Received 26 April 2024, accepted 29 May 2024, date of publication 3 June 2024, date of current version 14 June 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3409051

Multi-Agent Systems: A Survey About Its

Components, Framework and Workflow
DIEGO MALDONADO 1 , EDISON CRUZ 2 , JACKELINE ABAD TORRES 1 ,
PATRICIO J. CRUZ 1 , (Member, IEEE), AND SILVANA DEL PILAR GAMBOA BENITEZ 1
1 Departamento de Automatización y Control Industrial, Facultad de Ingeniería Eléctrica y Electrónica, Escuela Politécnica Nacional, Quito 170525, Ecuador
2 Department of Electrical, Computer and Biomedical Engineering, Southern Illinois University, Carbondale, IL 62901, USA

Corresponding author: Diego Maldonado ([email protected])

This work was supported by the Escuela Politécnica Nacional under Grant PIS-21-01.

ABSTRACT With the rapid technological advancements and the ever-evolving complex systems, the
identification and integration of the components and resources for the functioning of multi-agent systems
(MAS) are crucial tasks. However, difficulties arise due to the complexity of not having reference frameworks
that normalize their implementation. Therefore, in this survey, we propose the FC-MAS (Framework-
Components in Multi-Agent System) model as a conceptual framework designed to simplify comprehension
and standardization in incorporating the required functions and components for the deployment and operation
of MAS in engineering applications. This model comprises five abstract layers, each of which serves
a specific purpose and encompasses the details and resources required to operate MAS. Furthermore,
we propose a structured workflow for centralized and distributed MAS schemes with a set of related activities
that integrate the fundamental steps and stages for the successful implementation of MAS. Finally, this work
discusses potential directions for future research, including a deeper exploration of essential components, the
establishment of terminology standards across various domains, and the refinement of the proposed model
to enhance its applicability and relevance across a broader spectrum of contexts.

INDEX TERMS Multi-agent system, complex system, components, framework, workflow.

I. INTRODUCTION Systems such as those described above are identified as

An overarching trend observed across various industries, Multi-agent Systems (MAS), which are generally defined
including robotics and power systems, is the increasing as a set of elements, called agents, interacting with each
integration of their components. This integration has been other to achieve a common purpose [5], [6], [7], [8], [9].
facilitated by the advent of transformative technologies like The design, development, and coordination of these systems
the Internet of Things (IoT), artificial intelligence (AI), is still an open challenge due to the complexity that they
machine learning (ML), Industry 4.0, and networked systems. entail. Therefore, several works have focused on studying
[1], [2], [3]. Thus, their components become a fundamental them in more depth. For instance, MAS is a type of complex
part of an interactive environment capable of putting together system [10], so the theory, terms, and concepts from this
different processes and coordinating its activities to meet active research topic can be applied to them.
more complex objectives and global tasks. In addition, these Applications involving MAS are very diverse and in con-
interconnected systems are able to use the available resources stant growth. There are several works about their application
(either local or global) without the constant interaction in areas such as: emergency response operations involving
with a human operator or a great amount of information drones and mobile robots [11], smart grids or micro-
exchange [4]. grids [12], [13], [14], industrial plant control with actuator
saturation [15], multi-vehicle coordination [16], data traffic in
The associate editor coordinating the review of this manuscript and computer science, transportation and communication systems
approving it for publication was Chao-Yang Chen . for smart cities [17], [18], among others. Due to the different
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
80950 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

fields of MAS’s applications, the names of the involved for comprehending and standardizing the amalgamation
elements and processes differ according to the taxonomy of diverse resources and components within multi-agent
terms typically used in their field of study. systems. This model abstracts the functions and components
Another important point to highlight from the recent use in a hierarchical five-layer structure: the physical network,
of MAS is the possibility of implementing novel techniques synchronization, network controller, assessment, and fault
that seek to optimize the operation of specific procedures of tolerance. Subsequently, we provide illustrative instances
the system. For this reason, many studies focus on collecting of how this proposed model can effectively delineate
relevant information about these techniques and applying various engineering applications, irrespective of their specific
them in specific processes within the operation of MAS. For objectives. Second, we propose a structured workflow that
example, [19] presents general knowledge and advances of a offers a systematic insight into the required steps and stages
MAS process known as consensus from a control perspective. that must be followed from the initialization of the operation
Related works about this process, such as [20], present of a multi-agent system.
some consensus algorithm alternatives, considering delays The general structure of this survey is as follows. Section II
and the agent’s dynamics, e.g. single and double integrator presents an overview of the MAS information, the taxonomy
dynamics. Another critical activity carried out in MAS is with the most commonly used terms, and some relevant
decision-making. For instance, [7] states that each agent is applications in which they stand out. Then, in Section III,
an essential part of the MAS whose performance defines the conceptual framework model proposed is analyzed and
the benefit of the whole system and optimal sequences of its elements are described. Continuously, section IV presents
actions. Other processes in MAS are task allocation and how FC-MAS can be used to describe the implementation
coalition formation. In this regard, there are algorithms based in complex systems, using Multi-robot system (MRS) and
on demands, resources and profit objectives such as the Power Systems (PS) as examples. Section V details the MAS
one proposed in [21], while [22] presents a combination deployment process with the stages and steps to achieve a
of a greedy algorithm with a distributed many-objective main objective. Finally, in Section VI the conclusions and
evolutionary algorithm to find the best solution. An additional future directions are discussed.
feature that becomes crucial in MAS is the development of
fault-tolerant behavior during its operation. Thus, different
related works have focused on the design and the operation II. MULTI-AGENT SYSTEMS (MAS)
principles of robust and resilient MAS [23], [24]. A. GENERAL DESCRIPTION
Having all this information, it is convenient to have a In general, a multi-agent system (MAS) is a set of so-called
general structure of a MAS, which could serve as a basis agents interacting among them within a certain environment,
for its analysis and implementation. This type of study helps exchanging information through some communication chan-
to: identify its elements, know its capabilities, and then nel [5], [25] and acting cooperatively among them to solve a
apply the appropriate methods and techniques for improving common problem with great skill [26]. Further, this group of
its performance. This is the main motivation behind this autonomous entities (agents) perceives information about the
work, which seeks to provide a general understanding of environment through sensors and carries out their activities
MAS based on its most important components, characteristics by using actuators [6], [27].
and analyze case studies developed from a more realistic We begin defining the most basic element in a MAS: an
perspective. Further, we seek to provide a workflow that agent. Many authors have oriented the definition of this term
describes the deployment and operation of a MAS. according to the objectives of the application on which their
The goal of this study is to propose a schematic that work focuses, even so, many of these definitions resemble
describes any general MAS, which leads to formulating the each other. Starting with concepts that address MAS in a
following research questions(RQs): general way, [28] defines an agent as: ‘‘an entity which is
placed in an environment and senses different parameters that
• RQ1: What taxonomy relates to the elements of a MAS?
are used to make a decision based on the goal of the entity.
• RQ2: What are the most common components within a
The entity performs the necessary action on the environment
MAS?
based on this decision’’. In [25], the general definition of an
• RQ3: Is there a framework that categorizes the compo-
agent is: ‘‘a computer system that is capable of independent
nents and elements of a MAS?
actions on behalf of its user or owner. The agent can figure
• RQ4: Is there a specific workflow for implementing a
out for itself what it needs to do in order to satisfy its design
MAS and putting it into operation?
objectives’’. In works focused on robotics, such as in [7],
• RQ5: How can a MAS begin operating effectively
the authors propose a general concept for an agent, which
following a workflow from its design?
is ‘‘an autonomous entity capable of performing actions on
To address these questions, in conjunction with an its environment and perceiving its environment, aiming to
exhaustive examination of the pertinent literature, allows accomplish a goal’’.
us to present our main contributions. First, we propose a In the field of AI, [29] proposes the following: ‘‘agents
conceptual model, FC-MAS, which serves as a reference are autonomous, computational entities that can be viewed

VOLUME 12, 2024 80951

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

as perceiving their environment through sensors and acting

upon their environment through effectors’’. Therefore, they
refer to agents as either software or hardware with a certain
level of intelligence that depends on its own experience
for achieving goals. Another field to take into account
for defining agents is Autonomous agents. In this area,
[30] defines an agent as: ‘‘a computational system that is
located in some environment and is capable of performing
autonomous actions in order to achieve objectives. These
agents take input from the environment towards the sensor
and produce output actions’’.
Regarding the goals in MAS correspond to the objectives
pursued by the agents within the system and can be either FIGURE 1. General architecture of an agent interacting in an
local or global. Global goals (or objectives) refer to the environment. *The Agent’s Data Exchange (ADE) depends on the physical
links, topology, and architecture used in the network.
overall goal or main purpose to be achieved or accomplished,
which is usually the fundamental reason for developing or
implementing the system. Conversely, each agent pursues
local goals according to their assigned tasks, and decisions of the most used terminology in this type of work. The
are made based on their objectives and locally available selected terms have been classified as a function of their
information. Even though each agent may only focus in main characteristics. Then sub-terms are associated with four
its individual goal, it impacts the system performance and major areas in engineering in which there are widely used:
supports to reach the global goal. For example, in a traffic Control Systems, Power Systems, Telecommunication, and
management system, a vehicle’s individual goal is to reach its Robotics; here diverse MAS with their schemes, control
destination as quickly as possible, while the system’s global architecture, and communication technology are applied.
goal is to minimize traffic congestion and reduce overall Through this relationship, it is possible to notice how MAS
travel times. has been analyzed in various fields of research. Therefore,
Other relevant element constantly mentioned in the its application can be potentially generalized in other fields
definitions of MAS is the environment in which the agents where they have not been studied yet.
interact with each other by executing their programming in
pursuit of a common goal. Due to the interaction between 1) AGENT
the agents and the environment, the former may be able to The previously reviewed definitions of an agent share similar
alter the environment. In order to modify the environment, ideas about what it should be and do. An autonomous
the agent must be able to completely or partially observe agent is capable of perceiving its environment, processing
it by gathering information via sensors, data prediction or information, making decisions based on its objectives, taking
computing systems. Moreover, this environment control, actions without direct external control, and can have reactive,
which is either physical or logical action, is clearly limited proactive and social behaviors [24], [25], [28], [33]. A reac-
by the application, e.g, changing the environment implies the tive agent reacts immediately as a direct response to stimuli
reconfiguration of a transmission network in the electricity or changes in the environment without carrying out a deep
sector [31], or changing transportation routes in logistics. analysis of the situation. A proactive agent takes the initiative
On the other hand, the interactions among the agents refer in making decisions to achieve its objectives. It not only
to the way in which they share information with each other responds to environmental stimuli but also predicts future
to achieve cooperative work within the environment. This situations and can plan actions without direct intervention
can occur through communication channels or physical links from other agents. A social agent interacts within the system
that couple the agents, known as the Agent’s Data Exchange to exchange data with its neighbors, seeking to cooperate,
(ADE). Figure 1 presents a general architecture of an agent negotiate, or even compete. This type of agent recognizes
that fits to the previous definitions and provides a clear idea of that interaction with other agents is sometimes the best way
how this element works and interacts with the environment. to acquire knowledge. For instance, in a traffic management
system, an autonomous agent can take the form of a self-
driving vehicle equipped with sensors and advanced control
B. TAXONOMY OF SOME COMMON TERMS IN MAS algorithms. This allows the vehicle to interpret traffic signals
Generally, MAS uses a large number of different terms and make decisions such as changing lanes or stopping at
corresponding to methods, features, or elements involved intersections without requiring direct communication with
in certain areas of the system’s operation. [32] provides an other vehicles. A reactive agent may be a traffic light
overview of agent terms and then delves into the key concepts control that adjusts signals based on predetermined timing
and theories of MAS technology, touching upon related patterns. It operates without making predictions about long-
concepts like expert systems. Table 1 shows a compilation term traffic flow. In contrast, a proactive agent is a vehicle

80952 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

that predicts potential congestion based on historical data, TABLE 1. Compilation of terminology most used in MAS and how is it
used in different research areas.
weather forecasts, and special events. It can also suggest
alternative routes to other vehicles. Lastly, social agents are
vehicles that collectively exchange data from their position,
speed, and intentions on the road. They can also announce
traffic events, allowing other vehicles to plan their route. This
optimizes traffic flow, mitigates the risk of collisions, and
enhances safety in general.

2) ORGANIZATION
Organization in a MAS refers to how agents interact
based on their roles, behavioral expectations, and authority
relationships. Some related terms include centralized [34],
where agents send information to a control center that has the
monopoly to make decisions, and decentralized [35], where
many distributed controllers work locally. Hierarchy refers
to decision-making authority [9], which can be distributed
to raise fault tolerance. Homogeneous systems have agents
with identical characteristics, while heterogeneous systems
different hardware or software can interact and exchange
have agents with different characteristics that complement
information correctly, a common agreement is required in
each other. Holonic systems are composed of holons, which
the choice of communication language. Additionally, the
are a set of agents grouped according to certain features
data transmission frequency must be synchronized to avoid
and communicate with other agents in the same or different
information overflow and ensure that the receiver is available
holons of the same level.
to process the information. However, factors such as the
transmission medium, the amount of information, and the
3) ALGORITHMS IN MAS environment can cause delays in communication.
Control algorithms or schemes are essential for managing
the elements and ensuring the correct operation of a MAS. 5) SIMULATION ENVIRONMENTS
These algorithms can serve as the primary control for system Simulation environments are essential for developing and
operation or as a secondary algorithm for performing addi- testing MAS. They offer tools for modeling, building, and
tional tasks that enhance overall performance. Some common simulating agent behavior and interactions in a controlled
algorithms used in MAS include consensus algorithms [36], virtual environment. These environments are valuable for
which seek to reach a mutual agreement with all system comprehending, analyzing, and validating MAS designs
elements through a continuous exchange of information. before real-world implementation. For instance, Java Agent
Artificial Intelligence (AI) [37], [38] is another critical area DEvelopment (JADE) [43], a platform for building multi-
with broad applications, involving the ability of a system to agent systems in Java provides a set of tools and libraries that
interpret data correctly, learn from it, and use that knowledge enable the development of distributed and intelligent software
to achieve specific goals. Swarm intelligence algorithms based on the concept of agents. MATLAB is also a popular
[39], inspired by group behavior in the animal kingdom, platform for developing engineering applications, which
are often used in task assignments for MAS. Game theory offers a variety of toolboxes and functions that are helpful
[40], specifically in control theory and distributed control, for designing, implementing, and analyzing the behavior of
is oriented towards designing games that evoke desirable individual agents and their interactions in a MAS. Robot
emergent behavior of the collective subsystems. Finally, Operating System (ROS), an open-source framework widely
Machine Learning algorithms [41], refers to the capability of used to control robotic systems with realistic environmental
computer programs to identify intricate patterns by utilizing characteristics. GIS and Agent-based Modelling Architecture
algorithms that analyze data. By constructing mathematical (GAMA) [44], a platform for building spatially explicit agent-
models of problems, these algorithms enable the prediction based simulations with the capacity to manage geographical
of future behaviors. data and carry out simulations with hundreds of thousands
of agents. These simulation environments allow researchers
4) COMMUNICATION to test and refine MAS models, ultimately leading to more
Effective communication is crucial for a MAS to function effective and efficient systems.
properly, as it involves information exchange and interactions
among agents. Three key factors that impact communication C. APPLICATIONS
in MAS are the agent communication language, transmission As mentioned in Section I, many applications have included
frequency, and delays [9], [42]. To ensure that agents with the MAS concept to model complex systems and facilitate a

VOLUME 12, 2024 80953

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

solution by dividing the global problem or objective into a It applies additional extensions for processes related to the
series of individual tasks that can be performed more easily. selection of trajectories and also using specific algorithms
Below, we mention some applications that use MAS models for internal process like task allocation. Another example
in the resolution and optimization of real problems. Among is the protection action against PS failure events, where a
the applications, we have not included PS or multi-robot coordinated set of remote units disconnect sections of the
systems (MRS) because they will be analyzed in Section IV. transmission network, isolating the failure from the rest of
the elements of interest [14]
1) SMART TRANSPORT SYSTEM
The current development of some urban areas have increased 4) EPIDEMIOLOGY
the expectations about services, infrastructure, and mobility Health care is an area in which the application of MAS
among other aspects in the cities. Talking specifically of becomes useful. In specific studies such as epidemiology,
mobility and daily transportation of people, some research which among other objectives is responsible for the analysis
works have used the application of MAS in order to give of the spread of contagious diseases, the inclusion of MAS
innovative solutions to this problem. has made it possible to improve the analysis processes, create
For instance, [45] shows the adaptation of a MAS to models through which health personnel can analyze different
an intelligent transport system for better management of scenes of disease contagions and thus provide effective
traffic in the streets using swarm algorithms. The intention solutions. Taking current situations such as the COVID-19
is to increase the quality of the entire transport network, pandemic, [50] mentions some situations in the health area
avoiding congestion due to many cars on the same routes, in which MAS can be useful, like the correct handling of the
organizing new routes and itineraries considering real traffic data related to the health status of the patients, which would
data, optimizing in this way travel time. With the same be used for a correct allocation of resources. In [51] considers
purpose, [46] developed another application, which uses the decision-making according to the patient’s condition and data
data obtained from a network sensor installed in vehicles analysis of remote patient monitoring. Other works, such
and roads to get accurate data, create the traffic scenario in as [52] focus more on the analysis of the model that represents
specific places, and then use this information in the MAS for the spread of the virus, and thus study its behavior with
traffic light management. different variables involved.

2) SENSOR NETWORKS
III. FC-MAS MODEL
They include a series of sensors and actuators that are To show a general structure in which different MAS can
responsible for collecting data for further processing and be described, we propose a framework based on five layers
providing useful information for a monitoring application. containing main components that could be both essential
In recent years and with current technological advances, and complementary elements, functions, and/or properties
the implementations of these sensor networks are more (Table. 2). Some authors have proposed similar schemes to
affordable. Thus, favoring the implementation of applications analyze MAS from their subject of study perspective; for
where a network of multiple elements is required [47]. These instance, [29] presents agents, interactions and the environ-
sensor networks are widely used in different applications such ment as principal attributes of MAS, and [33] proposes a
as: monitoring and security for highly important sites, caring terminology of subjects of study for MAS, but oriented to the
for animals to detect patterns of their behavior, controlling study of PS.
efficient consumption in industrial and residential facilities, In this case, as one of the main contributions of this paper,
monitoring the environmental behavior of volcanoes, ecolog- we present the FC-MAS (Framework-Components in Multi-
ical reserves and the sea, among others [48]. Agent System) model as a framework that will allow us to
understand the integration of components of the MAS used in
3) EMERGENCY RESPONSE ACTIVITIES the applications according to their characteristics. The layers
Natural disasters are phenomena that seriously affect people’s are hierarchically organized, with this purpose, the physical
lives every time they occur. The devastating effects after network and the fault tolerance are defined as the first and last
their appearance such as fires, floods, collapse of buildings, layer respectively.
among others, cause that any rescue operation to be high It is worth clarifying that the proposed model tries to
risk and sometimes it is not efficient enough to help cover all the possible applications and configurations of the
all the people involved in the accidents. In this type of MAS, so the order and use of the components will depend
situation, the use of MAS, applied to sets of robots, on the application itself or the problem to be solved. Fig. 2
becomes a solution that allows improving search and rescue shows each proposed layer with its components in detail.
activities in emergency situations to help people as quickly Since some basic applications only require fundamental
as possible, without putting in risk to third parties. For components, while other complex applications integrate
example, [49] simulates multi-agent terrestrial and aerial high-level components for greater self-regulation capacity,
robotic systems for earthquake rescues in ROS environment. they (marked with an asterisk) are not fundamental and their

80954 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

TABLE 2. Description of Layers and components of the FC-MAS model. Further, the term node represents a greater number of
Items with an asterisk are optional and their use will depend on their
application. elements in many applications [10].
The term node acquires different interpretations depending
on the area of study, like: computers (Computer Science),
cells (Biology), stations (Transportation Systems), mobile
robots (Robotics), and generators (Power Systems), just
to mention a few examples [54]. In the technological
field, specifically in robotics, the term node commonly
represents unmanned ground vehicles (UGV) or unmanned
aerial vehicles (UAV), which are in charge of a task like:
mapping, transporting, searching, etc [55]. In PS, a node
could represent a group of generators, an agent that controls
and supervises a set of generators, or a set of components
necessary for producing energy that includes the generator,
governor, sensors, and actuators, among the most important
[5], [56].

2) LINKS
Even though a node can be a system itself, it also correlates
with its peers through the called links. In general, a link
represents the interactions between two nodes [57]. One of
the strengths of a MAS is the constant interaction between its
elements, so suitable management of the links can improve
the performance of the system. Hence, in some applications,
the use of links increases the scope of the sensory capabilities
of its agents [58].
In a MAS, or complex systems or networks, a link could
have a binary or a complex nature. In the former case,
the link represents a connection or relationship between
nodes, to mention a few examples: a physical communication
channel, a bridge, a road, a pipe, or cables that connect
the nodes. Meanwhile, in the latter case, the link carries a
numerical value that quantifies the strength of the interaction
or gives a measure about some features of the connection,
some examples may be the bandwidth of the communication
channel, data traffic with system information, or sending and
receiving states between agents, etc. [59], [60]. Graph theory
is widely used to describe and explain the interaction between
nodes through their links.

3) TOPOLOGY
use depends on the application. Each of the layers together
with their respective components is described below. This component refers to the system connectivity, which
explains how nodes interact with each other, or what
information is available for each node [28], [61]. The
A. PHYSICAL NETWORK
topology is generally known as the configuration of a system.
This layer focuses on the elements intervening in the It can be static or fixed, established from the beginning,
infrastructure of the system and the interaction between them. and remain unchanged throughout its life cycle. On the
other hand, some systems have a dynamic topology, also
1) NODE known as switching topology, whose configuration changes
The term node is often used in the field of complex networks as the system evolves or accordingly to its requirements [62].
to represent their more basic entities [53]. More importantly, the topology helps to determine some
Considering the definition of agent proposed in the dynamical features of the system and has a critical relevance
previous section, there is a certain similarity between agent in its robustness [63]. There are several topology alternatives,
and node, the latter being a broader term that allows a such as strongly connected, bus, mesh, star, tree, ring, and
symbolic representation of an entity operating in the MAS. mixed.

VOLUME 12, 2024 80955

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

B. SYNCHRONIZATION 3) TASK ALLOCATION

Synchronization occurs when the elements of a system Nodes or coalitions in MAS have a variety of capabilities
have the same or common behavior over time, which could that allow the system to achieve its objectives. So that all
be the result of internal interactions or also the induction nodes/coalitions contribute to achieving the system’s goal,
of external coordination forces [59]. This layer focuses various tasks are defined and distributed among them. The
on the elements that intervene on the synchronization so former process is known as task decomposition, which
that the system reaches a coupling on some variables or consists of dividing a complex task (or objective) into simpler
objectives. sub-tasks such that they can be carried out by a single node or
a coalition [70]. The later process, known as task allocation,
1) INFORMATION EXCHANGED assigns the sub-task to the nodes and coalitions according
The flow of information between agents through the links to their capabilities to maximize and optimize the overall
is the key to reach a common agreement. So the nodes performance of the system and quickly fulfill the complex
involved in the system are responsible for sensing and task with the least amount of resources [22], [71]. This
processing the information corresponding to the environ- process can be carried out several times while the system
ment in which they operate. This information is shared is running. In the event of a new task or if a node presents
either among all, or their closest, agents to establish a problems, a new task allocation will allow the system to
synchronization for cooperation and achieve a multi-agent restructure the distribution of sub-tasks based on the new
objective. This information has different nature, origin, and requirements.
representation depending on the specific applications. For According to [22] the main goals of employing task
example, in fire emergency scenarios related to search and allocation are:
rescue operation [11], the agents share information such as
victim status, victim location, fire status, fire location, and • Maximizing the number of successfully execute tasks
available resources, among others as numerical, Boolean, • Maximizing the benefits of executing tasks
or arrays variables. In multi-robot SLAM [64], the robots • Minimizing the resources consumed in executing tasks
either exchange positions obtained from odometry or a laser • Minimizing the maximum time spent by agents to
sensor, or a relevant area of a map obtained from the local execute tasks
measurements. Also, as highlighted in many of the studies, the main
elements involved in task allocation are: nodes, available
2) COALITION
resources in each node (i.e. energy, sensors, etc), task
Due to the limited resources and capabilities of a single requirements, sequence of task execution, and constraint
node, it can not complete a task by itself, or it is forced to conditions (i.e. time or resources). For task allocation,
reduce its performance. To avoid this situation, the nodes many methods have been studied in order to improve the
not only exchange information but also form coalitions that performance of the system in specific situations including:
are composed of a group of cooperative nodes, which have Ant Colony Optimization algorithm (ACO), PSO, con-
some abilities or resources that help to solve a task. [65]. tract network protocol, auction algorithm, cost matrices,
A coalition can be defined as an alliance between agents or negotiation according to the availability of the agent
when forming teams that combine their resources and forces [13], [21], [71].
to meet a common goal.
In this component, nodes should generally have access to
all the information and resources necessary to calculate the 4) CONSENSUS
conditions for optimal action. However, there can be resource The convergence of the actions carried out by the nodes
and communication constraints between them that affect the is the product of having consented to and consummated
formation of groups [27]. an agreement between them and is generally known as
Generally, the number of nodes, the task requirements, consensus [72]. Within a MAS, consensus can be established
and the agent’s capabilities determine the number of possible as a law where interaction rules are defined for the exchange
shapes of coalitions in the MAS. All these parameters must be of information between a node with its network neighbors
considered and evaluated to find the best coalition shape. The with whom it can interact. Its implementation difficulty
various possibilities for coalitions are exponentially related depends on the application requirements and their solution for
to the number of nodes that the system has and the number convergence to agreement is generally the key to distributed
of tasks to be accomplished. Thus, the coalition formation coordination [73]. For example, the finite-time consensus
process consumes time and resources of the system [66]. issue can be solved locally in a MAS conformed by n
To avoid a bottleneck in this process, there are seeveral meth- agents. This resolution is feasible provided that a certain
ods such as Particle Swarm Optimization (PSO), game-based condition, denoted as ν, is met. This stipulates that for any
methodology, and genetic and evolutionary algorithms [67], initial state xi (0), located within the neighborhood defined by
[68], [69]. xi (0) − xj (0) < ν, there exists a predetermined time interval

80956 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

T > 0 that satisfies the requirement [72]. behaving improperly by ensuring that local actions do not
lead to undesired behaviors.
lim xi (0) − xj (0) < ν (1)
t→T
3) CONTROL NETWORK ALGORITHM
Additionally, consensus can occur naturally as the result of This component is used to regulate the behavior of the nodes
the coupling of the internal forces of the MAS itself, or also and hence the whole system through the manipulation of the
be the effect of external manipulation through external orders control variables, typically using controllers. Currently, dif-
(forced consensus). The second one is a highly relevant issue ferent investigations propose options for controllers that work
in multi-agent systems [26], and it requires the appropriate at the network level. These control strategies use information
design of a control input that allows several nodes to converge from the entire system to generate the appropriate control
to a common agreement [73]. signals in each node and thus fulfill its tasks aimed at the
global objective. Some approaches propose algorithms based
C. NETWORK CONTROLLER on common control techniques, such it is the case of [76],
To coordinate the nodes and get a good system performance, which uses a common PID control scheme applied to a leader-
the MAS requires a controller other than node’s local one, follower multi-agent system, while others use more complex
or at least to coordinate the later ones. Components involved controllers as a sliding mode control (SMC) or a fuzzy
in this layer work together in the system to determine a control control [77].
signal for each node focusing its behavior on the global
objective. 4) CONTROL ARCHITECTURE
The architecture of a system is often related to its topology
1) ESTIMATION because the connectivity and the nodes intervene in both
The good performance of the MAS is largely due to the concepts. However, the architecture focuses on how the
correct determination of control orders. This, in turn, depends control devices are integrated for the operation of the system.
on the nodes knowing information about their neighbors This term also includes how the control actions, control
or at least having an unbiased estimate of the unknown signals, or local references are assigned to these elements.
states of the system. [74]. The challenge lies in the fact Two architectures are commonly studied and implemented
that, in many complex systems, obtaining comprehensive and in different types of systems: centralized and distributed
reliable information about their states and the environment architectures. When referring to the centralized case, the
is often unattainable due to factors like their substantial participation of a single element is taken into account
scale and the absence of sensors covering the entire system. as the main entity that defines control actions of all the
Consequently, accurately gauging system performance and rest of the nodes in the system [34]. Meanwhile, in a
predicting when a node can complete its designated task distributed architecture the entities are capable of defining
becomes a demanding endeavor. control actions with a certain level of autonomy without
Hence, it is necessary to have tools with the capacity to depending on a central element. According to his philosophy,
find an approximate value of the inaccessible states of the a decentralized architecture can be considered part of a
MAS. Here, the estimation process determines suitable values distributed one. The difference is that in a decentralized
of the states of the MAS based on math models and analyzing architecture, the nodes make their decisions based on local
indirect measurements of a small fraction of the previously information, while in a distributed architecture there is a
sensed variables. Some steps to estimate the states are: define negotiation with information between nodes being a feature
the state variables, collect sensor data, determine the system that makes it more efficient.
model that describes how the state variables change over time, Several works focus on analyzing and implementing
and evaluate the results [75]. the most convenient architecture for a given system. [78]
analyzed the advantages and disadvantages of central-
2) MONITORING ized and distributed configurations applied to the case
Task compliance inspection by system nodes is essential to of servers. While [17] studies the field of smart cities
achieve the global objective. For that reason, it is necessary and uses a hybrid configuration described as centralized
that each node collects information, analyzes it, and tracks locally and distributed globally. In this case, the idea of
the progress of the performed activities, achieving in this integrating these architectures is to take advantage of the
way local monitoring, which is carried out at the level of centralized hierarchy together with the distributed expansion
each agent. In the local monitoring, each node needs to versatility.
establish appropriate indicators like health status, autonomy,
events, alarms, percentage of the task completed, etc. All D. ASSESSMENT
of this is for progress measurement on assigned tasks and In this evaluation layer, the system seeks to reach a new level
providing information on changes over time about the node. of interaction in which the user intervenes and it is possible
Considering that the purpose is to prevent nodes from to assess the functioning of the system.

VOLUME 12, 2024 80957

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

1) DATA ANALYZING 4) DECISION-MAKING

The interaction of a system with its environment involves The decision-making component is the final process carried
working with a large amount of data, obtained by the out in the system in which the obtained information,
nodes’ sensors, and later transforming it into conclusive optimization algorithms, and even the users’ experience are
information, known as data analysis, that helps to improve used to generate a sequence of actions that will achieve the
the system by: solving or explaining problems, predict- general objective of the system. This provides a solution and
ing anomalies or in general providing useful information corrects the deficiencies found [7]. In this component, after
during the operation of it. As in any data management evaluating whether or not the MAS is aligned concerning its
application, there might be cases of bad measurements or objective, the consecutive actions and tasks to be carried out
missing data, so the result after the analysis is inconsistent, based on its current situation are defined. The result of this
which causes additional problems. For this reason, the can be both typical orders if the MAS is behaving as expected,
methods used in this process must be chosen appropriately or they can also be orders for the total restructuring of the
according to several factors such as the type of data, sam- tasks of each agent if the MAS has deviated from its objective
pling rate, resolution, and system resources, among others and can no longer perform it.
[79], [80], [81]. Decision-making is considered a difficult task due to
the existing uncertainties related to dynamic factors in the
2) MONITORING OBJECTIVE environment where the system is located. These unknowns
The monitoring process focuses on the current system’s can come also from the actions that each agent performs
state with respect to the defined goal in the MAS. The to meet objectives at a local level [27]. This has been
MAS’s objectives are the axis that delineates the system’s the motivation for developing various methods applied
operation, and any failure in their specification of leads to its for decision making, such as game theory, reinforcement
malfunction [82]. Therefore, a process is required to monitor learning, swarm intelligence, and evolutionary computing
the use of resources and the general operation of the system among others, which seek to optimize system resources to
while reaching the main system goal. Some works do not the maximum [70].
delve into the definition of monitoring, since it is understood.
However, in [83] a clear definition is presented that can be E. FAULT TOLERANCE
applied to a system with a main goal: monitoring is ‘‘a process This is one ability of the system to continue functioning
that fully and precisely identifies the root cause of an event by correctly and effectively, even in the presence of faults,
capturing the correct information at the right time and at the errors, abnormal behavior, or failures in one or more of
lowest cost in order to determine the state of a system and its agents or components. For use, several factors must
to surface the status in a timely and meaningful manner’’. be considered in its design so that can withstand failures
Monitoring tools have been widely used for tracking of different kinds such as limited agent autonomy, failure
resource utilization and the performance of systems and and loss of communication between agents, changing or
networks. inaccessible environments, limited energy resources, etc.
[24]. Although this property is not mandatory, consideration
3) DEVIATION DETECTION of its implementation is important. It should cover the most
The term deviation is usually seen as outliers, errors, or even common failures, especially in vital applications such as the
noise that come from the data collection carried out by military, health, or PS.
sensors or other elements in the system [84]. The detection of The methods used for fault tolerance will depend on the
deviations arises from the need to compare the measurements system’s requirements. For example, systems can handle
with respect to nominal values and detect which data faults through the concept of fault detection, isolation,
is not consistent or does not comply with the initially and recovery (FDIR) [87], [88], [89]. Studies in computer
established criteria [85]. In short, deviation detection attempts science and robotics analyze faults as threats on hardware,
to determine the error together with the causes that triggered firmware/OS, and application levels, which are studied as
it. Then, a diagnosis is made that determines an inconsistency cybersecurity [90]. In PS, the idea of fault tolerance is
in the obtained information, which can be one of the first lines explored as Remedial Actions Schemes (RAS) [91], [92].
of defense to detect erroneous data. Also, it provides infor- In all of these applications, fault tolerance seeks the proper
mation about a potential system’s operational fault profile. management of system resources in front of failures, and
In this fashion, it is possible to recommend corrective actions also focuses on three main components: fault detection,
to solve deficiencies in the system. System diagnosis is often robustness, and resilience.
based on user experience and lately, it has become automated
using techniques such as information processing, data 1) ROBUSTNESS
mining, and machine learning that provide fast and accurate This is the ability of the system to maintain its functionality
diagnoses [86]. and performance, even in the presence of uncertainties,

80958 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

disturbances, or unexpected events. A robust MAS can cope carried out to ensure the stability of the system in the presence
with changes and variations in the environment and continue of an event. Here, the output of the system or its response
providing the desired services and functionalities. The goal of is constantly monitored and if changes occur outside of the
robustness is to increase the system’s performance, reliability, normal operation, these actions are executed [91], [92].
availability, and dependability, and ensure that the desired
services and functionalities are provided even under adverse IV. FC-MAS IN TECHNOLOGICAL APPLICATIONS
or challenging conditions. During the design and operation Due to the versatility and benefits of MAS, its implementa-
of a MAS, one pays special attention to the robustness of all tion in different technological areas has been carried out in
its layers. This property is considered a pre-event or off-line recent years. Table 3 represents a dispersion of some works
concept since the system is generically designed to satisfy this reviewed on various applications of MAS in the last years in
characteristic [93]. A robust system’s design considers not different technological areas including some applications of
only the conditions in which it operates, but also additional MAS in PS, robotics, and even the Internet of Things. This
disturbances that it can suffer during its normal operations section demonstrates the adaptability and generalization of
such as hardware failures, load variations, or malicious the FC-MAS model through two cases. In the first example,
software. Some works have focused on implementing robust we describe the transport of loads application using a multi-
systems for specific situations or under specific considera- robot system in a distributed operation. In the second case,
tions [94]. we describe the Load-Frequency Control (LFC) developed
by the Automatic Generation Control (AGC) in a centralized
manner in PS.
2) FAULT DETECTION
This is considered a logical component and an essential part A. MULTI-ROBOT SYSTEM
of fault tolerance. It is capable of predicting the state of nodes
A multi-robot system (MRS) is a group of two or more robots
and detecting unusual conditions of the system, through the
capable of interacting with each other to achieve a common
use of the interaction between the elements in the MAS. Its
goal or set of goals such as cooperative formation, leader-
main goal is to minimize the impact of faults on the system’s
follower, rendezvous problem, sweep coverage, etc. One of
performance, by quickly detecting and isolating the faulty
the challenges of designing MRS is ensuring coordination
agents or components, and taking appropriate corrective
and communication between the robots to perform tasks
actions. This term could be studied accompanied by the term
that would be difficult or impossible for a single robot to
isolation and recovery, Fault Detection and Isolation (FDI),
accomplish alone. This can be achieved through various
and FDIR systems. Whose aim is to determine the presence of
techniques, such as centralized control, distributed control,
faults events in a system, occurrence times, and location, and
and consensus-based control. Additionally, designing MRS
depending on the type and severity of them apply strategies
requires considering factors such as task allocation, motion
to restore the functionality of a part or the whole system [89].
planning, and obstacle avoidance. MRS can be applied in a
variety of fields, including search and rescue, exploration,
3) RESILIENCE agriculture, transportation, and manufacturing. The advan-
Refers to the system’s ability to maintain its functionality tages of using MRS include increased efficiency, robustness,
and adapt to changes, perturbations, or failures, even in flexibility, scalability, and the ability to perform tasks that
the face of unexpected or uncertain events. A resilient would be dangerous for human workers.
MAS can recover from disruptions, adapt to changing As a case study, we will describe transports loads within
circumstances, and continue providing the desired services a given area by a ground heterogeneous MRS [6], [12],
and functionalities. Since the resilience methods used to analyzed from the perspective of our proposed framework
correct node malfunction is different than the resilience presented in Section III. The global goal of MRS is to
methods against communication failures, the actions and transport the load from one point to another under certain
behaviors of a resilient system depend on the type of failure, restrictions, ensuring their safety. Meanwhile, the local goal
its location, or its magnitude. According to the field of is to maintain a relative position with respect to its neighbors
study, the term resilience is usually found in conjunction with by following the planned trajectory at a given speed. Suppose
terms such as robustness or FDIR. In robotics, for example, that each robot is assigned a specific task, such as picking
the description of FDIR is normally used to encompass the up a load, moving it to a specific location, or handing it off
aforementioned terms of fault detection and resilience. In this to another robot. The ground robot group is equipped with
case, the system uses observers to estimate output values and enough sensors to know some variables such as location,
take decisions if the real values are not within the estimated route, transported load, and speed. Additionally, everybody
values. With this criteria also any type of information that has built-in communication devices for sending and receiving
may be harmful or malicious is isolated and strategies are information. The control system is based on a distributed
considered to continue with the operation of the system [87], architecture, where each robot has its own autonomy and
[88], [89]. Meanwhile, in PS, the concept of resilience is communication with other robots to coordinate their actions
related to RAS, which by definition are event-based actions and the development of the main task.

VOLUME 12, 2024 80959

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

TABLE 3. Recent MAS applications according to different areas in the last seven years.

1) PHYSICAL NETWORK the task, such as the number of robots, the size of the load,
a: NODE the environment, and the desired level of coordination and
The ground robots’ are the most basic components in this flexibility. For this specific case, a mixed topology between
MRS, known as nodes or agents. Here, it is taken into account tree and mesh topology is the most suitable. These provide
all the mechanisms that allow them to interact with the the robustness and fault tolerance necessary to overcome the
environment and with each other, such as infrared sensors or possibility of changing a topology when the nodes are in
cameras for obstacle detection and evasion, GPS to determine motion or the system requires reconfiguration.
its location, grippers to hold objects engines to move around,
antennas for communication, and local commands to interact 2) SYNCHRONIZATION
with the user [68]. a: INFORMATION EXCHANGED
Robots need to exchange information to effectively coordi-
b: LINK nate their actions and complete the task. Since this example
For this application, the links can be considered in two types, focuses on the mobilization of loads to meet logistics
physical and communication. The physical links consist of objectives, the position and speed of the agent are the
a series of interconnected mechanisms, such as a conveyor main data to be exchanged in the form of coordinates
belt or a series of rollers, that allows the robots to move the or through latitude and longitude points. This data is
load along a designated path. This is with the aim of unifying provided by a GPS, and allows the avoidance of colli-
forces. While communication links are used to communicate sions and maintaining efficient paths. Additionally, other
with one another and coordinate their movements to ensure relevant information to be shared are load weight and
the load is transported safely and efficiently. It is common dimensions, load destination, operating status, battery-life
to use wireless communication, e.g. WiFi or radio frequency charging status, and environmental conditions. With this,
due to the range of coverage. Here, information transmission the robots can know your surroundings and they function
by the link can be subject to the bandwidth of the channel smoothly.
used [79]. Additionally, links can be customized to fit the
specific needs of a particular application, making them a b: COALITION
versatile solution for a wide range of industries. Inside coalitions, each robot has a specific role or task in
the transportation process. For example, some robots may
c: TOPOLOGY form groups for carrying the load, while others may be
Due to the nature of the links, the MRS can have a variety of responsible for navigation or communication. This allows
topologies, which can even change over time. Each of these for the transport of more loads in fewer trips. Here,
configurations has its benefits and is chosen according to the robots work together to transport the load efficiently and
application and will depend on the specific requirements of safely, taking into account factors such as the weight

80960 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

FIGURE 2. MAS framework applied to MRS and PS.

of the load, the terrain, and any obstacles that may d: CONSENSUS
be present. In this way, it increases efficiency, reduces This component plays a crucial role to load transport tasks in
labor costs, and minimizes the risk of injury to human MRS, as they enable the robots to coordinate and make deci-
workers. sions on how to distribute the load among themselves. They
aim to minimize the total transportation time or maximize
c: TASK ALLOCATION
the total load transported. Some consensus algorithms for
load transportation are Auction-based consensus, max-sum
After the coalition, task allocation is the next step. This
algorithm consensus, Task allocation-based consensus, Route
involves determining which robots should be assigned to
planning consensus, Swarm intelligence-based consensus,
which tasks in order to optimize the overall performance of
Collision avoidance consensus, Load balancing consensus,
the system. For this application should be considered factors
and Gradient descent-based consensus. The choice of
such as the weight and shape of the object being transported,
algorithm depends on the specific requirements of the load
the terrain and obstacles in the environment, the capabilities
transport application and the characteristics of the MRS.
of the individual robots, and the overall efficiency and effec-
tiveness of the system. General tasks assigned to the majority
of robots are area sensing, picking up the load, moving it to a 3) NETWORK CONTROLLER
specific location, or handing it off to another robot. Addition- a: ESTIMATION
ally, some specific robots may be managed for navigation or To estimate the states several factors need to consider such as
communication. the number and types of robots involved, the characteristics of

VOLUME 12, 2024 80961

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

the loads being transported, and the environment in which the they collectively find the path with the minimum energy
robots are operating. Typically, the position and orientation consumption.
of each robot, the position and orientation of the loads, and
the velocities of the robots and loads are the state variables 4) ASSESSMENT
to estimate. For this purpose, kinematic or dynamic models a: DATA ANALYZING
are used to describe the motion of the robots and loads. This application can generate a significant amount of data
Estimation techniques such as Kalman filtering, particle that can be analyzed to improve the system’s performance.
filtering, or model predictive control are the most used to For example, some data that can be collected and analyzed
estimate the states of the complete system [75]. are load weight and size to optimize the number of robots
required, robot speed and direction to ensure that they are
b: MONITORING moving efficiently and avoiding possible collisions, power
To develop this activity in MRS, local sensors are used, consumption to reduce the load on certain robots, route
which provide information about the positions, speeds, and optimization to identify areas where congestion is likely to
charges of the robots. In addition to monitoring the robots occur, and finally performance metrics to evaluate the overall
themselves, monitoring the loads that they are transporting is effectiveness of the system.
important too. Sensors can be placed on the loads to measure
their weight, temperature, and other relevant parameters. This b: MONITORING GLOBAL OBJECTIVE
information is presented in real-time on a local monitor and is The information on the current state of the system is
used to track the progress of each robot, detect any anomalies compared with the reference value that it must achieve. Here,
or failures, ensure that the loads are being transported safely, robots are programmed to prioritize certain tasks or to adjust
and optimize the system’s performance [108]. their behavior based on the current state of the system. For
example, information such as the trajectory followed or the
c: CONTROL ARCHITECTURE percentage of packages transported correctly can normally be
The current system works with a distributed architecture viewed in a graphical interface and thus know how much is
because each agent is in charge of generating its control missing to achieve the overall objective.
actions, so it does not require a central entity to send it some
type of signal to function. Each robot uses its sensors to c: DEVIATION DETECTION
perceive its environment and the objects it needs to transport. Focusing on the global objective of this robotic system,
Additionally, they can communicate with each other to it is necessary to have procedures that allow for identifying
exchange information about the objects being transported, deviations from the objective. Some common deviations are
their locations, and the paths they are taking. For this alterations in the trajectory of the agents, agents not able to
application, the distributed architecture is ideal to facilitate withstand the load, or non-fulfillment to transporting the load
the scaling of the system when the integration of new agents for a certain lapse of time, to name a few. In case the robots
is required to transport heavier loads. observe these anomalies in the measurements provided by the
environment, they will use relevant information to conclude
d: CONTROL NETWORK ALGORITHM the possible causes of this erroneous behavior in the system.
For the effective transport of loads in MRS several control Subsequently, the consecutive action is to minimize said
algorithms have been proposed to coordinate all the agents deviation by executing a set of control rules in the predefined
[109]. The most common are: algorithms.
• Distributed consensus algorithm, where each robot
sends its value to its neighbors, and they update their d: DECISION MAKING
values based on the received values. This process In the application, the robots need to consider possible
continues until all the robots have the same value. external factors such as robot failure, the weight change of
• Auction-based consensus algorithm, robots bid for the the loads, the distance between locations, and any obstacles
task of carrying the load. The robot with the lowest bid is in their path, that may affect the objective. If an adverse
selected to carry the load, and the payment is distributed event occurs, commonly one robot takes hierarchical orders
among the other robots based on their bids. to take corrective actions that benefit the entire system. This
• Swarm intelligence-based consensus algorithm, robots agent charge of processing and sending new instructions to
work together as a swarm to transport the load. The the other agents so that they can execute them according to
swarm can adjust its behavior based on the load weight their capacities and the agreements they have reached.
and shape to optimize the transport process.
• Gradient descent-based consensus algorithm, robots use 5) FAULT TOLERANCE
gradient descent to find the optimal path for transporting a: ROBUSTNESS
the load. Agents communicate with each other to share The key to robustness is to design the system with redundancy
information on the load’s weight, shape, and terrain, and and flexibility in mind so that it can adapt to unexpected

80962 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

events and continue to function effectively even in the face frequency to its rated value, (2) remove the transferred power
of individual robot failures or environmental disruptions. deviation in transmission lines.
To solve the load transportation in MRS, adaptive algorithms An important point to keep in mind is that PS is an
that can adjust robot behavior in response to changing interconnected system by nature in which AGC forces an
conditions are used. For example, if a robot encounters additional level of synchronization. Therefore, we can talk
a particularly heavy load or an unexpected obstacle, the about two kind of interconnections, first one is a natural
algorithm could adjust the robot’s speed or path to ensure that connection created by transmission lines and the second one
it can safely navigate the environment and complete its task. that is forced by the AGC’s control schemes, which together
enable generators in PS to describe a collaborative behavior
b: FAULT DETECTION against power unbalance. Finally, it is important to highlight
For this application, the goal of fault detection is to minimize that primary regulation is implemented by the using of local
the risk of failures or errors that could lead to delays or controller (speed governor), and supplementary regulation is
damage to the transported goods. Each robot is equipped implemented at dispatch center level.
with weight and location sensors that measure the load and
position to detect any changes or deviations from what was 1) PHYSICAL NETWORK
agreed. If a sudden drop in weight is detected, the agents a: NODE
automatically trigger an alarm or alert to investigate the issue. In PS, generator and its governor is considered as the node
In some cases, is necessary to send a replacement robot to from frequency control problem viewpoint. This is because
take over the task of carrying the load, while the faulty robot LFC scheme performs variations in the mechanical input
is taken offline for maintenance. power to drive a generator speed variation that contributes
to regulating the system frequency. Under this consideration
c: RESILIENCE
the generator can be identified as an elemental component
In the context of load transportation in MRS, resilience can because any change in its operating point has an effect on the
be achieved through several approaches, such as redundancy system frequency value.
to ensure that if one robot fails, others can take over.
Implementing Adaptive control and learning in each of the b: LINK
robots, task allocation, and considering in the design step
In PS there are two kind of links that enables information
some extra physical or communication links.
exchange between nodes. The first one are the transmission
B. POWER SYSTEMS
lines which allows to exchange power that couples the
generators and drive them to converge to the same speed
The second study case analyzed from MAS viewpoint is the
thus defining the frequency on whole PS. The second one
Power System which we refer as PS. As is well known,
is instead provided through a communication network that
PS is a system designed to provide a continuous and reliable
enables to send the desired generator operating point from
electric power supply that meets certain quality requirements.
dispatch center to governors mainly. Since the different nature
And although the power requirements could be provided
of this links, we will define the first one as natural links and
by an isolated generator connected to the loads directly,
the second one as artificial links.
this operation way can compromise the service continuity
during failures in the generation sources. In this regard, the
interconnection of generators not only increases the system c: TOPOLOGY
capacity, but also provides a backup mechanism between As was just mentioned, the PS presents two kind of links
generators during fault of one generator. that could define the topology. But in LFC the prevailing
During PS operation, the global goal is to keep the balance topology is related to transmission lines, i.e., to natural
between generated and consumed power, which is why when links. In this regard, PS presents a meshed topology in
a power variation is presented it is necessary to implement order to gain robustness and reliability. This is justified by
a system reconfiguration that restores such balance. The the fact that PS arranges several generators interconnected
mentioned reconfiguration is accomplished by LFC and is between them to distribute the generated power requirements.
related directly with local objective of nodes that is to As well as power unbalances that may appear suddenly could
regulate the power provision according to power imbalance. be counteracted too returning PS to an stable state after
Regarding LFC, it is implemented in two stages. The first perturbation.
one called primary regulation is performed by the governors
to achieve an steady state after power imbalance, although 2) SYNCHRONIZATION
with frequency and transferred power deviated from their a: INFORMATION EXCHANGED
rated values. The second stage called secondary regulation Regarding the information exchange LFC control scheme
is accomplished by AGC that change the power setpoint of first takes advantage of natural coupling provided by
generators to restore balance of active power, and as result transmission lines to perform the primary regulation through
two main objectives are accomplished: (1) restore system the governors. This first regulating stage is triggered by

VOLUME 12, 2024 80963

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

an active power unbalance that modifies the speed of its operating scenario by creating a deviation on the power
closer generator and that in turn produces a variation on reference according to the regulation R of every generator.
exchange power with the rest of generators. As result, Meanwhile secondary regulation eliminates frequency and
this power variation is propagated through the transmission transferred power deviation by the convergence of the
lines making known to remaining generators about power deviation of the generators angles, i.e.,
unbalance. In answer each governor changes the mechanical
input power in its generator helping counteract the power 1δ1 → 1δ2 → · · · → 1δn . (4)
unbalance leading to the PS to an steady state operating point
This behaviour could be considered as an artificial consensus
with a new frequency value. It is important to highlight that
because it is accomplished by dispatch center through
the primary regulation is accomplished by the governors in a
communication links in order to modify the power output
distributed way thanks to the natural consensus. After in order
reference in each generator.
to restore the rate value of system frequency, the secondary
regulation will be performed by using its artificial links, i.e.
3) NETWORK CONTROLLER
the communication network. In this case the information is
mainly send by dispatch center to governors to establish the a: ESTIMATION
new power reference on every system generator. Even though estimation could be associated to state esti-
mation applications based on system measurements, in our
b: COALITION analysis it refers to the PS or node capability to know the state
From our analysis we conclude that the control area (CA) variables of neighboring nodes or the whole system. About
defined to regulate the power exchange between power variable estimation on LFC problem, we separated the state
companies is the coalition in LFC, since it is a group of variables that are involved in primary frequency regulation
generators which are governed under a common law based that are speed and angle of generators and variables related
on the Area Control Error (ACE). ACE can be described by to secondary regulation which are system frequency and
the expression (2): transferred powers. For primary regulation, every generator
knows its own state variables (angle and speed) and the
n
X transferred power to the neighboring generators. These
ACEn = Bn 1ωn + 1Pnij (2)
information allows to generator to ‘‘estimate’’ state variables
i=1
of its neighboring from local variables only. In the other
The control law of an area establishes that when the ACE is hand, dispatch center can directly measure system frequency
different from zero, control actions will be performed inside in some PS locations as well as transferred power on
the area to restore the condition ACE = 0 and in turn to transmission lines to know such variables, i.e. the value
counteract PS unbalance. of required variables are available without any complex
estimation procedure.
c: TASK ALLOCATION
As mentioned above, if the ACE is different from zero a b: MONITORING
control action must be established for whole area according Since monitoring stage aims to enable to every node the
to active power unbalance. This control action should define knowledge about its own and the system situation, in the
the new power set-point for every generator inside the LFC context monitoring must detect the occurrence of a
area. This additional amount of generated power in each power unbalance. This monitoring is different in primary and
generator depends on speed regulation or droop R that is a secondary regulation because the first one depends on natural
parameter configured in every governor. This parameter acts connection and the later on an artificial connection.
as a participation factor that define the ‘‘task allocation’’ for From generator viewpoint in primary regulation, unbalance
every generator by defining the power percent by which each is detected as an active power variation in the generator
generator contributes to counteract the power unbalance. terminals because this variation is a result whether by a local
or a remote active power variation. For secondary regulation
d: CONSENSUS ALGORITHM in the dispatch center the key variable that must be monitored
In the PS could be identified two kind of consensus, a natural is system frequency, because a deviation from its rated value
or intrinsic consensus and an artificial or forced consensus. could be understood as an active power unbalance in PS.
Natural consensus is observed when two or more generators
which are interconnected converge to same speed without c: CONTROL ARCHITECTURE
external control actions although no rated value necessary, The LFC control architecture could be considered as a two-
i.e., level hierarchical architecture because a power unbalance is
ω1 → ω2 → · · · → ωn . (3) counteracted in two stages as was mentioned before. In the
first level is the primary regulation that is the closest level to
This advantage of natural consensus AGC is exploited by the PS and works under a distributed architecture since the
primary regulation to lead the system to a first stable control actions are taken by the governors. In the other hand,

80964 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

at the second level in the architecture is the secondary control system frequency variations and deviations from scheduled
whose actions are defined in the dispatch center, therefore its values of transferred power between areas. In this regard,PS
architecture could be considered as centralized. From this we unbalance is detected from ACE value, i.e., if ACE value
can conclude that the frequency regulation is performed under is different to zero, then control actions are needed in order
a hybrid architecture. to return frequency and transferred power to their scheduled
values.
d: CONTROL NETWORK ALGORITHM
From the fact that at the PS level the objective is to counteract d: DECISION MAKING
the active power unbalance, therefore the main aim is to reach After an unbalance of active power is detected, control actions
the balance between generated power and consumed power, are taken to counteract it by changing set-point of power
i.e., output on some generators in the affected control area. In this
n
X m
X stage, dispatch center defines the new operation points that
Pgi = Plj (5) are sent to generation station by a communication system,
i j i.e. by an artificial link. After a new set-point for generators
could be considered the basic relationship that will define at station is configured according with characteristic of every
any control algorithm for LFC. Under this consideration one.
the classical LFC drives a change in mechanical input
power of generators that in turn change the electric power 5) FAULT TOLERANCE
generated to restore the power equilibrium. The additional a: ROBUSTNESS
amount of generated energy is defined on base of ACE error About robustness it must be taking in account that LFC leads
minimization that in an ideal scenario must be zero, i.e., PS generators to respond to a disturbance in a cooperative
way by distributing the ‘‘responsibility’’ to each according
ACE1 → ACE2 → . . . → ACEn (6)
its capability. Therefore it is possible to conclude that LFC
that in turn accomplishes scheme itself is a mechanism that provides robustness to PS.
In addition, procedures for PS operation have established
1δ1 → 1δ2 → . . . → 1δn → 0 (7) reliability criteria such as N − 1 criterion that pretends that
resulting in a new power balance as well as a system PS is able to withstand at all times an unexpected failure
frequency restoration and transferred power deviation elim- or outage of a single system component, has an acceptable
ination. reliability level.

4) ASSESSMENT b: FAULT DETECTION

a: DATA ANALYZING From LFC viewpoint fault detection could be accomplished
Since data analyzing is an assessment stage at ‘‘supervisory’’ by a similar procedure of PS unbalance detection, that
level, secondary regulation is considered only. That means is by detecting frequency and transferred power deviation
data analyzing will provide useful information about system from their planned values. But it is important to consider
frequency and transferred power through transmission lines that LFC is designed for counteracting power unbalances
that enables the monitoring global objective. In this regard that are normal during PS operation an in which variables
frequency and power through transmission lines are measured present a tolerable variation unlike fault conditions in
and centralized in dispatch center to calculate the current which the variables could be deviated considerably. In this
ACE to evaluate if a power unbalance is presented. regard, although LFC can detect the fault, it cannot face
a considerable power unbalance produced by it. Therefore
b: MONITORING GLOBAL OBJECTIVE fault detection task must be accomplished by a systemic
PS must monitoring the active power balance by checking protection scheme that operates in parallel to LFC and stays
that system frequency must be maintained constant and ACE as a passive function while PS maintains its normal operating
error must be zero. If any of both conditions isn’t true, and performs its control action under a fault condition only.
it could be understood as a power unbalance that requires
corrective actions that should be applied and maintained c: RESILIENCE
while variation persists since it will mean that PS unbalance In order to increase PS resilience several generators with
is not compensate yet. different power production characteristics could be con-
sidered because they have the capability to deal different
c: DEVIATION DETECTION kinds of weather conditions, natural disasters, etc. Further,
As mentioned in Control network algorithm, global objective a coordinated scheduling of different power sources is
in LFC is the active power balance in order to let the effective to improve generation resilience. Also transmission
system to an stable operating scenario. In this regard, networks need to have resilience to face disturbances caused
PS monitoring if its objective is accomplished by detecting by natural disasters or operating reconfiguration. Thus

VOLUME 12, 2024 80965

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

alternative corridors should be established in the event of Here, the agents execute the control commands issued
a main transmission path faults. Also with a future vision, by the agent master For example, a Power System’s
smart loads could be an alternative because they would have frequency control coincides with being a multi-agent
regulating abilities for contributing PS resilience [97]. system with a centralized architecture [110]. Where, its
control center decides and establishes the power of each
V. MAS DEPLOYMENT FLOW of the generation units, while these change their power
Implementation of MAS phases like execution and start-up until reaching an energy balance according to the order
can be a quite complex operation due to the heterogeneity received.
of the interconnected agents and the conditions of the Distributed : The development of MAS applications with
environment in which they interact. Each application requires a distributed architecture has gained importance in
the integration of a set of algorithms, procedural protocols, the last decade because of the limitations of the
and frameworks for its correct operation, but up to date, centralized architecture. This architecture distributes the
there is not a proposed organized scheme that guides what main functions, previously performed by a master, to the
are the basic resources that researchers should consider to autonomous agents (senior agents). This is achieved by
implement an application in MAS. As a consequence of the defining a set of interaction rules between the agents.
framework presented in section III, this work complements The challenge is to coordinate the decentralization
the proposal with a workflow for the practical implementation and distribution of decision-making processes among
of multi-agent systems in any area. For this, workflows individual autonomous agents, so that each agent actions
based on a sequence of stages and steps for centralized do not interfere with the activities of their neighbors,
and distributed architectures in MAS are proposed. Fig.3 while achieving a common objective. This analysis
presents a general description of the necessary stages for the considers new aspects of centralized architecture, such
process of execution and commissioning of a MAS, seen as instruction execution, primary resource distribution,
from the engineering perspective. This proposal is based on communication channel usage, command hierarchy,
defining certain implementation steps within three stages: establishment of interaction rules, etc. Several proposals
Initialization, Supervision, and Control. Next, we proceed to from different areas have attempted to solve this
describe the architectures, their stages, and the variants of the challenging task. Topics such as energy efficiency,
necessary steps according to the operation requirements for real-time performance, coexistence, interoperability,
each application. security, and privacy have been addressed in relation
to how distributed control has facilitated the rapid
A. ARCHITECTURES development of Industry 4.0 [111], [112]. Similarly,
Centralized: In the 60’s the first MAS applications began methodological proposals demonstrate the feasibility of
with a centralized architecture, which is characterized distributing control models to achieve greater benefits
by depending on a main agent called the control center from MAS [113], [114]. Its implementation opens new
or master, who is in charge of coordinating all the challenges such as the difficulty in distributing and
tasks of the rest of agents [34], [78]. This architecture coordinating the tasks of each agent, the synchronization
is suitable for solving applications where they require problem, and obtaining consensus on the variables of
a limited number of agents. Its greatest limitation is interest [35]. This architecture allows the scalability of
its capacity to expand, since by having to centralize a MAS system with the incorporation of new agents
all the information in a central agent, it must have [115]. In case the application requires a master agent
a great processing and decision-making capacity to as a global supervisor, this will be responsible for
generate the control orders of the rest of the agents. executing only the initialization and supervision stages.
For this case, centralized workflow is shown in Fig.4. While in the control stage, autonomous agents now with
Here, the proposed stages and those responsible for greater processing capacities perform main tasks such as
establishing a centralized MAS system are: the master coordination, decomposition, and decision-making on
agent, who coordinates the actions of the rest of the their own. Through the implementation of distributed
agents, will be in charge of executing the initialization algorithms, the agents establish the adequate exchange
and supervision stages, and also the global control sub- of information and each one is capable of coordinating
stage. The master will establish the global objective to be their own actions for the fulfillment of their individual
achieved and will be in charge of monitoring the status tasks, which in turn converge in accomplishing suc-
of the deviation from the progress of the agents’ tasks. cessfully the common global objective. The proposed
In the event of adverse circumstances within the MAS, workflow for this case is shown in Fig. 5. In this case,
the master will execute a decision-making function since each of the autonomous agents determines its
to try to return the system to operating conditions. own orders, the system itself can work faster because
While in the local control system stage, agents sense it no longer waits for the master to make and send the
and manipulate the environment to fulfill their tasks. decisions. However, the master will now take care of

80966 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

FIGURE 3. Centralized and distributed architectures.

FIGURE 4. Workflow centralized.

hierarchical functions such as objective definition and include initialization, supervision, and control. Figure 3
also monitoring for system convergence. illustrates these stages arranged hierarchically from top to
bottom. The utilization of individual steps within these stages
depends on the specific requirements of the MAS application
B. PROPOSED STAGES AND STEPS under development. This reliance is influenced by factors
In the proposed workflow, a stage refers to a group of steps such as architecture, available resources, number of agents,
that collectively aim to achieve a specific goal, and they need communication infrastructure, etc. [78]. The design of the
to be followed sequentially to establish a MAS. Depending proposed workflow aims to accommodate a wide range of
on the architecture type, these stages can be executed either applications, but it also maintains the flexibility to exclude
by the master agent in a centralized scheme or by the certain steps that may not apply to specific MAS applications
senior agents in a distributed scheme. The proposed stages or when their execution is carried out by an external operator.

VOLUME 12, 2024 80967

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

FIGURE 5. Workflow distributed.

The following sections describe each stage in detail, along of the equipment should be compiled, including the
with their corresponding steps. number and current status of agents, communication
channels and infrastructure, computational resources of
1) INITIALIZATION the supervisor, primary energy sources, and, if feasi-
Like any process that requires initialization, a MAS is no ble, additional resources within the agents’ operating
exception. This stage is typically carried out by a supervisor environment.
(sometimes referred to as a global master) in either a Set-up agent and MAS: After defining the resources, the
centralized or distributed scheme. Its primary objective is next step involves activating the main components of
to construct a MAS application by leveraging available the MAS. This involves too initiating communication
resources and their initial states. In this regard, launching any between the agents. The initial activation of these
application involves an analysis of the available resources to components will provide information on the initial
ensure the operability and autonomy of each agent. Following conditions of each entity within the MAS [28], [34].
that, the system variables are fully initialized at both the agent Global objective Definition: In this step, the design of
and global levels. Lastly, the MAS establishes or proposes functions and activities to be implemented within a
the global objective to be achieved [108]. The recommended Multi-Agent System (MAS) is undertaken. The primary
steps for this stage are as follows: aim is to establish the operational objective, which
defines how MAS agents will interact with their
Resource check-up: A crucial aspect of initializing any environment to accomplish specific tasks [5], [28]. The
Multi-Agent System (MAS) is gaining a comprehensive scope of the overall objective may be constrained by
understanding of all the physical and logical resources at factors such as resource availability, component state,
hand. By assessing their availability, it becomes possible and the creative abilities of the MAS implementation
to define objectives with varying degrees of scope. The team. Depending on the engineering application, the
aim of this step is to quantify the MAS’s state of health, overall goal can range from a complex set of multiple
offering insights into the scope and limitations when tasks to a relatively simple exercise. Regardless, it is
formulating global objectives based on the available crucial to regularly evaluate the MAS’s progress towards
resources [8]. To accomplish this, a detailed inventory achieving the proposed global objective. Additionally,

80968 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

it should be noted that advancements in technology, the proper functioning of the MAS, and its implementation
techniques, and algorithms provide new opportunities to steps depend on the chosen architecture. If the application
flexibly update the global objective based on potential is centralized, the master agent takes full responsibility for
new requirements arising during the decision-making generating orders for all agents. On the other hand, in a
process of the MAS [6]. distributed architecture, this stage is carried out by the agents
coordinated by the control strategy [61], [63]. This process
2) SUPERVISORY involves several steps, including task decomposition and
We can analyze the second stage of superior hierarchical evaluation. It can occur after the initialization or during the
order from two perspectives. Firstly, a central entity (referred operation of the MAS, in which case the system may have
to as the master agent) oversees and evaluates the overall intentionally or unintentionally changed, resulting in a shift
alignment of the MAS in achieving the proposed objective. in its global goal or consensus state.
Secondly, each agent is responsible for local monitoring During this stage, the performance of each agent in
internally to validate the fulfillment of its assigned task fulfilling their assigned task is individually analyzed. In the
if is a distributed architecture. Regardless of the scenario, centralized case, local-level controllers are implemented
a continuous monitoring function is active to assess the within the agents to manipulate their variables of interest
condition and progression of the MAS and its environment. and accomplish their tasks [56]. In contrast, in a distributed
This evaluation of information helps determine whether or architecture, this stage is considered as part of the General
not the MAS will achieve its objective [59], [108]. Control System stage. The internal controllers can be simple
or complex and are implemented based on the character-
Monitoring: In this particular step, the focus is on the istics and computational resources available in the agents.
execution of a fundamental workflow function. This Commonly used controllers include comparators, hysteresis,
function is responsible for supervising the acquisition and proportional-integral-derivative (PID) controllers due to
and processing of data in order to determine the current their ease of implementation. However, agents with advanced
state of the agents and their environment [82], [108]. capabilities can now include controllers such as fuzzy, fuzzy-
To carry out this task, at least two layers of a Multi- PID, neural networks, and sliding mode control (SMC) [28].
Agent System (MAS) are involved: the physical network
layer and the synchronization layer. These layers rely Data Acquisition: This step involves measuring the envi-
on the measurement acquisition infrastructure installed ronmental variables either through sensors installed in
either in the agents themselves or within the MAS the agents or strategically distributed sensors in the
environment. The primary objective of this step is to work area [47]. The collected data is then shared with
consistently identify patterns of deviation or errors in the supervisory entity or neighboring agents within
task completion or the overall objective. This is crucial the network using various communication channels.
for issuing accurate control commands. Within a MAS, It is important to note that since the data is exposed
the monitoring function is executed, and the results are to the work environment, it is susceptible to noise.
presented through a visualizer, display, or a comprehen- Therefore, if noise is present, a filtering stage should
sive SCADA system. This ensures that operators have be implemented to mitigate its impact. This approach
access to up-to-date information regarding the agents’ helps reduce the bandwidth usage of the communication
behavior and information exchange. network and minimizes computational efforts required
Decision Making: The decision-making capacity is a crucial for other processes [80].
element that applies to both individual agents and the Data Processing: This step initiates the processing of the
overall MAS [7]. It is an integral part granted to the gathered information concerning the environment and
elements of the MAS to fulfill their assigned tasks effec- the progress made by the agents in reaching their
tively. Various prominent algorithms in this field include objectives. Depending on the task assigned to the
Markov decision processes (MDP), Partially Observable agents in aMAS, one or more of these agents will be
MDP, game theory, swarm intelligence, graph theoretic responsible for processing the data and transforming it
models, reinforcement learning, dynamic programming, into meaningful information that has been predefined.
evolutionary computing, and neural networks [8]. The This relevant information will then be disseminated
primary goal of this process is to assess the cost/benefit across the network [80]. If there is any loss or alteration
objective function and make optimal decisions that of information during communication, data mining
enable each agent to fulfill its designated task while algorithms can be employed to eliminate or rectify
ensuring the MAS as a whole achieves its intended erroneous data.
objective. System and environment state determination: In this step, the
current state of the system and the environment in which
3) CONTROL the agents are performing their tasks is determined
To achieve the global objective, the tasks can be distributed through the implementation of a series of techniques
among the components of a MAS. This stage is crucial for and strategies. This information plays a crucial role

VOLUME 12, 2024 80969

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

in supervision as it enables the assessment of whether considering the environment [65], [66]. Together with
the MAS is progressing towards its objective or if the resource and task allocation, this sub-process can
it has deviated from it [25], [28]. During the initial accomplish an optimal performance when creating
execution, after establishing the global objective, this teams. As in resource allocation, the algorithms used for
step quantifies the availability and operational status of this purpose are based on AI and ML.
resources. This enables the identification of resources
and facilitates the distribution of activities. c: TASK ALLOCATION
Objective evaluation: This step focuses on identifying any In this step, the global task or the specific task are assign
discrepancies or errors in the MAS concerning the to each team or agent. In other words, this part of the
defined global objective if it is under wide supervision. process defines the action that must take each team or
Too, it aims to detect deviations in the variables agent to achieve the global goal or a consensus [21],
of interest compared to their reference values if the [71].
supervision is local. If the system is nearing its objective, Task evaluation: The tasks in the MAS are assigned to
the workflow will return to the monitoring step without one or several agents according to their capacity and
requiring any control actions. Conversely, if deviations their willingness to operate [84]. This step focuses on
exist, the workflow will proceed to implement algo- evaluating the fulfillment of these tasks and achieving
rithms that facilitate achieving the necessary consensus. the desired objective. In the event that the agents are
The algorithms used to assess deviations can range from capable of fulfilling their previously assigned tasks,
basic methods, such as comparing the reference value the workflow will return to the monitoring step since
to the actual measurement, to more intricate approaches everything is working as expected. But in the event that
involving numerous calculations and estimations [57]. for internal reasons (exhaustion of useful life, lack of
The latter plays a crucial role in understanding state fuel, lack of communication, etc.) or external reasons
deviations, changes in the global task, or the MAS itself, (such as changes in the environment with conditions not
and taking appropriate actions accordingly [10]. foreseen in the design) to the agents, the tasks may be
Task Decomposition: In this step, the analysis of the global redefined by a supervisor based on available resources.
objective takes place, and its distribution into specific This last case will give way to decision making.
tasks is planned to incorporate it into the MAS based Actuators: This step corresponds to the execution of orders
on the unique capacities and functions of each agent. by the agents in order to achieve their objective or
In a basic MAS, this task is typically handled by the proposed goal, as appropriate [47], [55]. Here the agents
designer beforehand. However, in certain cases, certain will see the need to interact with the environment and,
algorithms related to task decomposition are employed if necessary, manipulate it at their convenience through
in subsequent steps of the process, such as coalition their actuators or final elements. Depending on the
formation and team allocation [5], [70]. This phase of control actions, the changes that the environment may
the process is crucial for attaining the global goal, as an undergo can be small or large.
optimal system response can only be achieved through
the proper allocation of activities among all components.
C. WORKFLOW EXAMPLE
a: RESOURCE ALLOCATION Below, we use the proposed workflow to describe the start-
Due to the MAS’s global objective and depending on up of the LFC in a PS carried out by AGC in a centralized
the agents and the environment, often one needs to way, specifically, for the problem of primary and secondary
analyze the resources available. This sub-process is in regulation. We consider that the PS starting operating
charge of analyzing the environment, distributing the conditions are a stable state, that is, all the generators have
resources and tasks, and creating teams of agents. This their power reference PrefP i and thereP is a balance between
step requires specific algorithms for this purpose which demand and generation ( ni Pgi = m
j Plj ). Furthermore,
includes Artificial Intelligence (AI), Machine Learning it is assumed that the generators have been synchronized
(ML), adaptive algorithms among others [22], [71]. One concerning the slack bar.
of the most important consideration when distributing
and optimizing resources is the amount of time or energy 1) INITIALIZATION
that an agent needs to reach certain state related with the Resource check-up: The first step consists of determining
global objective. the exact resources that PS has, such as the number of
generators, transmission lines, and auxiliary equipment.
b: COALITION It is necessary to know the operating status of each one
This sub-process is in charge of creating groups or teams of them to determine whether or not it is in a condition
of agents to maximize the performance of the agents to be integrated into the LFC.
in the MAS while achieving the global goal. Here, the Set-up agent and MAS: The control center executes the
main focus is the global objective (or consensus) without optimal power flow to establish the operating scenario of

80970 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

FIGURE 6. Workflow example of PS. The red line represents the feedback to close the control loop. All
generators send their measurements to the SCADA.

each of the generators, thus giving its initial condition. their competition. The local goal of each generator is
With this, the power of each generator is known before to increase its power production. It only takes what it
a load variation event in the system. is entitled to.
Global objective Definition: The global goal of the LFC Task evaluation: Compliance with frequency regulation
in any PS is to keep the balance between demand occurs when the ACE is equal to zero, that is, there are no
and generation, and also maintain the frequency of the frequency deviations, so the required balance between
system constant, so the control center deploys the AGC generation and demand has been reached. If the ACE
described above into action. This controls centrally, the is different from zero, the system determines the power
primary and secondary regulation stages so that the that each generator must absorb, which is done in the
system always remains within its nominal operating decision making step. This task is constantly performed.
values. Here, the references of the system are defined, Decision Making: It is activated when there are deviations
both in Power (Pref ) and frequency (fref ). with respect to the objective, for instance, when the
consumers’ consumption is higher or lower than the
2) SUPERVISORY power produced by the generators, or when the power
Monitoring: It is necessary to know the current state consumed exceeds the maximum power that the genera-
of the PS to determine if it effectively meets the tors can provide. Depending on the situation, the system
balance condition between generation and demand. should decide to turn on/off generators, redistribute the
The variables that are measured in each ACE in the power generation, trip load, or reconfigure the network
system correspond to the power generated, the power so the system can continue operating.
transferred, and its current frequency. These variables Actuators: Hydraulic and thermal energy are typically used
are monitored in the SCADA and in a later phase they as primary resources, which are then transformed into
will be used to control the balance of power exchange. electrical energy as they pass through the mechanical
Objective Evaluation: Corresponds to the evaluation of and electrical components of the generator. In this case,
ACEn to eliminate power deviations in each of the actuators such as valves and gates, are used to regulate
generation centers and finally maintain the balance resource flow and control the turbine’s speed, which in
between generation and demand. If frequency variations turn moves the generator’s rotor. These actuators are
concerning the reference are detected, the ACE calcu- located inside the governor and interact with the turbine.
lates the total power that must be compensated in the
overall system. VI. CONCLUSION AND FUTURE DIRECTIONS
This study introduces FC-MAS, which offers an effective
3) CONTROL abstraction for component integration and resources, facil-
Task Decomposition: The power that has been determined itating the understanding and development of engineering
by the ACE to compensate for frequency variations applications in the MAS. Each layer focuses on specific
is sent to the generators through a participation factor aspects, such as physical network, synchronization, network
Bi to distribute the power among them thus avoiding controller, assessment, and fault tolerance. Furthermore,

VOLUME 12, 2024 80971

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

it outlines a comprehensive workflow that describes the [11] V. Nunavath and A. Prinz, ‘‘Visualization of exchanged information
essential stages and steps multi-agent systems must follow with dynamic networks: A case study of fire emergency search and
rescue operation,’’ in Proc. IEEE 7th Int. Advance Comput. Conf. (IACC),
to accomplish their objectives. To illustrate the practicality of Jan. 2017, pp. 281–286.
this layer model, we apply it to the domains of robotics and [12] G. H. Merabet, M. Essaaidi, H. Talei, M. R. Abid, N. Khalil, M. Madkour,
power systems, showcasing its effectiveness. Additionally, and D. Benhaddou, ‘‘Applications of multi-agent systems in smart grids:
A survey,’’ in Proc. Int. Conf. Multimedia Comput. Syst. (ICMCS),
we provide a comprehensive compilation of key terms Apr. 2014, pp. 1088–1094.
pertinent to the study of multi-agent systems, accompanied [13] D. Maldonado, J. Cepeda, S. Gamboa, J. A. Torres, and
by real-world examples demonstrating MAS applications. P. J. Cruz, ‘‘Optimum frequency response with decentralized
drivers in power system,’’ IET Gener., Transmiss. Distribution,
During the development of this project, some basic vol. 17, no. 9, pp. 1985–1998, May 2023. [Online]. Available:
approaches have not been addressed in depth, so its full https://ietresearch.onlinelibrary.wiley.com/doi/abs/10.1049/gtd2.12774
development is proposed for future work. Such is the case [14] M.-T. Kuo and S.-D. Lu, ‘‘Design and implementation of real-time
intelligent control and structure based on multi-agent systems in
of the selection of essential components, since for this microgrids,’’ Energies, vol. 6, no. 11, pp. 6045–6059, Nov. 2013.
work these have been chosen according to their incidence [15] B. Wang, J. Wang, B. Zhang, and X. Li, ‘‘Global cooperative control
and relevance in the information collected. Similarly, it is framework for multiagent systems subject to actuator saturation with
necessary to mention the existing discrepancy with the industrial applications,’’ IEEE Trans. Syst., Man, Cybern., Syst., vol. 47,
no. 7, pp. 1270–1283, Jul. 2017.
terms that intervene in a multi-agent system in the fields of [16] B. Wang, J. Wang, B. Zhang, W. Chen, and Z. Zhang, ‘‘Leader–follower
study analyzed; although an agreement was established by consensus of multivehicle wirelessly networked uncertain systems subject
naming the components of the model presented, future works to nonlinear dynamics and actuator fault,’’ IEEE Trans. Autom. Sci. Eng.,
vol. 15, no. 2, pp. 492–505, Apr. 2018.
could focus on improving this foundation, assigning terms [17] P. K. Sharma and J. H. Park, ‘‘Blockchain based hybrid network
capable of being used in more areas of application of MAS. architecture for the smart city,’’ Future Gener. Comput. Syst., vol. 86,
Furthermore, future research may concentrate on the topic of pp. 650–655, Sep. 2018.
[18] B. Wang, W. Chen, B. Zhang, and Y. Zhao, ‘‘Regulation coop-
interoperability and compatibility when integrating the FC- erative control for heterogeneous uncertain chaotic systems with
MAS model with existing or developing MAS technologies. time delay: A synchronization errors estimation framework,’’ Auto-
This will entail consideration of potential integration issues matica, vol. 108, Oct. 2019, Art. no. 108486. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0005109819303383
or conflicts that may arise, as well as the necessity to [19] J. Qin, Q. Ma, Y. Shi, and L. Wang, ‘‘Recent advances in consensus of
balance MAS performance when system resource constraints multi-agent systems: A brief survey,’’ IEEE Trans. Ind. Electron., vol. 64,
or limitations have to be taken into account. no. 6, pp. 4972–4983, Jun. 2017.
[20] W. Ren, R. W. Beard, and E. M. Atkins, ‘‘Information consensus in
Being FC-MAS a new model based on the compilation
multivehicle cooperative control,’’ IEEE Control Syst. Mag., vol. 27,
of different works, no applications were found in which no. 2, pp. 71–82, Apr. 2007.
each and every one of the proposed components has been [21] H. Chuin Lau and L. Zhang, ‘‘Task allocation via multi-agent coalition
implemented. So the systems proposed in the case studies formation: Taxonomy, algorithms and complexity,’’ in Proc. 15th IEEE
Int. Conf. Tools Artif. Intell., 2003, pp. 346–350.
have been adjusted in such a way that all the components [22] J. Zhou, X. Zhao, X. Zhang, D. Zhao, and H. Li, ‘‘Task allo-
intervene in its operation. It can be noted that both the user cation for multi-agent systems based on distributed many-objective
and the tasks are not treated as explicit components of the evolutionary algorithm and greedy algorithm,’’ IEEE Access, vol. 8,
pp. 19306–19318, 2020.
model. [23] A. M. Farid, ‘‘Multi-agent system design principles for resilient operation
of future power systems,’’ in Proc. IEEE Int. Workshop Intell. Energy Syst.
REFERENCES (IWIES), Oct. 2014, pp. 18–25.
[1] S. Greengard, The Internet of Things. Cambridge, MA, USA: MIT Press, [24] K. Potiron, A. E. F. Seghrouchni, and P. Taillibert, From Fault
2021. Classification to Fault Tolerance for Multi-Agent Systems, vol. 67. Cham,
[2] M. Ghobakhloo, ‘‘Industry 4.0, digitization, and opportunities for Switzerland: Springer, 2013.
sustainability,’’ J. Cleaner Prod., vol. 252, Apr. 2020, Art. no. 119869. [25] M. Wooldridge, An Introduction to Multi-Agent Systems. Hoboken, NJ,
[3] B. W. Wirtz, J. C. Weyerer, and C. Geyer, ‘‘Artificial intelligence and USA: Wiley, 2009.
the public sector—Applications and challenges,’’ Int. J. Public Admin., [26] Y. Li and C. Tan, ‘‘A survey of the consensus for multi-agent systems,’’
vol. 42, no. 7, pp. 596–615, May 2019. Syst. Sci. Control Eng., vol. 7, no. 1, pp. 468–482, 2019.
[4] S. A. DeLoach, M. F. Wood, and C. H. Sparkman, ‘‘Multiagent systems [27] P. G. Balaji and D. Srinivasan, ‘‘An introduction to multi-agent systems,’’
engineering,’’ Int. J. Softw. Eng. Knowl. Eng., vol. 11, no. 3, pp. 231–258, in Innovations in Multi-Agent Systems and Applications-1. Berlin,
2001. Germany: Springer, 2010, pp. 1–27.
[5] A. M. Uhrmacher and D. Weyns, Multi-Agent Systems: Simulation and [28] A. Dorri, S. S. Kanhere, and R. Jurdak, ‘‘Multi-agent systems: A survey,’’
Applications. Boca Raton, FL, USA: CRC Press, 2009. IEEE Access, vol. 6, pp. 28573–28593, 2018.
[6] L. Busoniu, R. Babuska, and B. De Schutter, ‘‘A comprehensive survey [29] G. Weiss, Multiagent Systems: A Modern Approach to Distributed
of multiagent reinforcement learning,’’ IEEE Trans. Syst., Man, Cybern., Artificial Intelligence. Cambridge, MA, USA: MIT Press, 1999.
C Appl. Rev., vol. 38, no. 2, pp. 156–172, Mar. 2008. [30] A. Kantamneni, L. E. Brown, G. Parker, and W. W. Weaver, ‘‘Survey
[7] Y. Rizk, M. Awad, and E. W. Tunstel, ‘‘Decision making in multiagent of multi-agent systems for microgrid control,’’ Eng. Appl. Artif. Intell.,
systems: A survey,’’ IEEE Trans. Cognit. Develop. Syst., vol. 10, no. 3, vol. 45, pp. 192–203, Oct. 2015.
pp. 514–529, Sep. 2018. [31] S. D. J. McArthur, E. M. Davidson, V. M. Catterson, A. L. Dimeas,
[8] M. Oprea, ‘‘Applications of multi-agent systems,’’ in Information N. D. Hatziargyriou, F. Ponci, and T. Funabashi, ‘‘Multi-agent systems
Technology. Berlin, Germany: Springer, 2004, pp. 239–270. for power engineering applications—Part I: Concepts, approaches,
[9] K. P. Sycara, ‘‘Multiagent systems,’’ AI Mag., vol. 19, no. 2, p. 79, 1998. and technical challenges,’’ IEEE Trans. Power Syst., vol. 22, no. 4,
[10] J. Leite, A. Omicini, L. Sterling, and P. Torroni, Declarative Agent pp. 1743–1752, Nov. 2007.
Languages and Technologies: First International Workshop, DALT 2003, [32] S. T. Goonatilleke and B. Hettige, ‘‘Past, present and future trends in
Melbourne, Australia, July 15, 2003, Revised Selected and Invited Papers, multi-agent system technology,’’ J. Européen des Systèmes Automatisés,
vol. 2990. Berlin, Germany: Springer, 2004. vol. 55, no. 6, pp. 723–739, Dec. 2022.

80972 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

[33] M. H. Moradi, S. Razini, and S. Mahdi Hosseinian, ‘‘State of art of [58] K.-K. Oh, M.-C. Park, and H.-S. Ahn, ‘‘A survey of multi-agent formation
multiagent systems in power engineering: A review,’’ Renew. Sustain. control,’’ Automatica, vol. 53, pp. 424–440, Mar. 2015.
Energy Rev., vol. 58, pp. 814–824, May 2016. [59] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang,
[34] A. Sujil, J. Verma, and R. Kumar, ‘‘Multi agent system: Concepts, ‘‘Complex networks: Structure and dynamics,’’ Phys. Rep., vol. 424,
platforms and applications in power systems,’’ Artif. Intell. Rev., vol. 49, nos. 4–5, pp. 175–308, 2006.
no. 2, pp. 153–182, Feb. 2018. [60] S. Pajevic and D. Plenz, ‘‘The organization of strong links in complex
[35] X. Ge, F. Yang, and Q.-L. Han, ‘‘Distributed networked control systems: networks,’’ Nature Phys., vol. 8, no. 5, pp. 429–436, May 2012.
A brief overview,’’ Inf. Sci., vol. 380, pp. 117–131, Feb. 2017. [61] K. You and L. Xie, ‘‘Network topology and communication data rate
[36] H. Peng, J. Wang, S. Wang, W. Shen, D. Shi, and D. Liu, ‘‘Coordinated for consensusability of discrete-time multi-agent systems,’’ IEEE Trans.
motion control for a wheel-leg robot with speed consensus strategy,’’ Autom. Control, vol. 56, no. 10, pp. 2262–2275, Oct. 2011.
IEEE/ASME Trans. Mechatronics, vol. 25, no. 3, pp. 1366–1376, [62] Y. Chen, J. Lu, X. Yu, and D. J. Hill, ‘‘Multi-agent systems with dynamical
Jun. 2020. topologies: Consensus and applications,’’ IEEE Circuits Syst. Mag.,
[37] M. Haenlein and A. Kaplan, ‘‘A brief history of artificial intelligence: On vol. 13, no. 3, pp. 21–34, 3rd Quart., 2013.
the past, present, and future of artificial intelligence,’’ California Manage. [63] X. F. Wang, ‘‘Complex networks: Topology, dynamics and synchroniza-
Rev., vol. 61, no. 4, pp. 5–14, Aug. 2019. tion,’’ Int. J. Bifurcation Chaos, vol. 12, no. 05, pp. 885–916, May 2002.
[64] M. T. Lazaro, L. M. Paz, P. Pinies, J. A. Castellanos, and G. Grisetti,
[38] C. E. Garcia, M. R. Camana, and I. Koo, ‘‘Machine learning-
‘‘Multi-robot SLAM using condensed measurements,’’ in Proc.
based scheme for multi-class fault detection in turbine engine disks,’’
IEEE/RSJ Int. Conf. Intell. Robots Syst., 2013, pp. 1069–1076.
ICT Exp., vol. 7, no. 1, pp. 15–22, Mar. 2021. [Online]. Available:
[65] W. Gruszczyk and H. Kwasnicka, ‘‘Coalition formation in multi-agent
https://www.sciencedirect.com/science/article/pii/S2405959521000096
systems—An evolutionary approach,’’ in Proc. Int. Multiconference
[39] M. Dorigo, G. Theraulaz, and V. Trianni, ‘‘Reflections on the future of Comput. Sci. Inf. Technol., Oct. 2008, pp. 125–130.
swarm robotics,’’ Sci. Robot., vol. 5, no. 49, Dec. 2020, Art. no. eabe4385. [66] J. Yang and Z. Luo, ‘‘Coalition formation mechanism in multi-agent
[40] J. R. Marden and J. S. Shamma, ‘‘Game theory and control,’’ Annu. Rev. systems based on genetic algorithms,’’ Appl. Soft Comput., vol. 7, no. 2,
Control, Robot., Auto. Syst., vol. 1, pp. 105–134, May 2018. pp. 561–568, Mar. 2007.
[41] J. Verbraeken, M. Wolting, J. Katzy, J. Kloppenburg, T. Verbelen, and [67] C. E. Garcia, M. R. Camana, and I. Koo, ‘‘Ensemble learning aided
J. S. Rellermeyer, ‘‘A survey on distributed machine learning,’’ ACM QPSO-based framework for secrecy energy efficiency in FD CR-NOMA
Comput. Surv., vol. 53, no. 2, pp. 1–33, 2020. systems,’’ IEEE Trans. Green Commun. Netw., vol. 7, no. 2, pp. 649–667,
[42] W. Qian, Y. Gao, and Y. Yang, ‘‘Global consensus of multiagent systems Sep. 2023.
with internal delays and communication delays,’’ IEEE Trans. Syst., Man, [68] R. J. Alattas, S. Patel, and T. M. Sobh, ‘‘Evolutionary modular robotics:
Cybern., Syst., vol. 49, no. 10, pp. 1961–1970, Oct. 2019. Survey and analysis,’’ J. Intell. Robotic Syst., vol. 95, nos. 3–4,
[43] F. L. Bellifemine, G. Caire, and D. Greenwood, Developing Multi-agent pp. 815–828, Sep. 2019.
Systems With JADE, vol. 7. Hoboken, NJ, USA: Wiley, 2007. [69] P. Kravets, ‘‘Game method for coalitions formation in multi-agent
[44] P. Taillandier, B. Gaudou, A. Grignard, Q.-N. Huynh, N. Marilleau, systems,’’ in Proc. IEEE 13th Int. Scientific Tech. Conf. Comput. Sci. Inf.
P. Caillou, D. Philippon, and A. Drogoul, ‘‘Building, composing and Technol. (CSIT), vol. 1, Sep. 2018, pp. 1–4.
experimenting complex spatial models with the GAMA platform,’’ [70] E. Lavendelis, A. Liekna, A. Nikitenko, A. Grabovskis, and
GeoInformatica, vol. 23, no. 2, pp. 299–322, Apr. 2019. J. Grundspenkis, ‘‘Multi-agent robotic system architecture for effective
[45] H. Hamidi and A. Kamankesh, ‘‘An approach to intelligent traffic task allocation and management,’’ in Proc. 11th WSEAS Int. Conf.
management system using a multi-agent system,’’ Int. J. Intell. Transp. Electron., Hardware, Wireless Optical Commun., Proc. 11th WSEAS
Syst. Res., vol. 16, no. 2, pp. 112–124, May 2018. Int. Conf. Signal Process., Robot. Automat., Proc. 4th WSEAS Int. Conf.
[46] L. Cruz-Piris, D. Rivera, S. Fernandez, and I. Marsa-Maestre, ‘‘Optimized Nanotechnol., Cambridge, U.K., 2012, pp. 167–174.
sensor network and multi-agent decision support for smart traffic light [71] L. Wang, Z. Wang, S. Hu, and L. Liu, ‘‘Ant colony optimization for
management,’’ Sensors, vol. 18, no. 2, p. 435, Feb. 2018. task allocation in multi-agent systems,’’ China Commun., vol. 10, no. 3,
[47] F. Derakhshan and S. Yousefi, ‘‘A review on the applications of multiagent pp. 125–132, Mar. 2013.
systems in wireless sensor networks,’’ Int. J. Distrib. Sensor Netw., [72] L. Kocarev, Consensus and Synchronization in Complex Networks.
vol. 15, no. 5, May 2019, Art. no. 155014771985076. Berlin, Germany: Springer, 2013.
[48] A. Barriuso, G. Villarrubia González, J. De Paz, Á. Lozano, and J. Bajo, [73] Y. Zheng, Y. Zhu, and L. Wang, ‘‘Consensus of heterogeneous multi-agent
‘‘Combination of multi-agent systems and wireless sensor networks for systems,’’ IET Control Theory Appl., vol. 5, no. 16, pp. 1881–1888, 2011.
[74] S. Das and J. M. Moura, ‘‘Distributed state estimation in multi-agent
the monitoring of cattle,’’ Sensors, vol. 18, no. 2, p. 108, Jan. 2018.
networks,’’ in Proc. IEEE Int. Conf. Acoust., Speech Signal Process.,
[49] A. V. Nazarova and M. Zhai, ‘‘The application of multi-agent robotic
2013, pp. 4246–4250.
systems for earthquake rescue,’’ in Robotics: Industry 4.0 Issues &
[75] C. Liu, Y. Wang, D. Zhou, and X. Shen, ‘‘Minimum-variance unbiased
New Intelligent Control Paradigms. Cham, Switzerland: Springer, 2020,
unknown input and state estimation for multi-agent systems by distributed
pp. 133–146.
cooperative filters,’’ IEEE Access, vol. 6, pp. 18128–18141, 2018.
[50] A. Sharma, ‘‘Multi-agent system applications to fight COVID-19 [76] A. Shariati and M. Tavakoli, ‘‘A descriptor approach to robust leader-
pandemic,’’ Apollo Med., vol. 17, no. 5, p. 41, 2020. following output consensus of uncertain multi-agent systems with delay,’’
[51] P. A. Cisse, J. M. Dembele, M. Lo, and C. Cambier, ‘‘Multi-agent systems IEEE Trans. Autom. Control, vol. 62, no. 10, pp. 5310–5317, Oct. 2017.
for epidemiology: Example of an agent-based simulation platform for [77] S. I. Han, ‘‘Prescribed consensus and formation error constrained finite-
schistosomiasis,’’ in Agents and Multi-Agent Systems for Health Care. time sliding mode control for multi-agent mobile robot systems,’’ IET
Cham, Switzerland: Springer, 2017, pp. 131–153. Control Theory Appl., vol. 12, no. 2, pp. 282–290, Jan. 2018.
[52] B. M. Castro, Y. de Abreu de Melo, N. F. dos Santos, A. L. da Costa [78] D. N. Serpanos and A. Bouloutas, ‘‘Centralized versus distributed
Barcellos, R. Choren, and R. M. Salles, ‘‘Multi-agent simulation model multimedia servers,’’ IEEE Trans. Circuits Syst. Video Technol., vol. 10,
for the evaluation of COVID-19 transmission,’’ Comput. Biol. Med., no. 8, pp. 1438–1449, Nov. 2000.
vol. 136, Sep. 2021, Art. no. 104645. [79] T. Van Gog, F. Paas, W. Savenye, R. Robinson, M. Niemczyk, R.
[53] E. Estrada, The Structure of Complex Networks: Theory and Applications. Atkinson, T. E. Johnson, D. L. O’Connor, R. M. Rikers, and P. Ayres,
London, U.K.: Oxford Univ. Press, 2012. ‘‘Data collection and analysis,’’ in Handbook of Research on Educational
[54] S. H. Strogatz, ‘‘Exploring complex networks,’’ Nature, vol. 410, Communications and Technology Third Edition. Taylor and Francis,
no. 6825, pp. 268–276, Mar. 2001. 2008, pp. 763–806, doi: 10.4324/9780203880869-61.
[55] J. Liu, Y. Pan, M. Li, Z. Chen, L. Tang, C. Lu, and J. Wang, ‘‘Applications [80] M. D. LeCompte, ‘‘Analyzing qualitative data,’’ Theory Pract., vol. 39,
of deep learning to MRI images: A survey,’’ Big Data Mining Analytics, no. 3, pp. 146–154, 2000.
vol. 1, no. 1, pp. 1–18, Mar. 2018. [81] S. Brandt and S. Brandt, Data Analysis. Cham, Switzerland: Springer,
[56] M. W. Khan, J. Wang, M. Ma, L. Xiong, P. Li, and F. Wu, ‘‘Optimal energy 1998.
management and control aspects of distributed microgrid using multi- [82] N. B. Harmancioglu, O. Fistikoglu, S. D. Ozkul, V. P. Singh, and
agent systems,’’ Sustain. Cities Soc., vol. 44, pp. 855–870, Jan. 2019. M. N. Alpaslan, ‘‘Definition of monitoring objectives,’’ in Water Science
[57] L. A. Amaral and J. M. Ottino, ‘‘Complex networks,’’ Eur. Phys. J. B, and Technology Library. Dordrecht, The Netherlands: Springer, 1999,
vol. 38, no. 2, pp. 147–162, 2004. pp. 95–110, doi: 10.1007/978-94-015-9155-3_5.

VOLUME 12, 2024 80973

 D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

[83] K. Fatema, V. C. Emeakaroha, P. D. Healy, J. P. Morrison, and T. Lynn, ‘‘A [103] R. Limosani, R. Esposito, A. Manzi, G. Teti, F. Cavallo, and P. Dario,
survey of cloud monitoring tools: Taxonomy, capabilities and objectives,’’ ‘‘Robotic delivery service in combined outdoor–indoor environments:
J. Parallel Distrib. Comput., vol. 74, no. 10, pp. 2918–2933, Oct. 2014. Technical analysis and user evaluation,’’ Robot. Auto. Syst., vol. 103,
[84] A. Arning, R. Agrawal, and P. Raghavan, ‘‘A linear method for deviation pp. 56–67, May 2018.
detection in large databases,’’ in Proc. KDD, 1996, vol. 1141, no. 50, [104] M. W. Khan and J. Wang, ‘‘The research on multi-agent system for
pp. 972–981. microgrid control and optimization,’’ Renew. Sustain. Energy Rev.,
[85] E. Toth and S. Chawla, ‘‘Group deviation detection methods: A survey,’’ vol. 80, pp. 1399–1411, Dec. 2017.
ACM Comput. Surveys, vol. 51, no. 4, pp. 1–38, Jul. 2019. [105] A. Farinelli, N. Boscolo, E. Zanotto, and E. Pagello, ‘‘Advanced
[86] S. P. Kavulya, K. Joshi, F. D. Giandomenico, and P. Narasimhan, approaches for multi-robot coordination in logistic scenarios,’’ Robot.
‘‘Failure diagnosis of complex systems,’’ in Resilience Assessment and Auto. Syst., vol. 90, pp. 34–44, Apr. 2017.
Evaluation of Computing Systems. Berlin, Germany: Springer-Verlag, [106] B. Arbanas, A. Ivanovic, M. Car, T. Haus, M. Orsag, T. Petrovic,
2012, pp. 239–261, doi: 10.1007/978-3-642-29032-9_12. and S. Bogdan, ‘‘Aerial-ground robotic system for autonomous delivery
[87] C.-H. Lin, H.-J. Chuang, C.-S. Chen, C.-S. Li, and C.-Y. Ho, ‘‘Fault tasks,’’ in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), May 2016,
detection, isolation and restoration using a multiagent-based distribution pp. 5463–5468.
automation system,’’ in Proc. 4th IEEE Conf. Ind. Electron. Appl., 2009, [107] B. Bona, L. Carlone, M. Indri, and S. Rosa, ‘‘Supervision and monitoring
pp. 2528–2533. of logistic spaces by a cooperative robot team: Methodologies, problems,
[88] T. J. Leung, J. H. Rife, P. Seiler, and R. Venkataraman, ‘‘Comparison of and solutions,’’ Intell. Service Robot., vol. 7, no. 4, pp. 185–202,
fault-tree models for fault detection, isolation, and recovery algorithms,’’ Oct. 2014.
J. Aerosp. Inf. Syst., vol. 14, no. 9, pp. 517–522, Sep. 2017. [108] G. Prennushi, G. Rubio, and K. Subbarao, ‘‘Monitoring and evaluation,’’
[89] M. Tipaldi and B. Bruenjes, ‘‘Survey on fault detection, isolation, and in A Sourcebook for Poverty Reduction Strategies. Washington, DC, USA:
recovery strategies in the space domain,’’ J. Aerosp. Inf. Syst., vol. 12, World Bank, 2002, pp. 30–107.
no. 2, pp. 235–256, Feb. 2015. [109] H. Farivarnejad and S. Berman, ‘‘Multirobot control strategies for
[90] G. W. Clark, M. V. Doran, and T. R. Andel, ‘‘Cybersecurity issues collective transport,’’ Annu. Rev. Control, Robot., Auto. Syst., vol. 5, no. 1,
in robotics,’’ in Proc. IEEE Conf. Cognit. Comput. Aspects Situation pp. 205–219, May 2022, doi: 10.1146/annurev-control-042920-095844.
Manage. (CogSIMA), Mar. 2017, pp. 1–5. [110] F. M. Gonzalez-Longatt and J. L. Rueda, PowerFactory Applications for
[91] Y. Zhang, M. E. Raoufat, and K. Tomsovic, ‘‘Remedial action schemes Power System Analysis (Power Systems). Springer, 2014.
and defense systems,’’ in Smart Grid Handbook. Wiley, 2016, pp. 1–10, [111] E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, ‘‘Industrial
doi: 10.1002/9781118755471.sgd032. Internet of Things: Challenges, opportunities, and directions,’’ IEEE
[92] J. Wen, P. Arons, and W. E. Liu, ‘‘The role of remedial action schemes in Trans. Ind. Informat., vol. 14, no. 11, pp. 4724–4734, Nov. 2018.
renewable generation integrations,’’ in Proc. Innov. Smart Grid Technol. [112] M. Wollschlaeger, T. Sauter, and J. Jasperneite, ‘‘The future of industrial
(ISGT), 2010, pp. 1–6. communication: Automation networks in the era of the Internet of Things
[93] J. A. Torres, P. J. Cruz, R. Vizuete, and R. Fierro, ‘‘On resilience and and industry 4.0,’’ IEEE Ind. Electron. Mag., vol. 11, no. 1, pp. 17–27,
heterogeneity in robotic networks,’’ in Cooperative Localization and Mar. 2017.
Navigation. Boca Raton, FL, USA: CRC Press, 2019, pp. 141–158. [113] R.-A. Luchian, S. Rosioru, I. Stamatescu, I. Fagarasan, and
[94] J. Lin, K. Dzeparoska, S. Q. Zhang, A. Leon-Garcia, and N. Papernot, G. Stamatescu, ‘‘Enabling industrial motion control through IIoT
‘‘On the robustness of cooperative multi-agent reinforcement learning,’’ multi-agent communication,’’ in Proc. IECON 47th Annu. Conf. IEEE
in Proc. IEEE Secur. Privacy Workshops (SPW), 2020, pp. 62–68. Ind. Electron. Soc., Oct. 2021, pp. 1–6.
[95] Y. Wang and H. Chen, ‘‘Blockchain: A potential technology to [114] V. Lesi, Z. Jakovljevic, and M. Pajic, ‘‘Reliable industrial IoT-based
improve the performance of collaborative emergency management distributed automation,’’ in Proc. Int. Conf. Internet Things Design
with multi-agent participation,’’ Int. J. Disaster Risk Reduction, Implement.. New York, NY, USA: Association for Computing Machinery,
vol. 72, Apr. 2022, Art. no. 102867. [Online]. Available: Apr. 2019, pp. 94–105, doi: 10.1145/3302505.3310072.
https://www.sciencedirect.com/science/article/pii/S2212420922000863 [115] R. Moghadam and H. Modares, ‘‘Resilient adaptive optimal control
[96] A. Khakimov, I. A. Elgendy, A. Muthanna, E. Mokrov, K. Samouylov,
of distributed multi-agent systems using reinforcement learning,’’ IET
Y. Maleh, and A. A. A. El-Latif, ‘‘Flexible architecture for deployment
Control Theory Appl., vol. 12, no. 16, pp. 2165–2174, Nov. 2018.
of edge computing applications,’’ Simul. Model. Pract. Theory, vol. 114,
Jan. 2022, Art. no. 102402.
[97] C. Wang, P. Ju, F. Wu, X. Pan, and Z. Wang, ‘‘A systematic review
on power system resilience from the perspective of generation,
network, and load,’’ Renew. Sustain. Energy Rev., vol. 167,
Oct. 2022, Art. no. 112567. [Online]. Available: https://www. DIEGO MALDONADO received the engineering
sciencedirect.com/science/article/pii/S1364032122004658 degree in electronics and control and the master’s
[98] J. Los, F. Schulte, M. T. J. Spaan, and R. R. Negenborn, ‘‘An auction-
degree in smart grids from the Escuela Politecnica
based multi-agent system for the pickup and delivery problem with
Nacional, Ecuador, in 2014 and 2020, respectively,
autonomous vehicles and alternative locations,’’ in Dynamics in Logistics,
M. Freitag, A. Kinra, H. Kotzab, and N. Megow, Eds. Cham, Switzerland:
where he is currently pursuing the Ph.D. degree
Springer, 2022, pp. 244–260. (researcher) in the field of distributed control. He is
[99] J. J. Damanik, H. Kasan, and H.-L. Choi, ‘‘Solving delivery assignment an Assistant Professor with the Departamento de
in hybrid-transit network using multi-agent reinforcement learning,’’ in Automatizacion y Control, where he has taught
Robot Intelligence Technology and Applications, vol. 6, J. Kim, B. Englot, electronics and automation engineering degree.
H.-W. Park, H.-L. Choi, H. Myung, J. Kim, and J.-H. Kim, Eds. Cham, His research interests include automatic systems
Switzerland: Springer, 2022, pp. 485–497. and distributed control applied to power systems.
[100] J. P. Queralta, J. Taipalmaa, B. C. Pullinen, V. K. Sarker,
T. N. Gia, H. Tenhunen, M. Gabbouj, J. Raitoharju, and T. Westerlund,
‘‘Collaborative multi-robot search and rescue: Planning, coordination,
perception, and active vision,’’ IEEE Access, vol. 8, pp. 191617–191643,
2020.
[101] S. M. Ghavami, J. Maleki, and T. Arentze, ‘‘A multi-agent assisted EDISON CRUZ received the engineering degree in
approach for spatial group decision support systems: A case electronics and control from the Escuela Politec-
study of disaster management practice,’’ Int. J. Disaster Risk nica Nacional, in 2020. He is currently pursuing
Reduction, vol. 38, Aug. 2019, Art. no. 101223. [Online]. Available: the Ph.D. degree with the School of Electrical,
https://www.sciencedirect.com/science/article/pii/S2212420919303218 Computer, and Biomedical Engineering, Southern
[102] N. Firdausiyah, E. Taniguchi, and A. G. Qureshi, ‘‘Modeling city Illinois University. His research interests include
logistics using adaptive dynamic programming based multi-agent multi-agent systems, optimal control, and mag-
simulation,’’ Transp. Res. E, Logistics Transp. Rev., vol. 125, netic field manipulation using permanent magnets.
pp. 74–96, May 2019. [Online]. Available: https://www.sciencedirect.
com/science/article/pii/S1366554518309049

80974 VOLUME 12, 2024

D. Maldonado et al.: MAS: A Survey About Its Components, Framework and Workflow

JACKELINE ABAD TORRES received the B.S. SILVANA DEL PILAR GAMBOA BENITEZ
degree from the Escuela Politecnica Nacional, received the degree in electronics and control
Quito, Ecuador, in 2008, and the M.S. and Ph.D. engineering and the master’s degree in control
degrees in electrical engineering from Washington systems from the Escuela Politecnica Nacional,
State University, in 2012 and 2014, respectively. Ecuador, in 2004 and 2008, respectively, and the
She held a Fulbright Grant, in 2010. She is Ph.D. degree in electrical engineering from the
currently an Associate Professor with the Depar- Electrical Energy Institute, Universidad Nacional
tamento de Automatizacion y Control Industrial, de San Juan, Argentina, in 2018. She is cur-
Facultad de Ingenieria Electrica y Electrnica, rently a Professor with the Faculty of Electrical
Escuela Politecnica Nacional. Her current research and Electronic Engineering, Escuela Politecnica
interests include structural analysis and controller design of dynamical Nacional. She is an Administrator of the Industrial Networks and the SCADA
networks with applications to sensor/vehicle networking, epidemic control, Laboratory. Her research interests include monitoring, protection, and
and power systems network control. control systems, SCADA, communications networks, and their application
in industrial processes and in power systems.

PATRICIO J. CRUZ (Member, IEEE) received the

B.S. degree in electronic and control engineering
from the Escuela Politecnica Nacional (EPN),
Quito, Ecuador, in 2005, and the M.Sc. degree
in electrical engineering and the Ph.D. degree in
engineering from The University of New Mexico
(UNM), Albuquerque, NM, USA, in 2012 and
2016, respectively. During his graduate studies,
he was a Research Assistant with the Multi-
Agent, Robotics, and Heterogeneous Systems
(MARHES) Laboratory, UNM. From 2019 to 2022, he was a Coordinator of
the Master Programs with the School of Electrical Engineering, EPN. He is
currently an Associate Professor with the Departamento de Automatizacion
y Control Industrial (DACI), EPN. He has been the Director and a
Collaborator of different research projects in areas, such as multi-agent
system coordination, control of unmanned aerial systems (UAVs), and the
development of platforms for rehabilitation robotics. He is the author or
coauthor of more than 25 papers in international and regional conferences
and journals. As part of his work, he published two book chapters focused
on cooperative control of robotic heterogeneous systems.

VOLUME 12, 2024 80975