Vous êtes sur la page 1sur 6

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the IEEE ICC 2009 proceedings

A mathematical perspective of self-optimizing wireless networks


Ingo Viering
Nomor Research GmbH Munich, Germany Email: viering@nomor.de

Martin D ottling, Andreas Lobinger


Nokia Siemens Networks Munich, Germany Email: {andreas.lobinger; martin.doettling}@nsn.com

AbstractWe present a mathematical framework for quantitative investigations of self-optimizing wireless networks (SON) with focus on the 3GPP Long-Term Evaluation (LTE) system. Basic target functions, such as the signal-to-noise ratio distribution, the number of satised users, or energy efciency are derived as a gure of merit, including the impact of adaptation of downlink transmit power adaptation, antenna tilt, and the handover parameter. The framework is exemplied by basic investigations on load balancing.

I. I NTRODUCTION Self-organization and self-optimization of mobile wireless systems is of utmost importance to operators and vendors and investigated for Long-Term Evolution (LTE) in research [1] and standardization [2]. Among others, important use cases are coverage and capacity optimization, energy savings and mobility load balancing optimization. The goal is two-fold: benets are expected from dynamic optimization (such as load variations and environmental changes), and the level of manual optimization steps should be minimized to reduce operational expenditures. Detailed discussions of SON requirements and assessment criteria can be found in [3]. The SON algorithms have to be sluggish in order not to collide with the highly adaptive radio resource management functionality. And there is actually no need for fast SON algorithms, since the underlying variations to be tracked are rather slow as well. Hence the time scale is at least in the order of minutes, more typically in the order of hours/days (load variations) or even weeks (environment). Conventional system level simulators for accurate performance simulations are not appropriate to investigate those SON mechanisms since they operate far beyond real-time. Hence, new methods are required. We propose a mathematical framework, which can be used for investigations alone, however which may also serve as the basis for a new class of simulation tools. This work concentrates on the evaluation methodology. We also present basic results for load balancing and coverage-related use cases. However, more elaborate investigations will be presented in future work. II. D EFINITIONS We will discuss the LTE downlink, which is a Localized OFDMA system. Adjacent OFDM subcarriers are grouped into so called physical resource blocks (PRBs). This is the smallest

unit which can be allocated, and the grouping is exactly the same over adjacent cells. Since the investigated optimization schemes operate at slow time scales based on statistics and averaged measurements, we have excluded small scale fading and all other fast effects and functions from the investigations at hand. The assumption is, that the optimizations described here inuence neither those algorithms, nor their impact on the performance. Absolute gures in this paper will deviate from reality, however the relative differences should give us clear indications about the functioning and potential of SON algorithms. A. General denitions We are assuming a network given by: C cells, every cell c is served by a base station (BS) at the position pc . Note that some cells will have the same BS position (sectorization). All BS antennas are mounted at height hBS . U user equipments (UE), each UE u is located at position qu , and desires to communicate with a bit rate Du . All users are located on the ground (height zero). An environment represented by C shadowing maps Mc (q ) (details below). A pathloss law given by the parameters lA and lB . Thermal noise N (per PRB, including UE noise gure). Furthermore, every cell c has a transmit power Pc for a single PRB (note that this power is cell specic, but not user specic, i.e. we are still assuming no power control) a 3-dimensional (3D) beampattern B (, ) which can be inuenced to some extent (details discussed below). We assume that user u is served by cell c = X (u), where X (u) is the connection function , and every user is connected exactly to a single BS. Finally we dene a cell load c of cell c, which denes the fraction of available PRBs used in cell c. In the following sections we will discuss the propagation models used in this work; all expressions will be given in dB. B. Distance dependent pathloss The distance dependent pathloss is straight forward, it is typically given by the expression LP L = lA + lB log10 (d/km) (1)

978-1-4244-3435-0/09/$25.00 2009 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2009 proceedings

where d is the distance |p q | between a BS and a UE. Note that the lB is ten times the pathloss exponent. Typical values are lA = 148.1dB and lB = 37.6dB [5]. C. Beampattern The 3D antenna pattern is typically given by a horizontal (azimuth) and a vertical (elevation) cut. Note that the real 3D pattern cannot be exactly extrapolated by the 2 cuts, however approximations are typically used [6]. In most 3GPP simulations so far only the azimuth has been considered, however results already have shown that the elevation offers an attractive degree of freedom to further optimize the system [8]. We assume an azimuth pattern B () as dened in [5] which is given by B () = max B0 , 12
2

i.e. Mc (q1 ) and Mc (q2 ) have some correlation if the distance |q1 q2 | is smaller than dS . Details are found in [5]. Note that for our investigations it is irrelevant where the (deterministic) shadowing maps come from, we could even use eld measurements. E. Overall attenuation Finally, the overall attenuation is the sum over distance dependent pathloss, shadowing and beampattern which depends on a number of parameters: L(pc qu , lA , lB , c , , , , B0 , hBS ) = LP L + B (, ) + Mc (qu ) F. Classication of parameters The parameters identied in the previous sections can be classied into different groups, i.e. parameters given by the environment (pathloss {lA , lB }, shadowing {Mc (q ), D, S }), by the deployment (BS positions pc and height hBS , azimuth orientation c , beamwidths and , and backward attenuation B0 ), and inherently (cell load c , will be discussed later). These parameters are considered as constants later on. In contrast, the most important group are parameters which we can adjust during operations and thereby used for optimization. Those are: the transmit power of each cell Pc , the downtilt of each cell c , and the user connection function X (u) We assume that the network operator has access to those parameters, and can adjust them during operations. Also beamwidth and orientation in azimuth can be optimized by beamforming techniques. However, this is not in the scope of this paper. With C cells and U users, we have a total of 2 C + U parameters which can be optimized. A third group of parameters comprises the UE positions qu and the UE rate requirements Du . In our model, those are the parameters which will vary over time, and the optimization of {Pc , c , X (u)} is purely based on those parameters. Note that the environment may also change slowly over time (vehicles, seasonal effects, construction etc.), however this is not considered here. With the previous discussion in mind, we will exclude the parameters of the rst group from the long list of inputs to the overall pathloss (6). Hence, for the sake of convenience we will abbreviate the complicated expression as (in the following we will use the linear form instead of the dB values): Lc (qu , c ) = 10 10 L(pc qu ,lA ,...,hBS ) III. O PTIMIZATION On a high level, the goal of this section is to sketch how the parameters {Pc , c , X (u)} can be optimized for a particular set of UE positions and rates {qu , Du }, given all the other parameters.
1

(6)

(2)

where , and B0 are the azimuth orientation, the azimuth beamwidth, and the backward attenuation. Typical parameters are = 70o and B0 = 20dB. A single BS typically serves three cells (sectors) which have the azimuth orientation = {0o , 120o , 120o}. For the elevation pattern B () we will assume the same equation B () = max B0 , 12
2

(3)

Note that we have assumed the same backward attenuation as in the azimuth. A typical value for the elevation width is = 10o . The orientation is also called downtilt, and this will be subject to optimization later on. We follow the simplest extrapolation approach in [6] and sum up the two pattern to get the 3D patterns B (, ) = B () + B (). (4)

Note that this approach for elevation modeling is based on electrical downtilt, and has also been followed in [7]. The denition of the angles and is straight forward: = (pc qu ) ; = arctan hBS |pc qu | (5)

For mechanical downtilt a coordinate transform is required. D. Shadowing Shadowing is included by use of so-called shadowing maps Mc (q ), which are deterministic functions of a particular UE position q in the network with respect to a cell c. These maps have been generated randomly from a log-normal distribution with zero mean and standard deviation S , and with the common mutual correlations where Mc1 (q ) and Mc2 (q ) are fully correlated if the cells c1 and c2 are served by the same BS, and correlated with coefcient 0.5 otherwise. Finally, there is a spatial correlation given by a decorrelation distance dS ,

(7)

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2009 proceedings

A. Conventional solutions Without the new SON features, the solutions for {Pc , c , X (u)} are agnostic against the actual positions and rates {qu , Du }. The parameters are statically optimized given a distribution, or some average/representative scenarios. Assume a scenario, where you have high user concentration in one place during the day (e.g. shopping mall), and in another place during the night. The conventional network planning would have to assume both load concentrations at the same time. The ability to dynamically optimize the parameters (i.e. separately for day and night) should lead to a more economic network design saving some BSs. In a homogenous case with same size of all cells, the typical solution looks as follows:

2) Constant Bit Rate: In general, we can dene the load as (10) Ntot where Nu and Ntot is the (actual) amount of resources occupied by user u and total number of resources, respectively. u to meet Du depend Unfortunately, the required resources N on the SIN Ru . Let us introduce a throughput mapping Ru = R(SIN Ru ) which expresses the data rate per PRB given SIN Ru . The most prominent example for such a function R() is Shannons equation, more accurate results are obtained if the function is taken from link level results. With this denition, the amount of required resources can be written as c = u = N Du R(SIN Ru ) (11)
u|X (u)=c

Nu

Pc = Pmax , where Pmax is the maximum transmit power of the BSs. c = 0 , where 0 is a network wide downtilt. X (u) = arg maxc Pc Lc (qu , c ), i.e. all users are connected to the cell which is received with highest power on average (best server).

u only holds for satised users, in general we have Nu = N u . Since cell load cannot exceed c = 1, we obtain Nu N in the CBR case: c = min 1,
Du u|X (u)=c R(SIN Ru )

This solution will be assumed as a baseline; it will also be the initialization of the iterative algorithms. B. Target functions The most critical aspect in an optimization problem with such a huge number of parameters is a suitable denition of the target functions, i.e. the metric which allows to judge that a solution is better than another one. In general there will be two classes of target functions: capacity-based targets would optimize gures such as cell throughput, cell edge throughput, number of satised users, etc.; energy-based targets would optimize power consumption of BSs and UEs. At least the capacity-based targets will be based on the signal-to-interference and noise ratio of every user. With the denitions above we can write the SINR for UE u on a single PRB as PX (u) Lc (qu , c ) SIN Ru = N+ c Pc Lc (qu , c )
c=X (u)

Ntot

(12)

Unfortunately we now observe that (8) is not a closed-form expression, since the each cell load c depends on all SIN Ru and vice versa. Instead, we have a system of Nu non-linear equations which is quite tough to resolve. In section III-B4, we will propose an approximation for this. 3) Resource fair scheduler with PRB limit per user: The full buffer model in section III-B1 shows a very poor dependency of the load on the number of users which is probably too unrealistic. Instead of a bit rate requirement Du we can dene a maximum number of PRB per user Nmax,u . Illustratively, we can argue that this is the required number of PRBs to achieve Du at an average position. If the number of UEs U is sufciently large, this approach should lead to similar results as in the CBR case of the previous section. Here, we can write the load as c = min 1,
u|X (u)=c

Nmax,u

Ntot

(13)

(8)

A particular issue in this equation is the interaction of cell load c with service requirements, scheduler design, and SINR distribution within the cell. Recall that c is dened as the fraction of used PRBs in cell c. We distinguish three cases. 1) Full buffer, resource fair scheduler: As soon as there is a single user in the cell, all PRBs would be used, and the load is c = 1, otherwise c = 0. The UEs served by one cell would share the available resources equally. Note that the UEs would not have a rate requirement Du in this case. We can write: c = 0 f or 1 f or u|X (u) = c . otherwise (9)

4) Iterative solution: Finally, we propose an iterative solution for (12). We could use the nite buffer approximation (13) as initialization, then calculate SIN Ru according to (8), then calculate the correct cell loads for those SINRs with (12), update SINRs again based on the new loads, and so forth. We will not pursue this approach in this work, we will study the convergence of such an iterative solution for SINR calculation in future work. C. Capacity-based target functions In the previous sections we have shown how we can calculate the SINRs of the users. However, the SINRs itself do not represent a suitable target function. Based on the SINRs we will now dene throughput-based target functions. First of all, exactly as in section III-B2 we calculate a per-PRBthroughput from the SINRs, i.e. Ru = R(SIN Ru ). In order

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2009 proceedings

to calculate a UE throughput, we still need the number of PRBs per UE Nu . 1) Resource fair - throughput percentiles: In the resource fair case, every UE gets the same amount of resources, i.e. every UE gets on average Nu = min Nmax,u , Ntot
u |X ( u)=X (u)

(14)

Note that focussing only on this target is critical, typically we combine this target with other targets. For instance, if we nd that all UEs are satised in a low-loaded scenario (Z = 0), then we can continue optimization with this target, until the rst UEs become unsatised. Furthermore, environments with high BS densities are typically heavily interference limited. Here we might be able to reduce the transmit powers without affecting any capacity-based target. E. Coverage holes In general, we have to take care that our optimization does not create coverage holes. Note that we have formulated the SON problem such that it optimizes the situation for the active users in the system which are located at given positions qu . Inactive UEs may request access to the system, so we have to ensure that at least broadcast channels are still readable everywhere in the network, and that random access is still possible. We can formulate this requirement as a side condition for our optimization problem: Pc Lc (q, c ) PRx,min (19)

where the denominator simply expresses the number of UEs connected to the same cell as u. Note that we have reused the maximum number of PRBs Nmax,u from section III-B3. Now we have an expression for the data rate for every UE u = Nu R(SIN Ru ) D (15)

and we can use typical data rate requirements in terms of percentiles, e.g. a particular rate which is exceeded by 95% of all UEs in the whole network (or in a neighborhood of cells). 2) Constant Bit Rate: If we look at CBR trafc models it doesnt make sense to look at throughputs, since the UEs either get exactly the CBR or they are unsatised. Hence, we will dene the number of unsatised users in the following. Equation (11) has already given an expression for the u of PRBs of UE u given SIN Ru . Based required number N on this we dene a virtual load c , whose only difference to (12) is that we allow it to exceed 1: . (16) Ntot This gives us a clear indication how overloaded a cell is. Note that c = 1 means all users in the cell are satised, c = 3 means one third of the users are satised. We can state the number of unsatised users in the whole network as (note that we need the max operator since the number of unsatised users cannot be negative in underloaded cells) 1 . Z= max 0, 1 1 (17) c c c =
u|X (u)=c Du u|X (u)=c R(SIN Ru )

for all positions q in the network, and PRx,min is the minimum received level at the UE. It is quite difcult to introduce this side condition into the optimization problem, however any found solution should be checked whether this is still fullled. F. Algorithms First of all we will reduce the U parameters of the connection function X (u) by abandoning the option of individually connecting the users. Instead we will use a cell-specic handover offset Tc such that the connection function becomes X (u) = arg max Pc Lc (qu , c ) Tc .
c

(20)

In the real system unsatised users would not be admitted to enter the system, or they would be dropped if they are already in. We could always achieve better performance if we prefer to drop (or not admit) edge users. In the previous equation it is inherently assumed that such a discrimination is not done. D. Energy-based target functions The recent environmental discussions have made the energy consumption an important topic. The energy consumption of a cell will be a non-linear function of the transmit power Pc . Note that even with Pc = 0W there will be a signicant consumption, due to supply voltage, site operation, etc. We introduce the energy consumption function E (Pc ), which is only zero if we switch a cell off entirely. We can formulate the energy consumption in the overall network as Etot =
c

E (Pc )

(18)

Increasing Tc > 1 effectively enlarges the area served by cell c, since users will be connected even if the received level is not the highest. Such a forced handover certainly will cause a very low SINR for the affected users. Note that with (20) we have decreased the parameter space from 2 C + U to 3 C . We have already mentioned that this paper is focused on the mathematical framework. We will present only a very simple algorithm which only modies the handover thresholds Tc , and leaves Tx power Pc and downtilt c with their default values. The following considerations are based on the CBR case described in section (III-C2). Assuming a current conguration {Pc , c , Tc } we will calculate the virtual load c according to (16) for every cell c. The most critical cells are those which exceed a certain upper load threshold U . The basic idea is to decrease the served area of those overloaded cells. We will accomplish this by decreasing the handover offsets Tc for those cells. Then we will recalculate the virtual load with the new conguration {Pc , c , Tc }, determine the most critical cells, adjust their handover offsets, and so forth. In the same way, we increase the offsets Tc in underloaded cells having a virtual load below a lower threshold L but not above Tc = 1. The algorithm will stop if all cells are below this load threshold

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2009 proceedings

50

average number of unsatisfied users Z

1000 800 600 400 200 0 200 400 600 800 1000 1000

28 30 29 25 27 26 22 24 23 55 57 56 52
Base Station nonHO UE HO UE 53 500

31 33 32 7 9 8 11 1 6 3 2 19 21 20 49 54 50
0

34 36 35 10 12 38 13 15 14 16 18 17 46 51 47
500 1000

45 40 35 30 25 20 15 10 5 0 0 5 10 15

S1 (init Tc=1) S2 (init with S1)

37 39 40 42 41 43 45 44 48

4 5

20

25

iteration
Fig. 2. Unsatised users Z averaged over 50 user drops

Fig. 1.

Cell layout with Scenario S1 load concentration

T (i.e. all users are satised), or if all overloaded cells have reached a minimum threshold Tmin (i.e. we declare this setup to be infeasible). Obviously, smarter algorithm are conceivable, e.g. which work cell-pair-wise (instead of cell-wise in our case), and which coordinate the adaptation of the handover thresholds with their neighbors. However, for the simple scenarios in this paper such an uncoordinated algorithm serves its purpose. More elaborate solutions will be presented in future work. IV. S IMULATIONS We will present results for a simple scenario which shows how load balancing can be achieved by very simple means. Note that all of the previous considerations are not based on a particular cell layout. They even cover heterogenous layouts with different and irregular spacing of the BSs. The same holds for the user positions. A. Layout, scenarios and parameters A regular hexagonal 57 cell layout with an inter site distance of 500m and wrap around is assumed (cf. Figure 1). Every cell represents one of 3 sectors served by a single BS. The hexagons are shown only for illustration, they have nothing to do with actual cell boundaries. The maximum transmit power is Pmax = 40W. We will not consider the impact of elevation, i.e. hBS = 0 and o = 0o . The pathloss law is given by lA = 148.1dB and lB = 37.6dB, thermal noise is N = 114dBm. Note that this is the reference setting for LTE simulations [5] which is a strongly interference limited scenario (Case 1). We are assuming that every user requests a constant bit rate of 512kbps. The LTE capacity in this case would be in the range of 19 UEs per cell (assuming 10MHz bandwidth). The algorithm described in III-F adjusts the handover offset Tc by 1dB per iteration, the minimum threshold is Tmin = 10dB, and the upper and lower load thresholds are U = 1 and L = 0.2.

B. User position In order to demonstrate load balancing, we will articially create heavy load concentrations in our 57 cell network. A total of 270 UEs are randomly dropped into concentration areas comprising all 3 sectors of 3 sites as shown in Figure 1 (every + and . marker represents a UE). 216 UEs are randomly dropped into the rest of the network. This leads to an average of 30 UEs per overloaded cell and 4 UEs per partially loaded cell. For every scenario we repeat our calculations for 50 different random user drops. We will refer to the scenario in Figure 1 with user concentrations in cells {4, 5, 6, 34, 35, 36, 43, 44, 45} as scenario S1. Please note that there are no network edge effects due to the wrap around model. C. Results 1) Scenario S1: Obviously, we cannot satisfy all 512kbps users in such a scenario when using the conventional homogenous solutions discussed in section III-A. The results of our experiments are given in Figure 2. It shows the number of unsatised UEs in the network Z according to (17) averaged over 50 user drops versus the iteration number. The blue solid curve represents scenario S1 which is initialized by the conventional homogenous solution (given by iteration 1). As expected we observe a signicant number of unsatised users, namely 40 out of 486. Adjusting the handover offset Tc increases user satisfaction signicantly; already after 5 iterations only 5 unsatised UEs are left. Note that Z does not converge to zero for all user drops. Figure 3 shows the average handover offsets Tc versus the cell index c. The magenta circles show the initialization of scenario S1 which is simply the homogenous solution Tc = 1 (i.e. 0dB). The blue squares show the result after 25 iterations. We clearly observe that the overloaded cells, as well as sometimes their neighbors, have reduced their handover threshold by a couple of dBs on average and thereby shrink their served area. As a consequence some UEs at the edge of the overloaded cells handover to a neighbor. Those UEs are marked as red x in Figure 1 (users

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2009 proceedings

average handover offset Tc in dB

0.5 1 1.5

1000 800 600 400

28 30 29 25 27 26 22 24 23 55 5 19 57 56 52
Base Station nonHO UE HO UE 53 1000 500

31 33 32 7 9 8 11 1 6 2 16 21 20 49 54 50
0

34 36 35 10 12 38 13 3 14 43 18 17 46 51 47
500 1000

37 39 40 15 41 45 44 48 42

200
2.5

0
3 3.5 4 4.5 5 Init S1 Solution S1 = Init S2 Solution S2 10 20 30 40 50

200 400 600 800 1000

cell index c
Fig. 3. Settings of cell-specic handover offset Tc after 25 iterations (averaged over 50 user drops)

Fig. 4.

Cell layout with Scenario S2 load concentration

with a blue + have not changed their connection during the iterations). 2) Scenario S2: Now let us assume that the load concentration changes from scenario S1 to a scenario S2 where the load concentrations have been moved to the left by the intersite distance of 500m. One of the user drops is shown in Figure 4. The overloaded cell indices now are {16, 17, 18, 22, 23, 24, 31, 32, 33}. The goal is to investigate how the network can track spatial load changes. To this end, we initialize our algorithms with the Tc solution we have found for scenario S1. Note that in this case the load changes abruptly which is a quite aggressive assumption. In reality those changes are much smoother. Certainly, the average number of unsatised users in Figure 2 initially is higher than in S1 (which has been initialized with the homogenous solution Tc = 1). However, the red dashed curve converges almost as fast as the blue solid curve. The solution for the handover thresholds in 3 after 25 iterations for scenario S2 is very similar to S1, if appropriately permuted. 3) Discussion: We could consider the scenarios S1 and S2 as load situations during day and night where the local concentration would shift from a commercial/business area (S1) to a leisure area (S2). Without an adaptive SON algorithm we denitely would need to put additional BSs into the network to serve the unsatised users. It is understood that both, the investigated scenarios and the applied SON algorithm, are only simplied examples to demonstrate the mathematical framework. More realistic and more elaborate results will be presented in future work. V. C ONCLUSION We have presented a mathematical frame work for evaluation of self-optimizing wireless network. For many use cases (such as coverage and capacity optimizations, load balancing, handover optimization and energy savings) transmit power, downtilt, and connection function are important optimization parameters [9]. We have expressed the SINRs as a function of those parameters and other, deployment-related parameters.

We have provided an approximation and an iterative solution dened different target functions such as number of unsatised users in the network for CBR users. As a rst application of this framework, we have described a very simple SON algorithm which adjusts the cell-specic handover thresholds for the sake of load balancing (leaving homogenous settings for transmit power and downtilt). This algorithm has been applied to a simple scenario where the users are concentrated in particular areas in the network. The number of unsatised users was signicantly reduced compared with the homogenous solution. In addition, we have shown that load concentrations can be tracked over time. This property can save BSs in areas where the location of load concentrations change, e.g. during day and night. R EFERENCES
[1] SOCRATES, Self-optimisation and self-conguration in wireless networks, European Research Project, http://www.fp7-socrates.eu.org. [2] 3GPP, Self-conguring and self-optimizing network use cases and solutions, Technical Report TR 36.902, available at http://www.3gpp.org. [3] M. Amirijoo, R. Litjens, K. Spaey, M. D ottling, T. Jansen, N. Scully, and U. T urke, Use Cases, Requirements and Assessment Criteria for Future Self-Organising Radio Access Networks, Proc. 3rd Intl. Workshop on Self-Organizing Systems, IWSOS 08, Vienna, Austria, December 10-12, 2008, accepted [4] Next Generation Mobile Networks, Use Cases related to Self Organising Network, Overall Description, available at http://www.ngmn.org. [5] 3GPP, Physical Layer Aspects for evolved Universal Terrestrial Radio Access (UTRA), Technical Report TR 25.814, available at http://www.3gpp.org. [6] R. Hoppe, Comparison and Evaluation of Algorithms for the interpolation of 3D Antenna Patterns Based on 2D Horizontal and 2D Vertical Patterns, Study AWE Communications GmbH, V1.0, 2003. [7] 3GPP, Rened antenna model for LTE advanced evaluations, R1-083070. [8] N. Zheng, P. Michaelsen, J. Steiner, C. Rosa, J. Wigard, Antenna Tilt and Interaction with Open Loop Power Control in Homogeneous Uplink LTE Networks, IEEE International Symposium on Wireless Communication Systems 2008, Reykjavik, Iceland, October, 2008. T. Isotalo, J. Niemela, J. Lempiainen, Electrical Antenna Downtilt in UMTS Network, European Wireless Conference 2004, Los Angeles, USA, September 2004. [9] U. T urke, and M. Koonert, Advanced site conguration techniques for automatic UMTS radio network design, Proc. Vehicular Technology conference VTC 2005 Spring, vol. 3, pp. 1960-1964, Stockholm, Sweden, May 2005.

Vous aimerez peut-être aussi