Sie sind auf Seite 1von 5

WCDMA Network Cost Function Minimization for Soft Handover Optimization with Variable User Load

John A. Flanagan
Nokia Research Center, PO Box 407 FIN-00045 NOKIA GROUP, FINLAND Adrian.Flanagan@nokia.com

Tomas Novosad
Nokia Networks, PO Box 301 FIN-00045 NOKIA GROUP, FINLAND Tomas.Novosad@nokia.com

AbstractWCDMA networks have many parameters which determine their performance. Ensuring a desired quality of service means a proper choice of parameters. One approach to choosing these parameters automatically is to minimize a cost function with respect to these parameters. The minimum of the cost function corresponds to an optimal network performance. The choice of cost function, in the simplest case can be some combination of Key Performance Indicators (KPIs) of the network. Here a cost function is described which can be used to optimize the window add handover parameter. A second order gradient method is derived which is used to minimize the cost function. The cost function and gradient algorithm are implemented on an advanced WCDMA radio network simulator. The robust nature is demonstrated using a time varying network trafc level.

II. C OST F UNCTIONS AND WCDMA N ETWORK P ERFORMANCE WCDMA networks have many parameters which determine their performance. Ensuring a desired quality of service means a proper choice of parameters. One approach to choosing these parameters automatically is to minimize a cost function with respect to these parameters. The minimum of the cost function corresponds to an optimal network performance. The choice of cost function, in the simplest case can be some combination of Key Performance Indicators (KPIs) of the network, for example
M

C = I. I NTRODUCTION The WCDMA radio interface for third generation mobile networks can carry voice and data services with various data rates, trafc requirements, and quality-of-service targets [1]. Moreover, the operating environments vary greatly from indoor cells to large macrocells. Efcient use of limited frequency band in the diverse conditions requires careful setting of numerous vital network and cell parameters. The parameter setting is referred to as radio network planning and optimization. Once a WCDMA network is built and launched, its operation and maintenance is largely monitoring of performance or quality characteristics and changing parameter values in order to improve performance. The operability of the network would greatly benet from automated monitoring and parameter changing. The automated parameter control mechanism can be simple but it requires an objectively dened performance indicator, or in this case a cost function, that unambiguously tells whether performance is improving or deteriorating. Conceiving of such indicators is a major task. WCDMA network autotuning and advanced monitoring are discussed in [2], for instance. This paper addresses the specic problem of deriving and optimizing a cost function based on the Key Performance Indicators (KPI), the Blocked Call Ratio (BKCR) and the down link transmission power (dlTxPower). The specic parameter to be optimized is the soft handover parameter window add (wadd). In the next section a more detailed discussion of the use of cost functions for optimizing network performance is given. In section III a description of a gradient algorithm suitable for optimizing a cost function is described. Following this section IV-A describes the simulator and section IV-B the simulation environment used to verify the cost function and gradient algorithm implementation. Section V describes the result of the simulation.
i=1

fi (KP Ii )

(1)

where KP Ii is the ith key performance indicator and fi is some positive function which can be used to transform, weight and/or scale the ith KPI. The network performance is optimized by minimizing this cost function C . The minimum of the cost is achieved by a proper choice of different network parameters W = (w1 , w2 , . . . , wN ) which would be considered the optimum value of the parameters. The cost function approach, implicitly assumes that the value of the KPIs are functions of the network parameters, that is, KP Ij = KP Ij (w1 , w2 , . . . , wN ) j (2)

and hence the cost function C is also a function of the network parameters, and can be rewritten as direct function of the parameters as,
N

C =
i=1

gi (wi , t)

(3)

where the gi are some functions which may vary with time t. The optimization of such a cost function is not so straightforward in a real situation. The main problems can be listed as follows, 1) In a real network there is much variation in trafc types, user distributions and load. These factors are not controllable by the network and can be considered as random external noise sources. Any optimization algorithm should be insensitive to such external random inuences. 2) Because of the variations in the load etc. at different times, the optimum choice of parameters for any given network can evolve over time (i.e. the functions gi of equation (3) vary with time). Any optimization algorithm should be able to adapt to such variations in the optimum

0-7803-7467-3/02/$17.00 2002 IEEE.

2224

operating point of a network and be able to track these changes. 3) Generating a model of the network which can be used for optimization in every different situation may not be feasible. This means that the functions gi of equation (3) are not known. The alternative, and the approach taken here, is not to assume any model for the network but to base the optimization uniquely on network measurements. With these different factors in mind a general type of optimization algorithm is derived in the next section suitable for use in the minimization of a cost function of KPIs. However before that the specic cost function used in this work to optimize the choice of the window add parameter is described. Soft handover allows for the seamless transfer of terminals moving between the coverage areas of adjacent cells. The idea is that the terminal can establish links with several base stations at the same time. The base station whose pilot power Ec /Io is the largest received by the terminal serves as a reference. The active set then consists of those base stations whose received Ec /Io is greater than the Ec /Io of the best base station minus the value of window add. The maximum size of the active set is usually limited for example to 3. The overhead refers to these extra connections to the base stations in the active set apart form the best received base station. In general as the value of window add increases then the overhead increases. However the use of soft handover can also improve the connection quality for the terminal, especially in the case of high transmission bit rates in the uplink. The reason for this is due to the recombination gain between the base stations in the active set. On the other hand the extra links cause problems in terms of the extra power needed to establish the links, particularly in the downlink. Hence there are conicting interests for increasing and decreasing the level of overhead. Too high overhead leads to large transmission powers in the downlink whereas too low means reduced quality in the uplink which leads to larger transmission power required in the uplink and hence higher blocking. The following cost function is then proposed to nd a compromise between these two opposing trends and an optimum value of window add. C = f (BlockingRatio)2 + dlT xP owerRatio2 (4)

for one single parameter. It is quite easy to extend the derivation to the case of several parameters. Consider the general case of a cost function C to be minimized with respect to a parameter denoted w . Let w0 be the value of w which minimizes C . Evaluating C (w0 ) using the Taylor series expansion about any value of w gives, C (w0 ) = C (w) + (w0 w)C (w) + (w0 w)2 C (w) (6) 2

where C (w) is the rst order differential of C with respect to w and C (w) the second order differential. As C (w0 ) is a minimum point of C then differentiating equation (6) with respect to w0 and letting the result equal 0, gives, w0 = w C (w) C (w) (7)

which is a classic Gauss-Newton algorithm optimization algorithm. In the case where C is a quadratic function of w then there is a one step convergence to the optimum point w0 . In the case where C is not quadratic then convergence is guaranteed so long as C (w) is always positive. In the case where it is not then it can be made positive and equation (7) resolves to a standard gradient algorithm. However as the minimum point of C is approached then the quadratic approximation becomes more accurate and the convergence speed is high. The algorithm described here is a classic and widely used. But now consider the problem in terms of the WCDMA cost function. From point 3 in the introduction we do not have a model of the network and it is difcult to nd a value for C (w) and C (w) hence equation (7) is difcult to use. It is shown how C (w) and C (w) can be derived from network measurements which allows for the use of equation (7). The theoretical evaluation of these terms is rst described and this is followed by a description of how they can be evaluated in real networks. Consider a small change w > 0 in the value of the parameter w to give a new parameter value w + w. The value of the cost function can then be approximated as, C (w + w) = C (w) + wC (w) + w2 C (w) 2 (8)

where BlockingRatio is the ratio of the number of blocked calls to the number of attempted calls and dlT xP owerRatio is the average downlink transmission power per base station divided by the maximum possible power transmission of the base stations. The equation f is a non-linear functional given by, f (x) = exp(12 x) 1 (5)

Next consider a decrease in the value of w by w to give a new parameter value w w. Once again the value of the cost function for this parameter value can be expressed as, C (w w) = C (w) wC (w) + w2 C (w) 2 (9)

The reason for this choice of functional is to punish large values of the blocking. For example for BlockingRatio = 0.05 then f (BlockingRatio) 1, this means that if the blocking becomes large then it will start to dominate the cost function and any changes to window add will be to decrease the level of blocking. III. G RADIENT A LGORITHM FOR C OST F UNCTION O PTIMIZATION The optimization algorithm to be used in the minimization of the cost function is described and derived from rst principles

By some simple algebraic manipulation it is possible to derive values or approximations to the values of C (w) and C (w) by knowing the values of C (w) C (w + w) and C (w w). The next question is how to evaluate these values at any particular time. The following algorithm is proposed. 1) At time t1 the parameter value is w and the value of the cost function C (w, t1 ) is evaluated from equation (1) based on network measurements of the appropriate KPIs at time t1 . 2) At time t1 the value of w is changed to w + w.

0-7803-7467-3/02/$17.00 2002 IEEE.

2225

3) At time t2 = t1 + t, t > 0, the value of the cost function is evaluated from equation (1) based on network measurements of the appropriate KPIs to give C (w + w, t2 ). 4) At time t2 the parameter w is changed to w w. 5) At time t3 = t2 + t the cost function is evaluated from equation (1) based on network measurements of the appropriate KPIs to give C (w w, t3 ). 6) At time t3 a new value of w is calculated using equation (7) with C (w) and C (w) respectively which can be derived from the values of C (w), C (w + w) and C (w w) given respectively by the measurements C (t1 ), C (t2 ), C (t3 ). These steps constitute one cycle of the algorithm and the cycle can be repeated over. IV. S IMULATION E NVIRONMENT A. The WCDMA Network Simulator The minimization of a cost function using the gradient algorithm is veried with an advanced WCDMA radio network simulator developed at Nokia Research Center in Helsinki [3]. For this analysis the simulator models 17 cells deployed over an area of Helsinki center. A set of mobile terminals move in the area with constant speed and, with random intervals, make calls of different services: voice, circuit-switched data, and packetswitched data. In the results presented here a speech and circuit switched load is considered. The simulation step is one frame or 100 ms, at which the transmission powers, received interference, and signal-to-interference ratios are recalculated for each connection in uplink and downlink. The method of [4] is used to obtain correctness of received frames from signal-tointerference ratios. The simulator implements many advanced features such as admission control, closed-loop and outer-loop power controls, soft and hard handover controls, packet scheduler, load control, and quality manager. Previous studies with the simulator are described in [1,7], for instance. Quality manager The quality manager is a logical unit in the radio network controller that collects statistics of various performance indicators. The quality manager calculates these statistics over a specied interval of time, which will be called qmInterval . Some of the statistics made available by the quality manager include : Quality - At every qmInterval, interval, the quality manager goes through all connections of the sector and checks the call quality. The number of bad quality calls and the total number of calls are accumulated in two counters over the control period. The quality is obtained as the ratio of the counter values. Blocking - The ratio of the blocked calls to the total number of admission requests during the previous qmInterval. Dropping - The ratio of calls ended by dropping to the total number of ended calls during the previous qmInterval. As this study only the blocking ratio was used, however the methods and simulations could be extended to include the other statistics returned by the quality manager. B. Simulation Scenario Fig 1 shows the general view of the simulated network. Green bars indicate sectors by pointing to the principal direc-

Fig. 1. Deployment of ten three-sector and two one-sector macrocell sites in Helsinki center area. Water areas appear as uniformly grey area. Simulated area is inside the rectangle.

tion of antenna pattern. The antenna of sector 26 was omnidirectional, however. The channel multipath prole was that of ITU Vehicular A [5] with 5-path propagation. The path gains are shown in Table 1. One half of the signal power came along the line of sight and the other half was a sum of powers from four reected signals. In the downlink, signals from the same base station propagating along the same path were totally orthogonal, that is, they did not interfere with each others. Thus, the downlink orthogonality factor [1] computed from the path gains was 60. The propagation loss was computed with the Okumura-Hata model [6]. The mobile stations were uniformly distributed along the streets of the simulated area and they made new calls according to a Poisson inter arrival distribution. The service of new calls was generated according to the probabilities shown in Table I. The network parameters are collected in Table I. V. R ESULTS To illustrate the operation of the optimization two simulations were carried out based on the simulation parameters described in the previous section. In the rst simulation the value of window add was kept constant at 1dB and the value of window drop was 3dB . In the second simulation the value of window add at the beginning of the simulation was 1dB and window drop was 3dB , but during the simulation the value of window add was adapted using the gradient algorithm of section III to minimize the cost function in equation (1) by varying window add. The variations of window add used to evaluate the terms for the gradient algorithm was w = 0.5dB . At all times window drop was given as 2dB higher than window add. To illustrate the robustness of the algorithm a varying load was used, the variations of the load are illustrated in gure 2. During the simulation the horizontal axis is scaled to the duration of the simulation which was 3600s. The number of mobile stations in the simulation at any time is given by the maximum number of terminals, 6000, multiplied by the value of this curve. Hence there are two load peaks during the simulation. The sampling rate and integration period of the quality manager for collecting the required statistics to evaluate the cost function, qmInterval,

0-7803-7467-3/02/$17.00 2002 IEEE.

2226

TABLE I
N ETWORK PARAMETERS WHERE , BS=BASE S TATION , MS=M OBILE S TATION , AC=A DMISSION C ONTROL , CS=C IRCUIT S WITCHED , UL= UPLINK , DL= DOWNLINK . Parameter Chip rate Frequency Bandwidth BS max. tx power macrocell Mobile station max tx power CPICH tx power macrocell Power control dynamic range Max link power in downlink BS antenna sector and gain MS antenna sector and gain UL system noise DL system noise Minimum coupling loss Propagation loss model Multipath prop. gains MS speed Max No of MS Call arrival rate Probability of voice service Average voice call length Voice data rate Voice outer loop FER target Probability of CS service Average CS call length CS data rate CS outer loop FER target AC noise rise target AC tx power target macrocell Initial window add Initial window drop qmInterval Value 3.84MHz 2.0 GHz 5.0 MHz 20 W 125 mW 1W 65 dB in UL, 20 dB in DL 1W 65, 17.5 dBi Omni, 11.0 dBi Omni, 0.0 dBi -102.9 dBm -99.9 dBm -50 dB Okumura-Hata 50, 30, 11, 6, 3 3 km/h 6000 5 calls/hour/user 0.8 120 s 8 kb/s 0.01 0.2 120s 64 kb/s 0.01 6 dB 10 W 1 dB 3 dB 10s
Load Scaling

0.8

0.6

0.4

0.2 0 0.2 0.4 0.6 0.8 1

Time Scaling

Fig. 2. Variable load curve used during the simulations.

8 7 6 5 4 3 2 1

was 10 seconds. This is also the period for updating the value of window add during the optimization simulation. Figure 3 shows the variations of window add during the two simulations. The dotted line indicates the constant value of 1dB for the simulation where the window add is not optimized. The continuous line indicates the optimum value of window add determined using the cost function and gradient algorithm. The maximum value of window add has been limited to 8.5dB. It is clear that the changes in window add in the optimized cases are correlated with the variations in the load. In the initial part of the simulation the value of window add increases quite quickly and then saturates at the 8dB level. This is followed by a decrease to a level of approximately 3dB and then it increases again. The decrease after the second peak is not so signicant. The variations of the cost function are shown in gure 4, it is clear that despite the noisy uctuations that on average the cost function in the optimized case is less than that in the nonoptimized case and there are fewer spikes in the curve. In fact for the simulation with optimization of window add the average value of the cost function over the whole simulation is 0.23 while for the second case its 0.41. This means the optimization of window add decreased the cost function, on average, by 40%. Figure 5 shows the blocking ratio for the two simulations, the increase in blocking for the two cases with an increase in load is obvious, however for the optimized case the increase is on average less than in the non-optimized case. For the com-

Window Add dB

0.5

1.5 2 Frames

2.5

3.5 x 10
5

Fig. 3. Variation of window add during the two simulations. - line indicates the optimization simulation, indicates non-optimized simulation.

plete simulation the average blocking in the optimized simulation was 1% while in the non-optimized case it was 2.5%, hence the blocking rate was halved by using optimization. The most signicant effect is on the uplink shown in gure 6. The compromise can be seen in the downlink blocking which increases in the optimized case however more signicantly is the increase in down link transmitted power shown in gure 7 where there is an increase in the transmitted power in the optimized case as a result of the increasing overhead during the higher load conditions. VI. C ONCLUSION The automatic selection of a single network parameter, window add, has been described based on the minimization of a cost function made up of some KPIs from the network. An adaptive second order gradient algorithm has been derived suitable for minimizing the cost function. Not only is the gradient algorithm robust in the presence of noisy inputs caused by the short sampling period of 10s for deriving network statistics, but

0-7803-7467-3/02/$17.00 2002 IEEE.

2227

0.45
2.5

0.4 Downlink TxPower Ratio 0.35 0.3 0.25 0.2 0.15 0.1

2 Cost Function C

1.5

0.5

0.05 0 0 0.5 1 1.5 2 Frames 2.5 3 3.5 x 10


5

0.5

1.5

2.5

3.5 x 10
5

Frames

Fig. 4. Variation of the cost function during the two simulations. - line indicates the optimization simulation, indicates non-optimized simulation.
0.08 0.07

Fig. 7. Variation of the down link transmitted power ratio during the two simulations. - line indicates the optimization simulation, indicates nonoptimized simulation..

600

500 No of Active Users

0.06 Blocking Ratio 0.05 0.04 0.03 0.02 0.01 0

400

300

200

100

0 0

0.5

0.5

1.5

2 Frames

2.5

3.5 x 10
5

1.5 2 Frames

2.5

3.5 x 10
5

Fig. 5. Variation of the blocking ratio during the two simulations. - line indicates the optimization simulation, indicates non-optimized simulation.

Fig. 8. Variation of the number of active users during the two simulations. - line indicates the optimization simulation, indicates non-optimized simulation.

0.1

Uplink Blocking Ratio

0.08

0.06

it also is robust in terms of adapting to changes in the operating point of the network, caused in this case by the changing number of terminals in the simulation. The cost function approach allows for a high level performance target setting of a WCDMA network sparing the need for operators to understand the lower level fundamentals of the network. R EFERENCES [1] H. Holma and A. Toskala, editors. WCDMA for UMTS. Wiley and Sons, 2001. [2] J. Laiho, A. Wacker, and T. Novosad, editors. Planning and Optimisation for UMTS. Wiley and Sons, 2002. [3] S. H am al ainen, H. Holma, and K. Sipil a. Advanced WCDMA radio network simulator. In Proc. PIMRC, pages 509604, 1999. [4] S. H am al ainen. A novel interface between link and system level simulations. In Proc. ACTS Summit, 1997. [5] Universal Mobile Telecommunications System (UMTS): selection procedures for the choice of radio transmission technologies of the UMTS. TR 101 112 V3.1.0 (1997-11), UMTS 30.03 version 3.1.0. UMTS30.03. [6] J. Lee and L. Miller. CDMA Systems Engineering Handbook. Artech House, London, 1998.

0.04

0.02

0 0

0.5

1.5 2 Frames

2.5

3.5 x 10
5

Fig. 6. Variation of the uplink blocking ratio during the two simulations. - line indicates the optimization simulation, indicates non-optimized simulation.

0-7803-7467-3/02/$17.00 2002 IEEE.

2228

Das könnte Ihnen auch gefallen