Sie sind auf Seite 1von 10

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO.

2, MARCH 2007 779

Multiobjective Distributed Power Control Algorithm


for CDMA Wireless Communication Systems
Mohammed Elmusrati, Member, IEEE, Riku Jäntti, Member, IEEE, and Heikki N. Koivo, Senior Member, IEEE

Abstract—Although power control has been explored since the interference to other (co)cross-channel communication systems
early 1990s, there still remains the need for further research. Most by minimizing the transmitted powers. For this and other rea-
of the algorithms so far consider either the problem of minimizing sons, the power control problem has been receiving a lot of
the sum of transmitted power under quality of service (QoS)
constraints given in terms of minimum signal-to-interference-plus- attention in the wireless communication literature. Although the
noise ratio (SINR) in a static channel or the problem of mit- power control problem is well understood, there still remain
igating fast fading in a single dynamic link. In this paper, we many related problems that are under active research such as
suggest a new approach to the power control by treating the speeding up the convergence of power control algorithms, de-
QoS requirement as another objective for the power control and termining the optimum implementation of the distributed power
a fully distributed method for solving the multiobjective power
optimization problem. The obtained solution is parameterized so control (DPC), designing DPC without snapshot assumption
that a tradeoff can be made between power consumption and QoS. (i.e., fixed channel gains), and joining the power control with
In the limit case, when only QoS is weighted, the algorithm reduces other resource control methods such as rate control and spatial
to the well-known distributed power control algorithm (IEEE processing.
Trans. Commun., vol. 42, no. 2/3/4, pt. 1, Feb./Mar./Apr. 1994). It can be shown that the minimum transmitted power vector,
In the other limit, the algorithm reduces to transmission with
fixed minimum power. The convergence properties of the proposed which is required to achieve the target SINR to users, is the
algorithm are studied both theoretically and with numerical sim- solution of a system of linear equations. Direct solution of
ulations. Although we only consider SINR and power sum, our this equation system is called centralized power control, and it
algorithm could be easily modified to take other objectives, such requires full knowledge of the channel gains between all active
as throughput, into account. transmitters and receivers. This required knowledge makes the
Index Terms—CDMA, multi-objective optimization, power con- centralized power control impractical method. There are many
trol, radio resource management. iterative algorithms to solve the power control problem in
distributed fashion; see, e.g., [1]–[11]. The commonly used
I. I NTRODUCTION methodology in the literature of DPC is the so-called snapshot
analysis: The system parameters such as channel gains and
P OWER control is essential in interference-limited capac-
ity communication systems such as the direct-sequence
code-division multiple access (DS-CDMA). The power control
noise levels of users are assumed to be fixed or slowly changing
compared to the rate at which power updates can be performed.
regulates transmitted powers of the users to be as close to This assumption is required to allow the DPC to converge
optimum as possible. The optimum transmission power is the to the solution of the centralized power control algorithm. In
minimum power needed to achieve some target quality of ser- practice, however, the link gains of mobile channels have fast
vice (QoS) level to users, which is usually expressed in terms of and random fluctuations that occur in the same time scale as
signal-to-interference-plus-noise ratio (SINR). The power con- power can be updated. These characteristics of mobile channels
trol mitigates the well-known near–far problem in DS-CDMA reduce the significance of the snapshot convergence property of
systems, which results in enhancing the system capacity and the power control algorithms. The work in [3] and [4] does not
performance [2]. Moreover, power control prolongs the oper- assume the snapshot analysis, but the resultant power control
ation time of the battery of mobile terminals and reduces the algorithm is relatively difficult to implement in a very limited
processing power handset.
The QoS constraint usually is not sharp; rather, it has some
Manuscript received February 23, 2004; revised October 29, 2005 margin in which the QoS remains acceptable. To be more
and January 31, 2006. The review of this paper was coordinated by
Prof. C.-J. Chang.
specific, the QoS is accepted if it falls within a set that is lower
M. Elmusrati was with the Department of Electrical and Engineering, limited by the minimum accepted QoS and upper limited by
Garyounis University, Benghazi, Libya. He is now with the Department of the supremum QoS. The minimum accepted QoS is the lowest
Computer Science, University of Vassa, 65101 Vaasa, Finland, and also with
the Control Engineering Laboratory, Helsinki University of Technology, 02015
acceptable connection quality; any level below that is unaccept-
Espoo, Finland. able. The supremum QoS is defined as the upper bound for the
R. Jäntti is with the Department of Computer Science, University of Vaasa, QoS. The preferred power control is that one that can achieve
65101 Vaasa, Finland, and also with the Control Engineering Laboratory,
Helsinki University of Technology, 02015 Espoo, Finland. an accepted QoS level (i.e., a level that is between minimum and
H. N. Koivo is with the Control Engineering Laboratory, Helsinki University supremum QoS levels) very fast at low power consumption. Our
of Technology, 02015 Espoo, Finland. proposed power control algorithm fast achieves an accepted
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. QoS level at very low power consumption. In this paper, we
Digital Object Identifier 10.1109/TVT.2006.889565 will focus on the uplink power control case. We assume that the

0018-9545/$25.00 © 2007 IEEE


780 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

SINR of each user can be estimated accurately and is available


for the power controller. A method to estimate the SINR at the
handset is proposed, e.g., in [11]. This estimation of the SINR
is based only on 1-bit feedback channel.
The proposed algorithm in this paper aims to achieve two
objectives by applying the multiobjective (MO) optimization
method. The first objective is to minimize the transmitted
power, and the second objective is to achieve an accepted
QoS level in terms of SINR. It is clear that both objectives are
generally conflicted. The typical way to deal with conflicted
objectives is to use MO optimization. The same objective has
been considered, e.g., in [12]. The difference between the MO
optimization problem suggested here and the joint optimization
considered, e.g., in [12], is the number of free parameters. In
joint optimization, one common objective is optimized with
respect to several resources such as transmit power and antenna
beam vector weights. In some cases, the original optimization
problem can be divided into subproblems that are then solved
Fig. 1. Convergence region for the MODPC algorithm.
with respect to one variable at a time. In MO optimization, the
situation is different; several possibly conflicting objectives are
nected. We are usually interested in one solution of the Pareto
to be optimized with respect to only a single resource or, in
optimal set. This solution is selected by a decision maker. There
our case, transmit power. Thus, although the objectives of this
are different techniques to solve the MO optimization problems.
paper coincide with the subobjectives of [12], the problems are
We adopted the method of “Weighted Metrics.” If the desired
quite different.
solution of the objectives is known in advance (for example, the
This paper is organized as follows: The derived MO power
target SINR), then problem (1) can be rewritten as follows:
control algorithm is presented in Section II. The convergence
analysis is discussed in Section III. Section IV shows numerical m  p1
   p
results. Finally, conclusions are given in Section V. Appendix I min λi fi (x) − zi∗ 
contains the derivation of the multiobjective distributed power i=1
control (MODPC) algorithm, Appendix II recapitulates the subject to x ∈ S (2)
necessary tools for convergence analysis, and Appendixes III
and IV contain the proofs of stated propositions. where 1 ≤ p ≤ ∞, zi∗ is the desired solution of the objective i,
and the tradeoff factors
II. MODPC A LGORITHM m

The MO optimization approach is a technique that optimizes λi ≥ 0, ∀i = 1, . . . , m and λi = 1.
different and usually conflicted objectives. In the MO opti- i=1

mization problem, we have a set of objective functions; each Other methods to handle the MO optimization problems can be
is a function of a decision vector [13]. It can be represented found in [13] and [14].
mathematically as follows: In the next formulation, the power control problem has been
  represented by two objectives given as follows: 1) minimizing
min f1 (x), f2 (x), . . . , fm (x)
the transmitted power and 2) keeping the SINR as close as pos-
subject to x ∈ S (1) sible to the supremum SINR. After convergence, the achieved
SINR is the target SINR if it falls inside the target set, i.e., larger
where we have m(≥ 2) objective functions fi : n → , x is than the minimum accepted SINR and less than or equal to the
the decision vector belonging to a (nonempty) feasible region supremum SINR. Fig. 1 explains this idea for two users’ case.
(set) S, which is a subset of the decision variable space n . The continuous lines correspond to the power values where the
The function min means that we minimize all the objectives SINR is equal to the minimum allowed level, while the dashed
simultaneously. The objectives are usually conflicted and pos- lines correspond to power values that achieve the supremum
sibly incommensurable. Thus, there is no single solution that SINR level. The shadowed area contains the accepted power
minimizes all objectives simultaneously. The MO optimization vectors at a certain time. The preceding objectives could be
provides different optimal solutions at different senses. They interpreted mathematically for user i by the following error
are called nondominated or Pareto optimal solutions. A decision function:
vector x∗ ∈ S is Pareto optimal if there does not exist another  
decision vector x ∈ S such that fi (x) ≤ fi (x∗ ) for all i = ei (t) = λi,1 |Pi (t) − Pmin | + λi,2 Γi (t) − Γsup
i
 (3)
1, 2, . . . , m and fj (x) < fj (x∗ ) for at least one index j [13],
[14]. The “Pareto optimal set” is the set of all possible Pareto where 0 ≤ λi,1 ≤ 1 and λi,2 = 1 − λi,1 are the tradeoff factors,
optimal solutions. This set can be nonconvex and noncon- Γsup
i is the supremum SINR, Pmin is the minimum transmitted
ELMUSRATI et al.: MODPC ALGORITHM FOR CDMA WIRELESS COMMUNICATION SYSTEMS 781

taps. By solving the optimization problem (5) for one tap (see
Appendix I for detailed derivations), we obtain the following
power control algorithm:

λi,1 Pmin + λi,2 Γsup


i
Pi (t) = Pi (t − 1)
λi,1 Pi (t − 1) + λi,2 Γi (t − 1)
i = 1, . . . , Q; t = 1, 2, . . . . (7)

In practice, the transmitter power is bounded above by some


maximum value Pmax determined by the properties of the
Fig. 2. Autoregressive model of power control. power amplifier of the handset. It is interesting to observe that if
λi,1 > 0, then the transmitted power of the MODPC algorithm
power that can be handled by the mobile station, and Γi (t) is (7) is naturally upper bounded such as
the SINR of user i at time slot t, which is given by
λi,2 sup
Pi (t)Gii (t) Pi (t) < Pmin + Γ = Psup , i = 1, . . . , Q. (8)
Γi (t) = , i = 1, . . . , Q (4) λi,1 i

Q
Pj (t)Gij (t) + δ(t) Thus, we can conclude that Pmax ≥ Psup , i.e., the power
j=1
j=i constraint cannot be violated, as long as
where Gij (t) is the channel gain between user j and receiver λi,2 Pmax − Pmin
i, Q is the number of active users at time slot t, and δ(t) is ≤ (9)
λi,1 Γsup
i
the average additive noise power at time slot t. Without loss
of generality, it has been assumed that user i is assigned to Since λi,2 = 1 − λi,1 , the preceding equation can be written in
base station i. We assume that the transmitted data of user i terms of λi,1 as
is correctly decoded at the receiver if the received SINR is
greater than some minimum value Γi,min . An interpretation for Γsup
i
λi,1 ≥ . (10)
the supremum SINR is that it is the minimum tolerable SINR Γsup
i + Pmax − Pmin
determined by the utilized coding and modulation scheme +
In the general case, we need to augment the MODPC
SINR margin to cope with fast fading.
algorithm to take the maximum power constraint into account
The obtained power control from minimizing the error
as follows:
function (3) is able to track the supremum SINR and with a


penalty of using a large power. The error function (3) can be λi,1 Pmin + λi,2 Γsup
i
Pi (t) = min Pmax , Pi (t − 1)
extended to include other objectives such as maximizing the λi,1 Pi (t − 1) + λi,2 Γi (t − 1)
throughput [16].
i = 1, . . . , Q; t = 1, 2, . . . . (11)
To generalize the optimization over all users for time window
of N slots, we define the following optimization problem: Some interesting observations of the MODPC algorithm (7)
Find the power vector P = [P1 , P2 , . . . , PQ ] (P denotes the are that the MODPC algorithm is a nonlinear algorithm. Its
transpose of vector P) that minimizes of the following cost behavior is determined by the values of the tradeoff factors
function: λi . It requires no extra measurements other than those used
 Q N in conventional algorithms [1]–[10]. At λi,1 = 0 and λi,2 =

N −t 2
J(P) = γ ei (t) , t = 1, . . . , N (5) 1 − λi,1 = 1, the first objective in (3) is relaxed, and only the
i=1 t=1 second objective is considered. It is interesting to observe that
in this case, the MODPC algorithm becomes the very well
where γ is an adaptation factor (also called forgetting factor)
known DPC algorithm [1], which is
and N is the number of time slots in the time window. If
the forgetting factor is not zero, then it means that the next Γsup
i
transmitted power will depend on all the accumulated past Pi (t) = Pi (t − 1). (12)
Γi (t − 1)
history of the transmitted powers.
Suppose further that power Pi (t) is described by an autore- Hence, we note that the DPC algorithm is a special case of the
gressive model, as shown in Fig. 2 [17]. The transmitted power MODPC algorithm. At the other extreme case where λi,1 = 1
can be given as and λi,2 = 0, the handset transmits at the minimum power,
n
regardless of the SINR situation (i.e., no power control). The
 proper values of tradeoff factors, which could be adaptive, can
Pi (t) = wi (k)Pi (t − k) = wi Xi (t) (6)
greatly enhance the performance of the algorithm, depending
k=1
on the scenario.
where wi = [wi (1), . . . , wi (n)] is the power adaptation weight In terms of MO optimization, the proper values of the trade-
vector, Xi (t) = [Pi (t − 1), . . . , Pi (t − n)] contains known off factors for a certain network situation are selected by the de-
old values of the transmitted power, and n is the number of cision maker, who determines the optimum point from a Pareto
782 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

optimal set. A simple but efficient decision rule can be derived Definition 1: The network configuration is said to be “feasi-
from the steady-state behavior of the proposed algorithm: At ble” if there exists a power vector P∗ ∈ {P : 0 ≤ P ≤ Pmax }
steady state (i.e., Pi (t + 1) = Pi (t) = Piss ), (7) results in the such that Γi ≥ Γi,min , i = 1, . . . , Q.
steady-state SINR of user i(Γss i ), which is given by Definition 2: The maximum achievable SINR µ∗ is de-
fined as
sup λi,1
Γss
i = Γi − (P ss − Pmin ) (13) µ∗ = max min Γi .
λi,2 i P≥0 i=1,...,Q
(16)

when λi,2 > 0. The case in which λi,2 = 0 denotes fixed The convergence analysis of the MODPC algorithm is dis-
transmission powers, and its steady-state analysis is omitted cussed in the next propositions.
as trivial. Proposition 1: For static channels and for any P (0) > 0, the
One of the key features of the MODPC algorithm can be MODPC algorithm (7) converges to a unique fixed point. This
observed in the steady-state solution given in (13). It is clear fixed point is determined by the values of the tradeoff factors
that the steady-state SINR equals the supremum SINR when the explained in the previous section.
steady-state power equals the minimum power. The penalty of The proof is given in Appendix III.
using any excessive power is a reduction in steady-state SINR. Proposition 2: For a noiseless feasible system, static chan-
This could be interpreted as dynamically changing SINR mar- nels with P (0) > 0, and proper values of tradeoff factors, the
gin. If the system is lightly loaded, a large margin could be used, MODPC algorithm converges to the SINR balance, i.e.,
and when the load increases, the SINR margin is decreased to
include more users. This behavior of MODPC resembles the lim P(t) = P∗
t→∞
gradual removal concept suggested by Yates et al. [22].
The decision maker should choose the tradeoff factors λi,1 lim Γi (t) = µ∗ ∀i = 1, 2, . . . , Q (17)
t→∞
and λi,2 so that all users can achieve at least the minimum
acceptable SINR level. In the worst-case situation, the steady- where µ∗ is the maximum achievable SINR, and P∗ is the cor-
state power is the maximum allowed power Pmax , which is responding eigenvector of the normalized channel gain matrix
determined by the power amplifier of the handset. A robust F = [fij ], fij = Gij /Gii for i = j and fii = 0∀i [2].
way to find the tradeoff factors values is to consider the worst See Appendix IV for the proof.
case Piss = Pmax . It should be noted that proposing the tradeoff
factors by this method is not necessarily optimum but gives an IV. MODPC A LGORITHM W ITH Q UANTIZED SINR
initial guess. From (13) and λi,1 + λi,2 = 1, the values of trade-
off factors can now be solved to yield (assuming Pmin = 0) Perfect estimation of the mobile’s SINR at the base station
is assumed in the MODPC algorithm. In the existing cellular
Pmax Γsup
i −Γi,min
communication system, only a quantized version of the SINR
λi,2 = , λi,1 = .
Pmax +Γsup
i −Γi,min Pmax +Γsup
i −Γi,min
is available at the mobile station. In the worst case, only a 1-bit
(14) quantizer is used to command the mobile to step up/down its
transmitted power. An estimation method for the SINR based
Let us consider the DS-CDMA case, where the relationship on the 1-bit command is presented in [11]. In this section, we
between the target SINR and the target Eb /I0 is given (in join this estimation method with the MODPC algorithm to see
decibel scale) by the quantization error effects on the algorithm. It can be shown
from [11] than the estimated SINR at the mobile station is
Γ̂sup
i = ρsup
i −Ω (15) given by
 t−1
where ρsupi is the supremum bit-energy-to-noise-power- sup

spectral-density ratio in decibels and Ω is the processing gain Γ̃i (t) = Γ − δi νi (t − k)ci (t, k) + νi (t)
in decibels. As typical values in voice applications, ρsupi = k=1

6 dB, Ω = 24 dB, and then Γ̂sup i = −18 dB. If the margin t = 0, 1, . . . ; i = 1, . . . , Q (18)
is 3 dB and Pmax = 1, then from (14), λi,1 ∼ = 0.01, and νi (t) = sign (Γsup − Γi (t)) EPC,i (t) (19)
λi,2 ∼
= 0.99. As stated before, the system designer decides i
k−1
the optimum values of the tradeoff factors based on targeted 1
objectives. The tradeoff factors can be adapted to achieve ci (t, k) = [1 + νi (t − n)νi (t − n − 1)] (20)
2k n=0
certain time-varying objectives. The convergence properties of
the MODPC algorithm are discussed in Section III. where EPC,i (t) is 1 with probability PPCE,i (t) and −1 with
probability 1 − PPCE,i (t). PPCE,i (t) is the probability of bit er-
ror in power control command transmission at time t. The same
III. C ONVERGENCE A NALYSIS OF THE
MODPC algorithm can be used with estimated SINR given in
MODPC A LGORITHM
(18) as with perfect SINR. We call the new algorithm as MO
Before proceeding to the convergence analysis, we consider totally DPC (MOTDPC) algorithm. The convergence property
the following definitions: of the MOTDPC algorithm is discussed in the next corollary.
ELMUSRATI et al.: MODPC ALGORITHM FOR CDMA WIRELESS COMMUNICATION SYSTEMS 783

TABLE I
AVERAGE TRANSMITTED POWER IN DECIBEL METERS

Corollary 1: For static channels and for any P (0) > 0, the
MOTDPC algorithm converges to a unique fixed point.
The proof is given in Appendix V.

V. N UMERICAL R ESULTS
In this section, we compare the convergence properties
MODPC with DPC [1], Foschini’s and Miljanic’s algorithm
(FMA) [6], and the second-order power control (SOPC) algo-
rithm [10]. We compare the algorithms in terms of convergence Fig. 3. Power path trajectories of two users using DPC and MODPC.
speed, outage probability, and power consumption. The com-
parisons are carried out for both static and dynamic channels.
In each channel scenario, we assumed 120 users uniformly
distributed in an area of 4 km2 with four base stations. The base
stations are regularly distributed, i.e., at (0.5,0.5), (0.5,1.5),
(1.5,0.5), and (1.5,1.5) km. Perfect handover is assumed, where
each user is assigned to the base station with the best instanta-
neous link quality. The maximum transmit power is normalized
to 30 dBm. An additive white Gaussian noise is assumed with
zero mean and an average power of −90 dBm. The radio link
gain is modeled as a product of three variables: 1) the large-
scale propagation loss that depends on the distance between the
mobile terminal and the base station; 2) lognormal shadowing
with a mean of 0 dB and a standard deviation of 5 dB; and
3) Rayleigh-distributed multipath fading generated by a cor-
related process [20]. In the dynamic scenario, the speed of
the users is 30 km/h. The carrier frequency is 2 GHz. In the
snapshot analysis, the channel gain of each user is set at a fixed
value during the simulation.
The tradeoff factors for the MODPC algorithm has been set Fig. 4. Power convergence differences in DPC and MODPC.
for all users to λi,1 = 0.01 and λi,2 = 0.99. In the dynamic
scenario, the same values of the tradeoff factors are used. faster toward the fixed point than that of the DPC algorithm.
The algorithms are compared in terms of the error norm and It is clear that the MODPC algorithm converges faster than
the outage percentage. The error is defined as the difference the DPC algorithm to reach the feasible region. The MODPC
between the actual transmitted power vector generated by the algorithm requires less number of iterations than other conven-
power control algorithm and the optimum power vector. The tional algorithms to converge to the adequate solution. It can be
optimum power vector is the solution of the centralized power explained, as indicated by (3) and (13), that there is a penalty in
control algorithm. The outage is the percentage of users that using the power. Fig. 3 illustrates this property of the MODPC
do not achieve the minimum tolerable SINR level. We follow algorithm. The convergence analysis in [18] showed that the
the approach taken, e.g., in [10], and consider outage in each Hessian matrix associated with the MODPC scheme had a
time slot instead of just observing the steady-state behavior. smaller spectral radius than the corresponding matrix for DPC.
The supremum SINR is set to −18 dB for all users. Note that This result indicates that MODPC converges asymptotically
SINR here denotes the SINR before dispreading. The minimum faster than DPC. This observation is also supported by our
allowed SINR has been set to 3 dB lower than the supremum simulation results here. It is, however, notable that the two
SINR. For comparison purposes, Table I lists the average algorithms are not converging to the same point. While DPC
transmitted power of all users in decibel meters of each power is converging toward Γsupi , MODPC is aiming to some lower
control algorithm for static and dynamic scenarios. SINR between Γi,min and Γsup i . This difference is clarified in
Fig. 3 shows the convergence behavior of the MODPC and Fig. 4. Fig. 4 shows the initial powers for both algorithms at
DPC algorithms for two users. The figure depicts that the power point (1); then, MODPC converges to point (2), while DPC
path trajectory of the MODPC algorithm is converging much converges to point (3), which achieves Γsupi .
784 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Fig. 5. Comparison between different power control algorithms in terms of Fig. 7. Comparison between different power control algorithms in terms of
error and outage for the static-channel scenario. error and outage for the dynamic-channel scenario.

Fig. 6. Achieved SINR (in decibels) for the best and worst users.
Fig. 8. Transmitted power versus time slots of different power control
algorithms.
Fig. 5 shows comparisons of the error norm of the average
outage percentage for the 120 users using MODPC, DPC, are quite similar in their trend as in the first scenario, as shown
FMA, and SOPC algorithms for the static-channel case. It is in Fig. 7. The MODPC still outperforms the other algorithms
clear in this scenario that the MODPC algorithm outperforms in terms of optimum power tracking capability in terms of
other algorithms in terms of power convergence speed and error norm and QoS tracing capability in terms of outage. The
outage convergence speed. Moreover, the average consumption average power consumption of MODPC is less than that of
power of the MODPC algorithm is considerably less than those other algorithms, as shown in Table I. Since this scenario is
of other simulated algorithms, as indicated by Table I. For for a dynamic channel, it is worth investigating the fluctuations
example, the average transmitted power of MODPC is lower of the transmitted power of one user and how the algorithms
than that of DPC by less than 4 dB. That means less than half of can track the optimum power. Fig. 8 depicts a sample of the
the power (on the average) is needed for the MODPC algorithm transmitted power of an arbitrarily selected user using optimal
to achieve the target QoS. Comparing the MODPC algorithm (centralized), DPC, FMA, and MODPC algorithms. The result
with other algorithms, it is clear that less power is consumed SOPC algorithm is omitted to enhance the visibility of the
by the MODPC algorithm. Fig. 6 shows the achieved SINR for graphs because it contains high overshoots (the transmitted
the best and worst users using the MODPC algorithm. It is clear power falls close to zero at iteration 19). Fig. 8 indicates that
that the aim is to be in the convergence region between Γi,min the MODPC algorithm can track the variations of the channel
and Γsup
i . better than other simulated algorithms. This result indicates
Dynamical channels with fast and slow fading are assumed that the MODPC algorithm can react faster to changes in
in the second scenario. In this scenario, the comparison results interference than the other algorithms. Finally, the convergence
ELMUSRATI et al.: MODPC ALGORITHM FOR CDMA WIRELESS COMMUNICATION SYSTEMS 785

The error function can be modified such as

ei (t) = λi,1 (Pi (t) − Pmin ) + λ̂i,2 (t) (Γi (t) − Γsup )
t = 0, 1, . . . (21)

where λ̂i,2 (t) = sign(Γi (t) − Γsup )λi,2 ; the sign function is
defined as

+1, x ≥ 0
sign(x) = (22)
−1, x < 0.

Suppose further that power Pi (t) is described by a linear


autoregressive model, as shown in Fig. 2 [17]. The transmitted
power is
n

Pi (t) = wi (k)Pi (t − k) = w i Xi (t), t = 0, 1, . . .
k=1
(23)
Fig. 9. Convergence behavior in terms of error norm and outage for MOTDPC
with different power control command error levels. where
 
wi = wi (1) · · · wi (n) , Xi (t) = Pi (t − 1) · · · Pi (t − n) .
behavior of the MOTDPC with estimated SINR and different (24)
power control command errors is shown in Fig. 9. The figure
Observe that Xi (t) contains known measured values of the
indicates that MOTDPC achieves close to ideal performance
transmitted power.
even if 1-bit feedback is used instead of perfect SINR values.
Substitute (23) into (21) and (4). Then, error ei (t) can be
The performance remains tolerable even with up to 20% power
written as
control bit error rate.
 
ei (t) = λi,1 w i Xi (t) − Pmin
 
VI. C ONCLUSION Gii (t)w i Xi (t)
+λ̂i,2 (t) − Γsup . (25)
Ii (t)
In this paper, we suggested a novel DPC algorithm that is
based on the MO optimization method. We considered two Denote
objectives: 1) QoS in terms of SINR and 2) power consumption.  
It should be noted, however, that the suggested power control Gii (t)
αt := λi,1 + λ̂i,2 (t) . (26)
scheme could be easily extended to take other objectives, such Ii (t)
as throughput, into account. The algorithm is parameterized so
Using this in (25), ei (t) becomes
that the designer can make a tradeoff between QoS and power
consumption. A simple guideline is provided for the choice of
ei (t) = αt w i Xi (t) − λi,1 Pmin − λ̂i,2 (t)Γsup . (27)
these parameters such that the given QoS constraints can be
met. In a snapshot case, the algorithm was proven to converge, From (5) and (25), a necessary condition for the minimum is
starting from an arbitrary initial value. The simulation results given as follows for all i = 1, . . . , Q:
indicated that the suggested MODPC algorithm outperformed
N

the DPC, FMA, and SOPC algorithms in convergence speed ∂ei (t)
and was able to track the QoS target better in the case of a 2 γ N −t ei (t) = 0. (28)
t=1
∂w
fading channel. The proposed algorithm has also been modified
to adopt estimated SINR (based on the up/down power control From (27)
command) instead of the ideal values. The modified method
is called the MOTDPC algorithm. It has been proven that in ∂ei (t)
= αt X i (t). (29)
the snapshot case, the power vector sequence generated by the ∂w
MOTDPC algorithm converges, starting from an arbitrary ini-
Substituting (27) and (29) into (28), we obtain
tial value. The simulations show that the MOTDPC converges
even in case of power control command errors. N
 ∂ei (t)
γ N −t ei (t)
t=1
∂w
A PPENDIX I N

D ERIVATION OF MODPC A LGORITHM = γ N −t (αt w i Xi (t) − λi,1 Pmin − λi,2 Γsup ) αt X i (t)
The absolute function of the first term in (3) is not needed be- t=1

cause the transmitted power cannot be less than the minimum. = 0. (30)
786 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Solving for wi To overcome these problems, only the positive values of λ̂i,2 (t)
 N
 are considered, i.e., λ̂i,2 (t) = λi,2 . This simplification has con-
  
γ N −t αt2 Xi (t)X i (t) wi siderably reduced the complexity of the MODPC algorithm
t=1 with the price of slight degradation in the convergence speed.
N The MODPC algorithm becomes

= γ N −t αt (λi,1 Pmin + λi,2 Γsup ) Xi (t) (31) λi,1 Pmin + λi,2 Γsup
t=1 Pi (t) = Pi (t − 1)
λi,1 Pi (t − 1) + λi,2 Γi (t − 1)
or i = 1, . . . , Q; t = 1, 2, . . . . (40)
wi (N ) = R−1
xx (N )Rx (N ), i = 1, . . . , Q (32)
A PPENDIX II
where I MPORTANT T HEORIES FOR C ONVERGENCE A NALYSIS
N
 In this Appendix, we introduce the definition of the standard
Rxx (N ) := γ N −t αt2 Xi (t)X i (t) (33)
power control algorithm as well as the two theorems related
t=1
to it, as given in [19]. The transmitted power of user i can be
N
   described mathematically as
Rx (N ) := γ N −t αt λi,1 Pmin + λi,2 Γsup Xi (t). (34)
t=1
Pi (t) = Ψi (P(t − 1), Γi (t)) (41)
Equations (32)–(34) are well known from least squares tech-
where Ψ(•) is the interference function, and Γi (t) is the SINR
niques. Equation (32) can be solved using the Recursive Least
of user i at time slot t.
Square method. To avoid the matrix inversion, Rxx (N ) may be
The interference function Ψ(•) is called “standard” when
computed recursively as
the following properties are satisfied for all nonnegative power
2
Rxx (N ) = γRxx (N − 1) + αN Xi (N )X i (N ). (35) vector P [19]:
• positivity, i.e., Ψi (P(t), Γi (t)) > 0;
Since the inverse of Rxx (N ) is needed, we can use the matrix • monotonicity, i.e., if P ≥ P , then Ψi (P(t), Γi (t)) ≥
inverse lemma to obtain [21] Ψi (P i (t), Γi (t)) > 0;
 • scalability, i.e., for all α > 1, αΨi (P(t), Γi (t)) >
−1 1 Ψi (αP(t), Γi (t)).
Rxx (N ) = R−1
xx (N − 1)
γ
A power control algorithm of form (38) is called standard if
R−1 2 −1 the utilized mapping Ψ(•) is the standard interference function.
xx (N − 1)αN Xi (N )X i (N )Rxx (N − 1)
− 2 X (N )R−1 (N − 1)X (N ) . (36) Standard power control algorithms have the two following
γ + αN i xx i
properties.
In addition, Rx (N ) can be computed recursively as
  Theorem 1 [19, Th. 1]
Rx (N ) = γRx (N − 1) + αN λi,1 Pmin + λi,2 Γsup Xi (N ).
(37) For a standard power control algorithm, solution of (41) has
a unique fixed point.
The power control algorithm is easy to implement and is
also computationally light to be applicable for existing wireless
Theorem 2 [19, Th. 2]
communication systems. Next, the simplest case, where n = 1
in (23) and γ = 0, is considered. Solving (32), we obtain If Ψ(•) is feasible, then for any initial power vector P o , the
standard power control algorithm converges to a unique fixed
λi,1 Pmin + λ̂i,2 (t)Γsup point P∗ .
wi (t) =
λi,1 Pi (t − 1) + λ̂i,2 (t)Γi (t − 1)
i = 1, . . . , Q; t = 1, 2, . . . . (38) A PPENDIX III
P ROOF OF P ROPOSITION 1
From (23), the transmitted power of user i at time slot t is
given by We will prove that the MODPC algorithm is a standard
power control algorithm discussed in Appendix II; then, by
λi,1 Pmin + λ̂i,2 (t)Γsup Theorems 1 and 2, the MODPC algorithm converges to a unique
Pi (t) = Pi (t − 1)
λi,1 Pi (t − 1) + λ̂i,2 (t)Γi (t − 1) fixed point. The interference function Ψi (P) of the MODPC
algorithm is given by
i = 1, . . . , Q; t = 1, 2, . . . . (39)
λi,1 Pmin + λi,2 Γsup
i
Due to the sharp changes in the λ̂i,2 (t) sign, the transmitted Ψi (P) = Pi (t − 1)
power in (39) may take negative values as well as very large λi,1 Pi (t − 1) + λi,2 Γi (t − 1)
power values, which are not part of the power feasible subspace. i = 1, . . . , Q (42)
ELMUSRATI et al.: MODPC ALGORITHM FOR CDMA WIRELESS COMMUNICATION SYSTEMS 787

and it can be represented as From (44), one can say that Ii (αP) > Ii (P), α > 1; then

Ii (P)a Ii (αP)a
Ψi (P) = (43) αΨ(P) > = Ψ(αP). (55)
λi,1 Ii (P) + λi,2 λ1 Ii (αP) + λ2

where Ii (P) is the normalized interference for user i such as Then, the “scalability” condition has been proven.
From (47), (52), and (55), one can say that the MODPC
Q
 Gij (t)Pj (t) δ(t) algorithm is a standard interference function. This means that
Ii (P) = + (44)
j=1
Gii (t) Gii (t) the MODPC algorithm converges to a unique fixed point. This
j=i concludes the proof.
and a = λi,1 Pmin + λi,2 Γsupi > 0. Note that for convenience,
the time symbol t has been omitted from the interference A PPENDIX IV
function as well as the normalized interference [e.g., use Ii (P) P ROOF OF P ROPOSITION 2
instead of Ii (P(t))]. In (7), let a = λi,1 Pmin +λi,2 Γsup
i . Since Γi (t) = Pi (t)/Ii (t),
From (44), it is clear that for any where Ii (t) is the normalized interference (44) with δ(t) = 0,
P ≥ 0, Ii (P) ≥ 0. (45) (7) can be rewritten as
a
In addition, if Pi (t + 1) = 
λi,1 + λi,2 Ii (t)
P ≥ Z ⇒ Ii (P) ≥ Ii (Z). (46) a
= Ii (t) = βi (t)Ii (t). (56)
λi,1 Ii (t) + λi,2
Since 0 ≤ λi,1 ≤ 1; λi,2 = 1 − λi,1 and from (43) and (45),
then for any Now, we have obtained the same form as the DPC
algorithm [1]. The proof of convergence of this algorithm is
P ≥ 0 ⇒ Ψ(P) ≥ 0. (47) presented in [1]. Since the transmitted power of each user is
bounded by (8), the interference Ii (t) will be upper bounded
Thus, the “positivity” condition has been proven.
for all t. For convergence, tradeoff factors should be selected to
The monotonicity condition will be proven by contradiction.
satisfy the following condition:
For any P ≥ Z, assume that Ψi (P) < Ψi (Z). Then, from (43)
t 

Ii (P)a Ii (Z)a a
< (48) lim (µ∗ )t <∞ (57)
λi,1 Ii (P) + λi,2 λi,1 Ii (Z) + λi,2 t→∞ λi,1 Ii (t) + λi,2
k=0

Ii (Z)a λi,1 Ii (P) + λi,2 where µ∗ is the spectral radius of the normalized channel gain
Ii (P)a < (49)
λi,1 Ii (Z) + λi,2 matrix [1], [2], [18]. This concludes the proof.
 
λi,1 Ii (Z) + λi,2 IIii(P)
(Z)

Ii (P)a < Ii (P)a . (50) A PPENDIX V


λi,1 Ii (Z) + λi,2 P ROOF OF C OROLLARY 1
However, from (46) Corollary 1 can be proven using the same method as the proof
  of Proposition 1. The MOTDPC algorithm is given by
(Z)
λi,1 Ii (Z) + λi,2 IIii(P)
0< <1 (51) λi,1 Pmin + λi,2 Γsup
i
λi,1 Ii (Z) + λi,2 Ψi (P) = Pi (t − 1)
λi,1 Pi (t − 1) + λi,2 Γ̃i (t − 1)
and (50) does not hold. Then, assumption (48) is not true. Then,
for any i = 1, . . . , Q (58)

P ≥ Z ⇒ Ψi (P) ≥ Ψi (Z). (52) where Γ̃i (t) is the estimated SINR at time slot t, which is given
by (18). The estimated SINR can be represented as
Thus, the “monotonicity” condition has been proven.
From (44), for any Γ̃i (t) = εi (t)Γi (t) (59)

α > 1 ⇒ αIi (P) ≥ Ii (αP). (53) where εi (t) ≥ 0 is the estimation error factor. By substituting
(59) in (58), we obtain
The equality is achieved if the additive noise is zero. From
(43), we get λi,1 Pmin + λi,2 Γsup
i
Ψi (P) = Pi (t − 1)
λi,1 Pi (t − 1) + λi,2 εi (t − 1)Γi (t − 1)
Ii (P)a Ii (αP)a
αΨ(P) = α ≥ . (54)
λi,1 Ii (P) + λi,2 λi,1 Ii (P) + λi,2 i = 1, . . . , Q. (60)
788 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

For a static environment, it can be assumed that the esti- [22] R. Yates, S. Gupta, C. Rose, and S. Sohn, “Soft dropping power
mation error factor is fixed and does not change with time. control,” in Proc. IEEE Veh. Technol. Conf., Phoenix, AZ, May 1997,
pp. 1694–1698.
Equation (43) can be modified in the MOTDPC case to

Ii (P)a
Ψi (P) = (61)
λi,1 Ii (P) + λ̃i,2
Mohammed Elmusrati (S’00–M’04) received the
B.Sc. (with honors) and M.Sc. (with high honors)
degrees in telecommunication engineering from the
where λ̃i,2 = εi λi,2 . The convergence conditions of the Department of Electrical and Electronic Engineer-
MODPC algorithm can be reproven for the MOTDPC algo- ing, Garyounis University, Benghazi, Libya, in 1991
rithm by repeating the same steps of Proposition 1. and 1995, respectively, and the Licentiate of Sci-
ence in technology (with distinction) and Doctor of
Science in Technology degrees in control engineer-
ing from Helsinki University of Technology (HUT),
Espoo, Finland, in 2002 and 2004, respectively.
R EFERENCES He was a Lecturer with the Department of Elec-
[1] S. Grandhi, R. Vijayan, and D. Goodman, “Distributed power control trical and Electronic Engineering, Garyounis University, during 1995–1999.
in cellular radio systems,” IEEE Trans. Commun., vol. 42, no. 2/3/4, He is currently a Senior Lecturer and Researcher with the Department of
pp. 226–228, Feb./Mar./Apr. 1994. Computer Science, University of Vaasa, Vaasa, Finland, and a Part-Time Senior
[2] J. Zander, “Distributed cochannel interference control in cellular radio Researcher with the Control Engineering Laboratory, HUT. He has also been a
systems,” IEEE Trans. Veh. Technol., vol. 41, no. 3, pp. 305–311, Visiting Researcher with the Radio Communication Systems Laboratory (now
Aug. 1992. Wireless@KTH), Royal Institute of Technology (KTH), Stockholm, Sweden.
[3] P. Ligdas and N. Farvardin, “Optimizing the transmit power for slow He has received several grants from Garyounis University, Nokia Foundation,
fading channels,” IEEE Trans. Inf. Theory, vol. 46, no. 2, pp. 265–276, and The ELLA JA Georg Ehrnrootin Foundation. His research interests include
Mar. 2000. resource management in wireless and optical communication networks, smart
[4] J. Chamberland and V. Veeravalli, “Decentralized dynamic power control antennas, ultrawideband wireless distributed actuator and sensor networks, and
for cellular CDMA systems,” IEEE Trans. Wireless Commun., vol. 2, data fusion.
no. 3, pp. 549–559, May 2003.
[5] T. Lee and J. Lin, “A fully distributed power control algorithm for
cellular mobile systems,” IEEE J. Sel. Areas Commun., vol. 14, no. 4,
pp. 692–697, May 1996.
[6] G. Foschini and Z. Miljanic, “A simple distributed autonomous power
control algorithm and its convergence,” IEEE Trans. Veh. Technol.,
Riku Jäntti (M’02) received the Ph.D. degree from the Control Laboratory,
vol. 42, no. 4, pp. 641–646, Nov. 1993.
Helsinki University of Technology, Espoo, Finland, in 2001.
[7] F. Gantmacher, The Theory of Matrices, vol. 2. New York: Chelsea, He is an Acting Professor and Docent (Adjunct Professor) of telecommuni-
1964.
cation engineering and the Vice Head of the Department of Computer Science,
[8] S. Grandhi and J. Zander, “Constrained power control in cellular radio
University of Vaasa, Vaasa, Finland. He is also a Docent (Adjunct Professor)
systems,” in Proc. IEEE VTC, 1994, vol. 2, pp. 824–828.
of control engineering with the Control Engineering Laboratory, Helsinki Uni-
[9] A. El-Osery and C. Abdallah, “Distributed power control in CDMA cel- versity of Technology (TKK), Espoo, Finland. From 1998 to 1999, he was with
lular systems,” IEEE Antennas Propag. Mag., vol. 42, no. 4, pp. 152–159,
the Radio Communication Systems Laboratory (now Wireless@KTH), Royal
Aug. 2000.
Institute of Technology (KTH), Stockholm, Sweden, as a Visiting Scholar.
[10] R. Jäntti and S. Kim, “Second-order power control with asymptoti-
He has also been a Visiting Professor with the Radio Resource Management
cally fast convergence,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, and Optimization Laboratory, Information and Communications University,
pp. 447–457, Mar. 2000.
Daejeon, Korea. His current research interests include various control and re-
[11] M. Elmusrati, M. Rintamäki, I. Hartimo, and H. Koivo, “Fully distributed
source management problems of wireless radio networks and their applications
power control algorithm with one bit signaling and nonlinear error esti-
to automation systems.
mation,” in Proc. IEEE VTC—Fall, Oct. 2003, vol. 2, pp. 727–731.
[12] A. Mercado and K. Liu, “Adaptive QoS for wireless multimedia networks
using power control and smart antennas,” IEEE Trans. Veh. Technol.,
vol. 51, no. 5, pp. 1223–1233, Sep. 2002.
[13] K. Miettinen, Nonlinear Multiobjective Optimization. Norwell, MA:
Kluwer, 1998.
[14] G. Liu, J. Yang, and J. Whidborne, Multiobjective Optimisation and Con- Heikki N. Koivo (S’67–M’71–SM’86) received
trol. Hertfordshire, U.K.: Research Studies, 2003. the B.S.E.E. degree from Purdue University, West
[15] J. Variendt, P. Laine, C. Lerouge, and X. Xu, “Mobile network Lafayette, IN, and the M.S. degree in electrical
evolution: A revolution on the move,” IEEE Commun. Mag., vol. 40, no. 4, engineering and Ph.D. degree in control sciences
pp. 104–111, Apr. 2002. from the University of Minnesota, Minneapolis.
[16] M. Elmusrati and H. Koivo, “Multi-objective distributed power and rate He is a Professor of control engineering with the
control for wireless communications,” in Proc. IEEE ICC, 2003, vol. 3, Control Engineering Laboratory, Helsinki Univer-
pp. 1838–1842. sity of Technology (HUT), Espoo, Finland. Before
[17] J. Candy, Signal Processing—The Modern Approach. New York: joining HUT in 1995, he served in various acad-
McGraw-Hill, 1987. emic positions at the University of Toronto, Toronto,
[18] M. Elmusrati, “Radio resource scheduling and smart antennas in cellu- ON, Canada, and Tampere University of Technology,
lar CDMA communication systems,” D.Sc. thesis, Control Eng. Lab., Tampere, Finland. He has been the Principal Investigator in more than 100
Helsinki Univ. Technol., Espoo, Finland. Rep. 142, 2004. research projects. He has authored more than 300 scientific publications. His
[19] R. Yates, “A framework for uplink power control in cellular radio research interests include the study of complex systems, adaptive and learning
systems,” IEEE J. Sel. Areas Commun., vol. 13, no. 7, pp. 1341–1347, control, mechatronics, microsystems, and wireless communication systems.
Sep. 1995. Dr. Koivo is a member of the Editorial Board of the Journal of Intelligent and
[20] F. Berggren and R. Jäntti, “Asymptotically fair transmission scheduling Fuzzy Systems, Intelligent Automation and Soft Computing, and the Journal of
over fading channels,” IEEE Trans. Wireless Commun., vol. 3, no. 1, Systems and Control Engineering. He was an Associate Editor of the IEEE
pp. 326–336, Jan. 2004. TRANSACTIONS ON ROBOTICS AND AUTOMATION and a member of the
[21] J. Proakis, Digital Communications, 3rd ed. New York: McGraw-Hill, Administrative Council of the IEEE Robotics and Automation Society. He is
1995. a Fellow of the Finnish Academy of Technology.

Das könnte Ihnen auch gefallen