Spectrum Inference in Cognitive Radio Networks 2018

150 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO.
1, FIRST QUARTER 2018
Spectrum Inference in Cognitive Radio Networks:

Algorithms and Applications
Guoru Ding, Senior Member, IEEE, Yutao Jiao, Jinlong Wang, Senior Member, IEEE,
Yulong Zou, Senior Member, IEEE, Qihui Wu, Senior Member, IEEE,
Yu-Dong Yao, Fellow, IEEE, and Lajos Hanzo, Fellow, IEEE
Abstract—Spectrum inference, also known as spectrum provide a comparative analysis of various spectrum inference
prediction in the literature, is a promising technique of infer- algorithms and discuss the metrics of evaluating the efficiency
ring the occupied/free state of radio spectrum from already of spectrum inference. We also portray the various potential
known/measured spectrum occupancy statistics by effectively applications of spectrum inference in CRNs and beyond, with
exploiting the inherent correlations among them. In the past an outlook to the fifth-generation mobile communications and
few years, spectrum inference has gained increasing attention next generation high frequency communications systems. Last
owing to its wide applications in cognitive radio networks (CRNs), but not least, we highlight the critical research challenges and
ranging from adaptive spectrum sensing, and predictive spectrum open issues ahead.
mobility, to dynamic spectrum access and smart topology control,
to name just a few. In this paper, we provide a comprehensive Index Terms—Spectrum inference, spectrum prediction, cog-
survey and tutorial on the recent advances in spectrum inference. nitive radio, 5G, HF communications.
Specifically, we first present the preliminaries of spectrum infer-
ence, including the sources of spectrum occupancy statistics, the
models of spectrum usage, and characterize the predictability of I. I NTRODUCTION
spectrum state evolution. By introducing the taxonomy of spec- A. Background and Motivation
trum inference from a time-frequency-space perspective, we offer
HE CONTRADICTION between spectrum shortage and
an in-depth tutorial on the existing algorithms. Furthermore, we
T spectrum under-utilization has motivated the emergence
of dynamic spectrum access (DSA) or opportunistic spec-
trum access (OSA) [1]. Cognitive radio (CR) [2] has been
Manuscript received July 12, 2016; revised February 11, 2017 well recognized as one of the crucial techniques of realizing
and July 8, 2017; accepted August 28, 2017. Date of publication the DSA/OSA concept. Since its conception, CR has been
September 11, 2017; date of current version February 26, 2018. The work of
G. Ding, J. Wang, and Q. Wu was supported in part by the National Natural designed for autonomous reconfiguration by learning from
Science Foundation of China under Grant 61501510 and Grant 61631020, and ultimately adapting to the continuously changing radio
in part by the Natural Science Foundation of Jiangsu Province under Grant environment [3].
BK20150717, in part by the China Post-Doctoral Science Funded Project, and
in part by the Jiangsu Planned Projects for Post-Doctoral Research Funds. The first step of implementing a CR is to capture the relevant
The work of Y. Zou was supported in part by the National Natural Science information about the spectral evolution. In the spectrum man-
Foundation of China under Grant 61401223 and Grant 61522109, and in agement framework relying on spectrum sensing, spectrum
part by the Natural Science Foundation of Jiangsu Province under Grant
BK20140887 and Grant BK20150040. The work of L. Hanzo was supported allocation decisions, spectrum sharing, and spectrum mobil-
in part by the European Research Council’s Advance Fellow Grant Beam- ity proposed in [4], spectrum sensing has the task of sensing
Me-Up, and in part by the Royal Society’s Wolfson Research Merit Award. its occupancy and capturing the characteristics of the primary
(Corresponding authors: Yutao Jiao; Lajos Hanzo.)
G. Ding is with the National Mobile Communications Research Laboratory, user (PU). However, in the practical sensing process, some
Southeast University, Nanjing 210096, China, and also with the College of inevitable problems arise concerning the sensing speed, the
Communications Engineering, PLA University of Science and Technology, potentially excessive energy consumption and the limited sens-
Nanjing 210007, China (e-mail: dr.guoru.ding@ieee.org).
Y. Jiao is with the School of Computer Science and Engineering, Nanyang ing scope, all of which hinder the efficient operation of the
Technological University, Singapore 639798 (e-mail: yjiao001@ntu.edu.sg). CR. The main reason for these problems is that each CR can
J. Wang is with the College of Communications Engineering, PLA only sense the current radio environment at its operating loca-
University of Science and Technology, Nanjing 210007, China (e-mail:
wjl543@sina.com). tion without any awareness of the unsensed bands or locations
Y. Zou is with the School of Telecommunications and Information and of the future trends of the spectral domain activities [5].
Engineering, Nanjing University of Posts and Telecommunications, This inevitably wastes precious information about the evolu-
Nanjing 210003, China (e-mail: yulong.zou@njupt.edu.cn).
Q. Wu is with the College of Electronic and Information Engineering, tion of spectral states between time slots, frequency bands,
Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China geographical locations, etc.
(e-mail: wuqihui2014@sina.com). Spectrum inference/prediction is known as an effective tech-
Y.-D. Yao is with the Department of Electrical and Computer Engineering,
Stevens Institute of Technology, Hoboken, NJ 07030 USA, and also with nique complementary to spectrum sensing for capturing the
Northeast University, Shenyang 110819, China (e-mail: yyao@stevens.edu). relevant information about the spectral evolution and iden-
L. Hanzo is with the School of Electronics and Computer Science, tifying spectrum holes. Briefly, spectrum sensing determines
University of Southampton, Southampton SO17 1BJ, U.K. (e-mail:
lh@ecs.soton.ac.uk). the spectrum state in a passive manner using various signal
Digital Object Identifier 10.1109/COMST.2017.2751058 detection methods. By contrast, spectrum inference/prediction
1553-877X c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
DING et al.: SPECTRUM INFERENCE IN CRNs: ALGORITHMS AND APPLICATIONS 151
is a promising technique of inferring the occupied/free state Moreover, the work of this survey paper is also related to
of radio spectrum from already known/measured spectrum the survey papers on various aspects of CRNs, such as spec-
occupancy statistics by effectively exploiting the inherent cor- trum sensing [25]–[30], spectrum decision [31]–[35], resource
relations among them, in a proactive manner. Fortunately, allocation [36]–[41], security and privacy [42]–[48], MAC pro-
some other fields of application, such as the atmosphere [6], tocols [49], [50], routing [51], network coding [52], various
finance [7], network traffic [8] and human mobility [9], infer- applications [53]–[56], to just mention a few. A list of the
ence/prediction techniques (including the popular big data related survey papers is provided in Table I for further refer-
technique [10]) have provided potential techniques of dis- ences. On one hand, the studies on spectrum inference can find
covering the usage patterns hidden in the data and have their applications in these aspects, which will be discussed in
succeeded in maintaining the stability of the economy, fore- detail in the following sections. On the other hand, spectrum
casting the weather conditions and so on. Similarly, spectrum inference to some extent is a parallel or complementary tech-
inference/prediction is a promising technique that can be uti- nique to those in Table I. This paper fills the gap between them
lized for acquiring precious unknown spectrum occupancy by providing a comprehensive survey of spectrum inference in
information in advance and for enhancing the performance CRNs.
of the CRs by accelerating the process of choosing the
best channel and expanding the sensing scope in the time- C. Organization and Notation
frequency-spatial-domain [11]–[13]. Therefore, an increas- This paper is organized as illustrated in Fig. 1. We introduce
ing number of researchers have been focusing on using or the necessary preliminary knowledge on spectrum inference in
developing effective inference/prediction techniques for CRs. Section II. Specific inference algorithms operating in the time-
Preliminary results (see [14]–[20]) demonstrate that spectrum frequency-spatial domain are detailed in Section III, while
inference/prediction algorithms operating across the time- Section IV proposes a macroscopic view of various infer-
frequency-spatial dimensions may equip CRs with an accurate ence methods. The current and potential future applications
forecasting capability. Consequently, there is an urgent need are introduced in Section V. Finally, a range of challenges
for a survey of the specific algorithms, since the infer- and future trends are presented in Section VI along with our
ence/prediction function is at the heart of the CR architecture conclusions. The acronyms used in this article can be found
and operates in a realm, where communications techniques in the Table VII for convenience.
meet artificial intelligence.
II. P RELIMINARIES FOR S PECTRUM I NFERENCE IN CRN S
A. Sources of Spectrum Data
B. Comparison With Related Survey Articles Empirical real world spectrum measurements constitute the
Historically speaking, there is only a brief survey rely- very important source of spectrum data and play a funda-
ing on 15 citations [21] on spectrum prediction in CRNs, mental role in supporting the research and development of
which is focused on a temporal scope. By contrast, the spectrum inference techniques. Numerous spectrum occupancy
goal of this paper is twofold. The first is to present an measurement campaigns have been conducted all over the
all-encompassing systematic survey of the recent advances world (see [23], [24], [59]–[87]). Excellent surveys and tuto-
in spectrum inference/prediction algorithms from a time- rials on the latest advances of worldwide spectrum occupancy
frequency-spatial domain perspective. The second is to discuss measurements can be found in [23] and [24]. Briefly, here we
the potential applications and prospects in order to empha- highlight some common features of the spectrum measurement
size the significance of inference/predicton techniques in the campaigns as follows.
context of the CR technology and to provide guidelines for Firstly, the campaigns have covered various frequency
researchers focusing on making CR more intelligent and ranges. For example, the probably earliest measurement cam-
efficient. paigns conducted by the Institute of Telecommunication
Notably, spectrum inference is also related to the subject Science in the USA around 1995 [59] measured very broad
of spectrum occupancy measurements [23], [24] and primary frequency bands spanning from 108 MHz to 19.3 GHz.
user activity modeling [22]. Briefly, primary user activity One the other hand, the very recent ones carried out in
models use mathematical or theoretical expressions to mimic Beijing [86] and London [87] focused on TV bands (e.g.,
the underlying PU activities, while spectrum inference algo- 470 MHz∼790 MHz) to identify TV white spaces. Moreover,
rithms use machine learning or data mining methods to predict there are many other campaigns that measured specific
or infer the future or unknown PU activities based on the licensed bands (like WiFi bands, TV bands, radar bands and
available historic data. References [22]–[24] are excellent sur- cellular bands) relying on diverse lengths of measurement peri-
vey and tutorial papers on spectrum occupancy measurements ods (from a few hours to several years) as well as different
and/or theoretical modeling relying on statistical analysis, places (e.g., Denver, Dublin, Aachen, and Singapore, etc) and
while our paper focuses on presenting various data min- scenarios (including rural, suburban and urban areas, indoor
ing algorithms conceived for spectrum inference/prediction. and outdoor cases) [23], [24], [59]–[87].
Indeed, the spectrum models in [22]–[24] are very useful for Second, the campaigns have had various measurement
theoretical performance analysis or simulated data generation. setups. In terms of hardware devices, the spectrum measure-
However, they cannot be directly used for prediction. ment systems generally consist of the antenna, the filter, the
152 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 1, FIRST QUARTER 2018
TABLE I
A L IST OF R ELATED S URVEY A RTICLES
amplifier and the spectrum analyser to collect the spectrum requires advanced databases. On the other hand, spectrum
data. Most of the spectrum measurement campaigns specify measurements ensures a data-driven approach, complementar-
several common setups. For example, both the sampling rate ily to the traditional propagation model-based approach [242],
and the measurement period, as well as the frequency span and when aiming for improving the accuracy and the update
frequency points, the measurement location, antenna polariza- speed of the spectrum availability provided by the geoloca-
tion and direction have to be chosen according to the particular tion spectrum databases [5], [88], [246]. Moreover, spectrum
applications considered. Moreover, since the various settings measurements are also coupled with the building of radio
may lead to different levels of measurement complexity and environment maps (REM) [254]–[256].
accuracy, the tradeoffs among of these potentially conflicting
factors have been carefully taken into consideration.
Third, the campaigns have a tendency to bridge real B. Models of Spectrum Usage
world spectrum measurements and the public/private spectrum Exploiting the statistics of spectrum use or the determin-
databases [247]–[250]. On one hand, the collection, storage istic status of PUs have been one of the critical issues for
and evaluation of massive amounts of spectrum measurements secondary use of the licensed spectrum. The research on the
Fig. 1. The structure of this paper.
PU’s behavior gleans lessons from the spectrum measurement 1) Time Domain Model: In the time domain, the spectrum
campaigns. A brief conclusion is that the obtained spectrum usage may simply be modeled by a Markov chain having two
occupancy results, e.g., in terms of the duty cycle, from spec- states, one representing that the channel is busy and the other
trum measurement campaigns run by different groups are one representing that it is idle. In terms of temporal continu-
not always the same at different measurement time, loca- ity, these models can be classified either into the discrete-time
tions, frequency bands with various measurement hardware Markov chain (DTMC) model or into the continuous-time
and softwares. However, the primary user occupancy mod- Markov chain (CTMC) model.
els (e.g., DTMC and CTMC) discussed in the following are a) DTMC model: Since the PU’s states can be described
widely used by different measurement campaigns associated as being either busy or idle and most measured data are usu-
with various parameter setups for specific bands, locations, ally represented in the form of binary sequences, the spectrum
and time. usage patterns are reflected by bits of zero value indicating
Specifically, in CRNs, spectrum usage models of the pri- having no traffic and logical ones indicating that the spectrum
mary system usually determine both the action and the is being used by the PU at a particular time instant [94], [95].
performance of a secondary network and thus they play a The state space of a primary radio channel obeying the DTMC
significant role in spectrum inference, since the exploration is denoted by S = {0, 1}. The time index set is discrete,
and evaluation of inference algorithms as well as techniques t = tk = kTs , where k is a non-negative integer representing
is often based on them [90]. Spectrum usage models can be the step index and Ts is the time period between consecutive
utilized to discover patterns of the PU’s activities by ana- transitions or state changes.
lyzing the spectrum data and by reconstructing the statistical The most important parameter of the DTMC model is the
properties of the spectrum usage in real radio communication transition probability pij (tm , tn ) ∈ {0, 1}, which represents the
systems. By taking advantage of these models, we can generate probability that channel state moves from state i in the m-th
simulated data for validating the inference algorithms. In order slot to state j in slot n. In this particular case, the transition
to prepare a solid basis for our discussions in the next section, matrix is given by:
we will briefly present the most common approaches of mod-
p p01
elling spectrum usage across the three orthogonal dimensions P = 00 . (1)
p10 p11
of time, frequency and space, as illustrated in Fig. 2, [19], [22],
[89], [91]–[93]. Explicitly, in Fig. 2, there are several primary The overall DTMC model describing the activity of a
transmitters having their own primary coverage area, which PU’s channel is illustrated in Fig. 3. We can use the above-
are carefully coordinated for mitigating the interference arising mentioned concept of DC, denoted by , as a metric of
from the others. A number of PUs and CRs are roaming either describing the ratio of time during which the channel is
within or outside the coverage area of these primary transmit- declared busy as well as idle [89], [91].
ters. Furthermore, every primary transmitter has its spectrum In terms of stationarity, the DTMC models fall into two cat-
for licensed or unlicensed use. egories: stationary and non-stationary DTMC models. In the
Fig. 2. An illustrative scenario in the time-frequency-spatial dimension.
where (t) is a variable, which is a function of the dis-

crete time t. Based on different patterns of the channel’s load,
DC models can be developed using both deterministic and
stochastic modelling approaches [89]. On one hand, in many
particular services, like GSM and TETRA, common human
habits and social behaviors impose a significant impact on the
load patterns, which deterministically shape the characteris-
Fig. 3. DTMC model. tics of (t). On the other hand, the stochastic DC modelling
approaches are presented to describe the random variable (t)
as a function of numerous random factors contributing to the
stationary DTMC model, the transition matrix can be formu- PU’s behavior in real scenarios. It was found that the empirical
lated as, PDFs of (t) can be modelled by the Beta distribution [96]
and the Kumaraswamy distribution [97].

1− Furthermore, extensive empirical measurement results have
P= , (2) shown that the DTMCs associated with the deterministic and
1−
stochastic DC models mentioned above are capable of repro-
where is a constant parameter and it can be selected as ducing the average DC value and the statistical properties of
= p01 = p11 , especially in the long term = P(s = 1). In the PU’s idle or busy duration in real-world channels. The
this way, the DTMC model becomes capable of reproducing DTMCs have also been widely used in the literature to ana-
the mean DC of a real channel. However, the stationary DTMC lyze the system throughput, the average packet delay and the
may only be valid for a limited time period, if the system end-to-end packet delay (see [98], [99]).
exhibits only an approximately stationary behavior during b) CTMC model: In contrast to the DTMC model, the
this period, where the reproduced average DC approximately CTMC model pays closer attention to modeling the length
matches the instantaneous DC at all times [89]. of a state’s holding time. Note that the traditional exponen-
If the modeled system does not appear to exhibit strictly sta- tial distribution used for characterizing the state holding time
tionary characteristics, a non-stationary DTMC model should in CTMC is not necessarily accurate enough, according to
be considered. In this case, the transition matrix should be a range of realistic measurements and analysis conducted
defined as, across the globe [91], [100]–[106]. In view of this, many
researchers considered the Continuous-Time Semi-Markov
1 − (t) (t) Chain (CTSMC) models, where the sojourn time may obey
P(t) = , (3)
1 − (t) (t) arbitrary distributions. In [104], [105], [107], and [108], the
idle duration was assumed to obey either a generalized Pareto such as energy detection [117], matched filtering, and wave-
distribution [108], or a mixture of uniform distribution and the form based sensing. Due to the inevitable noise, incomplete
generalized Pareto distribution, or alternatively, obey a hyper- observations, and the ambiguity of parameters like the energy
Erlang distribution [104], [105] and so on. López-Benítez [89] decision threshold and the undiscovered spectrum features,
employ goodness-of-fit metrics to evaluate several common such formulations are unable to accurately infer the PUs’ true
distributions and to demonstrate that in real-world systems status and to capture the real characteristics of the PU’s activ-
CTSMC models have a better performance in terms of repro- ity [117]–[119]. In this sense, all the models, including those
ducing the statistical properties of the busy and idle durations. introduced above may be imperfect, which necessitates further
Additionally, the correlation between the busy and idle dura- research.
tions, which the CTSMC fails to reproduce has been analyzed
in [91] and [109]. C. Spectrum Inference in Various CRNs
2) Frequency Domain Model: In this paper, the frequency- Depending on the type of available side information and
domain model can also be termed as the time-frequency model, on the regulatory constraints, there are three main CRNs
since we do analyze the properties of spectrum data over the paradigms: underlay, overlay, and interweave [120]. Briefly,
allocated frequency band based on the channel’s DC which is the underlay CRN allows CRs to operate if the interference
a temporal parameter. Although the channels are considered to they cause is below a given threshold. In overlay CRNs, the
be mutually independent within a spectrum band [90], [110], CRs overhear the transmissions of the PUs, and then use
this may not be the case for adjacent frequency-domain chan- this information along with sophisticated signal processing
nels [92]. Hence, when modeling the spectrum usage in the techniques to improve the performance of PUs, while also
frequency domain, the dependency among neighboring chan- obtaining some additional bandwidth for the CRs’ own trans-
nels should be taken into consideration. With respect to the mission. In interweave CRNs, the CRs sense the absence of
statistical correlation across the frequency band, two aspects a PU signal in the time, frequency, and/or spatial domains,
can be investigated through the analysis of real-world spectrum and opportunistically communicate during the PUs’ absence.
data. The first one is the DC distribution within an allocated Generally, underlay, overlay, and interweave can be used in
band, while the second is the cumulative distribution function different bands (e.g., TV white space, radar bands and cellu-
(CDF) of the DC distribution, which are closely matched by lar bands) according to the specific regulations in a separate
the Beta distribution, by the Kumaraswamy distribution and or hybrid manner [121], [122].
so on [89]. Most of the existing studies on spectrum inference focus
Another aspect to consider is the DC clustering across the their attention on interweave CRNs, i.e., on predicting the
frequency bands. In [111], the similar DC values of the adja- spectrum occupancy state in terms of being idle (i.e., the
cent channels were classified into the same group, and the absence of a PU signal) or occupied (i.e., the presence of
distribution of the size of the cluster group was described by an a PU signal), by licensed users. There are relatively few stud-
appropriately shifted version of the geometric distribution [96]. ies on spectrum inference in other types of CRNs such as
3) Spatial Domain Model: In recent years, numerous stud- underlay, overlay and hybrid CRNs. It is an interesting and
ies considered the relationships among the channel states fruitful research direction to extend the research from inter-
observed by CRs in different scenarios [112]–[114]. Making weave CRNs to other types of CRNs. Technically, a key
use of the location information beneficially supports the CR’s difference is that the inference of binary spectrum state (i.e.,
initial awareness of the environment and its ability of taking idle or busy) in interweave CRNs should be extended to the
actions efficiently [115], [116]. Based on the above-mentioned inference of multi-level or even continuous spectrum state val-
DC models in the time- and frequency- domains, the spatial ues in underlay and overlay CRNs, relying on the knowledge
DC models and the modelling of simultaneous observations of the channel state information between PU transmitters and
are briefly introduced below. Reference [89] discussed the spa- PU receivers.
tial DC models under different conditions, namely when con-
cerning either time-varying or constant power and continuous III. S PECTRUM I NFERENCE A LGORITHMS IN CRN S
or discontinuous transmitters.
The DC values, , are calculated after the process of energy Numerous prediction algorithms have been proposed for
detection. López-Benítez [89] also discussed the conditional forecasting the channel state, the spectrum occupancy, the
and the joint probabilities of the state observed at an arbi- potential interference imposed, the PU’s coverage area and so
trary location and the simultaneously observed state at the on. In this section, the common spectrum inference/prediction
reference location. Given these two probabilities, we will be algorithms found in the literature will be surveyed in the
able to model the spectrum usage at another location based time, frequency and spatial domains by adopting a multi-
on the model at the reference location. Therefore, the spatial dimensional approach. The taxonomy of these spectrum infer-
modelling procedure attaches importance to the correlation of ence algorithms is illustrated in Fig. 4.
the spectrum usage at various locations for supporting spec-
trum inference/prediction. However, it should be noted that A. Temporal Spectrum Inference Algorithms
generally spectrum measurement campaigns are conducted by The three main branches of Fig. 4 will be surveyed in
measuring the received signals and then estimating the sta- Sections III-A–III-C respectively, commencing with the tem-
tus of the PUs using various spectrum sensing techniques, poral techniques. In the time domain, spectrum prediction
Fig. 4. Taxonomy of spectrum inference algorithms in the literature (See Table VII for the acronyms).
infers the status of spectrum according to the historical infor- Reference [15] used a second-order autoregressive model, i.e.,
mation on the evolution of the spectral occupancy [128]. AR(2), for performing spectrum prediction in the time domain,
The information gleaned is represented by a series of num- where the coefficient αj has been obtained as the solution of
bers extracted from the original data, with the aid of linear the Yule-Walker equations. Then a Kalman filter has been
prediction, Markov modelling, neural networks, pattern min- used for predicting the state of a spectrum hole. In [130],
ing, etc. After acquiring the regular patterns of the PU, we the channel occupancy status is converted into a binary form
can sense the spectrum in less time than usual, conserve pre- and the AR model is used for spectrum occupancy charac-
cious energy and make collisions among PUs and CRs more terization as well as for prediction. Finally, artificial signals
infrequent. complying with the Global System of Mobile (GSM) commu-
1) Linear Prediction (LP): Let us commence with LP algo- nication standard were generated for testing the performance
rithms seen at the top of the temporal branch. Linear prediction of prediction. Furthermore, AR models have also been applied
is an important branch of mathematical statistics, where future for predicting both the channel state transitions in fading
values are predicted as a linear function of previous samples. channels [134], [135] and the channel occupancy [93].
The LP based approach is widely used in digital signal pro- b) MA model: The MA model is similar to the autore-
cessing for predicting the signal power as a benefit because gressive model, except that the predicted value for the obser-
of its remarkable simplicity, and it has also been invoked for vation in the t-th slot depends on the error values observed in
implementing temporal-domain spectrum prediction [15], [93], the past [136], rather than on the current values.
[128]–[130]. Other applications include speech- and audio- The MA model of order n, namely MA(n), can be formu-
compression [131] as well as video compression [132].The lated as [133],
most commonly used linear prediction models include the n
autoregressive (AR) model, the moving average (MA) model, X(t) = βi e(t − i), (5)
the autoregressive moving average (ARMA) model and the i=0
autoregressive integrated moving average (ARIMA) model. where, the notations X(t) and e(t) are the same as in the AR
a) AR model: With reference to the left column of the model, while βj is the parameter of the MA model and β0 = 1.
middle branch in Fig. 4, the AR model of order m, AR(m), In [137], when analysing the Chinese TV-band channels
can be formulated as [133], spanning from 603.25 MHz to 843.25 MHz, the results indi-
m cate that the MA model is more suitable for predicting the
X(t) = αj X(t − i) + e(t). (4) strength of television signals than the ‘experience based’
i=1 method. Additionally, in [129], a non-linear exponential mov-
In the spectrum prediction context, X(t − i) represents the ing average (EMA) model is proposed, whose weighting factor
past observation before the t-th slot, while X(t) and e(t) rep- decreases exponentially for each older data point, in order to
resent the observation
and error terms of the t-th slot. The put more emphasis on the recent observations, while giving
expression m α
i=1 j X(t − i) aims for weighting the historical some cognizance to former observations.
observations, where m is the length of the prediction memory c) ARMA model: Still referring to the left column of the
window and αj is the weighting parameter of the AR model. middle branch in Fig. 4, the MA model is combined with the
Fig. 5. Linear prediction (LP) structure.
AR model to form an autoregressive moving average model

ARMA(m, n) [16] as follows: Fig. 6. HMM prediction process.

m
n
X(t) + αj X(t − i) = βi e(t − i), (6)
i=1 i=0 the partially observable Markov decision processes (POMDP),
also featuring in Fig. 4.
where the notations of X(t) and e(t) are the same as in the AR
a) HMM: According to [141], an HMM is usually for-
and MA models. Furthermore, αj and βj are the parameters π , A, B), where π represents the initial state
mulated as λ = (π
of the ARMA model, β0 = 1. In [138], the ARMA model
probability vector, defined as:
shows a better prediction performance in channels which have
cyclic behaviors compared to the performance in forecast- π i ] = P(qt = si ), 1 ≤ i ≤ N,
π = [π (8)
ing the Bluetooth and WiFi channels that have more random
where S = {s1 , s2 , , sN } denotes N different states in a Markov
behaviors.
chain and qt ∈ S represents the state at time instant
N t, where
d) ARIMA modeling: The most important feature of the
πi satisfies the conditions 0 ≤ πi ≤ 1, i=1 πi = 1.
more flexible ARIMA model is that it is capable of handling
Furthermore, A = [aij ] is the state transition matrix, which
non-stationary time series as well as rapidly decaying auto-
defines the probability of traversing from one state to another,
correlation functions [139], etc. The general ARIMA(m, n, d)
formulated as:
model is formulated as [136],
aij = P qt = sj |qt−1 = si , 1 ≤ i, j ≤ N. (9)

m
n
W(t) = αj W(t − i) + βi e(t − i), (7) Furthermore, B = [bik ] is the observation probability matrix
i=1 i=0 that offers the option of producing different observed values,
where we have W(t) = X(t) − X(t − d), while αj and βi are while being in a particular state, which is encapsulated in:
the parameters of the ARMA model of order m, e(t) repre-
bjk = P ot = vk |qt = sj , 1 ≤ j ≤ N, 1 ≤ k ≤ M, (10)
sents random errors, and d is the degree of non-stationary
homogeneity. where V = {v1 , v2 , . . . , vM } represents the space containing M
The overall process of using these models is illustrated in observed symbols and ot is the value observed attime instant
Fig. 5 adopted from [128]. The samples X of the time series t, ot ∈ V . Note that we have 0 ≤ aij , bjk ≤ 1, N j=1 aij = 1

constitute the input of this procedure. As seen in Fig. 5, we and M a
k=1 jk = 1.
first judge the degree of stationarity for the input data, and if Traditionally, a standard HMM predictor is based on the
the samples are not deemed to be stationary, they will have to discrete-time model introduced in Section II, where the time
be processed by different techniques. Following this step, we is slotted and the state space is binary, with the spectrum
calculate both the covariance function as well as the partial slot being either in state si = 1 or si = 0. The set
autocorrelation function, and then classify X into a specific O = {o1 , o2 , . . . , oT } may be used for representing the past
model. Finally, we compute m and n. As seen in Fig. 5, we observation sequence in T consecutive slots. In order to predict
also have to estimate the parameters αj and βi . Finally, we use the state in the (T + 1)-st slot, a HMM predictor of the stan-
the selected model to forecast the prospective values. dard form may be designed by obeying the following four
2) Markov Model (MM) Based Methods: Let us now con- steps [20], [111], which are illustrated in Fig. 6 and detailed
sider the second column of the middle branch in Fig. 4. The as follows:
Markov chain based modelling of sub-band PUs was validated Step 1 (Initialization): Set the initial parameters of the
by analyzing real-world measurements in the paging spectrum HMM λ0 = (π π 0 , A 0 , B 0 ).
band in [140]. The Markov model has an appealingly sim- Step 2 (Observation): The CR senses the sub-band spectrum
ple structure characterized by its state-transition matrix. In and collects the observed data O = {o1 , o2 , . . . , oT }.
this subsection, we will discuss a pair of temporal prediction Step 3 (Training): Given the observed sequence, the param-
methods, namely the Hidden Markov models (HMM) and eters of the HMM will be adapted by invoking an appropriate
training approach, such as the Monte-Carlo method, and the spectral slot occupancy pattern is assumed to be deter-
Baum-Welch algorithm [141], for maximizing the likelihood ministically distributed, when the standard HMM model
associated with the model and defined as P(O O|λ). is used. However, in reality the PU’s behaviors obey
Step 4 (Prediction): The state observed at slot (T +1), oT+1 , time-varying non-stationary DTMC [89]. Under these cir-
can be predicted by the following rule: cumstance, conventional HMMs may fail to adequately
characterize the PU’s dwelling time distributions [17]. For
1, if P(OO, 1|λ) ≥ P(OO, 0|λ),
oT+1 = (11) this reason, the NS-HMM was employed for modeling
0, if P(OO, 1|λ) < P(OO, 0|λ).
the channel’s status in [17] in order to characterize
This standard HMM approach has been investigated the idle/busy duration of the PU’s channel. The NS-
in [14], [20], [95], [111], [142], [143], and [140] for the sake HMM can be described as λNS = [π π , A (t), B ], which
of predicting the future states and the idle/busy durations of the replaces the static transition probabilities of the conven-
channel. In [95], the HMM is used for predicting the spectrum tional HMMs with dynamic ones. In [17], the parameters
usage patterns, which are assumed to be deterministic Poisson of the NS-HMM are inferred through Bayesian inference
distributed. Given a sequence of predicted states, the CR will with the aid of Gibbs sampling, and the impact of differ-
dynamically choose these surmised frequencies for its use, ent model parameters on the model’s accuracy has been
even though the spectrum occupancy should generally be mod- investigated. It was found that as expected the channel
eled by stochastic distributions, as previously mentioned in quality experienced by the CR is an increasing func-
Section II. This algorithm was also advocated in [142], match- tion of the sensing accuracy and of the estimated idle
ing its output to the data collected from the 450-470 MHz duration. In [118], the so-called expectation maximiza-
band in Australia, but the specific implementation of the algo- tion based algorithm was developed for calculating the
rithm was not detailed in [142]. Meanwhile, Chen et al. [14] parameters of a NS-HMM. Numerical experiments rely-
considered a realistic propagation environment that took into ing on real spectrum measurement data have been carried
account the time delay incurred both by the hardware and out for demonstrating that the NS-HMM outperforms the
software. It should be mentioned that in [14] the parameters traditional HMM-based approaches.
of the HMM used are obtained statistically, without employ- • Hidden Bivariate Markov Model (HBMM): As an
ing any training for the model. Instead, the authors verify the extension of the standard HMM, HBMMs have been
proposed method using measured WiFi data instead of artifi- proposed for more accurately characterizing the transmis-
cial data. Both the model complexity and the computational sion pattern of a PU [144]. In contrast to the standard
complexity have been studied in order to provide a qualitative univariate HMM, the HBMM incorporates a number of
characterization of the HMM performance, which were com- additional variables by introducing a state duration dis-
plemented by a range of implementation issues arising during tribution for modeling the channel usage in CR, rather
the design [118]. As the number of states N increases, the than using the geometric duration distribution of a stan-
number of model parameters to be estimated also increases. dard HMM. In [144], the HBMM was described in form
The computational complexity of the HMM based predictor is of the function λB = (π π , G , μ , R ), where π is the initial
directly related both to the number of states N and to the size state probability vector reminiscent of that in the standard
of the observation sequence T. HMMs, G is the state transition matrix, while μ and R
Moreover, MM or HMM-based methods work well under represent the vector of observed receive signal strength
the assumption of having memoryless or Markovian spectrum average and variance. Repeatedly using the definitions
state evolution, where the future state depends only on the rel- of the standard HMM parameters, let Z = O × S =
evant information about the current, not on information from [Zt ] = (ot , qt ) denote the specific value returned by the
distant the past. The Markovian property has been validated bivariate Markov chain. The state transition matrix can
by analyzing real-world measurements in the paging spectrum be defined as:
band in [140], which motivates the research of HMM-based

spectrum inference methods. Most of the related studies have G = gab (ij) = P(Zt+1 = (b, j)|Zt = (a, i)) . (12)
focused on first-order and stationary HMM. However, in prac-
tical systems, the spectrum state of the future may depend on The parameter λB may be estimated by extending the
a relatively long historical information (see the tide effects Baum algorithm [145], which is simpler than an explicit
in the GSM bands [209]). The state distribution of PUs may duration estimating algorithm. When determining the
also dynamically change. In order to improve the performance model parameters, Thao et al. [144] apply forward-
of the standard HMM, in recent years researchers have devel- backward recursions for predicting the spectral slot state
oped several sophisticated HMMs by carefully considering the at a future time instant. The performance of the HBMM
stationarity and the high-order Markovian nature of the spec- in spectrum sensing and prediction was characterized
tral slot occupancy pattern. Some representative advances are by numerical results in [144], demonstrating that the
summarized below as follows: HBMM is capable of more accurate state predictions than
• Non-stationary Hidden Markov model (NS-HMM): a standard HMM.
Here we are still considering the Markov-modeling • Higher-order HMM: Still referring to Fig. 4, we note
belonging to the second branch of the temporal algo- that although the first-order standard HMM has been
rithms seen in Fig. 4. As mentioned above, the PU’s broadly adopted for predicting the channel’s state, this
model does not make full use of the historical infor-

mation, since a state only depends on the immediate
preceding state. To alleviate this problem, a higher-order
HMM was also proposed for predicting the next channel
state [146]. A higher-order HMM can be formulated as
λHO = (π π , A HO , B ). In contrast to the standard HMMs,
A HO = [aHOij ] is the state transition matrix representing
the transition from the states in the previous D slots to
the current state, which can be defined as [146]:

A HO = aHO ij

= P qt+1 = siD+1 |qt = siD , . . . , qt−D+1 = si1 ,
i1 , i2 , . . . , iD , iD+1 = 1, . . . , N.
When considering the computational complexity versus
model accuracy trade-offs, there are several variants of
the higher-order HMMs [146]–[149]. By relying on real- Fig. 7. POMDP based prediction.
world WiFi signals recorded, it was shown that the
performance of the proposed approach is significantly bet-
ter than that of the nearest neighbor prediction relying on state s is idle can be calculated by the Bayesian formula as
first-order HMMs, especially when the order of the HMM follows

increases [146].

S O ot+1 |s , a T s |s, a b(s)
b) POMDP: The last member of the MM-based tem- b (s ) = P s |a, b, o = .
P(ot+1 |b, a)
poral methods seen in Fig. 4 is constituted by the POMDP (13)
family, which is a generalization of a Markov decision pro-
cess that tolerates uncertainty about the state of a Markov The numerical results provided in [152] and [156] have
process and allows the acquisition of environmental infor- shown that the performance of the spectrum access approach
mation [150]. Based on DTMC model of Section II [89], using the POMDP based CR improves over time upon glean-
a POMDP applied in CRNs is defined by the six-tuple ing increasingly more accurate information concerning the
(SS , A , T , R , Z , O ) [151]–[154], where channel’s state inferred from the accumulated observations.
S is the channel state space {0(idle), 1(busy)}; As a benefit, the collisions of CRs with PUs were substan-
A is a discrete and finite set denoting the CR actions, i.e., tially reduced. However, a problem concerning the POMDP
as {a1 (access), a2 (no access)}; based approach is its potentially excessive complexity, espe-
T : S × A → (SS ) is the state-transition function. If s cially when the number of channels is high. To tackle this
is the current channel state and action a is chosen by the problem, more research is required for developing efficient
decision maker, the process will move to a new state s with algorithms for solving POMDP problems.
the probability of T (s |s, a). 3) Artificial Neural Network (ANN) Based Methods: Let
R : S × A → R is the reward function that defines the us now move on to the third family of the temporal meth-
expected immediate reward R (s, a) received, when the process ods portrayed in Fig. 4. As a key technique in machine
is in state s and action a is taken; learning, the ANNs have been extensively applied to com-
Z is a finite set of observations the CR infers from the radio munications, signal processing, intelligent control, etc. They
environment; represent a class of flexible nonlinear models based on par-
O : S × A → (Z Z ) is the observation function and for allel computation, which consist of both input and output
each action and resultant state, O(s |a, o) denotes the prob- layers as well as intermediate layers composed of artificial
ability of observing o and moving to state s after taking neurons [157]. In CRNs, an ANN allows for learning the PU’s
action a; behavioral patterns by invoking a training set of spectrum data,
The POMDP based CR engine [155] is depicted in the learn these patterns, and then capitalize on it for classifying
Fig. 7, with the state predictor being the core component in new patterns and for making forecasts. Because of its high
the process [151], [153]. In this procedure, b is the ‘belief prediction accuracy and capability of learning, ANNs con-
state’, which describes the probability that the channel’s state stitute a popular approach of predicting the spectrum state’s
is idle, summarizing all the past information necessary for evolution. Generally, depending on their architecture, we can
formulating the allocation decision π [154]. The channel state categorize ANNs into two classes: feedforward and feedback
predictor computes the probability b based on a combination neural networks [158]. The architecture of ANNs invoked for
of the most recent value, on the current observation and on the predicting channel states is depicted in Fig. 8.
previous action. More specifically, Zhao et al. [153] proposed a) Feedforward neural networks (FFNNs): Feedforward
an approach for channel state prediction based on POMDP by neural networks enter the spectrum data strictly into the input
finding the optimal policy that maximizes some aspect of the layer and give the states predicted by the output layer without
reward. In this approach, the probability b (s ) that the future any feedback. Referring to the rich literature concerning FFNN
Once the training process is completed, the future spectrum

state can be predicted by capitalizing on the observations with
the aid of the MFNNs. In [20], the predicted value yot+1 is
decided by invoking a threshold at the output layer, which
can be expressed as follows:

0, if yot+1 ≥ 0,
yt+1 =
o
(19)
1, if yot+1 ≤ 1.
It is noteworthy that the learning speed of the traditional
FFNNs, like MFNNs, is generally slower than desired, which
would naturally evolve the attainable prediction efficiency.
Fig. 8. Artificial neural network (ANN) based prediction. This problem was solved by the extreme learning machine
(ELM) of [177]. Furthermore, Miche et al. [159] optimized
the original ELM, while in [158] the optimally pruned extreme
based algorithms invoked for spectrum prediction [20], [157], learning machine (OP-ELM) was used for spectrum prediction.
[159]–[168], we will first provide a brief overview of typical OP-ELM based spectrum prediction also relies on the struc-
multilayer feedforward neural networks (MFNNs), also known ture of single hidden layer neural networks (SLFNs), just
as the multilayer perceptron (MLP). Then we will describe the like the original ELM, where the input weights are randomly
overall spectrum prediction procedure of neural networks. assigned, but the output weights are calculated using a dif-
Based on the PU’s activity modeled by DTMC, where the ferent procedure. According to the experiments conducted
input and output data are represented by binary sequences, in [158] and [159], the OP-ELM algorithm is more robust
MFNNs rely on neurons as their basic computing unit for and flexible than the original ELM, showing a higher learning
calculating a weighted sum of the inputs and perform a non- speed and a higher accuracy compared to other FFNNs.
linear transformation of the sum [20], [160], [168], which can b) Feedback neural networks (FBNNs): The only differ-
be expressed as follows [20], [168]: ence between FBNNs and FFNNs is the presence or absence
of feedback links that span from the outputs of neurons to
ylj = f alj , (14) the inputs of neurons in the previous layer [178], [179], as
depicted in Fig. 8. The FBNN requires the data to be passed
alj = yil−1 wlji , (15)
both forward as well as backward, hence imposing a high com-
i
plexity and often making the FBNNs ‘confused’ or unstable. In
where ylj is the output of a neuron j in the lth layer; wlji repre- this sense, the forecasting performance of FBNNs may become
sents the adaptive weights connecting the neuron j in the lth less attractive than that of FFNNs [180]. FBNNs gleam infor-
layer and neuron i in the (l − 1)st layer; alj is the weighted mation from the sequence or time dependence of the inputs and
sum of the inputs yielding the outputs of the former layer, the inputs themselves, which means that the features detected
and f (alj ) is the activation function, often represented by the in previous patterns form a part of each new pattern.
tangent sigmoid function [160]: Therefore, when we want to take the previous spectrum
2 observations into consideration in order to sense or predict
ylj = − 1. (16) the spectrum, the FFNNs discussed above are less pragmatic
1 + e−2aj
l
for us. In [178], Elman recurrent neural networks (ERNN) is
Before performing spectrum prediction, there should be a used for radio frequency (RF) multivariate time series mod-
training process for MFNNs. First, the observed spectrum elling in order to predict the spectral evolution. This feedback
data vectors X (t − τ + 1, t) = {xt , xt−1 , . . . , xt−τ +2 , xt−τ +1 } method leads to more intelligent CR decisions for exploit-
are imported into the input layer and the estimated output ing the expected spectrum opportunities, thereby leading to
data yot+1 is calculated using the activation function. Then, the optimized spectrum usage and interference avoidance [178].
difference between the desired value xt+1 and its estimation Although ERNN is powerful in predicting the time domain
yot+1 is denoted as the error et , which can be expressed as sequences, it has slow convergence and does not excel in
follows [20], [160]: terms of determining the network structure [181]. The Echo
State Network (ESN) concept is proposed as a new tech-
et = xt+1 − yot+1 . (17) nique of overcoming these problems in [182]. Taking this as
Finally, in order to minimize the prediction error et , the a basis, Yang et al. [183] further proposed a method based
adaptive weight wlji will be updated by repeatedly using the on an improved form of ESN for spectrum prediction. The
back propagation algorithm [169] until the prediction error proposed ESN structure is also visualized in [183]. In contrast
reaches its minimum. In [160], the prediction error is measured to the ERNN, the intermediate hidden layers are replaced by
by the mean squared error (MSE), denoted as E(w), which can a cycle reservoir associated with fixed feedback connections
be computed as follows [160]: and randomly connected neurons [183]. The parameters of
the improved ESN are calculated using particle swarm opti-
1 2 1 2 mization (PSO). It should be noted that the PSO is known
E(w) = et = xt+1 − yot+1 . (18)
2 t 2 t to outperform the traditional back propagation algorithms in
training ANNs, as a benefit of their faster convergence and based both on the observations and on the a posteriori distri-
lower complexity [165], [184]. Experiments have been con- bution. BIF-based prediction is proposed in [17] and [187] for
ducted in [183] to demonstrate that the performance of the estimating the distribution of idle durations in the PU’s traffic.
optimal ESN is better than that of the ERNN in predicting both The algorithm proposed in [17] combines BIF with NS-HMM
the specific state and the state duration of the primary spec- to predict the channel quality, which is capable of improv-
trum, since it reduces the training time required for obtaining ing the efficiency of dynamic spectrum access. Compared to
a higher prediction accuracy. the traditional maximum likelihood technique, the BIF tech-
4) Pattern Mining (PM) Based Methods: PM is a classic nique is capable of performing better, when processing a
data mining technique conceived for discovering underlying limited number of samples, and the channel availability may
patterns in large amounts of data. In CRNs, the patterns usu- be reliably predicted after estimating the prevalent spectrum
ally reflect the rules hidden in the spectrum occupancy of a usage patterns [188]. In [157], a LA technique is utilized
channel. The motivation of finding the rules arises from the for predicting the spectrum opportunities in CRNs. The LA
desire to analyze historical spectrum data for predicting the technique first generates the PU’s activity model according
future channel states, so as to improve the attainable spectral to the characteristics of the PU’s behavior. Then, the training
efficiency and hence to reduce the collision rate. In previous and testing stage updates the parameters of the models, and
research, the spectrum usage prediction has relied on frequent finally the performance of the model is assessed in order to
pattern mining (FPM) [119] and partial periodic pattern min- correct the prediction of the system’s behavior. In comparison
ing (PPPM) [16]. Both methods are based on a binary channel to the MFNNs, the LA has shown a high performance, despite
state model. its simple structure [157].
• FPM: In the context of CRNs, the channel state val- The SVM technique has been widely used to make time
ues observed are delivered in a sequence and serve as series predictions in many scientific fields, such as finan-
the spectrum occupancy history database. The mission cial marketing, power supply and medical sciences [189] as
of FPM is to find the specific sequences that appear in well as for spectrum prediction in CRNs. The application of
the data set with a frequency above a threshold, which SVM to solve regression problems is termed as support vec-
are referred to as a frequent sequential patterns [185]. tor regression (SVR). In [190], SVR is used for predicting the
Thus, it becomes possible to predict the next state or next probability density of the idle/busy CR state duration. In [191],
sequence. Although the FPM of [119] is two-dimensional, SVR is applied for processing the spectrum occupancy data.
its ability to predict the future spectral states is only
known in the time domain. B. Spectrum Inference in Frequency Domain
• PPPM: In contrast to FPM, the partial periodic patterns Section III-A has reviewed some of the popular techniques
observe the PU’s features during certain time periods conceived for predicting the future spectrum occupancy in the
instead of the entire time span [186]. Since the prop- time domain. Further studies are well dedicated to inferring
agation environment is stochastic and the PU’s behav- the states of other channels in the light of already acquired
iors may not show an obvious regularity. The patterns sensing or predicted results [93], [119], [130], [192], [193].
only become relevant for certain time periods of the The spectral occupancy correlation between adjacent channels
day [16]. The procedure of PPPM based prediction can has been evaluated by experiments in [194]–[196]. This cor-
be described as follows [16]: relation serves as the most important information representing
Step 1: Process the binary channel state values and the relationship of different channels within the same services.
generate candidate patterns; In other words, the more correlated two channels are, the more
Step 2: Count the support of each candidate pattern; accurate predictions will be made. Yin et al. [192] exploited
Step 3: Eliminate the redundant candidate patterns; the spectral correlations for inferring the availabilities of other
Step 4: Generate the prediction rules and calculate the channels in order to improve the throughput. The new concept
probability of the future channel states. of channel availability vector was introduced for characterizing
Additionally, the algorithm based on PPPM was also used the spectral occupancy information.
for predicting the channel state duration in [16] so as to In [193], predicting the states of unsensed channels is for-
reduce the collision probability owing to the PU’s short mulated as a matrix completion problem. The classic technique
absence. of belief propagation (BP) [197] is applied to fill the matrix
5) Other Algorithms: Other relevant temporal methods with predicted states. It is found that the BP scheme is more
found in the literature include Bayesian inference (BIF), learn- suitable for specific matrix types, where the adjacent elements
ing automata (LA) and support vector machines (SVM). BIF are highly correlated [29]. In [198], Bayesian networks are
is known as a classic method of solving the state prediction proposed for jointly modeling the spectral-domain and spatial-
problem in CRNs as a critical part of the Markov systems, domain correlations, where the authors introduced statistical
like the aforementioned NS-HMM and POMDP. It provides inference for evaluating the spectrum occupancy. The correla-
a general unifying framework for sequential state estimation across the frequency dimension was exploited to enhance
tion. The procedure of BIF based prediction continuously the estimation of spectrum occupancy in wideband spectrum
updates the a posteriori distribution in a recurrent manner, sensing [195], [196].
based on the influx of spectrum data. Assuming that the evo- Yin et al. [119] has proposed a frequency pattern match-
lution of the state is Markovian, BIF can make predictions ing algorithm operating in two-dimensions (FPM-2D) for
spectrum inference, which searches through all relevant 2D mitigating the performance loss caused by high-speed CRs
patterns [119]. Once these patterns have been obtained, we having erratic movements. In [203], the SVM was invoked
can compute the probabilities of future channel states and both for predicting the handoff point and the idle channels
estimate the channel’s availability. Their algorithm outper- with a high precision. In addition to the spectrum availabil-
forms the first-order HMM-based predictor in terms of its ity, the link availability between the CRs was also inferred
prediction accuracy. Similarly, frequency domain correlation by Hou et al. [204]. By invoking this novel approach, a more
techniques were also introduced in [93] by modeling the neigh- reliable path can be found for dynamic routing in CRNs.
boring channels in pairs. Simultaneously considering multiple 2) The Second Direction: For the TV services, the primary
frequency bands, the classic AR models which do not need receivers must be protected from harmful interference [18].
a priori knowledge on the communication environment were This approach makes full use of the effective antenna height,
also employed to reduce the complexity [93]. This algorithm effective isotropic radiated power and terrain information to
performs well for deterministic usage patterns. map the primary users’ coverage contour through Fresnel
The investigations of [93], [119], [192], and [196] clearly diffraction theory. The Google Earth software plays an indis-
quantify the performance improvements achieved by spectral pensable role in TV white space prediction. Based on the
correlation based inference. primary users’ coverage contour inference, the potential
interference and the collisions imposed by the CRs may be sig-
nificantly reduced. In [205], a location predictor is proposed,
C. Spectrum Inference in the Geographic Spatial Domain where the historical changes of the PU’s geographic locations
From a spatial-domain perspective, there are two directions are represented by a directed graph having weighted edges.
for spectrum inference: Once a spectrum occupancy prediction is requested, all the
• One of them aims to infer the spectrum state information, edges originating from the starting point are listed and then the
while taking both the positions and movements of CRs destination is predicted according to the calculated maximum
into consideration; weights.
• The other one infers geographical information, including
the primary system’s coverage contours, service areas and D. Spectrum Inference in Joint Multiple Domains
so on.
In addition to spectrum inference techniques in single
1) The First Direction: In the first direction, both the
domain (i.e., time, frequency and space) mentioned above,
BP technique and the Bayesian networks discussed in the
there are also recent studies on joint multiple-domain spec-
frequency domain are also capable of taking the spatial corre-
trum inference. Specifically, the work in [119] is one of
lations into account for inferring the channel states from those
the first studies to develop joint time-frequency spectrum
of other locations.
prediction algorithm, where two-dimensional frequent pat-
In the static spectrum environment, the interference powers
tern mining is proposed to analyze binary historical spectrum
experienced at different CR nodes are inferred by exploiting
occupancy data. Ding et al. [210] develop an algorithm for
the spatial variation of interference. Specifically, both semi-
spectral-temporal two-dimensional spectrum prediction with
supervised dictionary learning [200] and compressive sensing
incomplete historical data by exploiting the approximate low-
are employed for the interpolation of unobserved interference
rank property of real-world spectrum measurement matrices.
in the spatial domain, incorporating the CR network topology.
In [211], a robust spectral-temporal two-dimensional spectrum
Since the movement of CRs directly affects the spectrum
prediction algorithm is proposed under an assumption on the
availability in a specific geographic area, spectrum prediction
sparsity of abnormal or corrupted historical data. In [212],
in the time-space domain constitutes and important research
a matrix completion-based algorithm is developed for TV
area. The traditional static model is extended to a dynamic
white space database construction via joint spectrum sens-
mobility model in [201]. The CRs’ mobility and the PUs’
ing in time domain and spectrum inference in spatial domain.
activity can be jointly considered to infer both the spec-
Tang et al. [213] propose the concept of spectrum tensor and
trum occupancy and the interference constraints for a certain
develop a multi-dimensional (including, time, frequency, and
period under different relative positions between the PUs
space) spectrum inference algorithm for spectrum map con-
and CRs. Based on this scheme, both a greedy and a fair
struction by invoking the recent advances in tensor completion.
prediction algorithms were proposed in [201] to make the
Moreover, the recent tutorial papers [214], [215] propose to
spectrum exploitation more effective, hence maintaining fair-
combine spectrum sensing and spectrum inference into spec-
ness in spectrum allocation. A SVM is used for inferring
trum database for providing spectrum services for users, and
the spectrum occupancy evolution considering both the time-
also highlight that joint multiple-domain spectrum inference
domain and geographic distribution characteristics [202]. A
is an active research trend.
joint feature-vector extraction method is designed by analyz-
ing both the CRs’ movement and the PUs’ behavior. Again,
a SVM based inference mechanism is introduced for expedit- E. Summary and Insights
ing the convergence speed, which is shown to have a better As discussed above, there are a wealth of contributions
inference performance than the algorithms solely depending on spectrum inference. In this subsection, we offer a brief
on the speed and location information, as in [202]. It was summary and discussions of the existing spectrum inference
found that carefully choosing the parameters is capable of algorithms. First of all, the major developments of spectrum
TABLE II
M AJOR D EVELOPMENTS OF S PECTRUM I NFERENCE T ECHNIQUES (PART 1) (T/E: T HEORETIC /E XPERIMENTAL
W ORK ; U/O/I: U NDERLAY /OVERLAY /I NTERWEAVE )
inference/prediction techniques are summarized at a glance in Furthermore, there are also several major insights based
Tables II and III in the chronological order. Then, the advan- on the comprehensive survey of the existing studies detailed
tages and technical challenges of various spectrum inference as follows. Firstly, while the majority of the existing stud-
algorithms are summarized in Table IV, according to a priori ies focused on spectrum inference in the time domain,
information concerning the processed data and to the resultant in the past decade, spectrum inference in joint multiple
complexity and accuracy constraints. domains has been gaining increasing research interests.
TABLE III
M AJOR D EVELOPMENTS OF S PECTRUM I NFERENCE T ECHNIQUES (PART 2) (T/E: T HEORETIC /E XPERIMENTAL
W ORK ; U/O/I: U NDERLAY /OVERLAY /I NTERWEAVE )
TABLE IV
D ISCUSSIONS ON A DVANTAGES AND C HALLENGES OF VARIOUS S PECTRUM I NFERENCE A LGORITHMS
For instance, Ding et al. [209]–[211] proposed various (time, frequency, and space) spectral map construction
joint two-dimensional spectral-temporal spectrum prediction method.
algorithms by leveraging the low-rank nature of the spec- Second, there are some quantitative comparisons among
tral data. Tang et al. [213] developed a multi-dimensional different spectrum inference techniques. Specifically, it was
reported in [20] that under the same traffic scenario, an ANN however, suffers from the hidden node problem because of
based predictor performs slightly better than the HMM predic- shadowing [239]. In the past decade, intensive studies have
tor owing to having a flexible number of states and a more effi- been carried out across the globe to improve the detection
cient training mechanism. Although having more intermediate performance of spectrum sensing. Excellent survey papers
layers in the ANN increases their computational complex- in this direction can be found in [25]–[27]. Briefly, exist-
ity, the prediction results are improved [157]. As expected, ing studies on spectrum sensing in CRNs can be classified
there is a tradeoff between the complexity and the accuracy. into two groups: quiet-periods-based (see [26]) and non-quiet-
Moreover, it is reported in [139] that the short-term prediction periods-based spectrum sensing (see [288]), also known as
performance of the ANN based predictor is better than that of spectrum monitoring in the literature. As the name implies,
the ARIMA predictor, but their performance degrades, when in the former studies, a cognitive radio first spends a time
predicting further into the future. The performance of the duration (known as quiet period) to perform spectrum sens-
HMM predictor suffers from two inherent limitations [218]. ing and then determines whether to transmit based on the
Since the radio environment is changing all the time, the fixed sensing results. By contrast, spectrum monitoring in CRNs
states assumed in Markov modeling fail to capture the rich is a relatively new technique, where the cognitive radios can
temporal variations of the channel activities. Another limita- continue their communications while simultaneously monitor-
tion of the HMM aided predictor is that discrete or univariate ing the band to detect any transmissions that are initiated by
Gaussian models of observations often fail to adequately char- primary radios. More specifically, the spectrum is monitored
acterize the general situations. The POMDP predictor gives by the cognitive radio receiver during reception and without
full cognizance to random actions, to incomplete informa- quiet periods [287]–[289]. So far, spectrum sensing has not
tion and to noisy observations of the environment, but it may been widely accepted by regulatory bodies to ensure non-
not be able to accurately calculate the probabilities of spe- harming the primary/licensed users. It is challenging to meet
cific action outcomes [219]. When it comes to pattern mining the strict rules required by the regulators from FCC, such as
based schemes, the partial periodic pattern mining predictor the detection of a primary signal at -114 dBm by FCC.
is more robust than the HMM predictor owing to extracting
more channel state prediction rules, which reduces the prob- C. Relationship With Geolocation Spectrum Database
ability of spectrum usage prediction compared to traditional
By contrast, according to the regulations of
frequent-pattern mining [16].
FCC [240], [241], the geolocation spectrum database
IV. A LGORITHM A NALYSIS OF S PECTRUM approach seems to provide a technically feasible and com-
I NFERENCE IN CRN S mercially viable solution in the near future. This approach
A. Performance Metrics provides a service ensuring that an unlicensed device can
inquire the spectrum availability from a geolocation database,
When evaluating or choosing a prediction algorithm, care- which predicts the spectrum availability at any location using
fully selected metrics should be considered. Typically the propagation modeling combined with terrain data [242], [246].
notion of detection probability (Pd ) representing the correct Based on the guidelines provided by FCC [240], several
prediction of the busy states and the false-alarm probabil- TV white space database systems have been developed by
ity (Pfa ) representing the rate of incorrect prediction of idle companies [247]–[250]. A specific limitation of this approach
states are used for studying the performance of algorithms in is that the accuracy of the spectrum availability provided
spectrum occupancy prediction. A combination of a higher Pd by the geolocation database depends crucially both on the
and lower Pfa suggests a better prediction performance. Some quality of the propagation modeling and on the granularity of
other statistical criteria like the normalized mean square error the terrain data. To resolve this issue, a data-driven approach
(NMSE), the mean absolute percentage error (MAPE), and is presented in [5] to build a spectrum database by learning
the root mean square error (RMSE) can also be introduced to the spectrum availability from mobile crowd sensing and big
quantify the adequacy of prediction. In order to demonstrate spectrum data analytics. Another limitation of this approach
the advantage of using spectrum prediction, both the percent- is that the update speed of the spectrum availability provided
age improvement in spectrum utilization and the percentage by the geolocation database is relatively slow (e.g., in the
reduction in sensing energy can also be invoked. time scale of hours or days).
Spectrum inference/prediction is an efficient technique com-
plementary to other techniques, such as spectrum sensing,
D. Relationship With Radio Environment Maps
geolocation spectrum database, and radio environment map,
for capturing the relevant information about the spectral evo- The radio environment map (REM), concept was originally
lution and identifying spectrum access opportunities. In this proposed by scientists from Virginia Tech [115], [251], [252],
subsection, we first briefly present the key ideas of each tech- which is a promising tool that provides a practical realization
nique one by one. Then, we discuss the relationships between of CRNs, explicitly, it constructs a comprehensive map of the
spectrum inference and the other techniques. CRN by utilizing multi-domain information from geolocation
databases, characteristics of spectrum use, geographical ter-
B. Relationship With Spectrum Sensing rain models, propagation environment and regulations [253].
Spectrum sensing determines spectrum state in a passive In contrast to the geolocation spectrum database, REM is
manner using various signal detection methods. This approach, an advanced knowledge base that stores live multi-domain
information on the entities in the network as well as on the network controL [293], is a project funded by the
environment [254], [255]. European Commission’s Horizon 2020 Programme with
Spectrum inference is an efficient technique complementary the following objectives: i) to offer open, flexible and
to the above techniques. Spectrum inference, with spec- adaptive software and hardware platforms for radio con-
trum prediction in the time domain as a special case, infers trol and network protocol development; ii) to offer
an unknown spectrum state from known spectrum data, by portable facilities that can be deployed at any location;
effectively exploiting the inherent statistical correlations of iii) to attract and support experimenters for wireless
spectrum data in time, frequency and spatial domains. A innovation.
key distinguishing feature of spectrum inference is that it • CREW, which stands for Cognitive Radio
presents a proactive view of the spectrum state. In terms Experimentation World [294], is a project with the
of the relation with spectrum sensing, spectrum inference main target of establishing an open federated test plat-
can reduce the sensing time and energy consumption [127]. form, which facilitates experimentally-driven research
Spectrum inference (e.g., in time domain) can also be fused on advanced spectrum sensing, cognitive radio and
with spectrum sensing to improve the detection performance cognitive networking in view of both horizontal and
of spectrum access opportunities [12]. In terms of the relation- vertical spectrum sharing in licensed and unlicensed
ship with a geolocation spectrum database, spectrum inference bands. Parts of the research outputs of this project focus
(e.g., in spatial domain) can be further used to calibrate the on constructing radio environment maps using various
propagation models and to improve the update speed of the spatial inference techniques [295].
database. In terms of its relationship with REM, spectrum • FARAMIR, which stands for Flexible and spectrum-
inference in multiple domains (i.e., time-frequency-space) can Aware Radio Access through Measurements and
be used to construct dynamic interference maps for each modeling in cognitive Radio systems [296], is a project
frequency at each location of interest [253], in a proactive making CRs a reality, with the main goal of develop-
and energy-efficient manner. ing techniques for increasing the radio environmental and
spectral awareness of future wireless systems. Parts of
V. E XPERIMENTATION -BASED A PPROACHES IN CRN S the research outputs of this project focus on constructing
radio environment maps using various statistical inference
Although the majority of the existing spectrum inference
techniques [297], [298].
algorithms are mainly studied from a theoretical perspective,
there are some studies that aim for bridging the theory and
the practice. In this section, we present discussions regard- B. Predictability Analysis of Real Spectrum Measurements
ing real experimentation-based spectrum inference approaches
With respect to spectrum inference/prediction, there is a
in CRNs. We begin with presenting several well known
fundamental question: To what degree is the spectrum state
spectrum-related experimental projects, which serve as the
evolution predictable? In other words, the predictability of
hardware/software basis for the development of spectrum
spectrum evolution reflects the upper-bound performance of
inference algorithms. Then, we briefly comment on the anal-
all potential spectrum inference/prediction techniques. When
ysis of real-world spectrum measurements. Furthermore, we
considering the temporal spectrum usage evolution for a given
discuss the experimental results reported in the literature.
frequency band in a given location as a random time series,
the entropy is the most fundamental quantity characterizing the
A. Spectrum-Related Experimentation Projects degree of predictability for this random variable. In general,
During the past two decades, quite a few spectrum-related having a lower entropy implies a higher predictability and vice
experimental projects have been carried out all over the world, versa. Entropy-based analysis has indeed been introduced in
which collected empirical real world spectrum measurements diverse prediction scenarios, such as for the atmosphere [6], in
and supported the development of spectrum inference tech- finance [7], for network traffic [8] and for human mobility [9].
niques. Excellent surveys and tutorials on the latest advances The basic idea is that the entropy offers a precise definition of
of worldwide spectrum projects and occupancy measurements the information content of predictions and it is renowned for its
can be found in [23] and [24]. In the following, we pro- generality due to relying on minimal assumptions concerning
vide a brief review of several well-known spectrum occupancy the model of the scenario studied.
evaluation projects to assist the readers. Recently, in [127], we have invoked an information theoretic
• ORCA, which stands for Optimization and Rational use methodology of using statistical entropy measures and Fano’s
of wireless Communication bAnds [290], is a project with inequality to quantify the degree of predictability underlying
the general scientific objective of extensively evaluating real-world spectrum measurements. We found that despite the
the spectrum occupancy in order to study the potential apparent randomness, a remarkable 90% predictability may
for its exploitation by innovative wireless services. This be achieved, in real-world RSS dynamics over diverse spec-
project developed DTV coverage prediction algorithms tral bands for several popular services, including the classic
by exploiting for example the Longley-Rice propagation TV bands, cellular bands, the (Industrial Scientific Medical)
model [291], [292]. ISM bands, etc. Similarly, Olivieri et al. [123] have applied the
• WiSHFUL, which stands for Wireless Software and information theoretic entropy as a measure of predictability in
Hardware platforms for Flexible and Unified radio and the process of generating the ON- and OFF-period durations.
Wellens et al. [124] and Riihijarvi et al. [125], [126] have used
state-of-the-art multi-scale entropy metrics in order to examine
the predictability of the spectrum measurement traces recorded
as a function of the prediction complexity.
C. Experimental Spectrum Inference Approaches

As shown in Tables II and III, although the majority of the
existing spectrum inference algorithms are predominately stud-
ied from a theoretical perspective, there are some studies in
the open literature that implement and test spectrum inference
approaches from an experimental perspective.
To the best of the authors’ knowledge, [140] is one of the
earliest studies of this kind, where the Markovian property of
the spectrum usage has been validated by analyzing real-world
measurements in the paging spectrum band (928-948 MHz).
Experimentally, the HMM is used for predicting the ON-OFF Fig. 9. Illustration of the applications and benefits of spectrum inference in
CRNs.
activity of the PU with a prediction accuracy of about 76%.
Similarly, Chen et al. [14] verify the HMM method using
measured WiFi data instead of artificial data, where the numer- context of: i) 5th generation mobile communications (5G)
ical results show that the prediction accuracy for a single spectrum sharing, ii) next-generation high-frequency (HF)
device is above 70 %, which can be improved by cooperative automatic link establishment (ALE), and iii) several other
prediction of multiple devices. Melián-Gutiérrez et al. [111] potential applications in various future wireless networks,
train and validate the HMM model with the aid of real mea- such as cognitive smart grid networks, cognitive radio sensor
surements collected from the 14 MHz HF amateur band, where networks, cognitive cellular networks (CCN), and cognitive
an average of 10.3% prediction error rate is achieved. machine-to-machine communications (CM2M).
In [162], a multi-layer perception ANN is applied to fore- Based on the spectrum management framework proposed
cast the idle or busy state of different channels, where the in [4], the existing applications of spectrum prediction in
performances are evaluated for a seven-day spectrum dataset CRNs can be classified into four basic categories, namely,
collected in a metro city located in south China and the spectrum sensing, spectrum decisions, spectrum evolution
empirical results show a high prediction accuracy having a upon handover and spectrum sharing [21]. The correspond-
mean RMSE as low as 0.04. In [119], based on spectrum ing example applications of the inference algorithms discussed
measurements ranging from 20 MHz to 3 GHz at four loca- in Section VI on each category are summarized in Table V.
tions in Guangdong province of China, the authors develop Moreover, the applications and benefits of exploiting spectrum
a 2-dimensional frequent pattern mining (FPM) algorithm to inference/prediction in CRNs are presented in Fig. 9, Table VI
predict the availability of multiple channels simultaneously, and briefly summarized as Sections VI-A–VI-D.
where the numerical results show that the prediction accuracy
is higher than 95% for the GSM900/1800 downlink and for A. Spectrum Inference for Sensing
the TV bands.
Spectrum sensing is a critical aspect of spectrum inference
Ding et al. [210] develop a joint spectral-temporal
applications that aim to explore idle spectrum slots. In the
prediction from incomplete historical observations, where real
emerging spectrum sensing paradigms in the time, frequency
measurements of TV bands (614 MHz-698 MHz) are used for
and spatial dimensions, inference techniques are widely used
quantifying the improvements over the one-dimensional neu-
to infer the most possible vacant channels for improving
ral network-based temporal prediction scheme. In an extended
the detection performance and reducing the energy and time
work [209], Ding et al. develop a robust online spectrum
consumption of sensing.
prediction algorithm relying on incomplete and corrupted
In terms of spectrum inference in the time domain for spec-
historical observations, where the real measurements of TV
trum sensing [15], [93], [128], [129], [139], [187], [220] use
bands, ISM bands and GSM bands are used for demonstrating
various linear prediction techniques, such as AR, MA, ARMA
that the proposed algorithm performs well in the presence of
and ARIMA to perform spectrum prediction, where the out-
missing data.
put is used to improve the sensing accuracy and reduce the
sensing cost. In parallel, Markov models such as HMM [14],
VI. A PPLICATIONS OF S PECTRUM I NFERENCE IN CRN S [17], [20], [95], [111], [118], [140], [142], [144], [146], [218],
The specific benefits of cognitive radio networks depend [221], [222] and POMDP [153] are also widely used to per-
on our ability to infer/predict the unknown/unmeasured spec- form similar tasks. These kinds of models work well under the
trum states from known/measured spectrum data. To por- assumption of memoryless or Markov property existing in the
tray the various applications of the spectrum inference spectrum state evolution. That is, the future state depends only
technology, in this section we first review the existing on the relevant information about the current, not on informa-
use cases and then we portray spectrum inference in the tion from further in the past. ANN models are another kind of
TABLE V
E XISTING A PPLICATIONS OF VARIOUS S PECTRUM I NFERENCE A LGORITHMS
TABLE VI
A L IST OF R EPORTED B ENEFITS /G AINS OF E XPLOITING S PECTRUM I NFERENCE IN CRN S
techniques that are widely investigated (see [20], [119], [139], states of unsensed channels is formulated as a matrix comple-
[158], [160]–[162], [167], [168], [178], [183], [223], [224]), tion problem and the technique of belief propagation (BP) is
which show relatively better prediction accuracy over other applied to fill the matrix with predicted states. Yin et al. [119]
models. In addition, pattern mining methods [119], [216], BIF has proposed a frequency pattern matching algorithm oper-
methods [17], [187] and SVM methods [190], [191] are also ating in two-dimensions (FPM-2D) for spectrum inference,
investigated. which searches through all relevant 2D patterns. Similarly,
In terms of spectrum inference in the frequency domain frequency domain correlation techniques were also introduced
for spectrum sensing, the spectral occupancy correlation in [93] by modeling the neighboring channels in pairs, where
between adjacent channels has been evaluated by experi- the classic AR models which do not need a priori knowledge
ments in [194]–[196]. Related studies are well dedicated to on the communication environment were also employed to
inferring the states of other channels in the light of already reduce the complexity.
acquired sensing or predicted results [93], [119], [192], [193]. The existing studies on spectrum inference in the spatial
Specifically, Yin et al. [192] exploited the spectral corre- domain for spectrum sensing are relatively limited. In [212],
lations for inferring the availabilities of other channels in a matrix completion-based algorithm is developed for TV
order to improve the throughput. In [193], predicting the white space database construction via joint spectrum sensing
in time domain and spectrum inference in spatial domain. of spectrum mobility management, user mobility management
Tang et al. [213] propose the concept of spectrum tensor and and inter-cell resource balancing.
develop a multi-dimensional (including, time, frequency, and Over the years, most of the existing studies have been
space) spectrum inference algorithm for spectrum map con- focused on so-called reactive spectrum mobility, where the
struction by invoking the recent advances in tensor completion. CR switches its communication once a PU becomes active,
where the detection of the PU relies either on spectrum sens-
B. Spectrum Inference for Decision ing or on monitoring. However, there are also some emerging
proactive spectrum mobility solutions, based on prediction or
Spectrum inference for decision has been investigated
inference techniques. For instance, Bütün et al. [205] pro-
in various aspects such as, centralized spectrum allocation,
pose a static neighbor graph (SNG) algorithm for predicting
decentralized channel selection, physical layer rate adaption,
the mobility of cognitive users. Wang et al. [202] proposed
dynamic spectrum access, to name just a few. In the following,
a support vector machine (SVM) based spectrum mobility
we introduce several representative examples.
prediction algorithm for mobile CRNs. In [203], the SVM
Akbar and Tranter [95] proposed to use the HMM based
was invoked both for predicting the handoff point and the
prediction algorithm for dynamic spectrum allocation, where
idle channels with a high precision. In addition to the spec-
the superiority in terms of system throughput is presented
trum availability, the link availability between the CRs was
by comparing with traditional CSMA based algorithms.
also inferred by Hou et al. [204], where by invoking this
Qin et al. [199] and Melián-Gutiérreza et al. [111] respectively
novel approach, a reliable path can be found for dynamic
studied the problem of spectrum inference-based channel
routing in CRNs. In [205], a location predictor is proposed,
selection in the HF spectrum. Zhao et al. [153] proposed
where the historical changes of the PU’s geographic locations
an approach for channel state prediction based on POMDP
are represented by a directed graph having weighted edges.
by finding the optimal policy that maximizes some aspect of
Once a spectrum occupancy prediction is requested, all the
the reward. Baldo et al. [167], [168] proposed to use neural
edges originating from the starting point are listed and then
networks based cognitive controller for dynamic channel selec-
the destination is predicted according to the approximately
tion and adaptation. In [178], Elman recurrent neural networks
calculated maximum weights. Huang et al. [174] propose a
was used for radio frequency multivariate time series mod-
Bayesian inference-based prediction algorithm for spotting the
elling in order to predict the spectral evolution, which leads
specific channel that is most likely to be available for the
to intelligent CR decisions for exploiting the expected spec-
CR-aided vehicular ad-hoc networks, where the most criti-
trum opportunities, optimized spectrum usage and interference
cal challenges are the high-speed mobility of vehicles and
avoidance.
the dynamically-fluctuating channel availability. As a further
advance, Thakur et al. [175] propose a proactive spectrum
C. Spectrum Inference for Mobility prediction technique, where the emergence of PU is predicted
The mobility terminology in the CRNs has double mean- before its true emergence, in order to avoid dropping even a
ings. On the one hand, it refers to ‘spectrum mobility or spec- single packet.
trum handoffs’ from one spectral band to another, for example,
due to the appearance of PUs or owing to interference avoid-
ance for other CRs. One the other hand, the ‘mobility of CRs D. Spectrum Inference for Sharing
and/or PUs’, for example in vehicular CRNs, may also affect In the context of spectrum sharing, there are different
the surrounding geographical spectral environment in terms understandings in the literature. In the generalized sense, the
of imposing additional interference, hence changing channel concept of spectrum sharing is inter-changeable with the con-
conditions and spectrum availability, etc. cepts of dynamic spectrum access or cognitive radio, which
An excellent survey paper on spectrum mobility in CRNs consists of three paradigms of spectrum usage: underlay,
was conceived by [170], which starts from the consideration overlay and interweave [120]. This kind of understanding
that in highly dynamic environments, CR-aided communi- of spectrum sharing makes its meaning too board to cover
cation is often interrupted and hence spectrum mobility is every aspects of CRNs. On the other hand, in the narrow
recognized as a pivotal feature of enabling continuous CR sense, spectrum sharing focuses on the underlay mode, which
data transmission, by transferring ongoing sessions to an allows CRs to operate in the same band at the same time
idle channel. A recent classification and survey was provided if the mutual interference among them is below a toler-
by [171]. In [172], a Poisson distribution based model of ant threshold. Spectrum inference for sharing is invoked for
spectral resources relying on a cross-layer spectrum handoff supporting/facilitating the coexistence of CRs with the PUs.
protocol optimized for the minimum expected transmission As shown in Table V, the research on inference-based
time was developed for cognitive LTE networks. However, spectrum sharing is relatively limited and mainly focused
the fluctuating nature of the available spectrum makes it diffi- on spatial domain [5], [18], [176], [200]. In [5], a sys-
cult to support seamless CR communications. To address this tematic approach is developed to enable efficient spectrum
problem, a spectrum-aware mobility management scheme is sharing between the PU (i.e., TV receivers) and the CR
proposed in [173] for CR cellular networks, where a unified (e.g., smart phone, tablet, mobile vehicles, etc) where matrix
mobility management framework is developed for supporting completion-based spectrum inference technique is invoked
the diverse mobility patterns in CR networks, which consists to serve as a spatial interpolator of unmeasured spectrum
Fig. 10. Spectrum inference for spectrum sharing in 5G wireless communication systems. The basic operation cycle is in light color and the functionality
of spectrum inference/prediction is in dark color.
data. Villardi and Harada [18] propose to predict the con- case, predictive schemes are expected to obtain the RSS of
tour and service areas of the PU for enabling unlicensed such wide spectrum bands and to enable wideband spectrum
CR systems operating in the TV white space without pro- sharing in a timely and cost-efficient manner based on time
ducing harmful interference to the PU receivers. Recently, domain or joint time-frequency domains spectrum inference
Kim and Giannakis [200] have proposed a dictionary learning algorithms discussed in Section IV. Specifically, in Fig. 10,
framework to predict the interference power levels in various a spectrum inference/prediction module is introduced, which
locations for enabling harmonious spectrum sharing between infers the future RSS based on the historical RSS data acquired
the PU and the CR. Zhou et al. [176] formulate the adaptive by spectrum sensing, which can further enhance 5G spectrum
vehicular data piping problem for dynamic spectrum sharing sharing for example by,
as a coalitional formation game and propose a near optimal • supporting adaptive PHY-layer spectrum sensing, i.e.,
coalitional formation approach for enabling vehicular data pipe adaptive optimization of the sensing parameters, such as
selection partition in a distributed way. the sensing time duration in each time slot,
• facilitating resource-efficient MAC-layer spectrum sens-
ing, for example by reducing the number of time-
E. Potential Use Case I: Spectrum Inference for Supporting slots required for multi-band sequential sensing as well
5G Spectrum Sharing as scheduling and by reducing the energy consump-
Radio spectrum sharing is an essential topic for 5G wire- tion of multi-sensor cooperative sensing as well as
less communications [258], [259]. The explosive growth of scheduling,
data rates offered by smart phones, tablets, laptops, vehi- • supporting high-data-rate spectrum sharing by combining
cles and many other wireless devices is about to overwhelm the outputs of spectrum sensing and spectrum prediction,
the allocated 2G-3G-4G radio spectrum. Spectrum sharing is for example by guiding the selection of spectrum bands
emerging as an affordable, near-term method of meeting the of high channel quality in carrier aggregation.
5G radio spectrum shortage and increasing the radio access Specifically, in 5G systems, prediction-based spectrum shar-
network capacities for supporting 5G content delivery [260]. ing can be arranged at the base stations (BSs) of macro/small
Spectrum sharing may occur between an incumbent user (e.g., cells and at the access points (APs) of WLANs in a central-
commercial TV or public radars) and a secondary 5G user, ized manner as well as more aggressively, by the autonomous
when a 5G device uses a spectrum band allocated for an user equipment in a self-organized manner by relying on game
incumbent usage in a geographic place, time and RF chan- theory and graph theory.
nel that the incumbent is not using, subject to the tolerable
interference imposed on the incumbent [5].
Fig. 10 presents a vision of spectrum inference/prediction- F. Potential Use Case II: Spectrum Inference for
based spectrum sharing conceived for 5G wireless communi- Next-Generation HF Automatic Link Establishment
cations. Specifically, in the basic operation cycle printed in High-frequency (HF) radio, also known as short-wave
light color, the spectrum sensing module infers the observable radio [261], operating in the 1.5-30 MHz spectrum band, is
RF stimuli from the radio environment and outputs the pre- now widely used, not only by the amateur community, but
dicted RSS in various spectrum bands of the current time slot. also by worldwide governmental and non-governmental agen-
Based on the input from spectrum sensing, the spectrum shar- cies as an alternative to satellites for over-the-horizon wireless
ing module activates spectrum reuse and interference control communications. Typical application scenarios of HF radios
to ensure the safe coexistence between commercial or public include ships at sea, aircraft out of range of line-of-sight radio
incumbent users and secondary 5G users. It may be deemed networks, disaster areas where the terrestrial communications
inevitable for 5G devices to explore and exploit the bene- infrastructure has been destroyed, and distant regions lacking
fits of multiple (non-continuous) spectrum bands of a wide other communications, to name just a few.
range, spanning from a few hundred MHz in the VHF/UHF One of the key challenges in using HF communications is
band to the 30-300 GHz millimeter-wave length band. In this finding a frequency that will support the desired tele-traffic
from a transmitter to a receiver. The reasons behind this

challenge are mainly two-fold: Firstly, over-the-horizon HF
communications often use skywave propagation paths pro-
vided by ionospheric refraction, which physically makes the
window of usable frequencies time-varying throughout the
day, the season, the subspot cycle, the weather environment,
and the radio locations, etc [262], which is quite different
from terrestrial wireless communications, such as cellular and
WLAN systems. Second, there are many governmental and
non-governmental HF radio systems in the HF band, which
make it a nontrivial task for each transmission to find a
frequency without interfering with other users [263].
To tackle the above critical challenge, ALE [262] is well-
recognized as a promising technology, which automates the
process of finding a usable frequency and setting up links
between two or more radios. Since the late 1970s, three gen-
erations of automatic link establishment techniques have been
developed. Briefly, the first generation ALE was independently
Fig. 11. A vision of spectrum inference/prediction-based wideband 4G HF
developed by different manufactures to automatically iden- ALE. The basic operating cycle is in light color and the functionality of
tify suitable transmission frequencies using microprocessors, spectrum inference/prediction is in dark color.
instead of the original manual operation. The second gener-
ation ALE focused on standardized, interoperable HF radio
systems, relying on the standards such as MIL-STD-188-141A sensing and LQA increases proportionally with the number
and FED-STD-1045. The third generation ALE operates at a of frequencies sensed and analyzed. To reduce the associated
lower SNR, carries more traffic and supports larger networks. delay, spectrum inference/prediction may be invoked as seen
The ALEs of the first three generations are in essence nar- in Fig. 11. Specifically, spectrum inference can be employed
rowband ALE operating over 3kHz HF channels. However, in two specific ways:
in recent years we have seen an increasing demand for • Inference/Prediction of spectrum occupancy, which out-
higher-data-rate transmission over HF links, supporting ser- puts the predicted spectrum occupancy ranking. Based on
vices ranging from voice and low-speed data to real-time video this output, the specific frequency of the particular bands
over HF skywave channels from aircraft, which motivates us to exhibiting a higher probability of being unoccupied will
develop wideband ALE operating over HF channels wider than be sensed first in the following process.
3kHz, and up to 24 kHz [262]. Wideband ALE, also termed • Inference/Prediction of link quality, which outputs the
as fourth generation (4G) ALE, introduces an increased band- predicted link quality ranking. Based on this output, the
width flexibility as well as additional automated capabilities: specific frequency of the links having a higher predicted
i) detect and characterize the occupancy or interference within link quality will be firstly selected for LQA.
a wideband channel, and ii) coordinate the allocation of the We note that in the basic operating cycle of wideband 4G
clear subchannel. ALE, the specific order in which spectrum sensing and LQA
The literature [263]–[266] reports on the potential applica- are carried out is generally random or predefined. By contrast,
tion of cognitive radio techniques in 4G ALE, where consensus the introduction of spectrum inference/prediction is capable
ensures that spectrum sensing will indeed be used for HF of providing an informative guidance considering the sensing
radios for detecting the occupancy or interference within any order and LQA order. Therefore, the link establishment time
portion of channels up to 24 kHz. Based on data link quality is expected to decrease in 4G ALE, provided that the spectrum
analysis (LQA), the associated frequency selection algorithm inference techniques summarized in the previous sections are
will be invoked for determining which channel(s) will be used invoked. To enable the vision shown in Fig. 11, time domain
for data transmission. The research and development of 4G or joint time-frequency domains spectrum inference algorithms
ALE is still in its infancy [267]. discussed in Section IV can be used.
Fig. 11 presents a vision of spectrum inference/prediction-
based wideband 4G ALE. Specifically, in the basic operating
cycle printed in light color, firstly sequential spectrum sens- G. Other Use Cases: The Applications of Spectrum
ing is performed for detecting the occupancy or interference Inference for Various Future Wireless Networks
within each channel. Based on the output of spectrum sensing, In addition to 5G and HF communications, there are
sequential LQA is then performed on the particular channels, many other potential applications of spectrum inference
which are deemed to be unoccupied. Based on the LQA result, for various future wireless networks, such as cognitive
the frequency band(s) having a high quality will be selected smart grid networks (CSGN) [56], [268], cognitive radio
as the active transmission link. One of the main problems sensor networks (CRSN) [40], [269]–[271], cognitive cellu-
of the basic operation cycle for wideband ALE is that the lar networks (CCN) [272]–[274], and cognitive machine-to-
amount of time required both by the process of spectrum machine communications (CM2M) [275].
Specifically, in cognitive radio-enabled smart grid networks, ON- and OFF-period durations. Wellens et al. [124],
spectrum sensing and spectrum inference techniques are Riihijarvi et al. [125], [126] have used the multi-scale entropy,
promising methods of providing timely smart grid wireless in order to examine both the complexity and the predictability
communications by utilizing all available spectrum resources. of the spectrum measurement traces recorded. In [127], from
Besides the prediction of spectrum state evolution, various sta- an information theory perspective, the authors have introduced
tistical inference techniques can also be utilized to predict a methodology of using statistical entropy measures and Fano
the system state of the smart grid itself and the potential inequality to quantify the degree of predictability underlying
faults resulting from environmental disasters, cyber attacks and real-world spectrum measurements. However, there is no com-
mechanical failures [56]. monly accepted limit on spectrum inference in the literature so
In CRSNs, wireless sensor networks benefit from the advan- far. Moreover, most of them have focused on the performance
tage of cognitive radio to overcome the spectrum scarcity limits of time-domain spectrum prediction. No work has con-
by enabling dynamic spectrum access [269]. Among others, sidered the limits of joint multi-domain spectrum inference,
the limited energy/power supply and processing capability of intuitively, which is expected to have a higher performance
wireless sensors impose additional constraints on the utiliza- upper bound.
tion of various cognitive radio techniques. More specifically,
the design of spectrum inference in CRSNs should care- B. Spectrum Inference in Various Domains
fully balance the tradeoffs among the inference accuracy, the
Besides the time, frequency and spatial domains, in other
computational complexity, and the memory requirements.
dimensions such as the code- and angular-domain, to the best
In future CCNs, spectrum sharing-oriented cognitive radio
of our knowledge, inference/prediction of spectrum usage has
techniques can be used to improve the spectrum utiliza-
not been proposed. Once the system becomes aware of the
tion and to meet the dramatic increase of wireless data rate
spreading code that the PU is going to use, the CRs become
requirements. Most of the studies on spectrum sharing in
capable of choosing orthogonal codes to simultaneously trans-
future CCNs assume that the network will be i) multi-tier,
mit information with little or no interference. In the context
including macro cells, small cells such as micro cells and
of multiple input multiple output (MIMO) techniques, the
pico cells, and device-to-device or machine-to-machine com-
PUs may transmit data within a narrow beam in a specific
munications [272]; ii) multi-band, including both licensed
direction [279]. This provides opportunities for CRs to simul-
and unlicensed bands [273]; and iii) software-defined, rely-
taneously transmit over the same frequency band in different
ing on various intelligent learning or data mining algo-
directions. Inference/prediction techniques may help the CRs
rithms operated in the cloud radio access networks [274].
to infer the PU’s transmit phase trajectory and pre-shift the
Inference of the spectrum usage state and the users’ content
phase accordingly.
demand will facilitate more proactive and efficient network
Moreover, most of the existing spectrum inference studies
operations.
focus on single domain spectrum inference, more specifi-
cally, time-domain spectrum prediction, but only very lim-
VII. O PEN I SSUES AND R ESEARCH T RENDS ited work consider joint multi-domain spectrum inference.
Although a number of studies have been carried out on Actually, no individual spectrum data exists in isolation,
spectrum inference in CRNs, this research topic has a great there are inherent correlations between this data and its
potential for future investigation. Actually, there are still a neighbors in time domain, frequency domain, and space
number of unsolved challenges waiting for solutions. Based domain. Consequently, joint multi-domain spectrum infer-
on the above survey and tutorial, this section presents a range ence is an active research direction, which is expected to
of potential open issues and future research trends as follows. bring more comprehensive, accurate and reliable spectrum
information.
A. Fundamental Performance Limits of Spectrum Inference
First of all, the fundamental performance limits on spec- C. Spectrum Inference in Various Bands
trum inference are still unknown. Most existing work on Although there have been extensive spectrum measurement
spectrum inference focus on applying various statistical infer- campaigns all over the world during the past decade, the
ence techniques for capturing the spectrum state’s evolution. majority of the efforts have been focused on TV bands to
However, to the best of our knowledge, there is a paucity identify TV white spaces (see [86], [87]), which to some
of studies focusing on the theoretical performance analysis extent represent the responses to FCC’s spectrum policy task
of spectrum inference. Just as the Shannon capacity gives the force in 2002 [57] and to FCC’s adoption rule for unlicensed
upper bound of various modulation and coding schemes, there use of television white spaces around 2010 [240]. There is
should be fundamental performance limits for the various spec- an urgent need to perform spectrum measurements and infer-
trum inference techniques. These fundamental performance ence, in order to document the usage of spectrum bands that
limits can guide the further development of spectrum inference have attracted the recent interest of the regulatory bodies. For
algorithms. instance, FCC issued a report and order in 2015, adopting rules
As mentioned above, recently, Olivieri et al. [123] for the commercial use of 150 MHz of spectrum in the 3550-
have applied the information theoretic entropy as a mea- 3700 MHz band (known as 3.5 GHz band), where the primary
sure of predictability in the process of generating the user is the U.S. Department of Defense (DoD) radar systems.
There are relatively few studies on the spectrum measurements and PUs. The current research on inference-based spec-
in the new bands [282]–[284]. Interestingly, spectrum shar- trum sharing is limited and mainly focused on spatial
ing between the incumbent radar systems and the Internet of domain [5], [18], [176], [200], where the inference of spa-
Things devices or secondary WiFi networks have been recently tial signal coverage is used to improve the spatial reuse. One
studied in [284]–[286]. interesting direction is to jointly predict/inference the spectrum
demand of CRs and the spectrum activities of PUs for proac-
D. Cooperative Spectrum Inference tively coordinating the spectrum sharing among them. Another
direction is to predict/inference the statistical channel gains of
In scenarios where there are multiple PUs and multiple CRs,
CR links, PU links and CR-PU interference links, which can
cooperative spectrum inference among CRs is a promising
facilitate the resource allocation, especially when the channel
research direction. Cooperative spectrum inference can ben-
gains cannot be obtained due to various practical constraints.
efit from multi-user diversity gains since different CRs may
face different shadowing and fading environments and have
H. Spectrum Inference for Mobility
different spectrum consequences for inference. Furthermore,
cooperative spectrum inference can enable different CRs As mentioned above, spectrum mobility in the CRNs has
to perform different inference subtasks, which may both a twin interpretation. On the one hand, it refers to spectral
reduce the computational complexity and the inference delay. handoff from one band to another, due to the appearance of
Moreover, the geographical movements as well as the data PUs or interference avoidance. One the other hand, the mobil-
transmission trends of the CRs could be better predicted ity of CRs and/or PUs, for example, in vehicular CRNs, may
by cooperative schemes. There are several studies in this also affect the geographically surrounding spectrum environ-
direction [257], [277], [278]. ment in terms of imposing additional interference, changing
channel conditions and spectrum availability, etc. Spectrum
E. Deep Spectrum Inference inference associated with in the former case has been exten-
sively studied. However, in the latter case, there are only a
With the escalation of spectrum demand, there is a press- few contributions.
ing need to recognize, classify and to activate the available Supporting high-throughput vehicular communications is
bands. In fields like public health, economic development important for safety applications, traffic management and
and climate forecasting, data mining has shown a beneficial mobile Internet access. One promising scenario of spectrum
predictive power [276]. Similar success may be anticipated inference for mobility can be found in railway, highway or
in the radio environment conceiving the required transmis- subway based cognitive communications, where the mobil-
sion power, spectrum state, PU/CR location, as long as there ity trajectory is fixed and thus the spectrum demand may be
is sufficient training data. To handle the flood of spectrum relatively regular for inference. One interesting research direc-
data, recent advances in artificial intelligence techniques, tion is to jointly consider the spectrum mobility with the user
like deep learning and reinforcement learning [280], [281], mobility since the spectrum evolution patterns are generally
are promising tools for improving the spectrum inference determined by the human’s usage of radio spectrum. Moreover,
performance. it is reported in [9] that a 93% potential predictability of human
mobility can be expected, which in-turn can be exploited for
F. Spectrum Inference for Sensing supporting accurate spectrum inference.
There are a number of interesting directions of exploiting
spectrum inference for sensing. Firstly, as is known, wire- I. Spectrum Inference for Decision
less spectrum has a (time, frequency, space, etc) multi-domain As mentioned above, spectrum inference for decision
space. Due to either the hardware limitations or sensor deploy- has been extensively investigated in various aspects such
ment cost, spectrum sensing can only capture the state of a as, centralized spectrum allocation, decentralized channel
partial of the spectrum space. To obtain a whole picture of the selection, physical layer rate adaption, dynamic spectrum
multi-domain spectrum space, spectrum inference can help to access, to name just a few. There are relatively few open
fill the unmeasured space via acting as an interpolator. Second, issues found on this topic. However, one direction is to
to find idle spectrum from a large number of candidate bands, design demos or systems that utilizes the specific tech-
the order of spectrum sensing is vital to minimize the sensing niques with the well known example as DAPRA’s spectrum
time overhead. Spectrum inference can provide a predictive challenges.
input to spectrum sensing for enabling a better sensing order.
In addition, spectrum inference can output a complementary J. Applications of Spectrum Inference
spectrum state estimation to spectrum sensing. It’s possible to
Last but not least, more investigations on specific appli-
fuze the output of spectrum inference and spectrum sensing
cations of spectrum inference techniques is also a fruitful
to get a more accurate state estimation.
research direction. Although there are several common rules
for various applications, the requirements of spectrum infer-
G. Spectrum Inference for Sharing ence in different application scenarios are rather diverse. For
In the context of spectrum sharing, spectrum inference is example, the spectrum occupancy state in TV bands changes
invoked for supporting/facilitating the coexistence of CRs relatively slowly on a time-scale of several hours, while the
TABLE VII
ACRONYMS
spectrum occupancy state in cellular bands or Wi-Fi bands a fruitful research area. Our hope is that this paper, with
changes within several milliseconds. Considering the fact that its interdisciplinary perspective, will stimulate the research
the outdated prediction or inference is useless, the tolerance and development of spectrum inference in future wireless
to time delay in the corresponding spectrum inference is sig- networks.
nificantly different, which imposes different design constraints
on the specific inference algorithms. ACKNOWLEDGMENT
The authors thank the Editor and the Reviewers for
VIII. C ONCLUSION their helpful comments and constructive suggestions, which
Spectrum inference is a promising technique of improving have helped them significantly improve this work. The
the spectrum exploitation in cognitive radio networks. In this authors also thank Ms. Jing Zhang, Ms. Mengyun Tang and
paper, we reviewed the recent advances in spectrum inference Mr. Guangming Nie in collecting and formatting parts of the
based on an extensive study of the existing literature. We first references.
introduced the preliminaries of spectrum inference, including
the sources of spectrum data, the models of spectrum usage R EFERENCES
and the predictability of spectrum evolution. Then we explored [1] “Notice of proposed rule making and order,” Federal Commun.
various spectrum inference algorithms from a time-frequency- Commission, Washington, DC, USA, Tech. Rep. ET Docket 03-322,
Nov. 2003.
spatial domain perspective and presented an in-depth tutorial. [2] J. Mitola, “Cognitive radio: An integrated agent architecture for soft-
We proceeded by offering a comparative analysis of the advan- ware defined radio,” Ph.D. dissertation, KTH Roy. Inst. Technol.,
tages and challenges of various spectrum inference techniques. Stockholm, Sweden, 2000.
[3] S. Haykin, “Cognitive radio: Brain-empowered wireless communica-
Additionally, we reviewed the applications of spectrum infer- tions,” IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201–220,
ence both in existing and in future wireless networks, including Feb. 2005.
5G cellular communications, next-generation HF communica- [4] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “A survey on
spectrum management in cognitive radio networks,” IEEE Commun.
tions, cognitive smart grid networks, cognitive radio sensor Mag., vol. 46, no. 4, pp. 40–48, Apr. 2008.
networks, etc. We also highlighted a range of open issues and [5] G. Ding et al., “Cellular-base-station-assisted device-to-device commu-
research trends influencing the actual deployment of spectrum nications in TV white space,” IEEE J. Sel. Areas Commun., vol. 34,
no. 1, pp. 107–121, Jan. 2016.
inference. We conclude that the main goal of the existing [6] R. Kleeman, A. J. Majda, and I. Timofeyev, “Quantifying predictability
and forthcoming studies on spectrum inference in CRNs is in a model with statistical features of the atmosphere,” Proc. Nat. Acad.
to achieve a compromise amongst the conflicting objectives Sci. USA, vol. 99, no. 24, pp. 15291–15296, 2002.
[7] N. Navet and S.-H. Chen, “On predictability and profitability: Would
of improving the prediction accuracy, reducing its compu- GP induced trading rules be sensitive to the observed entropy of time
tational complexity and memory requirement. This forms series?” Nat. Comput. Comput. Financ., vol. 100, pp. 197–210, 2008.
[8] R. Li, Z. Zhao, X. Zhou, J. Palicot, and H. Zhang, “The prediction [32] L. Gavrilovska, V. Atanasovski, I. Macaluso, and L. A. DaSilva,
analysis of cellular radio access network traffic: From entropy the- “Learning and reasoning in cognitive radio networks,” IEEE Commun.
ory to networking practice,” IEEE Commun. Mag., vol. 52, no. 6, Surveys Tuts., vol. 15, no. 4, pp. 1761–1777, 4th Quart., 2013.
pp. 234–240, Jun. 2014. [33] E. Z. Tragos, S. Zeadally, A. G. Fragkiadakis, and V. A. Siris,
[9] C. Song, Z. Qu, N. Blumm, and A.-L. Barabási, “Limits of predictabil- “Spectrum assignment in cognitive radio networks: A comprehensive
ity in human mobility,” Science, vol. 327, no. 5968, pp. 1018–1021, survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 3, pp. 1108–1135,
2010. 3rd Quart., 2013.
[10] J. Manyika et al., Big Data: The Next Frontier for Innovation, [34] E. Ahmed, A. Gani, S. Abolfazli, L. J. Yao, and S. U. Khan, “Channel
Competition, and Productivity. New York, NY, USA: McKinsey Glob. assignment algorithms in cognitive radio networks: Taxonomy, open
Inst., May 2011. issues, and challenges,” IEEE Commun. Surveys Tuts., vol. 18, no. 1,
[11] R. Tandra, S. M. Mishra, and A. Sahai, “What is a spectrum hole pp. 795–823, 1st Quart., 2016.
and what does it take to recognize one?” Proc. IEEE, vol. 97, no. 5, [35] W. Liang, S. X. Ng, and L. Hanzo, “Cooperative overlay spectrum
pp. 824–848, May 2009. access in cognitive radio networks,” IEEE Commun. Surveys Tuts.,
[12] J. Yang and H. Zhao, “Enhanced throughput of cognitive radio vol. 19, no. 3, pp. 1924–1944, 3rd Quart., 2017.
networks by imperfect spectrum prediction,” IEEE Commun. Lett., [36] M. Naeem, A. Anpalagan, M. Jaseemuddin, and D. C. Lee, “Resource
vol. 19, no. 10, pp. 1738–1741, Oct. 2015. allocation techniques in cooperative cognitive radio networks,” IEEE
[13] Q. Wu, G. Ding, J. Wang, and Y.-D. Yao, “Spatial–temporal opportunity Commun. Surveys Tuts., vol. 16, no. 2, pp. 729–744, 2nd Quart., 2014.
detection for spectrum-heterogeneous cognitive radio networks: Two- [37] Y. Xu, X. Zhao, and Y.-C. Liang, “Robust power control and beamform-
dimensional sensing,” IEEE Trans. Wireless Commun., vol. 12, no. 2, ing in cognitive radio networks: A survey,” IEEE Commun. Surveys
pp. 516–526, Feb. 2013. Tuts., vol. 17, no. 4, pp. 1834–1857, 4th Quart., 2015.
[14] Z. Chen, N. Guo, Z. Hu, and R. C. Qiu, “Experimental validation of [38] A. Ahmad, S. Ahmad, M. H. Rehmani, and N. U. Hassan, “A survey
channel state prediction considering delays in practical cognitive radio,” on radio resource allocation in cognitive radio sensor networks,” IEEE
IEEE Trans. Veh. Technol., vol. 60, no. 4, pp. 1314–1325, May 2011. Commun. Surveys Tuts., vol. 17, no. 2, pp. 888–917, 2nd Quart., 2015.
[15] Z. Wen, T. Luo, W. Xiang, S. Majhi, and Y. Ma, “Autoregressive [39] G. I. Tsiropoulos, O. A. Dobre, M. H. Ahmed, and K. E. Baddour,
spectrum hole prediction model for cognitive radio systems,” in Proc. “Radio resource allocation techniques for efficient spectrum access in
IEEE Int. Conf. Commun. Workshops, Beijing, China, May 2008, cognitive radio networks,” IEEE Commun. Surveys Tuts., vol. 18, no. 1,
pp. 154–157. pp. 824–847, 1st Quart., 2016.
[16] P. Huang, C.-J. Liu, X. Yang, L. Xiao, and J. Chen, “Wireless spectrum [40] S. H. R. Bukhari, M. H. Rehmani, and S. Siraj, “A survey of channel
occupancy prediction based on partial periodic pattern mining,” IEEE bonding for wireless networks and guidelines of channel bonding for
Trans. Parallel Distrib. Syst., vol. 25, no. 7, pp. 1925–1934, Jul. 2014. futuristic cognitive radio sensor networks,” IEEE Commun. Surveys
[17] X. Xing, T. Jing, Y. Huo, H. Li, and X. Cheng, “Channel quality Tuts., vol. 18, no. 2, pp. 924–948, 2nd Quart., 2016.
prediction based on Bayesian inference in cognitive radio networks,” [41] M. El Tanab and W. Hamouda, “Resource allocation for underlay cog-
in Proc. IEEE INFOCOM, Turin, Italy, Apr. 2013, pp. 1465–1473. nitive radio networks: A survey,” IEEE Commun. Surveys Tuts., vol. 19,
[18] G. P. Villardi and H. Harada, “Computer prediction of primary con- no. 2, pp. 1249–1276, 2nd Quart., 2017.
tour and service areas for unlicensed radio systems operating in TV
[42] G. Baldini et al., “Security aspects in software defined radio and cog-
white space,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC),
nitive radio networks: A survey and a way ahead,” IEEE Commun.
Shanghai, China, Apr. 2013, pp. 4528–4533.
Surveys Tuts., vol. 14, no. 2, pp. 355–379, 2nd Quart., 2012.
[19] M. Lopez-Benitez and F. Casadevall, “Statistical prediction of spectrum
[43] A. Attar, H. Tang, A. V. Vasilakos, F. R. Yu, and V. C. M. Leung, “A
occupancy perception in dynamic spectrum access networks,” in Proc.
survey of security challenges in cognitive radio networks: Solutions
IEEE Int. Conf. Commun. (ICC), Kyoto, Japan, Jun. 2011, pp. 1–6.
and future research directions,” Proc. IEEE, vol. 100, no. 12,
[20] V. K. Tumuluru, P. Wang, and D. Niyato, “Channel status prediction for
pp. 3172–3186, Dec. 2012.
cognitive radio networks,” Wireless Commun. Mobile Comput., vol. 12,
no. 10, pp. 862–874, 2012. [44] H. Rifà-Pous, M. J. Blasco, and C. Garrigues, “Review of robust
[21] X. Xing, T. Jing, W. Cheng, Y. Huo, and X. Cheng, “Spectrum cooperative spectrum sensing techniques for cognitive radio networks,”
prediction in cognitive radio networks,” IEEE Wireless Commun., Wireless Pers. Commun., vol. 67, no. 2, pp. 175–198, Nov. 2012.
vol. 20, no. 2, pp. 90–96, Apr. 2013. [45] A. G. Fragkiadakis, E. Z. Tragos, and I. G. Askoxylakis, “A survey on
[22] Y. Saleem and M. H. Rehmani, “Primary radio user activity models for security threats and detection techniques in cognitive radio networks,”
cognitive radio networks: A survey,” J. Netw. Comput. Appl., vol. 43, IEEE Commun. Surveys Tuts., vol. 15, no. 1, pp. 428–445, 1st Quart.,
pp. 1–16, 2014. 2013.
[23] M. Höyhtya et al., “Spectrum occupancy measurements: A survey and [46] L. Zhang et al., “Byzantine attack and defense in cognitive radio
use of interference maps,” IEEE Commun. Surveys Tuts., vol. 18, no. 4, networks: A survey,” IEEE Commun. Surveys Tuts., vol. 17, no. 3,
pp. 2386–2414, 4th Quart., 2016. pp. 1342–1363, 3rd Quart., 2015.
[24] Y. Chen and H.-S. Oh, “A survey of measurement-based spectrum occu- [47] R. K. Sharma and D. B. Rawat, “Advances on security threats and coun-
pancy modeling for cognitive radios,” IEEE Commun. Surveys Tuts., termeasures for cognitive radio networks: A survey,” IEEE Commun.
vol. 18, no. 1, pp. 848–859, 1st Quart., 2016. Surveys Tuts., vol. 17, no. 2, pp. 1023–1043, 2nd Quart., 2015.
[25] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for [48] M. Grissa, B. Hamdaoui, and A. A. Yavuz, “Location privacy in cogni-
cognitive radio applications,” IEEE Commun. Surveys Tuts., vol. 11, tive radio networks: A survey,” IEEE Commun. Surveys Tuts., vol. 19,
no. 1, pp. 116–130, 1st Quart., 2009. no. 3, pp. 1726–1760, 3rd Quart., 2017.
[26] Y. Zeng, Y.-C. Liang, A. T. Hoang, and R. Zhang, “A review on spec- [49] A. De Domenico, E. C. Strinati, and M.-G. Di Benedetto, “A survey on
trum sensing for cognitive radio: Challenges and solutions,” EURASIP MAC strategies for cognitive radio networks,” IEEE Commun. Surveys
J. Adv. Signal Process., vol. 2010, Dec. 2010, Art. no. 381465. Tuts., vol. 14, no. 1, pp. 21–44, 1st Quart., 2012.
[27] I. F. Akyildiz, B. F. Lo, and R. Balakrishnan, “Cooperative spectrum [50] L. Gavrilovska, D. Denkovski, V. Rakovic, and M. Angjelichinoski,
sensing in cognitive radio networks: A survey,” Phys. Commun., vol. 4, “Medium access control protocols in cognitive radio networks:
no. 1, pp. 40–62, Mar. 2011. Overview and general classification,” IEEE Commun. Surveys Tuts.,
[28] E. Axell, G. Leus, E. G. Larsson, and H. V. Poor, “Spectrum sensing vol. 16, no. 4, pp. 2092–2124, 4th Quart., 2014.
for cognitive radio: State-of-the-art and recent advances,” IEEE Signal [51] M. Youssef, M. Ibrahim, M. Abdelatif, L. Chen, and A. V. Vasilakos,
Process. Mag., vol. 29, no. 3, pp. 101–116, May 2012. “Routing metrics of cognitive radio networks: A survey,” IEEE
[29] G. Ding, Q. Wu, Y.-D. Yao, J. Wang, and Y. Chen, “Kernel-based Commun. Surveys Tuts., vol. 16, no. 1, pp. 92–109, 1st Quart.,
learning for statistical signal processing in cognitive radio networks: 2014.
Theoretical foundations, example applications, and future directions,” [52] A. Naeem, M. H. Rehmani, Y. Saleem, I. Rashid, and N. Crespi,
IEEE Signal Process. Mag., vol. 30, no. 4, pp. 126–136, Jul. 2013. “Network coding in cognitive radio networks: A comprehensive sur-
[30] A. Ali and W. Hamouda, “Advances on spectrum sensing for cogni- vey,” IEEE Commun. Surveys Tuts., vol. 19, no. 3, pp. 1945–1973,
tive radio networks: Theory and applications,” IEEE Commun. Surveys 3rd Quart., 2017.
Tuts., vol. 19, no. 2, pp. 1277–1304, 2nd Quart., 2017. [53] X. Chen, H.-H. Chen, and W. Meng, “Cooperative communications
[31] M. T. Masonta, M. Mzyece, and N. Ntlatlapa, “Spectrum decision in for cognitive radio networks—From theory to applications,” IEEE
cognitive radio networks: A survey,” IEEE Commun. Surveys Tuts., Commun. Surveys Tuts., vol. 16, no. 3, pp. 1180–1192, 3rd Quart.,
vol. 15, no. 3, pp. 1088–1107, 3rd Quart., 2013. 2014.
[54] X. Huang, T. Han, and N. Ansari, “On green-energy-powered cog- [72] M. Wellens, J. Wu, and P. Mahonen, “Evaluation of spectrum occu-
nitive radio networks,” IEEE Commun. Surveys Tuts., vol. 17, no. 2, pancy in indoor and outdoor scenario in the context of cognitive radio,”
pp. 827–842, 2nd Quart., 2015. in Proc. 2nd Int. Conf. Cogn. Radio Orient. Wireless Netw. Commun.,
[55] H. Ding et al., “Cognitive capacity harvesting networks: Architectural Orlando, FL, USA, Aug. 2007, pp. 420–427.
evolution toward future cognitive radio networks,” IEEE Commun. [73] R. B. Bacchus, A. J. Fertner, C. S. Hood, and D. A. Roberson, “Long-
Surveys Tuts., vol. 19, no. 3, pp. 1902–1923, 3rd Quart., term, wide-band spectral monitoring in support of dynamic spectrum
2017. access networks at the IIT spectrum observatory,” in Proc. 3rd IEEE
[56] A. A. Khan, M. H. Rehmani, and M. Reisslein, “Cognitive radio for Symp. New Front. Dyn. Spectr. Access Netw. (DySPAN), Chicago, IL,
smart grids: Survey of architectures, spectrum sensing mechanisms, and USA, Oct. 2008, pp. 1–10.
networking protocols,” IEEE Commun. Surveys Tuts., vol. 18, no. 1, [74] M. H. Islam et al., “Spectrum survey in Singapore: Occupancy mea-
pp. 860–898, 1st Quart., 2016. surements and analyses,” in Proc. 3rd Int. Conf. Cogn. Radio Orient.
[57] Spectrum Policy Task Force, “Report of the spectrum efficiency Wireless Netw. Commun., Singapore, May 2008, pp. 1–7.
working group,” Federal Commun. Commission, Washington, [75] M. Lopez-Benitez, A. Umbert, and F. Casadevall, “Evaluation of
DC, USA, Tech. Rep., Nov. 2002. [Online]. Available: spectrum occupancy in Spain for cognitive radio applications,” in
https://www.fcc.gov/sptf/files/SEWGFinalReport_1.pdf Proc. IEEE 69th Veh. Technol. Conf., Barcelona, Spain, Apr. 2009,
[58] A. Azarfar, J. F. Frigon, and B. Sanso, “Improving the reliability of pp. 1–5.
wireless networks using cognitive radios,” IEEE Commun. Surveys [76] K. A. Qaraqe et al., “Empirical results for wideband multidimensional
Tuts., vol. 14, no. 2, pp. 338–354, 2nd Quart., 2012. spectrum usage,” in Proc. IEEE 20th Int. Symp. Pers. Indoor Mobile
[59] F. H. Sanders, “Broadband spectrum surveys in Denver, CO, San Diego, Radio Commun., Tokyo, Japan, Sep. 2009, pp. 1262–1266.
CA, and Los Angeles, CA: Methodology, analysis, and compara- [77] R. Bacchus, T. Taher, K. Zdunek, and D. Roberson, “Spectrum utiliza-
tive results,” in Proc. IEEE Int. Symp. Electromagn. Compat., vol. 2. tion study in support of dynamic spectrum access for public safety,” in
Denver, CO, USA, Aug. 1998, pp. 988–993. Proc. IEEE Symp. New Front. Dyn. Spectr. Access Netw., Singapore,
[60] F. H. Sanders, B. J. Ramsey, and V. S. Lawrence, “Broadband spectrum Apr. 2010, pp. 1–11.
survey at San Diego, California,” U.S. Dept. Commerce, Washington, [78] R. Schiphorst and C. H. Slump, “Evaluation of spectrum occupancy
DC, USA, Tech. Rep. NTIA 97-334, 1996. in Amsterdam using mobile monitoring vehicles,” in Proc. IEEE 71st
[61] M. Biggs, A. Henley, and T. Clarkson, “Occupancy analysis of the Veh. Technol. Conf., Taipei, Taiwan, May 2010, pp. 1–5.
2.4 GHz ISM band,” IEE Proc. Commun., vol. 151, no. 5, pp. 481–488, [79] V. Valenta et al., “Survey on spectrum utilization in Europe:
Oct. 2004. Measurements, analyses and observations,” in Proc. 5th Int. Conf.
[62] M. McHenry and D. McCloskey, “Spectrum occupancy measure- Cogn. Radio Orient. Wireless Netw. Commun., Cannes, France,
ments: Chicago, Illinois, November 16-18, 2005,” Shared Spectr. Jun. 2010, pp. 1–5.
Company, Vienna, VA, USA, Tech. Rep., 2005. [Online]. Available: [80] S. Contreras, G. Villardi, R. Funada, and H. Harada, “An investigation
http://www.sharedspectrum.com/wp-content/uploads/NSF_Chicago_ into the spectrum occupancy in Japan in the context of TV white space
2005-11_measurements_v12.pdf systems,” in Proc. 6th Int. ICST Conf. Cogn. Radio Orient. Wireless
[63] A. Petrin and P. G. Steffes, “Analysis and comparison of spectrum Netw. Commun., Osaka, Japan, Jun. 2011, pp. 341–345.
measurements performed in urban and rural areas to determine the [81] T. Harrold, R. Cepeda, and M. Beach, “Long-term measurements of
total amount of spectrum usage,” in Proc. Int. Symp. Adv. Radio spectrum occupancy characteristics,” in Proc. IEEE Symp. New Front.
Technol. (ISART), Boulder, CO, USA, Mar. 2005, pp. 9–12. Dyn. Spectr. Access Netw. (DySPAN), Aachen, Germany, May 2011,
[64] A. J. Petrin, “Maximizing the utility of radio spectrum: Broadband pp. 83–89.
spectrum measurements and occupancy model for use by cognitive [82] T. M. Taher, R. B. Bacchus, K. J. Zdunek, and D. A. Roberson, “Long-
radio,” Ph.D. dissertation, Georgia Inst. Technol., Atlanta, GA, USA, term spectral occupancy findings in Chicago,” in Proc. IEEE Symp.
2005. New Front. Dyn. Spectr. Access Netw. (DySPAN), Aachen, Germany,
[65] M. A. McHenry, P. A. Tenhula, D. McCloskey, D. A. Roberson, and May 2011, pp. 100–107.
C. S. Hood, “Chicago spectrum occupancy measurements & anal- [83] J. Kokkoniemi and J. Lehtomäki, “Spectrum occupancy measure-
ysis and a long-term studies proposal,” in Proc. 1st Int. Workshop ments and analysis methods on the 2.45 GHz ISM band,” in
Technol. Policy Accessing Spectr., Boston, MA, USA, Aug. 2006, Proc. 7th Int. ICST Conf. Cogn. Radio Orient. Wireless Netw. Commun.
pp. 1–5. (CROWNCOM), Stockholm, Sweden, Jun. 2012, pp. 285–290.
[66] D. A. Roberson, C. S. Hood, J. L. LoCicero, and J. T. MacDonald, [84] R. Urban, T. Korinek, and P. Pechač, “Broadband spectrum survey mea-
“Spectral occupancy and interference studies in support of cogni- surements for cognitive radio applications,” Radioengineering, vol. 21,
tive radio technology deployment,” in Proc. 1st IEEE Workshop no. 4, pp. 1101–1109, 2012.
Netw. Technol. Softw. Defined Radio Netw., Reston, VA, USA, 2006, [85] L. F. Pedraza, F. Forero, and I. Paez, “Metropolitan spectrum
pp. 26–35. survey in Bogota Colombia,” in Proc. 27th Int. Conf. Adv. Inf.
[67] V. Blaschke, H. Jäkel, T. Renk, C. Kloeck, and F. K. Jondral, Netw. Appl. Workshops (WAINA), Barcelona, Spain, Mar. 2013,
“Occupation measurements supporting dynamic spectrum allocation for pp. 548–553.
cognitive radio design,” in Proc. 2nd Int. Conf. Cogn. Radio Orient. [86] K. Chen et al., “Spectrum survey for TV band in Beijing,” in
Wireless Netw. Commun. (CrownCom), Orlando, FL, USA, Aug. 2007, Proc. 21st Int. Conf. Telecommun. (ICT), Lisbon, Portugal, May 2014,
pp. 50–57. pp. 267–271.
[68] R. I. C. Chiang, G. B. Rowe, and K. W. Sowerby, “A quantitative [87] O. Holland et al., “Changing availability of TV white space in the
analysis of spectral occupancy measurements for cognitive radio,” in U.K.,” IEE Electron. Lett., vol. 52, no. 15, pp. 1349–1351, Jul. 2016.
Proc. IEEE 65th Veh. Technol. Conf. Spring, Dublin, Ireland, Apr. 2007, [88] B. Bochow, O. Holland, and K. Katzis, “Spectrum sensing infrastruc-
pp. 3016–3020. ture support for IEEE 1900.6b sensing-assisted spectrum databases,”
[69] T. Erpek, M. Lofquist, and K. Patton, “Spectrum occupancy mea- in Proc. IEEE Conf. Stand. Commun. Netw. (CSCN), Berlin, Germany,
surements: Loring commerce centre, limestone, Maine September Oct./Nov. 2016, pp. 1–6.
18-20, 2007,” Shared Spectr. Company, Vienna, VA, USA, Tech. [89] M. López-Benítez, “Spectrum usage models for the analysis, design
Rep., 2007. [Online]. Available: http://www.sharedspectrum.com/ and simulation of cognitive radio networks,” in Cognitive Radio and
wp-content/uploads/Loring_Spectrum_Occupancy_Measurements_v2_ Its Application for Next Generation Cellular and Wireless Networks.
3.pdf Dordrecht, The Netherlands: Springer, 2012, pp. 27–73.
[70] T. Erpek, K. Steadman, and D. Jones, “Dublin Ireland spectrum [90] C. Ghosh, S. Pagadarai, D. P. Agrawal, and A. M. Wyglinski, “A frame-
occupancy measurements collected on April 16-18, 2007,” Shared work for statistical wireless spectrum occupancy modeling,” IEEE
Spectr. Company, Vienna, VA, USA, Tech. Rep., 2007. [Online]. Trans. Wireless Commun., vol. 9, no. 1, pp. 38–44, Jan. 2010.
Available: http://www.sharedspectrum.com/wp-content/uploads/ [91] M. Wellens, J. RiihijäRvi, and P. MäHöNen, “Empirical time and
Ireland_Spectrum_Occupancy_Measurements_v2.pdf frequency domain models of spectrum use,” Phys. Commun., vol. 2,
[71] S. D. Jones, E. Jung, X. Liu, N. Merheb, and I.-J. Wang, nos. 1–2, pp. 10–32, 2009.
“Characterization of spectrum activities in the U.S. public safety band [92] P. F. Marshall, “Closed-form analysis of spectrum characteristics
for opportunistic spectrum access,” in Proc. 2nd IEEE Int. Symp. for cognitive radio performance analysis,” in Proc. 3rd IEEE Symp.
New Front. Dyn. Spectr. Access Netw., Dublin, Ireland, Apr. 2007, New Front. Dyn. Spectr. Access Netw. (DySpan), Chicago, IL, USA,
pp. 137–146. Oct. 2008, pp. 1–12.
[93] A. Gorcin, H. Celebi, K. A. Qaraqe, and H. Arslan, “An autore- [116] Y. Zhao, B. Le, and J. H. Reed, “Network support the radio
gressive approach for spectrum occupancy modeling and prediction environment map,” in Cognitive Radio Technology. Amsterdam,
based on synchronous measurements,” in Proc. IEEE 22nd Int. Symp. The Netherlands: Newnes, 2006, pp. 325–366.
Pers. Indoor Mobile Radio Commun. (PIMRC), Toronto, ON, Canada, [117] J. J. Lehtomaki, R. Vuohtoniemi, and K. Umebayashi, “On the mea-
Sep. 2011, pp. 705–709. surement of duty cycle and channel occupancy rate,” IEEE J. Sel. Areas
[94] K. W. Sung, S.-L. Kim, and J. Zander, “Temporal spectrum shar- Commun., vol. 31, no. 11, pp. 2555–2565, Nov. 2013.
ing based on primary user activity prediction,” IEEE Trans. Wireless [118] X. Chen, H. Zhang, A. B. MacKenzie, and M. Matinmikko, “Predicting
Commun., vol. 9, no. 12, pp. 3848–3855, Dec. 2010. spectrum occupancies using a non-stationary hidden Markov model,”
[95] I. A. Akbar and W. H. Tranter, “Dynamic spectrum allocation in cog- IEEE Commun. Lett., vol. 3, no. 4, pp. 333–336, Aug. 2014.
nitive radio using hidden Markov models: Poisson distributed case,” in [119] S. Yin, D. Chen, Q. Zhang, M. Liu, and S. Li, “Mining spectrum usage
Proc. IEEE SoutheastCon, Richmond, VA, USA, 2007, pp. 196–201. data: A large-scale spectrum measurement study,” IEEE Trans. Mobile
[96] A. Papoulis and S. U. Pillai, Probability, Random Variables, and Comput., vol. 11, no. 6, pp. 1033–1046, Jun. 2012.
Stochastic Processes. Boston, MA, USA: McGraw-Hill, 2002. [120] A. Goldsmith and I. Maric, “Capacity of cognitive radio networks,” in
[97] P. Kumaraswamy, “A generalized probability density function for Principles of Cognitive Radio Networks. Cambridge, U.K.: Cambridge
double-bounded random processes,” J. Hydrol., vol. 46, nos. 1–2, Univ. Press, Nov. 2012.
pp. 79–88, 1980. [121] F. Paisana, N. Marchetti, and L. A. DaSilva, “Radar, TV and cellular
[98] A. U. Rehman, V. A. Thomas, L.-L. Yang, and L. Hanzo, “Performance bands: Which spectrum access techniques for which bands?” IEEE
of cognitive selective-repeat hybrid automatic repeat request,” IEEE Commun. Surveys Tuts., vol. 16, no. 3, pp. 1193–1220, 3rd Quart.,
Access, vol. 4, pp. 9828–9846, 2016. 2014.
[99] A. U. Rehman, C. Dong, L.-L. Yang, and L. Hanzo, “Performance of [122] F. Akhtar, M. H. Rehmani, and M. Reisslein, “White space: Definitional
cognitive stop-and-wait hybrid automatic repeat request in the face of perspectives and their role in exploiting spectrum opportunities,”
imperfect sensing,” IEEE Access, vol. 4, pp. 5489–5508, 2016. Telecommun. Policy, vol. 40, no. 4, pp. 319–331, 2016.
[100] Z. Liang and D. Zhao, “Quality of service performance of a cogni- [123] M. P. Olivieri, G. Barnett, A. Lackpour, A. Davis, and P. Ngo, “A scal-
tive radio sensor network,” in Proc. IEEE Int. Conf. Commun. (ICC), able dynamic spectrum allocation system with interference mitigation
Cape Town, South Africa, May 2010, pp. 1–5. for teams of spectrally agile software defined radios,” in Proc. IEEE
[101] L. Yang, L. Cao, and H. Zheng, “Proactive channel access in dynamic Symp. New Front. Dyn. Spectr. Access Netw. (DySPAN), Baltimore,
spectrum networks,” Phys. Commun., vol. 1, no. 2, pp. 103–111, 2008. MD, USA, Nov. 2005, pp. 170–179.
[102] M. Hamid, A. Mohammed, and Z. Yang, “On spectrum sharing and [124] M. Wellens, J. Riihijarvi, and P. Mahonen, “Evaluation of spec-
dynamic spectrum allocation: MAC layer spectrum sensing in cognitive trum occupancy using approximate and multiscale entropy metrics,”
radio networks,” in Proc. Int. Conf. Commun. Mobile Comput. (CMC), in Proc. 5th IEEE Annu. Commun. Soc. Conf. Sensor Mesh Ad Hoc
vol. 2. Shenzhen, China, Apr. 2010, pp. 183–187. Commun. Netw. Workshops, San Francisco, CA, USA, Jun. 2008,
[103] C. Guo, T. Peng, Y. Qi, and W. Wang, “Adaptive channel searching pp. 1–8.
scheme for cooperative spectrum sensing in cognitive radio networks,” [125] J. Riihijarvi, P. Mahonen, and M. Wellens, “Metrics for characteriz-
in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), Budapest, ing complexity of network traffic,” in Proc. Int. Conf. Telecommun.,
Hungary, Apr. 2009, pp. 1–6. St. Petersburg, Russia, Jun. 2008, pp. 1–6.
[104] S. Geirhofer, L. Tong, and B. M. Sadler, “Dynamic spectrum access [126] J. Riihijarvi, M. Wellens, and P. Mahonen, “Measuring complexity and
in WLAN channels: Empirical model and its stochastic analysis,” in predictability in networks with multiscale entropy analysis,” in Proc.
Proc. 1st Int. Workshop Technol. Policy Accessing Spectr., Boston, MA, IEEE INFOCOM, Rio de Janeiro, Brazil, Apr. 2009, pp. 1107–1115.
USA, Aug. 2006, pp. 1–10. [127] G. Ding et al., “On the limits of predictability in real-world radio
[105] S. Geirhofer, L. Tong, and B. M. Sadler, “Cognitive radios for dynamic spectrum state dynamics: From entropy theory to 5G spectrum sharing,”
spectrum access—Dynamic spectrum access in the time domain: IEEE Commun. Mag., vol. 53, no. 7, pp. 178–183, Jul. 2015.
Modeling and exploiting white space,” IEEE Commun. Mag., vol. 45, [128] J. Su and W. Wu, “Wireless spectrum prediction model based on time
no. 5, pp. 66–72, May 2007. series analysis method,” in Proc. ACM Workshop Cogn. Radio Netw.,
[106] L. Stabellini, “Quantifying and modeling spectrum opportunities in a Beijing, China, Sep. 2009, pp. 61–66.
real wireless environment,” in Proc. IEEE Wireless Commun. Netw. [129] Z. Lin, X. Jiang, L. Huang, and Y. Yao, “A energy prediction based
Conf. (WCNC), Sydney, NSW, Australia, Apr. 2010, pp. 1–6. spectrum sensing approach for cognitive radio networks,” in Proc. 5th
[107] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Int. Conf. Wireless Commun. Netw. Mobile Comput., Beijing, China,
Functions: With Formulas, Graphs, and Mathematical Tables. Sep. 2009, pp. 1–4.
New York, NY, USA: Dover, 1972. [130] S. Yarkan and H. Arslan, “Binary time series approach to spectrum
[108] S. Geirhofer, L. Tong, and B. M. Sadler, “A measurement-based model prediction for cognitive radio,” in Proc. IEEE 66th Veh. Technol. Conf.,
for dynamic spectrum access in WLAN channels,” in Proc. IEEE Baltimore, MD, USA, Sep./Oct. 2007, pp. 1563–1567.
Mil. Commun. Conf. (MILCOM), Washington, DC, USA, Oct. 2006, [131] L. Hanzo, C. Somerville, and J. Woodard, Voice and Audio
pp. 1–7. Compression for Wireless Communications. New York, NY, USA:
[109] M. Wellens, J. Riihijarvi, and P. Mahonen, “Modelling primary system Wiley, 2008.
activity in dynamic spectrum access networks by aggregated ON/OFF- [132] L. Hanzo, P. Cherriman, and J. Streit, Video Compression and
processes,” in Proc. 6th Annu. IEEE Commun. Soc. Conf. Sensor Mesh Communications: From Basics to H. 261, H. 263, H. 264, MPEG4
Ad Hoc Commun. Netw. Workshops, Rome, Italy, Jun. 2009, pp. 1–6. for DVB and HSDPA-Style Adaptive Turbo-Transceivers. Chichester,
[110] Q. Zhao and K. Liu, “Detecting, tracking, and exploiting spectrum U.K.: Wiley, 2007.
opportunities in unslotted primary systems,” in Proc. IEEE Radio [133] J. G. Proakis and D. K. Manolakis, Digital Signal Processing:
Wireless Symp., Orlando, FL, USA, Jan. 2008, pp. 491–494. Principles, Algorithms and Applications, 4th ed. New York, NY, USA:
[111] L. Melián-Gutiérrez et al., “HF spectrum activity prediction model Pearson, 2006.
based on HMM for cognitive radio applications,” Phys. Commun., [134] P. H.-Y. Wu and A. Duel-Hallen, “Multiuser detectors with dis-
vol. 9, pp. 199–211, Dec. 2013. joint Kalman channel estimators for synchronous CDMA mobile
[112] D. Willkomm, S. Machiraju, J. Bolot, and A. Wolisz, “Primary users radio channels,” IEEE Trans. Commun., vol. 48, no. 5, pp. 752–756,
in cellular networks: A large-scale measurement study,” in Proc. 3rd May 2000.
IEEE Symp. New Front. Dyn. Spectr. Access Netw. (DySPAN), Chicago, [135] P. Sharma and K. Chandra, “Prediction of state transitions in
IL, USA, Oct. 2008, pp. 1–11. Rayleigh fading channels,” IEEE Trans. Veh. Technol., vol. 56, no. 2,
[113] X. Hong, C.-X. Wang, H.-H. Chen, and Y. Zhang, “Secondary spectrum pp. 416–425, Mar. 2007.
access networks,” IEEE Veh. Technol. Mag., vol. 4, no. 2, pp. 36–43, [136] R. Matias, A. M. M. Carvalho, V. A. Teodoro, D. Tes, and L. B. Araujo,
Jun. 2009. “Performance comparison of forecasting models applied to LAN/MAN
[114] Z. Chen et al., “Interference modeling for cognitive radio networks traffic prediction,” in Proc. IEEE Workshop Local Metropolitan
with power or contention control,” in Proc. IEEE Wireless Commun. Area Netw. (LANMAN), Chapel Hill, NC, USA, Oct. 2011,
Netw. Conf. (WCNC), Sydney, NSW, Australia, Apr. 2010, pp. 1–6. pp. 1–6.
[115] Y. Zhao et al., “Applying radio environment maps to cognitive wireless [137] J. Su and W. Wu, “Wireless spectrum prediction model based on time
regional area networks,” in Proc. IEEE Symp. New Front. Dyn. Spectr. series analysis method,” in Proc. ACM Workshop Cogn. Radio Netw.,
Access Netw. (DySPAN), Dublin, Ireland, Oct. 2008, pp. 115–118. Beijing, China, Sep. 2009, pp. 61–66.
[138] A. A. Tabassam, M. U. Suleman, S. Khan, and S. H. R. Tirmazi, [161] S. Bai, X. Zhou, and F. Xu, “‘Soft decision’ spectrum prediction based
“Spectrum estimation and spectrum hole opportunities prediction on back-propagation neural networks,” in Proc. Int. Conf. Comput.
for cognitive radios using higher-order statistics,” in Proc. Wireless Manag. Telecommun. (ComManTel), Da Nang, Vietnam, Apr. 2014,
Adv. (WiAd), London, U.K., Jun. 2011, pp. 213–217. pp. 128–133.
[139] C. Stolojescu-Crisan, “Data mining based wireless network traffic fore- [162] L. Yin, S. Yin, W. Hong, and S. Li, “Spectrum behavior learning in cog-
casting,” in Proc. 10th Int. Symp. Electron. Telecommun. (ISETC), nitive radio based on artificial neural network,” in Proc. Mil. Commun.
Timişoara, Romania, Nov. 2012, pp. 115–118. Conf., Baltimore, MD, USA, Nov. 2011, pp. 25–30.
[140] C. Ghosh, C. Cordeiro, D. P. Agrawal, and M. B. Rao, “Markov chain [163] V. K. Tumuluru, P. Wang, and D. Niyato, “A neural network based
existence and hidden Markov models in spectrum sensing,” in Proc. spectrum prediction scheme for cognitive radio,” in Proc. IEEE Int.
IEEE Int. Conf. Pervasive Comput. Commun., Galveston, TX, USA, Conf. Commun. (ICC), Cape Town, South Africa, May 2010, pp. 1–5.
Mar. 2009, pp. 1–6. [164] Z. Jianli, W. Mingwei, and Y. Jinsha, “Based on neural network
[141] L. R. Rabiner, “A tutorial on hidden Markov models and selected appli- spectrum prediction of cognitive radio,” in Proc. Int. Conf. Electron.
cations in speech recognition,” Proc. IEEE, vol. 77, no. 2, pp. 257–286, Commun. Control (ICECC), Ningbo, China, Sep. 2011, pp. 762–765.
Feb. 1989. [165] O. Winston, A. Thomas, and W. OkelloOdongo, “Optimizing neural
[142] T. Black, B. Kerans, and A. Kerans, “Implementation of hidden network for TV idle channel prediction in cognitive radio using particle
Markov model spectrum prediction algorithm,” in Proc. Int. Symp. swarm optimization,” in Proc. 5th Int. Conf. Comput. Intell. Commun.
Commun. Inf. Technol. (ISCIT), Gold Coast, QLD, Australia, Oct. 2012, Syst. Netw. (CICSyN), Madrid, Spain, Jun. 2013, pp. 25–29.
pp. 280–283. [166] O. Winston, A. Thomas, O. Patrick, and O. William, “PSO of neural
[143] E. Chatziantoniou, B. Allen, and V. Velisavljevic, “An HMM- networks to predict busy times of cellular traffic for assignment to TV
based spectrum occupancy predictor for energy efficient cognitive idle channels by cognitive radio,” in Proc. Eur. Model. Symp. (EMS),
radio,” in Proc. IEEE 24th Int. Symp. Pers. Indoor Mobile Radio Manchester, U.K., Nov. 2013, pp. 48–52.
Commun. (PIMRC), London, U.K., Sep. 2013, pp. 601–605. [167] N. Baldo, B. R. Tamma, B. S. Manoj, R. R. Rao, and M. Zorzi, “A
[144] N. Thao, B. L. Mark, and Y. Ephraim, “Spectrum sensing using a hid- neural network based cognitive controller for dynamic channel selec-
den bivariate Markov model,” IEEE Trans. Wireless Commun., vol. 12, tion,” in Proc. IEEE Int. Conf. Commun. (ICC), Dresden, Germany,
no. 9, pp. 4582–4591, Sep. 2013. Jun. 2009, pp. 1–5.
[145] L. E. Baum and T. Petrie, “Statistical inference for probabilistic func- [168] N. Baldo and M. Zorzi, “Learning and adaptation in cognitive radios
tions of finite state Markov chains,” Ann. Math. Stat., vol. 37, no. 6, using neural networks,” in Proc. 5th IEEE Consum. Commun. Netw.
pp. 1554–1563, 1966. Conf., Las Vegas, NV, USA, Jan. 2008, pp. 998–1003.
[146] Z. Chen and R. C. Qiu, “Prediction of channel state for cogni- [169] S. Haykin, “Neural networks: A comprehensive foundation,” in Neural
tive radio using higher-order hidden Markov model,” in Proc. IEEE Networks, vol. 2. Basingstoke, U.K.: Palgrave, 2004.
SoutheastCon, Concord, NC, USA, Mar. 2010, pp. 276–282. [170] I. Christian, S. Moh, I. Chung, and J. Lee, “Spectrum mobility in cogni-
[147] H. A. Engelbrecht, “Efficient decoding of high-order hidden tive radio networks,” IEEE Commun. Mag., vol. 50, no. 6, pp. 114–121,
Markov models,” Ph.D. dissertation, Stellenbosch Univ., Stellenbosch, Jun. 2012.
South Africa, 2007. [171] K. Kumar, A. Prakash, and R. Tripathi, “Spectrum handoff in cognitive
[148] J. A. du Preez, “Efficient training of high-order hidden Markov models radio networks: A classification and comprehensive survey,” J. Netw.
using first-order representations,” Comput. Speech Lang., vol. 12, no. 1, Comput. Appl., vol. 61, pp. 161–188, Feb. 2016.
pp. 23–39, 1998. [172] Y.-S. Chen, C.-H. Cho, I. You, and H.-C. Chao, “A cross-layer pro-
[149] Z. Chen, “A study and implementation of speech recognition system tocol of spectrum mobility and handover in cognitive LTE networks,”
for mobile communication terminals (in Chinese),” M.S. thesis, College Simulat. Model. Pract. Theory, vol. 19, no. 8, pp. 1723–1744, 2011.
Commun. Eng., Hangzhou Dianzi Univ., Hangzhou, China, 2003. [173] W. Y. Lee and I. F. Akyildiz, “Spectrum-aware mobility management
[150] G. E. Monahan, “State of the art—A survey of partially observable in cognitive radio cellular networks,” IEEE Trans. Mobile Comput.,
Markov decision processes: Theory, models, and algorithms,” Manag. vol. 11, no. 4, pp. 529–542, Apr. 2012.
Sci., vol. 28, no. 1, pp. 1–16, 1982. [174] X.-L. Huang et al., “Historical spectrum sensing data mining for cog-
[151] L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, “Planning and act- nitive radio enabled vehicular ad-hoc networks,” IEEE Trans. Depend.
ing in partially observable stochastic domains,” Artif. Intell., vol. 101, Secure Comput., vol. 13, no. 1, pp. 59–70, Jan./Feb. 2016.
nos. 1–2, pp. 99–134, 1998. [175] P. Thakur, A. Kumar, S. Pandit, G. Singh, and S. N. Satashia,
[152] H. Fang, L. Xu, and C. Huang, “Dynamic opportunistic spectrum “Spectrum mobility in cognitive radio network using spectrum
access of multi-channel multi-radio based on game theory in wire- prediction and monitoring techniques,” Phys. Commun., vol. 24,
less cognitive network,” in Proc. IEEE 9th Int. Conf. Mobile Ad Hoc pp. 1–8, Sep. 2017.
Sensor Netw. (MSN), Dalian, China, Dec. 2013, pp. 127–132. [176] H. Zhou et al., “Toward multi-radio vehicular data piping for dynamic
[153] Q. Zhao, L. Tong, A. Swami, and Y. Chen, “Decentralized cognitive DSRC/TVWS spectrum sharing,” IEEE J. Sel. Areas Commun., vol. 34,
MAC for opportunistic spectrum access in ad hoc networks: A POMDP no. 10, pp. 2575–2588, Oct. 2016.
framework,” IEEE J. Sel. Areas Commun., vol. 25, no. 3, pp. 589–600, [177] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning
Apr. 2007. machine: Theory and applications,” Neurocomputing, vol. 70, nos. 1–3,
[154] A. T. Hoang, Y.-C. Liang, D. T. C. Wong, Y. Zeng, and R. Zhang, pp. 489–501, 2006.
“Opportunistic spectrum access for energy-constrained cognitive [178] M. I. Taj and M. Akil, “Cognitive radio spectrum evolution prediction
radios,” IEEE Trans. Wireless Commun., vol. 8, no. 3, pp. 1206–1211, using artificial neural networks based multivariate time series mod-
Mar. 2009. elling,” in Proc. 11th Eur. Wireless Conf. Sustain. Wireless Technol.
[155] C. Clancy, J. Hecker, E. Stuntebeck, and T. O’Shea, “Applications (Eur. Wireless), Vienna, Austria, 2011, pp. 1–6.
of machine learning to cognitive radio networks,” IEEE Wireless [179] A. Katidiotis, K. Tsagkaris, and P. Demestichas, “Performance evalu-
Commun., vol. 14, no. 4, pp. 47–52, Aug. 2007. ation of artificial neural network-based learning schemes for cognitive
[156] Q. Zhao and B. M. Sadler, “A survey of dynamic spectrum access,” radio systems,” Comput. Elect. Eng., vol. 36, no. 3, pp. 518–535, 2010.
IEEE Signal Process. Mag., vol. 24, no. 3, pp. 79–89, May 2007. [180] G. Dematos, M. S. Boyd, B. Kermanshahi, N. Kohzadi, and I. Kaastra,
[157] M. Golestanian, S. Iranmanesh, R. Ghazizadeh, and M. Azimi, “Feedforward versus recurrent neural networks for forecasting monthly
“A learning automata based spectrum prediction technique for cognitive Japanese yen exchange rates,” Financ. Eng. Jpn. Markets, vol. 3, no. 1,
radio networks,” Sci. Educ., vol. 2, no. 3, pp. 93–97, 2014. pp. 59–75, 1996.
[158] L. Yang, N. Lv, and Z. X. Xu, “Spectrum prediction for cognitive [181] D. Verstraeten, B. Schrauwen, M. D’Haene, and D. Stroobandt, “An
radio system based on optimally pruned extreme learning machine,” experimental unification of reservoir computing methods,” Neural
Appl. Mech. Mater., vols. 536–537, pp. 430–436, Apr. 2014. Netw., vol. 20, no. 3, pp. 391–403, 2007.
[159] Y. Miche et al., “OP-ELM: Optimally pruned extreme learning [182] H. Jaeger, “The ‘echo state’ approach to analysing and training recur-
machine,” IEEE Trans. Neural Netw., vol. 21, no. 1, pp. 158–162, rent neural networks-with an erratum note,” German Nat. Res. Center
Jan. 2010. Inf. Technol. GMD, Bonn, Germany, Tech. Rep., p. 34, 2001.
[160] H. Li, X. Xu, B. Wu, and X. Chen, “Multilayer feedforward neu- [183] L. Yang, X. Liang, T. Ma, and K. Liu, “Spectrum prediction based
ral network based efficient spectrum occupancy prediction scheme on echo state network and its improved form,” in Proc. 5th Int. Conf.
for cognitive radio system,” J. Comput. Inf. Syst., vol. 10, no. 10, Intell. Human Mach. Syst. Cybern. (IHMSC), vol. 1. Hangzhou, China,
pp. 4017–4028, 2014. Aug. 2013, pp. 172–176.
[184] X. Hu, Y. Shi, and R. C. Eberhart, “Recent advances in particle swarm,” [207] Z. Shi, X. Jiang, and L. Huang, “A prediction approach for MAC
in Proc. IEEE Congr. Evol. Comput., vol. 1. Portland, OR, USA, layer sensing in cognitive radio networks,” in Proc. IET 2nd Int.
Jul. 2016, pp. 90–97. Conf. Wireless Mobile Multimedia Netw. (ICWMMN), Beijing, China,
[185] J. Han, H. Cheng, D. Xin, and X. Yan, “Frequent pattern mining: Oct. 2008, pp. 317–320.
Current status and future directions,” Data Min. Knowl. Disc., vol. 15, [208] B. Houlding, A. Bhattacharya, S. P. Wilson, and T. K. Forde, “A
no. 1, pp. 55–86, 2007. fast Bayesian model for latent radio signal prediction,” in Proc. 7th
[186] J. Han, G. Dong, and Y. Yin, “Efficient mining of partial periodic Int. Symp. Model. Optim. Mobile Ad Hoc Wireless Netw., Seoul,
patterns in time series database,” in Proc. 15th Int. Conf. Data Eng., South Korea, 2009, pp. 1–7.
Sydney, NSW, Australia, Mar. 1999, pp. 106–115. [209] G. Ding et al., “Robust online spectrum prediction with incomplete
[187] Z. Wen et al., “A learning spectrum hole prediction model for cognitive and corrupted historical observations,” IEEE Trans. Veh. Technol., to
radio systems,” in Proc. 10th IEEE Int. Conf. Comput. Inf. Technol. be published, doi: 10.1109/TVT.2017.2693384.
(CIT), Bradford, U.K., Jun./Jul. 2010, pp. 2089–2093. [210] G. Ding et al., “Joint spectral-temporal spectrum prediction from
[188] H. Kim and K. G. Shin, “Optimal online sensing sequence in incomplete historical observations,” in Proc. IEEE Glob. Conf. Signal
multichannel cognitive radio networks,” IEEE Trans. Mobile Comput., Inf. Process. (GLOBALSIP), Atlanta, GA, USA, Dec. 2014, pp. 1–6.
vol. 12, no. 7, pp. 1349–1362, Jul. 2013. [211] G. Ding, S. Zhai, X. Chen, Y. Zhang, and C. Liu, “Robust spectral-
[189] N. I. Sapankevych and R. Sankar, “Time series prediction using support temporal two-dimensional spectrum prediction,” in Proc. EAI Int.
vector machines: A survey,” IEEE Comput. Intell. Mag., vol. 4, no. 2, Conf. Mach. Learn. Intell. Commun. (MLICOM), Shanghai, China,
pp. 24–38, May 2009. Aug. 2016, pp. 1–9.
[190] S. Ni and S. Shen, “Frequency spectrum access mechanism of [212] M. Tang, Z. Zheng, G. Ding, and Z. Xue, “Efficient TV white space
cognitive radio based on spectrum prediction,” in Proc. IET Int. database construction via spectrum sensing and spatial inference,” in
Conf. Commun. Technol. Appl. (ICCTA), Beijing, China, Oct. 2011, Proc. IEEE Int. Perform. Comput. Commun. Conf. (IPCCC), Nanjing,
pp. 556–560. China, Dec. 2015, pp. 1–5.
[191] C.-J. Yu, Y.-Y. He, and T.-F. Quan, “Frequency spectrum prediction [213] M. Tang, G. Ding, Q. Wu, Z. Xue, and T. A. Tsiftsis, “A joint tensor
method based on EMD and SVR,” in Proc. 8th Int. Conf. Intell. Syst. completion and prediction scheme for multi-dimensional spectrum map
Design Appl., vol. 3. Kaohsiung, Taiwan, Nov. 2008, pp. 39–44. construction,” IEEE Access, vol. 4, pp. 8044–8052, 2016.
[192] S. Yin, D. Chen, Q. Zhang, and S. Li, “Prediction-based throughput [214] Q. Wu et al., “A cloud-based architecture for the Internet of spec-
optimization for dynamic spectrum access,” IEEE Trans. Veh. Technol., trum devices over future wireless networks,” IEEE Access, vol. 4,
vol. 60, no. 3, pp. 1284–1289, Mar. 2011. pp. 2854–2862, 2016.
[193] H. Li, “Reconstructing spectrum occupancies for wideband cognitive [215] J. Qiu, Q. Wu, G. Ding, Y. Xu, and S. Feng, “A survey of machine
radio networks: A matrix completion via belief propagation,” in Proc. learning for big data processing,” EURASIP J. Adv. Signal Process.,
IEEE Int. Conf. Commun. (ICC), Cape Town, South Africa, May 2010, vol. 2016, no. 81, p. 67, Dec. 2016.
pp. 1–6. [216] P. Huang, C.-J. Liu, L. Xiao, and J. Chen, “Mining frequent partial peri-
[194] S. Arkoulis, E. Anifantis, V. Karyotis, S. Papavassiliou, and N. Mitrou, odic patterns in spectrum usage data,” in Proc. IEEE 20th Int. Workshop
“Discovering and exploiting spectrum power correlations in cogni- Qual. Service, Coimbra, Portugal, Jun. 2012, pp. 1–4.
tive radio networks: An experimentally driven approach,” EURASIP
[217] M. T. Ozden, “Adaptive multichannel sequential lattice prediction fil-
J. Wireless Commun. Netw., vol. 2014, no. 1, p. 17, 2014.
tering method for ARMA spectrum estimation in subbands,” EURASIP
[195] K. Hossain, B. Champagne, and A. Assra, “Cooperative multiband
J. Adv. Signal Process., vol. 2013, no. 1, pp. 1–36, 2013.
joint detection with correlated spectral occupancy in cognitive radio
[218] D. W. Browne, “Predicting communications activity in the radio
networks,” IEEE Trans. Signal Process., vol. 60, no. 5, pp. 2682–2687,
spectrum,” in Proc. 46th Conf. Rec. Asilomar Conf. Signals
May 2012.
Syst. Comput. (ASILOMAR), Pacific Grove, CA, USA, Nov. 2012,
[196] K. Hossain and B. Champagne, “Wideband spectrum sensing for cog-
pp. 1069–1073.
nitive radios with correlated subband occupancy,” IEEE Signal Process.
Lett., vol. 18, no. 1, pp. 35–38, Jan. 2011. [219] J. Qadir, “Artificial intelligence based cognitive routing for cognitive
[197] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of radio networks,” Artif. Intell. Rev., vol. 45, no. 1, pp. 25–96, Jan. 2016.
Plausible Inference. San Francisco, CA, USA: Morgan Kaufmann, [220] Z. Wang and S. Salous, “Spectrum occupancy statistics and time series
1988. models for cognitive radio,” J. Signal Process. Syst., vol. 62, no. 2,
[198] H. Li and R. C. Qiu, “A graphical framework for spectrum modeling pp. 145–155, 2011.
and decision making in cognitive radio networks,” in Proc. IEEE Glob. [221] S. D. Barnes and B. T. Maharaj, “Prediction based channel alloca-
Commun. Conf. (GLOBECOM), Miami, FL, USA, Dec. 2010, pp. 1–6. tion performance for cognitive radio,” AEU Int. J. Electron. Commun.,
[199] Z. Qin et al., “Link quality analysis based channel selection vol. 68, no. 4, pp. 336–345, 2014.
in high-frequency asynchronous automatic link establishment: A [222] G. Ning, J. Duan, J. Su, and D. Qiu, “Spectrum sharing based on spec-
matrix completion approach,” IEEE Syst. J., to be published, trum heterogeneity and multi-hop handoff in centralized cognitive radio
doi: 10.1109/JSYST.2017.2717702. networks,” in Proc. Wireless Opt. Commun. Conf. (WOCC), Newark,
[200] S.-J. Kim and G. B. Giannakis, “Cognitive radio spectrum prediction NJ, USA, Apr. 2011, pp. 1–6.
using dictionary learning,” in Proc. IEEE Glob. Commun. Conf. [223] C. Xu, Y. Li, Y. Yang, and Y. Xian, “A novel spectrum prediction
(GLOBECOM), Atlanta, GA, USA, Dec. 2013, pp. 1–6. algorithm for cognitive radio system based on chaotic neural network?”
[201] Y. Wang, Z. Z. Zhang, Q. Y. Yu, and J. M. Chen, “A novel prediction- J. Comput. Inf. Syst., vol. 9, no. 1, pp. 313–320, 2013.
based spectrum allocation mechanism for mobile cognitive radio [224] Q.-L. Ma, Q.-L. Zheng, H. Peng, T.-W. Zhong, and L.-Q. Xu, “Chaotic
networks,” KSII Trans. Internet Inf. Syst., vol. 7, no. 9, pp. 2101–2119, time series prediction based on evolving recurrent neural networks,” in
2013. Proc. Int. Conf. Mach. Learn. Cybern., vol. 6. Hong Kong, Aug. 2007,
[202] Y. Wang, Z. Zhang, L. Ma, and J. Chen, “SVM-based spectrum mobil- pp. 3496–3500.
ity prediction scheme in mobile cognitive radio networks,” Sci. World [225] Z. Yang, L. Yang, and Y. Fu, “Spectrum prediction for high-frequency
J., vol. 2014, Mar. 2014, Art. no. 395212. radar based on extreme learning machine,” in Proc. 7th Int. Conf. Adv.
[203] J. Guo, H. Ji, Y. Li, and X. Li, “A novel spectrum handoff manage- Comput. Intell., Fujian, China, Mar. 2015, pp. 235–239.
ment scheme based on SVM in cognitive radio networks,” in Proc. 6th [226] S. Bai, X. Zhou, and F. Xu, “Spectrum prediction based on
Int. ICST Conf. Commun. Netw. China (CHINACOM), Harbin, China, improved-back-propagation neural networks,” in Proc. 11th Int.
Aug. 2011, pp. 645–649. Conf. Nat. Comput. (ICNC), Zhangjiajie, China, Aug. 2015,
[204] L. Hou, K.-H. Yeung, and K.-Y. Wong, “Link availability prediction for pp. 1006–1011.
cognitive radio ad hoc networks,” J. Comput. Commun., vol. 2, no. 6, [227] Z. Zhang, K. Zhang, F. Gao, and S. Zhang, “Spectrum prediction and
pp. 18–24, 2014. channel selection for sensing-based spectrum sharing scheme using
[205] I. Bütün, A. Ç. Talay, D. T. Altilar, M. Khalid, and R. Sankar, online learning techniques,” in Proc. 26th Annu. Int. Symp. Pers.
“Impact of mobility prediction on the performance of cognitive radio Indoor Mobile Radio Commun. (PIMRC), Hong Kong, Aug./Sep. 2015,
networks,” in Proc. Wireless Telecommun. Symp. (WTS), Tampa, FL, pp. 355–359.
USA, Apr. 2010, pp. 1–5. [228] H. Eltom, S. Kandeepan, B. Moran, and R. J. Evans, “Spectrum occu-
[206] Z. Wang and S. Salous, “Time series ARIMA model of spectrum occupancy prediction using a hidden Markov model,” in Proc. Int. Conf.
pancy for cognitive radio,” in Proc. IET Seminar Cogn. Radio Softw. Signal Process. Commun. Syst. (ICSPCS), Cairns, QLD, Australia,
Defined Radios Technol. Techn., London, U.K., Sep. 2008, pp. 1–4. Dec. 2015, pp. 1–8.
[229] A. Eltholth, “Forward backward autoregressive spectrum prediction [249] Microsoft Research WhiteSpaceFinder. Accessed: Jul. 2, 2017.
scheme in cognitive radio systems,” in Proc. 9th Int. Conf. Signal [Online]. Available: http://whitespaces.msresearch.us/WSWebGUI/
Process. Commun. Syst. (ICSPCS), Cairns, QLD, Australia, Dec. 2015, whitespaces.aspx
pp. 1–5. [250] White Space Database-Capitalize on Dynamic Spectrum Access
[230] Y. Zhao, Z. Hong, G. Wang, and J. Huang, “High-order hidden bivari- Opportunities, Telcordia, Piscataway, NJ, USA, 2010. [Online].
ate Markov model: A novel approach on spectrum prediction,” in Available: https://prism.telcordia.com/tvws/main/home/index.shtml
Proc. 25th Int. Conf. Comput. Commun. Netw. (ICCCN), Waikoloa, [251] Y. Zhao, J. Gaeddert, K. K. Bae, and J. H. Reed, “Radio environment
HI, USA, Aug. 2016, pp. 1–7. map enabled situation-aware cognitive radio learning algorithms,” in
[231] H. Eltom, S. Kandeepan, Y. C. Liang, B. Moran, and R. J. Evans, Proc. Tech. Conf. Softw. Defined Radio Forum (SDRF), 2006, pp. 1–6.
“HMM based cooperative spectrum occupancy prediction using hard [252] Y. Zhao, “Enabling cognitive radios through radio environment maps,”
fusion,” in Proc. IEEE Int. Commun. Workshops (ICC), Kuala Lumpur, Ph.D. dissertation, Virginia Polytech. Inst. State Univ., Blacksburg, VA,
Malaysia, May 2016, pp. 669–675. USA, 2007.
[232] A. A. Eltholth, “Spectrum prediction in cognitive radio systems using [253] H. B. Yilmaz, T. Tugcu, F. Alagöz, and S. Bayhan, “Radio environment
a wavelet neural network,” in Proc. 24th Int. Conf. Softw. Telecommun. map as enabler for practical cognitive radio networks,” IEEE Commun.
Comput. Netw. (SoftOOM), Split, Croatia, Sep. 2016, pp. 1–6. Mag., vol. 51, no. 12, pp. 162–169, Dec. 2013.
[233] A. Saad, B. Staehle, and R. Knorr, “Spectrum prediction using hidden [254] J. Perez-Romero et al., “On the use of radio environment maps for
Markov models for industrial cognitive radio,” in Proc. IEEE 12th Int. interference management in heterogeneous networks,” IEEE Commun.
Conf. Wireless Mobile Comput. Netw. Commun. (WiMob), New York, Mag., vol. 53, no. 8, pp. 184–191, Aug. 2015.
NY, USA, Oct. 2016, pp. 1–7. [255] S. Debroy, S. Bhattacharjee, and M. Chatterjee, “Spectrum map and
its application in resource management in cognitive radio networks,”
[234] F. Hou, X. Chen, H. Huang, and X. Jing, “Throughput performance
IEEE Trans. Cogn. Commun. Netw., vol. 1, no. 4, pp. 406–419,
improvement in cognitive radio networks based on spectrum
Dec. 2016.
prediction,” in Proc. 6th Int. Symp. Commun. Inf. Technol. (ISCIT),
[256] B. Sayrac and S. B. Jemaa, “Radio environmental maps (REMs):
Qingdao, China, Sep. 2016, pp. 655–658.
A cognitive tool for environmental awareness,” Pascal Cordier
[235] M. Huk and J. Mizera-Pietraszko, “Contextual neural-network based Orange Labs, Paris, France, Tech. Rep. [Online]. Available:
spectrum prediction for cognitive radio,” in Proc. 4th Int. Conf. http://www-syscom.univ-mlv.fr/~najim/gdr-ecoradio/sayrac.pdf
Future Gener. Commun. Technol. (FGCT), Luton, U.K., Jul. 2015, [257] Z. Quan, S. Cui, and A. H. Sayed, “Optimal linear cooperation for
pp. 1–5. spectrum sensing in cognitive radio networks,” IEEE J. Sel. Topics
[236] L. Yu, J. Chen, G. Ding, and Z. Qin, “Fast automatic link estab- Signal Process., vol. 2, no. 1, pp. 28–40, Feb. 2008.
lishment: A new metric and the value of spectrum prediction,” in [258] J. G. Andrews et al., “What will 5G be?” IEEE J. Sel. Areas Commun.,
Proc. 7th Wireless Commun. Signal Process. (WCSP), Yangzhou, vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
China, Oct. 2016, pp. 1–6. [259] J. Zhao and J. Wang, “Joint optimization algorithm based on central-
[237] J. Zhang, G. Ding, Y. Xu, and F. Song, “On the usefulness of spec- ized spectrum sharing for cognitive radio,” in Proc. IEEE Int. Conf.
trum prediction for dynamic spectrum access,” in Proc. 7th Wireless Commun. (ICC), London, U.K., Jun. 2015, pp. 7653–7658.
Commun. Signal Process. (WCSP), Yangzhou, China, Oct. 2016, [260] J. Mitola, III, et al., “Accelerating 5G QoE via public-private spec-
pp. 1–4. trum sharing,” IEEE Commun. Mag., vol. 52, no. 5, pp. 77–85,
[238] M. Tang, G. Ding, Z. Xue, J. Zhang, and H. Zhou, “Multi-dimensional May 2014.
spectrum map construction: A tensor perspective,” in Proc. 7th Wireless [261] E. E. Johnson, S. Cook, R. I. Desourdis, G. Earle, and J. C. Ostergaard,
Commun. Signal Process. (WCSP), Yangzhou, China, Oct. 2016, Advanced High-Frequency Radio Communications. Boston, MA, USA:
pp. 1–5. Artech House, 1997.
[239] K.-M. Kang et al., “Deployment and coverage of cognitive radio [262] E. E. Johnson, E. Koski, W. N. Furman, M. Jorgenson, and J. Nieto,
networks in TV white space,” IEEE Commun. Mag., vol. 50, no. 12, Third-Generation and Wideband HF Radio Communications. Norwood,
pp. 88–94, Dec. 2012. MA, USA: Artech House, 2012.
[240] FCC Adopt Rule for Unlicensed Use of Television White Spaces, [263] J. Wang, Research and Development of HF Digital Communications
Federal Commun. Commission, Washington, DC, USA, 2010. [Online]. (in Chinese). Beijing, China: Science Press, 2013.
Available: http://www.fcc.gov [264] E. Koski and W. N. Furman, “Applying cognitive radio concepts to
[241] “Office of engineering and technology authorizes TV white space HF communications,” in Proc. 11th Int. Conf. Ionospheric Radio Syst.
database administrators to provide service to unlicensed devices Tech. (IRST), Edinburgh, U.K., 2009, pp. 1–6.
operating on unused TV spectrum nationwide,” Federal Commun. [265] W. N. Furman and E. Koski, “Next generation ALE concepts,” in
Commission, Washington, DC, USA, Tech. Rep. ET Docket 04-186, Proc. 11th Int. Conf. Ionospheric Radio Syst. Tech. (IRST), Edinburgh,
Mar. 2013. U.K., 2009, pp. 1–5.
[242] R. Murty, R. Chandra, T. Moscibroda, and P. Bahl, “SenseLess: A [266] A. Shahid, S. Ahmad, A. Akram, and S. A. Khan, “Cognitive ALE
database-driven white spaces network,” IEEE Trans. Mobile Comput., for HF radios,” in Proc. 2nd Int. Conf. Comput. Eng. Appl. (ICCEA),
vol. 11, no. 2, pp. 189–203, Feb. 2012. vol. 2. Bali, Indonesia, Mar. 2010, pp. 28–33.
[267] W. N. Furman, E. Koski, and J. W. Nieto, “Design concepts for a
[243] “System architecture for information exchange between different
wideband HF ALE capability,” in Proc. 12th IET Int. Conf. Ionospheric
geo-location databases (GLDBs) enabling the operation of white
Radio Syst. Tech. (IRST), York, U.K., May 2012, pp. 1–5.
space devices (WSDs),” ETSI, Sophia Antipolis, France, Tech.
[268] R. C. Qiu et al., “Cognitive radio network for the smart grid:
Rep. TS 103 143 V1.1.1 (2015-01), 2015. [Online]. Available:
Experimental system architecture, control algorithms, security, and
http://www.etsi.org/deliver
microgrid testbed,” IEEE Trans. Smart Grid, vol. 2, no. 4 pp. 724–740,
[244] “Interference and dynamic spectrum access,” CSMAC Committee, Dec. 2011.
Nat. Telecommun. Inf. Admin., Washington, DC, USA, Tech. [269] O. B. Akan, O. B. Karli, and O. Ergul, “Cognitive radio sensor
Rep., Nov. 2010. [Online]. Available: http://www.ntia.doc.gov/ networks,” IEEE Netw., vol. 23, no. 4, pp. 34–40, Jul./Aug. 2009.
legacy/advisory/spectrum/reports [270] A. O. Bicen, V. C. Gungor, and O. B. Akan, “Delay-sensitive and
[245] Institute for Telecommunication Sciences, “Developing forward think- multimedia communication in cognitive radio sensor networks,” Ad
ing rules and processes to fully exploit spectrum resources: Case Hoc Netw. J., vol. 10, no. 5, pp. 816–830, Jul. 2012.
study exploring approaches for real-time federal spectrum sharing,” in [271] M. Ozger, E. A. Fadel, and O. B. Akan, “Event-to-sink spectrum-aware
Proc. 13th Annu. Int. Symp. Adv. Radio Technol., Boulder, CO, USA, clustering in mobile cognitive radio sensor networks,” IEEE Trans.
Jul. 2012, pp. 1–192. Mobile Comput., vol. 15, no. 9, pp. 2221–2233, Sep. 2016.
[246] J. Wang, G. Ding, Q. Wu, L. Shen, and F. Song, “Spatial-temporal [272] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for
spectrum hole discovery: A hybrid spectrum sensing and geolocation modeling, analysis, and design of multi-tier and cognitive cellular wire-
database framework,” Chin. Sci. Bull., vol. 59, no. 16, pp. 1896–1902, less networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 3,
Jun. 2014. pp. 996–1019, 3rd Quart., 2013.
[247] SpectrumBridge. Accessed: Jul. 2, 2017. [Online]. Available: http:// [273] M. Chen, W. Saad, and C. Yin, “Optimized uplink-downlink decou-
whitespaces.spectrumbridge.com pling in LTE-U networks: An echo state approach,” in Proc. IEEE Int.
[248] Google Spectrum Database. Accessed: Jul. 2, 2017. [Online]. Conf. Commun. Mobile Wireless Netw. Symp. (ICC), Kuala Lumpur,
Available: http://www.google.org/spectrum/whitespace/ Malaysia, May 2016, pp. 1–6.
[274] M. Chen, W. Saad, C. Yin, and M. Debbah, “Echo state networks [297] J. Riihijärvi, J. Nasreddine, and P. Mähönen, “Demonstrating radio
for proactive caching and content prediction in cloud radio access environment map construction from massive data sets,” in Proc. IEEE
networks,” in Proc. IEEE Glob. Commun. Conf. Workshop 5G RAN DySPAN, Bellevue, WA, USA, Oct. 2012, pp. 266–267.
Design (GLOBECOM), Washington, DC, USA, Dec. 2016, pp. 1–6. [298] A. Palaios, J. Riihijärvi, O. Holland, A. Achtzehn, and P. Mähönen,
[275] Y. Zhang et al., “Cognitive machine-to-machine communications: “Measurements of spectrum use in London: Exploratory data anal-
Visions and potentials for the smart grid,” IEEE Netw., vol. 26, no. 3, ysis and study of temporal, spatial and frequency-domain dynam-
pp. 6–13, May/Jun. 2012. ics,” in Proc. IEEE DySPAN, Bellevue, WA, USA, Oct. 2012,
[276] A. McAfee, E. Brynjolfsson, T. H. Davenport, D. Patil, and D. Barton, pp. 154–165.
“Big data,” Manag. Revolution Harvard Bus. Rev., vol. 90, no. 10,
pp. 61–67, 2012.
[277] W.-B. Chien, C.-K. Yang, and Y.-H. Huang, “Energy-saving cooper-
ative spectrum sensing processor for cognitive radio system,” IEEE
Trans. Circuits Syst. I, Reg. Papers, vol. 58, no. 4, pp. 711–723, Guoru Ding (S’10–M’14–SM’16) received the B.S.
Apr. 2011. (Hons.) degree in electrical engineering from Xidian
[278] T. Jing, X. Xing, W. Cheng, Y. Huo, and T. Znati, “Cooperative University, Xi’an, China, in 2008 and the Ph.D.
spectrum prediction in multi-PU multi-SU cognitive radio networks,” (Hons.) degree in communications and informa-
in Proc. 8th Int. Conf. Cogn. Radio Orient. Wireless Netw. tion systems from the College of Communications
(CROWNCOM), Washington, DC, USA, Jul. 2013, pp. 25–30. Engineering, Nanjing, China, in 2014. Since 2014,
[279] R. Umar and A. U. H. Sheikh, “A comparative study of spectrum he has been an Assistant Professor with the
awareness techniques for cognitive radio oriented wireless networks,” College of Communications Engineering and a
Phys. Commun., vol. 9, pp. 148–170, Dec. 2013. Research Fellow with the National High Frequency
[280] T. O’Shea and T. C. Clancy, “Deep reinforcement learning radio control Communications Research Center of China. Since
and signal detection with KeRLym, a Gym RL agent,” arXiv preprint 2015, he has been a Post-Doctoral Research
arXiv:1605.09221, 2016. Associate with the National Mobile Communications Research Laboratory,
[281] R. McGwier, T. O’Shea, K. Karra, and M. Fowler, “Recent develop- Southeast University, Nanjing. His research interests include cognitive radio
ments in artificial intelligence applications of deep learning for signal networks, massive MIMO, machine learning, and big data analytics over wire-
processing,” in Proc. Virginia Tech Wireless Symp., 2016. less networks.
[282] F. H. Sanders, R. L. Hinkle, and B. J. Ramsey, “Measurement He was a recipient of the best paper awards from EAI MLICOM 2016, the
procedures for the radar spectrum engineering criteria (RSEC),” IEEE Vehicular Technology Conference (VTC) 2014-Fall, and IEEE WCSP
Nat. Telecommun. Inf. Admin., Washington, DC, USA, Tech. 2009, the Alexander von Humboldt Fellowship in 2017, and the Excellent
Rep. 05-420, 2005. [Online]. Available: https://www.its.bldrdoc.gov/ Doctoral Thesis Award of China Institute of Communications in 2016. He
publications/download/TR-05-420.pdf has served as a Guest Editor for the IEEE J OURNAL ON S ELECTED A REAS
[283] F. H. Sanders, E. F. Drocella, and R. L. Sole, “Using on-shore IN C OMMUNICATIONS (Special issue on spectrum sharing and aggregation
detected radar signal power for interference protection of off- in future wireless networks). He is currently an Associate Editor of the
shore radar receivers,” Nat. Telecommun. Inf. Admin., Washington, Journal of Communications and Information Networks, the KSII Transactions
DC, USA, Tech. Rep. 16-521, Mar. 2016. [Online]. Available: on Internet and Information Systems, and the AEU-International Journal
https://www.its.bldrdoc.gov/publications/\download/TR-16-521.pdf of Electronics and Communications. He has acted as a Technical Program
[284] Z. Khan et al., “IoT connectivity in radar bands: A shared access model Committees Member for a number of international conferences, including the
based on spectrum measurements,” IEEE Commun. Mag., vol. 55, no. 2, IEEE Global Communications Conference, the IEEE International Conference
pp. 88–96, Feb. 2017. on Communications, and VTC. He is a Voting Member of the IEEE 1900.6
[285] S. Bhattarai, J.-M. J. Park, B. Gao, K. Bian, and W. Lehr, “An overview Standard Association Working Group.
of dynamic spectrum sharing: Ongoing initiatives, challenges, and a
roadmap for future research,” IEEE Trans. Cogn. Commun. Netw.,
vol. 2, no. 2, pp. 110–128, Jun. 2016.
[286] F. Hessar and S. Roy, “Spectrum sharing between a surveillance radar
and secondary Wi-Fi networks,” IEEE Trans. Aerosp. Electron. Syst., Yutao Jiao received the B.S. degree from the
vol. 52, no. 3, pp. 1434–1448, Jun. 2016. College of Communications Engineering, Nanjing,
[287] S. W. Boyd, J. M. Frye, M. B. Pursley, and T. C. Royster, IV, “Spectrum China, in 2013. He is currently pursuing the Ph.D.
monitoring during reception in dynamic spectrum access cognitive degree with the School of Computer Science and
radio networks,” IEEE Trans. Commun., vol. 60, no. 2, pp. 547–558, Engineering, Nanyang Technological University. His
Feb. 2012. research interests include resource management in
[288] A. Ali and W. Hamouda, “Spectrum monitoring using energy ratio the Internet of Things and economics of big data.
algorithm for OFDM-based cognitive radio networks,” IEEE Trans.
Wireless Commun., vol. 14, no. 4, pp. 2257–2268, Apr. 2015.
[289] T. Ihalainen, A. Viholainen, T. H. Stitz, and M. Renfors, “Spectrum
monitoring scheme for filter bank based cognitive radios,” in Proc.
Future Netw. Mobile Summit, Florence, Italy, Jun. 2010, pp. 1–9.
[290] NATO SfP Project ORCA. Accessed: Jul. 2, 2017. [Online]. Available:
http://orca.feit.ukim.edu.mk/index.html
[291] O. Fratu et al., “Comparative study of radio mobile and ICS telecom Jinlong Wang received the B.S. degree in wire-
propagation prediction models for DVB-T,” in Proc. IEEE Int. Symp. less communications and the M.S. and Ph.D. degrees
Broadband Multimedia Syst. Broadcast. (BMSB), Ghent, Belgium, in communications and electronic systems from the
Jun. 2015, pp. 1–6. Institute of Communications Engineering, Nanjing,
[292] P. Lazaridis et al., “Evaluation of prediction accuracy for the Longley- China, in 1983, 1986, and 1992, respectively. He
rice model in the FM and TV bands,” in Proc. IX Int. Conf. EATI, is currently a Professor with the PLA University of
Ohrid, Macedonia, Sep. 2013, pp. 1–4. Science and Technology, Nanjing, China.
[293] WiSHFUL Project. Accessed: Jul. 2, 2017. [Online]. Available: He is the Co-Chair of the IEEE Nanjing Section.
http://www.wishful-project.eu/ His current research interests span a wide range
[294] CREW Project ORCA. Accessed: Jul. 2, 2017. [Online]. Available: of topics in wireless communications and signal
http://www.crew-project.eu/ processing, including cognitive radio networks, HF
[295] M. Pesko et al., “The indirect self-tuning method for constructing communications, cooperative communications, and wireless security. He has
radio environment map using omnidirectional or directional transmit- published extensively in internationally renowned journals, including the
ter antenna,” EURASIP J. Wireless Commun. Netw., vol. 2015, no. 1. IEEE Communications Magazine, the IEEE Signal Processing Magazine, the
pp. 1–12, Dec. 2015. IEEE T RANSACTIONS ON C OMMUNICATIONS, the IEEE T RANSACTIONS
[296] FARAMIR Project. Accessed: Jul. 2, 2017. [Online]. Available: ON W IRELESS C OMMUNICATIONS , and the IEEE T RANSACTIONS ON
http://www.ict-faramir.eu V EHICULAR T ECHNOLOGY in the above areas.
Yulong Zou (S’07–M’12–SM’13) received Yu-Dong Yao (S’88–M’88–SM’94–F’11) is cur-

the B.Eng. degree in information engineer- rently a Professor and the Department Director of
ing from the Nanjing University of Posts and electrical and computer engineering. He is also the
Telecommunications (NUPT), Nanjing, China, Director of Stevens Wireless Information Systems
in 2006, the first Ph.D. degree from the Stevens Engineering Laboratory. He has been with the
Institute of Technology, NJ, USA, in 2012, and the Stevens Institute of Technology, Hoboken, NJ, USA,
second Ph.D. degree from NUPT, in 2012, where since 2000. From 1989 and 1990, he was with
he is a Full Professor. Carleton University, Ottawa, Canada, as a Research
He is currently serving as an Editor for the IEEE Associate researching on mobile radio communi-
C OMMUNICATIONS S URVEYS & T UTORIALS, cations. From 1990 to 1994, he was with Spar
the IEEE C OMMUNICATIONS L ETTERS, the Aerospace Ltd., Montreal, Canada, where he was
EURASIP Journal on Advances in Signal Processing, and KSII Transactions involved in research on satellite communications. From 1994 to 2000, he was
on Internet and Information Systems. He is also serving as a Lead with Qualcomm Inc., San Diego, CA, USA, where he participated in research
Guest Editor for special issue on Security Challenges and Issues in and development in wireless code-division multiple-access systems.
Cognitive Radio Networks in the EURASIP Journal on Advances in He received the B.Eng. and M.Eng. degrees from the Nanjing University
Signal Processing. He has acted as symposium chairs, session chairs, and of Posts and Telecommunications, Nanjing, China, in 1982 and 1985, respec-
TPC members for a number of IEEE sponsored conferences, including tively, and the Ph.D. degree from Southeast University, Nanjing, in 1988, all
the IEEE Wireless Communications and Networking Conference, IEEE in electrical engineering. He holds one Chinese patent and 12 U.S. patents.
Global Telecommunications Conference, IEEE International Conference on His research interests include wireless communications and networks, spread
Communications, IEEE Vehicular Technology Conference, the International spectrum and CDMA, antenna arrays and beamforming, cognitive and soft-
Conference on Communications in China, and International Conference on ware defined radio, and digital signal processing for wireless systems. He was
Communications and Networking in China. His research interests span a wide an Associate Editor of the IEEE C OMMUNICATIONS L ETTERS and the IEEE
range of topics in wireless communications and signal processing, including T RANSACTIONS ON V EHICULAR T ECHNOLOGY, and an Editor of the IEEE
the cooperative communications, cognitive radio, wireless security, and green T RANSACTIONS ON W IRELESS C OMMUNICATIONS.
communications. He has published extensively in internationally renowned
journals, including the IEEE T RANSACTIONS ON S IGNAL P ROCESSING,
the IEEE T RANSACTIONS ON C OMMUNICATIONS, the IEEE J OURNAL ON
S ELECTED A REAS IN C OMMUNICATIONS, the IEEE T RANSACTIONS ON
W IRELESS C OMMUNICATIONS, the IEEE T RANSACTIONS ON V EHICULAR Lajos Hanzo received the master’s degree in elec-
T ECHNOLOGY, and the IEEE Communications Magazine in the above areas. tronics and the Doctorate degree from the Technical
University of Budapest in 1976 and 1983, respec-
tively, the Honorary Doctorate (Doctor Honoris
Causa) degree in 2010, and Doctor of Sciences
(D.Sc.) degree from the University of Southampton,
U.K., in 2004. Since 1986, he has been with
the University of Southampton. During his 36-year
career in telecommunications he has held var-
ious research and academic posts in Hungary,
Germany, and the U.K. Since 1986, he has been
an Academic Staff Member with the School of Electronics and Computer
Science, University of Southampton, where he currently holds the Chair in
Qihui Wu (SM’13) received the B.S. degree in Telecommunications and the Head of the Communications Research Area. He
communications engineering and the M.S. and is also a Chaired Professor with Tsinghua University, Beijing, China.
Ph.D. degrees in communications and informa- He has co-authored 20 Wiley/IEEE Press books totalling about 10 000
tion systems from the Institute of Communications pages on mobile radio communications, and published over 1200 research
Engineering, Nanjing, China, in 1994, 1997, and papers and book chapters at IEEE Xplore. He was a recipient of the num-
2000, respectively. From 2003 to 2005, he was a ber of distinctions, most recently the IEEE Wireless Technical Committee
Post-Doctoral Research Associate with Southeast Achievement Award in 2007, the IET Sir Monti Finniston Achievement Award
University, Nanjing. From 2005 to 2007, he across all disciplines of engineering in 2008, and the Honorary Doctorate
was an Associate Professor with the College of of the Technical University of Budapest in 2010. He has also organized
Communications Engineering, PLA University of and chaired major IEEE conferences, such as WCNC2006, WCNC2009,
Science and Technology, Nanjing, where he served VTC2011, presented Tutorial/overview lectures at international conferences.
as a Full Professor from 2008 to 2016. Since 2016, he has been a Full He presented a number of named lectures and keynotes. He is also an IEEE
Professor with the College of Electronic and Information Engineering, Nanjing Distinguished Lecturer of both the Communications Society and the Vehicular
University of Aeronautics and Astronautics, Nanjing. In 2011, he was an Society and a fellow of the IEEE, IEE/IET, and the Royal Academy of
Advanced Visiting Scholar with the Stevens Institute of Technology, Hoboken, Engineering. He is acting as a Governor of the IEEE VTS as well as of
USA, for six months. His current research interests span the areas of wireless ComSoc. He is the Editor-in-Chief of the IEEE P RESS. His most recent paper
communications and statistical signal processing, with emphasis on system awards are: WCNC2007 in Hong Kong, ICC’2009 Dresden, and ICC’2010
design of software defined radio, cognitive radio, and smart radio. Cape Town.

Spectrum Inference in Cognitive Radio Networks 2018

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Spectrum Inference in Cognitive Radio Networks 2018

Hochgeladen von

Copyright:

Verfügbare Formate

150 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO.

1, FIRST QUARTER 2018

Spectrum Inference in Cognitive Radio Networks:

Fig. 1. The structure of this paper.

Fig. 2. An illustrative scenario in the time-frequency-spatial dimension.

where (t) is a variable, which is a function of the dis-

Fig. 5. Linear prediction (LP) structure.

AR model to form an autoregressive moving average model

model does not make full use of the historical infor-

Once the training process is completed, the future spectrum

C. Experimental Spectrum Inference Approaches

from a transmitter to a receiver. The reasons behind this

Yulong Zou (S’07–M’12–SM’13) received Yu-Dong Yao (S’88–M’88–SM’94–F’11) is cur-

Das könnte Ihnen auch gefallen