Cognitive Internet of Things: A New Paradigm …(IoT) vision [11]. Later, in 2005 IoT was formally introduced as the theme of the seventh in the series of International Telecommunication

1

Cognitive Internet of Things:A New Paradigm beyond Connection

Qihui Wu, Senior Member, IEEE, Guoru Ding, Student Member, IEEE, Yuhua Xu, Student Member, IEEE,Shuo Feng, Zhiyong Du, Jinlong Wang, Senior Member, IEEE, and Keping Long, Senior Member, IEEE

Abstract—Current research on Internet of Things (IoT) mainlyfocuses on how to enable general objects to see, hear, and smellthe physical world for themselves, and make them connectedto share the observations. In this paper, we argue that onlyconnected is not enough, beyond that, general objects shouldhave the capability to learn, think, and understand both physicaland social worlds by themselves. This practical need impels usto develop a new paradigm, named Cognitive Internet of Things(CIoT), to empower the current IoT with a ‘brain’ for high-level intelligence. Specifically, we first present a comprehensivedefinition for CIoT, primarily inspired by the effectiveness ofhuman cognition. Then, we propose an operational frameworkof CIoT, which mainly characterizes the interactions among fivefundamental cognitive tasks: perception-action cycle, massivedata analytics, semantic derivation and knowledge discovery,intelligent decision-making, and on-demand service provisioning.Furthermore, we provide a systematic tutorial on key enablingtechniques involved in the cognitive tasks. In addition, we alsodiscuss the design of proper performance metrics on evaluatingthe enabling techniques. Last but not least, we present theresearch challenges and open issues ahead. Building on thepresent work and potentially fruitful future studies, CIoT has thecapability to bridge the physical world (with objects, resources,etc.) and the social world (with human demand, social behavior,etc.), and enhance smart resource allocation, automatic networkoperation, and intelligent service provisioning.

Index Terms—Cognitive Internet of Things, massive data an-alytics, semantic, knowledge discovery, decision-making, serviceprovisioning, cognitive radio network

I. INTRODUCTION

A. Background and Motivation

The Internet of Things (IoT), firstly coined by Kevin Ashtonas the title of a presentation in 1999 [1], is a technological

This work has been accepted for publication by IEEE Journal of Internet ofThings. Personal use of the material in this work is permitted. Permission fromIEEE must be obtained for all other uses, including reprinting/republishingthis material for advertising or promotional purposes, collecting new collectedworks for resale or redistribution to servers or lists, or reuse of any copyrightedcomponent of this work in other works.

This work is supported by the National Natural Science Foundation ofChina (Grant No. 61172062, 61301160) and in part by Jiangsu ProvinceNatural Science Foundation (Grant No. BK2011116).

Q. Wu, G. Ding, Y. Xu, S. Feng, Z. Du, and J. Wang are withthe College of Communications Engineering, PLA University of Scienceand Technology, Nanjing 210007, China (email: [email protected];[email protected]; [email protected]; [email protected];[email protected]; [email protected]). G. Ding is the correspondingauthor.

K. Long is with Institute of Advanced Network Technologies and NewServices (ANTS) and Beijing Engineering and Technology Center for Con-vergence Networks and Ubiquitous Services, University of Science andTechnology Beijing (USTB), No. 30, Xueyuan Road, Haidian District, Beijing,China 100083 (e-mail: [email protected]).

revolution that is bringing us into a new ubiquitous connec-tivity, computing, and communication era. The developmentof IoT depends on dynamic technical innovations in a numberof fields, from wireless sensors to nanotechnology [2]. Forthese ground-breaking innovations to grow from ideas tospecific products or applications, in the past decade, we havewitnessed worldwide efforts from academic community, ser-vice providers, network operators, and standard developmentorganizations, etc (see, e.g., the recent comprehensive surveysin [3]–[5]). Technically, most of the attention has been focusedon aspects such as communication, computing, and connectiv-ity, etc, which are indeed very important topics. However, weargue that without comprehensive cognitive capability, IoT isjust like an awkward stegosaurus: all brawn, no brains. Tofulfill its potential and deal with growing challenges, we musttake the cognitive capability into consideration and empowerIoT with high-level intelligence. Specifically, in this paper,we develop an enhanced IoT paradigm, i.e., Brain-EmpoweredInternet of Things or Cognitive Internet of Things (CIoT), andinvestigate the involved key enabling techniques.

Before gonging deep into the new concept CIoT and itsenabling techniques, let’s first share two interesting applicationscenarios that will probably come into our daily life in future:

Application scenario 1: Let’s imagine that it’s Friday, afterfive days’ hard work, I’d like to relax myself and watch aTV Soap Opera tonight. When time goes to the midnight,I become more and more sleepy and finally fall asleep onmy sofa. Generally, I will wake up late on Saturday and feelvery tired since I do not sleep well with the TV noise, theuncomfortable sofa and the fluctuating temperature all nightlong. Consequently, I have a dream that one day the TV, thesofa, and the air conditioner in my room could individuallyor cooperatively sense my movement, gesture, and/or voice,based on which they analyze my state (e.g., ‘sleepy’ or ‘notsleepy’), and make corresponding decisions by themselves tocomfort me, e.g., if I am in the state of ‘sleepy’, the TV itselfgradually lowers or even turns off the voice, the sofa slowlychanges itself to a bed, and the air conditioner dynamicallyadjusts the temperature suitable for sleep.

Application scenario 2: Living in a modern city, trafficjams harass many of us. With potential traffic jams intoconsideration, every time when the source and the destinationis clear, it is generally not easy for a driver to decide whatthe quickest route should be, especially when the driver isfresh to the city. Among many others, the following schememay be welcome and useful for drivers: Suppose that thereare a city of crowdsourcers, such as pre-deployed cameras,

arX

iv:1

403.

2498

v1 [

cs.A

I] 1

1 M

ar 2

014

2

(a) Application scenario 1. (b) Application scenario 2.

Fig. 1. Motivational and illustrative application scenarios.

vehicles, drivers, and/or passengers, intermittently observe thetraffic flow nearby and contribute their observations to a datacenter. The data center effectively fuses the crowdsourcedobservations to generate real-time traffic situation map and/orstatistical traffic database. Then, every time when a drivertells his/her car the destination, the car will automaticallyquery the data center, deeply analyze the accessed trafficsituation information from the data center and meanwhile othercars/drivers’ potential decisions, and intelligently selects thequickest route or a few top quickest routes for its driver.

Just like the aforementioned examples, you may be familiarwith a lot of other blueprints of “intelligent life” and lookforward to that the dreams could become true soon. But thingsare not that simple as they look like. Just as a sportscar withoutengine is only a gorgeous waste, the IoT without a ‘brain’ isnot enough to bring the expected convenient and comfortablelife to us. These observations motivate us to develop the newparadigm Cognitive Internet of Things (CIoT).

But, first and foremost, what do we mean by CognitiveInternet of Things? Before responding to this question, it isin order that we first address the meaning of the related term“cognition.” Referring from the well-known books [6]–[9], itis more appropriate to refer to “cognition” as an “integrativefield” rather than a “discipline” since the study on “cogni-tion” integrates many fields that are rooted in neuroscience,cognitive science, computer science, mathematics, physics,and engineering, etc. Specifically, in this paper, the authorstake the operational process of human brain as the referenceframework for cognition [9], and offer the following definitionfor cognitive internet of things:

Cognitive Internet of Things (CIoT) is a new networkparadigm, where (physical/virtual) things or objects are in-terconnected and behave as agents, with minimum humanintervention, the things interact with each other following acontext-aware perception-action cycle, use the methodologyof understanding-by-building to learn from both the physicalenvironment and social networks, store the learned semanticand/or knowledge in kinds of databases, and adapt themselvesto changes or uncertainties via resource-efficient decision-making mechanisms, with two primary objectives in mind:

• bridging the physical world (with objects, resources, etc)and the social world (with human demand, social behav-

ior, etc), together with themselves to form an intelligentphysical-cyber-social (iPCS) system;

• enabling smart resource allocation, automatic networkoperation, and intelligent service provisioning.

B. Historical Notes

The history of Internet goes back to the development ofcommunication between two computers through a computernetwork in the late 1960s [10]. Since then, the evolutionof the Internet has passed three main phases: Internet ofComputers, Internet of People (mainly via social networking),and Internet of Things (including computers, people and anyother physical/virtual objects).

As mentioned above, the term ‘Internet of Things’ wasfirstly coined by Kevin Ashton in 1999 [1]. Then, in 2001the MIT Auto-ID center presented their Internet of Things(IoT) vision [11]. Later, in 2005 IoT was formally introducedas the theme of the seventh in the series of InternationalTelecommunication Union (ITU) Internet reports [2]. In 2008,the first international conference on the internet of things washeld in Zurich [12]. In 2009, China government advocatedthe idea of “Sensing China” and Wuxi city became oneof the leading centers of IoT-related research and industryin China [13]. At the same year, IoT European ResearchCluster (IERC) presented a document of IoT strategic re-search roadmap on future research and development until2015 and beyond 2020 [14], and one year later, publisheda comprehensive document on the vision and challenges forrealizing the IoT [15]. In the past couple of years, the IoT hasgained significantly increasing attention from academia as wellas industry, comprehensive surveys can be found in [3]–[5].Briefly, so far IoT is a very broad paradigm and many visions(e.g., “Internet oriented visions,” “Things oriented visions,”and “Semantic oriented visions” [3]) coexist.

Unlike (conventional) IoT, the research on Cognitive In-ternet of Things (CIoT) is very limited. In [16], a cognitivemanagement framework is presented to empower the IoTto better support sustainable smart city development, wherecognition mainly refers to the autonomic selection of themost relevant objects for the given application. In [17], CIoTis viewed as the current IoT integrated with cognitive andcooperative mechanisms to promote performance and achieve

3

intelligence, where the cognitive process is made up of a three-layer cognitive ring. Alternatively, in this paper we coin theCIoT by integrating the operational process of human cogni-tion into the system design, and we also provide systematicdiscussions on the key enabling techniques for the fundamentalcognitive tasks involved in the research and development ofCIoT.

Another related topic is Cognitive Radio Networks (CRN),which was firstly proposed by Joseph Mitola III in 1999 [18]and recoined by Simon Haykin in 2005 from a signal pro-cessing perspective [19], and since then the research on CRNhas been one of the hottest topics in the field of wirelesscommunications (see, e.g., [20]–[24]). One common point ofCIoT and CRN is that both of them benefit from the recentadvances in cognitive science [6]–[9]. The differences betweenCIoT and CRN are much more than the common. CRN is well-known as a promising paradigm to improve the utilization ofradio electromagnetic spectrum, by allowing unlicensed radiosto opportunistically access the idle spectrum licensed to theprimary radios [18]–[25]. CRN is in essence a radio systemwith the objective to improve wireless network throughput.However, CIoT generally consists of (massive) heterogeneousgeneral objects, not just radios, with various objectives fordifferent applications. Moreover, the technical research onCIoT should not focus on specific applications, instead, itshould be general enough to support as many applications aspossible, and consequently face much more unique challenges,which will be discussed in detail in the following sections.

C. Purpose of this Paper

The original motivation of the concept ‘Internet of Things’was explained by Kevin Ashton as follows [1]:

“Today computers-and, therefore, the Internet-are almostwholly dependent on human beings for information ... Theproblem is, people have limited time, attention and accuracy... We need to empower computers with their own means ofgathering information, so they can see, hear and smell theworld for themselves...”

The primary purpose of this paper is to build on KevinAshton’s visionary insights and enhance them by empowergeneral objects to learn, think, and understand physical andsocial worlds by themselves, by effectively integrating theoperational process of human cognition into the design ofIoT and presenting detailed expositions of cognitive processingtechniques that lie at the heart of Cognitive Internet of Things.

D. Organization and Potential Applications

The reminder of this paper is organized as follows. SectionII presents an overview of CIoT. Section III-V sequentiallyaddress the key enabling techniques for the fundamental cog-nitive tasks. Section VI provides discussion on the design ofperformance evaluation metrics for CIoT. Section VII presentsthe research challenges and open issues and Section VIIIconcludes the paper.

The work in this paper can be applied to many practicalapplications, e.g., the two application scenarios (i.e., smart TVand intelligent transportation) described in Section I-A. Taking

the second application scenario as an example, the frameworkof CIoT developed in this Section can be applied to buildthe architecture of an intelligent transportation system, andthe enabling techniques introduced in Section III-V can beembedded to CIoT companies’ products, such as the softwareor Apps.

II. COGNITIVE INTERNET OF THINGS: AN OVERVIEW

A. From Internet of Things to Cognitive Internet of Things

Currently, one of the most distinguished characteristics ofInternet of Things is that: with the increasing inter-connectivityamong general things or objects, a number of interestingservices or applications are emerging. However, so far manyof the existing Internet of Things applications are still de-pendent highly on human beings for cognition processing.This observation serves as one of the primary motivations ofthis paper to introduce ‘Cognitive Internet of Things’, wheregeneral objects behave as agents, and interact with physicalenvironment and/or social networks, with minimum humanintervention. Briefly, Cognitive Internet of Things enhancesthe current Internet of Things by mainly integrating the humancognition process into the system design. The advantages aremulti-fold, e.g., saving people’s time and effort, increasingresource efficiency, and enhancing service provisioning, to justname a few.

B. Framework of CIoT and Fundamental Cognitive Tasks

Fig. 2 presents a framework of CIoT. Generally, CIoT servesas a transparent bridge between physical world (with generalphysical/virtual things, objects, resources, etc.) and socialworld (with human demand, social behavior, etc.), togetherwith itself form an intelligent physical-cyber-social (iPCS)system. From a bottom-up view, the cognitive process of theiPCS system consists of four major layers:• Sensing control layer has direct interfaces with physical

environment, in which the perceptors sense the environ-ment by processing the incoming stimuli and feedbacksobservations to the upper layer, and the actuators act soas to control the perceptors via the environment.

• Data-semantic-knowledge layer effectively analyzes thesensing data to form useful semantic and knowledge.

• Decision-making layer uses the semantic and knowledgeabstracted from the lower layer to enable multiple oreven massive interactive agents to reason, plan and selectthe most suitable action, with dual functions to sup-port services for human/social networks and stimulateaction/adapation to physical environment.

• Service evaluation layer shares important interfaces withsocial networks, in which on-demand service provision-ing is provided to social networks, and novel performancemetrics are designed to evaluate the provisioned servicesand feedback the evaluation result to the cognition pro-cess.

With a synthetic methodology learning-by-understandinglocated at the heart, the framework of CIoT includes fivefundamental cognitive tasks, sequentially, Perception-action

4

Perception/Sensing(Large-scale, heterogeneous

perceptors/sensors)

Data Analytics(Heterogeneous/nonlinear/high-

dimensional/parallel data processing...)

Physical World

Cyber World

Social World

Decision-Making(Cognitive selection, reasoning,

planning...)

Performance Evaluation(QoS/QoE/QoD/QoI, resource-

efficiency)

Service Provisioning(Sensing-as-a-service, … everything-as-a-service)

Action/Adaptation

Semantic Derivation and Knowledge Discovery

(Context, ontology, association analysis, outlier analysis...)

Learning-by-Understanding

Service Evaluation Layer

Decision-Making Layer

Data-Semantic-Knowledge Layer

Sensing Control Layer

Social Networks(Human demand, social behavior)

Physical Environment(Things, objects, resources)

Fig. 2. Framework of Cognitive Internet of Things (CIoT).

cycle, Massive data analytics, Semantic derivation and knowl-edge discovery, Intelligent decision-making, and On-demandservice provisioning. Briefly, perception-action cycle is themost primitive cognitive task in CIoT with perception asthe input from the physical environment and action as theoutput to it. On the other hand, on-demand service provi-sioning directly supports various services (e.g., Infrastructure-as-a-Service (IaaS), Platform-as-a-Service (PaaS), Sensing-as-a-Service (SaaS), and more broadly Everything-as-a-service(EaaS) [26]) to human/social networks, which has been inves-tigated recently (see, e.g., [27]–[29]). In the following sections,we will focus on the key enabling techniques involved in theother three fundamental cognitive tasks.

III. MASSIVE DATA ANALYTICS IN COGNITIVE INTERNETOF THINGS

The future CIoT will be highly populated by large numbersof heterogeneous interconnected embedded devices, which aregenerating massive data in an explosive fashion. The data wecollect may not have any value unless we analyze, interpret,understand, and properly exploit it. Taking the applicationscenario 2 introduced in Section I-B as an example, the trafficdata is collected from massive crowdsourcers, including pre-deployed cameras, vehicles, drivers, and passengers, which aregenerally noisy, corrupted, heterogeneous, high-dimensional,and nonlinear separable. To exploit the value of the massive

data, the development of effective algorithms on massive dataanalytics is urgently needed.

As shown in Fig. 3, in this section we propose a systematictutorial on the development of effective algorithms for mas-sive data analytics, which are grouped into four classes: 1)heterogeneous data processing, 2) nonlinear data processing,3) high-dimensional data processing, and 4) distributed andparallel data processing.

Organized Data

Massive Data Analytics

––Heterogeneous data processing

––Nonlinear data processing

––High-dimensional data processing

––Distributed and parallel data processing

... ...

Raw Sensing Data

Massive, noisy, corrupted, heterogeneous, high-dimensional, and nonlinear separable...

...

Massive

Heterogeneous

sensors

Fig. 3. The framework of massive data analytics in CIoT.

5

A. Heterogeneous Data Processing

In practical CIoT applications, the massive data are gen-erally collected from heterogeneous sensors (e.g., cameras,vehicles, drivers, and passengers), which in turn may provideheterogeneous sensing data (e.g., text, video, and voice).Heterogeneous data processing (e.g., fusion, classification)brings unique challenges and also offers several advantagesand new possibilities for system improvement.

Mathematically, random variables that characterize the datafrom heterogeneous sensors may follow disparate probabilitydistributions. Denote zn as the data from the n-th sensor andZ := znNn=1 as the heterogeneous data set, the marginalsznNn=1 are generally non-identically or heterogeneouslydistributed. In many CIoT applications, problems are oftenmodeled as multi-sensor data fusion, distribution estimation ordistributed detection. In these cases, joint probability densityfunction (pdf) f(Z) of the heterogeneous data set Z is neededto obtain from the marginal pdfs f(zn)Nn=1.

For mathematical tractability, one often chooses to assumesimple models such as the product model or multivariateGaussian model, which lead to suboptimal solutions [30]. Herewe recommend another approach, based on copula theory,to tackle heterogeneous data processing in CIoT. In copulatheory, it is the copulas function that couples multivariate jointdistributions to their marginal distribution functions, mainlythanks to the following theorem:

Sklar’ Theorem [31]: Let F be an N -dimensional cumula-tive distribution function (cdf) with continuous marginal cdfsF1, F2, ..., FN . Then there exists a unique copulas function Csuch that for all z1, z2, ..., zN in [−∞,+∞]

F (z1, z2, ..., zN ) = C(F1(z1), F2(z2), ..., FN (zN )

). (1)

The joint pdf can now be obtained by taking the N -orderderivative of (1)

f(z1, z2, ..., zN )

=∂N

∂z1∂z2 ...∂zNC(F1(z1), F2(z2), ..., FN (zN )

)= fp(z1, z2, ..., zN )c

(F1(z1), F2(z2), ..., FN (zN )

), (2)

where fp(z1, z2, ..., zN ) denotes the product of the marginalpdfs f(zn)Nn=1 and c(·) is the copula density weights theproduct distribution appropriately to incorporate dependencebetween the random variables. The topic on the design or se-lection of proper copula functions is well summarized in [32].

B. Nonlinear Data Processing

In CIoT applications, such as multi-sensor data fusion,the optimal fusion rule can be derived from the multivariatejoint distributions obtained in (2). However, it is generallymathematically intractable since the optimal rule generallyinvolves nonlinear operations [33]. Therefore, linear dataprocessing methods dominate the research and development,mainly for their simplicity. However, linear methods are oftenoversimplified to deviate the optimality.

In many practical applications, nonlinear data processingsignificantly outperforms their linear counterparts. Kernel-based learning (KBL) provides an elegant mathematical means

to construct powerful nonlinear variants of most well-knownstatistical linear techniques, which has recently become preva-lent in many engineering applications [34].

Briefly, in KBL theory, data x in the input space X isprojected onto a higher dimensional feature space F via anonlinear mapping Φ as follows:

Φ : X → F , x 7→ Φ(x). (3)

For a given problem, one now works with the mappeddata Φ(x) ∈ F instead of x ∈ X . The data in the inputspace can be projected onto different feature spaces withdifferent mappings. The diversity of feature spaces providesus more choices to gain better performance. Actually, withoutknowing the mapping Φ explicitly, one only needs to replacethe inner product operator of a linear technique with anappropriate kernel k (i.e., a positive semi-definite symmetricfunction),

k(xi,xj) := 〈Φ(xi),Φ(xj)〉F , ∀xi,xj ∈ X . (4)

The most widely used kernels can be divided into twocategories: projective kernels (functions of inner product, e.g.,polynomial kernels) and radial kernels (functions of distance,e.g., Gaussian kernels) [34].

C. High-Dimensional Data Processing

In CIoT, massive data always accompanies high-dimensionality. For example, images and videos observedby cameras in many CIoT applications are generally veryhigh-dimensional data, where the dimensionality of eachobservation is comparable to or even larger than the numberof observations. Moreover, in kernel-based learning methodsdiscussed above, the kernel function nonlinearly maps the datain the original space into a higher dimensional feature space,which transforms virtually every dataset to a high-dimensionalone.

Mathematically, we can represent the massive data in acompact matrix form. Many practical applications have ex-perimentally demonstrated the intrinsic low-rank property ofthe high-dimensional data matrix, such as the traffic matrixin large scale networks [35] and image frame matrix in videosurveillance [36], which is mainly due to common temporalpatterns across columns or rows, and periodic behavior acrosstime, etc.

Low-rank matrix plays a central role in large-scale dataanalysis and dimensionality reduction. In the following, weprovide a brief tutorial on using low-rank matrix recoveryand/or completion1 algorithms for high-dimensional data pro-cessing, from simple to complex.

1) Low-rank matrix recovery with dense noise and sparseanomalies: Suppose we are given a large sensing data matrixY, and know that it may be decomposed as

Y = X + V, (5)

1Matrix completion aims to recover the missing entries of a matrix, givenlimited number of known entries, while matrix recovery aims to recover thematrix with corrupted entries.

6

where X has low-rank, and V is a perturbation/noise ma-trix with entry-wise non-zeros. We do not know the low-dimensional column or row space of X, not even their di-mensions. To stably recover the matrix X from the sensingdata matrix Y, the problem of interest can be formulated asclassical principal component analysis (PCA) [36]:

minX

||X||∗ subject to ||Y −X||F ≤ ε, (6)

where ε is a noise related parameter, || · ||∗ and || · ||F standsfor the nuclear norm (i.e., the sum of the singular values) andthe Frobenious norm of a matrix.

Furthermore, if there are also some abnormal data Ainjected into the sensing data matrix Y, we have

Y = X + V + A, (7)

where A has sparse non-zero entries, which can be of arbitrarymagnitude. In this case, we do not know the low-dimensionalcolumn and row space of X, not know the locations of thenonzero entries of A, and not even know how many there are.To accurately and efficiently recover the low-rank data matrixX and sparse component A, the problem of interest can beformulated as the following tractable convex optimization [35]:

minX,A

||X||∗ + λ||A||1

subject to ||Y −X−A||F ≤ ε, (8)

where λ is a positive rank-sparsity controlling parameter, and|| · ||1 stands for the l1-norm (i.e., the number of nonzeroentries) of a matrix.

2) Joint matrix completion and matrix recovery: In prac-tical CIoT applications, it is typically difficult to acquireall entries of the sensing data matrix Y, mainly due to i)transmission loss of the sensing data from the sensors to thedata center, and ii) lack of incentives for the crowdsourcers tocontribute all their sensing data.

In this case, the sensing data matrix Y is made up of noisy,corrupted, and incomplete observations,

Y := PΩ(Y) = PΩ(X + A + V), (9)

where Ω ⊆ [M ] × [N ] is the set of indices of the acquiredentries, and PΩ is the orthogonal projection onto the linearsubspace of matrices supported on Ω, i.e., if (m,n) ∈ Ω,PΩ(Y) = ym,n; otherwise, PΩ(Y) = 0. To stably recover thelow-rank and sparse components X and A, the problem canbe further formulated as [37]

minX,A

||X||∗ + λ||A||1

subject to ||PΩ(Y)− PΩ(X + A + V)||F ≤ ε. (10)

The problems formulated in (6), (8), and (10) show thefundamental tasks of the research on “matrix completion andmatrix recovery” for high-dimensional data processing, whichis receiving growing attention ranging from mathematiciansto engineers (see e.g., [35]–[38]). To efficiently solve theproblems in (6), (8), and (10), existing algorithms mainlyinclude augmented Lagrange multipliers (ALM) algorithmand accelerated proximal gradient (APG) algorithm, which

have been explained in [38] in detail. Readers can tailor thetheoretical results in [35]–[38] and references therein to theirspecific CIoT applications of interest.

D. Parallel and Distributed Data Processing

So far, all the data processing methods introduced aboveare in essence centralized and suitable to be implemented ata data center. However, in many practical CIoT applications,where the objects in the networks are organized in an ad hocor decentralized manner, centralized data processing will beinefficient or even impossible because of single-node failure,limited scalability, and huge exchange overhead, etc. Now,one natural question comes into being: Is there any way todisassemble massive data into groups of small data, and trans-fer centralized data processing into decentralized processingamong locally interconnected agents, at the price of affordableperformance loss?

In this subsection, we argue that alternating directionmethod of multipliers (ADMM) [39], [40] serves as a promis-ing theoretical framework to accomplish parallel and dis-tributed data processing. Suppose a very simple case with aCIoT consisting of N interconnected smart objects. They havea common objective as follows

minxf(x) =

N∑i=1

fi(x), (11)

where x is an unknown global variable and fi refers to theterm with respect to the i-th smart object. By introducing localvariables xi ∈ RnNi=1 and a common global variable z, theproblem in (11) can be rewritten as

minx1,...,xN ,z

N∑i=1

fi(xi)

subject to xi = z, i = 1, ..., N. (12)

This is called the global consensus problem, since theconstraint is that all the local variables should agree, i.e.,be equal. The augmented Lagrangian of problem (12) can befurther written as

Lµ(x1, ...,xN , z,y)

=

N∑i=1

(fi(xi) + yTi (xi − z) +

µ

2||xi − z||2F

). (13)

The resulting ADMM algorithm directly from (13) is thefollowing:

xk+1i := argminxi

(fi(xi) + ykTi (xi − zk) +

µ

2||xi − zk||2F

)(14)

zk+1 :=1

N

N∑i=1

(xk+1i + 1/µyki

)(15)

yk+1i := yki + µ(xk+1

i − zk+1). (16)

The first and last steps are carried out independently ateach smart object, while the second step is performed at a

7

fusion center. Actually, when the smart objects are multi-hopconnected, the second step can be replaced by

zk+1i :=

1

|Ni|∑i∈Ni

(xk+1i + 1/µyki

), (17)

where Ni denotes the one-hop neighbor set of the i-th objectand | · | is the cardinality of a set. Eq. (17) means that thesecond step can also be carried out at each smart object byfusing the local data from one-hop neighbors.

This is a very intuitive algorithm to show the basic prin-ciple of ADMM. ADMM serves as a good general-purposetool for optimization problems arising in the analysis andprocessing of massive datasets in a parallel and distributedmanner. Apart from the intuitive ADMM algorithm for globalconsensus problem, more advanced topics include (but notlimited to) [39], [40]:• Developing ADMM algorithms for distributed large-scale

model fitting, where each update in subproblems (14)-(16) reduces to a model fitting problem on a smallerdataset. These subproblems can be solved using anystandard algorithm suitable for small to medium sizedproblems. In this sense, ADMM builds on existing al-gorithms for single machines, and so can be viewed asa modular coordination algorithm that coordinates a setof simpler algorithms to collaborate to solve much largerglobal problems together than they could on their own.

• Implementation of ADMM algorithms in a MapReduceframework, where each iteration of ADMM can easilybe represented as a MapReduce task: The parallel localcomputations are performed by Maps, and the globalaggregation is performed by a Reduce. MapReduce isa popular programming model for distributed processingfor very large datasets [41].

IV. SEMANTIC DERIVATION AND KNOWLEDGEDISCOVERY IN COGNITIVE INTERNET OF THINGS

With massive data analytics, tremendous perceived dataabout physical world, cyber world, and social world in CIoTare well processed into an organized manner. However, asCIoT envisions trillions of objects to be connected and func-tion cooperatively, it is still not feasible to utilize these ana-lyzed data for decision-making directly due to both complexityand inefficiency. As one can imagine, only if the objectswithin CIoT are able to understand correctly and reasonproperly can they behave appropriately. For instance, the signallamp in future smart transportation system may be able tounderstand how many vehicles and passengers are waitingat the intersection, whether there is an ambulance amongthem, which directions they are heading, and how long havethey been waiting. These kinds of information are taken intoconsideration by the lamp, so as to decide how to change thetransportation signal would be the most effective and fairestoption. Besides, this signal lamp may figure out a few patternsafter serving couple of months, such as the average time for20 adult passengers to pass is about 30 seconds, or a busalways runs faster than a truck. This knowledge can not onlybe utilized by the lamp in future decision, but also be exportedto the ones in social world.

Fig. 4. The framework of semantic derivation and knowledge discovery.

Therefore, to make the objects in CIoT understand andbe aware, it is necessary to enable them to automaticallyderive the semantic from analyzed data. Besides, based on theanalyzed data and semantic, some valuable patterns or rulescan be discovered as knowledge as well, which is a necessityfor everyday objects in CIoT to be, or appear to be intelligent,as illustrated in Fig. 4.

A. Semantic Derivation in CIoT

Generally, semantic refers to the meaning of any (set of)object, situation, symbol, language, etc., and semantic deriva-tion in CIoT is defined as the process of deriving semantic byadopting various kinds of semantic technologies from analyzeddata. In this subsection, we introduce and discuss several keyconcepts in semantic derivation, i.e., context, ontology, andsemantic standardization.

1) Context in CIoT: To date, there is no standard definitionof context. A well known definition for context is provided byAbowd et al. [42] as follows:

“Context is any information that can be used to characterizethe situation of an entity. An entity is a person, place, orobject that is considered relevant to the interaction between auser and an application, including the user and applicationsthemselves.”

After massive data analytics, tremendous perceived dataabout physical world, cyber world, and social world are wellorganized. Once we put these analyzed data in such a waythat they represent the situation of an entity, they are viewedas the context in CIoT. The context can be location, identity,time, activity, and so on. For example, the crowdsourced ob-servations on the city road are perceived data. After analyzing,these data are organized and used to construct real-time trafficsituation map and/or statistical traffic database, which are/isidentified as context that characterizes the traffic situation ofthe city.

Although context in CIoT contains the semantic desired bythe devices, it still needs to be further processed. One of thereasons is that in CIoT, the sources of context are massiveand heterogeneous. As a result, one identical situation can beexpressed in plenty of contexts from different sources, whichin fact contains the same semantic. It promotes the difficultyof understanding the meanings for everyday devices. Since

8

CIoT is envisioned to be capable of combining information ofdifferent contexts, it is necessary to apply effective semantictechnologies to obtain the semantic from different contexts.Among others, ontology is treated as one of the most importantcomponent in semantic technologies, and will be discussedhereafter.

2) Ontology in CIoT: In the existing literatures (see,e.g., [4] and [43]), it is established that one of the mostappropriate formats to manage context is ontology. In philos-ophy, an ontology is a theory about the nature of existence, ofwhat types of things exist; as a discipline ontology studiessuch theories. In artificial intelligence and Semantic Webresearches, the term of ontology refers to a document or filethat formally defines the relations among terms. In CIoT, thedefinition of ontology is adopted as Studer et al. [44] havedefined:

“An ontology is a formal, explicit specification of a sharedconceptualization. A conceptualization refers to an abstractmodel of some phenomenon in the world by having identifiedthe relevant concepts of that phenomenon.”

Ontology offers an expressive language to represent therelationships and context, and has provided a solution toidentify the same semantic came from different contexts inCIoT. For example, take identity as the context type in smarttransportation, the contexts obtained by different objects mightbe: the highest buildings nearby, the most symbolic one,the one around corner, or the one with ATM machines init. As a matter of fact, it is highly possible that all thesecontexts indicate the same semantic (meaning), which is theconstruction as headquarter of bank of communications. Notethat by using various kinds of semantic technologies (whoseextensive discussion falls out of the scope of this article) suchas ontology, the semantic is derived from contexts comprisedby analyzed data.

3) Semantic Standardization in CIoT: Undoubtedly, seman-tic standardization is an important enabler for the success ofsemantic derivation in CIoT paradigm, since it may effectivelyincrease the semantic inter-operability and extendibility. CIoTsupports interactions among massive heterogeneous sourcesof data and contexts through standard interfaces and modelsto ensure a high degree of semantic interoperability amongdiverse systems [14]. Although many different semantic stan-dards may coexist, the use of ontology based ones will enablemapping and cross-referencing between them, in order toenable information share/exchange.

In CIoT, semantic standards play an increasingly importantrole with every everyday objects connected. Its status isemphasized with two big changes occurred in CIoT: one ismassive and heterogeneous sources in physical world, theother is tremendous and personalized application demands insocial world. As a result, the objects are required to shareand/or exchange semantic information continuously. Semanticstandards make it possible for the objects to communicate themeanings with each other efficiently with minimum ambiguity.

Besides, semantic standardization can draw from, as well asserve for the intersected research field of CIoT and CRN. Stan-dards regarding spectrum allocation, nodes selection, transmitpower control, and communication protocols will ensure that

the objects connected in CIoT/CRN can share the valuableradio spectrum with each other harmoniously. As greaterreliance is placed on CIoT as the global infrastructure forprocessing information, it will be essential to deploy andfurther develop semantic standardization in future.

B. Knowledge Discovery from Analyzed Data in CIoT

As aforementioned in Section I-A, one of the major char-acteristics of CIoT compared to classic IoT is the emphasison high-level intelligence. It is not accomplished in anyseparate part of CIoT, and should be considered throughoutall the stages of design, development, implementation, andevaluation.

To achieve intelligence for the objects in CIoT, the most im-portant way is to realize knowledge discovery from analyzeddata, and then apply it in the following. In CIoT, knowledgeis actually a broad concept that includes the general principlesand natural laws related to every object. For example, thebehavior (even thinking) patterns of human in social world,the correlations and functional mechanisms among all thecomponents of cyber world, and the dynamic characteristicsand common laws of physical world, and so on.

In general, knowledge is valid, certain, and potentially use-ful [45]. It is also consolidated, contextualized, and more stablein time than data, context, or semantic in CIoT. As previous,take the smart transportation as an example. The crowdsourcedobservations on the street are regarded as raw data, the real-time traffic situation map and/or statistical traffic databaseare/is identified as analyzed data (context), the meaning ofwhether it is jammed on the way to destination currently issemantic, and the rules about that the average time for 20 adultpassengers to pass is about 30 seconds, or a bus always runsfaster than a truck are viewed as the knowledge in CIoT.

It is recognized that tremendous techniques from areas suchas artificial intelligence, machine learning, pattern recogni-tion, database technology, etc., can be applied to discoverknowledge from analyzed data in CIoT. In this article, severalknowledge discovery techniques which are well established inthe above listed disciplines are introduced as follows [46].

1) Association Analysis: One of the feasible knowledgediscovery techniques is association analysis. It is very usefulfor knowledge discovery from analyzed data, as there are manyassociation types existing in CIoT.• Multilevel associations involve semantic at different ab-

straction levels (such as the relation between street, city,and country understood by the objects in CIoT). To avoidachieving commonsense knowledge at high abstractionlevels as well as avoid achieving trivial patterns at low orprimitive abstraction levels, it is important to develop ef-fective methods using multiple minimum support thresh-olds to discover this kind of association, with sufficientflexibility for easy traversal among different abstractionspaces.

• Multidimensional associations involve more than onedimension (e.g., rules that relate how much time it costsfor a passenger passing a street to width of the streetand/or the passenger’s age. Here, time, street width, and

9

age are different dimensions). As the sources in CIoT areheterogeneous, the analyzed data obtained by differentkinds of objects usually focus on different aspects. Tech-niques for analyzing such kind of associations should bechosen according to the difference in how they handlerepetitive predicates.

• Quantitative association rules involve numerical at-tributes or measures which have an implicit orderingamong values (e.g., time, street width, or age as men-tioned before). The quantitative attributes can be dis-cretized into multiple intervals, and then be treated asnominal data in the discovery of this kind of associationrules.

2) Clustering Analysis: Clustering analysis, as one of theclassic knowledge discovery techniques, is the process ofpartitioning a set of analyzed data into subsets. Each subset isa cluster, such that the analyzed data in a cluster is similar toone another, yet dissimilar to the ones in other clusters. Theset of clusters resulting from a cluster analysis can be referredto as a clustering. In CIoT, different clustering methods maygenerate different kinds of clusters on the same set of analyzeddata. A simple example is that, the clusters formed whentrying to count the number of vehicles and the number ofpassengers separately, are different from the clusters formedwhen the vehicles and passengers are partitioned by theirheading directions. Since the partitioning is performed in cyberworld rather than social world, clustering analysis is veryuseful to lead to the discovery of previously unknown groupswithin the analyzed data.

3) Outlier Analysis: Another useful knowledge discoverytechnique is outlier analysis. In CIoT, some of the objectsmay not comply with the general behavior or action model likeothers. These objects are considered as outliers. As the threatsregarding security and privacy are extremely crucial in CIoT(such as falsifying the identity as an ambulance to pass thestreet early, or making credit card fraud to evade the paymentfor vehicles), outlier analysis plays an important role to keepCIoT from being compromised. To be specific, outliers may bedetected using distance measures where objects that are remotefrom any cluster are considered as outliers, or using reputationmechanism where the reputation of one object is calculatedbased upon its neighbors’ opinion or its historical behavior.After the outliers are differentiated from the normal ones, theywould be discarded from CIoT to keep a pure environment.

Besides, CIoT also provides comprehensive reasoningmechanisms based upon ontology [47], which allows knowl-edge discovery from the derived semantic when it is neces-sary. Furthermore, the knowledge discovery process shouldbe highly interactive. Thus, it is important to build flexibleinterfaces with social world, facilitating the user’s interactionwith the system in cyber world of CIoT. For instance, auser may like to first access to the derived semantic ordiscovered knowledge, evaluate its correctness and utility, andthen modify or just regenerate it. Interactive discovery shouldallow the requests coming from social world to dynamicallyrefine the discovery process, while taken the current situationof physical world into account. Knowledge that has beendiscovered as well as the input from social world should

also be incorporated into the following knowledge discoveryprocess as guidance. In addition, we should point out thatthe influence of knowledge is quite far-reaching [48], suchas the influence of facilitation in data analytics, the influenceof expectation in semantic derivation, and the influence ofsupplementation in decision-making, to name just a few.

V. INTELLIGENT DECISION-MAKING FOR COGNITIVEINTERNET OF THINGS

Generally, decision-making in CIoT includes reasoning,planing and selecting. For reasoning and planing, the keyconcerns are analyzing the collected data and inferring usefulinformation, which belong to data analysis in essence. To avoidillegibility, we refer to decision-making as selecting in thisarticle. The task of selecting is common in CIoT, e.g., selectingthe path in the smart traffic systems, choosing the channelsfor wireless transmission, and selecting the optimal servicewhen there are multiple services simultaneously available.To summarize, selecting can be defined as the process ofchoosing an action from the action set. Motivated by thelearning ability in cognitive radio networks [19], we studycognitive selecting in CIoT, which is characterized by havingthe ability to intelligently adjust the selecting based on thehistory information.

Methodically, three kinds of cognitive selecting have beenstudied in the literature [49]: Markovian decision process,multi-bandit armed problem and multi-agent learning. In com-parison, the first two kinds are mainly for single decision-maker while the third one is for multiple decision-makers.Since it is expected that there are a large number of decision-makers (human or machine) in CIoT, we focus on multi-agent learning. Since the selections of the decision-makersare interactive, we can formulate the multiple decision-makingsystem as a game and then study multi-agent learning ap-proaches. Specifically, we establish a framework for intelligentdecision-making in CIoT, study intelligent decision-making inlarge-scale CIoT, and investigate the learning approaches withuncertain, dynamic and incomplete information.

A. A Framework for Intelligent Decision-Making in CIoT

We establish a framework for intelligent decision-makingin CIoT, which is shown in Fig. 5. Each decision-maker hassemantic information and/or knowledge from the environment.

Decision-maker

Semantic information

Information from other decision-makers

Decision output

Information from other decision-makers

Decision output

Interactions among decision-makers

Knowledge

Decision-maker

Knowledge Semantic information

Fig. 5. The framework of intelligent decision-making in CIoT.

10

Note that semantic information is generally generated bysemantic derivation, while knowledge can be obtained fromknowledge discovery or be given from social world in advance.In addition, it may have information about other decision-makers if information exchange is available. However, ifthe information exchange is resource-consumed or even notavailable in some scenarios, a decision-maker does not haveinformation about others. Using the information from theenvironment and others and taking into account its servicedemand, a decision-maker performs the cognitive selecting andoutputs the decision result.

Since there are multiple decision-makers in CIoT, the se-lections of the decision-makers are interactive. To capturethe interactions among multiple decision-makers, one wouldformulate the problems of cognitive selecting as game models,which were originally studied in economy and have been suc-cessfully applied into several engineering fields. To get a betterunderstanding of game models, we briefly present the gamemodel. Formally, a game is denoted as G = N , An, un,where N is the set of players, An is the selection set of playern and un is its utility function. Due to the feature of distributedand autonomous decision-making in CIoT, non-cooperativegame models can characterize the interactions among decision-makers well. In a non-cooperative game, each player maxi-mizes its individual utility function and Nash equilibria arethe well-known stable solutions for non-cooperative games.A pure strategy Nash equilibrium is a selection profile suchthat if and only if no player can improve its utility functionby deviating unilaterally. Other concepts of stable solutionsin non-cooperative games are correlated equilibria [50] andBayesian Nash equilibria. For more analysis of game models,refer to [51].

Technically, there are two important issues for using game-theoretic learning for CIoT [52], [53]:

• Designing utility function. It is emphasized that a gamemodel only addresses the interactions among multipledecision-makers, whereas it does not guarantee the per-formance. In some worse cases, the selfish nature ofplayers may lead to inefficiency and dilemma, which isknown as tragedy of commons [54]. Thus, one shouldcarefully design the utility functions such that somemetrics can be improved, e.g., the aggregate quality ofexperience (QoE), fairness and resource utilizing effi-ciency. Moreover, the designed utility function shouldadmit some stable solutions, which are important forpractical applications.

• Achieving the stable solutions. With the multipledecision-making problem in CIoT now formulated asgame models, learning procedures are needed to convergeto the desirable stable solutions. In particular, appropriateapproaches are desirable for solving different informationconstraints, e.g., the network may be static or dynamic,the system parameters may be known or unknown, theinformation about the environment and others may becomplete or incomplete.

The interactive range

Fig. 6. An illustrative example of local interactions in CIoT.

B. Intelligent Decision-Making in Large-Scale CIoT

An interesting feature of CIoT is that there are always largenumber of spatially distributed decision-makers. Moreover, theselection of a decision-maker only has direct impact on itsnearby decision-makers; that is, the decisions in CIoT arelocal interactive. Examples of local interactions are given by:the vehicles in the smart traffic systems only affecting othervehicles in proximity and a sensor contending for resources(channel, energy and time) only with its neighbors. An illustra-tive example of local interactions in CIoT is shown in Fig. 6.The interactive range is context-dependent. There may be over-lapping in the interactive ranges, which is determined by thenetwork topology; moreover, when all other decision-makersare located in the interactive range of each decision-maker, thelocal interaction becomes ordinary global interaction.

The feature of local interaction makes the correspondinggame models different. Specifically, the game is called spatialgame, which is denoted as Gl = N ,Jn, An, un, where Nis the set of players, Jn is the player set in the interactiverange of n, An is the selection set of player n and un isits utility function. In global interactive games, the utilityfunction is determined by the selection profiles of all players,i.e., the utility function is expressed as un(an, a−n), wherea−n is the selection profiles of all other players except n. Incomparison, the utility function in spatial game is expressedas un(an, aJn

), where aJnis the selection profiles of players

in the interactive range of n.

It is seen that spatial game models are more suitable forlarge-scale CIoT. However, due to the spatially distributionof players, spatial game is generally hard to analyze. Themain reasons are: 1) although the direct interactions arelocal, there exists inherent mutual interaction among any twoarbitrary players, 2) the number of all players may be huge,3) the interactive neighbors of players are different. Thus, tomake the game-theoretic approaches feasible in large-scaleCIoT, some efforts are needed. One promising approach isintroducing local cooperation into spatial games. Specifically,although global information exchange among all players isnot possible in large-scale CIoT, local information in theinteractive range is feasible. Based on this, the player behavesaltruistically by taking its interactive neighbors into account. Itwas shown in [55] that local cooperation leads to near-optimaloptimization.

11

( )nr t

( 1) [ ( ), ( )]n n na t F a t r t+ =

( 1)na t +

Learning

Decision-maker

( )mr t

( 1) [ ( ), ( )]m m ma t F a t r t+ =

( 1)ma t +

Learning

Decision-maker

Interactions

Distributed networks with dynamic, unknown

and incomplete information

Fig. 7. The illustrative diagram of learning with uncertain, dynamic andincomplete information in CIoT.

C. Decision-Making with Uncertain, Dynamic and IncompleteInformation

In this subsection, we investigate information constraintsin CIoT, which are important for decision-making problems.Specifically, we study intelligent decision-making with uncer-tain, dynamic and incomplete information. The presented threeinformation constraints are common in CIoT. Taking the smarttraffic systems as an example, the arrival of vehicles is random,the congestion level of a path is dynamic and a sensor mayhave partial information for an event and have no informationabout others sensors.

To deal with the above information constraints, the op-timization metrics should be carefully designed. Generally,there are two optimization metrics with uncertain, dynamic andincomplete information. The first is to maximize the expectedpayoff, i.e. max E[rn(t)], where rn(t) is the random payoffafter each play. The second is to minimize the outage prob-ability, i.e., min Prrn(t) > ηn, where ηn is the thresholdfor achieving certain service. To illustrate, one may want tominimize the expected traveling time from home to the office,or minimize the probability that the traveling time is large thanthirty minutes.

To solve the uncertain, dynamic and incomplete informationconstraints, learning is a promising approach. The illustrativediagram of learning is shown in Fig. 7. The ideas are asfollows: 1) for a given selection profile (an(t), a−n(t)), eachplayer gets a random payoff rn(t), which is jointly determinedby the selection profile and the environment, 2) the playersemploy a rule to update its next selection based on the currentselection and payoff, i.e., an(t+ 1) = F [an(t), rn(t)], 3) thisprocedure is repeated until some stopping criterion is met. Itis noted that the proposed learning scheme is autonomous andfully distributed, since it only relies on the individual historyinformation of a player; moreover, its convergence propertycan be analyzed by the theories of Markovian process andstochastic approximation [56].

VI. PERFORMANCE EVALUATION METRICS IN COGNITIVEINTERNET OF THINGS

Evaluating the performance of CIoT service is a challengingtask, since a lot of considerations and factors are involved. In

Profit dimension Cost dimension

Data layer: QoD

Information layer: QoI

User layer: QoE

Device Utilization Efficiency

Computational Efficiency

Energy Efficiency

Storage Efficiency

Fig. 8. Considered metric structure.

order to fully cover the issue, we broadly divide the metricsinto two dimensions: profit and cost. The profit dimensioncorresponds to appealing results in CIoT, while the costdimension considers the cost efficiency aspect. The overallstructure of considered metrics is presented in Fig. 8.

A. Profit dimension

As shown in Fig. 8, we expect to characterize the profitdimension from the following three layers: data layer, informa-tion layer and user layer. Corresponding to these three layers,we use three metrics quality of data (QoD), quality of infor-mation (QoI) and quality of experience (QoE), respectively.

1) Data layer-QoD: The data layer metric aims to evaluatethe quality of sensed data, the process of data acquiring andthe possible data distribution at the Perception/Sensing stage.Data plays a fundamental role in the CIoT cycle and evaluatingits quality is desirable.

In IoT, the acquired data may not meet system require-ment resulted from the following factors. Firstly, the data iscommonly noisy in practice, due to the environment noiseand sensing devices’ deviation and limited sensing accuracy.The sensing accuracy and deviation may vary from deviceto device. Secondly, the data may be corrupted by maliciousdata. Thirdly, the data can be incomplete since not all datacan be collected considering the limited number of sensordevices and constrained sensing cost. Furthermore, even thedata is accurate and complete, it can be outdated for demand.In response to above considerations, we propose a new metric,quality of data (QoD). The QoD consists of data accuracy, datatruthfulness, data completeness and data up-to-dateness [57].Apparently, the data accuracy reflects the precision of collecteddata. The data truthfulness indicates the reliability degree ofthe data resource. The data completeness corresponds to theratio of collected data amount to the amount of all requireddata. The data up-to-dateness reflects the validity of data to thedecision making, i.e., if the data is too late to assist decision-making, it is meaningless. The four aspects jointly determinethe overall quality of data.

2) Information layer-QoI: Since CIoT is marked with theintelligent decision-making, where information plays a keyrole in functional Cycle of CIoT, the quality of information indecision-making needs to be evaluated. We treat the informa-tion as the input to decision making and resort to the conceptof quality of information (QoI) in [64]. We believe that QoIis a satisfactory metric at present, as it tries to concern theinformation that meets decision maker’s need at some place,location, social setting and specific time. Existing QoI metric

12

Level 1

Level 2

Level 3

Level 4

Communication

Computation

Application

Access

(a) QoE framework in IoT.

Service provisioning level

QoE

Binary mapping

Log-like mapping

Sigmoid mapping

Linear mapping

(b) Different service provisioning level-QoE mappings.

Fig. 9. QoE in user layer.

is defined as

QoI = Q ∗ P ∗R ∗A ∗D ∗ T ∗ V, (18)

where Q denotes quantity, P denotes precision, R denotesrecall, A denotes accuracy, D denotes detail, T denotestimeliness, V denotes validity. All values are normalized into[0, 1] with 1 representing the corresponding best case.

In the above metric, quantity represents how much usefulinformation the decision maker has obtained for a specifictask. If all needed information is available, Q = 1. Precisionhere may refer to the proportion of relevant information toall information gathered by sensors, networks or services.On the other hand, recall refers to the proportion of relevantinformation without the assistant from sensors, networks orservices. Accuracy represents the accuracy degree of infor-mation to decision maker’s requirement. Note that quantity,precision, recall and accuracy jointly characterize the qualityof the information quantity provided. Detail characterizes thecomplete degree of the information to the decision maker.Timeliness is used to measure the decision maker’s timelinealong which the information is to be employed. We denotethe time delay as the gap between the instant the informationavailable and the instant the information employed. Then, thetimeliness can be treated as inversely proportional to the timedelay. If the information is available before the decision-makerusing it, the timeliness is 1. Validity reflects the trueness of theprovided information. We may find that the QoD and QoI sharesome similar properties. However, the involved objectives aredifferent, that is, QoD is used for data quality evaluation, whilethe QoI is used for information quality evaluation.

3) User layer-QoE: QoE is defined by the InternationalTelecommunication Union (ITU) as “the overall acceptabilityof an application or service, as perceived subjectively bythe end user” [59]. Since IoT mainly concerns applicationsfor human, we believe the ideal of QoE is suitable formeasuring user profit in IoT applications. While existing QoEin communications and networking is mostly derived fromcommunication quality provisioning [60], neglecting the roleof upper layer computation resource and application quality.Therefore, we extent the QoE concept and derive a new QoEframework as shown in Fig. 9(a).

In the proposed framework, QoE is evaluated from factorsin four levels. Specially, level 1 “Access” focuses on thebasic Internet connection ability of application related thingsand objects, since without the Internet connection, the IoTalmost losses its spirit. Upon the access ability, level 2 turnsto the communication capability to guarantee the runningof application. Clearly, different applications or traffic mayimpose diverse communication capabilities. Both level 1 andlevel 2 capture the impact of communication on QoE, whichgenerally corresponds to existing QoE modeling methods. Onthe other hand, level 3 turns the focus to computation ability,which brings the computation resource, the new emergingresource from cloud computing, into consideration. Note thatthis is important for computation-intensive applications inCIoT. Finally, applications directly deliver the service tohuman. For example, whether the user interface is friendlyand whether the service is custom for human can greatly affecthuman’s perception. Thus, “application” constitutes level 4.

Define the overall performance of the above five levelsas service provisioning, different service provisioning levelto QoE mappings can reflect users’ heterogeneous demandsto some extent. As shown in Fig. 9(b), the different curvesindicates users’ diverse elasticity [61] [62] or sensibility toservice provisioning.

B. Cost dimension

There is “no free lunch” in that for every gain we make inpractice there is a price to be paid. Hence, we also considerthe cost dimension metric in terms of resource efficiency.In particular, the resource efficiency embodies four types asfollows.• Device Utilization Efficiency: Hardware resource, espe-

cially device resource is commonly constrained for tworeasons. On one hand, some devices, for example, thespectrum analyzer, are expensive. On the other hand,even for some relatively cheap sensors, ideally densedeployment is unrealistic. Therefore, given limited deviceresource, maximizing the utilization or exploring the ca-pability of devices is indispensable. The device utilizationefficiency evaluates the degree of utilization efficiencygiven limited devices resource. For example, given the

13

same amount of carbon dioxide sensors, increasing thesensors’ sample points on the geographic area by movingsensors can increase the gathered data amount, comparedwith static sensors deployment cases. Correspondingly,the sensors’s utilization efficiency is improved in themoving sensors case.

• Computational Efficiency: Computation and decisionmaking in CIoT incur computational load. The computingresource will become scarce when a large number of userswith various computation intensive tasks are involvedin CIoT. Although the cloud computing paradigm canrelieve the computing resource scarce problem, improv-ing the computational efficiency is still a fundamentalrequirement.

• Energy Efficiency: The energy consumption in CIoT mayoccur at all cognitive tasks from perception to decision-making, action/adaptation and service provisioning. Notethat besides the conventional energy consumption con-siderations in communication systems such as cellularnetwork, additional energy is needed to support massiveand ubiquitous wireless access for connected things,considerable computation and storage ability. In CIoT,the number of connected things on the global will benumerous, resulting in a considerable consumption inenergy. Therefore, there is a urgent demand to improvethe energy efficiency in IoT. The energy efficiency metrichas to be able to reflect overall energy utilization levelincorporating communication, computation, storage, etc.,in order to provide quantified information to reflectenergy consumption of certain configurations, to compareenergy consumption performance of different applicationsand solutions, to set research and development targets onenergy efficiency [63].

• Storage Efficiency: Storage cost is another important as-pect of cost, as the storage and update of data, informationand knowledge rely on physical storage in CIoT. With theincreasing larger amount of information and data in theemerging big data era, the storage demand will increasein CIoT, and the storage problem is becoming a newchallenge. Storing the largest amount of data with theleast physical storage and without performance loss to theservice is always preferred. Thus, the storage efficiencyevaluates the ability to store and manage data given fixedamount of physical storage space.

VII. RESEARCH CHALLENGES AND OPEN ISSUES

CIoT has truly drawn a beautiful and exciting future, thoughcurrent researches and developments are still far away fromthat vision. Several major research challenges and open issuesinclude (but not limited to):• In practical CIoT applications, it is much more challeng-

ing to process the obtained massive sensing data thatcan be of mixed characteristics, including heterogeneity,high-dimensionality, and nonlinear separability, etc.

• For different applications in large-scale CIoT applica-tions, the game models and the multi-agent learningalgorithms should be carefully designed. In particular,

the local interaction and the uncertain, dynamic andincomplete information constraints should be taking intoaccount for decision-making.

• In most existing multi-agent learning algorithms, theplayers update their strategies based on the history action-payoff information. This procedure may take long timeto converge since the players need to explore all the pos-sible selections. In CIoT, some new knowledge-assistedlearning technologies should be developed to increase theconverging speed and achieve better performance.

• Developing effective semantic technologies and knowl-edge discovery techniques that are more suitable for CIoTapplications is still a fundamental task.

• Most of current studies on QoE are limited on singleuser case, there is lack of study on system-level QoE,especially for large scale CIoT systems with massiveusers.

• Generic approaches in CIoT research mainly focus onabstracting common techniques involved in various appli-cations. However, generic approaches cannot be directlyused for each specific situation. To apply the generic ap-proaches for specific situations, more practical constraintsshould be further considered.

• Last but not least, more attention should be focused onbuilding the bridge from theory to practice. For example,how and where might the theoretical studies in CIoTresearch actually be applied? What does it mean for thecompany implementing a smart city?

VIII. CONCLUDING REMARKS

A new network paradigm, named Cognitive Internet ofThings (CIoT), was developed in this paper to empower thecurrent IoT with a ‘brain’ for high-level intelligence, wheregeneral objects can not only see, hear, and smell the physicalworld for themselves, but also learn, think, and understandphysical and social worlds by themselves. Inspired by humancognition process, we first presented a comprehensive defini-tion for CIoT. Based on this definition, we further providedan operational framework of CIoT, which characterizes thefundamental cognitive tasks. Then, we addressed the keyenabling techniques involved in the cognitive tasks in detail. Inaddition, we also discussed the design of proper performanceevaluation metrics and the research challenges and open issuesahead.

Finally, we envision that the presented research is offeredas a mere baby step in a potentially fruitful research direction.We hope that this article, with interdisciplinary perspectives,will stimulate more interests in research and development ofCIoT, to enable smart resource allocation, automatic networkoperation, and intelligent service provisioning.

ACKNOWLEDGMENT

We thank the editor and anonymous reviewers for their con-structive comments and suggestions, which help us improvethe presentation of this paper. We also thank Prof. YuemingCai, Mr. Ducheng Wu, Long Wang, Junfei Qiu, Liang Yue,Zhen Xue, and Ms. Yijie Luo for their helpful discussions.

14

REFERENCES

[1] K. Ashton, “That ‘internet of things’ thing in the real world, things mat-ter more than ideas,” RFID Journal, June 2009, http://www.rfidjournal.com/article/print/4986 [Accessed on: 2013-10-25].

[2] International Telecommunication Union (ITU), “ITU internet reports2005: The internet of things,” ITU, Executive Summary, Nov. 2005,http://www.itu.int/pub/S-POL-IR.IT-2005 [Accessed on: 2013-10-25].

[3] L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,”Computer Networks, vol. 54, no. 15, pp. 2787-2805, Oct. 2010.

[4] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, “Contextaware computing for the internet of things: A survey,” IEEE Communi-cations Surveys & Tutorials, Accepted for Publication.

[5] M. R. Palattella, N. Accettura, X. Vilajosana, et al, “Standardized proto-col stack for the internet of (important) things,” IEEE CommunicationsSurveys & Tutorials, vol. 15, no. 3, pp. 1389-1406, Third Quarter 2013.

[6] J. M. Fuster, Cortex and Mind: Unifying Cognition. Oxford, UK: OxfordUniv. Press, 2003.

[7] R. Pfeifer and C. Scheier, Understanding Intelligence. Cambridge, MA:MIT Press, 1999.

[8] M. A. Fischler and O. Firschein, Intelligence: The Brain, and theComputer, ser. MA. Reading: Addison-Wesley, 1987.

[9] S. Haykin, Cognitive Dynamic Systems: Perception-Action Cycle, Radar,and Radio. London, UK: Cambridge Univ. Press, 2012.

[10] N. Olifer and V. Olifer, Computer Networks: Principles, Technologiesand Protocols for Network Design. John Wiley & Sons, 2005. [On-line]. Available: http://au.wiley.com/WileyCDA/WileyTitle/productCd-EHEP000983.html

[11] D. L. Brock, “The electronic product code (epc) a namingscheme for physical objects,” Auto-ID Center, White Paper, Jan-uary 2001, http: //www.autoidlabs.org/uploads/media/MIT-AUTOID-WH-002.pdf [Accessed on: 2013-10-25].

[12] http://www.iot-conference.org/iot2008/ [Accessed on: 2013-10-25].[13] Wuxi Development and Reform Commis-

sion, “Wuxi Sensing China,” May 2010,http://www.nacsa.com/archives/files/wuxi event 20100526.pdf[Accessed on: 2013-10-25].

[14] O. Vermesan1,P. Friess, P. Guillemin, et al, “Internet of thingsstrategic research roadmap,” The Cluster of European ResearchProjects, Tech. Rep., September 2009, http://www.internet-of-things-research.eu/pdf/IoT Cluster Strategic Research Agenda 2009.pdf [Ac-cessed on: 2013-10-25].

[15] H. Sundmaeker, P. Guillemin, P. Friess, and S. Woelffle,“Vision and challenges for realising the internet of things,”European Commission Information Society and Media,Tech. Rep., March 2010, http://www.internet-of-things-research.eu/pdf/IoT Clusterbook March 2010.pdf [Accessed on:2013-10-25].

[16] P. Vlacheas, R. Giaffreda, V. Stavroulaki, et al, “Enabling smart citiesthrough a cognitive management framework for the internet of things,”IEEE Communications Magazine, vol. 51, no. 6, pp. 102-111, June 2013.

[17] M. Zhang, H. Zhao, R. Zheng, et al, “Cognitive internet of things:Concepts and application example,” International Journal of ComputerScience Issues, vol. 9, no. 3, pp. 151-158, Nov. 2012.

[18] J. Mitola III and G. Q. Maguire, “Cognitive Radio: Making SoftwareRadios More Personal,” IEEE Personal Communications, vol. 6, no. 4,pp. 13-18, Aug. 1999.

[19] S. Haykin, “Cognitive radio: Brain-empowered wireless communica-tions,” IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201-220,Feb. 2005.

[20] I. F. Akyildiz, W. Y. Lee, M. C. Vuran, and S. Mohanty, “NeXtgeneration/dynamic spectrum access/cognitive radio wireless networks:A survey,” Computer Networks Journal (Elsevier), Sep. 2006.

[21] N. Devroye, M. Vu, and V. Tarohk, “Cognitive radio networks: High-lights of information theoretic limits, models and design,” IEEE SignalProcess. Mag., vol. 25, no. 6, pp. 12-23, Nov. 2008.

[22] Z. Quan, S. Cui, H. V. Poor, and A. H. Sayed, “Collaborative widebandsensing for cognitive radios,” IEEE Signal Process. Mag., vol. 25, no.6, pp. 60-73, Nov. 2008.

[23] F. Gao, R. Zhang, Y.-C. Liang, and X. Wang, “Design of Learning BasedMIMO Cognitive Radio Systems,” IEEE Trans. Veh. Technol., vol. 59,no. 4, pp. 1707-1720, May 2010.

[24] Q. Wu, G. Ding, J. Wang, and Y.-D. Yao, “Spatial-temporal opportunitydetection in spectrum-heterogeneous cognitive radio networks: Two-dimensional sensing,” IEEE Trans. Wireless Commun., vol. 12, no. 2,pp. 516-526, Feb. 2013.

[25] R. Li, Z. Zhao, X. Zhou, et al, “The prediction analysis of cellular radioaccess network traffic: From entropy theory to networking practice,”IEEE Communications Magazine, accepted, 2014.

[26] P. Banerjee, R. Friedrich, C. Bash, et al, “Everything as a service:Powering the new information economy,” Computer, vol. 44, no. 3, pp.36-43, March 2011.

[27] D. Guinard, V. Trifa, S. Karnouskos, et al, “Interacting with the SOA-based Internet of Things: Discovery, query, selection, and on-demandprovisioning of web services,” IEEE Transactions on Services Comput-ing, vol. 3, no. 3, pp. 223-235, July-Sep. 2010.

[28] D. Bimschas, H. Hasemann, M. Hauswirth, et al, “Semantic-serviceprovisioning for the Internet of Things,” WowKiVS 2011, ElectronicCommunications of the EASST, vol. 37, 2011

[29] A. Zaslavsky, C. Perera, and D. Georgakopoulos, “Sensing as a serviceand big data,” in International Conference on Advances in CloudComputing (ACC-2012), Bangalore, India, July 2012, pp. 21-29.

[30] D. Mari and S. Kotz, Correlation and Dependence. London, U.K.:Imperial College Press, 2001.

[31] R. B. Nelsen, An Introduction to Copulas. New York: Springer-Verlag,1999.

[32] S. G. Iyengar, P. K. Varshney, and T. Damarla, “A parametric copula-based framework for hypothesis testing using heterogeneous data,” IEEETransactions on Signal Processing, vol. 59, no. 5, May 2011

[33] I. Goodman, R. Mahler, and H. Nguyen, Mathematics of Data Fusion.Kluwer Academic Publishers, 1997.

[34] G. Ding, Q. Wu, Y.-D. Yao, et al, “Kernel-based learning for statisticalsignal processing in cognitive radio networks: Theoretical foundations,example applications, and future directions,” IEEE Signal ProcessingMagazine, vol. 30, no. 4, pp. 126-136, July 2013.

[35] M. Mardani, G. Mateos, and G. B. Giannakis, “Dynamic Anomalogra-phy: Tracking Network Anomalies Via Sparsity and Low Rank,” IEEEJ. Sel. Topics Signal Process., vol. 7, no. 1, pp. 50-66, Feb. 2013

[36] E. Candes, X. Li, Y. Ma, and J. Wright, “Robust principal componentanalysis?,” J. ACM, vol. 58, pp. 1-37, 2011.

[37] Y. Chen, A. Jalali, S. Sanghavi, and C. Caramanis, “Low-rank matrixrecovery from errors and erasures,” IEEE Trans. Inf. Theory, vol. 59,no. 7, pp. 4324-4337, July 2013.

[38] Z. Lin, M. Chen, L. Wu, and Y. Ma, “The augmented Lagrangemultiplier method for exact recovery of a corrupted low-rank matrices,”Mathematical Programming, 2010.

[39] D. Bertsekas, Constrained Optimization and Lagrange MultiplierMethod. Academic Press, 1982.

[40] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributedoptimization and statistical learning via the alternating direction methodof multipliers,” Foundations and Trends in Machine Learning, vol. 3,no. 1, pp. 1-122, 2010.

[41] J. Dean and S. Ghemawat, “MapReduce: Simplified data processing onlarge clusters,” Communications of ACM, vol. 51, no. 1, pp. 107-113,2008.

[42] A. K. Dey, G. D. Abowd, and D. Salber, “A conceptual frameworkand a toolkit for supporting the rapid prototyping of context-awareapplications,” Hum.-Comput. Interact., vol. 16, pp. 97-166, Dec. 2001.

[43] T. Strang and C. Linnhoff-Popien, “A context modeling survey,” inWorkshop on Advanced Context Modelling, Reasoning and Management,Nottingham/England, 2004.

[44] R. Studer, V. Benjamins, and D. Fensel, “Knowledge engineering:Principles and methods,” Knowledge Engineering, vol. 25, no. 12, pp.161-197, 1998.

[45] A. Somov, C. Dupont, and R. Giaffreda, “Supporting smart-city mobilitywith cognitive Internet of Things,” Future Network & MobileSummit2013 Conference Proceedings, pp. 1-10, 2013.

[46] J. W. Han, M. Kamber, and J. Pei. Data Mining: Concepts andTechniques, 3rd ed. MA, USA: Elsevier, 2012.

[47] H. Chen, T. Finin, A. Joshi, L. Kagal, F. Perich, and D. Chakraborty,“Intelligent agents meet the semantic web in smart spaces,” IEEEInternet Computing, vol. 8, no. 6, pp. 69-79, Nov. 2004.

[48] M. W. Eysenck, and M. T. Keane. Cognitive Psychology: A Student’sHandbook, 6th ed. New York: Psychology Press, 2010.

[49] Y. Xu, A. Anpalagan, Q. Wu, et al., “Decision-theoretic dis-tributed channel selection for opportunistic spectrum access: Strate-gies, challenges and solutions,” IEEE Commun. Surveys Tuts..DOI:10.1109/SURV.2013.030713.00189, to be published.

[50] D. Foster, V. Rakesh, “Calibrated learning and correlated equilibrium,”Games Econ. Behav., vol. 21, no. 1, 1997.

[51] R. Myerson, Game Theory: Analysis of Conflict. Cambridge, MA:Harvard Univ. Press, 1991.

15

[52] P. Young, “Learning by trial and error,” Games Econ. Behav., vol. 65,pp. 626-643, 2009.

[53] B. Pradelski and H. Young, “Learning efficient Nash equilibria indistributed systems,” Games Econ. Behav., vol. 75, no. 2, pp. 882-897,July 2012.

[54] H. Kameda and E. Altman, “Inefficient noncooperation in networkinggames of common-pool resources,” IEEE J. Sel. Areas Commun., vol.26, no. 7, pp. 1260-1268, July 2008.

[55] Y. Xu, J. Wang, Q. Wu, et al, “Opportunistic spectrum access in cognitiveradio networks: Global optimization using local interaction games,”IEEE J. Sel. Signal Process., vol. 6, no. 2, pp. 180-194, 2012.

[56] P. Sastry, V. Phansalkar and M. Thathachar, “Decentralized learningof Nash equilibria in multi-person stochastic games with incompleteinformation,” IEEE Trans. Syst., Man, Cybern. B, vol. 24, no. 5, pp.769-777, 1994.

[57] P. Bellavista, A. Corradi, M. Fanelli et al. “A survey of context datadistribution for mobile ubiquitous systems”, ACM Computing Surveys(CSUR), vol. 44, no. 4, 2012.

[58] J. Zhang. N. Ansari, “On assuring end-to-end QoE in next generationnetworks: challenges and a possible solution,” IEEE Commun. Mag.,vol. 49, no. 7, pp. 185-191, July 2011.

[59] ITU-T Recommendation G.1070, “Opinion model for video-telephonyapplications.”

[60] M. Fiedler, T. Hossfeld and P. Tran-Gia, “A generic quantitative rela-tionship between quality of experience and quality of service,” IEEENetwork, vol. 24, no. 2, pp. 36-41, 2010.

[61] S. Shenker, “Fundamental design issues for the future internet,” IEEEJ. Sel. Areas Commun., vol. 13, no. 7, 1995.

[62] Q. T. Nguyen-Vuong, Y. Ghamri-Doudane, N. Agoulmine, “On utilitymodels for access network selection in wireless heterogeneous net-works,” in Network Operations and Management Symposium (NOMS),2008, pp. 144-151.

[63] T. Chen, H. Kim, Y. Yang, “Energy efficiency metrics for green wirelesscommunications,” in Proc. 2010 International Conference on WirelessCommunications and Signal Processing (WCSP), 2010, pp. 1-6.

[64] J. Mitola, “Cognitive radio architecture evolution,” Proc. IEEE, vol. 97,no. 4, pp. 626-641, 2009.

[65] J. Mitola, “Cognitive radio for flexible mobile multimedia communi-cations,” in Proc. IEEE International Workshop on Mobile MultimediaCommunications (MoMuC), 1999, pp. 3-10.