Opportunities in Communication

Hitachi Professor Emeritus of Engineering

J OHN M . C IO F F I

Communicationusesandapplications

April 20, 2018 2

EE 102April 20, 2018

BasicCommunication(digital)

§ Communicationisfundamentaltoallwedo(informationisthenaturalresource)• Uptherewithwater,electricity,food,…

• Senderchoosesamessagetosend(canbefromlargeset)

• Channelwilldistortthismessage(maybealotormaybeonlyslightly)

• Receiverattemptstodecidewhatthemessagewas(mathematically“detection”)

§ Talking,texting,sending/viewingvideo,searchingoninternet,facialrecognition,radar,lidar,….

3

EE 102

§ Messages• Internet• Email• Text• video,audio• Sensor/cameraimages

TraditionalInternetServiceProviderComm Channels

§ ChannelØ Fiber, copperØ WirelessØ Many or singleØ Mesh

Networkedge

April 20, 2018 4

OSI = ?

Open Systems InterconnectHow much money spent on connectionSubscriptions by consumers globally?

$1.3T

EE 102

§ Delayedsignalcouldbe180degreesoutofphaseatsomefrequencies• Whathappensthen?

§ Solution:Exploitthedimensions• use/leverageotherfrequenciesthatreinforce(frequencydimensions)• Lookonlyinonedirection(spatialdimensions)

April 20, 2018

TraditionalMobileand“multi-path”

L1: 5

Any one spot a problem?

This could be a reflectionAlso (building, mountain, etc)

EE 102

Notsotraditional– SearchEngine

• Yes,asearchengineisalsoreceiver/detector– decideswhichpagestodisplay(“pagerank”)Ø Searchesamongmanypossiblemessages

• ThechanneloutputisthetypedphraseØ Thatchannel-outputphraseforthisuserhasamessage(pageorgroupofpages)detectedforit

• Dimensionsarethewords,phrases,context,etc (alongwithsomehistoryanduserknowledgeor“cookies”)

April 20, 2018 L1: 6

www.stanford.eduChannel

outputReceiver

Estimated message

Messageset

The ultimate receiver is the woman herself – she decides to click on one of the displayed pages“iterative decoding” – the system may refine the decision with additional passes (more dimensions added)

Layer 7 = Layer 1?

EE 102

Self-DrivingCars

• Soself-drivingcarshavemanycommunicationchannelsandreceiversØ Multi-dimensionalinmanyways

§ Messages to be estimatedØ Relative position

Ø Chan out = Lidar reflection Ø delay = distance = position

Ø RouteØ Channel out = destination

Ø Location: current, anticipatedØ Chan out = GPS coordinates, recent locationsØ Message is the position

Ø Driver IDØ Chan out – facial image, finger prints (cameras in or out)

Ø TimeØ Chan out = electronic ref, GPS, base station signal

Ø Info to/from other vehiclesØ Chan out = Wi-Fi, Bluetooth, other

April 20, 2018 L1: 7

EE 102

Drones/Satelites andCommunication

• MessagesØ Emails/webpagesØ Tests/chatsØ imagesØ VideoØ Audio

April 20, 2018

§ ChannelØ airØ Moving fastØ attenuatesØ noise

L1: 8

Dimensions are in time, possibly space (multiple antennas)

EE 102

3DCameras(imagesandvideos)

• Messages=Ø SportingeventØ TrafficconditionsØ Security/intrudersØ People/faces

April 20, 2018

§ CHANNEL OUTPUTS• Visiblespectrum

• Obstacles(opaque)affectthis• Multiplepaths• 2Darraysofpixels

• Multipleplaces

9

Who is in the photo?

EE 102

Manufacturing/Factories

• MessagesØ Statusof“widget”Ø TestresultsØ PositionofmachinesØ Positionof”widget”Ø faults

April 20, 2018

§ ChannelØ Wired and/or wirelessØ Interference from othersØ Sensor and other low-power limitations

10

EE 102

HealthCare

• MessagesØ Diagnostics

SensorsaroundthebodySensorsinthebody

April 20, 2018 11

§ Channel .. could beØ Half-way around worldØ From internal organ to cellphoneØ Unusual propagationØ Very high frequencies within body

Ø Or very low frequencies (pills with radios)

EE 102

Agriculture– Feedtheworldbetter

• MessageØ PlantstatusØ AnimalstatusØ Water,fertilizerlevelsØ Machineposition

April 20, 2018 12

§ ChannelØ wirelessØ To drone, satellite, local AP

EE 102

SmartCitiesand/orSmartHomes

• Messages– appliancestatus,typicalinternetmessages,• Channel– wirelessandwires,manypaths

April 20, 2018 13

EE 102

§ Despiteallourcommunicationalgorithms/knowledge

April 20, 2018 14

Consumer/User/ThingHappiness?

Tech Mahendra 2016

Over 50% of under-35 internet consumers experience it ~ daily

“a state of uncontrollable fury or violent anger induced by the delayed or interrupted enjoyment of streaming video content from over-the-top (OTT)

services.”

EE 102

§ Self-DrivingAutoRagecoming?~😡§ PatientRageevenworse?

§ Securityasservice,privacyasservice?• Whathappensifthesehaveoutage?

April 20, 2018 15

Fifth-GenerationCommuncations:Applicationtrends(or“verticals”)

CurrentlyTelecom

Services is$2T/yearGlobal

industry

BasicCommunication

April 20, 2018 16

EE 102April 20, 2018

3BasicProblemstoSolve

§ CHANNELIDENTIFICATION– whatkindoflimitsdoesithave,doesitvary,varywithwhat?

§ CODING - Whatisagood(best)setXforagivenchannel?

§ DETECTION – Whatisagood(best)receiverfordeciding X?

L1: 17

The same 3 problems occur in Machine Learning

EE 102April 20, 2018

SimpleAdditiveWhiteGassian NoiseChannel

§ 4messagesor16messages§ QAMà 2dimensional§ EquallyLikely§ Cos/Sine

L1: 18

X̂

X

2d

23d

2d-

23d-

2d

23d

2d-2

3d-

4QAM

16QAM

y

§ Addnoise§ Zeromean§ Variance

§ Pickclosestpoint§ ForAWGN§ MaximumLikelihood

• Uniforminputdistn

Detection Problem First

+

ny

Pick closest point is maybe obvious,But also mathematically optimum

EE 102April 20, 2018

TheIntersymbol Interference(ISI)Channel

§ TheISIchannelwithadditivewhiteGaussianNoise(AWGN)

L1: 19

p(t)xk +

n t( )y t( )

§ Mostdetectiondesignedfornobandlimitations,buthowaboutthisbandlimitedISIChannel?

12T− 1

2T

Every T seconds, a symbol is transmitted – the channel “stretches” them and causes “intersymbol interference (ISI)”

Channel sums message with .9 x message delayed T sec

§ Equalization?? “Enhances” the noise

EE 102

Multiplecarriers/dimensionsinfrequency

April 20, 2018 20

• HowdowelearnandadjustØ Dynamically

• SomeofveryfirstAImethodsincom(fromStanford)

Use “bit-swapping” AI/MLMethod – SU patent #3 in Engineering (#7 overall)

“water-filling”

EE 102

§ Thereiscrosstalkbetweendimensions(kindoflikeISI,justdifferentdimensions)

• Canactuallydoublethedatarateifrightsignalprocessingused(andantennasarenottooclose)

April 20, 2018

2x2AntennaSystem(MIMO)

21

y1

y2

x1

x2

WirelessChannel

P

σ 2 = .1

EE 102

MultipledirectionsinSpace

April 20, 2018 L1: 22

Best Energy will also be water-fill overThe singular vectors of the channel

Essentially matrix form of machine learning From earlier

EE 102

MU-MIMO(multi-user)involvesmorelearning/factoring

April 20, 2018 L1: 23

Virtually all communication problems can be cast into the Same theory of finding the critical modes of channel

AnExamplefromthefield

April 20, 2018 24

EE 102

ANetworkStateMachine

• Netmaybeinastate

• GYRØ Moredetail

• Relatestomoretraditionalmodels

April 20, 2018 25

15-25 Mbpsclean

10-20 Mbpsclean

5-12 MbpsWeak signal

2-6 MbpsLow signal

1-2 Mbps~ nothing

15-25 Mbpsinterference

10-20 Mbpsinterference

5-12 MbpsWeak and Int

2-6 MbpsBad int

1-2 MbpsReal Bad

2-6 Mbpsboth

1-2 MbpsUgly

EE 102

• Strongcorrelationbetween

WiFi problemsand

ComplaintRate

• 48%ofAPshaveWiFi issues

• Customersidentifiedas

havingpoorWi-Fi

performanceare5times

morelikelytocomplain0.00% 2.00% 4.00% 6.00% 8.00%

10.00% 12.00% 14.00% 16.00% 18.00% 20.00%

NoProblem WiFiNeedsAttention

WiFiproblems

CustomerComplaintRate

Selling more mesh points to ISP often increases instability percentage

Stable not quite unstable

Wi-FiandQoE (thedesiredresponse)

26

GYR

April 20, 2018

EE 102

RootCausesofWi-FiQoEDegradation– 2.4GHz

• TopThreeWi-FiProblemDriversØ CorrelatedwiththemainUserQoE

April 20, 2018 27

Interference Coverage Noise

EE 102

Wi-Fi“Hot”SpotsandProblems

April 20, 2018 28

=

Coverage (noise)& signal strength

Crosstalk

who controls phy

EE 102

CocktailPartyEffect(crosstalk)

• Solution: All speak politely at low volume (lower power)

Ø All send more information (more power and/or higher data rate)

• This can be learned and optimized

• EG: Wi-Fi Box/Chips blasting at 1-10 Gbps !Ø Or worse yet – install repeaters/mesh and have them all blastØ Or use MU-MIMO when 3 people are collinear (BAD)

April 20, 2018 29

TALK LOUDER Sorry, can’t hear,Talk louder

I NEED TOTALK VERY

LOUD

OK,I’ll SHOUT

EE 102

Channel-optversustop-sellingchipauto-select

ChipX’s“Auto-Channel”vsMLCloudOptimization

April 20, 2018 30

Throughput Pre-OptimizationPopulation%

Post-Optimizationpopulation%

%Gain

≤3M 12.3% 4.7% 61.5%

≤5M 20.9% 11.4% 45.5%

≤10M 42.8% 28.6% 33.3%

62%reductioninWi-Fibelow3Mbps45%reductioninWi-Fibelow5Mbps

12%

9%

22% 43%

14%

AutoChannel–MLOptimization 5% 7%

17%

54%

17%

AfterMLOptimization

0-3M

3-5M

5-10M

10-20M

EE 102

Optimization,QoE,andStabilitybasedonseveralhundredmilliondailyinternetusers

• UnmanagedInternetconnectionsØ Globally~20%ofunmanagedfixedconnectionsareunstableeachmonth!

• Dynamic(optimized)ManagedConnectionsØ Oftenseetheabovenumbersdropby½ormoreØ Fromexperiencebothinelectronicmeasuresandconsumerfeedback

April 20, 2018 31

0% 20% 40% 60% 80% CzechRepublic

RomaniaGermanyHungaryCanada

USAFrance

UnitedKindom

Stabilityimprovement

53%

8%

40%

DidyouroverallperformanceimprovewithBooster?

Yes

No

Neutral/Notsure

management?

Consumer Survey Machine-Learned Stability

EE 102

QoETimeTrendlearned(300kIPTVcustomers)

• Notethedynamiclearningandadjustingongoing

April 20, 2018 32

0% 20% 40% 60% 80%

100%

3/1/20

16

4/1/20

16

5/1/20

16

6/1/20

16

7/1/20

16

8/1/20

16

9/1/20

16

10/1/201

6

11/1/201

6

12/1/201

6

1/1/20

17

2/1/20

17

QoETracker

BadQoE

WarningQoE

GoodQoE

EE 102April 9, 2018

DataWarsGame§ UsesFMloadingonseveralcross-talkinglinks/connections

• SignalfromotheruseristreatedasGaussiannoise

L4: 33

y1

y2

x1

x2

WirelessChannel

P

§ IfLink2coulduselowerenergy,thenLink1seeslessnoiseà Link1loadstohigherdatarate

§ ButLink2mightwantLink1tosaveenergysoitcantransmitfaster

§ Whowins– sometimescalled“theprisoner’sdilemma”ingametheory

EE 102April 9, 2018

Examplefor25usersallsharingsamespatialchannel

§ 25usersallsharingsamespectrum

• Near-end(“echo”)aswell– growswithf1.5 andfar-end(“xtalk”)attenuateswithsametransferbuthasadditionalf2 coupling.

• Eacheffectstheother24users

• Allareofsamedistancebetweenxmit/rcvr

L4: 34

25 Crosstalking systems versus connection distance

connection distance

Iter-W-Fillall-max rate

§ Speedsroughlydouble

EE 102April 9, 2018

Spectraafterconvergence

§ Theecho-crosstalkisstrong– see“FDM”effectbetweendownandup

§ Itwasalllearnedbywater-filling• Nocleverdesigner• NoFCC/regulator

§ Called“cognitiveradio”bysome

§ Notbest,butprettygood!

§ Canworkwithdifferenttopology,rates,etc

§ How’sthatforMachineLearning,eh?

L4: 35

UPLINK

DOWNLINK

4000’

Virtualization– SoftwareDefinedNetwork

April 20, 2018

EE 102

ConvergedVirtual5Gnetworks(fixed&wireless lines)

April 20, 2018 37

Virtual Network Operators Need “functions” (VNFs)

Infrastructure ProviderSoftware-

Defined-Network (SDN)

• Singleinfrastructuresharedbymultipleoperators

EE 102

X-Haul– basicallyallowsforjustADC/DACatantennas

April 20, 2018 385GPP X-Haul Group

• IncreasinglyeverythinginsoftwareØ Formost/allconnections

EE 102

TheOpportunity

• 4Bpeopleusinginternetà hopefullyallØ Theyallneed,andincreasinglydependseriously,ontheconnection

• ItaffectsallapplicationsØ ThenewhotonesoftodayØ Theoldvoice,videoØ Theoneswedon’tknowyet

• ThesenetworksandproblemswillalwaysbethereØ Andneedhelp,design,improvementfromEE’s– likethoseinthisclass!

April 20, 2018 39

Thank you

Hitachi Professor Emeritus of Engineering

J OHN M . C IO F F I