Presentation by Yue Wang, Samsung Research UK at IET 5G - the Advent conference on 30 January 2019 | IET London: Savoy Place
This presentation looks scalable and deployable AI solution in the 5G Infrastructure along with updates from ETSI. Use cases such as Energy Saving, Cell Selection, Fronthaul Management, Orchestration & more are discussed in this presentation.
*** SHARED WITH PERMISSION ***
Long journey of Ruby standard library at RubyConf AU 2024
Yue Wang: AI in 5G –the Why and How - Jan 2019
1. AI in 5G – the Why and How
Yue Wang, Samsung Research UK
5G - the Advent, 30 Jan, 2019
2. What has changed in 5G?
2
• Smart phones vs cars,
cameras…
• End customers vs.
Verticals - B2B, B2B2C
B2B2X
• Best efforts vs
customised SLA
3. What has changed in 5G?
3
• One Network
(with new technologies)
• Network
• SBA
• Slicing
• Access - 5GNR
• New spectrum
• massive MIMO
• numerology
• Deployment – SA/NSA
A complicated network
RAN, TRANSPORT, CORE
ORCHESTRATOR
4. How Can AI Help?
4
• High efficiency
• Context aware
• Future proof
AI
• Scalability
• Adaptability
• Flexibility
• Real-time
• Dynamicity
• Flexible and scalable to support
variety technologies and KPIs
• Fast adaptable to new (sometimes
real time) network contexts
• Optimised operations and efficiently
use of resources – OPEX
5G
Better Service, More customers, with Lower Cost
6. AI for Energy Saving
6
Configuration for peak hour
* ETSI ENI
• Servers take 70% of the total power
consumption
• Deployed and running to meet the
requirement of peak hour service - 100%
powered-up full time
• Learn and update service pattern
• Autonomously turn spare servers to idle
state
• Predict peak hours and wake up the
necessary number of servers
7. AI for Cell Selection
UE
Actions (the
selected TRP)
cell selection
with AI
UE location,
speed,
measured
signal strengths
(RSRP/RSRQ)
Feedbacks from
the network
UE
7
36.304, cell reselection procedure:
• 35 parameters for system information
• 10 parameters for speed dependent selection
• 13 parameters for interworking
• The list is getting larger:
• New technologies - beam sweeping
• New services
Without AI With AI
Periodically measured Triggered measurement
Threshold based (lots
of parameters/
configurations)
No threshold
Static configurations Real-time, adapted to
changes of the context
(e.g., speed)
Faster selection
Reduced latency
8. AI for Fronthaul Management
8
• Multiple factors
• Changing contexts
• Large dimensionality of
solution space
• Flexible and dynamic resource
slicing and functional split
• Real-time optimisation
RAU
Fronthaul
Cluster
FronthaulOrchestration
andManagement
Load Estimation
RCC
Data-plane Fronthaul traffic per slice
Functional split, cluster size designation and resource reservation
Low MAC
High PHY
Low PHY
RF
PDCP
RLC
High MAC
9. Elastic Resource Management
Elastic Intra-slice
Orchestration
Elastic Cross-slice
Orchestration
A
B
C
D
E
F
Network
Control
Network
Orchestration
Multi-slice
awareness
Single slice
Optimization
VNF elasticity
• Computational aspects of VNFs
• Orchestration of the computational
resources across slices
• Optimise VNF migration
using intelligence on multiple
resource utilization data (CPU, RAM,
storage, bandwidth)
• Elastic resource provisioning
to network slices
*5G-MoNArch
*ETSI ENI
10. Putting Together
10
Core Network/Cloud NFV
RCC
Orchestrator
AI
AI
AI
AI
VNFs
AI
AI
AI
AI
Initial Access
Fronthaul
management
Slice Management
AI
Power
management
11. But there is more…
11
Service orchestration and
management
• Service deployment
• Service optimisation and
prioritisation
• Service assuranceRAN Transport Core
AI
12. Problem Domain
12
Simple AI for self-
contained
problems
Local contexts
Immediate benefit
Trickier AI for
linked problems
Network contexts
Compatibility and
enhancement on the
E2E network
New
problems/systems
with advanced AI
Significant change of
the system with
advanced AI (Neural
network, deep
learning)
Disruptive innovations,
longer term
OR just for the fun of
research
13. In the AI space…
13
Artificial Intelligence
Machine learning
Reinforcement learning
Classification/regression
…
…
Supervised learning
clustering
…
…
Un-Supervised learning
Abnormality detection …
Knowledge representation
and reasoning
Natural language
processing
Multi-agent systems Robotics
15. Think more on the practical side
15
Put an ML
in the network
Results/enhancement
• Synthesized data?
• Accurate representation of real data?
• How is one ML better than the other?
• The integration to the network?
• Isolated AI?
• Network contexts?
• The inter-link of the network?
A scalable and
deployable AI
solution
How to turn
concept to practice?
Open questions
• Specific data for specific solutions vs unified data
• Standardization vs open source
• Integration and compatibility to the (‘legacy’) network