Towards Cognitive Autonomous

Networks in 5G

26-28 November

Santa Fe, Argentina

Stephen S. Mwanje

Nokia Bell Labs, Germany

Stephen.mwanje@nokia-bell-labs.com

26-28 November

Santa Fe, Argentina

Need for Cognition in Network Automation

• Multi-RAT/Band Networks higher

complexity, CapEx, OpEx

• Solution = Automation e.g. SON

• SON is inadequate

– Static, rule-based mapping of KPIs to actions

– inflexible to the infinitely many contexts

Customer Customer

OSSPlanning

Network Engineering

Service fulfilmentCustomer

experience Mgt

Data & Analytics

BSS

Operator Network Management

Network Management Automation (NMA) maximizes value of automation

Why is this discussion important?

SON - Self-Organizing NetworksKPI - Key Performance Indicator

26-28 November

Santa Fe, Argentina

Automation, self-organization and cognition

• Nature motivates Self-Organization

– E.g. insect colonies on foraging for food

– Non-deterministic self-organization - “..

foraging … is also a learning process.” [6].

– Behavior emerges from internal cognitive

process

• Cognition (in humans): the collection of

mental sub-processes to support acquisition

of knowledge and understanding through

sensory stimuli, experience and thought

But what is Cognition?

Problem solving

Learning

thought

Intelligence

perceptionsensation attention

memory

Four basic processes as core;

Higher processes as combinations

of basic and other higher processes

26-28 November

Santa Fe, Argentina

Cognition: A perception-reasoning pipeline

How then should cognition be modelled?

Memory

Perception Reasoning

Sense

subconscious

(e.g. stop)

R

LSF

R

LSF

R

LSF

R

LSF

Reflex

(e.g. blink)

conscious (e.g. run,

fight, get a weapon)A

ctio

ns

Conceptualize

Contextualize

OrganizeInfer

26-28 November

Santa Fe, Argentina

Taxonomy -

• Two dimensions – decision & action

– Scripted network – human-controlled

scenario specific execution of scripts

– Self-Organizing Network – automatically

interprets events to determine cause-effect

relations, selects and executes of actions

– Autonomous network - acts on its own, but

may not be able to reason suitably

– Cognitive network - reasons to formulate

decision/actions, although actions may

require operator approval before execution

What levels of automation can be achieved in networks?

Different capabilities are achieved by

the different form of automation

Tak

e o

ne o

r m

ore

acti

on(s

)

Diagnose events

Contextualize events/data

Anticipate standalone events

Anticipate correlated events

Correlate events/data

Detect a network event

Scri

pte

d

Co

gn

itiv

e-A

uto

no

mo

us

Self

-Org

an

izin

g

Au

ton

om

ou

s

Preset time/sequence trigger

Au

tom

ate

d

26-28 November

Santa Fe, Argentina

Transition trajectory towards CANs

How do we get to CANs?

Environment

Network

BSS/OSS

configure

OAM

NM

interact

Day

s-

Wee

ks

User

Operator

Set Network

Parameters

Network

BSS/OSS

configure

SON

Day

s-

Mon

ths

User

KPIs

Min

s

Configure

SON

Network

BSS/OSS

configure

CNM

Secs-

Mon

ths

User

KPIs

Min

s

Define

contexts

OAMOAM/

History

mon

ths

contextualize

Environment

Network

BSS/OSS

configure

CAN

Secs-

Mon

ths

User

KPIsM

ins

Update/

upgrade

OAM/

History

Derive contexts

SON Function selects from

manually pre-set actions

Operator defines contexts; CF

learns context-specific actions

Cognitive-Autonomous Function

learns contexts and their actions

Operator decides how and

when to execute actions

Manual Automated Cognitive - automated Cognitive - autonomous

26-28 November

Santa Fe, Argentina

CAN Framework

How will the CAN system look like?

Environment Modelling & Abstraction

Decision Applications

(Dapps)

Configuration Management

Engine (Controller)

Interfaces:a - Technical targetsb - Raw network data c - Network statesd - Preferred config’ns e - Network config’ns f - Config’n reports g - Coordination decisionsh - DApps config’ns

NETWORK & ENVIRONMENT

Static nature, e.g.

Locality, BS location,

RAT, Spectrum

KPIs

Dynamic nature, e.g.

Velocity, User count,

Applications, etc

Coordination Engine

Network Objectives Manager

Decision Applications

(Dapps)

Decision Applications

(DApps)

Network data

b

ca

d

e

f, h f

g

Operator Goals

• Cognitive Function (CF) may contain:

– DApps - Learners for specific objectives

– CME (controller) - set boundaries and

decide if recommendation is executed

– CE - resolve conflicts among DApps

recommendations

– EMA - define and label observed

network events and contexts

– NOM - translate the operator’s goals to

technical targets for CFs

26-28 November

Santa Fe, Argentina

Example CF – Mobility Robustness

• Objective: learn optimal behavior

per mobility (velocity) state in cell

– Control - Hysteresis (Hys) & Time-

to-Trigger (TTT)

• KPIs – Rates of

– HO Ping-Pongs (PPR)

– RLF due to early HO (FER)

– RLF due to late HO (FLR)

• Scenario: LTE network of 21 cells

– Velocity: 3,10,30,60,90,120 kmph

Implementation – Supervised learning

on Velocity, Hys, TTT and rates

– Nearest Neighbors Regression

(KNN)

– Decision Trees (DT )

– Random Forests (RFR)

– Extremely Randomized Trees (ETR)

Evaluation: RE - absolute error relative to expectation of rate E{y}

𝑅𝐸 =|𝑦𝑖 − 𝑦𝑖|

𝐸{𝑦}

Can we concretize these ideas?

on

HO: handover; RLF: Radio Link Failures

26-28 November

Santa Fe, Argentina

Learning network responseML algorithms predict behavior in unknown states using response in known

states

What has been observed so far?

26-28 November

Santa Fe, Argentina

Not simply off the shelf

Achieve better outcomes through smart combinations of algorithms

Any lessons from the study?

26-28 November

Santa Fe, Argentina

Learning methods can automate network

management not only to learn the best

configurations for specific states but also the

underlying network response which can then be

used to predict response in unknown states

Thank you