+ All Categories
Home > Documents > Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis,...

Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis,...

Date post: 08-Jul-2020
Category:
Upload: others
View: 1 times
Download: 0 times
Share this document with a friend
18
Decoupling Prediction and Control in Modern Networks Sandeep Chinchali, Alex Anemogiannis, Tianshu Chu, Marco Pavone, Sachin Katti Forecaster Controller
Transcript
Page 1: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

DecouplingPredictionandControlinModernNetworks

SandeepChinchali,AlexAnemogiannis,Tianshu Chu,MarcoPavone,SachinKatti

Forecaster Controller

Page 2: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

Time-seriesControlisUbiquitous

[email protected] 2

RoboticTaxiFleets

OptimalControl

CellCongestion

StockPrice

VideoStreaming

QuantitativeFinance

Page 3: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

ChallengesofTime-seriesControl

1. Data-drivenforecasts• Whatfeatures/statistics areneededforcontrol?

2. ManyInputVariables• ForecasterandController

3. Increasingly:• Databoundaries

CellCongestion

Controller

NetworkOperator

Page 4: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

Claim:Decouple butco-design predictorandcontroller

Whynotend-to-end learning?

WhyDecouple?1. NaturalDataBoundaries2. Modularity(Re-useforecaster)

WhyCo-design?1. Tuneforecaststocontrolrisk2. RobustAdversarialTraining

4

End-to-end

State Action

End-to-end

State Action

ControlAPI

Reward

Forecaster Controller

PrivateInfo PrivateInfo

Page 5: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

ControlReward

ForecastedFeatures

Action

ControllerPrivateStateForecasterPrivateState

Joint(Public)State

(ControlAPI)

GeneralApproach

Forecaster Controller

Page 6: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

VideoQoE

Private:Usermobility Private:BufferState

PastThroughputs

Bitrate

FutureThroughputs(Risk-adjusted,~30s)

Forecaster Controller

VideoStreaming

NetworkOperator

CloudVideoServices

Page 7: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

PassengerWaitTime,RideEfficiency

FutureCellCongestion,Anomalies (~hrs)

TaxiRoutes

Private:UserLocation,CellDemand

Private:Taxilocations,OutstandingDemand

Forecaster Controller

RoboticTaxiFleet City-wideCellCongestion

NetworkOperator

TaxiOperator

Page 8: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

Approach:ReinforcementLearning(RL)

Forecaster Controller

Page 9: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

ReinforcementLearning(RL)

Goal:Maximizethetotalreward

Agent Environment

Observestate𝑠"

Action𝑎"

Reward𝑟"

9AdaptedfromPensieve (Sigcomm 18,Maoet.al.)

Page 10: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

ControllerForecaster

RLFormulation

Adversarial

Page 11: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

Application:MobileVideoStreaming

ABRagent

state

NeuralNetwork

240P480P720P1080P

policyπθ(s,a)

Observestates

parameterθ

Figure fromPensieve (Sigcomm 18,Maoet.al.)

CanwebeatPensieve byforecastingnetworkconditions?

Page 12: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

PaloAltoCellThroughputDiversity

Insight: Foresightoftrue networkconditionhelpsSolution: Dynamicallysplicespecializedcontrollers(metaRL)

SwitchCellFreeway

Page 13: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

High-LevelTrace

Statistics[µ, 𝝈](API)

VLO

LO

MID

HI

metaRL

Bitrate

metaRL controllerForecaster

VLO

LO

MID

HI

Infer networkconditionfromforecaststatsPrivate:UserLocation

QoE

Page 14: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

PaloAlto(Ourdata)+FCC/Norway(Pensieve)

metaRL

Page 15: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

GeneralizetoFCC/NorwaydatafromPensieve

metaRL

Page 16: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

Re-analysisofPensieve (Sigcomm 18,Maoet.al.)

LinearQoE (hi-thpt) HDQoE (vlo-thpt)

OptimizeTail

Page 17: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

QuantifyingSub-optimalityGap

Withoracleknowledgeofnetworkcondition Havetolearnnetworkcondition

Value/PriceofTimeseriesVariables?

Gap

Page 18: Decoupling Prediction and Control in Modern NetworksSandeep Chinchali, Alex Anemogiannis, TianshuChu, Marco Pavone, Sachin Katti Forecaster Controller. Time-series Control is Ubiquitous

Futurework

1. Broad-visionforTime-SeriesControl• Data-drivenforecasts/controlstrategies

• Intrinsicdataboundaries

2. Value/PriceofInformationusedforLong-TermControl?

3. Privacy/InformationLeakage

Questions:[email protected]

Forecaster Controller


Recommended