evolutionary Programming. Springer, 1998, pp. 601–
[50] J. M. Whitacre, R. A. Sarker, and Q. T. Pham, “The self-organization of 610.
interaction networks for nature-inspired optimization,” IEEE Transac- [73] S. Cheng and Y. Shi, “Diversity control in particle swarm optimization,”
tions on Evolutionary Computation, vol. 12, no. 2, pp. 220–230, 2008. in 2011 IEEE Symposium on Swarm Intelligence. IEEE, 2011, pp. 1–9.
[51] C. Huepe, G. Zschaler, A.-L. Do, and T. Gross, “Adaptive-network [74] Z.-H. Zhan, J. Zhang, Y. Li, and H.-H. Chung, “Adaptive particle swarm
models of swarm dynamics,” New Journal of Physics, vol. 13, no. 7, optimization,” IEEE Transactions on Systems, Man, and Cybernetics,
p. 073022, 2011. Part B (Cybernetics), vol. 39, no. 6, pp. 1362–1381, 2009.
[52] M. Oliveira, C. J. A. Bastos-Filho, and R. Menezes, “Assessing particle [75] R. Mendes, J. Kennedy, and J. Neves, “The fully informed particle
swarm optimizers using network science metrics,” in Complex Networks swarm: Simpler, maybe better,” IEEE Transactions on Evolutionary
IV, ser. Studies in Computational Intelligence, 2013, vol. 476, pp. 173– Computation, vol. 8, no. 3, pp. 204–210, 2004.
184. [76] R. Mendes, “Population topologies and their influence in particle swarm
[53] ——, “Towards a network-based approach to analyze particle swarm performance,” Ph.D. dissertation, University of Minho, 2004.
optimizers,” in 2014 IEEE Symposium on Swarm Intelligence, 2014, pp. [77] D. Bratton and J. Kennedy, “Defining a standard for particle swarm
166–173. optimization,” in 2007 IEEE Swarm Intelligence Symposium, 2007, pp.
[54] ——, “Using network science to assess particle swarm optimizers,” 120–127.
Social Network Analysis and Mining, vol. 5, no. 1, pp. 1–13, 2015. [78] M. Clerc and J. Kennedy, “The particle swarm - explosion, stability, and
[55] M. Oliveira, D. Pinheiro, B. Andrade, C. Bastos-Filho, and R. Menezes, convergence in a multidimensional complex space,” IEEE Transactions
Communication Diversity in Particle Swarm Optimizers. Springer on Evolutionary Computation, vol. 6, no. 1, pp. 58–73, 2002.
International Publishing, 2016, pp. 77–88. [79] Y. Shi and R. Eberhart, “A modified particle swarm optimizer,” in 1998
[56] M. Oliveira, D. Pinheiro, M. Macedo, C. Bastos-Filho, and R. Menezes, IEEE International Conference on Evolutionary Computation Proceed-
“Better exploration-exploitation pace, better swarm: Examining the ings. IEEE World Congress on Computational Intelligence, 1998, pp.
social interactions,” in 2017 IEEE Latin American Conference on 69–73.
Computational Intelligence (LA-CCI), 2017, pp. 1–6. [80] R. Eberhart, P. Simpson, and R. Dobbins, Computational intelligence
[57] J. Janostik, M. Pluhacek, R. Senkerik, and I. Zelinka, “Particle swarm PC tools. Academic Press Professional, Inc., 1996.
optimizer with diversity measure based on swarm representation in [81] A. P. Engelbrecht, “Particle swarm optimization: Global best or local
complex network,” in Proceedings of the Second International Afro- best?” in 2013 BRICS Congress on Computational Intelligence and 11th
European Conference for Industrial Advancement AECIA 2015, 2016, Brazilian Congress on Computational Intelligence, 2013, pp. 124–135.
pp. 561–569. [82] M. Clerc, Particle swarm optimization. John Wiley & Sons, 2010,
[58] M. Pluhacek, J. Janostik, R. Senkerik, I. Zelinka, and D. Davendra, “Pso vol. 93.
as complex network—capturing the inner dynamics—initial study,” in [83] J. Janostik, M. Pluhacek, R. Senkerik, I. Zelinka, and F. Spacek, “Cap-
Proceedings of the Second International Afro-European Conference for turing inner dynamics of firefly algorithm in complex network—initial
Industrial Advancement AECIA 2015, 2016, pp. 551–559. study,” in Proceedings of the Second International Afro-European Con-
[59] W.-B. Du, W. Ying, and G. Yan, “The impact of population structure on ference for Industrial Advancement AECIA 2015, 2016, pp. 571–577.
particle swarm optimization: A network science perspective,” in Lecture [84] K. Tang, X. Li, P. N. Suganthan, Z. Yang, and T. Weise, “Benchmark
Notes in Computer Science, 2016, pp. 341–349. Functions for the CEC’2010 Special Session and Competition on Large-
[60] P. Kromer, P. Gajdo, and I. Zelinka, “Towards a network interpretation of Scale Global Optimization,” University of Science and Technology of
agent interaction in ant colony optimization,” in 2015 IEEE Symposium China (USTC): China, Tech. Rep.
Series on Computational Intelligence, 2015, pp. 1126–1132.
 11

Marcos Oliveira is a postdoctoral researcher at the Computational Social Carmelo Bastos-Filho is an Associate Professor at the Polytechnic School of
Science department in GESIS–Leibniz Institute for the Social Sciences. the University of Pernambuco and Scientist-in-Chief of the Technological Park
He received his PhD in Computer Science from the Florida Institute of of Pernambuco for Electronics. His interests are related to optical networks,
Technology. His research interests lie in complex systems, human dynamics, swarm intelligence, evolutionary computation, multi-objective optimization
and self-organizing mechanisms. and biomedical applications.

Diego Pinheiro is a postdoctoral scholar at the Department of Internal

Medicine in the University of California, Davis. He received his Ph.D. in
Computer Science from the Florida Institute of Technology. His research
interests include complex systems, data science, and computational health. Ronaldo Menezes is a professor of Data and Network Science, the director
of the BioComplex Laboratory, and the Head of the Computer Science
Department at the University of Exeter, UK. In addition to Data and Network
Science, his research interest also include self-organization, human dynamics,
and predictability of complex behavior.
Mariana Macedo is a Ph.D. student at the Computer Science Department
in the University of Exeter. She received a B.Sc. and M.Sc degree in
Computer Engineering from the University of Pernambuco in 2016 and 2017,
respectively. Her research interests are related to computational intelligence,
multi and many-objective, human mobility, complex networks, and machine
learning.