Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Performance Analysis of Combiners for Cooperative

Interweave-Spectrum Sharing Network with Imperfect


Channel Sensing

Journal: IEEE Signal Processing Letters

Manuscript ID Draft

Manuscript Type: Letter

Date Submitted by the


n/a
Author:

Complete List of Authors: Balaji, Rushabha; Electrical and electronics engineering


Pai, Vinay; Electrical and electronics engineering
Bitragunta, Sainath; Electrical and electronics engineering

Signal processing for communications, Emerging Signal Processing


Subject Category:
Applications

COM-ESTI Communication signal transmission, reception and detection


algorithms, channel estimation and modeling < COM Signal processing
for communications, COM-PERF Performance analysis of communication
systems < COM Signal processing for communications, COM-MACOG
EDICS:
Multiple Access, Scheduling, Cognitive Radios < COM Signal processing
for communications, SAS-STAT Detection, estimation and classification
theory and methods, statistical signal processing < MLSAS Machine
Learning and Statistical Signal processing
Page 1 of 5

1
2
3
Author’s Responses to Custom Submission Questions
4
5
6 Does your paper contain five (5) or more No
7 authors?
8 If yes, please provide a brief description of n/a
9 each author's contributions in the text field
10 below. If no, please enter "n/a" into the text
11 field below.
12 Is this manuscript related to any other papers, No
13 including but not limited to those of the
14 authors, that are either (a)
15 published/accepted, (b) currently under
16 review, or (c) being concurrently submitted
17 for review?
18 If yes, please list those manuscripts below. n/a
19 Also include these manuscripts as "Supporting
20 Documents" in the submission system. Please
21 note that this question covers papers in any
22 publication venues, not just in this journal;
23 you do not need to include papers already
24 included in response to prior questions. If no,
25 please enter "n/a" into the text field below.
26 Does your manuscript contain supplementary No
27 materials such as multimedia, extended
28 objects, or any other item intended for
29 publication but not included in the main body
30 of the paper? Supplementary materials must
31 go through review and may not be added after
32 acceptance.
33 Is this manuscript a re-submission or modified No
34 version of a previously rejected manuscript?
35 (Please note that this question refers to
36 previous submission to ANY journal, not just
37 to this journal or other IEEE journals.) If you
38 answer "Yes", specify the journal and
39 manuscript ID of the previous submission and
40 upload a response to the previous review.
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Page 2 of 5
1

1
2
3 Performance Analysis of Combiners for Cooperative
4
5
6
Interweave-Spectrum Sharing Network with
7
8
Imperfect Channel Sensing
9
Rushabha Balaji, Vinay U Pai and B. Sainath, Senior Member, IEEE
10
11
12
13 Abstract—A spectrum sharing or cognitive radio network scale fading and shadowing effects, the sensing result by a
14 enables spectrum sharing and increases spectral utility by single SU might not be sufficient to make a reliable decision
15 enabling secondary users to transmit whenever the primary about the presence of a PU accessing the channel. A distributed
16 user is dormant. In this paper, we investigate the effect of
framework for sensing called cooperative spectrum sensing
practical channel estimation on the performance of cooperative
17 sensing interweave-spectrum sharing networks. Specifically, we (CSS) has been proposed to enhance this reliability.
18 compare least squares (LS) channel estimation and minimum Relevant literature review: CSS techniques are used widely
19 mean square error (MMSE) channel estimation techniques in in the literature on spectrum sharing networks to address the
20 a flat fading scenario. We verify that a MAP-based combining
problem of degraded spectrum sensing performance during
21 scheme performs better than a majority voting system over a
large scale and small scale fading effects [3]. The work in [4]
larger range of energy-detection threshold and fading averaged
22 signal power to noise power ratio (SNR). However, we see the is an early work that demonstrates the benefits of coopera-
23 trade-off between the cost of complexity. Furthermore, we also tion between CR systems. Specifically, the authors proposed
24 find that the majority combiner is robust to channel estimation algorithms for multi-user spectrum sharing networks with
25 errors for large SNRs. On the other hand, the MAP-based
an improved probability of detection and reduced detection
26 combiner is sensitive to estimation errors if either the threshold
or SNR changes. Finally, we notice a trade-off in changing the time. A key assumption in the paper is the cognitive users’
27 ED threshold to achieve a higher probability of detection but at knowledge of the primary transmitter’s location. Also, channel
28 the cost of a higher probability of false alarm. state information (CSI) is assumed to be known. In [5], the
29 authors proposed a two-stage cooperative spectrum sensing
Index Terms—Spectrum sharing, interweave mode, cooperative
30 sensing, LS estimation, MMSE estimation, MAP combining, ma- scheme to sense the PU’s signals in multiple frequency bands.
31 jority combining, detection probability, false alarm probability. These stages are Energy detection (ED) and cyclostationary
32 feature detection (CFD). Furthermore, The authors further in-
33 vestigated the proposed model to verify that it simultaneously
34 improves energy efficiency and saves time. However, they do
I. BACKGROUND AND M OTIVATION
35 not consider the impact of imperfect CSI at the second stage.
36
37
38
T HE electromagnetic (EM) spectrum is a precious resource
for various fixed and mobile wireless communication sys-
tems. According to the Federal communications commission
The authors in [6] developed the performance analysis of
spectrum sharing systems with imperfect channel estimation.
They considered Minimum Mean Squared Error (MMSE) and
39 (FCC), geographical and time variations present in the utiliza-
linear-MMSE techniques to estimate the channel and the per-
40 tion of the assigned spectrum. These variations are ranging
formance of these estimation methods under imperfect sensing
41 from 15%−85% [1]. Spectrum sharing networks (SSN) aim to
conditions. However, the authors do not exploit the advantages
42 address the issue of spectral scarcity and under-utilization by
of cooperation in CR networks. In [7], the authors considered a
43 enabling the opportunistic use of frequency bands that licensed
fusion-center (FC) based cooperative spectrum sensing system
44 or primary users do not occupy. This method of CR is called
model. Specifically, the employed the maximum a posteriori
45 the interweaving technique. The interweave spectrum sharing
(MAP) estimation technique at the fusion center (FC) to
46 mode is an attractive solution due to its similarity with time
combine the decisions of all the individual SUs. However, the
47 division duplexing systems and collision avoidance protocols
impact of channel estimation and imperfect CSI is missing.
48 like CSMA/CA. Hence, making the integration of this novel
Recent interests in cooperative sensing include learning-
49 idea into current standards very easy [2].
aided approaches for SSN. In [8], the authors proposed re-
50 In interweave mode-based spectrum sharing networks, un-
inforcement learning (RL)-enabled CSS for SUs in an SSN.
51 licensed secondary users (SUs) should reliably sense the
However, the performance of RL depends on the quality
52 spectrum to check whether it is used by a primary user (PU).
of feedback data. For instance, in highly dynamic wireless
53 Several different approaches are proposed for identifying the
channel environments, the CSI that is fed back may not be
54 presence of PU signal transmission. However, due to the small
useful. In practice, we get imperfect CSI which will degrade
55
The first and the second authors are formerly with the Dept. of Electrical the detection performance.
56 and Electronics Eng. at the Birla Institute of Technology & Science (BITS),
57 Remarks on novelty: To the best of the authors’ knowledge,
Pilani, India. The third author is with BITS Pilani, India.
58 Emails: f20170220p@alumni.bits-pilani.ac.in, f20170131p@alumni.bits- investigation on the effects of imperfect channel sensing on the
59 pilani.ac.in, sainath.bitragunta@pilani.bits-pilani.ac.in performance of majority and MAP-based combiners for coop-
60
Page 3 of 5
2

1
2 erative interweave SSN (CIW-SSN) has not been explored. ED
Fusion Center
3 The key contributions are as follows.
SU1
4 • CIW-SSN with imperfect channel sensing: We propose 1011...X1
Channel Estimation of:
h1
5 an algorithm with imperfect sensing and study the ef-
PU
ED h 1 , h2 , h3
6 fects of imperfect channel estimation on a cooperative PU
SU2
0011...X2 LS MMSE
7 interweave-SSN. Investigating and compensating for the
h2
8 effects of imperfect channel sensing help better spectrum ED
1001...X3 Detection Based On:
Codeword: X1 X2 X3
9 management in cooperative SSNs. PU SU3
Majority MAP
10 • Performance benchmarking: The authors compare the
h3
11 performance of the MAP combiner and the majority com-
12 biner that employ estimated CSI. The least squares (LS) Fig. 1: CIW-SSN model with imperfect channel sensing.
13 and minimum mean square error estimation (MMSE)
14 techniques are considered. Specifically, the authors obtain
15 numerical results for the probability of detection (D) and III. D ETECTION S TAGE ONE
16 the probability of false alarm (FA). These results help The SUs use the energy detection (ED) technique to make
17 understand the impact of SNR and threshold on the key the local decision of whether the PU is transmitting on the
18 two probability measures. channel or not. The ED technique is chosen since it does not
19 require any prior information about the PU, such as its data
Notation and outline: We denote the probability of an event
20 rate and corresponding modulation.
by P (.), a vector is denoted by y and a matrix by Y . The
21 The energy of the received signal is calculated and com-
diagonalization of a vector is denoted by diag(.) and the
22
23
Hermitian transpose of a vector is denoted by y † . Further, for pared to a predefined threshold √ √ the bits 1 and
(TH). In BPSK,
any real b and y < 0, γinc (b, y) denotes the incomplete Gamma 0 are translated to the symbols Es and − Es , respectively,
24 where Es is the symbol energy. The threshold, shown in [7],
function [9]. Rest of the presentation is as follows. Section II
25   2
develops the CIW-SSN model with imperfect channel sensing. is given by THn = Q−1 FA2 n
P 2
26 σPU n
, where PFAn is
Sections III and IV present details on the two-stage cooperative
27 the desired local false alarm probability at the nth SU where
sensing system. Section V presents the simulation results and
28 2
n ∈ {1, 2, · · · , N }, σPU is the channel noise power between
the conclusions follow in Section VI.
29 n
the PU and the n SU, and Q−1 (.) denotes the inverse Q
th
30 function defined on standard Gaussian distribution.
31
32 II. C OOPERATIVE I NTERWEAVE -SSN M ODEL
IV. D ETECTION S TAGE TWO
33
Consider a system of N SUs, one base station (BS), one Assuming perfect synchronization between the SUs and the
34
FC, and PUs who get priority access to the channel. Figure 1 FC, the FC constructs a codeword (CW) from the data received
35
depicts this system model. The SUs can only transmit on the from all the SUs. The FC then combines the decision bits
36
channel when the PU is not transmitting. The FC decides of the CW to decide about the status of the PU. The data
37
whether the channel is idle or not once all the SUs have received at the FC is the channel corrupted version of the CW.
38
transmitted their sensing results to it. The cooperative sensing Hence, we need to recover the transmitted data by removing
39
process occurs in two stages: i) The decisions of each SU are the channel effects on the transmitted CW.
40
transmitted as a bit, with 1 denoting the presence of the PU Our paper uses pilot-aided channel estimation, zero-forcing
41
and 0 its absence. ii) The individual decisions of the SUs are (ZF) equalization, and hard thresholding to help us detect the
42
combined at the FC to arrive at the final decision. received CW. To help us achieve this, we send the data from
43
The SUs first detect the PU status, that is, active or idle, each SU to the FC in the form of packets where each packet
44
using the energy detection technique. Each SU then sends a consists of a pilot bit followed by Nd data bits. The channel
45
data bit which is 1 if the energy detected is greater than a estimation is done for every packet sent out using the pilot
46
predefined threshold, or a 0 if the energy is lesser than this bits of all the SUs. We then recover the data from each packet
47
threshold. The SUs also transmit a pilot bit for every block using the corresponding channel estimates.
48
of Nd data bits transmitted to the FC. Note that the pilot We assume that the packet transmission time is much less
49
bits are useful for the FC estimate the channel. At the FC, than the channel coherence time. Furthermore, the channel
50
the channel coefficients for each SU-FC link are estimated estimation has been done using two methods, namely LS and
51
based on the pilot bit. The FC uses the transmitted data bit by MMSE, and the subsequent comparison of the performance
52
each SU to construct a N × 1 vector, called a codeword. The between the two is shown in Section V.
53
54 codeword is used to decide the status of the PU employing the
55 MAP and majority-based combining techniques. We assume A. Least Squares (LS) Estimation
56 that the data and pilot bits are transmitted using the BPSK For a packet i sent out by all the users, the received N × 1
57 modulation scheme. Both the PU-SU and SU-FC links are pilot vector, y p,i , is given by
58 modeled as Rayleigh frequency-flat fading channels. In the
59 following sections, we present the two detection stages. y p,i = Xp hi + wp,i , (1)
60
Page 4 of 5
3

1
2 where Xp is the N ×N diagonal matrix formed from the N ×1
3 transmitted pilot vector, xp . The noise vector, wp,i , consists of
4 circularly symmetric complex Gaussian (CSCG) i.i.d. random PU Status Codeword P (CWr )
5 variables, with mean 0 and variance σ 2 .
6 Since Xp is a square matrix, the LS estimated channel CW1 1, 1, 1 PD1 , PD2 , PD3
7 vector, ĥLS,i , is given by ĥLS,i = Xp−1 y p,i . Using the ZF CW2 1, 1, 0 PD1 , PD2 , PMD3
CW3 1, 0, 1 PD1 , PMD2 , PD3
8 method, the recovered CW is given by Active CW4 1, 0, 0 PD1 , PMD2 , PMD3
9 ˆ LS −1 CW5 0, 1, 1 PMD1 , PD2 , PD3
CW b,i = [diag(ĥLS,i )] y b,i , b ∈ {1, · · · , Nd }. (2)
10 CW6 0, 1, 0 PMD1 , PD2 , PMD3
CW7 0, 0, 1 PMD1 , PMD2 , PD3
11 B. Linear Minimum Mean Square Estimation (LMMSE) CW8 0, 0, 0 PMD1 , PMD2 , PMD3
12
The linear MMSE estimate is given by [10]
13
E[|h|2 ] CW9 1, 1, 1 PFA1 , PFA2 , PFA3
14 u= x, (3) CW10 1, 1, 0 PFA1 , PFA2 , 1 − PFA3
15 E[|h|2 ]Es + σ 2 p CW11 1, 0, 1 PFA1 , 1 − PFA2 , PFA3
16 where Es and E[|h|2 ] are the energy of the pilot symbols and
Idle CW12 1, 0, 0 PFA1 , 1 − PFA2 , 1 − PFA3
CW13 0, 1, 1 1 − PFA1 , PFA2 , PFA3
17 the average channel gain, respectively. Since we are taking a CW14 0, 1, 0 1 − PFA1 , PFA2 , 1 − PFA3
18 single tap for the Rayleigh channel whose power is normalized CW15 0, 0, 1 1 − PFA1 , 1 − PFA2 , PFA3
19 1
to 1, (3) is then u = Es +σ 2 xp . Further, the MMSE estimate,
CW16 0, 0, 0 1 − PFA1 , 1 − PFA2 , 1 − PFA3
20
ĥMMSE , for a packet i sent out by each SU is given by
21 TABLE I: List of probabilities for MAP combiner for both
ĥMMSE,i = u† Yp,i , where Yp,i is the N × N square matrix
22 active and idle PU.
obtained after diagonalizing y p,i . The codewords are recovered
23
from the received data are given by
24
25 ˆ MMSE
CW b,i = [diag(ĥMMSE,i )]−1 y b,i , b ∈ {1, · · · , Nd }. (4)
26
We use two techniques to combine the data bits in the
27 Algorithm 1 MAP based combiner with channel estimation
codeword to arrive at the final decision about the PU status.
28 Inputs: xp , P (CWr ), σ 2 , Es
29 C. Majority combiner Output: PULS status, PUMMSE status
30 Preprocessing: Xp = diag(xp )
31 This combiner is analogous to a majority voting system. 1: for each packet i do
After the channel estimation, if the estimated codeword has 2: Yp,i = diag(y )
32 3:
p,i
LS estimate: ĥLS,i = Xp−1 y , HLS,i = diag(ĥLS )
more 1s than 0s, the PU is active. Similarly, if there are more
33 4:
p,i
LMMSE estimate: ĥMMSE,i = u† Yp,i , HMMSE,i = diag(ĥMMSE,i )
0s than 1s, the PU is said to be inactive. Suppose the codeword
34 5: for each data bit b ∈ {1, · · · , Nd } do
contains an equal number of 1s and 0s. In that case, the PU is
35 6: Yb,i = diag(y )
b,i
declared active because it is safer to falsely classify the PU to 7: for r ∈ {1, · · · , 2N } do
36 8:
be active, even if it is idle, than to misclassify it as idle and
37
LS 2
KA,b,i (r) ← ||y − HLS,i xb,i ||
b,i
risk causing interference with the PU.
38 2
− σ log(P (CWr ))
39 D. Maximum a Posteriori (MAP) combiner 9:
40
LS 2
KI,b,i (r) ← ||y − HLS,i xb,i ||
b,i

41 The MAP combiner uses the local information at each SU 2


− σ log(P (CWr+2N ))
42 to help us determine the status of the PU at the FC. The
parameters, other than PFAn , which are calculated at each SU, 10:
43  √
MMSE
KA,b,i (r) ← ||y
b,i
− HMMSE,i xb,i ||
2

44 are given by PMDn = 1 − γinc 1, √ 2 THn


, PDn = 2
− σ log(P (CWr ))
σPUn (1+Es /σPU
2
)
45 n
1 − PMDn . Here, PMDn denotes the probability of misdetection 11:
46 at the nth SU. PDn is the probability of detection at the nth
MMSE
KI,b,i (r) ← ||y
b,i
− HMMSE,i xb,i ||
2

47 SU. For our illustration, we have taken N = 3. Note that it is


2
− σ log(P (CWr+2N ))
48 possible to determine the results to large N . The probability 12: end for
49 map of all possible codewords from three SUs, with their 13: LS
if min(KA,b,i LS
) < min(KI,b,i ) then
50 indicated PU status, is as shown in Table I.
14: PULS
b,i status ← PU is active
15: else
51 The pseudocode of the algorithm, which utilizes the local 16: PULS status ← PU is idle
52 probabilities, data sent from each SU, and the estimated
17: end if
MMSE MMSE
18: if min(KA,b,i ) < min(KI,b,i ) then
53 channel matrix, is shown in Algorithm 1. The MAP cost 19: PUMMSE status ← PU is active
b,i
54 function, derived in [7], is used to determine the status of 20: else
21: PUMMSE
55 the PU. The cost function is given as 22: end if
b,i status ← PU is idle

56 23: end for


ˆ = argmin(||y − Hi x ||2 − σ 2 log(P (CWr )), (5)
CW
57 b,i b,i 24: end for
CWr
58
59 where Hi is the diagonalized LS or MMSE estimated vector.
60
Page 5 of 5
4

1
2 V. N UMERICAL RESULTS the threshold increases in MAP combiner, a trade-off arises
3 We conduct simulations on MATLAB for 10 iterations. In 2 between lower PD and lower PFA . We can also make similar
4 each iteration, we send 104 CWs to the FC. To compare the remarks for the probability of misdetection (PMD = 1−PD ). In
5 performance between the two combiners and between the two addition to the threshold, SNR is another important parameter
6 estimation techniques, we first vary the local PFA of the SUs as that shows the impact on PD and PFA . Below, we present a plot
7 shown in Table II, which presents the simulation parameters. showing the impact of SNR on PD in for the three scenarios.
8 The transmitted SNR at the PU and the SUs are kept constant
9 at 10 dB when the PFA varies. Next, we keep the local PFA 1

10 at a constant 5% and then change the SNR of the PU and the 0.95

11 SUs as shown in Table II. 0.9

12
0.85
MAP Combiner
0.8
13 TABLE II: Simulation parameters
0.75

14 0.7
Majority Combiner

Parameter Value
15 Number of SUs 3
0.65

16 SNR 0 − 20 dB
0.6

0.55
17 σ 2 , σPU
2
n
10 dB 0.5

18
0 2 4 6 8 10 12 14 16 18 20
mean channel power gain 0 dB
Nd 4 bits
19
PFA 10−4 − 0.99 Fig. 4: Probability of detection as a function of SNR.
20
21
22 Impact of SNR on PD : Figure 4 plots the probability of
23 detection as a function of SNR for the two combiners, namely,
1.05

24 MAP and majority combiner. The MAP combiner outperforms


1

25 the majority combiner in terms of both PD and PFA for the


0.95

26 entire range of SNR. Figure 4 shows that there is no significant


0.9

27
MAP Combiner difference between the MMSE and LS estimation techniques
0.85

28 when the majority combiner is used, whereas the MMSE


0.8

29 Majority Combiner estimate performs slightly better than the LS estimate in the
0.75

30 low SNR regime for the MAP combiner.


0.7

31 Remarks: The trade-off between PD and PFA as the ED


0.65

32
10-3 10-2 10-1 100 threshold varies is fundamental to SSN. It illustrates the
33 need to select the threshold to be low enough to detect the
34 Fig. 2: Probability of detection as a function of threshold. PU transmission but high enough to ensure SUs can access
35 the channel regularly. The results also show that the MAP
36 combiner is sensitive to channel estimation errors when the
37 1
threshold or SNR changes. On the other hand, the majority
38 0.9 combiner depends only on the accuracy of the SU’s channel
39
0.8 MAP Combiner
sensing. Therefore, it is sensitive to SNR change and nearly
0.7

40 Majority Combiner insensitive to channel estimation errors. We see that the


0.6

41 0.5
performance of the MAP combiner is superior. The improved
42 0.4 performance comes at the cost of higher complexity. As the
43
0.3 number of SUs increases, the number of CWs that have to be
44
0.2
tracked increases exponentially.
0.1

45 0
10-3 10-2 10-1 100
46 VI. C ONCLUSIONS
47
Fig. 3: Probability of false alarm as a function of threshold. In this paper, we studied the effect of channel estimation on
48
49 the performance of CSS interweave-SSN. We verified that a
50 Impact of threshold on PD and PFA : Figure 2 plots the MAP-based combining scheme performs better than a majority
51 probability of detection as a function of the threshold for voting system over a larger range of energy-detection threshold
52 the MAP and majority combiners. Further, Figure 3 plots the and fading averaged SNR, albeit at the cost of complexity.
53 probability of false alarm as a function of the threshold for We also found that the majority combiner is robust to channel
54 different combiners. Both figures plot three different scenarios, estimation errors if the SNR is high enough. On the other hand,
55 namely, LS, MMSE, and ideal. the MAP-based combiner is sensitive to estimation errors
56 We see from Figure 2 and Figure 3 that the MAP combiner if either the threshold or SNR changes. We also noticed a
57 outperforms the majority combiner for both PD and PFA . trade-off in changing the ED threshold to achieve a higher
58 However, we see that the majority combiner is relatively robust probability of detection but at the cost of a higher probability
59 to channel estimation techniques. Further, we also see that as of false alarm.
60
Page 6 of 5
5

1
2 R EFERENCES
3 [1] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “Next
4 generation/dynamic spectrum access/cognitive radio wireless networks:
A survey,” Computer Networks, vol. 50, no. 13, pp. 2127–2159, 2006.
5 [2] E. Biglieri, A. J. Goldsmith, L. J. Greenstein, N. B. Mandayam, and
6 H. V. Poor, Principles of Cognitive Radio. Cambridge University Press,
7 2012.
[3] S. M. Mishra, A. Sahai, and R. W. Brodersen, “Cooperative sensing
8 among cognitive radios,” in Proc. ICC, 2006, pp. 1658–1663.
9 [4] G. Ganesan and Ye Li, “Cooperative spectrum sensing in cognitive radio
10 networks,” in First IEEE International Symposium on New Frontiers in
Dynamic Spectrum Access Networks, 2005. DySPAN 2005., 2005, pp.
11 137–143.
12 [5] M. Mehdawi, N. Riley, A. Fanan, and O. Bentaher, “Proposed system
13 model for wideband cooperative spectrum sensing with multi-bit hard
decision using two-stage adaptive sensing,” in 2019 27th Telecommuni-
14 cations Forum (TELFOR), 2019, pp. 1–4.
15 [6] S. Akin and M. C. Gursoy, “Performance analysis of cognitive radio
16 systems with imperfect channel sensing and estimation,” IEEE Trans.
Commun., vol. 63, no. 5, pp. 1554–1566, 2015.
17 [7] A. Tohamy, U. S. Mohamed, M. M. Abdellatif, T. A. Khalaf, and M. Ab-
18 delraheem, “Cooperative spectrum sensing using maximum a posteriori
19 as a detection technique for dynamic spectrum access networks,” IEEE
Access, vol. 8, pp. 156 408–156 421, 2020.
20 [8] W. Ning, X. Huang, K. Yang, F. Wu, and S. Leng, “Reinforcement learn-
21 ing enabled cooperative spectrum sensing in cognitive radio networks,”
22 Journal of Communications and Networks, vol. 22, no. 1, pp. 12–22,
2020.
23 [9] L. S. Gradshteyn and L. M. Ryzhik, Tables of Integrals, Series and
24 Products. Academic Press, 2000.
25 [10] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
USA: Cambridge University Press, 2005.
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

You might also like