Course Introduction

Rohit Budhiraja

Machine Learning For Wireless Communications (EE698Z)

Jan 13, 2021

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 1

 Introduction

How machine learning can be applied to wireless systems

Will teach machine learning fundamentals and pick-up wireless examples to show their applications

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 2

 Wireless

Begin with single-antenna digital communication system which operate in AWGN channel

System is mathematically modeled as y = x + n

Example - satellite communications systems
Single-antenna digital communication system which operate in fixed wireless AWGN channel
Mathematically modeled as y = hx + n
Example - cellular communications

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 3

 Machine learning and wireless – classification and detection (1)
Consider single-antenna AWGN channel y = x + n
x can be BPSK/4-QAM/16-QAM/64-QAM/256-QAM/1024-QAM
If x is a BPSK signal then it can either take value +1 or −1 with let’s say equal probability
Received noisy signal will look like:

ML parlance - we need to classify receive signal either as +1 or −1

Wireless parlance - we need to detect the receive signal and declare it either as +1 or −1
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 4
 Machine learning and wireless – classification and detection (2)

If x is a 4-QAM signal then it can take following values with let’s say equal probability
√1 + √j , √1 − √j , − √12 + √j , − √12 − √j
2 2 2 2 2 2
Receive noisy signal for AWGN channel y = x + n will look like:

ML/Wirelss parlance - we need to classify/detect receive signal

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 5

 Machine learning/wireless – clustering/modulation detection
We do not even know whether x is a BPSK or QPSK modulation
Blind modulation detection
Example: Stealth communication systems
Receive noisy signal will be either of these constellations

We need to first classify which modulation was transmitted

Wireless system uses clustering technique from machine learning to count the number of clusters
Two-widely popular technique - K-means and Gaussian mixture modeling
We are building a communication system where we need to classify such modulation schemes
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 6
 Clustering example in wireless systems

Consider the following wireless system

We need to cluster the user to employ emerging modulation schemes

Routinely use clustering techniques in our research

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 7

 Multiple-input multiple-output (MIMO) wireless systems
Consider a transmitter with M antennas and receiver with N antennas.

MIMO system simultaneously transmits M different symbols. Received signal is

y1 = h11 x1 + h12 x2 + · · · + h1M xM + n1
y2 = h21 x1 + h22 x2 + · · · + h2M xM + n2
.. ..
. = .
yN = hN1 x1 + hN2 x2 + · · · + hNM xM + nN
y = Hx + n
Receive signal vector y = [y1 , · · · , yN ]T , transmit signal x = [x1 , · · · , xM ]T , receiver noise
n = [n1 , · · · , nN ]T
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 8
 MIMO systems and machine learning (1)
Channel  
h11 ··· h1M
 .. .. .. 
H= . . . 
hN1 ··· hNM
Receive signal vector is
y = Hx + n
All M symbols in transmit vector x interfere with each other at the receiver
We need to first design a receiver to recover x from y
Most common receiver is zero forcing W = (HT H)−1 HT
Wy = ỹ = WHx + Wn
= (HT H)−1 HT Hx + |{z}
Wn
ñ
= x + ñ
ML parlance: process is called regression/least squares
For this receiver to work, N ≥M - number of equations ≥ number of symbols
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 9
 MIMO systems and machine learning (2)

For W = (HT H)−1 HT , inversion of (HT H) will be numerically unstable when its
Condition number, which is ratio of largest to smallest eigenvalues, is large
λmax
Matrix (HT H) is ill-conditioned e.g. λmax = 10, λmin = 0.01 then λmin = 1000
T
Add a scaled identity matrix γI to (H H)
λmax
If we add I then λmax = 11, λmin = 1.01 then λmin
≈ 11
T −1 T
Receiver W = (H H + γI) H is called regularized zero forcing
ML parlance: process is also called regularization

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 10

 Machine learning and 5G mMTC systems (1)
Consider a mMTC system with M single-antenna mMTC devices and N-antenna base-station (BS)

Only few mMTC active devices transmit data which BS need to process
BS does not know which devices are active. All active M mMTC devices transmit simultaneously
Total number of mMTC devices M  N
Number of active mMTC devices K < N  M, where N ≥ 2K
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 11
 Machine learning and 5G mMTC systems (2)
Received signal assuming all devices are active

y1 = h11 x1 + h12 x2 + · · · + h1M xM + n1

y2 = h21 x1 + h22 x2 + · · · + h2M xM + n2
.. ..
. = .
yN = hN1 x1 + hN2 x2 + · · · + hNM xM + nN
y = Hx + n

We can view transmit signal x = [x1 , · · · , xM ]T , receive signal y = [y1 , · · · , yN ]T , and receiver noise
n = [n1 , · · · , nN ]T
To recover x from y, N ≥ M, which is not applicable here
Recall number of active mMTC devices K < N  M , where N ≥ 2K
Transmit vector x contains only K  M non-zero values x = [1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 · · · , 0]T
Transmit signal is sparse - recovery of this vector in
ML parlance - relevance vector machine
Wireless parlance - compressive sensing, sparse Bayesian learning
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 12
 Course outline
Linear modelling: least Squares Approach /zero-forcing:
Linear modelling
Regularized least squares
Linear modelling: different avatars
Maximum likelihood approach
Bayesian approach
Classification
Probabilistic/non-probabilistic classifiers
Support vector machine and relevance vector machine (sparse Bayesian learning)
Clustering
K mean, Gaussian mixture modelling
Variational Bayes
Principal component analysis
Gaussian processes
Will not teach deep learning - neural networks, re-inforcement learning, Q learning
Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 13
 Books

Text book - A First Course In Machine Learning Second Edition [FCML]

Simon Rogers, Mark Girolami, CRC Press, 2017.
Reference book - Pattern Recognition and Machine Learning by Christopher Bishop [PRML]
Christopher Bishop, Springer, 2010.

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 14

 Course evaluation and logistics

Mini-quiz-1 – 10%
Mini-quiz-2 – 10%
Mid-sem – 20%
End-sem – 50%
MATLAB simulation assignments based on class material - 10%
Take home tutorials – not graded, will hold tutorial classes to clarify doubts
In camera exams
Cheating: First offense: award zero in the quiz/assignment. Second offense - fail the course.
Videos/books/lectures/tutorial/assignments will be uploaded in Google class-room - link is shared

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 15

 Brief Introduction

Lead the Wireless System Design and OptiMization (WiSDOM) group and 5G Testbed Lab
Prototyping and algorithm design for 5G systems
Massive MIMO, mMTC systems
Millimeter and cloud radio systems
Non-orthogonal multiple (NOMA) access systems
Practising engineer
Designed 3G, 4G and 5G systems – nine years in the industry
Designing 5G testbed at IIT Kanpur
Designing a wireless system for armed forces
Work closely with Qualcomm, Nokia, Intel and startups
Courses
Machine learning for wireless communications
MIMO wireless communications
Simulation-based design of 5G NR systems
Digital communications

Machine Learning For Wireless (Rohit Budhiraja, IITK) Course Introduction 16