The Scope for Robotic Process Automation & Machine Learning in Telecom Operations

The Scope for Robotic Process Automation
& Machine Learning in Telecom Operations
James Crawshaw
Senior Analyst
Heavy Reading

First, the earth cooled …
Strowger switch (1891):
“Girl-less, cuss-less, out-of-order-less, wait-less telephone".
First exchange (1877):
Built from "carriage bolts, handles from teapot lids and bustle wire"

Evolution of automation
Time
Degree of automation
Electromechanical switch
Electronic switch
OSS
Telephone exchange
SDN, NFV, IBN, OMG
Brutal Automation
Ferocious, Savage, Bloodcurdling Automation

Intelligent process automation per McKinsey
1. RPA: software tool that automates routine tasks such as data
extraction and cleaning. Robot has user ID and performs rules-
based tasks such as accessing programs and systems, performing
calculations, creating reports, and checking files.
2. ML: algorithms that identify patterns in structured data, through
supervised or unsupervised learning, and provide insights or make
predictions.
3. NLP: a way for computers to analyze, understand, and derive
meaning from human language.

Typical RPA use cases
Use case Challenge Outcome
Customer Service &
Support Desk
Average time for 1st line support to
execute user admin tasks = 6 mins
Task time reduced to 50 seconds with
2 week project saving $1m annually
IT & Infrastructure
Support
Managed service provider
Needed a skilled support solution expert
for a 24/7 operation.
Bot created to handle admin tasks
including validation, regular testing,
diagnostics and fault remediation.
Data Migration &
Management
12 employees assigned to manually rekey
records from one system to the other.
2 bots created to replace 12 FTE
Back Office
Administration
300 staff using 18-year old mainframe
system and spreadsheets to process
queries in 20 minutes
Process time reduced to 2 minutes; 12
FTE saving; investment payback 2
months.
Connecting Process
Islands
HR starter/leaver process used disparate
systems with email handoff between
departments
Processing times reduced by 90%
Source: Thoughtonomy

Examples of RPA in telecoms
• Telefonica UK - SIM swaps, credit checks, order processing, customer
reassignment, unlatching, porting, ID generation, customer dispute
resolution and customer data updates.
• AT&T - migration of customer accounts from DirectTV, completing
requirement documents for Ethernet services, automation of sales
order entry, reconciliation of revenue against assets and inventory.
• Deutsche Telekom - 1,000 robots supporting 30m transactions per
year and saving 800 FTE
• Tele Danmark (TDC) - Customer facing (order handling and customer
care), Operations (updating drawings used in network planning),
Finance
• Elisa - automated pay-TV ordering; payback 3 months.

Specialized software versus BPMS vs RPA
Source: Blue Prism

RPA conclusions
• RPA is effective for processes that require predictable and high
frequency interactions with multiple applications.
• RPA is highly scalable - a robotic workforce can be doubled almost
instantly when new products are about to be launched, and then
scaled back after the surge.
• RPA can be a cost-effective alternative to Business Process
Outsourcing of back office functions.
• Telecom operations include many mundane and repetitive but
essential processes that require multiple systems to be queried
and/or updated to complete the task. The tasks must be completed
reliably and accurately - a textbook case for RPA.

ML vs AI
“Machine learning is a subfield of AI, but it’s grown so large and
successful that it now eclipses its proud parent.” – Pedro Domingos,
The Master Algorithm
AI
ML
Deep
learning
E.g. pattern recognition, statistical modeling,
data mining, predictive analytics, data
science, self-organizing systems, etc.
E.g. Multilayer
Perceptron
E.g. Knowledge base

Machine Learning types
• Supervised learning: in addition to the data inputs we also show the
system the desired outputs and ask the system to create its own
mapping of inputs to output conditions.
• Classification systems e.g. Bayesian networks; Decision trees; Logistic
regression; Random Forest; Support Vector Machine.
• Regression systems e.g. KNN; Linear regression. A regression algorithm
indicates a statistical relationship between two, or more, variables (e.g.
temperature and noise).
• Unsupervised learning: system tries to find patterns in the input data
without knowing of any specific output conditions of the data.
Clustering techniques are normally applied to unsupervised learning
systems. Examples of clustering algorithms include Apriori;
Distribution-based (Gaussian mixture models); K-Means.

Why the resurgence of interest in ML?
• Breakthroughs in neural network theory around 2006
• Improvements in computing capacity: x86 CPUs, GPUs, FPGAs and
custom ASICs designed specifically for ML e.g. Google's Tensor
Processing Unit (TPU) and associated Tensor Flow software libraries.
• The cloud makes computing capacity highly available and cheap.
• Massive data sets: online photos, email, video, gaming, search,
messaging, mapping and shopping are fertile ground for ML.
• Success stories: AlphaGo, Google Pixel Buds, image recognition, lip
reading, etc.

Adoption of ML in telecoms
29%
31%
22%
18%
Are you deploying AI?
We're taking a wait and see approach
We have started proofs of concept
We are working with suppliers to build AI intro some products and services
We have built some internal AI expertise and are incoporating it into product and service roadmaps
Source: TMForum, 2017

Potential use cases for ML
• Customer facing
• chatbots
• contact center optimization and compliance
• sentiment analysis
• churn prediction and prevention
• fraud detection and prevention
• Networking
• anomaly detection for OAM&P (operations, administration, maintenance and
provisioning)
• security – e.g., threat detection
• automated resolution of trouble tickets
• prediction of network faults
• performance monitoring and optimization
• route optimization – e.g. in IP transport or SDN switching
• traffic identification – e.g. for policy enforcement
• SLA monitoring and enforcement

AI use cases in telecoms

ML and analytics for network management

ML in network management
• Vodafone trialed machine learning in a Centralized Self-
Organizing Network to identify the optimal settings to
deliver VoLTE and to predict traffic hotspots.
• KDDI developed an AI-based monitoring system to predict
anomalies caused by hardware and software.
• Zhejiang Telecom implemented AI engine for route
optimization, capacity planning, traffic prediction and
dynamic optimization of the network.

Barriers to ML in networking
• Network engineers typically don’t have backgrounds that
include the kinds of mathematical training and experience
that are essential in ML.
• Lack of good data sets: most data sets are noisy,
unnormalized. Most data sets are proprietary. There is no
standard way to integrate network data with other data
sources like CPU usage, memory, etc.
• No useful “Theory of Networking”

ML in DT’s RT-NSM Architecture
Billions of
events
Entropy
filtering,
deep learning
neural
networks
Filter
Event
categorization
Advanced
clustering:
K-Means, k-NN,
mixture models,
CURE, deep
learning, etc.
Recomm
endation
engine
Operations &
optimization
engine based
on deep
learning
algorithms
Reinforcement
learning
Continuous
improvement of
operational
performance

ML code is only a small part of the job
Source: Hidden Technical Debt in Machine Learning Systems, Sculley et al.

Knowledge-Defined Networking
Source: Knowledge-Defined Networking, Mestres et al

ETSI’s Experiential Networked Intelligence

Telecom Infra Project AI/ML working group
• Will focus on application of machine-based decisioning and auto-
remediation to help carriers keep pace with the growth in network
size, traffic volume and service complexity.
• Three work streams:
• ML-based network operations, optimization and planning to enhance
intelligence in network operations areas through, for example, predictive
maintenance and dynamic resource allocation
• Customer behavior-driven service optimization to enhance the overall
customer experience, particularly for bandwidth-intensive, latency-sensitive
and/or data-heavy applications
• Multi-Vendor ML-AI Data Exchange Formats to ensure the development of
generalized ML models that are applicable across the industry.

The Route To Automation
Traditional networking Cloud networking
Manual Software-driven
Delivery time 10-20 days (enterprise service) Delivery time 5-10 minutes
Static (>1 year contracts) Dynamic (on demand services)
Source: Colt

Automation end game
If we virtualize all
our networks
functions and use
big data analytics
and deep learning
algorithms to run
the operations …
… the entire
business can be
managed by just
one monkey …
… plus a second
monkey to look at
the PowerPoint
slides of the first
monkey.

Paradox of Automation
• “The more advanced a control system is, so the more crucial may be
the contribution of the human operator.” - Dr. Lisanne Bainbridge,
Department of Psychology, University College London.
• “We appear to be locked into a cycle in which automation begets the
erosion of skills or the lack of skills in the first place and this then
begets more automation” - William Langewiesche, ex professional
airplane pilot.
• “To err is human, to really foul things up requires a computer.” -
William E. Vaughan, newspaper columnist.

The Scope for Robotic Process Automation & Machine Learning in Telecom Operations

Recommended

Recommended

More Related Content

Similar to The Scope for Robotic Process Automation & Machine Learning in Telecom Operations

Similar to The Scope for Robotic Process Automation & Machine Learning in Telecom Operations (20)

More from James Crawshaw

More from James Crawshaw (19)

Recently uploaded

Recently uploaded (20)

The Scope for Robotic Process Automation & Machine Learning in Telecom Operations