A study on machine learning and regression based models...

MotivationProblem statement

Simulation scenarioResults of performance prediction

Conclusions

A study on machine learning and regressionbased models for performance estimation of

LTE HetNets

B. Bojovic1, E. Meshkova2, N. Baldo1,J. Riihijärvi2 and M. Petrova2

1Centre Tecnològic de Telecomunicacions de Catalunya (CTTC)Castelldefels, Spain

2Institute for Networked Systems, RWTH Aachen University,Aachen, Germany

ACROPOLIS 3rd Annual Workshop - London, 2013Bojovic, Meshkova, Baldo, Riihijärvi, Petrova Machine Learning and regression for performance estimation



Conclusions

Outline

1 MotivationFemtocell network deploymentsInter-cell interference awarenessDynamic frequency allocation for LTE system

2 Problem statementLTE physical layerProposed prediction methods and covariates

3 Simulation scenario

4 Results of performance predictionIssues in predicting performanceMachine learning techniques for performance estimation

Bojovic, Meshkova, Baldo, Riihijärvi, Petrova Machine Learning and regression for performance estimation



Conclusions

Femtocell network deploymentsInter-cell interference awarenessDynamic frequency allocation for LTE system

Femtocell network deploymentsGoals and issues.

Recently, there was much interest in femtocellsdeployments for different network technologies.The most of attention was in shown for WCDMAtechnology in 3G network deploymentsThe same concept could be applied also to more recenttechnologies such as WiMAX, LTE,..The goal is to improve the performance of the mobilenetwork by:

extending indoor coverageachieving higher data rate for the indoor and short rangeoutdoor communications

The main challenge is how to manage the availablespectrum resources




Conclusions


Femtocell network deployments.Goals and issues.

One approach could be to divide the available spectruminto several frequency bands and then to assign to eachfemtocell a different oneIn previous case, the drawback are: low or no frequencyreuse, not feasible for dense deployments and requiresmuch maintenanceAnother approach, which is typically considered in recenttechnologies, is to share available frequency bands amongfemtocellsIn the previous case the main issue is the interferencebetween femtocells, between femtocell and macro cellsConsequence: the network performance degradation andunfairness among users




Conclusions


Inter-cell interference awarenessPrevious work

Many studies have identified significant performance gainsin the systems that are aware of the inter-cell interferenceObjective: to maximize the system performance and toprovide fairness among users by minimizing the inter-cellinterferenceSome of proposed techniques are based on: power controlschemes, static and adaptive fractional frequency reuseschemes, cancellation techniques, intelligent scheduling,network power coordination,...




Conclusions


Dynamic frequency allocation for LTE systemOur approach

Our approach it to improve the system performance byenabling the dynamic frequency allocationWe consider an LTE system, which is designed to be usedwith frequency reuse factor of 1LTE technology includes advanced features such asOFDMA, SC-FDMA, AMC, dynamic MAC scheduling andHARQ,..More difficult than in previous technologies to predict thesystem capacity for specified system configurationOur approach is to apply machine learning and regressionanalysis for system capacity estimation, that will enableefficient dynamic frequency allocation




Conclusions

LTE physical layerProposed prediction methods and covariates

LTE physical layerConfigurable network parameters

The OFDMA multicarrier technology provides flexiblemultiple-access scheme that allows different spectrumbandwidths to be utilized without changing thefundamental system parameters or equipment designIn this work we consider:

1 the carrier frequency fc ; and2 he bandwidth B

In frequency domain resources are grouped in units of 12subcarriers that are called resource blocks centeredaround fc and occupy 180 kHz in frequency domain




Conclusions


Configuring carrier frequency and bandwidth.An example

The channel raster is 100 kHz for all bandsLTE eNB operates using a set of B contiguous resourceblocks(RB); the allowed values for B are 6, 15, 25, 50, 75,100 RBs ( 1.4 , 3, 5, 10, 15, 20 MHz)

EARFC

Nfc1

EARFC

N

fc2

600kHz

Femtocell 1

Femtocell 2

180kHz

fb1 - fo2

fo1

fb2fo2

fb1




Conclusions


The optimization problem in broader context

The quality of the achieved performance depends onvarious factors: amount of available inputs, their usability inthe context of particular models and optimization methods

Inputs

Models &Decisions

Parameters & KPIsTopologyTime granularitySpatial samplingPer user/femto reportingSystem information/dependencies

Abstractions/simpli�cationsMethodsOnline adaptation and/or trainingRequired parameteres and prior info

ComplexityTraining timeRobustnessReusabilityAccuracyOptimality

Performance

Propagation losses, MAC and application layer throughput2-4 node topologySchedulers

Coloring, Graph abstraction,SINR, and MAC throughput estimation Regression analysis (several forms)Genetic algorithmsDerived system dependencies (PHY - MAC(incl. schedulers) - Application)

Num. of samples for trainingPrediction accuracyComments on computation e�ordsPer network/femto/user predictions

General criteria

Criteria speci�callydiscussed in our work




Conclusions


The specific optimization problemAn example

The specific problem that we are solving is: select for eachfor each deployed eNB i = 1, ...,N , the frequency fc andthe system bandwidth B that will provide the best networkperformance.In this work we consider capacity as performance metric(we are interested to consider delay, fairness, differentQoS metrics)Another aspect to consider is the number of possiblesolutions, which is exponential with NAn example for 4 femtocell scenario, total availablebandwidth of 5 MHz, fc muliple of 300 kHz and B that cantake value 6,15,25 there are 4625 physically distinctsolutions




Conclusions


LTE femtocell capacity estimationdesign the cost function

Common approach is to use some variation of Shannonformula, but those approaches typically do not considereffects of AMC, MAC scheduling,..The TB size for each given AMC scheme and number ofRBs is defined by the LTE specificationWe consider effect of different schedulers and for thatpurpose we use Round Robin and Proportional Fair

−10 −5 0 5 10 15 20 25 300

500

1000

1500

SINR

Cap

acity

inbi

ts/R

B

LTEShannonmod.Shannon




Conclusions


prediction methods and covariates

Our goal is to predict the performance of the femtocellaccurately by using regression analysis and machinelearning techniquesThe information used: basic pathloss and configurationinformation, a limited number of feedback measurementsthat provide the throughput and the delay metrics for aparticular frequency settingsDifferent regressors: SINR, SINR/MAC throughputmapping and different sampling technique affects theprediction performanceThe performance metrics considered are: network-wideand per-user throughputThe methods considered: Bagging tree, Boosted tree,Kohonen network, SVM radial, K-nearest neighbor,Projection pursuit regression, LinearBojovic, Meshkova, Baldo, Riihijärvi, Petrova Machine Learning and regression for performance estimation



Conclusions

Simulation scenario.

To simulate this scenario we use LENA, LTE-EPC networksimulatorWe choose a typical LTE urban scenario with buildings (inLTE literature known as dual stripe scenario)The HeNbs are randomly distributed in the buildings andeach HeNb has equal number of usersFor the four femtocell scenarios we simulate:

1 three users per node allocation with the total of 2 MHzbandwidth, and

2 two users per node allocation with the total of 5 MHzbandwidth




Conclusions

Dual-stripe scenarioRadio environmental map in a scenario with 2 buildings and 2 HeNbs

200

220

240

260

280

300

220 240 260 280 300 320 340

y-co

ordi

nate

[m]

x-coordinate [m]

-5

0

5

10

15

20

SIN

R[d

B]




Conclusions

Issues in predicting performanceMachine learning techniques for performance estimation

Issues in predicting performanceScenario 1

Scenario with 4 femtocells and 3 users each femto, using 2MHz overall bandwidth, showing actual measured MACthroughput vs. Shannon models. Sum SINR means a sumover all RBsEven if correlations between the different SINR relatedmetrics and the achieved throughputs are apparent, thedispersion of the results is rather substantial




Conclusions



1000 1200 1400 1600Sum SINR [dB]

MAC

Thr

ough

put [

Mbp

s]

6.0 7.0 7.5 8.5Sum SINR/THR NS3 Mapping [Mbps]

MAC

Thr

ough

put [

Mbp

s]

MAC

Thr

ough

put [

Mbp

s]

120 240 360 480

100

150

200

250

300

MAC

Thr

ough

put [

kbps

]

600 720 840 960 1080 1200

600

800

1000

1200

1400

MAC

Thr

ough

put [

kbps

]

(c) Measured MAC thoughput for two selected users vs. SINR/MAC THR mapping for a scenario with 4 femtos 3 users each, Proportional Fair (left) and Round Robin schedulers (right) for TCP traffic.

6.0

7.0

8.0

9.0

MAC

Thr

ough

put [

Mbp

s](b) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right)

for a scenario with 4 femtos with 3 users each, Round Robin scheduler, and TCP traffic.

(a) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario with 4 femtos with 3 users each, Proportional Fair scheduler, and TCP traffic.

Sum SINR/THR NS3 Mapping [Mbps]

Sum SINR/THR NS3 Mapping [kbps]Sum SINR/THR NS3 Mapping [kbps]

6.5 8.06.

07.

08.

09.

0

6.0 7.0 7.5 8.5

5.5

6.0

7.0

7.5

6.5 8.0

6.5

5.5

6.0

7.0

7.5

6.5

Sum SINR [dB] 1000 1200 1400 1600




Conclusions



Scenarios with 2 femtocells and 5 users each and 4femtocells with 2 users each, using a bandwidth of 5 MHz.Different behavior is noted for UDP and TCP traffic, and fordifferent schedulersFrom these results we can expect effective performanceprediction to be a challenging problem especially for TCPtraffic and a small number of users per femtocell.




Conclusions



6.0

8.0

10.0

MAC

Thr

ough

put [

Mbp

s]

6.0

8.0

10.0

MAC

Thr

ough

put [

Mpb

s]

0 5 10 15 204.

0Min SINR per RB [dB]

4.0

Sum SINR to MAC Throughput Mapping [Mbps]

(a) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario with 4 femtos with 2 users each, Proportional Fair scheduler, and TCP traffic.

10.0

15.0

20.0

25.0

Sum SINR to MAC Throughput Mapping [Mbps]M

AC T

hrou

ghpu

t [M

pbs]

0 5 10 15 20Min SINR per RB [dB]

10.0

15.0

20.0

25.0

MAC

Thr

ough

put [

Mpb

s]

0.8 1.2 1.6 2.0 2.4

2.0

3.0

4.0

5.0

2.0 3.0 4.0 5.0

4.0

6.0

8.0

10.0

Sum SINR to MAC Throughput Mapping [Mbps]Sum SINR to MAC Throughput Mapping [Mbps]

MAC

Thr

ough

put [

Mpb

s]

MAC

Thr

ough

put [

Mpb

s](b) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario

with 4 femtos with 2 users each, Proportional Fair scheduler, and UDP traffic.

(c) Measured MAC throughput vs. SINR/MAC thoughput mapping for a selected femtocell (left) and a whole network (right) for a scenario with 2 femtos with 10 users each,

Round Robin scheduler, and UDP traffic.

4.0 8.0 12.0 16.0 20.0

4.0 8.0 12.0 16.0 20.0




Conclusions


Pursuit regression technique

Scenario with 4 femtocell and 3 users per each femto andbandwidth of 2MHzRandom sampling with 10% of 337 permutations beingexploredThe regression technique is appliedThe earlier expectation that UDP traffic with more primitivescheduler results in higher predictability is confirmed, theusers the predictions are very accurate




Conclusions


Pursuit regression techniqueFigure

0 50 100 150 200 250 300 350

00.

51.

5Th

roug

hput

[Mbp

s]1.

0

Permutation

(a) Measured and predicted MAC throughput for a selected user with permutations ordered after measured throughput for the scenario with TCP traffic, and proportional fair scheduler.

MeasuredPredicted

(c) Achieved performance ratio (predicted vs. measured MAC throughput) for the scenario with UDP traffic,

and round robin scheduler.

1 2 3 4 5 6 7 8 9 10 11 12

0.4

0.6

0.8

1.0

User ID

Ach

ieve

d pe

rform

ance

ratio

1 2 3 4 5 6 7 8 9 10 11 12

0.4

0.6

0.8

1.0

User ID

Ach

ieve

d pe

rform

ance

ratio

(b) Achieved performance ratio (predicted vs. measured MAC throughput) for the scenario with TCP traffic,

and proportional fair scheduler.




Conclusions


A simple linear regression model

Scenario with 4 femtocell scenario and 3 users per femtoand bandwidth of 2 MHzTCP traffic and proportional fair schedulerThis method results in poorer prediction average both interms of median error, as well as the variability of theresultsThe increased amount of information available for thepredictor results in both improved median predictionperformance, as well as substantial reduction in themagnitude of the outliers




Conclusions


A simple linear regression modelFigure

(c) Achieved performance ratio for the linear regression method with 10% of samples and the stratified sampler.

1 2 3 4 5 6 7 8 9 10 11 12

0.4

0.6

0.8

1.0

User ID

Ach

ieve

d pe

rform

ance

ratio

1 2 3 4 5 6 7 8 9 10 11 12

0.2

0.4

0.6

0.8

1.0

User ID

Ach

ieve

d pe

rform

ance

ratio

(c) Achieved performance ratio for the PPR regression method with 70% of samples and the stratified sampler.

1 2 3 4 5 6 7 8 9 10 11 12

0.6

0.7

0.8

0.9

1.0

(c) Achieved performance ratio for the Kohonen regression method with 10% of samples and the random sampler.

User ID

Ach

ieve

d pe

rform

ance

ratio




Conclusions

Conclusions

The initial results obtained here show that advancedregression techniques can result in good performancepredictionsIn future work we plan to investigate an active samplingbased on graph coloring


A study on machine learning and regression based models...

Documents

Transcript of A study on machine learning and regression based models...