towards cognitive autonomous networks (can): 5g and beyond
TRANSCRIPT
© 2021 Nokia1Public
Towards Cognitive Autonomous Networks (CAN): 5G and Beyond
Henning Sanneck, Stephen Mwanje, Janne Ali-Tolppa, Marton Kajo & Team
Network Automation Research,Nokia StandardsMunich, Germany
© 2021 Nokia2
Modelling Cognitive Decision MakingA Perception-Reasoning pipeline
Trigger
action(s)
reflex actions subconscious actions conscious actionsExternal closed loop
Memory
Fetch(F), Read labels(R), (re)label(L) (if needed
depending on perception), and (re)store (S)
Sense Conceptualize
Contextualize
Organize
Infer
Reasoning
Perception
Public
© 2021 Nokia3
The Need for Cognitive Autonomy in Communication Networks Definitions
Tak
e one
or
more
act
ion(s
)
Diagnose events
Contextualize events/data
Anticipate standalone events
Anticipate correlated events
Correlate events/data
Detect a network event
Preset time/sequence trigger
Auto
no
mous
Cogn
itiv
e-A
uto
nom
ous
Cogn
itio
n i
ncr
ease
s
Self-Organizing – Achieve steady state without external control
o automates the selection and execution of actions
o interprets events & context to determine the cause-effect relations.
Autonomous - able to act on its own, without dictation or rules from anyone else
o not necessarily able to reason based on its environment or even smart enough to make the best static decisions
Cognitive - able to reason and formulate recommendations for subsequent behaviour
o may however require the operator’s approval
Public
© 2021 Nokia4
Static model vs. learning from dataCognition in Communication Networks
Classic AI: Model is available
o Static – fixed rule set on discrete or continuous valued logic
o Adaptive to deployment – by probabilities
Learning: Happens if performance at task T as measured by performance measure P, improves with experience E [1]
o Parameterise a model (θ) with loss function J → P=J(θ)
o Iteratively update model with experience E from data
[1] adapted from Tom Mitchell (1997))
𝜃:= 𝜃 − 𝛼𝜕
𝜕𝜃𝐽(𝜃)
Expanded & New interfaces for standardization – data & model transfer
Problem
inputsModel
Human develops
model
Process
Classic AI
std std
Problem
inputsModel
Learn patterns on data
structure or that match
data to labels
Fit inputs to patterns
Process
Build
Model
Training
Data
Machine Learning
std
std std
std
Public
© 2021 Nokia5
Autonomy
Manual Assisted Partially automated Automated Partially
autonomous
AutonomousC
og
nit
ion
Machine: None
Human execution &
supervision
Machine: assisted
execution &
supervision
Human: partial
execution
Machine: partial
execution;
Human: supervision via
policy
Machine: execution;
Human: supervision
via policy
Machine: execution &
partial supervision
Human: policy &
intent
Machine: execution &
supervision;
Human: intent-only
+ Anticipate
correlated
events
+ Anticipate
individual
events
+ Context-
ualize
Machine2human
visualization
+ Machine profiling
+Automatic re-act
+ Machine
automatic re-action
selection
+ Machine learning
of new policies
+ Machine
reasoning
+ General learning;
+Trustworthiness
+ Diagnose
eventsMachine2human
selective exposure
+ Machine mapping to
causes (rules)
+Automatic re-act
+ Human labelling
of causes identified
by machine
+ Model-free
(Reinf- Learning)
+ Transfer learning
+ Machine
explanation
+ Correlate
eventsHuman correlation Machine correlation +Automatic re-action (n.a. – due to limited
- scalability of automation- feasibility of machine supervision
in a system with low cognitive capability)Detect an
event Human detection Machine detection +Automatic re-action
Cognition vs. AutonomyTaxonomy
Human diagnosis
Human experience+ Machine prediction
+ Automatic pro-action
+ Machine automatic
pro-action selection
+ Machine
prediction of new
policies
+ Machine
reasoning
PLANARCaseStudy
© 2021 Nokia6
A live testbed created in the 5G MoNArch projectdemonstrating 5G slicing at the Hamburg Harbor:
• Three slices
• eMBB: Local applications in the harbor
• URLLC: Traffic light control
• IoT: Emission sensor readings from barges
• Data collection
• Slice-specific BTS KPIs: PRB usage, throughput, latency etc.
• UE measurements from up to three ships including position (by GPS), RSRP, RSRQ, ping etc.
• Collected for 6 months every 5 seconds~3M records
The 5G network slicing testbed at Hamburg HarborPredictive Location-Aware Network Automation for Radio
Local Applications(HPA, Veddeler Damm 18)
Heinrich HertzTower
Sensor measurements
on barges
Public
PLANARCaseStudy
© 2021 Nokia7
Problem statementPredictive Location-Aware Network Automation for Radio
• IoT requires high reliability
• In certain areas of the testbed, coverage and mobility issues are observed in the IoT slice
• Shadowing effects and/or
• Long distances from the base station
• Reliable service must be guaranteed, but without overprovisioning of resources or compromising the performance of the other slices
RSRP of USRP1 [dBm]
Ping [ms]
LOS
BTS
Non-LOS
Public
© 2021 Nokia8
Prediction of Mobility and QoS/RSRPPredictive Location-Aware Network Automation for Radio
Radio Propagation Map:
• Created based on UE measurements (reported GPS position, RSRP)
• Using a FNN
Mobility Pattern Prediction (MPP):
• Positions reported by the barges
• Prediction of barge movement using a convolutional neural network Combining the mobility prediction
with the coverage model, of 62200 sequences in a validation set, we were able to predict up to 90% of the low-RSRP events and RLFs 40 seconds ahead
RSRP, serving cell: Cell1
RSRP, serving cell: Cell2
Input sequence
Ground truth
Prediction
Public
© 2021 Nokia9
Closed-Loop Automation Evaluation with SimulationPredictive Location-Aware Network Automation for Radio
A digital twin of the testbed setup is mirrored in a simulator
• Full 3D model of the city of Hamburg and especially the harbor area
• Network topology and configuration as in the real testbed
• Traces of the movement of the real barges are collected from the testbed and imported into the simulation scenario
• The coverage issues of the real testbed can be reproduced
Public
© 2021 Nokia10
Closed-Loop Automation Evaluation with SimulationPredictive Location-Aware Network Automation for Radio
The digital twin is extended to simulate NR with beam forming
• Four beams per each of the two cells, with two antenna elements for two simultaneous beams
Public
© 2021 Nokia11
Predictive Location-Aware Network Automation for Radio Demo @ IEEE NOMS 2020
Actors
ML model inference
PLANAR
Management
Hamburg Testbed
Scenario
Recorded
movement
Trx Power
Beam Adaptation
System-
Level RAN
Simulator
pre
dictio
ns
Data Collector Deployment
Public
https://www.youtube.com/watch?v=nMdBbLv2G98
Mobility Prediction QoS prediction
© 2021 Nokia1212
Predictive Location-Aware Network Automation for Radio (PLANAR)Preventive Closed-Loop Optimization
RS
RP
[DB
]
-100
-120
-140
-160
Thro
ughput
[kbits/s
ec] 150
100
50
0Time
Measurement
Prediction
(40 sec. ahead)
Predictied RLF
Corrective action deployed
40 s
RLFHO
Public
QoS and RLFs can be predicted and prevented by:
• Optimizing the transmission power
• Beam forming
• Activate or de-active beams
• Tilting of individual beams
• Prediction allows higher thresholds to be used in the optimization algorithms
• Minimized overprovisioning of resources
• Minimized compromises between different network slicesTX Power change
Demo GUI
Ce
llID
Beam tilt
© 2021 Nokia13
Project (kick-project.de)
Application of AI methods in a highly dynamic wireless network and flexibly reconfigurable factory environment for monitoring and controlling communication and production
Artificial Intelligence for Campus Communication
Stuttgart Munich
Public
© 2021 Nokia14
SECURITY
ARCHITECTURE
ETHICS
SUSTAINABILITY
AI
GOVERNANCE INNER AI
EXECUTIONACTION
INTER AI
AI in Cognitive Autonomous NetworksEnabling areas
Public
PLANARCaseStudy
PLANARCaseStudy
PLANARCaseStudy
© 2021 Nokia15
Towards Cognitive Autonomous Networks
Rule-based management functions,Software Management
LTE / early 5GSelf Organizing
Network
Learning management functions, Software Management (DevOps)
5G EvolutionCognitive
Network Management
Management of training + Comprehension (C)
Towards 6GCognitiveNetwork
(C)NF: (Cognitive) Network Functions
Learning network functions(embedded management)
Human2machine i/f
CNF
Management of training
CNF CNF
C
CNM CNMCNM
Rule-based / data-driven network functions
NF+ NF+ NF+
SW Mgmt. (DevOps)
Human2machine i/f
SON SON SON
SON operation
Rule-based network functions
NF NF NF
Human2machine i/f
SW Mgmt.
5G and Beyond (RAN)
Public
© 2021 Nokia16
Towards Cognitive Autonomous Networks5G and Beyond (RAN)
CNF
Management of training +
Comprehension
LTE / early 5GSelf Organizing
Network
5G EvolutionCognitive
Network Management
Towards 6GCognitive Autonomous
Network
SON
CNM
SON
CNM
OAM connectivity / interface setup
Resource ID allocation (beam/cell ID/RS)
Neighbour relationship setup (ANR)
Mobility mgmt. / robustness / load balancing
Anomaly Detection → Diagnosis → Healing
Mobility load balancing / traffic steering
VNF Scaling down/up; Placement
Coverage and Capacity Optimization
Resource ID allocation (beam/cell ID/RS)
Mobility robustness (MRO)
Sleeping Cell / Outage Detection
Mobility load balancing / traffic steering
VNF Scaling in/out, down/up; Placement
Basic Coverage and Capacity Optimization
OAM connectivity / interface setup
Neighbour relationship setup (ANR)
OAM connectivity / interface setup
VNF Scaling in/out
RRM & Coverage and Capacity Optimization
Traffic steering
Mobility robustness (MRO)
VNF Scaling in/out, down/up; Placement
SON
Anomaly Detection → Diagnosis → Healing
Public
PLANARCaseStudy
© 2021 Nokia17
Towards Cognitive Autonomous NetworksConclusions
• Cognition & Autonomy are crucial, separable properties• Understand system concepts and contexts to enable decision making at machine level• Actions can be taken at any step in the cognitive process
• AI Enablers: Data → Execution → Action; inter-AI & Governance• “Inner AI”: telco- / NM-specific choice and adaptation of AI Algorithms• AI requires expanded (training data) and new (model management) standardized interfaces
• LTE/early 5G: SON → 5.5G: Cognitive NM (CNM) → 6G: Cognitive Autonomous Network (CAN)• Cognition & Autonomy penetrate the u- and c-plane
• AI is the technology area to enable this• m-plane takes a new role (managing the AI on the u- and c-plane)
• Process (integration / replacement of legacy functions)• Some SON and CNM functions remain
→ Performance improvement of existing network (management) functions & enabling of new functions → Simplification from operator perspective (shift from “execution” to “supervision”)
Public
© 2021 Nokia18
Towards Cognitive Autonomous NetworksNetwork Management Automation for 5G and beyond
• Published October 2020
https://www.wiley.com/en-us/Towards+Cognitive+Autonomous+Networks+%3A+Network+Management+Automation+for+5G+and+Beyond-p-9781119586388
Book structure and outline
Background Core argumentation Solutions/Application Open challenges
1. Introduction2. Network Evolution
4. Modeling Cognition
3. SON in pre-5G Networks
5. Classical AI for reasoning
7. Cognitive Auto-Configuration
8. Cognitive Autonomy in Optimization
9. Cognitive Self-Healing
11. System Challenges of CANs
12. Towards realizing CANs
6. ML for cognitive decision making
10. Cognitive Self-Operation
© 2021 Nokia19
Towards Cognitive Autonomous NetworksRecent journal / conference publications & tutorials
• S. Mwanje, M. Kajo, J. Ali-Tolppa, Environment Modeling and Abstraction of Network States using DeepClustering, IEEE Network special issue "AI for mobile networks“ 2020
• M. Kajo, B. Schultz, Deep clustering of mobile network data with sparse autoencoders, IEEE NOMS 2020• M.L. Mari-Altozano, S. Mwanje, S. Luna-Ramirez, M. Toril, H. Sanneck, C. Gijon, A Service-centric Q-learning
Algorithm for Mobility Robustness Optimization in LTE, IEEE TNSM, 2021.• A. Masri, T. Veijalainen, H. Martikainen, J. Ali-Tolppa, S. Mwanje, M. Kajo, Machine Learning Based Predictive
Handover, IEEE IM 2021• A. Banerjee, S. Mwanje, G. Carle, Optimal configuration determination in cognitive autonomus networks, IEEE
IM 2021• A. Banerjee, S. Mwanje, G. Carle, RAN Cognitive Controller, 2nd KuVS Fachgespräch on Machine Learning and
Networking, 2020 • A. Banerjee, S. Mwanje, G. Carle, Game theoretic Conflict Resolution Mechanism for Cognitive Autonomous
Networks, IEEE SPECTS 2020• P. Kalmbach, P. Diederich, W. Kellerer, R. Pries, A. Blenk, sfc2cpu: Operating a Service Function Chain Platform
with Multi-Agent Reinforcement Learning, IEEE IM 2021• S. Mwanje, Towards Cognitive Autonomous Networks, Tutorial, IEEE NoF 2020• S. Mwanje, Towards Cognitive Autonomous Networks, Tutorial, IEEE CNSM 2020
Public