Sei sulla pagina 1di 6

Optimal Sequential Channel Estimation for

Multi-channel Cognitive Radio


Raied Caromi and Lifeng Lai
Department of Systems Engineering
University of Arkansas, Little Rock, AR 72204, USA
Email: { rmcaromi, lxlai}@ualr.edu

Abstract-In this paper, we consider how to quickly obtain be empty, through channel scanning algorithms for example
accurate channel gain estimates for multi-channel cognitive the one developed in [6]. The goal of the cognitive radio
radio systems. A sequential setup, in which the cognitive users
is to get accurate estimates of the channel gains of these
estimate available channels one by one, is considered. In addition,
the channel estimation for each channel is also sequential. I n
channels with a minimal number of training symbols. We
particular, cognitive users will determine whether t o finish the consider a sequential estimation setup, in which the cognitive
estimation process of the current channel and switch to next radio pair obtain estimates of the channel gains one by one.
channel for estimation based on the training symbols received Furthermore, the channel estimation process within a channel
so far. To balance the tradeoff between the estimation error and
is also a sequential one. More specifically, in stead of fixing
time spent on the channel estimation, we aim to design a switch
function and an estimator that minimize a linear combination of
the number of training symbols used, the cognitive transmitter
these two factors. We obtain the optimal solution for the general keeps sending training symbols over a channel. Once the
case. Numerical results are provided to show the substantial gain receiver obtains a reasonably good estimate of the channel
of our solution over that of uniformly allocating training time to gain, it will sends a switch signal to the transmitter. After
each channel.
receiving the switch signal, the pair then move to estimate
Index Terms-Bayesian estimation, cognitive radio, channel the next available channel. The process terminates once all
estimation, sequential analysis, spectrum sharing.
the channel gains have been estimated. To strike a balance
between the estimate error and the time required for training,
I. INTRODUCTION
we use a linear combination of the estimation variance and
Opportunistic spectrum access is one of the key aspects of total training time as the cost function. Our goal is to design
cognitive radio technology. Real measurements have shown switching rules and estimators that jointly minimize this cost.
that spectrum is under utilized even in the most crowded We first consider the special case in which there is only one
areas. While demand is increasing, the spectrum under uti channel in Section III. Due to the special property of Mini
lization problem could be mitigated by allowing opportunistic mum Mean Square Error (MMSE) estimation, the sequential
spectrum access without interrupting the primary system [1], estimation problem under consideration can be converted into
[2]. There has been extensive studies that focus on detecting a fixed size estimation problem. The problem of minimizing
spectrum opportunities [3]-[5], i.e., the spectrum sensing the cost function is equivalent to the problem of finding the
problem. However, so far, the channel estimation problem has optimal training size problem, which can be easily solved. In
not received much attention. In wireless fading channels, the Section IV, we then consider the general case in which we
channel gain between the secondary transmitter and receiver is have multiple channels to estimate. Using the results obtained
time-varying. Hence to utilize the identified empty channel, the from the single channel case, the problem at hand is converted
secondary transmitter needs to send a training sequence so that to a problem of allocating training symbols among multiple
the secondary receiver can obtain an estimate of the channel channels. It is found that the optimal solution that minimizes
gain. Clearly, a longer training sequence will lead to a smaller the total risk for the multiple channels case has two regions,
estimation error. However, in cognitive radio systems, the time namely region one when the cost per sample is relatively small
period before the primary user returns (i.e, the transmission and region two when the cost per sample is large. In region
opportunity) is often short. If too much time is spent on the one, the optimal solution uses all the available training time
training, the time left for real data transmission is then limited. divided between channels based on their qualities. While in
The problem is even more involved if there are multiple bands region two, the optimal total number of training symbols used
need to be estimated. depends directly on the cost per sample. The larger the cost,
In this paper we consider the channel estimation problem for the smaller total number of training symbols will be used.
multiple-channel cognitive radio systems. In the setup consid In both regions, one will allocate more training symbols to
ered, there are multiple channels that have been identified to channels with better quality. We derive the optimal allocation
scheme. In Section V, we provide several numerical examples
The work of R. Caromi and L. Lai was supported by the National Science
Foundation CAREER award under grant CCF- 1O-S4338 and by the National to compare the performance of the optimized system to that of
Science Foundation under grant DMS-1l- 18822. allocating training symbols uniformly over multiple channels.

978-1-4673-3140-1/12/$31.00 2012 IEEE


I I. S YSTE M M ODEL

We consider a primary communication system with Mp Ch 1


licensed channels. Suppose M of them have been identified N2
to be empty!, the SU wishes to estimate the channel gains Ch 2
between the secondary transmitter and receiver. We consider
=N3
---
_ . . " " " , . . . . .

a block channel fading model and assume that the channels Ch 3

are independent. Furthermore, we assume uncorrelated WGN


over each channel and the channels are not identical in terms
of gain. The received signal for channel k at each time slot i
is ChM
(1)
in which Ski is the transmitted signal in channel k at time slot
i and is known in the training session, Hk channel coefficients Fig. 1. System Model
with distribution N (0, O") and Wki is i.i.d Gaussian noise
'"

with distribution N (0,0";).


'"

The channel estimation problem has two effects on the linear combination of both
overall achievable throughput. First, we want to reduce the
p(Hk;XkNk) + CNk,
A

(3)
number of symbols required by channel training process so
that more time is left for data transmission. Second, we want in which c the cost per training symbol. We use
to reduce the channel estimation error, which however requires
a longer training period. It is well known that sequential
procedure could yield to a better performance in terms of
U{!k(X{!k) =
! { p(Hk;xf) + cNk } (4)

number of symbols required for training [7]. Hence we adopt to denote the minimal cost among all estimators that use the
a sequential estimation process. Assume Nk is the number observation X{!k.
of training symbols used for channel k and the maximum Our goal is to design switch function and estimates H
training session time is upper bounded by aTt, where a is that minimize the total cost incurred for these M channels,
a constant 0 < a < 1 and Tt is the total transmission that is we aim to solve the following optimization problem
session time. Fig. 1 shows a typical operation of the proposed
sequential estimation model. More specially, without loss of
generality, the cognitive users start from the first channel. For
r= inf IE
,H
{t [
k =l
p(Hk;X{!k) + cNk ]}, (5)

each channel k, the receiver makes two decisions based on the s.t. L Nk ::::: aTt.
received training symbols X{!k = (xk(l), ,xk(Nk)). The
first one is a switch decision which determines when the users To simplify the solution more we will first investigate the
should switch to the next available channel for training. We single channel case then generalize for the multi-channel case.
use k to denote the switch rule used, in which k(X{!k) = 1
I I I. SING LE CHANNE L CASE
denotes a switch while k(X{!k) = 0 implies to stay on the
same channel for estimation. We assume that the users will We first consider the single channel case, in which no switch
not come back to a channel once they switch from it. We use is required. Once the estimation is completed, the procedure
to denote (l,. ,M ). Note that M = 1 indicates that is terminated. Furthermore, for simplicity we will drop the
the whole estimation process is finished. The second one is an time constraint in (5) and the k notation for this case. Using
estimate Hk of Hk, which has estimation error cost associated the channel model of (1) and choosing the squared error loss
with it: L(H,H) = (H -H ) 2 as the loss function (hence, we focus on
MMSE estimate), then from (2) the posterior risk p(H;XN)
p(Hk;X{!k) = l [Lk(Hk, Hk)] (2) is

= I: Lk(Hk,Hk)Jr(Hk I X{!k) Hk d '


p(H;XN) = I: (H _H )2Jr(H I XN) H d = var(H I XN),

'JJ
where Lk(Hk,H is the loss function for estimating channel
where var(H I XN) is the conditional variance of H. We
can easily evaluate the resulting MMSE estimation H, which
k and Jr(Hk I Xk k) is the posterior density of Hk after Nk
is the conditional mean, and var(H I XN), which is the mean
observations. We use H to denote (HI, ,HM).
square error [8]. For the general model, the conditional mean
To balance the estimation error and the number of training
and covariance are
symbols required, the total cost associate with channel k is a
E(H I XN) = /-lH+(Ci/+STC;;/S)-ISTC;;/(XN -H/-lH),
1 In this paper, we assume that the detection is perfect. The impact of
detection error will be considered in the future work. CH1xN = (Ci/ + STC;;/S)-l,
where fJH is the mean and CH is the covariance matrix. For as in the single channel case. Then, the optimization problem
the prior PDF's mentioned earlier and assuming S = 1, after in (5) is reduced to
taking N symbols the conditional mean and variance can be
evaluated as Minimize r = f
k=1
( k
1 + kl(J"2Hk
+ Nk )
.
C , (11)

(6) M
S.t. L Nk ::; aTt
A
(J"(J"; k=1
var(H) =
2 + N(J"2 . (7)
(J"w H The objective function in (11) is convex in Nk and the effective
value of a defined as aeff =
I:M . that
kT Nk. , one should notice
Since var(H) and hence UN (XN) in (4) does not depend on
minimizing over Nk is the same as minimizing over aeff,
Xn for all n 2: 1, the problem will reduce to a fixed sample
since aeff constraints to a as the upper bound. Using convex
size problem [7]. The total risk in this case is
optimization techniques to solve this problem, we first write
(J"2H(J"2 the Lagrangian function as
r = ;; + N c.
(J"w+ (J"2H
2 M 2 M

The optimal number of symbols N that minimizes the risk ().., Nk) = L( 1 + Hk k (J"2 + cNk) + )..(L Nk - aTt).
k=1 l Hk k=1
can be found by setting trv = 0 ,

iN*l
{-f 2
(J"w _ (J"w
,
The KKT conditions are
a().., Nk) -((J"k)2,
+ c + ).. = 0,
=
(J" (12a)
[1 + Nkl(J"2]2
C
aNk Hk
where i.l represents a positive integer approximate. To sim M
plify the notation, we will drop this symbol from the optimal )..(L Nk - aTt) = 0, (12b)
number of symbols N* and always consider it equal to a k=1
the nearest positive integer or zero. The estimation error
if = H - H is gaussian with zero mean and variance (12c)
(J"7 = var(H). For the data transmission the received symbols k=1
will suffer additional error that can be added to the original ).. 2: 0, (12d)
WGN where X = H +w and w = if +w. The effective noise Nk 2: 0, k = 1,2, ... M. (12e)
. 2 'IS
vanance (J"ef f
Solving for these equations we first find
2
2
(J"eff =
2
(J"w
(J"(J";
+ N (J"H
2 + (J"2w =
((J";) +
.- -
-'...O:::":"
(J"w
(N +
" -; 2:-'-----:::-
+ N
1)(J"(J";
:- ' 2.;-"'--"'-
(J"H N;; =
(J)..
I
- _1_),
(J"k
I
+C
(13a)

IV. MULTI-CHANNEL OPTIMIZATION PROBLEM I _ c,


)..*
[aTn + -L 2: +]2
= (13b)
The problem of multi-channel is formulated as a joint M M J-1 (JH
j
optimization problem that minimizes the total risk as in (5) . M
Assume the transmitter SNR I = S/(J"; , where S is the aTt _1_ '" _1_ __ 1_
N*k = + (13c)
average power of the signal that is equal on each channel and M M I (J"2H 1 (J" 2'
H k
j=1 j
(J"; is the noise power. The estimate Hk, the variance var(Hk),
To simplify the notation define a as
(J" ; ( )
and effective estimation error variance ff k along channel
k are shown in (8), (9) and (10) respectively
a =
[ aTti
+
_1
2
1
M M
j=1 (J"Hj
Examining the condition in (12b) we find that if ).. > 0 then
the second term is equal to zero and the solution in (l3c) for
Nk satisfies optimality conditions for c ::; l/a2. On the other
hand if c 2: 1/a2 , then).. should equal zero to satisfy (12b)
and (12d) , then optimal solution for N;; is

The objective in the joint optimization problem is to minimize _ _1_ .


N*k
.JYC ,(J"k
=
__ 1
(14)
the total risk jointly over all available channels subject to a
time constraint, hence maximizing transmission rate. Since the To summarize, the optimal solution for Nk in (13a) and (14)
Uf: (Xf:) function in (4) does not depend on xl.: for all nk 2: 1 represents the optimal number of symbols for each channel
then the problem will reduce to a fixed sample size problem that minimize the total risk for region one when c ::; 1/a2
and for region two when e 2: r / a2 respectively. The solution total number of allowed symbols during the estimation session
for both regions is equal at the critical point e = r / a2. The is set to constant cxTt. Furthermore, the number of symbols is
total minimized risk for each region is found as divide equally between all available channels.

rtl =
M
a
+ caTt ,e:S;
a
, (15a) V. N UMERICAL RESULTS
M N umerical tests are performed to evaluate the performance
rt2 = 2M - 2: L _1_ ,e 2: . (15b) of the proposed model. In all of the following simulations, we
V r r j=l (Jj a assume M = 16, that is 16 channel were identified free in the
It is obvious that the optimal number of symbols per channel detection session ignoring sensing errors. The transmitter SN R
for region one changes with r(Jk proportionally since the is assumed to be fixed during training and data transmission
solution first divide the total time uniformly then allocate session which mean no power allocation is performed. Channel
prior variances are selected randomly between 0.1 and 1 to
some of the time according to channel quality, i.e. give more
test the proposed model performance upon unequal channels'
time for better channel. At region two, unlike region one the
characteristics. In these tests we compare the jointly optimal
solution depends on the cost e directly since the effect of
the cost is more than the effect of channel quality when the model with the uniform model that is to divide the available
cost per sample is relatively high. The optimal number of time cxTt equally between all channels. We set T = 10000 and
symbols decreases inversely with .JC but still give more time cx = 0.15, note that choosing alpha with higher values will
for better channel through the term r(Jk' that is give more more degrade the performance of the uniform model since the
time for better channel at this region. On the other hand, one time is fixed for each channel, therefor we chose a reasonable
should notice that as the constraint cx is more relaxed then the value of cx that guarantee a fair comparison for broad values
critical point e = r / a2 shifts to the left and intuitively that of SN R.
is convenient since the no time constraint solution is that of Fig. 2 depicts the risk of the proposed model against the
region two.
uniform model for different values of cost c and two values
To analyze the data transmission performance of the opti of SN R. In the uniform model, the risk vs cost is very close
mized problem, we first compute cxeff that represents the total to be linear since N is fixed and the variance term in (11) is
very low, hence the the cost term eN dominates. In general
number of symbols N = Ll Nk
the risk increases as the cost increases and is higher for lower
CXeff = CX, SN R. The rate of risk increase is much lower in the optimized
case because of the minimized N as the cost increases, that
M
1 M 1 is to maintain the risk minimization as the cost increases. At
cxeff =
Tt ;;:yc
( -L 2 ), >2
e-
k=l r(JHk a2' the lower values of c, the performance of the optimized and
v I
uniform models is very close since the optimized model uses
where N = cxeffTt, the channel gain estimation and effective all the available time cxTt in region one. Comparing the case
noise are of OdB and -10dB SN R's, the critical point a'2 = e is shifted
.
up ill the case of S NR = -10dB. Intuitively that is expected
2
(Jeff = e+
(Jw2 , since at lower SN R the model needs more time to perform
, e bk
Hk = (1 - -2- ) 1:::72 ' estimation and hence it uses all the available time until the
(JHk v (J
. r critical point.
where bk is the sample mean of the k h channel and
t e is the
error term defined as 9 ,-----------

-Optimized SNR-IO dB
- - - Uniform SNR -\ 0 dB
-Optimized SNR 0 dB
- - - Unifonn SNR 0 dB

-i:l5
i024

e=
The data transmission rate for M parallel channels is

R = (1 - cxeff) t log ( 1+ IS ) .
2 3 4
k=l eff(k) Cost (c per sample)
To evaluate the spectral efficiency of the proposed model, we
will compare it to a uniform model. In the uniform model, the Fig. 2. Risk vs cost SNR=O,-l0 dB
After showing the risk optImIzation problem in Fig. 2, outperforms the uniform model, as the cost decreases the
now we will investigate the performance of proposed model optimized model become closer to the uniform model. If the
for data transmission to see the effect of risk minimization cost equals to zero, the only enhancement that the optimized
on the transmission rate. In all these results, a Monte Carlo model could offer is when the channels have a very broad
simulation of 10000 is used. Fig. 3 shows the transmission rate characteristics, i.e. different channels have high and low loss.
of the optimized and uniform case vs SN R when c = 0.001 However, from practical point of view, the cost should not be
per sample. The effect of minimizing the risk and hence the zero because as the time spent on channel training goes up,
required number of symbols is obvious in maximizing the the total cost or transmission loss increases.
rate. To be more specific, the total value of the variance was
jointly optimized between all channels to maximize the rate. 1600 ,-------,--,--,
Furthermore, beside that the performance of the optimized case
is better than the uniform case, it can also be noticed from Fig.
3 that as SN R becomes larger, the optimized rate diverge more
1400
from the uniform rate because of the high reduction in N value
when the symbols are less noisy. The overall transmission 1200
rate gain is the sum of the effect of optimization between
the channels and the reduction in the total number of symbols 1000
N.
800
40 ,------.----._--_.,_--_,
600

30
-'-
4
---------'
4000L---------2----------'- -
6
------ ---------'
8 1O
- --
- -- SNR
:E 25 ..

--- .. ....
.... Fig. 4. ASN vs SNR, c=O.OOOI
20 ....
....
"
......
15 ..
....
..
.. 40 ,----.----._--_.,_--_,
10 '#

O
L- 30
o 4 6 8 10--------L---
SNR :E 2' ..-
:> .. ..--
'"' ....
20 ....
Fig. 3. Transmission Rate vs SNR, c=O.OOI
.. ....
" ....
15 .. ..
..
In this test, we will first investigate Fig. 4 and Fig. 5 when
the cost was decreased to c = 0.0001 per sample. Fig. 4 10
represents the total average sample number (ASN ) for different
values of SN R while Fig. 5 shows the rate with SN R for
the same case. The total ASN is highly reduced when SN R
2 4 6 8 10
become larger while it stays at the same level of exTt like SNR
the uniform case before the critical point of c = "( / a2 , but
unlike the uniform case, different channels are optimized to Fig. 5. Transmission Rate vs SNR, c=O.OOOl
use different values of number of symbols in this region. At
region one before the critical point the rate performance of the
uniform and optimized case is very close since larger time is VI. CONCLUSIONS

required due to noisy symbols. However, increasing the time In this paper, we have considered a channel estimation prob
, i.e the value of ex, is recommended in the optimized case but lem in multi-channel cognitive radio systems. A sequential
not for the uniform case since the model optimizes the time setup, in which the cognitive users estimate available channels
for different values of SN R's, but it will highly degrade the one by one, has considered. Optimal switching and estimator
performance of the uniform case since the time is fixed. To have been derived to minimize a cost function that is a linear
evaluate the effect of the cost reduction, we now compare Fig. combination of estimation error and time spent on estimation.
3 and Fig. 5. When the cost was higher the optimized model N umerical results have been provided to show the substantial
gain of our solution over that of uniformly allocating training
sequence to each channel.

REFERENCE S

[1] S. Haykin, "Cognitive radio: brain-empowered wireless communications,"


IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201 - 220, Feb. 2005.
[2] W. Wang and X. Liu, "List-coloring based channel allocation for open
spectrum wireless networks," in Proc. IEEE Veh. Technology Corif.,
Dallas, TX, Sep. 2005, pp. 690--694.
[3] Z. Quan, S. Cui, A. H. Sayed, and H. V. Poor, "Wideband spectrum
sensing in cognitive radio networks," in Proc. IEEE Int. Can! Commun.,
Beijing, China, May 2008, pp. 901-906.
[4] S. J. Kim and G. B. Giannakis, "Rate-optimal and reduced-complexity
sequential sensing algorithms for cognitive OFDM radios," in Proc. Can!
Inform. Sci. and Syst., Johns Hopkins Univ., Baltimore, MD, Mar 2009,
pp. 14 1- 146.
[5] S. J. Kim, G. Li, and G. B. Giannakis, "Minimum-delay spectrum sensing
for multi-band cognitive radios," in Proc. IEEE GLOBECOM Corif.,
Miami, FL, Dec. 2010, pp. 1-5.
[6] R. Caromi, Y Xin, and L. Lai, "Fast multi-band spectrum scanning for
cognitive radio systems," IEEE Trans. Commun. Submitted, 2011.
[7] M. Ghosh, N. Mukhopadhyay, and P. K. Sen, Sequential Estimation. New
York: Wiley and Sons, 1997.
[8] S. M. Kay, Fundamentals of Statistical Processing, Volume I: Estimation
T heory. New Jersey: Prentice Hall, 1993.

Potrebbero piacerti anche