The Role of Machine Learning in Network Automation Alberto - - PowerPoint PPT Presentation
The Role of Machine Learning in Network Automation Alberto Leon-Garcia University of Toronto alberto.leongarcia@utoronto.ca Acknowledgment to: Dr. Saeideh Parsaei Fard and Iman Tabrizian Outline Context: Network Automation
Alberto Leon-Garcia University of Toronto alberto.leongarcia@utoronto.ca Acknowledgment to: Dr. Saeideh Parsaei Fard and Iman Tabrizian
2
3
AI engines:
Machine learning algorithms
Neural networks Deep Learning: MLP, CNN, RNN
source : www.atis.org. report 2018, Evolution to an Artificial Intelligence Enabled Network
https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp22_ENI_FINAL.pdf
4
5
6
Urbanization Challenges
7
Source Domain Data Gathering and Storage Data Preparation Analyzing, Optimizing and Learning Preparing a plan and related parameters for action Execution and Implementation Destination Domain
Step 1: Monitoring Step 2: Analysis
Step 3: Planning Step 4: Execution
Knowledge
10
Evolution to Network Intelligence
Evolution to Mgmt & Ops Intelligence
Actuation Sensing
Intelligent Components https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp22_ENI_FINAL.pdf
12
13
Phys. Resources Cloud Controllers (SD) Network Controllers Access/Things Controllers
SDI Resource Management
SDI Manager Topology Manager Monitoring & Analytics Multi-Tier Software Defined Infrastructure
PaaS
End-To-End, Multi Domain, Orchestration
Information-Centric Data Dissemination
BIaaS Publish/Subscribe Overlay
Algorithmic Engines Analytics Engines
APIs SaaS
Portal Custom KPIs Urban Planning Congestion pricing 3rd Party Apps
15
Service-based architecture for 5G control plane:
& communicate with each other
NWDAF
analytics to other Network Functions
NWDAF
Intelligent Components
17
Intelligent Components
18
ML Overlay
combined to form analytics fcn
for ML fcns & interfaces
heterogeneous networks
Technology-specific realization
technoolgy
collector pre-processor policy distributor
https://www.itu.int/en/ITU-T/focusgroups/ml5g/Documents/ML5G-delievrables.pdf
Management Subsystem
Multi-level ML pipeline
Closed-loop subsystem
https://www.itu.int/en/ITU-T/focusgroups/ml5g/Documents/ML5G-delievrables.pdf
ML pipeline 3 → 6
predictions at NMS affect configurations in different domains (e.g.
9→2→4→ML pipeline1
from RAN & UE/RAN to make predictions at CN (e.g., MPP). 10→7→ ML pipeline 2 → 8
from MEC platform to make predictions at the edge and apply them to MEC. Could also use side information from the UE and RAN (e.g., caching decisions made at the MEC). 3→4→ ML pipeline1 → 5
from CN and possibly UE/RAN inputs to make predictions at CN ,and apply to NMS parameters, that in turn affect configurations in different domains (e.g., SON decisions made at the CN). https://www.itu.int/en/ITU-T/focusgroups/ml5g/Documents/ML5G-delievrables.pdf
Challenges;
requirements onto the network resource availability?
Solution: 5G Network Slice Broker
interposed between external tenants and mobile network management
Hidden Technical Debt in Machine Learning Systems https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf
25
Conference on Network Softwarization and Workshops (NetSoft), Montreal, QC, 2018, pp. 168-176.
26
Use cases:
Clustering, Forecasting, and Management," 2018 4th IEEE Conference on Network Softwarization and Workshops (NetSoft), Montreal, QC, 2018, pp. 168-176. doi: 10.1109/NETSOFT.2018.8460129
29
Transport subdomain
Core subdomain
Open Interface-1 Chain of network functions of Slice Application n Chain of network functions of Slice Health Chain of network functions of Slice AR/VR Chain of network functions of Slice URRLC AI-aaS Orchestrator (s) X-MKL- Manager (s) SDN Controllers
CF1 CF3 CF2 CF4 CF5
CF1
CF3 CF2 CF4 CF5
MKL –chains of NAL MKL –chains of OTT
Intra AI-NAL Sandbox Intra Slice Sandbox
X-MKL Sandbox
SDI
Internal subdomain Third Parties subdomain
Computation domain Controllers Slice Controllers Open InterfaceI-2 SDI- manager(s)
AI-aaS Management Plane
AI-aaS Application Plane
AI-aaS Training Plane
Networking Application Layer (NAL)
e.g., load balance, QoS assurance, Anomaly detection, SLA management
Business Application Layer or Over the top (OTT-Layer)
Drones, e-factory, robotics, haptic, image processing, Big data analysis
AO- MM MM-SM AO-SM
TO
T1 T2 T3
TNAL T- OTT
34
SAVI EDGE Victoria
VM 1 VM 2 Metric Exporter
SAVI EDGE Waterloo
VM 1 VM 2 Metric Exporter
SAVI EDGE Carleton
VM 1 VM 2 Metric Exporter
SAVI Core Toronto
VM 1 VM 2 Metric Collector
Node 1 Node 2 Node 3
Kubeflow
monitoring tools preinstalled
server
”Core”
Encoder Decoder Bottleneck
Learned Encoding
the main issues for a cognitive network management is how to compress diverse types of data in more efficient manner and reduce a huge volume
around 30% (Bottleneck layer has 70 output)
Step 1: Data Monitoring
Transformer Serve (TFX)
Report
Autoencoders
Use Case 1 Step 2: AI Engines Step 3 &4: Policy & execute
applications Evaluation is based on the reconstruction error
Encoder Decoder Bottlenec k
Learned Encoding
A good tradeoff between error and compression ratio
Data ML Model Optimization problem System setting to solve optimization problem Cloud- based training Federated Learning Data is sent to the cloud to derive the model or solve the
problem
the cloud
solves its own
problem Data
Data Data Data
45