Professional Documents
Culture Documents
Optimal Resource Allocation For Cooperative Orthogonal Frequency Division Multiplexing-Based Cognitive Radio Networks With Imperfect Spectrum Sensing
Optimal Resource Allocation For Cooperative Orthogonal Frequency Division Multiplexing-Based Cognitive Radio Networks With Imperfect Spectrum Sensing
Research Article
ISSN 1751-8628
Optimal resource allocation for cooperative Received on 18th January 2014
Accepted on 23rd October 2014
orthogonal frequency division multiplexing- doi: 10.1049/iet-com.2014.0605
www.ietdl.org
based cognitive radio networks with
imperfect spectrum sensing
Wanming Hao 1 ✉, Showyi Yang 1, Bing Ning 1, Wanliang Hao 2
1
School of Information Engineering, Zhengzhou University, Zhengzhou, People’s Republic of China
2
College of Navigation and Aerospace Engineering, Information Engineering University, Zhengzhou, People’s Republic of China
✉ E-mail: wmhao@hotmail.com
Abstract: This study investigates sensing-based spectrum sharing access (SSSA) and sensing-based spectrum
opportunistic access (SSOA) schemes in cooperative orthogonal frequency division multiplexing (OFDM)-based
cognitive radio networks with imperfect spectrum sensing. The optimal resource allocation strategy, including sensing
time and transmit power, is designed to maximise the ergodic throughput of the secondary system. For a two-hop
cooperative communication in the secondary system, the authors adopt the amplify-and-forward relay protocol and
enable the source and relay to sense the state of the primary user jointly. To protect the PU effectively from harmful
interference, they consider the average interference power constraint in each hop. The total average transmit power
constraint of the source and relay is considered. Two simplified versions for the SSSA and SSOA schemes are
employed because of the complexity of the problem. They then propose two algorithms that acquire the optimal
sensing time and power allocation for both schemes. Finally, simulation results are presented to compare the
performance of the two schemes.
2 System model
1n |hnss,sd |2 |hnss,sr |2
Pdn (t, 1n ) = Q − g − 1 2f t/N (gn + 1)2 (4) an,3 = , an,1 = ,
s2 n
|hnpt,sd |2 Pp,n + d2 |hnpt,sr |2 Pp,n + d2
1 |hnsr,sd |2
Pfn (t, 1n ) = Q n
− 1 2f t/N (5) an,2 =
s 2 |hnpt,sd |2 Pp,n + d2
In the SSSA scheme, the source and relay adapt their transmit power RnSSSA = P(H1n )Pdn r11,n + P(H1n )(1 − Pdn )r10,n
on the basis of the detection outcome. If the nth sub-channel is
detected to be idle (H0n ), then the source and relay transmit with + P(H0n )Pfn r01,n + P(H0n )(1 − Pfn )r00,n (10)
(0) (0)
high power Ps,n and Pr,n , respectively. However, if the nth
sub-channel is detected to be active (H1n ), the source and relay where P(H1n ) and P(H0n ) denote the probability that the nth
(1) (1)
transmit with low power Ps,n and Pr,n , respectively, to ensure the sub-channel is active and idle, respectively.
QoS of the PU. To maintain keep the long-term budget of the secondary system,
In fact, the result of spectrum detection is not always correct we consider the total average transmit power constraint for the
because of the limitations of the spectrum sensing techniques and source and relay. This constraint can be found in several studies
the nature of the wireless communication. Moreover, a [8, 9] and can be written as
miss-detection or a false alarm can occur in the nth sub-channel.
Miss-detection occurs when the nth sub-channel is detected to be
T −t N
idle but is used by the PU, whereas false alarm indicates that the E P(H1n )Pdn (Ps,n
(1)
+ Pr,n
(1)
) + P(H1n )(1 − Pdn )(Ps,n
(0)
+ Pr,n
(0)
)
nth sub-channel is sensed to be busy but is actually idle. As a T n=1
(0)
result, four possible scenarios can be distinguished on the basis of
the sensing decision and the actual status of the PU on each + P(H0n )Pfn (Ps,n
(1)
+ Pr,n
(1)
) + P(H0n )(1 − Pfn )(Ps,n
(0)
+ Pr,n ) ≤ Pav
sub-channel. These scenarios are listed in Table 1. (11)
We rewrite the rate of the second system as
where Pav is the maximum average transmit power for the source and
(1) (1)
relay.
1 an,1 Ps,n an,2 Pr,n
r11,n = log2 1 + an,3 Ps,n
(1)
+ (1) (1)
(6) In a CR, the PU is prioritised; thus, its QoS should be ensured at
2 an,1 Ps,n + an,2 Pr,n all times. Methods that protect the PU from harmful interference
include the average interference power constraint [14, 27, 28] and
the peak interference power constraint [29, 30]. In fact, the
(0) (0)
1 an,1 Ps,n an,2 Pr,n advantage of the average interference power constraint, as shown
r10,n = log2 1 + an,3 Ps,n +
(0)
(0) (0)
(7)
2 an,1 Ps,n + an,2 Pr,n in [27], is that it provides not only higher ergodic throughput for
the secondary system but also better protection for the primary
(1) (1)
system. Thus, we apply an average interference power constraint in
1 bn,1 Ps,n bn,2 Pr,n this work.
r01,n = log2 1 + bn,3 Ps,n +
(1)
(1) (1)
(8) In sensing-based cooperative communication, the data
2 bn,1 Ps,n + bn,2 Pr,n
transmission phase includes two hops in a frame, and the PU can
suffer from harmful interference in each hop. Therefore the
1 b P(0) b P(0) average interference power constraint should be considered in each
r00,n = log2 1 + bn,3 Ps,n
(0)
+ n,1 (0)s,n n,2 r,n(0) (9) hop. Under the SSSA scheme, interference on the nth sub-channel
2 bn,1 Ps,n + bn,2 Pr,n affects the PU in two cases, missed detection and correct
detection. Consequently, the average interference power constraint
in each hop can be written as
Table 1 Four possible scenarios for SSSA
PU’s state Sensing results Related probability Power Rate
T −t
The first hop
T
Active (H1n ) H1n Pdn (1)
Ps,n (1)
Pr,n r11,n
× E |hss,pr | Ps,n P(H1n )(1 − Pdn ) + |hnss,pr |2 Ps,n
n 2 (0) (1)
P(H1n )Pdn ≤ Qnav
Active (H1n ) H0n 1 − Pdn (0)
Ps,n (0)
Pr,n r10,n
Idle (H0n ) H1n Pfn (1)
Ps,n (1)
Pr,n r01,n n = 1, . . . , N
Idle (H0n ) H0n 1 − Pfn (0)
Ps,n (0)
Pr,n r00,n
(12)
In fact, problem (14) is complicated to solve. We will thus simplify function of problem (16) can be written as
this problem according to a practical situation. In practice, the
detection probability Pdn is usually high. The value is set to be g(l, m, n) W max L(P, l, m, n) (22)
(1) (0) (1) (0)
larger than 0.9 in the IEEE 802.22 wireless regional area network. {P s,n ,P s,n ,P r,n ,P r,n }
Meanwhile, the false alarm probability Pfn is very low, that is,
usually <0.1. Thus, the item including 1 − Pdn and Pnf in (10) is where the Lagrangian is
relatively small. Moreover, the active probability of the PU is small
at <0.4. This assumption is reasonable because a recent report [1]
revealed that most of the licensed spectrum is underutilised. T − t N
Under the aforementioned assumptions, the optimal problem was L(P, l, m, n) = E (a r + a00,n r00,n )
T n=1 11,n 11,n
approximated by
T −t N
T −t N
+ l Pav − E (a01,n + a11,n )(Ps,n
(1)
max E (a11,n r11,n + a00,n r00,n ) (16) T n=1
(1) (0) (1) (0)
{t,P s,n ,P s,n ,P r,n ,P r,n } T n=1
T −t
s.t. + Pr,n ) + (a10,n + a00,n )(Ps,n + Pr,n )
(1) (0) (0)
T
N
×E (a01,n +a11,n )(Ps,n +Pr,n )+(a10,n +a00,n )(Ps,n +Pr,n ) ≤Pav
(1) (1) (0) (0)
N
T −t
+ mn Qnav − (1)
E s11,n Ps,n + s10,n Ps,n
(0)
n=1 n=1
T
(17)
N
T −t
T −t (0)
+ vn Qnav − (1)
E w11,n Pr,n + w10,n Pr,n
(0)
(1)
E s11,n Ps,n + s10,n Ps,n ≤ Qnav n = 1, . . . , N (18) n=1
T
T
(23)
T −t
(1)
E w11,n Pr,n + w10,n Pr,n
(0)
≤ Qnav n = 1, . . . , N (19)
T
The dual optimisation problem is given by
(0)
Ps,n ≥ 0, (1)
Ps,n ≥ 0, (0)
Pr,n ≥ 0, (1)
Pr,n ≥ 0, 0≤t≤T (20)
min g(l, m, n)
where the parameters a11,n, a10,n, a01,n, a00,n s11,n, s10,n, w11,n and l,m,n (24)
w10,n are given by s.t. l ≥ 0, m ≥ 0, n ≥ 0
⎧
⎪
⎪ a11,n = P(H1n )Pdn According to [31], a dual function is always convex, such that we
⎪
⎪ 10,n = P(H1 )(1 − Pd )
n n
⎪
⎪ a can use the gradient or sub-gradient-based method to minimise g
⎪
⎪
⎪
⎪ a01,n = P(H0 )Pf
n n
(l, μ, n). Dual variables l, μ and n are updated in parallel as
⎪
⎪
⎪ a00,n = P(H0n )(1 − Pfn )
⎨ follows (see (25))
s11,n = |hnss,pt |2 P(H1n )Pdn (21)
⎪
⎪ where ο(t), ζ(t) and j(t) are diminishing stepsizes, and t is the
⎪
⎪ s10,n = |hss,pt | P(H1 )(1 − Pd )
n 2 n n
⎪
⎪ iteration index. The stepsizes are chosen on the basis of the
⎪
⎪
⎪
⎪ 11,n = |hsr,pt | P(H1 )Pd
n 2 n n diminishing stepsize policy [31] to guarantee convergence.
⎪
⎪
w
⎪
⎩ However, computing for the dual function (25) involves
w10,n = |hnsr,pt |2 P(H1n )(1 − Pdn ) determining the optimal P at the given dual point l, μ and n. We
⎧
+
⎪
⎪ T −t N
⎪
⎪ l(m + 1) = l(m) + o(t) Pav − (a01,n + a11,n )(Ps,n + Pr,n ) + (a10,n + a00,n )(Ps,n + Pr,n )
(1) (1) (0) (0)
⎪
⎪ E
⎪
⎪ T
⎪
⎨
n=1
T −t + (25)
m
⎪ n (m + 1) = m (m) + 6 (t) Q n
− E s P (1)
+ s P (0)
⎪
⎪
n av
T 11,n s,n 10,n s,n
⎪
⎪
⎪
⎪ T −t +
⎪
⎪
⎩ nn (m + 1) = nn (m) + j(t) Qav − E w11,n Pr,n + w10,n Pr,n
n (1) (0)
T
T − t N
T −t N
g(l, m, n) W max E a r −l E (a00,n + a10,n )(Ps,n + Pr,n )
(0) (0)
(1) (0) (1) (0)
{P s,n ,P s,n ,P r,n ,P r,n } T n=1 00,n 00,n T n=1
T −t N (0) T − t N (0) T − t N
− mn E s10,n Ps,n − vn E w10,n Pr,n +E a11,n r11,n
T n=1 T n=1 T n=1
(26)
T −t N
T − t N
−l E (a01,n + a11,n )(Ps,n
(1)
+ Pr,n
(1)
) − (1)
mn E s11,n Ps,n
T n=1
T n=1
T −tN N N
− (1)
vn E w11,n Pr,n + lPav + mn Qnav + vn Qnav
T n=1 n=1 n=1
⎧
+
⎪
⎪ 1 a00,n a1,n a22,n + (a3,n − 2(l(a00,n + a10,n ) + mn s10,n ) ln 2/a00,n )(a1,n mn + a2,n )2
⎪
⎪
⎨ 2 ln 2 · m (l(a + a ) + m s )(a m + a )(a a m + a a + a a )
r,n =
p(0) n 00,n 10,n n 10,n 1,n n 2,n 3,n 1,n n 2,n 3,n 1,n 2,n
(30)
⎪
⎪ if (l(a00,n + a10,n ) + mn s10,n )a2,n . a3,n (l(a00,n + a10,n ) + nn w10,n )
⎪
⎪
⎩
0 if (l(a00,n + a10,n ) + mn s10,n )a2,n ≤ a3,n (l(a00,n + a10,n ) + nn w10,n )
a2,n ( ((l(a00,n + a10,n ) + mn s10,n )a1,n a2,n − (l(a00,n + a10,n ) + nn w10,n )a1,n a3,n + (l(a00,n + a10,n ) + mn s10,n )a2,n a3,n )
+a3,n (l(a00,n + a10,n ) + yn w10,n ))
mn = (31)
a1,n ((l(a00,n + a10,n ) + mn s10,n )a2,n − (l(a00,n + a10,n ) + yn w10,n )a3,n )
⎧
+
⎪
⎪ 1 a11,n a1,n a22,n + (a3,n − 2(l(a01,n + a11,n ) + mn s11,n ) ln 2/a11,n )(a1,n m∗n + a2,n )2
⎪
⎪
⎨ 2 ln 2 · m∗ (l(a + a ) + m s )(a m∗ + a )(a a m∗ + a a + a a )
r,n =
p(1) n 01,n 11,n n 11,n 1,n n 2,n 3,n 1,n n 2,n 3,n 1,n 2,n
(33)
⎪
⎪ if (l(a01,n + a11,n ) + mn s11,n )a2,n . a3,n (l(a01,n + a11,n ) + mn w11,n )
⎪
⎪
⎩
0 if (l(a01,n + a11,n ) + mn s11,n )a2,n ≤ a3,n (l(a01,n + a11,n ) + mn w11,n )
a2,n ( ((l(a01,n + a11,n ) + mn s11,n )a1,n a2,n − (l(a01,n + a11,n ) + nn w11,n )a1,n a3,n + (l(a01,n + a11,n ) + mn s11,n )a2,n a3,n )
+a3,n (l(a01,n + a11,n ) + nn w11,n ))
m∗n = (34)
a1,n ((l(a01,n + a11,n ) + mn s11,n )a2,n − (l(a01,n + a11,n ) + nn w11,n )a3,n )
T −t N
max E a00,n r00,n (41)
{t,P s,n ,P r,n } T
(0) (0)
n=1
T −t N
s.t. E (a00,n + a10,n )(Ps,n + Pr,n ) ≤ Pav
(0) (0)
(42)
T n=1
T −t (0)
E s10,n Ps,n ≤ Qnav n = 1, . . . , N (43)
T
Fig. 3 Optimal sensing time and power allocation algorithm for the SSSA
scheme
T −t
(0)
E w10,n Pr,n ≤ Qnav n = 1, . . . , N (44)
T
4 Sensing-based spectrum opportunistic access
nature of the wireless communication make missed detection However, for sensing time t, we can obtain the optimal value by
inevitable. Thus, the secondary system accesses a sub-channel in using a 1D exhaustive search. Therefore we focus on finding the
two cases. One is a correct detection (the sub-channel is idle), and optimal power allocation that maximises the ergodic throughput
the other is missed detection (the sub-channel was active). The for a given sensing time under the SSOA scheme.
average throughput of the nth sub-channel for the SSOA scheme is The Lagrange dual function of problem (41) can be written as
given by
g(l, m, n) W max L(P (0)
s,n , P r,n , l, m, n)
(0)
(46)
(0) (0)
RnSSOA = P(H1n )(1 − Pdn )r10,n + P(H0n )(1 − Pfn )r00,n (35) {P s,n ,P r,n }
Similar to the case in Section 3, we can derive the average transmit where the Lagrangian is
power constraint and average interference power constraint as
follows (see (36))
T − t
N
s,n , P r,n , l, m, n) = E
L(P (0) (0)
(a00,n r00,n )
T − t n 2 (0) T n=1
The first hop E |hss,pr | Ps,n P(H1n )(1 − Pdn ) ≤ Qnav
T T −t N
n = 1, . . . , N (37) + l Pav − E (a10,n + a00,n )(Ps,n + Pr,n )
(0) (0)
T n=1
N
T −t (0)
T − t n 2 (0) + mn Qav −
n
E s10,n Ps,n
The second hop E |hsr,pr | Pr,n P(H1n )(1 − Pdn ) ≤ Qnav n=1
T
T
n = 1, . . . , N
N
T −t
+ vn Qnav − (0)
E w10,n Pr,n (47)
(38) n=1
T
Finally, we formulate the optimisation problem that maximises the The dual optimisation problem is given by
ergodic throughput under the SSOA scheme as
min g(l, m, n)
l,m,n
T −t N (48)
max E n
RSSOA (39) s.t. l ≥ 0, m ≥ 0, n≥0
{t,P s,n ,P r,n } T
(0) (0)
n=1
To find the optimal values l, μ and n that minimise the dual function
g(l, μ, n), the gradient or subgradient-based method can be used
(36) − −(38), 0 ≤ t ≤ T [30]. Dual variables l, μ and n are updated in parallel as (see (49)
s.t. (40)
(0)
Ps,n ≥ 0, Pr,n
(0)
≥0 at the bottom of the next page)
where ο(t), ζ(t) and j(t) are diminishing stepsizes and t is the
Problem (39) is actually difficult to solve. The above analysis shows iteration index. The stepsizes are chosen following the diminishing
that the item 1–Pdn is very small in practice for sensing-based CR step sizepolicy [30] to guarantee convergence.
T −t N
E P(H1n )(1 − Pdn )(Ps,n
(0)
+ Pr,n
(0)
) + P(H0n )(1 − Pfn )(Ps,n
(0)
+ Pr,n
(0)
) ≤ Pav . (36)
T n=1
⎧
⎪
⎪ m p(0) , r,n . 0
if p(0) 6 Simulation results
⎨
n r,n +
s,n =
p(0) a00,n 1
⎪
⎩ 2(l(a + a ) + m s ) ln 2 − a
⎪ r,n = 0
if p(0) In this section, we present the simulation results for the two access
00,n 10,n n 10,n 3,n schemes: SSSA and SSOA. We consider quasi-static
(50) frequency-selective Rayleigh fading channels with a 6-tap
equal-gain equal-spaced delay profile. The delay interval between
adjacent taps is equal to the inverse of the OFDM system
(see (51 and 52))
bandwidth. For simplicity, we consider N = 4 subcarriers. The
channel variances of PT → SR, PT → SD, SS → PR, SS → SD and
The derivation is shown in the Appendix.
SR → PR links are all set to be −10 dB, whereas the channel
By updating the dual variables shown in (49), the optimisation
variances of PT → PR, SS → SR and SR → SD links are all set to
problem in (41) can finally be solved. Based on the above
be 0 dB. The target detection probability Pdn for every sub-channel
discussion, a detailed algorithm is given in Fig. 4.
is set to be 0.9, with γn = −15 dB. The transmit power of the PU
in every sub-channel is assumed to be 10 dB.
5 Complexity analysis
6.1 Imperfect sensing scenario
The computational complexity of the proposed two schemes is
determined by the complexity of solving the dual problem. The Fig. 5 shows the ergodic throughput against the sensing time under
complexity of the subgradient method is polynomial in the number the SSSA and SSOA schemes. We assume that the P(H0n ) is the same
of dual variables. Given the total power constraint for the source for every sub-channel and is replaced with P(H0). The total transmit
and relay, 2N + 1 dual variables are derived, and the complexity is power of the source and relay is set to be 15 dB, whereas the
O(|2N + 1|2). However, we need to find the optimal value between interference power constraint is set to be −10 dB for every
(0 T ) by using 1D exhaustive search. Assuming the search interval sub-channel. From Fig. 5, we find an optimal sensing time for the
is t, we need (T/t + 1) times search, with the complexity of each SSSA and SSOA schemes to achieve the maximum ergodic
search being O(|2N + 1|2). throughput of the secondary system. In addition, the ergodic
⎧
+
⎪
⎪ T −t N
⎪
⎪ l(m + 1) = l(m) + o(t) Pav − (a10,n + a00,n )(Ps,n + Pr,n )
(0) (0)
⎪
⎪ E
⎪
⎪ T
⎪
⎨
n=1
T −t + (49)
⎪ mn (m + 1) = mn (m) + 6(t) Qnav − (0)
E s10,n Ps,n
⎪
⎪ T
⎪
⎪
⎪
⎪ T −t +
⎪
⎪
⎩ nn (m + 1) = nn (m) + j(t) Qav −
n (0)
E w10,n Pr,n
T
⎧
+
⎪
⎪ 1 a00,n [a1,n a22,n + (a3,n − 2(l(a00,n + a10,n ) + mn s10,n ) ln 2/a00,n )(a1,n mn + a2,n )2 ]
⎪
⎨ ·
2 ln 2 mn (l(a00,n + a10,n ) + mn s10,n )(a1,n mn + a2,n )(a3,n a1,n mn + a2,n a3,n + a1,n a2,n )
p(0) = (51)
r,n
⎪
⎪ if (l(a00,n + a10,n ) + mn s10,n )a2,n . a3,n (l(a00,n + a10,n ) + mn w10,n )
⎪
⎩
0 if (l(a00,n + a10,n ) + mn s10,n )a2,n ≤ a3,n (l(a00,n + a10,n ) + mn w10,n )
a2,n ( ((l(a00,n + a10,n ) + mn s10,n )a1,n a2,n − (l(a00,n + a10,n ) + nn w10,n )a1,n a3,n + (l(a00,n + a10,n ) + mn s10,n )a2,n a3,n )
+a3,n (l(a00,n + a10,n ) + nn w10,n ))
m∗n = (52)
a1,n ((l(a00,n + a10,n ) + mn s10,n )a2,n − (l(a00,n + a10,n ) + nn w10,n )a3,n )
9 References
1 Federal Communications Commission.: ‘Spectrum policy task force report, FCC 10 Appendix
02-155’, 2002.
2 Hossain, E., Niyato, D., Han, Z.: ‘Dynamic spectrum access and management in
cognitive radio networks’ (Cambridge University Press, 2009) The derivative of SP1 with respect variable P (0)
s,n and P r,n replaced by
(0)
3 Haykin, S.: ‘Cognitive radio: brain-empowered wireless communications’, IEEE P1 and P2 for simplicity is given by
J. Sel. Areas Commun., 2005, 23, (2), pp. 201–220
4 Zhao, Q., Swami, A.: ‘A decision-theoretic framework for opportunistic spectrum
access’, IEEE Trans. Wirel. Commun., 2007, 14, (4), pp. 14–20 ∂L a a (P a + P2 a2 )2 + P22 a1 a22 1
5 Ghasemi, A., Sousa, E.S.: ‘Fundamental limits of spectrum-sharing in fading = 00 · 3 1 1 · − l(a00 + a10 ) − ms10
environments’, IEEE Trans. Wirel. Commun., 2007, 6, (2), pp. 649–658 ∂P1 2 ln 2 P1 a1 + P2 a2 c
6 Bansal, G., Hossain, M.J., Bhargava, V.K., Tho, L.N.: ‘Subcarrier and power
allocation for OFDMA-based cognitive radio systems with joint overlay and
(53)
underlay spectrum access mechanism’, IEEE Trans. Veh. Technol., 2013, 62,
(3), pp. 1111–1122
7 Hsu, C.-N., Su, H.-J., Lin, P.-H.: ‘Joint subcarrier pairing and power allocation for ∂L a P12 a21 a2 1
OFDM transmission with decode-and-forward relaying’, IEEE Trans. Signal = 00 · · − l(a00 + a10 ) − nr10 (54)
Process., 2011, 59, (1), pp. 399–414 ∂P2 2 ln 2 P1 a1 + P2 a2 c
8 Dang, W.B., Tao, M.X., Mu, H., Huang, J.W.: ‘Subcarrier-pair based resource
allocation for cooperative multi-relay OFDM systems’, IEEE Trans. Wirel.
Commun., 2010, 9, (5), pp. 1640–1649 where c = P1α1 + P2α2 + P1(P1α1α3 + P2α2α3 + P2α1α2).
where +
a00 1
p1 = − (58)
2 ln 2bs a3
a ( br (a1 a2 bs − a1 a3 br + a2 a3 bs ) + a3 br )
m= 2
a1 (a2 bs − a3 br ) Thus, (29) and (30) are proven.