Machine learning for Wireless Networks: Challenges and ...jfro/Journees... · HUAWEI TECHNOLOGIES...
Transcript of Machine learning for Wireless Networks: Challenges and ...jfro/Journees... · HUAWEI TECHNOLOGIES...
1
Machine learning for Wireless
Networks: Challenges and
opportunities
Mérouane Debbah
Mathematical and Algorithmic Sciences Lab, Huawei, France
8th of June, 2018
HUAWEI TECHNOLOGIES CO., LTD. Page 2华为保密信息,未经授权禁止扩散
Huawei [Wow Way]
180,000Employees
1580,000 No. 70170+ No. 83R&D Institutes
& Centers
R&D employees
Interbrand'sTop 100 Best Global Brands
Countries Fortune Global 500
HUAWEI TECHNOLOGIES CO., LTD. Page 3华为保密信息,未经授权禁止扩散
HQ in Boulogne-
Billancourt
1 Open Lab (opening
April 2018)
85 researchers in
applied
mathematics
4 R&D centers,
Academic
cooperation with
Major Universities
950+ employees,
among with 150
in R&D
Huawei [in France]
HUAWEI TECHNOLOGIES CO., LTD. Page 4华为保密信息,未经授权禁止扩散
The cost of Understanding
HUAWEI TECHNOLOGIES CO., LTD. Page 5华为保密信息,未经授权禁止扩散
Why it works
•Machine-learning algorithms have progressed in recent years, especially through the development of deep learning and reinforcement-learning techniques based on neural networks.
•Computing capacity has become available to train larger and more complex models much faster. Graphics processing units (GPUs), originally designed to render the computer graphics in video games, have been repurposed to execute the data and algorithm crunching required for machine learning at speeds many times faster than traditional processor chips. Key Trend Emerging: Specially design chips and Hardware for Machine Learning workloads (Tensor Units).
•Massive amounts of data that can be used to train Machine Learning models are being generated, for example through daily creation of billions of images, online click streams, voice and video, mobile locations, and sensors embedded in the Internet of Things devices.
HUAWEI TECHNOLOGIES CO., LTD. Page 6华为保密信息,未经授权禁止扩散
From Learning to recognize
HUAWEI TECHNOLOGIES CO., LTD. Page 7华为保密信息,未经授权禁止扩散
Why would we learn to optimize in Communications?
• « A Mathematical Theory of Communication », Bell System Technical Journal, 1948, C. E. Shannon
HUAWEI TECHNOLOGIES CO., LTD. Page 8华为保密信息,未经授权禁止扩散
Deep Communications now?
• Models are expensive to obtain
• The E2E objective function is not defined mathematically
• High Dimensional space with many parameters
• Optimization is complex and difficult to perform
9
Networks are becoming very complex
B2C
B2H
B2V
WTTx
4GHD Video
Voice
MBB
Data
Smart
Transportation
Smart Energy
AR/VR
WTTx
VoLTE
Smart
manufacturingSmart
meter
3D
Video
DroneuRLLC
eMBB
mMTC
One Network 4 Services One Network Hundreds of
Services
V
R
Tactile
internetAR
Multi-person video
call
Autonomous car
Real time gaming
Disaster alert
Automotive
ecall
Sen
sor
s
Device remote control
Source: GSMA/Huawei X Labs
Diversified Services SLA
Requirement
HISILICON SEMICONDUCTORHUAWEI TECHNOLOGIES CO., LTD. Page 10
Wireless AI: Key Technology AI/ML
MLCV
Robo NLP
Exp. Sys.…
DL
What is Wireless AI
ms
seconds
minutes
hours
days
months
Execution
AI Algorithm Wireless Algorithm
Value Data Link
Scenario Automatically Manually
Target Global probability optimization
Local determined optimization
Scope E2E network Locally Modelling
method Big data, learning Formula , optimization
Usage Set the target goal Tune parameters manually
Big Data
Network
TRMRRM
RTT IRF ANT
NLPSMBB
RNP/O
Chipset
Product
Wireless Alg.
AI in Wireless Network
Comparison
RTT
RRM
MBB
OSS
Wireless Brain
Goal oriented and self control in network
management and optimization solution, can
overcome the problem when the network cannot
be accurately expressed with formula based on
big data and machine learning technology.
RB
Feature
Feature Carrier Cell RAT Slice
RTTRRM
OSS/SON
MBB/Core Conf. and Opti.
Policy & Monitor
HISILICON SEMICONDUCTORHUAWEI TECHNOLOGIES CO., LTD. Page 11
Failure Detection and Analysis
Network Planning and Optimization
More Accurate
More Intelligent
More Faster
Wireless Brain
Network Resource Management
Architecture
Platform
Dataset
AI in wireless network
Physical Sub-health detection VoLTE root cause analysis Failure prediction Network security risk analysis
Traffic prediction Experienced network AI in SON Use behavior analysis
CA policy selection Slice resource management Intelligent base station MEC deployment
AI/ML technique is designed into the network pipe, to
enable wireless network autonomic
Improve the network operations
Reduce the complexity of network fault diagnosis
New deep learning network architecture has
been proposed
Rapid development of unsupervised learning
Development of AI chipset /TPU
Technique trend in AI/ML
Trend when AI/ML used in wireless network
Reconstruct Wireless network using AI technique
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
AI in Wireless
AI NE AINetwork
AI Data AI chipset AI service
AI base station
Network auto configuration
Failure detectionPolicy management
RAN minimal deployment
RAN AI chipset
DTX adaptation
Self operation
MBB AI Slice
E2E Performance learning
Comp mode selectionLink Adaption
PAPR non-linear compensation
• Regression• Clustering• Classification
• Deep Learning• RDL• Transfer Learning• Graphic algorithms
Slice resource management
ASFN adjustment
• Re-enforcement learning• Dynamic optimization• GMM/HMM• Association rule mining
HF/LF collaboration
LTE power control
HF channel map construction
AI management platform
Alg
orith
mA
rchite
cture
AI Center Trainer• Learner• Policy• Explorer
eNB AI Agent• Decides action• Determine state
Feature Statistic Collector• 1 Collects Statistics• 2 Applies Action
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Example
Page 13
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Why do we need energy efficiency in communications?
Page 14
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Page 15
Why do we need energy efficiency in communications?
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Page 16
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Page 17
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Page 18
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Page 19
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Deep Forward Neural Network
Page 20
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
GEE Maximization in Interference Limited Networks
Page 21
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
GEE Maximization Problem
Page 22
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
ANN Training & Validation
Page 23
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Online Implementation
Page 24
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Building the training set
Page 25
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Neural Network Architecture
Page 26
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
GEE Comparison. LMMSE reception is used.
Page 27
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Power Consumption. LMMSE is used
Page 28
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Example
Page 29
30
Edge User THP Affect User Experience Than Average THP
25
%
<1 1~
5
5~10 10~1
5
>15Mbps
15
%
20
%21
%19
%
Edge User Throughput and VoLTE Background
100 + 50 + + =Feature Parameter
> 𝟏𝟎𝟑𝟎Combination
Too Many Factors Affect Edge Throughput and VoLTE, Make
Optimization Too Difficult
VoLTE PLR(Packet Loss Rate) Correlated with
MOS
4.15 4.08 3.84 3.73.02 2.51 2.04
0
2
4
6
MOS
PLR
31
AI Assisted VoLTE and Edge User THP Optimization
> 𝟏𝟎𝟑𝟎Combination
240Combination
150 Factors TopN Factors
Combination
1
Optima
Data Modeling
Recognition
Extracting Feature
Agent Evaluation
Reinforcement Learning
Transfer Learning
1st Cell, 14 Round Converge
…
Experience
Library
Agent Evaluation
Reinforcement Learning
N Cell, 7 Round Converge
∙∙∙Importance
0 1
0.75
0.65
0.1
0.08
Factor1
Factor2
Importance Ratio Evaluation Importance Ratio Rank
Top N Factors AnalysisOptimal Value
Search
32
Innovation in Turkey – Edge THP Optimization
Test Area: Ankara Scope: 699 Site, 2281LCell
Ankara Top 10% Gain Cells
50%
40%
32%
Throughput Gain Distribution
0 - 10% 10 - 30%
Whole Network Gain:
User Num
4
6
8
10
9/7
/20
17
9/9
/20
17
9/1
1/2
01
7
9/1
3/2
01
7
9/1
5/2
01
7
9/1
7/2
01
7
9/1
9/2
01
7
9/2
1/2
01
7
9/2
3/2
01
7
9/2
5/2
01
7
9/2
7/2
01
7
9/2
9/2
01
7
10/1
/201
7
10/3
/201
7
10/5
/201
7
10/7
/201
7
10/9
/201
7
10/1
1/2
01
7
User Throughput < 1Mbps Ratio (%)
Optimization
Optimization
<1 Mbps Use Ratio Decrease while User Num Increase.
33
Innovation in Turkey – VoLTE Optimization
Test Area: Bursa Scope: 188 Site, 877 LCell
Whole Network
Gain:
58%
10%
32%
DL PLR Gain Distribution
0 - 10% 10 - 30% >30%
62%20%
18%
UL PLR Gain Distribution
0 - 10% 10 - 30% >30%
Bursa Top 10% Gain Cells
00.5
11.5
22.5
33.5
9/7
/20
17
9/9
/20
17
9/1
1/2
01
7
9/1
3/2
01
7
9/1
5/2
01
7
9/1
7/2
01
7
9/1
9/2
01
7
9/2
1/2
01
7
9/2
3/2
01
7
9/2
5/2
01
7
9/2
7/2
01
7
9/2
9/2
01
7
10/1
/201
7
10/3
/201
7
10/5
/201
7
10/7
/201
7
10/9
/201
7
10/1
1/2
01
7
DL Packet Loss Rate (%)
0.30.35
0.40.45
0.50.55
0.60.65
9/7
/20
17
9/9
/20
17
9/1
1/2
01
7
9/1
3/2
01
7
9/1
5/2
01
7
9/1
7/2
01
7
9/1
9/2
01
7
9/2
1/2
01
7
9/2
3/2
01
7
9/2
5/2
01
7
9/2
7/2
01
7
9/2
9/2
01
7
10/1
/201
7
10/3
/201
7
10/5
/201
7
10/7
/201
7
10/9
/201
7
10/1
1/2
01
7
UL Packet Loss Rate (%)
Optimization
Optimization
DL/UL PLR Decrease Dramatically
PLR: Packet Loss Rate
HUAWEI TECHNOLOGIES CO., LTD.
35pt
32pt
) :18pt
Model versus Data
Page 34