SELF-TUNING ADAPTIVE ALGORITHMS IN THE POWER CONTROL OF

WCDMA SYSTEMS
Matti Rintamki1, Kai Zenger2 and Heikki Koivo3
1

Helsinki University of Technology, Signal Processing Laboratory, P.O.BOX 3000,

FIN-02015 HUT, Finland, [email protected]
2,3
Helsinki University of Technology, Control Engineering laboratory, P.O.BOX 5400,
FIN-02015 HUT, Finland, [email protected], [email protected]
ABSTRACT
Power control is an essential radio resource function in
WCDMA systems. It is needed to compensate the nearfar effect, i.e., a strong signal overriding a weaker one at
a receiver. In practical systems power control is handled
so that a receiver measures the signal-to-interference ratio (SIR) and compares it to a target value. Based on this
comparison, the receiver requests the transmitter to either decrease or increase its transmitter power by a fixed
amount, typically 1 dB. A more sophisticated approach
has been proposed in [0], where an adaptive self-tuning
controller (STC) with generalized minimum variance
(GMV) criterion was applied in the closed-loop power
control. Simulations indicated that the proposed method
outperformed the conventional bang-bang type power
control algorithm proposed in [1]. However, it is not
clear how to select the various polynomials included in
the design of the GMV controller to maximize the controller performance. In this paper we use system identification methods for modeling the WCDMA closed-loop
power control. The power control process is modeled using a convenient parameterized linear structure, which
can be used to tune the GMV controller performance in
the design phase. The model is identified using inputoutput data collected from a radio network simulator by
opening the power control loop of a randomly selected
user. The results give significant insight to the power
control process and are useful in the design of adaptive
power control algorithms.
1. INTRODUCTION
Transmitter power control (TPC) is one of the important
radio resource management functions that are shown to
increase the capacity of cellular communication systems
(see e.g. [2-4]). Power control is essential in codedivision multiple access (CDMA) or wideband-CDMA
(WCDMA) systems, where the users share simultaneously the same transmission medium and thus interfere
with one another. Due to channel variations, the levels of
several transmitters' signals might vary considerably at
the receiver if the transmitters use the same transmitter
power. This is called the near-far effect. By controlling
the transmitter powers, the aim is to provide sufficient
signal quality for all the users. This basically means that
the users must have sufficient signal-to-interference ratios (SIRs).

SIR target
(dB) +
Received SIR (dB)
Channel
variation
(dB)

Transmitter
power

TPC command
1 dB

error (dB)

OUT

IN
Return channel
error 1

Interference
(dB)

BASE
STATION

RADIO
CHANNEL

Loop kT
p
delay

Tp

+-

Step
size

MOBILE
UNIT

Integrator

Figure 1. CDMA uplink closed-loop power control

as presented in [1].
A conventional power control algorithm for practical
applications is shown in Figure 1 for uplink. This was
proposed in [1]. The SIR is measured at each receiver,
and compared to a target value, which is set by a quality
controller to achieve a sufficient signal quality. Based on
this measurement, a command is sent to the transmitter
to either increase or decrease its transmitter power by a
fixed amount. We refer to this contol scheme as the
Bang-Bang.
A more sophisticated power control algorithm was
proposed in [0]. The general idea there is to precede the
relay (or comparator) in the conventional algorithm with
an adaptive self-tuning controller with generalized
minimum variance (GMV) criterion. However, to tune
the controller to best performance, more insight is
needed of the properties of the controlled system.
In this paper, we apply system identification methods
to model the uplink closed-loop power control process in
WCDMA systems. The idea is to collect input and output data from a radio network simulator and use them to
find a linear model structure that sufficiently describes
the properties of the power control process. The insight
gained from this process can then be used for the design
of the GMV controller to improve its performance in the
power control algorithm proposed in [0].
The paper is organized as follows. In Section 2 the
system model is defined. Section 3 describes the identification methods and results. In Section 4 an adaptive
GMV controller is designed for the identified process
model. A brief introduction to adaptive control is also
given. In Section 5 the controller performance is simu-

lated in the radio network simulator. Finally, some conclusions are drawn in Section 6.

e(t )

C
A

2. SYSTEM MODEL

u (t )

q k B
A

2.1. Cellular Network Simulator and Data Collection

The data for the identification is collected from a radio
network system simulator, which is described here. We
consider a two-dimensional seven-cell hexagonal pattern, where the cell radius is 50 m. 200 users are uniformly distributed over the seven cells. In the beginning
of the simulation, the users are assigned velocities uniformly between 0 km/h and 30 km/h and a random direction of movement. These are not changed during simulation. Ideal handovers are assumed in the sense that each
user is connected to the base station with the least channel attenuation at all times. The radio link attenuation is
modeled as a product of three variables: the large scale
propagation loss that depends on the distance between
the transmitter and the receiver, log-normal shadowing
with a mean of 0 dB and standard deviation of 8 dB, and
motion-induced Rayleigh-distributed multipath fading
generated by Jakes' model [5]. The log-normal shadowing component has also a correlation model first proposed by Gudmundson [6]. The model is extended to
two dimensions as in [7].
In the beginning of the simulation, a randomly selected user, initially connected to the central cell, is selected for observation and its velocity is set to 5 km/h.
Each user except the observed user is power controlled
using the power control method in [2] as illustrated in
Figure 1. The users transmit at a constant data rate without voice activity.
To collect the input-output data for the system identification, the power control loop of the observed user is
opened at points "IN" and "OUT" in Figure 1. The point
"IN" is then connected to a pseudo-random binary signal
(PRBS) generator, and the resulting SIR is measured at
the base station (point "OUT" in Figure 1). In other
words, the selected mobile unit receives power control
commands from the PRBS generator and adjusts its
transmitter power according to those commands. The
generated PRBS sequence and the resulting SIR sequence (both in decibels) are then regarded as input and
output, respectively, of an unknown system to be identified. Bit errors in the power control commands were not
considered. The sampling rate of the simulator matches
the sampling rate of the power control, which is 1.5 kHz
for WCDMA.
2.2. Model of the Closed Loop Power Control
We attempt to model the closed-loop power control with
a linear parameterized model of the following form:

y (t )

Figure 2. ARX model of the closed-loop power

control.

Ay (t ) = q k Bu (t ) + Ce(t ) ,

(1)

where y (t ) is the output of the model at time t (in our

case, the SIR at the receiver), u (t ) is the input of the
model at time t (in our case, the power control command), {e(t )} is a sequence of white Gaussian noise
samples with variance e2 , q 1 is the backward shift
operator defined by q 1 x(t ) = x(t 1) for an arbitrary
discrete-time signal x(t ) , k is the delay of the system,
and A , B and C are polynomials in q , given by
A = 1 + a1 q 1 + K + a n q n ,

(2)

B = b0 + b1 q 1 + K + bn q n ,

(3)

C = 1 + c1 q 1 + K + c n q n .

(4)

If n c = 0 , this model is referred to as an autoregressive model with an exogeneous variable, denoted by

ARX(na, nb, k). Otherwise it is referred to as an autoregressive moving average model with an exogeneous
variable, denoted by ARMAX(na, nb, nc, k) [8]. The
model is illustrated in Figure 2. We consider only the
ARX model in this paper.
3. MODEL IDENTIFICATION

The input-output data collected from the radio network

simulator according to Section 2.1. is used to identify the
model parameters, i.e., the coefficients of the polynomials A and B . Since the system is time-varying, we use
the Recursive Least Squares (RLS) algorithm with exponential forgetting to take this variation into the model.
Equation (1) with C = 1 can be cast in the form
y (t ) = T x(t ) + e(t ) ,
where

(5)

Variance of RLS residual signal versus model order

0.36

1.5

0.34

b0

1
a2

0.32
0.5
0.3
0

0.28
0.26

-0.5

0.24

b1

-1

0.22
a1

-1.5
0.2
-2

0.18
0.16

(1,1)

(2,1)

(2,2) (3,1) (3,2) (3,3) (4,1) (4,2)

Model polynomial orders (na ,nb )

(4,3)

-2.5

(4,4)

Figure 3. Variance of the residual signal (t ) .

= a1 K a n b 0 K b n
a

0.2

0.4

0.6

0.8
1
1.2
time (seconds)

1.4

1.6

1.8

Figure 4. Identified model parameters.

(6)
1

x (t ) =

[ y(t 1) K y(t n a )

u (t k ) K u (t k nb )]

0.8

(7)

0.6

Equation (5) is the familiar linear regression form,

which can be used for recursive estimation of the parameter vector . The RLS algorithm equations are

(t ) = y (t ) x T (t 1) (t 1) ,

0.2

0.4

0.6

0.8

1
(a)

1.2

1.4

1.6

1.8

0.2

0.4

0.6

0.8

1
(b)

1.2

1.4

1.6

1.8

0.8

(8)

0.6

(t ) = (t 1) + L(t ) (t ) ,
L (t ) =

(9)

P(t 1)x(t 1)

f + x T (t 1)P(t 1)x(t 1)

P (t ) = I L(t )x (t 1) P(t 1)

(11)

where P is the inverse of the covariance matrix of the parameter estimates, L is the Kalman gain, f is the forgetting factor and (k ) is a vector containing the parameter estimates, i.e.,

T
(t ) = a1 (t ) K a n (t ) b0 (t ) K bn (t ) .
a

Figure 5. Absolute values of the zeros (a) and poles

(b) of the identified model.

(10)

(12)

3.1. Selection of Model Order and delay

To select the model order, i.e., the orders of the model
polynomials n a and n b , we ran the RLS algorithm with
forgetting factor 0.99 with various values for n a and n b .
Figure 3 shows the variance of the residual signal (t )
in various cases. The "knee-point" in the figure occurs
with the configuration (n a , n b ) = (2,1) , and there is prac-

tically no advantage in increasing the model order beyond this point.

The delay is already defined in the simulator, and it is
assumed to be one sampling period (1/1500 seconds).
Thus our model structure will be an ARX(2,1,1)
model.
3.2. Identification Results
Figure 4 shows the parameters that resulted from the
RLS-identification of the model of the selected structure.
The coefficients of the model were updated every 30th
sample in order to catch the time-variation of the system
and to neglect the unwanted noise components from the
parameters, since the identified time-varying parameters
are later to be used for controller design for the system.
Figure 5 shows the time-development of the absolute
values of the poles and zeros of the identified model.
Remarks:
the model is unstable around time 0 ... 0.5 s and after that only marginally stable (the absolute value of
one pole stays around 0.995)

the model is minimum-phase, altoung the zero is

relatively close to the unit circle around time 0.3 ...
0.5 s

4. CONTROLLER DESIGN
4.1. Brief History of Adaptive Control
Adaptive controllers are considered to be one special
type among general nonlinear controllers. The theory
behind adaptive control has its origins in early 1950s,
when there was a need to design well-operating autopilots to high-performance aircraft. The fundamental
problem was how to control a system, which has several
operation points and which may even be continuously
varying between different operation modes. The fundamental issue then is to combine closed loop identification and control a combination which makes the problem nonlinear and extremely complex. For example, the
question of stability has been, and to some extent still is,
a difficult problem to deal with in adaptive control systems.
In 1960s control theory developed significantly,
when state-space theory was formulated and the stability
methods by Lyapunov became known in the control engineering community. Optimal control was developed in
the state-space framework, and major new ideas like dynamic programming, dual control, and stochastic control
were invented. Also, major advances in the identification theory were made. The basic concepts in adaptive
control model reference adaptive control and selftuning regulator or pole-placement adaptive control
were introduced.
The connections to the theory of linear quadratic control were established, which led naturally to minimumvariance control and its extensions also in the adaptive
context.
In 1970s and 1980s major stability results for adaptive controllers were found. A wide amount of practical
applications were reported. However, adaptive control
had a reputation of being a difficult control scheme,
which was often difficult to apply such that the stability
and performance specifications could be proved to hold.
A major drawback came, when it was noticed that many
adaptation schemes could lead to unstable control because of disturbances in the process or unmodelled dynamics in the process. As the theory of robust control
has been a hot topic for research in the 1990s and from
here on, the new concept of robust adaptive control has
been created to overcome the robustness problems. This
work is still going on.
Minimum-variance control and its generalizations,
moving-average control and general predictive control
are methods, which have been studied extensively in the
literature for different kinds of plants. Adaptive versions
of these control methods are also available and their
problems are well understood. Usually the problems are
related to difficulties in the on-line identification of the
plant in closed-loop and to the ringing effect of the control signal. No general solution to these problems exists
today, and the implementation of these controllers is a
combination of theoretical research and practical engineering expertise.

r (t ) +

GC

e(t )

C
A

u (t )

q k B
A

y (t )

Figure 6. A system with a feedback controller.

More information of adaptive control can be found

from [8-11].
4.2. Generalized Minimum Variance (GMV) Controller
The GMV controller is briefly introduced here. More information can be found from [0],[8],[9].
Figure 6 illustrates the concept of feedback control. It
is desired to design the controller GC in such a way that
the output y (t ) of the system follows the reference signal r (t ) . The original idea of minimum variance control
is to design a feedback controller to minimize the variance of the output of the controlled system with the reference signal set to zero. This is referred to as the regulator problem, whereas the nonzero reference signal case is
called a servo problem [8].
It is well known that pure minimum variance control
is not applicable to nonminimum phase systems. The
GMV approach circumvents this problem by including a
penalty for the control signal u (t ) in the cost function.
Thus, instead of minimizing the variance of y (t ) , the
following cost is minimized:

I 1 = E [ y (t + k ) r (t + k )] + Su 2 (t ) ,
2

(13)

is the expectation operator and

where E{}
S = s 0 + s1 q 1 + K + s n q n .

(14)

The controller can be further modified to force the

steady state error of y (t ) r (t ) to zero by addition of an
integrator into the feedback loop [8]. In this case the cost
function is

I 2 = E [ y (t + k ) r (t + k )] + Su 2 (t ) ,
2

(15)

where = 1 q 1 .
As can be seen in (15), the GMV controller no longer
minimizes the variance of the output, but instead minimizes the variance of a combination of the output, the
reference signal, and the control signal. The control signal penalty S is usually chosen as a constant in practice.
The GMV controller is [8],[9]

Fu (t ) + Gy(t ) Cr (t ) = 0 ,

(16)
1.5

1.5

y-r

(17)

F = BT + SC ,

y-r

where
0.5

and T and G are defined via the identity

0.5
0

C = TA + q G ,

(18)

(a)

0.5
s

1.5

1.5

(b)

0.5
s

(d)

0.5
s

T = 1 + t1 q 1 + K + t k 1 q ( k 1) ,

(19)

y-r

y-r

where
0.5

0.5
0

G = g 0 + g 1 q 1 + K + g n q

ng

n g = max(n a 1, n c k ) .

(20)
(21)

Our interest in this paper is to investigate the effect

of the tuning polynomial S to the controller performance.
4.3. Performance of the GMV controller designed
for the identified process

For the process model identified in Section 3.2, the

GMV controller is, from (1) and (13)-(21),
u (t ) =

a1 y (t ) + a 2 y (t 1) b1u (t 1) + r (t )
b0 + s

(22)

for cost function I 1 and

u (t ) =

a1 y (t ) + a 2 y (t 1) (b1 s )u (t 1) + r (t )
b0 + s

(23)

for cost function I 2 . Here we have used S = s. Note that

the indirect version of the adaptive GMV controller was
used, instead of the direct version as in [0], where the
controller parameters were directly identified.
We simulated an ARX(2,1,1) system, the parameters
of which were varied according to those in Figure 4. The
system parameters were estimated in real time using the
RLS algorithm, and the parameter estimates were at each
time fed to the GMV controller (22) or (23). The reference signal was a sawtooth wave of amplitude one, mean
value -17 and 1 second period, resembling the variation
of the SIR target by WCDMA outer loop power control.
Figure 7 shows the variance of y (t ) r (t ) versus the
tuning parameter s for the controllers (22) and (23). Subplots (b) and (d) show the results for unlimited control
signal u (t ) . Subplots (a) and (c) show the results of the
case where the control signal calculated from equations
(22) and (23) were quantized by taking only its sign, thus
resembling the power control method in [0]. The dashed
lines correspond to the Bang-Bang control scheme, were
the controller is
Sign (r (t ) y (t ) ) ,
where

(24)

(c)

0.5
s

Figure 7. Variance of y(t)-r(t) versus the tuning parameter with (a) cost I1, 1-bit command, (b) cost I1,
unlimited command, (c) cost I2, 1-bit command, (d)
cost I2, unlimited command. Solid lines: GMV controller, dashed lines: Bang-Bang controller.

1, if x 0
Sign (x ) =
1, if x < 0

(25)

As seen in Figure 7, the variance grows with the tuning parameter when the command is unlimited, as expected from (13) and (15). Since the system is minimum
phase, the controller works even with s = 0 . However,
including the nonlinear component (Sign-function) in the
GMV controller brings more ambiguity to the results.
With controller (22) the tuning parameter is not affecting
the results at all. With controller (23) there is still the
trend of growing variance with respect to s, altough not
so clear as with unlimited commands. Note that by selecting s properly the variance is always smaller than
with the Bang-Bang controller (24).
5. RADIO NETWORK SIMULATION RESULTS

Having observed the controller performance in a somewhat idealized linear world, we now check the performance of the designed controller in the radio network
simulator described in Section 2.1. Now we let all the
mobile users in the system, including the observed user,
be controlled by the designed adaptive GMV controller.
Figure 8 shows the receiver SIR and SIR target for
the observed user. The SIR target (in decibels) was the
same sawtooth signal as the reference signal used in Section 4.3, but with a slightly different mean value (since
the mean value was set in the beginning of the simulation to such a level that the radio network can still support all users). Three different controllers were used,
namely the GMV controller (23) with s = 0.1 , MV controller (corresponds to GMV controller with s = 0 ) and
the Bang-Bang controller of [1]. In all controllers, the
control signal was limited to 1 dB.
It is seen that the variance minimizing controllers indeed achieve smaller variance than the Bang-Bang controller. With the minimum variance controller there are

-12
-14
-16
-18
-20
-22
0

0.1

0.2

0.3

0.1

0.2

0.3

(a)

0.4

0.5

0.6

0.7

0.4

0.5

0.6

0.7

-12
-14
-16
-18
-20
-22

(b)

performance. We designed a GMV controller for the

identified process model, and verified its performance by
simulation of the linear model and by simulation in the
WCDMA network simulator that was used for data collection. The WCDMA network simulations showed good
correspondence to simulations with the linearized model.
However, the optimal selection of the tuning parameters
of the adaptive controller for this particular case
(WCDMA power control) is still an open problem and is
a subject for further research.

-12
-14
-16

7. REFERENCES

-18
-20
-22
0

0.1

0.2

0.3

(c)

0.4

0.5

0.6

0.7

Figure 8. Receiver SIR and SIR target for the observed user, when all users are power-controlled
using (a) GMV controller, (b) MV controller and
(c) Bang-Bang controller.

Table 1. Variance of SIR SIR target of the observed user

Algorithm
GMV ( s = 0.1 )
MV
Bang-Bang

Var(SIRSIR target)
0.8854
0.8789
1.2961

problems in the beginning (between 0 ... 0.1 seconds),

which is due to the identified system being nonmimimum phase. By including a small penalty for the control
signal, we see that the GMV controller avoids these
problems with insignificant change in the variance. The
variance of SIR minus SIR target (with the first 0.15
seconds ignored) for the three cases are shown in Table
1.
6. CONCLUSIONS

We investigated the properties of WCDMA power control by modeling the power control process with a linear
parameterized ARX model. The model was identified recursively using signal data exported from a WCDMA
network simulator. The time-varying nature of the radio
network was taken into account by using RLS algorithm
with exponential forgetting for the identification. Using
this linear model, significant insight to the properties of
the system can be observed, and one can easily design
adaptive controllers to drive the system into a desired

[0] M. Rintamki, I. Virtej and H. Koivo, "Two-mode

fast power control for WCDMA systems," in Proc.
IEEE Semiannual Vehicular Tech. Conf. VTC2001
Spring, Rhodos, Greece, May 2001.
[1] S. Ariyavisitakul, "SIR based power control in a
CDMA system," in Proc. GLOBECOM92, Orlando, Florida, USA, Dec. 1992, pp. 868-873.
[2] J. Zander, "Performance of optimum transmitter
power control in cellular radio systems", IEEE
Trans. Vehic. Technol., Vol. 41, No. 1, pp. 57-62,
Feb. 1992.
[3] G. J. Foschini and Z. Miljanic, A simple distributed
autonomous power control algorithm and its convergence, IEEE Trans. Vehic. Technol., Vol. 42, No.
4, pp. 641-646, Nov. 1993.
[4] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. J.
Viterbi, L. A. Weaver, Jr. and C. E. Wheatley III,
On the capacity of a cellular CDMA system, IEEE
Trans. Vehic. Technol., Vol. 40, No. 2, pp. 303-312,
May 1991.
[5] W. C. Jakes, Ed., Microwave Mobile Communications, New York: John Wiley, 1974.
[6] R. Gudmundson, "Correlation model for shadow
fading in mobile radio systems", Electronic Letters,
Vol. 27, No. 23, pp. 2145-2146, Nov. 1991.
[7] B. Makarevitch, Reinforcement Learning in CDMA
Admission Control, Licentiate thesis, Helsinki University of Technology, May 2001.
[8] R. Iserman, K. Lachman and D. Matko, Adaptive
Control Systems, Prentice Hall International Ltd,
1992.
[9] P. E. Wellstead and M. B. Zarrop, Self-Tuning Systems, Control and Signal Processing, John Wiley &
Sons, 1991.
[10] K. J. strm and B. Wittenmark, Adaptive Control,
2nd ed., Addison-Wesley, 1995.
[11] P. A. Ioannou and J. Sun, Robust Adaptive Control,
Prentice-Hall, 1996.