Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
doi:10.11916 / j.issn. 1005⁃9113.2015.05.010
Estimation for Traffic Arrival Rate and Service Rate of Primary Users
in Cognitive Radio Networks
Xiaolong Yang and Xuezhi Tan ∗
( Communication Research Center, Harbin Institute of Technology, Harbin 150001, China)
Abstract: In order to estimate the traffic arrival rate and service rate parameters of primary users in cognitive
radio networks, a hidden Markov model estimation algorithm ( HMM⁃EA) is proposed, which can provide better
estimation performance than the energy detection estimation algorithm ( ED⁃EA ) . Firstly, spectrum usage
behaviors of primary users are described by establishing a preemptive priority queue model, by which a real
state transition probability matrix is derived. Secondly, cooperative detection is utilized to detect the real state of
primary users and emission matrix is derived by considering both detection and false alarm probability. Then, a
hidden Markov model is built based on the previous two steps, and evaluated through the forward⁃backward
algorithm. Finally, the simulations results verify that the HMM⁃EA algorithm outperforms the ED⁃EA in terms of
convergence performance, and therefore the secondary user is able to access the unused channel with the least
busy probability in real time.
Keywords: cognitive radio; hidden Markov model; cooperative detection
CLC number: TN929 5 Document code: A Article ID: 1005⁃9113(2015)05⁃0061⁃08
1 Introduction
Cognitive radio has been widely considered as a
promising technique to solve spectrum scarcity
problem, which is caused by the fixed spectrum policy
and the underutilization of the allocated spectrum at a
time, frequency, or space [1] . Cognitive radio network
( CRN ) allows a secondary user ( SU ) to
opportunistically access or share the licensed bands
with the primary user ( PU ) , thereby extremely
improving the spectrum efficiency [2] . Specifically, in
CRNs, a spectrum management framework includes
four interrelated sections: spectrum sensing, spectrum
decision, spectrum sharing, and spectrum mobility [3] .
Spectrum sensing is interpreted as an action, by which
SUs search for available spectrum holes among licensed
spectrum bands for communication. Based on sensing
information of spectrum holes, a SU selects the best
channel for accessing by spectrum decision. Then,
spectrum sharing will coordinate channel access among
multiple SUs. As long as a PU reclaims a licensed
channel temporarily occupied by a SU, spectrum
mobility will immediately suspend the transmission,
vacate the channel, and resume ongoing communication
or retransmit data on another fallow channel.
Spectrum mobility mainly refers to spectrum
handoff which has been widely considered as a
challenge in different research programs related to
heterogeneous networks and CRNs [4] . Up to now, the
vast majority of researches on spectrum handoff focus on
establishing handoff models and studying handoff
strategies, which are generally divided into two
categories: proactive spectrum handoff and reactive
spectrum handoff [5] . In proactive spectrum handoff, the
target channels prepared for future handoff have already
been determined by intelligent prediction techniques or
long⁃term statistics before a SU starts to establish a
connection. In reactive spectrum handoff, as a SU is
required to vacate a licensed channel, the SU will
suspend the ongoing communication and execute
spectrum sensing to seek for idle channel first, and
then the available channel would be selected according
to the sensing results. Essentially, both spectrum
handoff schemes are aimed at resolving the same
problem, i.e. the optimal approach to choose the target
channel when more than one idle channel exists.
Apparently, the SU gives preference to the channel
with the lowest busy probability which depends on the
traffic arrival rate and the traffic service rate of the PU.
So, it is very important to estimate these parameters
before data transmission.
In the existing works, the hidden Markov model
( HMM) is widely applied to solve the spectrum
Received 2014-08-01.
Sponsored by the National Natural Science Foundation of China ( Grant No. 61071104) .
∗Corresponding author. E⁃mail: tanxz1957@ hit.edu.cn.
· 61·
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
prediction and parameter estimation problem in CRNs.
Park proposed the HMM⁃based channel status predictor
to implement a channel predictor and predict next
channel status based on past channel states and analyze
the disadvantage of the predictor [6] . In Ref. [ 7 ] , a
modified HMM⁃based single SU prediction was
proposed and examined based on two hardware
platforms ( the universal software radio peripheral 2 and
the small form factor software defined radio
development platform ) . In addition, cooperative
prediction with two stages was also proposed. In order
to convince the theory results, real⁃world Wi⁃Fi signals
are applied in simulation environment. In Ref. [ 8] , a
HMM⁃based channel prediction algorithm was proposed
as well as a channel allocation algorithm, resulting in
the improvement of throughput in the frequency⁃
hopping scheme. All these literatures above, however,
predict the states of the PU ( busy or idle) but do not
estimate the traffic arrival rate and the traffic service
rate of the licensed channels, further. Therefore, the
SU may not be able to access the optimal channel,
which directly results in more interruptions and longer
handoff delay.
Additionally, Ref. [ 9 ] presented an analytical
framework to study the proactive handoff strategies by
evaluating the latency performance of spectrum
handoffs in CRNs. Furthermore, the authors proposed a
traffic⁃adaptive spectrum handoff strategy based on
traffic parameters, i. e. the traffic arrival rate and the
traffic service rate of the PU. But, they discussed little
about how to obtain the traffic conditions. Similarly,
Ref.[10] researched the handoff characteristics on the
latency performance in the case of reactive handoff
strategy. According to traffic parameters, the authors
also provided an important insight to study the channel
utilization. However, they still did not mention the way
to get traffic parameters. In Ref. [ 11] , the CRN was
modeled as a HMM, and the traffic arrival rate and the
traffic service rate of the licensed channels were
estimated by energy detection estimation algorithm
( ED⁃EA ) . However, the SUs could not accurately
detect the real state of the PU in the case of low signal
to noise ratio or hidden terminal problem. In order to
solve aforementioned problems, a hidden Markov
model estimation algorithm ( HMM⁃EA ) based on
cooperative detection was proposed to estimate the
traffic arrival rate and service rate parameters of the
PU, which is better than ED⁃EA algorithm in terms of
convergence performance.
signals while the other one is omnidirectional antenna
for spectrum detection. In addition, each channel is
considered with its own high⁃priority and low⁃priority
queues as discussed in Ref. [ 12] . In order to transmit
data, PUs and SUs enter the high⁃priority and low⁃
priority queues, respectively. Then, the primary
connection and secondary connection are established,
respectively. Here, in spite of uplink or downlink, the
connection with the same priority is considered to
follow the first⁃come⁃first⁃served ( FCFS ) scheduling
policy in a centralized manner. The CRN is supposed to
adopt time⁃slotted scheme as implemented in
Refs.[13 - 14] .
In order to avoid interfering with PUs, the SU
must sense the operation state of current channel at the
beginning of every time slot. In single⁃user spectrum
detection process, energy spectrum detection is
adopted and licensed channel state is directly
determined by the SU according to its detection result.
In multiple⁃user spectrum detection process,
cooperative energy spectrum detection is applied and
licensed channel state is judged by fusion center based
on detection results from cooperative SUs. If the state is
busy, the SU has to execute spectrum handoff to
retransmit its unfinished data. Otherwise, the SU is
allowed to transmit data in remaining duration of this
time slot.
Spectrum usage behaviors of PUs are analyzed by
M / M / 1 / N queuing network model, which requires the
following traffic parameters. It is assumed that the
arrival processes of primary connection on the queue of
each channel are Poisson. Furthermore, the
corresponding service time follows exponential
distribution. N represents the cache capacity of each
channel, namely queue length. Let n denote the number
of PU in high⁃priority queues, and n ∈ {0,1,2,…,N}.
Fig.1 shows the schematic diagram of M / M / 1 / N
queuing network model, where λ( n) ( arrivals / slot )
denotes the arrival rate of transition from staten to n +
1, and μ( n) ( slots / arrival) denotes the service rate of
transition from staten to n + 1. From the point of SU,
when queuing network model satisfies n = 0, the
channel stays idle, presented by S 0 . Otherwise, the
channel stays busy, presented by S 1 .
2 Cognitive Radio System Model
Fig.1 Schematic diagram of M/ M/ 1/ N queuing network model
In this paper, it assumes that the CRN has M SU
SUs, and each SU equips with two sets of antenna, one
is directional antenna for transmitting or receiving
Define the probability function P n( t) as the
steady⁃state probability that the given process is in state
n at timet. Based on the above assumptions and
· 62·
λ(
0)λ(
1) λ(
2) λ(
N-2)λ(
N-1)
0
1
2
…
N-1
N
μ(
1) μ(
2) μ(
3) μ(
N-1) μ(
N)
Bus
y
(
S1)
I
dl
e
(
S0)
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
analysis, the equilibrium equations of the given process
can be written as follows,
ìï∂P n( t)
= - λ(0) P 0( t) + μ(1) P 1( t) , n = 0
ï ∂t
ï∂P (t)
ï n
= λ(n - 1)Pn-1(t) - (λ(n) + μ(n))Pn(t) +
í ∂t
ï μ( n + 1) P ( t) , 1 ≤ n ≤ N - 1
n+1
ï
∂P
(t)
ï N
ï ∂t = λ(N - 1)PN-1(t) - μ(N)PN(t), n = N
î
(1)
When the process goes steady, for all n(n ∈ {0,1,
2,…,N}), the limit of the steady⁃state probability as t
approaches to infinity exists. That is, lim∂Pn(t) / ∂t = 0
t→∞
and limPn(t) = Pn . Then, the equilibrium equations can
t→∞
be solved and the solutions are expressed as follows:
1
ìï =
P0
N
ï
λ(0) λ(1) …λ( n - 1)
1 +∑
ï
μ(1) μ(2) …μ( n)
n=1
ï
λ(0) λ(1) …λ( n - 1)
í
ï
μ(1) μ(2) …μ( n)
ïP n =
, for all n > 0
N
λ(0) λ(1) …λ( n - 1)
ï
1 +∑
ï
μ(1) μ(2) …μ( n)
î
n=1
(2)
From the point of SU, we define the real state
transition probability as a S i→S j , which denotes the
probability that high⁃priority queue transfers from state
S i to S j between two time slots. For ∀i,j ∈ {0,1} ,
∑ a Si→Sj = 1 and ∑ a Si→Sj = 1. Thus, the real state
i
j
transition probability matrix can be expressed as,
é a S0→S0 a S0→S1 ùú
A= ê
(3)
ê a S1→S0 a S1→S1 ú
ë
û
Let Pnn′(t;Δt) denote the probability that given that
the high⁃priority queue stays in staten at time t, and then
a time slot Δt later, it will stay state n′. That is,
P nn′( t;Δt) = P[ X( t + Δt) = n′,X( t) = n] (4)
Therefore, the state transition probability a S0→S0
can be expressed as follows,
P 00( t;Δt)
= e -λ(0) Δt
a S0→S0 = lim
(5)
t→∞
P 0( t)
And a S0→S1 = 1 - a S0→S0 . In addition,
N
a S t➝S0 = lim
t→∞
P n0( t;Δt)
∑
n=1
N
P n( t)
∑
n=1
=
P 1 μ(1) Δt
N
Pn
∑
n=1
+ o( Δt) =
λ(0)
·μ(1) Δt
μ(1)
+ o( Δt)
N
æ λ(0) λ(1) …λ( n - 1) ö
ç
÷
∑
μ(1) μ(2) …μ( n) ø
n=1 è
(6)
And a S1→S1 = 1 - a S1→S0 . In this paper, Δt is viewed
as the length of a time slot, which is defined to be
10 ms in the IEEE802 22 standard. Therefore,
o( Δt) ≈ 0. In a steady queue model, for simplicity,
we assume that λ(0) = λ(1) = … = λ( N - 1) = λ,
μ(1) = μ(2) = … = μ( N) = μ, so that the real state
transition probability matrix between two time slots can
be expressed as follows,
-λΔt
-λΔt
éê e 1 - e
ùú
-
-
1
ρ
1
ρ
=
A ê
(7)
ú
-
êë 1 - ρ N μΔt 1 1 - ρ N μΔt úû
where ρ = λ / μ is defined as the busy⁃probability,
which is a positive lesser than 1 in a steady queue
model. Obviously, the higher the busy probability is,
the busier the channel is. Especially, when the busy
probability exceeds 1, SUs will have no chance to
access the licensed channel. However, this is rarely
seen
in
traditional
networks
because
the
underutilization of spectrum resource is up to 70%
based on the report of Berkeley Wireless Research
Center. After deriving the real state transition
probability matrix, we will further study A as an
important element of HMM in section 4.
3 Spectrum Detection Methods
3 1 Single User Energy Detection
In the existing research on cognitive radio, energy
detection is primarily applied to detect whether if the
licensed channel is busy. But, in this paper, we will
further study the energy detection to adequately utilize
the relationship between the detection result and the
real state of licensed channel. Let D 0 ,D 1 present the
detection consequences which are respectively non⁃
presence and presence of PUs, and let B = { b S j→D k } ,j,
k ∈ {0,1} describe the relationship between the
detection result and the real state of licensed channel.
Specifically, b S j→D k donates the probability that the
channel stays state S j with the detection result D k . For
∀i,j ∈ {0,1} , ∑ b S j→D k = 1 and ∑ b S j→D k = 1.
j
k
According to the definitions of the detection probability
P d and the false alarm probability P f ,
P d = P( D 1 | S 1 ) = b S1→D1
(8)
P f = P( D 1 | S 0 ) = b S0→D1
{
So, the matrix B can be written as,
é 1 - P f P f ùú
B = êê
ú
ë1 - P d Pd û
(9)
In the following, the concrete expression will be
derived. Firstly, we suppose that the licensed channels
are additive white Gaussian noise ( AWGN) channels.
During the energy detection period, Let r′( t) present
the received RF signal, which is first pre⁃filtered by an
· 63·
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
ideal filter. Then, r( t) , the output of the filter is
sampled, squared, and summed over a sensing time
interval to finally produce a measure of the energy of
the received waveform. That is,
Ns
|
∑
k=1
Y=
(10)
r( k) | 2
where finally has Ns terms to sum over, and Y acts as a
test statistic to test the two hypotheses S1 and S0 . It then
follows that under S1 , Y has a non⁃central chi⁃square
distribution. Likewise, given S0 , will be central chi⁃
square distribution. With variance σ 2 , non⁃centrality
parameter 2γ, and Ns degrees of freedom, the probability
density function of Y can be expressed as follows[15] ,
Ns - 2
ìï 1 æ y ö 4 - 2γ +y N æ 2γy ö
ç
÷
e 2σ 2 I 2s - 1 ç 2 ÷ , S1
ïσ 2 è 2γ ø
è σ ø
f Y(y) = í
(11)
y
1
ï
Ns / 2 - 2
ïσ Ns 2Ns / 2 Γ(N / 2) y e 2σ , S0
î
s
where Γ(·) is the gamma function, and I ν (·) is the
νth order modified Bessel Function of the first kind.
Therefore, the detection probability and the alarm
false probability of the mth SU can be written as Pmd =
Pm(D1 | S1 ) and Pmf = Pm(D1 | S0 ), respectively. Set the
parameter η m as the energy detection threshold of the mth
SU. So, Pm(D1 | S1 ) can be expressed as follows,
P m(D1 | S1 ) = P(Ym > η m | S1 ) =
∫
1 æ ym ö
ç
÷
2
η m 2σ è 2γ m ø
m
∞
Nm - 2
4
e-
2γ m +y m
2σ 2m
æ 2γ m y ö
I N2m - 1 ç
÷ dy m =
2
è σm ø
æ 2γ m
ηm ö
Q N2m ç
(12)
÷
,
2
σ 2m ø
è σm
where σ 2m and γ m donate the noise power and signal
power of receiver, respectively; N m represents sampling
number; Q N2m(·, ·) is the generalized Marcum
the mth SU, respectively. For single user spectrum
detection, the matrix B can be expressed as,
éê
æ Nm η m ö
æ Nm η m ö
ùú
Γç , 2 ÷
Γç , 2 ÷
ê
ú
è 2 2σ m ø
è 2 2σ m ø
ê 1 -
ú
æ Nm ö
æ Nm ö
ê
ú
Γç ÷
Γç ÷
B=
ê
ú
è2 ø
è2 ø
ê
ú
æ
ö
æ
ö
ê 1 - Q N m ç 2γ m , η m ÷ Q N m ç 2γ m , η m ÷ ú
2
2
2
2
ê
σ 2m ø
σ 2m ø úû
è σm
è σm
ë
(14)
3 2 Cooperative Energy Detection
Recalling aforementioned assumptions about
system model, for each sensing slot, multiple SUs, at
the same time, perform distributed cooperation
detection, in which each SU adopts energy detection,
and then delivers the detection consequence to fusion
center. Afterward, the fusion center judges whether the
primary channel is busy or not based on a certain
fusion criterion. Assuming that report channels are
perfect, and OR criterion is employed, for improving
the detection performance of cognitive radio networks,
as well as reducing the detection accuracy requirement
of single SU. Based on the OR criterion, the final
detection probability and the false alarm probability can
be expressed as follows,
M SU
ìï
P d = 1 - ∏ (1 - P md )
ï
m=1
í
M SU
ï
=
-
P
1
(1 - P mf )
∏
ï f
=
m 1
î
(15)
Obviously, the cooperation detection probability is
higher than single SU energy detection with the same
channel situation. However, the false alarm probability
is becoming higher, too. Although this will increase the
possibility that SU has to perform handoff, it is
absolutely necessary to guarantee the quality of service
of PU. Based on the aforementioned analysis, the
expression of B can be got as follows,
Q⁃function. Similarly, P m( D 1 | S 0 ) can be written as,
P m( D 1 | S 0 ) = 1 - P( Y m < η m | S 0 ) =
Nm
ηm
y
1
-
2 e 2σ 2 dy
=
1 -
y
m
m
m
M SU
Nm
-∞
Nm ö
éê M SU
ù
æ
Nm
m
σm 2 2 Γç ÷
(1 - P f ) 1 - ∏ (1 - P mf ) ú
∏
2
êm=1
ú
è ø
m=1
B = ê M SU
M SU
ú (16)
æ Nm η m ö
æ Nm ö
m
m
Γç , 2 ÷ Γç ÷
(13)
ê ∏ (1 - P d ) 1 - ∏ (1 - P d ) ú
êm=1
ú
è2 ø
è 2 2σ m ø
m=1
ë
û
where Γ(·, ·) is the incomplete gamma function.
Then,
substituting
Eqs.
(
12)
and
(
13)
into
Eq.
Without loss of generality, let P md and P mf present the
(16) , we can get the final expression of B. That is,
detection probability and the false alarm probability of
ηm ö
ηm ö
é M SU æ
ù
M SU æ
ê
ú
Γ( N m / 2,
)÷
Γ( N m / 2,
)÷
ç
ç
2
2
ê∏ ç
ú
2σ m ÷ 1 - ∏ ç
2σ m ÷
m = 1 ç1 -
÷
÷
ê m = 1 ç1 -
ú
Γ( N m / 2) ø
Γ( N m / 2) ø
B= ê
(17)
è
è
ú
M SU
M SU
ê
æ
æ 2γ m
æ
æ 2γ m
ηm ö ö
ηm ö ö ú
÷ ÷ 1 - ∏ ç1 - Q Nm / 2 ç
÷÷ú
ê ∏ ç1 - Q Nm / 2 ç
,
,
2
2
2
êë m = 1 è
m=1 è
σ øø
σ 2 ø ø úû
è σ
è σ
∫
m
· 64·
m
m
m
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
This is a very useful conclusion. The matrix B will
be viewed as an important element of HMM in the
following discussions.
4 HMM⁃EA Algorithm
It is well⁃known that the hidden Markov process is
a doubly embedded stochastic process with an
underlying stochastic process. This process is not
observable, but can only be observed through another
stochastic process which can produce a sequence of
observations. In this paper, from the perspective of
SU, the variation process of the real state ( S 1 or S 0 ) of
channel is viewed as an underlying stochastic process
which cannot be obtained directly. But, the channel
detection will produce a sequence of observations ( D 1
or D 0 ) so as to judge the real states. Consequently, it
has definite physical significance to describe the
aforementioned cognitive radio system as a HMM.
Fig.2 shows the schematic diagram of the HMM,
which perfectly describes all possible variation. From
the schematic diagram, we can see three basic
elements in the HMM. As one of the basic elements,
the real state transition probability matrix A depicts the
probability that the licensed channel transfers from one
state to another state between two time slots and it has
been obtained in section 2. The matrix B, another basic
element, presents the relationship between hidden
states ( i. e. real states ) and observations ( i. e.
detection results ) and it is called emission matrix,
derived in detail in section 3. Thus, the HMM can be
formulated as ψ = ( A,B,π) . Among the compact
notation, π = { π i } denotes the initial real state
distribution probability.
aS
0
aS
aS
S
0
S0
b
S
0
b
S
0
D
1
0
S
1
1
S
0
S1
b
S
0
aS
1
S
1
Hi
dde
ns
t
a
t
e
s
S
1
Obs
e
r
v
a
bl
es
e
que
nc
e
s
D0
D0
D1
Fig. 2 The relationship between the basic elements in the
HMM
Generally speaking, there are three basic
problems which need to be solved in practical
applications. The first one, evaluation problem, is to
compute the probability of the observed sequence,
given a model and a sequence of observations. The
second one, decoding problem, is to find out the
unobservable part of the HMM to discover the real state
sequence. The third one, learning problem, is to
optimize the model parameters so that it can accurately
describe how a given observation comes out. In this
paper, the last problem will be solved to optimally
adapt model parameters to observed sequence, namely
to estimate best model ψ for a detection sequence of
SU. Therefore, the arrival rate and the service rate of
PU can be obtained by estimating the state transition
matrix A and the emission matrix B. In order to solve
this problem, firstly considering such a known
observed sequence obtained by spectrum detection,
D = d 1 d 2 …d T
(18)
where T denotes the number of time slots, and d t( d t ∈
{ D 0 ,D 1 } ) is the detection consequence of SU in the
tth time slot. Let s t( s t ∈ { S 0 ,S 1 } ) present the real
state of PU in the tth time slot. As shown in Fig.3, the
arrows indicate the direction from one state to another
state between two time slots.
t
=0
S0(
π0)
Hi
dde
n
π1)
s
t
a
t
e
s S1(
Obs
e
r
v
a
bl
e
Se
que
nc
e
s
t
=1
S0
t
=2
S0
…
…
t
=T
S0
S1
S1
…
S1
1
2
d
d
dT
Fig.3 Diagram of the HMM along time slots
In order to solve learning problem, the forward⁃
backward procedure will be firstly introduced.
Considering the forward valuable ϑ t( i) defined as the
probability of the partial observation sequence,
d 1 d 2 …d t , ( until time t) and state s t at time t, given the
model ψ, namely,
ϑ t( i) = P( d 1 d 2 …d t ,s t = S i | ψ)
(19)
The forward valuable can be inductively solved,
as follows
ìï ϑ1(i) = π i bi(d1 ), 0 ≤ i ≤ 1 and t = 1
ï
1
í
t +1
ïϑt +1(i) = [ ∑ ϑt(i)aij ] bj(d ),0 ≤ j ≤ 1,2 ≤ t ≤ T
i =0
î
(20)
where b i( d t ) = b S i→d t ,a ij = a S i→S j . Then, in the similar
manner, a backward variable εt( i) can be defined as
the probability of the partial observation sequence from
t + 1 to the end, given state s t at time t and the mode
ψ, namely,
t+1 t+2
T
t
ω
t( i) = P( d d …d | s = S i ,ψ) (21)
Therefore, the backward valuable can be
inductively solved as follows,
T( i) = 1, 0 ≤ i ≤ 1
ìïω
1
ï
=
ω
(
i)
a ij b j( d t + 1 ) ω
ít
t + 1( j) ,
∑
j=0
ï
ï t = T - 1,T - 2,…,1 and 0 ≤ j ≤ 1
î
(22)
Based on the definitions above, ξ t( i,j) can be
defined as the probability of being in state S i at time t,
and state S j at time t + 1, given the model and
observation sequence, expressed as,
· 65·
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
ξ( i,j) = P( d t = S i ,d t + 1 = S j | D,ψ) =
P( d t = S i ,d t + 1 = S j ,D | ψ)
=
P( D | ψ)
ϑ t( i) a ij b j( d t + 1 ) ω
t + 1( j)
1
1
∑
∑ ϑt( i) a ij b j( d
i=0 j=0
t+1
)ω
t + 1( j)
(23)
So, given the observation sequence and the
model, the probability of being in state S i at time t can
1
T-1
1
be written as ∑ ξ t( i,j) . Thus, ∑ ∑ ξ t( i,j) can be
j=0
t=1 j=0
obviously interpreted as the expected number of
transitions made from S i while
T-1
ξ t( i,j)
∑
t=0
can be
viewed as the expected number of transitions from S i to
S j . By maximizing the Baum’ s auxiliary function,
Q(ψ,
ψ) = ∑ P(S | D,ψ)log[P(D,S |
ψ)] (24)
S
the re⁃estimation formulas can be derived
over ψ,
directly, as follows,
πi =
aij =
1
ξ 1( i,j)
∑
j=0
T-1
ξ t( i,j)
∑
t=0
T-1
1
∑ ∑ ξt( i,j)
(25)
(26)
t=1 j=0
T
1
=
b
j( k)
∑
∑ ξt( i,j)
t=1 j=0
s.t. d t = D k T
1
∑
∑ ξt( i,j)
t=1 j=0
(27)
It has been proved in Refs. [ 16 - 17] that the
maximization of the Baum ' s auxiliary function leads to
B,
increased likelihood, namely model
ψ = ( A,
π) is
=
more likely than model ψ
(A,B,π). Through
iteratively using
ψ in place of ψ and repeating the re⁃
estimation calculation, the most likely model will be
found and the parameters of the HMM will be estimated,
i.e. the traffic arrival rate and service rate of PU will be
obtained by solving the equation sets as follows,
Δt
-λ
= aS0→S0
ìïe
ï 1 - ρ
μΔt = aS1→S0
ï
N
(28)
í1 - ρ
ï
ïρ= λ
ï
μ
î
The proposed HMM⁃EA algorithm’ s complexity
consists of two parts. One is the complexity of HMM
model estimation and can be written as (6H 1 Z 1 Z 2 Z 3 +
2H 2 Z 1 Z 2 Z 3 ) [8] , where Z 1 means the number of states;
Z 2 is the number of observation; Z 3 is the total number
of transitions divided by the total number of states; H 1
means dimension of the observation vector and H 2 is
· 66·
maximum number of observations for a single transition.
The other one is the complexity of cooperative detection
and can be written as mO( Z) , where O( Z) means the
complexity of every SU’ s energy detection, and Z is
the number of sample andmis the number of cooperative
SUs. According to Ref.[11] , the ED⁃EA algorithm’ s
complexity can be written as (14Z 1 Z 2 Z 3 + O( Z) ) . In
the following section, we will compare algorithm
complexity between the HMM⁃EA algorithm and ED⁃
EA algorithm by simulations.
5 Simulation and Result Analysis
In section 4, we do the analysis about the HMM⁃
EA algorithm. In this part, the convergence
performance will be discussed by simulations. We
suppose that the licensed channels are additive white
Gaussian noise channels and set degrees of freedom of
chi⁃square distribution to be 5, signal to noise ratio of
the detected channel SNR = γ / σ 2 = 8 dB, detection
threshold TH = η / σ 2 = 10 dB, convergence tolerance of
likelihood estimation Δδ = 1e - 6, and the maximum
number of iterations NUM = 500. In addition, only one
licensed channel is detected at a time and the real
traffic arrival rate and the traffic service rate of the PU
queue are set to be λ = 0 3 and μ = 0 5, respectively.
Simulation results and analysis are as follows.
Fig.4 shows that, for constant channel conditions,
the estimation values of real state transition probability
matrix A gradually converge to the real values with the
increase of number of detection time slots. Because
more detection results can more accurately reflect the
transformation law of licensed channel state. Let
A( i,
j) present the estimation value of A( i,j) . Specifically,
when Tadds up to 2 000, the estimation error
|
A(1,1) - A(1,1) | is only 0 032 and |
A(2,1) -
A(2,1) | is 0 017. Additionally, it is found that the
curve of
A(1,1) and the curve of
A(1,2) are
symmetrical because of ∑ a S i→S j = 1, as well as the
j
curve of
A(2,1) and the curve of
A(2,2) .
Fig.5 shows that, for constant channel conditions,
the estimation values of traffic arrival rate and service
rate gradually converge to the real values with the
increase of number of detection time slots. Because
these two parameters are obtained by solving Eq.(28) .
Especially, when T adds up to 2 000, the estimation
- λ | is 0 007.
error | μ- μ | is only 0 018 and | λ
Hence, the traffic parameters can be estimated
accurately by HMM⁃EA algorithm.
In order to describe the convergence performance of
traffic parameters, let |
A(i,j) - A(i,j) | denote
estimation error of A(i,j). Fig. 6 shows that the
estimation error of A(1,1) converge to zero with the
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
0.
8
0.
6
0.
4
100
0.
2
0
Fig. 4 A(
1
,
1
)
(
Tr
u
e
)
A(
1
,
1
)
(
Es
t
i
ma
t
e
d
)
A(
1
,
2
)
(
Tr
u
e
)
A(
1
,
2
)
(
Es
t
i
ma
t
e
d
)
Es
t
i
ma
t
i
o
ne
r
r
o
r
1.
0
A(
2
,
1
)
(
Tr
u
e
)
A(
2
,
1
)
(
Es
t
i
ma
t
e
d
)
A(
2
,
2
)
(
Tr
u
e
)
A(
2
,
2
)
(
Es
t
i
ma
t
e
d
)
denote the estimation error of traffic parameter λ and μ,
respectively. For convenience, the estimation error
curves are also fitted by using exponential function.
The simulation results are shown in Figs.8 and 9. From
these two figures, it can be seen that the convergence
performance of HMM⁃EA algorithm is better than that
of ED⁃EA algorithm. Furthermore, the convergence
rate of HMM⁃EA gradually accelerates with the
increase of the number of cooperative users. However,
with the increase of the number of cooperative users,
signaling overhead and complexity also increase. It is
worth mentioning the optimal number of cooperative
users is studied in Ref.[18] .
1
05
01
0
0 20
0
0 40
0
0 60
0
0 80
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
Convergence of
A according to the number of
detection time slots by HMM⁃EA ( m = 2)
Pa
r
a
me
t
e
rv
a
l
ue
s
1.
0
0.
8
λ(
Tr
ue
)
λ(
Es
t
i
ma
t
e
d)
μ(
Tr
ue
)
μ(
Es
t
i
ma
t
e
d)
0.
6
0.
4
0.
2
10-2
A(
1
,
1
)
e
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
S
NR=8d
B
A(
1
,
1
)
e
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
S
NR=4d
B
A(
1
,
1
)
e
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
S
NR=0d
B
10-4
1
05
01
0
0 20
0
040
0
0 60
0
080
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
10-3
100
0
1
05
01
0
0 20
0
040
0
0 60
0
080
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
, μ
Fig.5 Convergence of λ
according to the number of
detection time slots by HMM-EA ( m = 2)
10-1
Fig.7 Estimation performance of A( 2,2) compared with
different SNR, that is, SNR = 8 dB, SNR = 4 dB and
SNR = 0 dB
Es
t
i
ma
t
i
o
ne
r
r
o
r
Va
l
ue
so
fe
l
e
me
nt
si
nA
increase of number of detection time slots. Furthermore,
the convergence rate gradually accelerates when the
channel condition becomes better. Obviously, when
SNR = 8 dB, the estimation error converges to zero
faster than that of SNR = 4dB and SNR = 0 dB. Based
on the above theoretic analysis, as the channel
condition is better, the detection result will be more
accurate. Consequently, the HMM model will be
estimated more accurately by solving learning problem.
In addition, for convenience, the estimation error curve
is fitted by exponential function. Similarly, Fig.7 shows
the same law.
10-1
10-2
λe
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d)
,
HMM EAm=4
λe
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d)
,
HMM EAm=2
λe
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d)
,
HMM EA
10-4
1
05
01
0
0 20
0
040
0
0 60
0
080
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
10-3
Fig. 8 Estimation performance of λ, compared between
HMM⁃EA and ED⁃EA algorithm
10-2
10-3
A(
1
,
1
)
e
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
S
NR=8d
B
A(
1
,
1
)
e
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
S
NR=4d
B
A(
1
,
1
)
e
s
t
i
ma
t
i
o
ne
r
r
o
r
-5
10
(
f
i
t
t
e
d
)
,
S
NR=0d
B
1
05
01
0
0 20
0
040
0
0 60
0
080
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
10-4
Fig.6 Estimation performance of A( 1,1) compared with
different SNR, that is, SNR = 8 dB, SNR = 4 dB and
SNR = 0 dB
In order to compare the convergence performance of
HMM⁃EA and ED⁃EA algorithm, we set channel
condition SNR = 4 dB. Then, let |
λ - λ | and | μ
- μ |
100
Es
t
i
ma
t
i
o
ne
r
r
o
r
Es
t
i
ma
t
i
o
ne
r
r
o
r
10-1
10-1
10-2
μe
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
HMM EAm=4
μe
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
HMM EAm=2
μe
s
t
i
ma
t
i
o
ne
r
r
o
r
(
f
i
t
t
e
d
)
,
HMM EA
10-4
1
05
01
0
0 20
0
040
0
0 60
0
080
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
10-3
Fig. 9 Estimation performance of μ, compared between
HMM⁃EA and ED⁃EA algorithm
Based on the previous analysis about HMM, Z 1 =
H 1 = 2, Z 2 = T and H 2 = 1. In addition, Z is set to be
· 67·
Journal of Harbin Institute of Technology ( New Series) , Vol.22, No.5, 2015
Co
mpl
e
x
i
t
yo
fa
l
g
o
r
i
t
hms
(
104)
5 000 and Z 3 can be obtained from the statistical results
within evaluating HMM process. Therefore, the
simulation results are shown as Fig. 10. As is seen,
with the increase of number of detection time slots, the
complexity of the HMM⁃EA and ED⁃EA algorithm are
both becoming more and more complex. Furthermore,
the complexity of the HMM⁃EA algorithm is more
complex than that of the ED⁃EA algorithm. From
previous simulation pictures, we know that more
detection time slots lead to better convergence
performance. That means the HMM⁃EA algorithm
sacrifices algorithm complexity performance to improve
the convergence performance.
16
14
EDEA
HMMEEAm=2
12
HMMEEAm=4
10
8
6
4
2
0
1
05
01
0
0 20
0
040
0
0 60
0
080
0
01
00
0
0
Numbe
ro
fde
t
e
c
t
i
o
nt
i
mes
l
o
t
sT
Fig.10 Complexity comparison between the HMM⁃EA and
ED⁃EA algorithm
6 Conclusions
In this paper, we build the cognitive radio system
as a preemptive priority queue, and then derive the
real state transition probability matrix. Subsequently,
the emission matrix is derived based on cooperative
energy detection. According to these two matrixes, we
build the HMM and thereby propose the HMM⁃EA
algorithm, which sufficiently utilizes the inner
collection among detection results to estimate traffic
parameters of licensed channels. Due to the fact that
cooperative energy detection can overcome the low
signal to noise ratio or hidden terminal problem, the
proposed HMM⁃EA algorithm has better convergence
performance than ED⁃EA algorithm. Therefore, SUs
can estimate busy probability of licensed channels, and
access the optimal target channel. In this paper, the
cooperative energy detection is discussed based on
perfect report channel. So, in our future work, we will
focus on imperfect report channel, which seriously
impacts detection results.
References
[1] Haykin S. Cognitive radio: brain⁃empowered wireless com⁃
munications. IEEE J. Selected Areas in Comm., 2005, 23
(2) : 201-220.
· 68·
[2] Akyildiz I F, Lee W Y, Vuran M C, et al. A survey on
spectrum management in cognitive radio networks. IEEE
Comm. Magazine, 2008, 46(2) : 40-48.
[3] Christian I, Moh S, Ilyong C, et al. Spectrum mobility in
cognitive radio networks. IEEE Comm. Magazine, 2012, 50
(6) : 114-121.
[4] Lee W Y, Akyildiz I F. Spectrum⁃aware mobility management
in cognitive radio cellular networks. IEEE Transactions on
Mobile Computing, 2012, 11(4): 529-542.
[5] Wang L C, Wang C W. Spectrum handoff for cognitive
radio networks: reactive⁃sensing or proactive⁃sensing.
Proceedings of IEEE International Conference on Computing
and Communication Performance. Austin. 2008. 343-348.
[6] Park C, Kim S, Lim S, et al. HMM based channel status
predictor for cognitive radio. Proceedings of IEEE Asia⁃
Pacific Microwave Conference. Bangkok. 2007. 1-4.
[7] Chen Z, Guo N, Hu Z, et al. Experimental validation of
channel state prediction considering delays in practical
cognitive radio. IEEE Trans. Veh. Technol., 2011, 60
(4) : 1314-1325.
[8] Sung H S, Sung J J, Jae M K. HMM⁃based adaptive
frequency⁃hopping cognitive radio system to reduce
interference time and to improve throughput. KSII
Transactions on Internet and Information Systems, 2010, 4
(4) : 475-490.
[9] Wang L C, Wang C W, Chang C J. Modeling and analysis
for spectrum handoffs in cognitive radio networks. IEEE
Transactions on Mobile Computing, 2012, 11( 9) : 1499-
1513.
[10] Wang C W, Wang L C. Analysis of reactive spectrum
handoff in cognitive radio networks. IEEE Journal on
Selected Areas in Communications, 2012, 30 ( 10 ) :
2016-2028.
[11 ] Kae W C, Hossain E. Estimation of primary user
parameters in cognitive radio systems via hidden Markov
model. IEEE Transactions on Signal Processing, 2013, 61
(3) : 782-795.
[12] Shiang H P, Van D S M. Queuing⁃based dynamic channel
selection for heterogeneous multimedia applications over
cognitive radio networks. IEEE Trans. Multimedia, 2008,
10(5) : 896-909.
[13] Wang P, Xiao L, Zhou S, et al. Optimization of detection
time for channel efficiency in cognitive radio systems.
Proceedings of IEEE Wireless Communications and
Networking Conference. Kowloon. 2011. 111-115.
[ 14 ] Lee W Y, Akyildiz I F. Optimal spectrum sensing
framework for cognitive radio networks. IEEE Transactions
Wireless Communications, 2008, 7(10) : 3845-3857.
[15] Fadel F D, Mohamed⁃Slim A, Marvin K S. On the energy
detection of unknown signals over fading channels. IEEE
Transactions on Communications, 2007, 55(1) : 21-23.
[16] Baum L E, Sell G R. Growth functions for transformation
on manifolds. Pac. J. Math., 1968, 27(2) : 211-227.
[17] Baker J K. The dragon system⁃An overview. IEEE Trans.
Acoust. Speech Signal Processing, 1975, 23(1) : 24-29.
[18 ] Wei Z, Ranjan K M, Khaled B L. Optimization of
cooperative spectrum sensing with energy detection in
cognitive radio networks. IEEE Transactions on Wireless
Communications, 2009, 8(12) : 5761-5766.
© Copyright 2026 Paperzz