1

RAN Intelligence Use Case,

Architecture and Interface

Dr. Qi Sun

China Mobile Research Institute

Sep. 25th 2020

2

2

Bringing AI to the RAN: From “on the top” to “embedded”

Network Planning

Network Deployment

Network Operation & maintenance

Protocol stack & Signaling

Radio Resource Management

PHY layer Optimization

Statistic/Semi-Statistic

Sensing

Multi-dimensional

/cross-layer context info

(user, application, network)

Radio Environment Map…

Intelligent

Decision Making

Machine/deep Learning

Offline model training &

online decision making

Operation & Management Plane Control & Data Plane

Classic Communication Theory Meets Data Technology

Prediction

User behavior

(trajectory, location)

Traffic fluctuation

Service type

……

Real-Time

Customized Network

Strategy

Data Driven

Machine Learning Based

Complex Network Optimization

Predication oriented

configuration &

Decision making

3

3

Use Cases: AI empowered RAN optimization-Time/Resource Categories

Time~100msms s min Hours Day/Month

RBs

Carrier

gNodeB

Slice

Near Real-Time

(Control & User Plane)

Network Plan & Deployment

Network Optimization & Configuration

Network Anomaly Detection

RF parameter optimization

Network Energy Saving

QoS/QoE optimization

Load Balance

Interference Management

Multi-connectivity

Mobility Management

Cell

Non Real-time

(Management Plane）

Slice Resource Management

Resources

Multi-user scheduling

Link adaptation (MCS)

AI empowered PHY design

AI DPD

Real-Time

Commercialized

Testing and Trial

Research

4

4

Standardization: Overview of Research, SDOs and Open Source Activity

Research Standardization Open Sources

WAIA FG ML5G

IEEE ETI

Network Intelligence

machine learning for Communications

RAN3, SA2, SA5

ENI, ZSM

O-RAN

ONAP

Acumos

Adlik

PNDA

……

WG1/2/3

5

5

Standardization: Overview of Research, SDOs and Open Source Activity

ITU-T ML5G:

AI/ML framework

ML

Archiectectural &

Data handling &

optimization/

deployment

framework for

future networks

ETSI ENI/ZSM: OSS/BSS layer Big Data/AI Architecture & API/Interface

gNB (CU/DU)

CN-MDAS

Producer

RAN-

MDAS

Producer

Cross Domain-MDAS Producer

5G Core

CP & UP

NWDAFNear-RT RIC

Non-RT RICPolicy

Mgnt

EI

Mgnt

ML

Mgnt

gNB (CU/DU) gNB (CU/DU)

gNB (CU/DU)

Mangement

RAN CN

ML model inference

UE

SA5 R17 SI(IDMS, MDAS)

SA2 WIR16 eNA

O-RAN RIC (Non-RT & Near-RT)

RAN3, R16 SI DC &WI SON&MDT

SA1 WI: Model distribution and transfer [under discussion]

E2fine granularity

data collection&

control&policy

3GPP R17

3GPP R16

O-RAN (RIC)

3GPP Discussions

6

6

Enable RAN Intelligence with Hierachical RIC

Orches-trator

Non-RT RIC

Intelligent Management and Orchestration

FCAPS

Fault CMPMsecu

rity

AI Model mgnt

PolicyMgnt

EI Mgnt

data analytics; AI/ML training

rApp 1 rApp 2 rApp N

Near-RT Radio Intelligent Controller

xApp 1 xApp 2 xApp N

Database

gNB/CU/DU

E2

O1 A1 O1*

Messaging Infrastructure

SubscriptionMgmt.

Mgmt. Services Conflict

Mitigation

Log

gin

g, Tra

cin

g,

Metr

ics

Secu

rity

Intent (Goal, what)

Declarative

Policy (/Behavior, How）Imperative

Enrichment Data

(UE Speed, service type)

AI/ML Model

(data driven algorithm)

Imperative Policy

Event, Condition, Action

Smart Control

Hierarchical Intents/Policy/Control

rApp 1

xApp 1

Non-real time ML application

Near-real time ML application

RAN Programmability

Data Collection, Processing, Storage & Sharing

7

7

Non-RT RIC and A1 interface

Main Use Cases

• QoE Optimization

• QoS Optimization

• Traffic Steering

• Network Slicing performance assurance

A1 interface support

A1-P – Policy Management Service

A1-EI – Enrichment Information Service

A1-ML – ML Model Management Service

Non-RT RIC framework

AI/ML model training

A1 policy management

Enrichment information management

Network Configuration Optimization

rApp: non-RT intelligence application, e.g. Carrier

license scheduling, energy saving, ...

Non-RT RICSMO

(FCAPS & Orch)

Service Management and Orchestration Framework

Near-RT RIC

policy &

Enrichment Information

(for UE/group of UE/

Applications)

AI/ML

model

deployment/

updates

O-CU/O-DU

O1 configuration

(e.g.AntennaHorizontal/vertical

angle/bandwidth)

O-RU

E2 (Control & Policy)

O1/O2

PolicyMgnt

Enrichment Info. Mgnt

AI/ML Model Training

Network Config. Opt.

A1

RAN intent Enrichment Information

rApp rApp rApp

Goal: Enable closed-loop automation and optimization of RAN elements & resources, making it more intelligent (ML/AI), more

granular (per-UE or group of UEs), more flexible (intents/policies).

8

8

Non-RT RIC & A1 Standardization Progress

• Non-RT RIC: Functional Architecture under discussion

• A1 interface:

• A1 specification v1.0 published in Nov. 2019, v1.1 published in Apr. 2020, v2.0 in July 2020.

No. Specification Contents

1 A1 General Aspects and Principles (A1 GAnP) • General principles, open A1 interface and interoperability• Define 3 types of services: A1 Policy, A1 Enrichment

Information, A1 ML

2 A1 Transport Protocol (A1TP) • A1 transport layer definition，HTTP/JSON

3 A1 Application Protocol (A1AP) v1.0 • A1 application layer definition，A1 API Definition and Data models

• Support UE/Slice/QoS/Cell level QoS/QoE targets and cell/carrier access policy

4 A1 Application Protocol (A1AP) v1.1 • Add YAML format A1-P OpenAPI(s) standard• Improve A1 status & Notification operations

• Ongoing activities

• A1 EI (enrichment information ） APIdefinition, exmaples including

• radio fingerprint

• video code rate/frame rate

• UAV path

• weather

A1-P Consumer

non-RT RIC

A1-ML Consumer

A1-EI Producer

A1-P Producer

A1-ML Producer

A1-EI Consumer

near-RT RIC

A1-P A1-ML A1-EI

near-RT RIC

non-RT RIC

A1

SMO

O1

E2 nodes

E2

RAN intent

O-RAN external information

sources

Information

O-RAN internal information

sources

Delivery of External Enrichment Information

9

9

A1 Policy Protocols, Procedures and Data Models

L1

L2

Physical layer

TCP

HTTPS

JSON

non-RT

L1

L2

IP

TCP

HTTPS

JSON

near-RT

IP

Data link layer

Network layer

Transport layer

Application delivery

RAN modeling language

(policy based)

A1

A1 interface protocol structure

A1 policy procedure

HTTP method

Create policy POST

Query policy GET

Update policy PUT

Delete policy DELETE

Feedback policy POST

Query policy type GET

A1 Policy Procedures

Policy Representation

Policy Objective/Resource Statements

Scope Identifier

Data Models

ueId

groupId

sliceId

qosId

cellId

QosObjectives：GFBR,MFBR,Priority Level, PDB

QoeObjectives:

qoeScore,initialBufferingreBuffFreq,stallRatio

TspResources:

cellIdList,preference,primary

10

10

Further Thoughts on the Intent/Policy Modeling Enhancement

Intent/Policy

Objects Operation Result/Goals

Expected

State

Avoid

State

Condition ActionConstraint

•Time/geographic/cell scenarios

•Cell Level context

•Traffic load (PRB usage, UE numbers)

•UE level context

• UE RSRP/CQI/buffer status

•UE speed

•UE service type

•UE levels (VIP, normal, high value/Low value)

•UE terminal type, e.g, URLLC, eMBB, mMTC

Node

Silce

UE/Flow

• Activate/Deactivate

• Handover/stay

• admission accept/reject

• connected/disconnected

• flow control

• increase/decrease scheduling priority

• change the RAN configuration

MaximizeObject ModelPreferred priority

Avoid

Minimize

Object Model Behavior Model+

Behavior Model

RAN Model

• KPI targets

•Average KPI

• variance of KPI

•Energy Efficiency

• SLAs:delay, throughput,

bandwidth, jitter, UE

connection number,

bitrate

• preferred/avoid carrier

bands or priority

Note: condition and action may apply to different objects

11

11

Business Value Driven Use Cases: RAN Optimization and Capability Exposure

data：L1/L2 measurement、MDT、signalingmechanism：use case driven data subscription

Service QoE Assurance

CT Enhancment

Wireless Capability Exposure

QoE/SLA prediction

QoE/SLA assurance

Wireless Positioning

Radio bandwidth prediction

Load balancing

Mobility Optimization

Interference Management

DC/CA

Smart MCS

Energy Saving

RAN data collection

RAN data analytics and AI/ML functional

Framework

Architecture、Inf & Procedures

Usage Scenarios Use Cases Key Techs

data collection、training、inference、decision、execution

centralized, distributed Archmodel distribution, update

To provide cutomized network capabilites and

service assurance, especially for the

diversified verticals

To improve the network resource and energy

efficiency

To provide value added services

12

12

Key Issues and Future Work

• Usage Scenarios

• RAN capability customization for the verticals

• to make the network easy to be customized by the diversified requirements

• e.g., Deterministic SLA assurance

• RAN automation

• to simplify the network operation and maintainance

• e.g.,Network planning, optimizaiton, maintainance

• Intent/Policy Modeling

• Hard to model the Intent/Policy Expression for varied usage scenarios

• How to model the hierachical level of the Intent/Policy?

• Intent/Policy Engine Design

• How to do the network control and optimization to fulfill the Intent and Policy?

• How to leverage the wireless big data and AI/ML technologies?

13

Thank you!

14

14

Discussion on the RAN programmability

➢ Level 1: Configure the parameters of an well designed algorithm to change the behavior (mainly

used in the current OAM configuration management)

➢ Level 2: Adding Policy rules/new Algorithm logics of the optimization problem solution to guide the

algorithm behavior (e.g. condition/action/constraint /parameter rules) [Imperative policy]

➢ Level 3: Express the intent and Model the RAN optimization problem by setting the objectives,

constraints. Let the system to figure out how to do accomplish the Intent. [Declarative Policy]

➢ Need a intent engine (e.g., algorithm framework) to automatically solve the problem .

➢ Level 4: Depoly/on-board algorithms (can be AI/ML assisted software, containers, etc.) directly

on the RAN to address specific RAN optimization

L1 L2 L3 L4

ConigurationImperative

PolicyDeclarative

Policy

(Intent)

Software

Programmability

Onboarding applications