Agentification of the Internet of Things: A systematic literature review

Abstract

The Internet of Things is aimed at controlling the physical real world using a global network of heterogeneous smart objects that are interconnected through the Internet. Nonetheless, this emergent paradigm still poses challenges in terms of intelligence, autonomy, security, privacy, and interoperability, all of which must be overcome in order for this approach to be consolidated. The integration of Internet of Things with agent technologies has been proposed as an alternative. In this article, we present a general overview of the main preliminary models, theoretical and experimental studies, and applications currently proposed to carry out the agentification of the Internet of Things through modeling smart objects and networks of Internet of Things objects using software agents. Furthermore, we present an analysis of the main strengths, opportunities, weaknesses, and threats of this approach in terms of computational intelligence. Finally, we discuss the main challenges that must be overcome in order for the agentification process of the Internet of Things to become a consolidated paradigm to build cognitive Internet of Things applications.

Introduction

The Internet of Things (IoT) is an emergent paradigm whose objective is to control the physical real world using a global network of objects that are interconnected through the Internet.^1–3 As a result of this approach, there is currently a wide range of “intelligent” objects on the market, with the goal of covering a broad spectrum of real applications in different domains of knowledge such as health care, ⁴ comfort management in smart homes, ⁵ management of public services in smart cities, ⁶ industrial process management, ⁷ among others. However, the high supply of devices of this nature has led to the proliferation of objects (e.g. smart appliances, IP video cameras, sensors, actuators, technological items of clothing, robots, smart phones, and micro-computers) that adopt different standards of wireless communication (e.g. Bluetooth, ZigBee, 3G, 4G, GPRS (General Packet Radio Service), Wi-Fi, and RFID (radio-frequency identification)) ⁸ and many of which are incompatible when integration is attempted.

In general, the majority of the objects that are compatible with the IoT interfaces currently on the market consist of black boxes that limit the embedding of new programming code in the physical device itself. This makes it difficult to include specialized intelligence mechanisms to cover specific needs. Due to this limitation, a new approach based on the integration process of the IoT ¹ and agents, ⁹ also known as the IoT agentification process, has arisen in the last decade. ¹⁰ Its objective is principally focused on optimizing the use of resources associated with IoT objects via the execution of intelligent, adaptive, and collaborative processes.^11,12 In this respect, software agents—independent of whether reactive, deliberative, ¹³ or hybrid ¹⁴ architecture is used—, as entities, are capable of executing autonomous tasks according to specific objectives, ¹⁵ and possessing innate characteristics such as autonomy, sociability, reactivity, and proactivity, and others such as mobility, adaptation, and collaboration, ¹⁶ are placed as potential units for modeling reconfigurable, autonomous, and intelligent IoT ecosystems. ¹⁰

This article presents a systematic review of the literature centered around the identification of the efforts made by the scientific community to carry out the modeling of IoT agentification. Therefore, the objective of the systematic review proposed in this article is focused on identifying the main approaches, experimental and non-experimental models, frameworks, real applications and simulations, as well as studies and reviews of the literature in which the integration process of the IoT and agents has been addressed. In this way, a general context of this emerging approach is provided; identifying its main strengths and weaknesses, as well as the threats and opportunities that this approach may have regarding the development of future intelligent systems based on IoT technologies.

In addition to the analysis of the approaches that arise from the fusion of the IoT and agents, this study performs an analysis of the impact—positive or negative—that networks of IoT objects may experience if they introduce software agents as part of their architecture to manage intelligent behaviors (e.g. the optimization of resources associated with IoT objects, the provision of resources in a precise manner, management of heterogeneous resources, and execution of collaborative tasks). Finally, the study concludes by determining the main success factors of the IoT agentification process to date, as well as the main challenges that must be addressed by the scientific community in order to face the current limitations associated with this approach.

This article is organized into five sections. The “Systematic review” section describes the process applied and the results obtained at each stage of the systematic review. In the “Results” section, we present an analysis of the results obtained in the previous section. This analysis is carried out in terms of the models that have been proposed to create networks of intelligent IoT objects. Section “IoT agentification from a practice view” describes the main practical cases in which the IoT and agents have been integrated, and also analyzes the main intelligence mechanisms used in each of these. The “Analysis of the IoT agentification” section describes the main strengths, opportunities, weaknesses, and threats associated with the IoT agentification process, as well as the main challenges and future lines of research that arise as a result. Finally, in section “Conclusion,” we describe the conclusions of the study in detail.

Systematic review

A systematic review of the literature consists of identifying, evaluating, and interpreting the most relevant studies of a topic, research question (RQ), or phenomenon of a different type. ¹⁷ Therefore, the general objective of a process of this nature is to provide evidence that responds to a set of RQs from primary studies in which the subject of study has previously been addressed. ¹⁸

To carry out the systematic review process proposed in this article, we followed the methodological guidelines proposed by Kitchenham and Charters ¹⁷ for the development of literature reviews in the field of Software Engineering. Nevertheless, due to the fact that the process detailed is quite general, its application in this study has been considered feasible after contrasting it with another methodological tool oriented toward the development of studies of the same nature in the area of Computer Sciences proposed by Kofod-Petersen. ¹⁸ The application of this process is detailed in the sections below.

Definition of RQs

In view of the recent interest of the scientific community in the integration process of IoT technologies with agents to achieve an intelligence/cognition level in objects and networks of IoT objects superior to that reached by the generic IoT, ¹⁰ this systematic review aims to answer seven RQs considered as relevant by the authors, which have not been answered by other similar studies. These questions have served as the basic guide for the revision process addressed in this study and are detailed as follows:

1. RQ₁: What is the existing trend regarding the use of agents to model systems based on IoT technologies?

2. RQ₂: What are the main approaches proposed to address the modeling of cognitive networks based on IoT technologies?

3. RQ₃: What are the main proposals in terms of models and technological tools that the scientific community has formulated in order to integrate the IoT and software agents?

4. RQ₄: Do IoT applications that are developed based on agents and implemented in scenarios to solve real problems exist?

5. RQ₅: What aspects, additional to the cognitive ones, have been experienced with the development of applications modeled through the IoT agentification process?

6. RQ₆: What are the main intelligence techniques used in IoT cognitive applications modeled using software agents?

7. RQ₇: What are the main strengths, opportunities, weaknesses, and threats (SWOT) of the agentification process of intelligent networks of IoT objects?

Selection of information sources

The previously defined RQs provided the guidelines to determine a set of terms that have allowed us to carry out the search process. The primary terms are constituted by the words (agent, software agent, multiagent), (internet of things), and (intelligent, smart). Based on these, we strategically defined four search strings (SS) that follow the structure described below:

1. SS₁ : (“agent” OR “software agent” OR “multiagent”) AND (“internet of things”);

2. SS₂ : (“agent” AND “integration” AND “internet of things”);

3. SS₃ : (“internet of things” AND “based” AND “agent”);

4. SS₄: (“smart” OR “intelligent”) AND (“internet of things”).

Once the SS were defined, we proceeded to determine the information sources used to execute the search equations, and thus retrieve the most relevant studies in which the process of integration between agents and IoT has been addressed and published in high-impact scientific publications. Based on the methodological frameworks proposed by Kitchenham and Charters ¹⁷ and Kofod-Petersen ¹⁸ for the development of literature reviews, a total of five primary information sources (PIS) were considered. Among these are three specialized digital libraries in the area of Computer Science (PIS₁ to PIS₃) and two documentary databases (PIS₁ and PIS₅). The last two sources have been included with the objective of extending the search range and thus retrieving other types of documents such as registered patents:

1. PIS₁: Institute of Electrical and Electronics Engineers (IEEE) Digital Library, available electronically at: http://ieeexplore.ieee.org/

2. PIS₂: Association for Computing Machinery (ACM) Digital Library, available electronically at: https://dl.acm.org/

3. PIS₃: Science Direct, available electronically at: http://www.sciencedirect.com/

4. PIS₄: Institute for Scientific Information (ISI) Web of Science Database, available electronically at: https://webofknowledge.com/

5. PIS₅: Scopus of Elsevier Database, available electronically at: https://www.scopus.com/

Selection of studies

The search process carried out in October 2017 was performed on each of the search engines specified above and applied to the titles of the publications. In this way, we discarded tens of studies classified in the research line related to both agent- and IoT-oriented technologies that did not focus on the process of agentification of IoT. Most of them were focused on simulating the behavior of IoT scenarios and therefore, they had no relevance whatsoever to this investigation. We were basically focused on studies where the agentification process was carried out to model aspects of cognition in IoT. Nonetheless, aspects addressed by the IoT through software agents, such as sociability, autonomy, proactiveness, mobility, and adaptation, are valid for this study as long as they contribute to the execution of intelligent services by the objects connected to IoT infrastructures.

Overall, the exploration and discovery process provided us with a total of 239 studies, of which 92 were not repeated—data obtained without grouping by SS or information source. These 92 studies were reduced to 69 after analyzing the title in detail and discarding a further 23 studies. Finally, after analyzing the summary of each article, 13 proposals were identified as not constituting relevant works in the area of study. Consequently, 56 studies remained to be analyzed comprehensively. Of these 56 studies, which were initially considered to be fundamental, 6 were considered to have little relevance after being read comprehensively, on the grounds that they were not aligned with the subject of study. Among the main reasons for the exclusion of these 6 articles are as follows: full text unavailable (one manuscript), English version not available (two written in Mandarin Chinese and one in Russian), and not related to the specific topic addressed in this study (the remaining two). Therefore, for the analysis of results in this systematic review, a total of 50 proposals were considered as primary sources. The details of the aforementioned studies are shown in Table 1 and the results of the selection process in Table 2.

Table 1.

Main primary information sources analyzed.

No.	Source	R. String	Year	T. Study	Reference
E1	Scopus	4	2015	Conference	Pizurica and Vandaele 19
E2	IEEE	1	2013	Conference	Elkhodr et al. 20
E3	Science Direct	1	2016	Journal	Schatten et al. 21
E4	ACM	4	2016	Conference	Thangaraj et al. 22
E5	Scopus	3	2017	Conference	Savaglio et al. 10
E6	Scopus	4	2015	Conference	Dyk et al. 23
E7	IEEE	1	2017	Journal	García-Magariño et al. 24
E8	IEEE	1	2013	Conference	Mzahm et al. 11
E9	IEEE	1	2015	Conference	Neisse et al. 25
E10	Science Direct	1	2013	Journal	Wang et al. 26
E11	IEEE	1	2014	Conference	Leppänen et al. 27
E12	Web of Science	1	2017	Journal	Choi and Koh 28
E13	Science Direct	1	2013	Conference	Do Nascimento and de Lucena 29
E14	Scopus	4	2014	Journal	Lin et al. 30
E15	IEEE	1	2013	Conference	Yu et al. 31
E16	Scopus	4	2015	Conference	Bosse 32
E17	IEEE	1	2014	Conference	Manate et al. 33
E18	IEEE	4	2012	Conference	Zhou and Lou 34
E19	IEEE	4	2016	Journal	Santos et al. 35
E20	Web of Science	2	2015	Conference	Al-Sakran 36
E21	Scopus	4	2016	Conference	Knol et al. 37
E22	IEEE	1	2015	Conference	Alexakos and Kalogeras 38
E23	IEEE	1	2017	Conference	Adhikaree et al. 39
E24	IEEE	1	2016	Conference	Carlier and Renault 40
E25	IEEE	4	2014	Conference	Pathak et al. 41
E26	ACM	1	2015	Journal	Alan et al. 42
E27	Web of Science	3	2013	Patent	Bhaksaran et al. 43
E28	IEEE	1	2013	Conference	Leppänen et al. 44
E29	IEEE	1	2013	Conference	Jarvenpaa et al. 45
E30	Scopus	4	2014	Conference	Leppänen et al. 46
E31	IEEE	1	2016	Conference	Bosse 32
E32	IEEE	1	2017	Journal	Fortino et al. 12
E33	Scopus	3	2016	Journal	Laghari and Niazi 47
E34	Web of Science	3	2017	Conference	Batool and Niazi 48
E35	IEEE	1	2017	Conference	Forestiero 49
E36	IEEE	1	2014	Conference	Leong and Lu 50
E37	Web of Science	1	2017	Journal	Kato et al. 51
E38	IEEE	1	2013	Conference	Godfrey et al. 52
E39	IEEE	1	2017	Journal	Gateau and Rykowski 53
E40	IEEE	1	2015	Conference	López et al. 54
E41	Web of Science	1	2011	Conference	Copigneaux 55
E42	IEEE	1	2014	Conference	Moin 56
E43	Scopus	1	2015	Conference	Fortino et al. 57
E44	Scopus	1	2016	Conference	Chamoso et al. 58
E45	Scopus	3	2015	Conference	Alan et al. 59
E46	ACM	1	2016	Conference	Calvaresi et al. 60
E47	ACM	1	2017	Conference	Singh and Chopra 61
E48	IEEE	1	2017	Conference	Ayala et al. 62
E49	Scopus	1	2015	Journal	Fortino and Russo 63
E50	Scopus	1	2014	Conference	Mzahm 64

Table 2.

Search results based on the strings and defined information sources.

Search strings	SS 1	SS 2	SS 3	SS 4	SS 5	Σ Studies	Repeat studies	Unique studies
SearchString1 (SS1)	24	22	5	50	44	145	58	87
SearchString2 (SS2)	2	0	0	3	2	7	4	3
SearchString3 (SS3)	6	1	3	21	20	51	23	28
SearchString4 (SS4)	10	5	1	10	10	36	8	28
Σ Studies	42	28	9	84	76	239
Repeat studies	11	9	4	34	33
Unique studies	31	19	5	50	43

Results

In this section, the contents of the primary studies listed in Table 1 are analyzed and discussed with the aim of extracting relevant information regarding the trends in the agent-based IoT approach in the last decade, as well as its impact on the solution of real problems where this approach has been applied. The results, detailed in Figure 1, show that the number of publications related to the IoT agentification process is growing. This emergent approach first appeared in 2011; however, its greatest impact has occurred in the last 5 years when 91% of the studies considered as sources of primary data by this study were published.

Figure 1.

Trend of the agentification of the IoT according to current scientific publications.

Of the studies published between 2011 and 2017, as detailed in Figure 2(b), the largest number of these revolve around models—experimental and non-experimental—and applications implemented in real environments (92%): the same studies that have been published mostly in journals and scientific conference proceedings (98%) as shown in Figure 2(a). There are insufficient literature reviews in this area (8%) that analyze the different proposals made to date. Of the majority of studies that revolve around the proposed models, 36% correspond to models in which an experimentation process has been carried out through the presentation of quantitative results, while another 36% of the theoretical models are only proposals where ideas, architectures, and frameworks are presented, including innovative ideas to carry out the integration of IoT and agents. In addition, one can appreciate that the developed models are beginning to be adopted in the solution of real problems (20%) that have led to the development of patents. This leads us to conclude that the IoT agentification process is moving forward slowly and there is a high probability that it will grow stronger over the next few years due to the progressive growth of the use of IoT devices and the need for them to operate autonomously, intelligently, and in an easily customizable way so that they are a more effective means for the development of work activities and daily life. ²¹

Figure 2.

(a) Nature of the studies and (b) types of proposals that have addressed the process of agentification of the IoT.

The experimental and non-experimental models published, which propose ideas regarding the integration process of the IoT and agents, demonstrate researchers’ growing interest in creating cognitive networks of IoT objects. Based on these models, it has been possible to identify two approaches to modeling the IoT based on agents. The first of these corresponds to the approach aimed at modeling and designing IoT objects by embedding an intelligence and reasoning mechanism in the object itself, on the basis of which it is feasible for the object to directly perform actions cognitively in order to optimize the use of its associated resources. ¹¹ However, there is a second approach that consists of modeling multiagent systems (MASs) to manage generic object networks compatible with IoT. Thus, these types of networks acquire social abilities as well as autonomy and intelligence that jointly enable objects to indirectly carry out collaborative tasks between various objects. ⁶⁵ These two approaches, together with their respective application in the solution of real problems, are analyzed with a greater level of details below.

Modeling of IoT objects based on agents

This approach is aimed at algorithms that define intelligent behavior at the level of the hardware used by the IoT; ⁴⁰ that is to say, to embed a software agent in the device itself so that it acquires the innate abilities inherent to the agents. Thus, the object can incorporate more dynamic, flexible, and cognitive behavior than those of generic IoT objects that are currently marketed. Next, Table 3 summarizes the main approaches proposed in this area, which include details such as the approach or proposed processing unit, the model of agent used for its execution, IoT technologies compatible with the proposed model, and scope (i.e. intelligence, sociability, autonomy, proactiveness, interoperability, mobility, adaptability, modular development of components, security and privacy, self-adaptation, and self-organization) of the applications created by means of the use of each of the proposals.

Table 3.

Main proposals focused on the modeling of IoT objects using agents.

No.	Approaches of agentification	Model of agent	Technology of IoT	Scope	References
1	Agent of Things	Theoretical	Theoretical	I, S, A	Mzahm et al.11,64
2	IoT-a	Triskell3 S	Raspberry Pi	S	Carlier and Renault 40
3	Iota	Theoretical	Theoretical	I, S	Singh and Chopra 61
4	Smart Object	JADE, Jadex	OMNeT++	I, S, A, P	Fortino et al.;12,57Fortino and Russo 63
5	Knol et al.	JADE	BeagleBone board	S	Knol et al. 37
6	Internet of Agents	Theoretical	Theoretical	T, I	Yu et al. 31
7	Manate et al.	Akka toolkit	General	T	Manate et al. 33
8	Sol	Self-StarMAS	Android devices,Sun SPOT mote	AA, O, I	Ayala et al. 62
9	NFC-agent	Reactive	NFC smartphones	S, P, R, I	Lin et al. 30
10	AIoT device	AIoT agent	LEGO MindstormsNXT, Drones	I, A, AO	Kato et al. 51
11	REST agent	REST agent	General	O	Leong and Lu 50
12	MAF	MAF agent	Non-defined	S	Bhaksaran et al. 43
13	REST agent	REST agent	Smartphone,ATmega 1284 P	I, O, M	Leppänen et al.27,44,46
14	HTML5 agent	HTML5 agent	Raspberry PI	M, T	Jarvenpaa et al. 45
15	AgentJS	Java Script agent	Smartphone	T, M, D, I	Bosse 66
16	PherCon	AgentSpace	Lego NXT	S, M, P	Godfrey et al. 52
17	DLDA agent	DLDA agent	Smartphone	S&P, I	Elkhodr et al. 20

I: intelligence; P: proactivity; M: mobility; S: sociability and collaboration; A: autonomy; O: interoperability; T: adaptability; D: modular; S&P: security and privacy; AA: self-adaptation; AO: self-organization.

Agent embedded in IoT objects

The first 10 items listed in Table 3 correspond to proposals focused on embedding a software agent in the architecture of the devices or objects compatible with the IoT. The process of embedding the agent in the architecture of the physical device can take two forms. The first of these consists of direct embedding and occurs when the object is of an open hardware type and supports the update of software that directs its behavior. This involves loading a firmware that can be dynamically updated. In this respect, one of the most popular initiatives is mangOH. ⁶⁷ However, the second way to embed agents in an IoT object is by means of alternative mechanisms; that is to say, by integrating a set of sensors/actuators, thus forming specific IoT objects by adding a SBC component (Single Board Computer). Hence, the microcontroller used in the first case is exchanged for a more powerful microprocessor with sufficient computing and storage power to execute the algorithms that an agent uses to control a set of physical components that can be integrated into its boards via analog and digital inputs. Some relevant projects in this vein are Raspberry PI and BeagleBoard, among others. ⁶⁸

Several proposals have been made in the attempt to define a standard basic processing unit for the approach associated with agent-based IoT modeling. Among the most relevant units are entities of the type agent of things, IoT-a, Iota, and AIoT, coined, respectively, by Mzahm et al.,^11,64 Carlier and Renault, ⁴⁰ Singh and Chopra, ⁶¹ and Kato et al. ⁵¹ It is important to note that, although these entities (processing units) have been set out under different names, they share the same objective—which is to integrate intelligence into objects to form networks of cognitive IoT objects. For this reason, each of the proposed units are being adopted by models, frameworks, and technological tools that seek to reach levels of intelligence and autonomy superior to those achieved by generic IoT applications. ⁶⁹

The studies of the previously specified authors propose, in a general way, modeling IoT objects by means of linking an individual agent. Nevertheless, although in a similar—and in some cases identical—way, each author defines this integration process from his or her own perspective. For example, Mzahm et al.^11,64 conceive this process as a device that gives objects the capacity to execute reasoning and communication processes with heterogeneous objects by themselves; thus providing a high level of interoperability and dynamic and flexible intelligence, personalized through software agents. Carlier and Renault ⁴⁰ and Singh and Chopra ⁶¹ meanwhile define it as a conceptual unitary machine that stores the control logic of the object that is focused on the execution of reasoning tasks in a similar way as posed by Mzahm et al. Along the same lines, Kato et al. ⁵¹ and Knol et al. ³⁷ propose the design of IoT objects using agents to improve their communication skills and establish cooperative processes with other devices ⁵¹ as well as extracting and storing data coherently, both in cyberspace and on the device. ³⁷

However, although the agent of things and Iota proposals supply the theoretical guidelines about how to carry out the agent-based object modeling process without defining the agent model to use, Kato et al. ⁵¹ recommends the use of rule-based agents IF-THEN to execute a certain action. Thus, objects can be configured in a similar way to how IFTTT (IF-This, Then-That) technologies operate. ⁷⁰ However, it is possible to define more complex cognitive processes where deliberative capacities of BDI (beliefs, desires, and intentions) type, processes of semantic reasoning based on ontologies, and processes of greater complexity based on artificial intelligence (AI) techniques (i.e. machine learning, artificial neural networks, genetic algorithms, Bayesian networks, and fuzzy logic) are applied. In these latter cases, the proposals are oriented toward more specific problems that have considered the type of problem and the resources associated with the objects, given that the use of a specific agent model (reactive, deliberative, or hybrid) and the intelligence technique(s) to adopt will be based on this.

Applying the same comparison as previously described, Fortino and colleagues^12,57,63 introduces the concept of smart object, which enables IoT objects to execute proactive, cooperative, and intelligent management of contextual awareness. This last characteristic is achieved thanks to the reactive capacity associated with the agents. In addition, to bring the proposed concept into being, the authors include mechanisms for objects to take advantage of Cloud Computing ⁷¹ in order to provide them with greater computing and storage capacity, ⁶³ which are limited resources in the majority of current IoT objects. Similarly, Lin et al. ³⁰ proposes the development of an intelligent agent model with the capacity to carry out communication through near-field communication (NFC), which combines personal and social profiles from which it recommends relevant services that support reactive action, proactive achievement, and social cooperation—relevant aspects to the approach called the Social Internet of Things (SIoT). ⁷²

Finally, Ayala et al. ⁶² propose embedding agents in heterogeneous IoT objects with the aim that these agents manage IoT resources in order to provide the appropriate services requested within IoT ecosystems. In summary, Ayala et al.’s proposal focuses on the use of Self-StarMAS type agents to support interoperability and heterogeneity between IoT networks, providing flexible communication between connected objects, just as Yu et al. ³¹ envision. In this analysis, the IoT is seen not only as a network of interconnected objects but also as a network of agents that manage these objects in a more flexible and intelligent way than is currently done. From there, arises an emergent paradigm called the Internet of Agents (IoA), which could integrate all the proposals analyzed in this study to achieve a level of consolidation and future standardization of proposals focused on the integration of IoT and agents. However, in order to fulfill the objective pursued by this paradigm, focused on the empowerment of agent networks to manage IoT resources, it is also essential to include mechanisms that facilitate the dynamic reconfiguration of the system by non-specialized users and communication between agents that are operating in heterogeneous platforms. All these are aspects that must still be addressed by scientists. Along these lines, Manate et al. ³³ proposes the use of a protocol that allows agents to exchange information from any geographical location, even when the infrastructure is exposed to a high degree of network traffic.

Web agents targeting IoT objects

Generally, IoT objects are electronic devices with limited computing capacity and storage resources to support the agents running on them. To face this limitation, several models of alternative agents based on web technologies have been proposed, in addition to the traditional models of reactive agents such as JADE ⁷³ and JADE Leap, ⁷⁴ and deliberative agents such as Jadex, ⁷⁵ which are currently the most widespread and popular platforms for creating agents.

The 11th proposal, detailed in Table 3, constitutes the main web agent model proposed to date that is developed for its use in IoT ecosystems. This model, proposed by Leong and Lu, ⁵⁰ is an agent based on REST technologies, which uses HTTP as a communication protocol as well as messages encoded in JSON (JavaScript Object Notation) language. This enables a better performance to carry out the process of message exchange between agents that run in a distributed manner. In this way, the agents provide a mechanism to support lightweight agents who do not need a central directory to find other objects that help them meet their objectives. Hence, these agents provide a high level of interoperability and ease of distributing intelligence, which the IoT can take advantage of in order to develop cognitive networks based on lightweight and interoperable software agents through the web. ⁷⁶ It is important to note that other proposals have also been developed based on the use of emerging web technologies, but that include a special component such as mobility. For this reason, the aforementioned agent models are analyzed in the following section where the process of object modeling is carried out via mobile agents.

Mobile agents embedded in objects

Following the same criteria of the need for lightweight components that contribute to achieving a higher level of autonomy, cognition, and flexibility of adaptation in the IoT, several proposals for models of lightweight agents with mobility support have been proposed. First, Leppänen et al.’s ^27,44,46 proposal is focused on a lightweight agent model based on REST technologies, aimed at heterogeneous IoT devices of low power resources and large-scale networks. These networks are displayed on the web and thus, the Web of Things^77,78 is positively benefited in terms of operationalization.

However, under the same criteria followed by Leppänen et al., Jarvenpaa et al.45 proposed a mobile agent model developed by means of the use of HTML5 web technology that allows agents to move in different IoT devices, even allowing their cloning in order to create more complex behaviors. These behaviors can be executed in two ways: through a user interface within a browser or from a server. The use of one alternative or another depends on the resources associated with the objects where the agent must be migrated. Along the same lines, Bosse66 proposes an additional mobile agent model called AgentJS, developed through JavaScript, which can feasibly be run on all host platforms including web browsers.

Finally, Godfrey et al. ⁵² propose an alternative mobile agent model, which not only focuses on enabling communication between objects connected to various IoT networks but also focuses, moreover, on enhancing the objects with capabilities that allow them to proactively search for the appropriate resources and provide specialized services to devices connected to a network of this nature. In conclusion, the lightweight mobile agents proposed in this section make a significant contribution to allowing IoT objects to integrate a level of expertise through components that do not necessarily have to be incorporated indefinitely in their architecture, but that contribute to the fulfillment of their objectives at certain times.

IoT modeling based on MASs

This second approach to carrying out the process of integration of IoT and agents focuses on the modeling of a MAS to manage the objects connected to IoT infrastructures and their associated resources. The management of communications between the agents that are part of the MAS with the physical devices that make up the network of IoT objects is possible thanks to the use of middlewares that are specialized in providing a low-level communication interface with the IoT objects. Some tools of this type detailed by Razzaque et al. ⁷⁹ are TinySOA, SOCRADES, ubiSOAP, ActorNet, UbiROAD, and openHAB. ⁸⁰ In the review process carried out in this study, the main proposals that have contributed to the origin of this specific approach are detailed in Table 4, which follows the same structure as that used by Table 3.

Table 4.

Main models of MASs modeled using agents.

No.	Approaches of agentification	Model of agent	Technology of IoT	Scope	References
1	MAS-based Machine Learning	JADE, Jade,Jade-Leap	PC and smartphones	I, AO, AA	Do Nascimentoand de Lucena 29
2	MAS-basedRecommendationSystem	General	General	I	Forestiero 49
3	LSMASs	General	General	O	Schatten et al. 21
4	Sensor Cloud	SeSAm	SoC microchip		Bosse 32
5	MAS-MINC	JADE	PC	B	Laghari and Niazi 47
6	DMA	CMA agent	General	M	Choi and Koh 28

I: intelligence; O: interoperability; M: mobility; T: adaptability; AO: self-organization; AA: auto-adaptation; B: scalability.

MASs modeling for IoT management, also known generically as agent-based modeling (ABM) of the IoT, is basically focused on the modeling of adaptation, social, and collaboration capacities. This is evidenced by the demonstration made by Kaminski et al. ⁸¹ and Savaglio et al., ¹⁰ who evidence in their respective studies and analyze the benefits of ABM as a technique recognized as an “effective and full support for the development of open, dynamic, cooperative, and decentralized IoT systems.”

The first two models summarized in Table 4 are agent systems based on sophisticated mechanisms. Both of these are approaches in which the agents integrated into the MAS are geared toward managing a network of IoT objects with more complex capabilities based on automatic learning techniques and AI optimization. The first of these is the approach of Do Nascimento and de Lucena ²⁹ who propose a model in which objects acquire the capacity for self-organization and self-adaptation by means of agents based on automatic learning techniques such as neural networks and evolutionary algorithms. In this vein, the second approach proposed by Forestiero ⁴⁹ proposes an agent model based on a multiagent recommendation algorithm—based on the mapping of similar objects—in the context of IoT, to provide a set of meaningful suggestions that meet the needs of a user with determined characteristics.

Some proposals have also been set out regarding the modeling of the social and cooperative component. Among the most relevant of these are the works of Ruan et al. ⁸² and Schatten et al. ²¹ (third item of Table 4). On the one hand, Ruan et al. propose a framework aimed at managing each agent’s level of confidence in order to help it assess the reliability of its counterparts. On the other hand, Schatten et al. propose a study which presents the guidelines to carry out the organizational design of large-scale MASs that have limitations to work and communicate in an interoperable way. In addition to these approaches, there are two additional studies regarding the management of the social component. The first of these is the study by Manate et al. ³³ who propose a hybrid multiagent platform where agents can be embedded in the devices themselves or in the cloud. In this way, agents can exchange information regardless of their location. Hence, the proposed platform performs a security audit on the distributed agent architecture to determine which geographic areas offer a better quality of services for end-users. The second paper, presented by Choi and Koh, ²⁸ included as the sixth item in Table 4, proposes a distributed mobility agent model (distributed mobility agent (DMA)) to manage the location of mobile sensors and to carry out transfer control operations with their distributed neighboring counterparts.

Finally, relative to the cognitive component, as detailed in the fifth item of Table 4, Laghari and Niazi ⁴⁷ explore the modeling of complex communication networks based on cognitive agents through the platform MAS-MINC (multiagent system for managing carbon footprint) that also considers the aspect related to cooperation. Against this backdrop, Singh and Chopra ⁶¹ propose the integration of MASs as a support mechanism for the distribution of intelligence supported by Cloud Computing, with the aim of providing flexible reasoning processes and optimizing the discovery and selection of IoT objects and associated resources. With a similar objective, as indicated in the fourth element of Table 4, Bosse ³² proposes a MAS based on mobile agents capable of operating heterogeneous wide-scale IoT networks allowing the development of clouds of interoperable sensors with their own reasoning capabilities.

In summary, MAS modeling can be addressed by combining it with the first approach related to object modeling based on agents as described above. Modeling objects as agents automatically enables the development of MAS for the management of IoT object networks and the resources associated with the infrastructure in a global way; that is to say, contemplating the cognition component, social capacity and collaboration, adaptation, and autonomy.

Frameworks and middlewares for supporting the agentification process

Both Tables 3 and 4 show the main models of agents and MASs that have been proposed to materialize the process of agentification of the IoT. Despite the existence of a variety of models of agents oriented to manage the objects and resources of IoT, it is evident that there are no standard frameworks that describe in an integral way the guidelines and conceptualization regarding the process of agentification nor the process to follow to build agentified IoT ecosystems. However, as the content of Table 5 shows, some frameworks to create applications directed by the IoT agentification process have been proposed.

Table 5.

Main frameworks for the agentification of the IoT.

No.	Framework	Concepts	Architecture	Processes	Pragmatics	Completeness	Implementation	References
1	SenseSim	Yes	No	Yes	No	Medium	Yes	Dyk et al. 23
2	EULA	Yes	Yes	Yes	Yes	Low	Yes	Neisse et al. 25
3	AR	Yes	Yes	Yes	Yes	Low	Yes	Leppänen et al. 27
4	FIoT	Yes	Yes	Yes	Yes	High	Yes	Do Nascimentoand de Lucena 29
5	NIA	Yes	Yes	Yes	Yes	High	Yes	Lin et al. 30
6	TSF	Yes	Yes	Yes	Yes	Low	Yes	Al-Sakran 36
7	Pathak et al.	Yes	Yes	Yes	Yes	Low	No	Pathak et al. 41
8	HTML5-Thing	Yes	Yes	Yes	Yes	Low	Yes	Jarvenpaa et al. 45
9	EABM	Yes	Yes	Yes	Yes	High	Yes	Laghari and Niazi 47
10	AIoT	Yes	Yes	Yes	Yes	High	Yes	Kato et al. 51
11	MAF	Yes	Yes	Yes	Yes	Low	Yes	Godfrey et al. 52
12	ACOSO	Yes	Yes	Yes	Yes	High	Yes	Fortino andcolleagues57,63,83
13	IoA	Yes	Yes	No	No	Low	No	Pico-Valenciaand Holgado-Terriza 84

Table 5 details the main studies that have raised one or several aspects that should include a framework to agentify the IoT. These aspects have been defined on the basis of the study by Sturm and Shehory ⁸⁵ who sets out four elementary criteria—concepts (guidelines that describe the concepts and properties employed by the framework), notations (symbols used to represent elements within a system), processes (actions and algorithms that a system can perform), and pragmatics (dealing with practical aspects of deploying and using in a specific scenario)—that an agent-oriented methodology should describe. However, in the case of the evaluation of a framework, only three of these criteria—concepts, processes, and pragmatics— have been considered as relevant and these have been complemented with three additional aspects that outline technical and proof of concepts associated with a framework. These concepts include the following: architecture (proposed architectural design that determines the elements of a system), completeness (degree of description of the programming model or modeling tool), and implementation (technological tool resulting from the proposed framework).

After analyzing all articles recovered in this work, it was found that only 12 studies proposed the introduction and description of a framework designed to agentify the IoT. The remaining 38 were only focused on describing an agent platform, performing a critical analysis on the subject or creating an application without going into detail on the aspects of Table 5. However, of the first 12 frameworks proposed, detailed in Table 5, 6 of them include too few technical details to allow a developer to create an agentificated IoT system. Also, one of them includes mid-level details and the remaining five include a high level of conceptual and technical details, that is, they describe each of the aspects of Table 5 including details of the proposed programming or modeling model. This therefore makes it easier for programmers to approach an IoT agentification project using such frameworks if tools associated with the framework were available. However, in all cases, the studies do not indicate where the middlewares involved can be accessed.

Table 6.

Main applications modeled from the agentification process of the IoT.

No.	Domain	M. of intelligence	Agent model	Technology of IoT	References
1	Transportation	Context-based rules	Reactive	RFID	Al-Sakran 36
2	Transportation	IA and probabilisticmodels	Non-specified	Wi-Fi, ZigBee,Bluetooth	Chamoso et al. 58
3	Transportation	IA model	JADE, Jade Leap	PC, smartphones	Do Nascimento and de Lucena 29
4	Entertainment	Context-based rules	Self-SmartMAS	NFC, RFID,smartphones	Ayala et al. 62
5	Health care	Semantic logic	JADE	TCP/UDP	Thangaraj et al. 22
6	Health care	Probabilistic model	Reactive	Non-specified	García-Magariño et al. 24
7	Comfort	IA model	Non-specified	AlertMe devices	Alan et al.42,59
8	Comfort	Rules	Reactive	Raspberry PI	Adhikaree et al. 39
9	Comfort	Reasoning	JASON	ZigBee, Wi-Fi,Bluetooth	Gateau and Rykowski 53
10	Manufacturing	Semantic logic	Gaia	RFID	Alexakos and Kalogeras 38
11	Agriculture	Context-based rules	JADE	RFID	Zhou and Lou 34
12	Logistic	Semantic andprobabilistic models	MAPS, Lopes et al.; 87 AFME, Muldoon et al. 88	PCs, PDAs, cameras	Wang et al. 26

However, in addition to the first 12 frameworks already mentioned, an additional framework has been covered in the literature which has been included as the 13th framework in Table 5. Although it is a purely conceptual framework based on dimensions—agent, environment, interaction, planning, organizational, and normative—, this type of study, as well as the 12 previous ones, constitutes a motivation that encourages researchers to create new tools and specialized application software that guide the process of agentification of the IoT. This is the case of the study proposed by Maamar et al. ⁸⁶ who present the modeling and proof of concept of an agenticated IoT system based on a normative model. Similarly, the different models of agents described in Tables 3 and 4 are also basic lines that can benefit the crystallization of the IoT agentification process.

Complementary to the frameworks presented in Table 5, it has also been proposed middlewares oriented to the development of IoT systems that provide specialized mechanisms to model IoT networks through the use of software agents. Their main function is aimed at developing abstraction mechanisms that facilitated low-level communication with the devices through the use of agents have been created, as Razzaque et al. ⁷⁹ present. In the above-mentioned study, 61 tools were presented, of which 10 (16.39%) have been developed to support software agents. Specifically, these agent-based tools—Impala, Smart messages, ActorNet, Agila, Ubiware, UbiROAD, AFME, MAPS, MASPOT, and TinyMAPS—support the use of mobile agents to support devices with limited computing resources in optimizing the use of energy (30%) and memory (60%), as well as carrying out the dynamic adaptation of networks of heterogeneous IoT objects (90%), discovery of services (80%), device discovery (70%) distributed, coherent context management (70%), interoperability (100%), and autonomous behavior (100%). Their main objective is optimizing the allocation and resource management processes of networks of IoT objects.

In addition to the middlewares already specified, an additional study proposed by Savaglio et al. ¹⁰ summarizes a list of middlewares oriented to realize the process of agentification of the IoT. The analysis of each tool is performed taking into account several indicators such as the cognitive one. Thus, it is possible to list some other tools in addition to those identified by Razzaque et al., ⁷⁹ among which are the following: iCore, CIoT, ASSIST, INTER-IoT, and VICINITY. ¹⁰

In synthesis, it is important to note that despite the existence of specialized frameworks and middlewares to apply the process of agentification of the IoT, none of them is standard or part of a recommendation (e.g. W3C and FIPA). This has been one of the main limitations by which the process of agentification of the IoT has not been sufficiently extended or popularized in the industrial, scientific, and business environment.

IoT agentification from a practice view

In section “Systematic review,” we presented the two main approaches proposed for the development of IoT applications based on software agents. The evidence presented and described shows that the subject of study in this research has grown substantially over the last 5 years. Although, as shown in Figure 2, several of the studies have been theoretical (8%) and non-experimental (36%). In contrast, several of the models have been validated through experiments and simulations (36%), and others have been validated in real-life environments and situations. In this section, proposals in this last category are analyzed. However, experiments that have modeled real-world situations (e.g. real-road intersection and real museum) have also been considered.

Table 6 includes the most relevant details of the applications developed using IoT technologies and agents. The most significant contributions, according to the researchers’ criteria, show 12 proposals that resolve problems in 7 different areas: intelligent transport management, intelligent management of entertainment services, intelligent assistance in health care, intelligent comfort management, intelligent manufacturing processes, management of activities in the field of agriculture, and finally, intelligent logistics management. Each of the proposals described is focused on IoT technologies such as RFID, micro-computers (Raspberry PI), IP cameras, smart phones, and sensors and actuators compatible with communication interfaces based on ZigBee, Wi-Fi, and Bluetooth Low Energy (BLE). Furthermore, it is important to highlight that the models of agents used are quite heterogeneous. These models range from JADE-based reactive models to more specialized or specific ones managed by certain agent methodologies (e.g. JASON, MAPS, ⁸⁷ and AFME ⁸⁸ ) and other models that emerge as a result of the proposal presented (e.g. Self-SmartMAS).

IoT agentification and cognition

The proposals of applications that introduce the concept of agentification in IoT technologies summarized in Table 6 take advantage of the intelligence capabilities inherent to the reactive (proposals 1, 3, 4, 5, 6, 8, and 11) and deliberative (proposals 9 and 12) models. However, these generic behaviors are increasingly complemented by advanced AI techniques (proposals 2, 3, and 6), models based on probabilities (proposals 2, 6, and 12), and semantic logic (proposals 5, 10, and 12) that allow reasoning processes using ontologies to be executed. In addition, it is important to note that it is increasingly common to combine several of the previously discussed techniques in the same system, such as the use of contextual information in the case of proposal 12, which combines ontologies and a Markov decision-making model in order to more accurately recommend IoT services according to user needs.

The analysis of the applications listed in Table 6 is presented below, grouping them according to the cognitive mechanism used by each of them, in order to define their scope in terms of intelligence level.

Cognitive capacity modeled by means of a set of rules

The reactive agent model, mostly implemented via the JADE tool, employs an event-based mechanism managed by a set of rules that automatically trigger when a change occurs in the context where the agents operate. The use of this type of agents in networks of IoT objects is basically oriented toward the modeling of rules based on contextual information of the devices connected to these types of networks in order to more effectively coordinate the services offered by the devices. In this ambit, there are some proposals presented by Al-Sakran, ³⁶ Adhikaree et al., ³⁹ Ayala et al., ⁶² and Zhou and Lou. ³⁴

On the one hand, Adhikaree et al. ³⁹ employ reactive smart household agents that act autonomously in order to minimize electricity costs in smart homes. Similarly, Ayala et al. ⁶² proposes the use of reactive agents in the management of the activities of an intelligent museum, based on contextual information provided by sensors in order to provide context-specific services such as establishing routes within the museum considering the presence of other groups in different physical areas of the facility, and sharing and disseminating information between the guide and their group of visitors. Supported by RFID technologies as proposed by Adhikaree et al., ³⁹ Al-Sakran. ³⁶ combine these technologies with reactive agents to monitor the identification of vehicles on roads in real time and thus minimize the number of accidents. Finally, Zhou and Lou ³⁴ propose an intelligent load-tracking system that supplies the user with real-time information about vehicles and cargo based on reflex agents that have an internal representation of their environment that helps them trigger the action–condition rules based on the location of the load, in addition to establishing negotiation processes when problems arise with the route of a specific load.

Cognitive capacity modeled from deliberative components

Of the proposals retrieved in the search process carried out in the systematic review, only one proposal within the framework of the introduction of BDI components for the modeling of IoT intelligence has been identified. This proposal is the work of Gateau and Rykowski, ⁵³ who present a model to provide intelligent comfort based on deliberative agents that formulate and negotiate the needs and expectations of users in order to provide comfort through Maslow’s hierarchy of needs (thermal comfort, ventilation, humidity, gas emissions, lighting, and acoustics). It is important to point out that the use of these types of agents in the IoT is minor inasmuch as they are complex to model and adapt and not oriented toward the modeling of interaction ecosystems in medium- and large-scale MASs. This is one of the main reasons why the deliberative agent model is not the most used in IoT agentification, although the level of cognition can reach a higher level than that of reactive agents. In fact, as Moin ⁵⁶ states, neither Jadex nor any other next-generation BDI agent framework could efficiently operate platforms with limited resources such as the IoT.

Cognitive capacity modeled from semantic logic

This approach proposes the design of agents with the capacity to exchange information based on ontologies which represent the semantics of the information processed by the system in order to execute automatic tasks by means of semantic reasoning processes. In this regard, Alexakos and Kalogeras ³⁸ design agents with these capabilities so that decisions are based on the meaning of the data and not only on the syntax. It is important to point out that in order to apply this strategy, specialized ontologies are needed in the domain in which the application is classified and, in addition to this, ontologies are used to describe the aspects of IoT such as devices and resources, among others. For the case proposed by Alexakos and Kalogeras ³⁸ limited to the manufacturing environment, they employed on the one hand, the P2Ontology ontology, which organizes knowledge in manufacturing environments, and on the other hand, the SSN ontology (semantic sensor network ontology) specialized in organizing IoT knowledge.

Along the same lines, Thangaraj et al. ²² propose a similar agent approach that acts on the basis of semantic logic and semantic descriptors to carry out the intelligent management of patients (i.e. monitoring vital signs and diagnosing diseases based on the patient’s clinical data). The ontology used includes only the concepts and properties of vital signs. Finally, Wang et al. ²⁶ enable the efficient organization and management of various services provided by third parties in general IoT ecosystems through the introduction of ontologies in the agent-based IoT model.

Cognitive capacity modeled from probabilistic models

The intelligence capacity in this approach is acquired by the agents by means of the use of probabilistic and stochastic models such as Markov decision models, Bayesian networks, and genetic algorithms, among others. Of the applications listed in Table 6, Chamoso et al. ⁵⁸ use the Bayesian network technique to estimate the most appropriate workflow to be executed by the agents in an IoT environment in order to automatically optimize traffic flow management in smart cities. However, Chamoso et al. also incorporate lightweight agents that are based on collective intelligence techniques modeled from the interactions between the agents to execute collaborative behaviors, which are represented numerically and based on the assignment of values to the different parameters of the model.

In the same vein, Do Nascimento and de Lucena ²⁹ employ genetic algorithms based on evolution processes using stochastic techniques in the processes of crossing and mutation of the population in order for the objects to acquire the capacity of self-adaptation. In addition, García-Magariño et al. ²⁴ present a smart bed ABS-BedIoT which recognizes a sleeper’s positions and generates text files to apply Big Data techniques and thus detect certain postures and the amount of movement during sleep, with the aim of detecting harmful postures in chronic patients so that nurses can use this information to help them care for these patients. Finally, Wang et al. ²⁶ employ a Markov decision process (MDP) that allows agents to efficiently execute, organize, and manage various IoT services.

Cognitive capacity modeled with sophisticated AI techniques

The cognitive capacity in this approach is accomplished by incorporating intelligent heavyweight agents that mimic certain human cognitive abilities studied in the area of AI. In applications 2, 3, and 6 summarized in Table 6, which adopt this philosophy to add the intelligence component in IoT, Chamoso et al. ⁵⁸ propose two agent models, one heavy weight and another lightweight (previously described in the category of agents that use probability models). The first of these is based on AI techniques. The agent represents an entity that is specialized in automatically fusing information through supervised learning and prior training; that is to say, the model employs RNA and classifiers that combine different detection technologies which deliver heterogeneous data and provide more accurate information about IoT resources. In this type of model, the intelligence resides in each of the agents that participate in the system; consequently, from a computational standpoint, they are less efficient and are consequently not targeted at IoT devices but at Cloud Computing infrastructures that are endowed with high computing power to execute these types of tasks. Therefore, the agents that adopt these techniques include mobility support to execute the processes of intelligence in the cloud, and afterwards, to execute decision actions in the devices, using the results obtained.

Along these same lines, other works that adopt other AI techniques have been presented, including the proposals of Do Nascimento and de Lucena, ²⁹ Ramchurn et al., ⁸⁹ and Alan et al. ⁵⁹ In the first case, RNAs are again combined with evolutionary algorithms in order for the objects to acquire the capacity for self-organization and self-adaptation. In the second case, the proposed agent model uses automatic learning techniques to obtain awareness of the situation or condition of an environment where a natural disaster has occurred, based on data captured by unmanned aerial vehicle (UAV). These represent two practical cases where the cognitive actions of the agents are modeled via sophisticated AI techniques. However, additional models following the same approach can be proposed. Finally, in the latter case, Alan et al. ⁵⁹ also use machine learning techniques to predict the daily energy consumption in a home automation system based on the previous day’s consumption and thus, plan the rate and when to implement changes to predict the best rate.

Analysis of the IoT agentification

The selected primary studies were analyzed, organized, compared, and contrasted, in order to identify the main strengths (S), opportunities (O), weaknesses (W), and threats (T) of the IoT agentification process to achieve cognitive networks of this nature. These four types of components are described as follows.

Strengths

• S₁: Availability of theoretical guidelines (processing unit and frameworks) to carry out the IoT agentification.

• S₂: Several proposals of lightweight agent models with mobility support based on standard web technologies supported by devices with limited resources (e.g. processing, energy, and memory) such as IoT objects.

• S₃: Support for the modeling of cognitive components based on homogeneous and heterogeneous techniques (e.g. rules, deliberative components, semantic logic, probabilistic models, and the application of AI techniques) whose use depends exclusively on the type of IoT technology and agentification approach employed.

• S₄: Support of distributed and dynamic intelligence mechanisms (e.g. mobile agents and Cloud Computing) as support to execute actions that the IoT objects do not support on their own.

• S₅: Mechanisms of dynamic customization at runtime aimed at users who abstract the complexity of communications with IoT physical devices.

• S₆: Mechanisms for the support of human–object and object–object communication at the level of agent abstraction that favors collaborative processes.

• S₇: Availability of standards for agent-oriented technologies such as FIPA-ACL (Foundation for Intelligent Physical Agents–Agent Communication Language) and FIPA communication protocols, which facilitate the development of the IoT agentification process and, in particular, the processes associated with the communication between heterogeneous objects.

Opportunities

• O₁: Availability of lightweight web technologies (e.g. REST, JSON, HTML5, JavaScript, Jquery, AJAX, and Node.js) that can be easily integrated into the IoT agentification models to make them more efficient and compliant.

• O₂: Maturity of SBC technologies (e.g. Raspberry PI and BeagleBoard) that allow the creation of objects that are compatible with the interfaces and standards managed by the IoT under the open hardware license.

• O₃: Maturity of the technologies based on Cloud Computing for the development of lightweight agents that support highly complex cognitive abilities (AI and probabilistic models).

• O₄: Feasibility of adopting the use of open data by agents to use a global vocabulary in order to achieve a better level of semantic interoperability in IoT networks.

Weaknesses

• W₁: IoT objects that are currently on the market are mostly distributed as black boxes that prevent the embedding of programming code to adapt their behavior according to new needs.

• W₂: Limitations of agent platforms regarding interoperability support between heterogeneous agent platforms.

• W₃: Existence of mobility models of light agents that have not been standardized for use in IoT environments.

• W₄: Lack of specialized software engineering tools such as methodologies, patterns and anti-patterns, and CASE tools (Computer Aided Software Engineering) to guide the process of agentification of IoT networks.

• W₅: Higher level of complexity than the modeling of generic IoT infrastructures due to the fact that it is necessary to have the modeling of an agent ecosystem that takes advantage of external agents that operate in ecosystems of external agents.

• W₆: Increase in cost due to the requirement of additional resources (i.e. servers, specialized tools for agent management, and cloud computing services) that allow the creation of the agent ecosystem that governs the IoT objects and each of the resources associated with them.

• W₇: Tools for the development of the most popular and commonly used agents have not been developed for devices with resource constraints, but for personal computers and powerful servers.

Threats

T₁: Agents are vulnerable to attack by computer experts, which could affect the privacy of user information and the behavior of their IoT objects.

Main trends of the IoT agentification based on embedded agents

The objective of this section is to perform an analysis of the SWOT components defined in the previous section in order to determine the main tendencies and future lines of research related to the IoT agentification process—based on the criteria of the researchers and the different analyses carried out in the “Results” section. This analysis has been carried out from two perspectives: the first case proposes carrying out the agentification of the IoT by embedding agents in the objects themselves. The main trends and areas of work in this field are described as follows.

Standard lightweight agent models for IoT technologies (S₁–O₁)

It has been possible to demonstrate that the current agent platforms are not sufficiently light to be able to develop components that must be executed in devices with limited resources such as the IoT objects that are currently on the market. This is one of the main weaknesses (W₇) and for this reason, some proposals have developed their own agent models adopting lightweight web technologies. It is necessary to advance in the development of new platforms with lightweight and ultralight agents, taking advantage, on the one hand, of the possibility of using operating systems with a smaller memory footprint—thanks to the capacity to componentize—and with the possibility of executing a multitask environment in devices with low-performance processors and limited storage capacity. Along these lines, the agents of the future must have the ability to work in fixed and mobile environments (S₂–O₁). As a consequence, agents must be provided with capabilities that allow them to move between heterogeneous IoT components to assist in processes where a common objective is met based on a set of specialized agents. Therefore, mechanisms to support light mobility are also required, and although some proposals of this nature have already been addressed, it is necessary to standardize them (W₃) and for the source code to be available to scientists and programmers to use and to contribute to its improvement. It is important to point out that the created models must also step up their efforts to provide security mechanisms and manage data privacy (T₁) in order to provide trusted agents.

Marketing of distributed IoT objects under open hardware license (F₁–O₂)

One of the main limitations of many of the currently marketed IoT objects (e.g. electrical appliances) is that they are distributed as black boxes where the user can generally customize the object to change its behavior but does not have the option to adapt it or re-program it to meet his requirements in their entirety (W₁) as happens with technologies such as NEST thermostats (https://nest.com/thermostats/). In this sense, taking advantage of the maturity and cost reduction of SBC technologies (e.g. Raspberry PI and BeagleBoard), a new line of work focused on the creation of IoT physical objects distributed under open hardware license is possible. In this way, the agentification process, based on the embedding of agents in the objects themselves, may materialize in the short/medium term, given that these objects can innately handle behaviors in a flexible, autonomous, reactive, social, and intelligent way.

Intelligence as a service (F₃, F₄–O₃)

It is problematic that many of the commercialized IoT objects are advertised as “intelligent” objects when, in reality, their level of intelligence is narrow or limited. Modeling generalized intelligence is highly complex and therefore many companies offer the user the option of personalizing the device (e.g. a microwave) using previously defined parameters. However, AI allows us to go beyond what is offered via the use of agents which have support for distributed intelligence modeling, either by applying a single technique (S₃) or several techniques (S₄). In this sense, and thanks to the benefits offered by cloud-based computing technologies for executing complex processes and storing large volumes of data (O₃), agents can base their behavior on both simple intelligence techniques (e.g. rules) and on others of greater complexity (e.g. AI techniques, based on stochastic models). From there, a possible line of study to develop is focused on providing intelligence as a service. Thus, agents could exploit the computing power of the cloud so that the data captured by the agents in the IoT ecosystem are processed and, based on these results, the agents make decisions aimed at satisfying the behavior of the global system.

User-friendly agents (S₅–O₁)

In the field of agent-oriented technologies, the interaction between the software agents and the user is simpler than the interaction between humans and IoT physical devices (S₆). Therefore, the line of research focused on covering this aspect must address aspects such as the development of mechanisms that allow the creation and personalization of agent behavior at runtime, in a simple manner and without the need for the user to have the technical knowledge to carry it out. However, although this ability to adapt is inherent to the agents (S₅), efforts must be made to create conflict-resolution techniques in multiuser environments, which originate after introducing changes in agent behavior.

Main trends of the IoT agentification based on MASs

The second analysis has been carried out considering the implementation of the IoT agentification process by means of the development of MASs that manage IoT resources through the use of specialized middleware to carry out the communications with the objects, without the need for any agent to operate on the device itself. The developments that revolve around this approach are detailed below.

Interoperable agent platforms (F₁–O₁)

Considering the scale and level of heterogeneity handled in IoT environments, it is essential that the agents of a platform have the ability to interoperate with other agents operating on different platforms and that have been developed with different tools. This aspect is impossible to achieve if standards are not defined. We consider it necessary for the FIPA to extend its range of standards so that agents, regardless of the tool used for their creation, can analogously communicate with other agents, as is done by service-oriented computing through service-oriented architecture (SOA). ⁹⁰ A line of investigation covering this aspect can be developed in order to establish standards that enable technical interoperability (shared resources and communication interfaces), syntactic (the format of shared messages), and semantics (ontology and representation of shared knowledge), which are feasible in the agentification of the IoT according to the study by Savaglio et al. ¹⁰ Although syntactic interoperability has been achieved thanks to FIPA (S₇), it is important to establish foundations in view of total interoperability. Therefore, the developments will be framed in the application of current standards such as those employed on the Web (O₁) along with semantic technologies such as ontologies and linked open data so that all agents handle the same vocabulary and reach better levels of semantic interoperability, thus supporting the development of intelligent applications based on semantic logic techniques.

Specialized software engineering tools for the agentification of the IoT (S₁, W₄)

The systematic review provides us with evidence of the existence of theoretical bases, processing units, frameworks, and even real applications where the IoT agentification process has been addressed (S₁). However, it is important that basic guidelines—and, if possible, standards—are developed in order to consolidate this approach, related to the design of MASs for the management of IoT objects and their resources, and for the software development community to begin to apply them in a substantial way. In this sense, we propose a line of research that addresses software engineering tools such as formal methods, methodologies, and CASE tools that automate the process of modeling MAS oriented for IoT.

Security mechanisms and data privacy (F₁–A₁)

Finally, security and privacy are indispensable aspects to be covered by environments that carry out large-scale interaction processes and transmit data, often of a private nature. Accordingly, a new line of investigation is raised in which communication protocols for agents at the inter- and intra-platform level are established, in order to provide secure mechanisms to the users of IoT agents. Moreover, it is important that agents work according to mechanisms that manage the consent of the data according to the user, taking into account the context in which the operations will be carried out. This is because, as we speak of large-scale and distributed agent ecosystems, it is possible for a user’s data to be provided in one country and the intelligence processes executed by the agent to be executed in another country, governed by different legislation. Hence, the importance of establishing consent mechanisms for accessing and handling the data at every moment that this is required and that the user has knowledge of what is happening with the information he or she generates.

Conclusion

From the systematic review carried out, it was possible to confirm the existence of several studies in which the guidelines to carry out the IoT agentification process have been introduced. The main studies proposed in the last 5 years (2013–2017) show an increasing trend (RQ₁) and it is expected that, with improvements in agent technologies regarding intra- and inter-platform communication, applications that adopt this approach will grow exponentially in the coming years. Furthermore, through the PIS analyzed in this study, it was possible to identify two basic approaches that can be used to integrate IoT and software agents. These approaches basically propose the modeling of objects by embedding an agent in the device itself and the modeling of complete networks of IoT objects through the use of MASs (RQ₂). From these approaches, it is not only possible to model cognitive behaviors but also to add social and collaborative behavior, privacy control, adaptation, interoperable communication, and autonomy—all of which represent necessary aspects for the current networks of IoT objects.

We also noted that the IoT has a tendency toward an emerging paradigm called the IoA. ³¹ The IoT is viewed as an approach supported by intelligent agents that efficiently manage the resources offered in IoT for both objects and users. Entities such as agent of things, IoT-a, Iota, and smart object are expected to facilitate the development of IoT object networks with a high degree of intelligence and that are also capable of interoperating with heterogeneous agents (RQ₃). Hence, the conceptual proposals being implemented in practical middlewares which allow the development of IoT cognitive applications are important since the current practical proposals implemented in real environments that have adopted this approach are relatively few—many of them remain at the level of experiments and simulations (RQ₄). Therefore, it is deemed appropriate to develop methodological frameworks and tools that adopt the approach studied in this article, so that the software development community has the guidelines to guide and motivate them to use this paradigm so that it can consequently be extended, popularized, and established as a standard for the development of IoT cognitive networks.

However, this research has also allowed us to investigate the main techniques used by the IoT agentification process. We identified five main alternatives: cognitive capacity modeled from rules, deliberative components, semantic logic, probabilistic models, and the application of AI techniques, which are used individually and in combination (RQ₆). The strategy employed basically depends on the objectives set, the complexity of the problems, the big data generated by the IoT objects, and the type of IoT technology used. In addition, it is important to highlight that the introduction of software agents in IoT architectures enables objects to be more autonomous, proactive, interoperable with heterogeneous objects, adaptive to new user needs, collaborative with heterogeneous IoT ecosystems, and self-adaptive and self-organizing, thanks to the learning mechanisms that the agents can include (RQ₅). These capabilities are not inherent to IoT objects, so the process of agentification—although it requires more technological resources and its operationalization—can bring a higher level of complexity and can provide the previously specified benefits for the fulfillment of the global objective of IoT—total autonomy.

Finally, this research presented a SWOT analysis in which the main strengths and weaknesses associated with the IoT agentification are analyzed, as well as the opportunities and lines of research that this approach will have in the future in order to become a more consolidated and specialized approach with an acceptable level of standardization for the development of IoT cognitive applications for real environments (RQ₇). In summary, regarding IoT object networks, not only are the limitations identified in previous years resolved but the field is also constantly evolving. Based on this, we share the vision of Yu et al., ³¹ viewing IoT systems as large ecosystems of agents. Consequently, we believe that the scientific community should enhance its efforts toward standardizing the IoT agentification process.