a roadmap fro umts optimization to lte self optimization

IEEE Communications Magazine • June 2011172 0163-6804/11/$25.00 © 2011 IEEE

INTRODUCTION

Third-generation (3G) mobile communicationsystems have seen widespread deploymentaround the world. The uptake in High SpeedPacket Access (HSPA) subscriptions, numberingwell over 85 million at the start of 2009 [1], indi-cates that the global thirst for mobile data hasjust begun: broadband subscriptions are expect-

ed to reach 3.4 billion by 2014, and about 80percent of these consumers will use mobilebroadband.

Operators are doing business in an increas-ingly competitive environment, competing notonly with other operators, but also with newplayers and new business models. In this context,any vision for technology evolution shouldencompass an overall cost-efficient end-to-endlow latency proposition. By the end of 2004,3GPP started discussions on 3G evolutiontoward a mobile communication system thatcould take the telecom industry into the 2020s.Since then, the 3GPP has been working on twoapproaches for 3G evolution: LTE and HSPA+.The former is being designed without backwardcompatibility constraints, whereas the latter willbe able to add new technical features preservingWideband Code Division Multiple Access(WCDMA) and HSPA investments.

Besides the evolution in radio technologiesand network architectures, the conception ofhow these networks should be operated andmanaged is evolving as well. The envisaged highdensity of small sites, e.g., with the introductionof femto-cells, and the pressure to reduce costs,clearly indicate that deploying and running net-works needs to be more cost-effective. SON,aiming to configure and optimize the networkautomatically, is seen as one of the promisingareas for an operator to save operational expen-ditures. Indeed, it can be stated that SON will berequired to make the business case for LTEmore attractive. Consequently, SON has receivedmuch attention in recent years. For example, inthe framework of the Next Generation MobileNetworks (NGMN) Alliance, operator use caseshave been formulated for different stages [2]:planning (e.g., eNode-B power settings), deploy-ment (e.g., transport parameter set-up), opti-mization (e.g., radio parameter optimization thatcan be capacity, coverage or performance driv-en), and maintenance (e.g., software upgrade).Similarly, the ICT-FP7 SOCRATES projectdefined a detailed use case taxonomy in [3].3GPP has also acknowledged the importance ofSON [4]. Progress is being reflected e.g., in [5],where use cases are being developed togetherwith required functionalities, evaluation criteria

ABSTRACT

Self-Organizing Networks (SON) are current-ly being introduced by the 3rd Generation Part-nership Project (3GPP) as part of the LongTerm Evolution (LTE) system as a key driver forimproving the operation of wireless networks.Given that many challenges have been identifiedwhen moving from the SON concept to practicalimplementation, this article claims the suitabilityto gain insight into the problem by taking advan-tage of optimization mechanisms in live Univer-sal Mobile Telecommunications System (UMTS)networks. From that perspective, this article out-lines a roadmap from actual manual optimiza-tion toward the inclusion of SON concepts infuture networks. In particular, an optimizationframework is developed considering differentstages that include the collection of inputs fromdifferent sources (network counters, measure-ment reports, drive tests), the tuning parametersto achieve optimization, and the optimizationprocedure itself. Even though the proposedframework is presented and evaluated for conve-nience in a UMTS context, it is rather genericand technology-agnostic and, therefore, it can beformulated as an initial reference for LTE self-optimization use cases. The proposed frameworkis applied to a practical coverage optimizationuse case supported by data extracted from a realUMTS network in a European city, illustratingthe capability to automatically identify a cellwith sub-optimal coverage and to provide a solu-tion to the problem. The extension of the casestudy to LTE is also analyzed. Finally, as a resultof the lessons learned, the article makes a pro-jection to the LTE context by identifying the keypoints to be solved for the materialization of aself-optimization procedure.

ACCEPTED FROM OPEN CALL

Oriol Sallent, Jordi Pérez-Romero, Juan Sánchez-González, and Ramon Agustí,

Universitat Politècnica de Catalunya (UPC)

Miguel Ángel Díaz-Guerra, David Henche, and Daniel Paul, Telefónica

A Roadmap from UMTS Optimization toLTE Self-Optimization

ROMERO LAYOUT 5/19/11 9:05 AM 72

IEEE Communications Magazine • June 2011 173

and expected results, impacted specifications andinterfaces, etc. Similarly, the ICT-FP7SOCRATES project is developing in deeperdetail different use cases, such as cell outage [6].Paradigms related to SON are discussed in anearly paper in this area [7]. Additionally, in [8]the self-configuration of newly added eNode-Bsin LTE networks is discussed, and a self-opti-mization load balancing mechanism is proposedas well.

Due to its ambitious objectives, the practicalrealization of SON is seen as challenging, and itcan be anticipated that SON will continue as ahot research topic in coming years requiring fur-ther research efforts. Some of the aspects thatmust be properly covered include: to find anappropriate trade-off between the performancegains and the implementation complexity (e.g.,required signaling and measurements, computingresources), to ensure the convergence to a stablesolution within a given timing requirement or toensure the robustness to missing, wrong or cor-rupted input measurements.

The radical changes needed from currentdeployed networks such as Second-generation(2G) and 3G HSPA, mainly managed by central-ized remote operations and maintenance (O&M)applications with intensive human intervention,to future SONs (LTE, HSPA+) undoubtedlyrequire the definition of a clear roadmap whenmoving from SON’s concepts to their practicalimplementation. The impact will be on allaspects of an operator’s radio engineeringdepartment (operational procedures, O&M soft-ware tools, radio engineer’s skills, etc.). In thisrespect, early actions enabling a smooth intro-duction of the “SON culture” may be highlybeneficial on a long-term perspective. This arti-cle outlines a roadmap from actual manualUMTS optimization toward the futuristic objec-tive of a self-optimized LTE.

In particular, this article first develops aUMTS optimization framework, discussing froma practical perspective the inputs available for anautomatic optimization process as well as theoptimization procedure itself. Even though theproposed optimization framework is presentedfor convenience in a UMTS context, it is rathergeneric and technology-agnostic and, therefore,it can be formulated as the initial reference forLTE self-optimization use cases. Second, thisarticle formulates a UMTS coverage optimiza-tion case study, supported by data extractedfrom a real UMTS network in a large Europeancity. The main objective of the proposed casestudy is to gain insight into the problem from apractical point of view, in order to identify anddraw conclusions on the key points to be solvedfor the materialization of a full self-optimizationprocedure. Given the differences in radio accesstechnology, we also analyze how the coverageoptimization use case could be extended to LTE.In turn, we collect the lessons learned from theUMTS case study and project them into theLTE context. This is accompanied by the up-to-date status of SON in research and standardiza-tion fora, so that the motivation of this article tocontribute to on-going efforts from a comple-mentary perspective is enforced. Finally, wesummarize the conclusions reached.

UMTS OPTIMIZATION FRAMEWORK

UMTS network optimization turns out to be atricky and complex task because the target objec-tives in terms of coverage, capacity, and qualitytend to be contradictory (e.g., capacity may beincreased at the expense of coverage or qualityreduction) [9]. Additionally, the number of tun-able parameters in a WCDMA network is signif-icantly higher than that of a 2G Time DivisionMultiple Access (TDMA)-based network. Insuch complex scenarios, efficient network man-agement is needed in order to guarantee theQoS requirements to the subscribed users. Inrecent years, an intensive effort has been madein the field of automated optimization [10–14].Two different phases can be distinguished in theoptimization of a 3G system [15]. The first phaseis RF optimization, whose objective is to guaran-tee the required coverage, avoiding excessivepilot pollution, cell overlap or cell overshoot byoptimizing the setting of RF parameters such aspilot power, antenna down-tilt, etc. The secondphase is service parameter optimization. Thisincludes the setting of admission and congestioncontrol thresholds, maximum downlink powerper connection, events to change to compressedmode, channel switching, etc.

INPUTS FOR THE OPTIMIZATION PROCESSIn order to perform the radio optimization pro-cess, the actual network status and performanceneeds to be captured. Clearly, the more accurateand complete this picture is the more efficientthe optimization process can be expected. Inpractice, the network operator can collect diverseinformation about the network status from dif-ferent sources:

Network counters: These are measurementstaken at Node-Bs and/or Radio Network Con-trollers (RNC), such as load levels, number ofusers in soft/softer handover, etc. that are trans-ferred to the O&M module (e.g., transmittedpower level can be recorded at the Node-B andforwarded to the O&M module every 15 min-utes in the form of average values or statisticaldistributions). Typically, the manufacturer pro-vides the operator with the software tool tomanage the collection, processing (e.g., measure-ment averaging), and transference from networkelements to O&M. The type of measurementsthat a given network element must be able toperform is usually defined in the standards[16–18].

Measurement reports: These are measure-ments taken by the mobile terminals while in anactive connection and transmitted live to theRNCs (e.g., received power of the pilot channel,pilot channel energy per chip over total receivedpower density for the serving and neighboringcells, etc.). The primary usage of these reports isas input for Radio Resource Management(RRM) algorithms, such as a handover algo-rithm. Nevertheless, these measurements canalso be used as an input for the optimizationprocess. In such a case, measurement reportsneed to be managed by a software tool, able tostore and forward them to the O&M.

Drive tests: These are measurements carriedout by one or several specialized terminals able

In order to perform

the radio

optimization process,

the actual network

status and

performance needs

to be captured.

Clearly, the more

accurate and

complete this picture

is the more efficient

the optimization

process can be

expected.

ROMERO LAYOUT 5/19/11 9:05 AM Page 173

IEEE Communications Magazine • June 2011174

to record a number of parameters while follow-ing a trajectory in the field. They are equippedwith a Global Positioning System (GPS), so thatthe position where each measurement has beentaken is also recorded. For example, when somemisbehavior of the network is identified, radiooptimization engineers can trigger a drive testover the affected area in order to obtain furtherinformation about the operating conditions.Nevertheless, drive tests can also be regularlyperformed as a basis of network quality controland as another input to the network optimiza-tion process. A drive test performed with a fre-quency scanner can also be very useful becauseit is able to capture true coverage predictionswithout the influence of the specific networkconfiguration. This can be of particular interestin, for example, a 3G/2G heterogeneous scenariowhere quality problems in 3G could be automat-ically overcome by diverting traffic to the 2Gnetwork, thus hiding the 3G problems.

It is worth noting that network counters areable to capture mainly the behavior of theuplink, whereas measurement reports are mainlyable to collect the status of the downlink. Inturn, drive tests acquire similar information asmeasurement reports, with the additional advan-tage that the precise position where each mea-surement is taken is recorded. Drive testmeasurements mainly illustrate the status of thedownlink for the particular trajectory followedby the testing vehicle. Therefore, it is mostimportant to specify such trajectory so that itensures the acquisition of measurements in theareas where the real traffic is generated.

OPTIMIZATION PROCEDUREThe overall network optimization procedure isillustrated in Fig. 1 and Fig. 2, providing themost generic view where possible inputs comingfrom network counters, measurement reports,and drive test are considered. The procedureacts as a loop that continuously interacts with

the real network based on observation andactions.

At the observation stage, the optimizationprocedure collects information from the differ-ent available Key Performance Indicators (KPIs).Then the optimization procedure will containdifferent processes for each of the optimizationtargets specified based on operator policies (e.g.,avoidance of coverage holes, interference reduc-tion, reduction of overlapping between cells,etc.). For each optimization target a three-stepprocess will be executed, consisting of:• Processing the KPIs coming from the obser-

vation phase (e.g., combining several KPIs,obtaining statistical measurements such asaverages or percentiles, etc.).

• Detecting the sub-optimal operation, i.e.,that the optimization target is not properlyfulfilled, based on some criteria dependenton the specific target.

• Carrying out an optimization search to findthe proper parameter setting to solve thesub-optimal operation.

The final result of the optimization will be theadequate configuration of selected networkparameters (e.g., antenna downtilts, pilot pow-ers, RRM parameters configuration, etc.) affect-ing either the cell where the sub-optimaloperation has been found and/or its neighboringcells. This change in the configuration parame-ter(s) turns out to be the action executed overthe network.

The optimization process is formulated in theform of hypotheses tests against the sub-optimaloperation for a number of established targets.Hypotheses are reinforced by a likelihood indexthat is increased every time a given conditionevaluated over a certain KPI (or combination ofKPIs) is met. Considering that a total of Nc con-ditions c(i) i = 1,… Nc are evaluated, each con-dition will have an associated weight α idepending on how relevant the condition is forthe optimization target under consideration.

Figure 1. Network optimization loop.

Drivetests

Networkcounters

O&M

Action

Parameters setting

Optimization

Opt

imiz

atio

n ta

rget

1 KPIprocessing

Optimizationsearch

Detection ofsub-optimaloperation

KPIprocessing

Optimizationsearch

Detection ofsub-optimaloperation

Observation KPIs: Network counters

KPIs: Measurement reports

KPIs: Drive tests

Measurementreports

Opt

imiz

atio

n ta

rget

N

Drive test measure-

ments mainly illus-

trate the status of

the downlink for the

particular trajectory

followed by the test-

ing vehicle. There-

fore, it is most

important to specify

such trajectory so

that it ensures the

acquisition of mea-

surements in the

areas where the real

traffic is generated.



With the degrees of freedom provided by theweight factors αi, the value of the likelihoodindex F is then simply defined as a linear combi-nation of the evaluated conditions:

(1)

where U(c(i)) takes the value 1 if condition c(i)is fulfilled, and 0 otherwise. As a result, F is arelative value between 0 and 1. The higher the Fvalue is, the more likely that the target is notsufficiently optimized. The likelihood index Fcan be defined for every optimization targetaccording to the specific conditions established,and the weight associated with every conditioncan be fixed according to the operator’s strategyand policies.

Given that not necessarily all the requiredKPIs to detect a given sub-optimal operationmay be available (e.g., measurement reportsfrom the mobile may not be available at O&M,drive tests may not have been run for a givenarea, etc.), a reliability index R in the detectionshould be defined in accordance with how manyconditions are actually evaluated with respect tothe total number of potential conditions Ntot thatwould be evaluated if all KPIs were available:

(2)

As a final result for the detection of subopti-mal operation with respect to a given target, thecell under study will be categorized as one of thefollowing:• If F < Fmin, it is likely that the target is

optimized. Reliability of this hypothesis willbe associated with the obtained value of R.

• If F ≥ Fmin, the most likely hypothesis isthat the target is not sufficiently optimized.In this case:–If the reliability index R ≥ Rmin, then theoptimization search stage is triggered. Inthis case, if drive test measurements areavailable, it may be possible to identify spe-cific geographical areas (denoted here asclusters) where the sub-optimal behavior isdetected. This is done by grouping the drivetest positions exhibiting the particular sub-optimal behavior based on their geographi-cal distance.–If the reliability index R < Rmin, then it isdecided that there is not sufficient informa-tion about the network status to perform anautomated optimization search procedure.In this case, an alarm will be triggered indi-cating that human intervention is requiredfor the optimization of this target in thiscell.It is worth noting that the thresholds Fmin

and Rmin will be set according to the operator’sstrategy and policies. A high value for Fmin willtend to trigger optimization search procedures

only for cases where the likelihood that sub-opti-mal operation exists is high. A high value forRmin will indicate that the operator relies on theautomatic operation when many inputs from thenetwork status are available; otherwise the oper-ator prefers analyzing the situation with theassistance of a radio engineer expert.

In case the suboptimal operation has beendetected, the role of the optimization search isto propose certain changes in the configurationparameters of one or several cells (e.g., the cellwhere the suboptimal operation was detectedand/or its neighbors) in order to provide an opti-mized operation. Each target will have a numberof associated parameters, which are related tothe resulting performance for such target. Sincea given parameter may be related to severaloptimization targets fixed by the operator andmay also have a direct influence on other param-eters, the choice of parameter and its new valuemust be carefully studied, requiring some degreeof coordination between the optimization search-es associated with different targets.

Given that parameter change on the live net-work is a very sensitive issue, optimization algo-rithms will usually require the support of a moduledevoted to estimate the expected impact of agiven parameter change (e.g., a planning tool, asystem level simulator, a software tool fed withreal data measurements, etc.) prior to the modifi-cation in the live network. In practice, since theautomatization of this stage may involve the inte-gration of very diverse software tools, some humanintervention may be required in the meantime.

The optimization search strategy can be basedon different approaches depending on the spe-cific target and the parameter(s) to optimize.Classical optimization methods, such as e.g.,genetic algorithms, linear programming, particleswarm algorithms, etc., or other heuristicapproaches, can be considered. Such methodscould be adequately extended to address theproblem from the multi-objective optimizationperspective, jointly taking into account the dif-

Ri

i

N

ii

N

c

tot= =

=

∑

∑

α

α

1

1

F

U c iii

N

ii

N

c

c= =

=

∑

∑

α

α

( ( ))1

1

Figure 2. Processes in the network optimization algorithm.

Param k cell Y

Param k cell Z

Param k cell X

Observation

Opt

imiz

atio

n ta

rget

1

Networkcounters

Drivetests

Measurementreports

KPI 1

KPIprocessing

KPI 2 KPI 3 KPI k KPI n

Optimizationsearch

Detection of sub-optimal operation

Opt

imiz

atio

n ta

rget

N

KPIprocessing

Optimizationsearch

Detection of sub-optimal operation

Action

Optim

ization

Param 2 cell Y

Param 2 cell Z

Param 2 cell XParam 1 cell Y

Param 1 cell Z

Param 1 cell X



ferent optimization targets [19, 20]. Further-more, learning mechanisms can also be appliedat this stage.

CASE STUDY:UMTS COVERAGE OPTIMIZATION

In order to illustrate the framework presented inthe previous section, the procedure for the opti-mization of UMTS coverage is presented. Theconsidered scenario is shown in Fig. 6. It consistsof six tri-sectorial cells in an urban area of amajor European city. Each cell is identified asCell_x_y where x corresponds to the Node-Bidentifier and y is the sector of this Node-B.

Table 1 shows the KPIs that can be useful forthe optimization of the UMTS coverage in a givencell, the corresponding conditions to detect sub-optimal behavior, and the source from which eachKPI can be obtained. As shown, these conditionsare based on statistics related to power measure-ments, handovers, or propagation delay measure-ments. Some hints about the rationale to considerthese specific KPIs are given in the following:

•In general, poor coverage will be associatedwith low values of Received Signal Code Power(RSCP) of the Common Pilot Channel (CPICH)and Received Signal Strength Indicator (RSSI)at the Node-B.

•High uplink transmitted power may indicateexcessive propagation losses and, consequently,poor coverage.

•A high ratio of handovers from UMTS toGlobal System for Mobile communications(GSM) cells may indicate UMTS coverage prob-lems, provided that the applied network selec-tion strategy is such that it favors continuity overUMTS once attached to this network. Besides,coverage problems may prevent the Active Set(AS) and Monitored Set (MS) lists from beingadequately updated (e.g., due to problems in thesignaling procedures for adding/removing cellsfrom the Active Set, derived from the fact thatsome messages can be lost due to coverage prob-lems).

•Finally, the statistical distribution of thepropagation delay of specific signals transmittedfrom mobiles to cell is considered to be a valu-able input because it reflects the statistical distri-bution of mobiles-to-cell distances whenconnected to the cell. More specifically, in casethat users tend to connect only at distancesmuch below the planned cell coverage radius,this might reflect suboptimal coverage at the celledge.

Table 1 also shows the configuration parame-ters that are needed in the suboptimal operationdetection process, which in turn will usually bethe tunable parameters to be considered in theoptimization search stage.

The algorithm for the detection of sub-opti-mal coverage operation is presented in Fig. 3,making use of the conditions presented in Table1. It is worth noting that the RSCP condition isevaluated not only for the cell under study but

Table 1. Considered KPIs and tunable parameters.

Considered KPIs Criteria/Condition Source

RSCP (Received Signal Code Power fromthe CPICH channel)

Prob(RSCP<RSCP*)>ThRSCP(%) Drive Test/Measurement Report

Uplink transmission power PT Prob(PT>PT*)>ThPT(%) Drive Test/Measurement Report

Active Set and Monitored SetThe Active Set is not full (AS < ASmax) and there exists acell in the Monitored Set that should be in Active Set.

Drive Test/Measurement Report

Number of handover attempts to GSMrelative to cell load

NumberHO GSM

Cell throughputThGSM

_

_> Network counters

Uplink RSSI (Received Signal StrengthIndicator at the Node-B)

Prob(RSSI<RSSI*)>ThRSSI(%) Network counters

Statistical distribution of the propagationdelay between mobiles and base station.

Prob(Propdelay<P*)>Thpropdelay(%) Network counters

Network configuration parameters/Tunable parameters

Positions of the different nodes B in the network.

Antenna azimuth of the different cells.

Neighbor lists.

CPICH transmission power in the different cells.

Antenna downtilt of the different cells



also for the neighboring cells, so as to increasethe likelihood of sub-optimal coverage in caseneither the cell under study nor any of its neigh-boring cells are detected with a high enoughRSCP. According to these conditions and itscorresponding weights αi shown in Fig. 3, thealgorithm determines the reliability index R andthe likelihood index F in the detection of cover-age problems. If the likelihood and reliabilityindex are higher than certain thresholds (e.g.,Rmin = 0.4 and Fmin = 0.6 are chosen in thiscase study), coverage problems in the cell understudy are assumed and the optimization search isactivated.

In order to gain some insight into the net-work status in the considered scenario, someexamples of measurements used for the detec-tion of sub-optimal cell coverage are presentedand discussed in the following. On the one hand,Fig. 4 shows the Cumulative Distribution Func-tion (CDF) of the RSCP for different cells wheneach considered cell is detected either in theactive set if the mobile is in connected mode oras best server if the mobile is in idle mode.While the RSCP in Cell_6_2 is below –85dBmfor around 90 percent of the samples available in

the cell, Cell_6_1 exhibits only 10 percent belowthat threshold and Cell_6_3 exhibits that almostall samples are above the threshold. Therefore,the observation of this plot may point out cover-age problems within Cell_6_2. On the otherhand, the correlation between the uplink trans-mitted power and the RSCP is presented in Fig. 5. A cell with coverage problems may have ahigh percentage of measurements with a highuplink transmitted power and a low RSCP (e.g.,Cell_6_2). In terms of the algorithm for thedetection of sub-optimal coverage, this linkagehas been captured by including one condition onRSCP and another condition on the uplink trans-mitted power. If both are fulfilled (i.e., lowRSCP in downlink and high transmitted powerin uplink are observed), the likelihood index Fthat the cell is affected by coverage problemsincreases. On the contrary, cells like Cell_6_3,with low values of uplink transmitted power andhigh RSCP values, are not expected to have cov-erage problems.

In order to assess the performance of theautomated coverage optimization process in thisreal scenario, the algorithm presented in Fig. 3 isexecuted. Considering the weights for each con-

Figure 3. Detection of sub-optimal operation algorithm.

F>Fmin

Computelikelihood index F

and reliabilityindex R.

Go to optimizationsearch algorithm.

Clusteridentification

N

Y

Cell without coverageproblems with likelihood

F and reliability R

Too low reliability.Manual tuning is

required.

N

N

Y

Y

Cell with coverageproblems with likelihood F

and reliability R

Evaluate RSSI condition(weight α5=2)

Evaluate RSSI propagationdelay condition(weight α6=2)

Evaluate handover to GSMcondition (weight α7=1)

Networkcounters

Measurementreports

Evaluate RSCP condition(weight α4=3)

Evaluate AS and MS condition(weight α1=1)

Evaluate uplink transmittedpower condition(weight α2=1)

Evaluate if the RSCP condition isfulfilled for all neighbour cells

(weight α3=6)

Drive testmeasurements

Evaluate RSCP condition(weight α8=3)

Evaluate AS and MS condition(weight α11=1)

Evaluate uplink transmittedpower condition(weight α10=1)

Evaluate if the RSCP condition isfulfilled for all neighbour cells

(weight α9=6)

R>Rmin

Drivetest measurements

available

The statistical

distribution of the

propagation delay

of specific signals

transmitted from

mobiles to cell is

considered to be a

valuable input

because it reflects

the statistical

distribution of

mobiles-to-cell

distances when

connected to

the cell.



dition indicated in Fig. 3 and the definitions oflikelihood index and reliability index indicated inEqs. 1 and 2, the result obtained with the avail-able data for Cell_6_2 is F = 0.93 and R = 0.48.Therefore, since F ≥ Fmin = 0.6, the most likelyhypothesis is that the coverage is not sufficientlyoptimized. Besides, since R ≥ Rmin = 0.4, theoptimization search stage is triggered. Further-more, since drive test measurements are avail-able in this case, the algorithm is able to identifythree clusters where the sub-optimal coverage islocalized (the identified clusters are shown bycircles in Fig. 6).

The optimization search considered here isbased on readjusting the antenna downtilt of thecell under study (or some neighbors) so as toimprove the coverage in the cluster locationswhile maintaining proper coverage in the rest ofthe cell. From a practical perspective, electricaltilt would be preferred, while mechanical tiltshould be limited to a certain maximum in orderto avoid deformation on the main radiation lobe.

The procedure makes use of the knowledgeabout the antenna radiation pattern of the dif-ferent cells together with the measurementsavailable from the drive tests and the informa-tion about the positions of the different clustersidentified in the detection step. The procedure

first considers downtilt changes for the cellwhere the problem was detected (Cell 6_2 in thisexample). An iterative search is carried out byshifting the downtilt in steps of Δ degrees up toa maximum Δmax. For each iteration and foreach sample corresponding to each identifiedcluster an estimation of the resulting RSCPvalue is computed using the antenna radiationpattern. As a result, an estimation of the per-centage of samples with expected RSCP abovethreshold RSCP* is obtained for each consid-ered downtilt. If a downtilt providing sufficientimprovement is found, then a recommendationfor parameter change is drawn. Otherwise, thealgorithm concludes that it is not possible tooptimize the coverage of this cell through down-tilting on the same cell and a solution searchover neighboring cells is triggered. In that case,a similar procedure is executed by consideringdifferent downtilts in neighboring cells. Thealgorithm will be able to provide a recommenda-tion about downtilt change in a neighboring cellprovided that the neighbor can significantlyimprove the coverage over the affected clusterswhilw still maintaining proper RSCP levels overits own coverage area.

Applying the optimization search algorithmin the analyzed scenario, it is found that nodowntilt change on Cell_6_2 would provide suffi-cient RSCP improvement over the cell. There-fore, the solution search was extended toneighboring cells. In this case, it was found thata downtilt change from 11° to 4° in Cell_4_1would improve the overall situation.

In order to understand the outcome of theoptimization algorithm, Fig. 7 shows, on the onehand, the CDF for the RSCP from Cell 6_2 cor-responding to the samples taken along the drivetest segment marked in Fig. 6. It can be clearlyobserved that poor coverage is provided overCell 6_2, since almost 85 percent of the samplesare below –88 dBm. In a closer view, measureddata over the drive test (not shown in the figuresfor the sake of brevity) reveal that Cell 6_2 is notdetected at all in Cluster 1, while it is detectedwith an average value of –92 dBm in Cluster 2and –93 dBm in Cluster 3. In turn, with respectto Cell 4_1, this becomes the best server alongClusters 1, 2 and 3 (which are within the theo-retical coverage area from Cell 6_2), although itis received with very low values of RSCP, inmost of the cases below -88 dBm.

Figure 7 shows, on the other hand, the CDFfor the RSCP received from Cell 4_1 along thedrive test for downtilt values of 4° and 11°.Thanks to the downtilt change from 11° to 4° theprobability of low RSCP values is reduced, whichillustrates that RSCPs from Cell 4_1 improve atClusters 1, 2 and 3. This is at the expense ofreducing the probability of high RSCP, althoughthis is not impacting negatively on the coverageprovided by Cell 4_1 along its own theoreticalcoverage area. After the downtilt adjustment inCell_4_1, the likelihood index and the reliabilityindex for Cell_6_2 are calculated again for thenew situation. F = 0.46 and R = 0.48 areobtained. F is reduced because the RSCP condi-tion for neighbor cells (Fig. 3) is not fulfilled inthe resulting scenario. Given that F < Fmin, theoutput of the algorithm is that coverage is suffi-

Figure 4. Cumulative distribution function of the RSCP.

RSCP (dBm)-105

0.2

0

0.1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-100 -95 -90 -85 -80 -75 -70 -65 -60 -55

Cell_6_1Cell_6_2Cell_6_3

Figure 5. Correlation between RSCP and uplink transmit power.

Uplink transmitted power (dBm)-35-40

-90

-100

RSC

P (d

Bm)

-80

-70

-60

-50

-30 -25 -20 -15 -10 -5 0 5 10

Cell_6_2Cell_6_3



ciently optimized and, therefore, the coverageproblem has been solved with this action.

In conclusion, this example has shown a prac-tical case where a cell provides poor coverage,which is eventually overcome by increasing theoverlapping from a neighboring cell. Indeed, inorder to verify the conclusions reached in thisexercise, a closer analysis in-the-field was con-ducted and it was found that there were somehigh buildings obstructing the antenna ofCell_6_2 in the direction of the clusters, so it wasnot possible to improve the coverage in thoseareas by acting on the downtilt of that particularcell.

EXTENSION TO LTEIn order to illustrate that the actual work onUMTS optimization can also provide valuablebackground for LTE, this section discusses howthe coverage optimization case study presentedfor UMTS could be extended to the LTE case.In this way, the usefulness of the presentedframework as a basis for LTE is emphasized,even though the type of optimization targets tobe considered and the specific KPIs can be dif-ferent as they are technology-specific. Similarly,there would be differences in terms of systemarchitecture, so that while in UMTS many of theKPIs can be available at the RNC, in LTE mostof the KPIs will be available at the eNode-Bs.

Table 2 presents the extension of the KPIsand tunable parameters from the UMTS case tothe LTE case. As can be observed, the condi-tions to detect the sub-optimal coverage are alsobased on statistics related to power measure-ments, handovers, and propagation delay. Never-theless, there is some specificity as indicated inthe following:

•The corresponding measurement that canbe equivalent to the RSCP in UMTS is the Ref-erence Signal Received Power (RSRP), whichcorresponds to the measured power by themobile terminal at the specific reference signalstransmitted by the eNode-B [21].

•As for the uplink case, given that in LTEthere also exists a power control scheme, theuplink transmitted power in either PhysicalUplink Control Channel (PUCCH) or PhysicalUplink Shared Channel (PUSCH) that can becollected through drive tests or measurementreports, can also be used to identify poor cover-age situations whenever the transmit power isabove some threshold, reflecting excessive prop-agation losses. Note also here that the RSSI inthe uplink that was considered in UMTS is nolonger considered in the LTE case, given thatthis measurement is actually not collected by theeNode-B [21].

•The condition regarding the number of han-dover attempts to GSM has been reformulatedto also include the handovers to UMTS, toreflect that a large number of handovers to thesetechnologies could be an indication of LTE cov-erage problems.

•The condition regarding the AS and MSconfiguration used in UMTS is no longer consid-ered in LTE, given that no soft handover is used.Instead, the considered condition corresponds tothe monitoring of the RSRP in the neighboringcells. In particular, if the best cell is not actually

the serving cell, this could reflect problems inthe signaling associated with the handover pro-cedure that can be due to coverage problems.

•As for the condition related to the propaga-tion delay, it can also be used in the context ofLTE. While in UMTS the estimation of thismetric is implementation-dependent and couldbe obtained from the delays in the randomaccess, the availability of time advance measure-ments in LTE makes even more straightforwardthe estimation of the propagation delay, with aresolution of 0.52 μs [22].

Based on the above conditions, the algorithmfor the detection of sub-optimal coverage inLTE would be very similar to the one presentedin Fig. 3 for UMTS with the corresponding mod-ifications in accordance with the KPIs and condi-

Figure 6. Scenario and location of the identified clusters.

Cell_1_1

Cell_6_1

cluster3

cluster2

cluster1

Cell_6_2Cell_6_3

Cell_5_1

Cell_5_2

Cell_4_1Cell_5_3

Cell_4_3Cell_4_2

Cell_3_3Cell_3_2

Cell_3_1

Cell_2_3

Cell_2_2

Cell_2_1

Figure 7. CPICH RSCP before and after the downtilt correction.

RSCP (dBm)

Cumulative distribution function of the RSCP

-105

0.1

0

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-100 -95 -90 -85 -80 -75 -70 -65 -60 -55 -50

Cell_4_1 with tilt 11 degreesCell_4_1 with tilt 4 degreesCell_6_2



tions from Table 2. Similarly, the concepts ofreliability R and likelihood index F would alsobe applicable, although perhaps with differentvalues depending on the weights to be set foreach condition.

INSIGHTS FORLTE SELF-OPTIMIZATION

SON is introduced as part of the 3GPP LTE as akey driver for improving O&M. SON concepts areintroduced in LTE starting from the first releaseof the technology (Release 8), and expanding inscope with subsequent releases. The progressiveinclusion of standardized SON features certainlyreflects the expected LTE network evolutionstages as a function of time. For example, Release8 includes functions covering different aspects ofthe eNode-B self-configuration (automatic inven-tory, automatic software download, automaticneighbor relation, automatic physical Cell Identifi-er (Cell ID) assignment [23]). The objective ofself-configuration is to provide “plug and play”functionality in the planning, integration, and con-figuration of new eNode-Bs. In turn, Release 9will provide SON functionalities addressing opti-mization. This phased approach, targeting auto-mated optimization for more maturing networkstages, seems fairly reasonable given the ambitiousobjectives of a full SON. Some of the considereduse cases in Release 9 are:

•Mobility robustness optimization, encom-passing the automated optimization of parame-ters affecting active mode and idle modehandovers to ensure good end-user quality andperformance.

•Load balancing optimization, in order tointelligently spread user traffic across the sys-tem’s radio resources as necessary in order toprovide quality end-user experience and perfor-mance, while simultaneously optimizing systemcapacity.

•Coverage and capacity optimization, whichis related to the case study presented earlier.While the goal is the same regardless of radiotechnology, the specific algorithms and parame-ters vary with technology. This has been illus-trated earlier, which has outlined a possiblesolution.

Release 9 also includes elements to increaseenergy savings, UE reporting functionality tominimize the amount of drive tests, etc. Never-theless, it is worth emphasizing that SON-relat-ed functionalities will continue to expandthrough the subsequent releases of the LTEstandard.

In this context, where a full LTE-SON can beset as the skyline ultimate objective, it is impor-tant to devise a roadmap starting from actualcommonly-used manual-based UMTS optimiza-tion. The practical experience gained so far, asreflected in this article, has brought to someobservations, sometimes over-sought when tack-ling the problem from a theoretical/simulationperspective, as well as to the identification ofsome key open points to be tackled in this evolu-tionary process toward LTE-SON. Similarly,lessons learned can also be of help and provide acomplementary and reinforcing approach, partic-ularly at the time that LTE-SON is still takingfull shape and while it cannot benefit from livenetwork experiences. In this respect, it is worthmentioning:

Table 2. Considered KPIs and tunable parameters for the LTE case.

Considered KPIs Criteria/Condition Source

RSRP (Reference Signal Received Power) ofserving cell

Prob(RSRP<RSRP*)>ThRSRP(%) Drive Test/Measurement Report

Uplink transmission power PT Prob(PT>PT*)>ThPT(%) Drive Test/Measurement Report

RSRP of neighboring cells The best cell is not the serving cell Drive Test/Measurement Report

Number of handover attempts toGSM/UMTS relative to cell load

NumberHO GSM NumberHO UMTS

Cell throughputThHO

_ _

_

+> Network counters

Statistical distribution of the propagationdelay between mobiles and base station.

Prob(Propdelay<P*)>Thpropdelay(%) Network counters

Network configuration parameters/Tunable parameters

Positions of the different e-nodes B in the network.

Antenna azimuth of the different cells.

Neighbor lists.

Transmit power in the different cells.

Antenna downtilt of the different cells



•Given that the practical implementation ofSON concepts puts strong requirements on thecapabilities offered by O&M tools and impactsall aspects of an operator’s radio engineeringdepartment, early actions enabling a smoothintroduction of the “SON culture” may be highlybeneficial on a long-term perspective.

•The optimization framework presented ear-lier can be considered as a reference startingpoint for LTE, given that the main principleshighlighted are technology-independent. In par-ticular, the different sources to acquire the net-work status, the network optimization loop andthe process in the network optimization algo-rithm are, indeed, of rather general applicability.

•Related to SON architectures, guiding prin-ciples and operational procedures, again theframework proposed in this article, which hasbeen derived from a practical perspective, can beuseful for the consolidation of SON in the frame-work of LTE [24].

•Given the uncertainty about the networkstatus, outputs from SON algorithms requireassociated likelihood/reliability indicators.Therefore, this approach as presented in SectionII with F and R parameters tightly coupled withthe detection of sub-optimal behavior could alsobe applicable in the LTE context.

•Given that changing the network configura-tion/parameterization is a critical action fromthe operator’s perspective, the automated opti-mization procedure is likely to be introduced ina step-by-step approach. While automatic sub-optimal behavior detection mechanisms can beconsidered mature enough (as reflected in thepresented UMTS case study), the automatizationof optimization search and parameter change isseen at this stage, and even for UMTS networks,far from mature. The confidence in such auto-matic procedures as required by the operatorneeds to be significantly enhanced, as discussedin the points below.

•Optimization search procedures need to bespecifically developed for each possible parame-ter. Criteria to anticipate the most suitableparameter to cope with a given sub-optimaloperation need to be established.

•Supporting tools and/or models for the opti-mization search stage require further research.The estimation of the impact of a given parame-ter change on the live network as part of theoptimization procedure is a critical point.

•The ultimate view of SON, where the jointself-optimization of several targets is performed,may raise complex interactions. The influence ofthe execution of a stand-alone optimization tar-get on the rest of the targets needs to be investi-gated.

•In the best case, SON will be able to auto-matically detect a sub-optimal operation andprovide a corrective action. However, in somecases the SON algorithm may fail in detectingthe problem or may be unable to provide a rec-ommendation for a corrective action.

•Besides, there are incipient market solutionsfor UMTS optimization that allow KPIs’ delocal-ization. This promising feature, which is expect-ed to mature in a LTE context, may reduce theneed for drive tests in the future and pave theway for a full self-optimization framework.

CONCLUSION AND FUTURE WORK

SON is introduced as part of the 3GPP LTE as akey driver for improving O&M. However, giventhat many challenges are identified when movingfrom the SON concept to practical implementa-tion, this article has claimed the suitability togain insight into the problem by taking advan-tage of optimization mechanisms in live UMTSnetworks.

The article has developed an optimizationframework considering different stages: inputsfrom different possible sources (network counters,measurement reports, drive tests), tuning parame-ters to achieve optimization and the optimizationprocedure itself (including KPIs processing, sub-optimal operation detection for a given target andoptimization search). The optimization processhas been formulated in the form of hypothesestests against the sub-optimal operation for a num-ber of established targets. Hypotheses are rein-forced by likelihood and reliability indexes, so thatthe number and relevance of diverse KPIs withrespect to the considered target can be weighted.

The optimization framework has been appliedto a UMTS coverage optimization use case in alarge European city. The algorithm has allowedthe automated identification of a cell with sub-optimal coverage and provided a solution to bet-ter optimize its operation. The extension to LTEhas been analyzed, identifying a number of speci-ficities to be considered due to technology change.

Finally, this article has provided some insightfor LTE self-optimization. On one side, it hasbeen discussed that the presented optimizationframework could also be taken as a reference inthe LTE context. On the other side, lessonslearned from the UMTS case study have enabledthe identification of a number of key issues to besolved in the evolutionary path toward LTE-SON. Besides, it has been highlighted that astep-by-step introduction of SON can be soundin practice, implementing self-detection withmanual self-optimization in a first stage and pro-gressively moving toward automatization of theoptimization search.

ACKNOWLEDGMENTSThis work has been supported by Telefónica.This article reflects the authors’ views and doesnot necessarily represent the view of Telefónica.

REFERENCES[1] UMTS Forum, Annual Report 2008 and Directions for

2009[2] NGMN Alliance Deliverable, “NGMN Use Cases Related

to Self Organizing Network, Overall Description,” May2007.

[3] INFSO-ICT-216284 SOCRATES project, “D2.1: Use Casesfor Self-Organizing Networks,” March 2008.

[4] 3GPP TR 32.816, “Study on Management of EvolvedUniversal Terrestrial Radio Access Network (E-UTRAN)and Evolved Packet Core (EPC)”

[5] 3GPP TR 36.902, “Evolved Universal Terrestrial RadioAccess Network (E-UTRAN); Self-Configuring and Self-Optimizing Network Use Cases and Solutions.”

[6] M. Amirijoo et al., “Cell Outage Management in LTENetworks,” 6th Int’l. Symp. Wireless Commun. Systems(ISWCS 2009), Italy, Sept. 2009, pp. 600–04.

[7] C. Prehofer and C. Bettstetter, “Self-Organization inCommunication Networks: Princip les and DesignParadigms,” IEEE Commun. Mag., vol. 43, no. 7, July,2005, pp. 78–85.

There are incipient

market solutions for

UMTS optimization

that allow KPIs’

delocalization. This

promising feature,

which is expected to

mature in a LTE

context, may reduce

the need for drive

tests in the future

and pave the way

for a full self-opti-

mization framework.



[8] H. Hu et al., “Self-Configuration and Self-Optimizationfor LTE Networks,” IEEE Commun. Mag., vol. 48, no. 2,Feb., 2010, pp. 94–100.

[9] J. Pérez-Romero et al., Radio Resource ManagementStrategies in UMTS, John Wiley & Sons, 2005.

[10] I. Siomina, P. Värbrand, and D. Yuan, “AutomatedOptimization of Service Coverage and Base StationAntenna Configuration in UMTS Networks,” IEEE Wire-less Commun., vol. 13, no. 6, Dec. 2006, pp. 16–25.

[11] J. Zhang et al., “Mathematical Modeling and Compar-isons of Four Heuristic Optimization Algorithms forWCDMA Radio Network Planning,” Proc. 8th IEEE Int’l.Conf. Transparent Opt. Networks, June 2006, pp.253–57.

[12] S. Ben Jamaa et al., “Steered Optimization Strategy forAutomatic Cell Planning of UMTS Networks,” IEEE 61stVehic. Tech. Conf. (VTC 2005 Spring), June, 2005, pp.359–62.

[13] S. Allen et al., “Smart Cell Planning and Optimizationfor UMTS,” 5th IEE Int’l. Conf. 3G Mobile Commun.Technologies (3G 2004), 2004, pp. 34–38.

[14] J. Laiho et al., Eds., Radio Network Planning and Opti-mization for UMTS, 2nd Edition, John Wiley & Sons.

[15] C. Chevalier et al., WCDMA Deployment Handbook:Planning and Optimization Aspects, John Wiley & Sons,2006.

[16] J. Sánchez-González et al., “A New RF Methodologyfor Failure Detection in UMTS Networks,” IEEE/IFIP Net-work Operations and Management Symp., Salvador daBahia, Brazil, April, 2008, pp. 718–21.

[17] 3GPP TS 25.215 “Physical Layer Measurements (FDD)(Release 5).”

[18] 3GPP TS 25.331 “Radio Resource Control (RRC)(Release 5).”

[19] D.E. Goldberg, Genetic Algorithms in Search, Opti-mization & Machine Learning, Addison Wesley, 1989.

[20] C.M. Fonseca and P. J. Fleming “Multiobjective Opti-mization and Multiple Constraint Handling with Evolu-tionary Algorithms — Part I: A Unified Formulation,”IEEE Trans. Systems, Man and Cybernetics — Part A:System and Humans, vol. 28, no. 1, Jan. 1998, pp.26–37.

[21] 3GPP TS 25.214 v8.7.0, “Evolved Universal TerrestrialRadio Access (E-UTRA); Physical layer — Measurements(Release 8)”

[22] 3GPP TS 25.213 v8.1.0, “Evolved Universal TerrestrialRadio Access (E-UTRA); Physical layer — Procedures(Release 8)”

[23] 3G Americas White Paper, The benefits of SON in LTE,December 2009

[24] T. Jansen et al., “Embedding Multiple Self-Organiza-tion Functionalities in Future Radio Access Networks,”69th IEEE Vehic. Tech. Conf. (VTC 2009 Spring) ,Barcelona, Apr. 2009.

BIOGRAPHIESORIOL SALLENT ([email protected]) is Associate Professor atthe Universitat Politècnica de Catalunya. His research inter-ests are in the field of mobile communication systems,especially radio resource and spectrum management forcognitive heterogeneous wireless networks. He has pub-lished more than 150 papers in international journals andconferences. He has participated in several research pro-jects of the 5th 6th and 7th Framework Program of theEuropean Commission and served as a consultant for anumber of private companies.

JORDI PÉREZ-ROMERO is currently associate professor at theDept. of Signal Theory and Communications of the Univer-sitat Politècnica de Catalunya (UPC) in Barcelona, Spain. Hereceived the Telecommunications Engineering degree andthe Ph.D. from the same university in 1997 and 2001. Hisresearch interests are in the field of mobile communicationsystems, especially packet radio techniques, radio resource

and QoS management, heterogeneous wireless networksand cognitive networks. He has been involved in differentEuropean Projects as well as in projects for private compa-nies. He has published papers in international journals andconferences and has co-authored one book on mobilecommunications. He is associate editor of IEEE VehicularTechnology Magazine and Eurasip Journal on WirelessCommunications Networks.

JUAN SANCHEZ-GONZALEZ received the Engineer of Telecom-munications and Ph.D. degrees from the UniversitatPolitècnica de Catalunya (UPC), Barcelona, in 2002 and2007, respectively. He is currently assistant professor in thedepartment of Signal Theory and Communications in UPC.He has participated in projects funded by the EuropeanUnion as well as in projects for private companies. Hisresearch interests include radio resource and QoS manage-ment for wireless communication networks, optimizationof cellular systems and cognitive networks.

RAMON AGUSTI received the Engineer of Telecommunica-tions degree from the Universidad Politécnica de Madrid,Spain, in 1973, and the Ph.D. degree from the UniversitatPolitècnica de Catalunya (UPC), Spain, 1978. He becameFull Professor of the Department of Signal Theory andCommunications (UPC) in 1987. After graduation he wasworking in the field of digital communications with par-t icular emphasis on transmission and developmentaspects in fixed digital radio, both radio relay and mobilecommunications. For the last fifteen years he has beenmainly concerned with aspects related to radio resourcemanagement in mobile communications. He has pub-lished about two hundred papers in these areas and co-authored three books. He participated in the Europeanprogram COST 231 and in the COST 259 as Spanish rep-resentative delegate. He has also participated in theRACE, ACTS and IST European research programs as wellas in many private and public funded projects. He receivedthe Catalonia Engineer of the year prize in 1998 and theNarcis Monturiol Medal issued by the Government of Cat-alonia in 2002 for his research contributions to themobile communications field. He is a Member of theSpanish Engineering Academy.

MIGUEL ÁNGEL DIAZ-GUERRA received the TelecommunicationsEngineering degree from the Universidad Politécnica deMadrid, Spain, in 1992. In 1994 he joined Telefónica. Hehas been working in Radio Planning and Radio Optimiza-tion during 16 years. He has collaborated in differentresearch projects with University. He has participated ininternational projects inside Telefónica Group in Latin-American and Europe. He currently is Access Network Engi-neering Manager.

DAVID HENCHE received the Telecommunications Engineeringdegree from the Universidad Politécnica de Cataluña,Spain, in 1993. In 1995 he joined Telefónica, where he cur-rently is Radio Access Network Engineer working in theplanning and deploying area. His main interests are in thefield of internal and external interference detection andremoval to increase network performance, and optimaldeployment of mobile networks for current and new futuredata services.

DANIEL PAUL received the Telecommunications Engineeringdegree from the Universidad Politécnica de Madrid, Spain,in 1989. In 2001 he joined Telefónica, where he currently isAccess Network Engineering Director. Prior to that, he waswith Motorola, where he has been working from 1990 par-ticipating in the initial development of the GSM systemand participating in the GSM deployment in a number ofoperators in Europe. His main interests are in the field ofnew techniques of radio access network optimization andthe optimal deployment of MBB networks adjusted for thenew future data services.

A step-by-step

introduction of SON

can be sound in

practice, implement-

ing self-detection

with manual

self-optimization in a

first stage and

progressively moving

toward automatiza-

tion of the

optimization search.


a roadmap fro umts optimization to lte self optimization

Documents