Application of ML in physical layer

Application of Machine Learning in Physical Layer
Communication
Domain: Wireless Communication Research
Dr. Varun Kumar
Domain: Wireless Communication Research Dr. Varun KumarDr. Varun Kumar 1 / 18

Outlines
1 Fundamental of Physical Layer Communication
2 Limitation of Classical Methodology
3 Introduction to Multi-layer Perceptron
4 Introduction to Recurrent Neural Network
5 Use Cases of ML in Physical Layer Communication
6 References

Introduction to Physical Layer :
OSI Layer
Key feature of physical layer
This data stream contain not only the information bit, but also the
frame synchronization bits, channel control bits etc.
The key role of physical layer is to transmit these data stream from
transmitter end to receiver end successfully, irrespective of the
information bit, channel control bit, etc.

Physical layer communication:
Basic of physical layer communication
1 Physical layer communication occurs either through the guided media
(UTP, co-axial cable, ﬁber optics) or unguided media (wireless).
2 There is more uncertainty to receive and properly detect the signal in
case of unguided media compare to guided media.
3 These digital bit stream is altered into a time varying analog signal by
the diﬀerent digital modulation techniques.
4 Antenna terminal propagate these electromagnetic signal from
transmitter end to receiver end.
5 When any base station (BS) is mounted in the new geographical
location (unguided media) then there is a requirement of network
training for proper physical layer communication.
6 There are various parameter, which either require for maximization or
minimization in physical layer communication. Ex- Maximization
(data rate, capacity), minimization (power, latency, bandwidth, etc)

Challenges of physical layer communication:
Nature based transmission impairment
Attenuation
Distortion or fading
Dispersion
Noise
Challenges
Interference suppression (ISI, ICI)
Acquisition of channel state information (CSI)
Minimization of outage, BER, mutual coupling, antenna correlation
Maximization of capacity or data rate at minimum utilization of
wireless resource (time, frequency, power, space, code)
Many more

Introduction to classical methodology
Introduction to classical methodology :
1. Classical approach is based on the statistical analysis.
2. Characteristic of dependent variable is modeled into a well deﬁned
mathematical form. Ex-
y = x + n ⇒ Received signal for guided media
y = hx + n ⇒ Received signal for unguided media
Note : h = f (v, f , d, g(θ)....)
v → Velocity
f → Operating carrier frequency
d → Physical separation between Tx and Rx .
g(θ) → Nature of the wireless media, where θ → (θ1, θ2, ....)
Many more
3. Dependent parameter is considered as a random variable.

Continued–
4. With exhaustive study of well deﬁned probability density function, a
closed-form expression is obtained.
5. This close-form expression provide the mean, variance and
information about higher order moment.
6. Statistical observation reduces the lots of computational burden with
signiﬁcant proportion.

Limitation of classical methodology:
Limitation
If dependent parameter has not well defined probability density
function (PDF).
If dependent variable depends on multiple independent parameters
and each parameter has well defined PDF then there is a maximum
chance that dependent variable has not the well defined PDF. EX
y =
x1 + x2
2
x3
+ x4
y → Dependent variable
x1, x2, x3, x4 → Independent variable
For convenience, let x1, x2, x3, x4 → all are Gaussian distributed random
variable then,
Q-What will be the PDF of dependent variable y ?

Case study:(received signal through unguided media)
Received signal expression (composite fading scenario)
y = hx + n (1)
h = gβ1/2 → Channel coefficient β = z
(r/r0)ν
x → Transmitted signal, r → Physical separation between Tx and Rx
r0 → Reference distance for far field communication, n → Noise
Statistical observation
g → Complex Gaussian RV in case of slow fading
n → Gaussian distributed random variable
β → Large scale fading coeff, β = f (z, r, ν)
z → Log-normal random variable
r → Poisson distributed random variable.
ν → Path loss coefficient (depends on nature of the media)

Continued–
Shannon’s Capacity Relation
C = log2(1 + SNR) = log2 1 +
P|h|2
σ2
n
(2)
where, P = E(xx∗) and σ2
n = E(nn∗)
Note: There may be possibility of many extensions for the above equation.
From (2), dependent variable C depends on multiple independent variables
having distributions are,
Gaussian distribution
Log-normal distribution
Chi-square distribution
Poison distribution and many more
Important conclusion:
Under speciﬁc condition, sub-optimal solution of C can be calculated.

Effective state of art for physical layer communication:
Exact PDF calculation is very difficult, considering independent
variable having different distribution and all independent variables are
related with different arithmetic and other mathematical operation.
Important specification of 5G:
1 Ultra-reliable low latency (URLLC)
2 Enhanced mobile broadband (eMBB) → 5G NR
3 Massive machine type communication (mMTC)
Approach towards 5G beyond
Tera-Hertz communication
Commercially viable mm-Wave communication
Free space optics
Many more

Multi-layer perceptron : A machine learning methodology
Terminology and key features :
Feed-forward neural network : It is an artiﬁcial neural network wherein
connections between the nodes do not form a cycle.
A multi-layer perceptron (MLP) is a class of feed-forward artiﬁcial
neural network (ANN).
It can distinguish data that is not linearly separable.

Recurrent neural network (RNN)
Recurrent neural network (RNN)
In RNN, there is requirement of feedback system.
A cycle/cycles makes between input-output mapping. Ex- hidden
layer outcome may be feed as an input to previous hidden layer from
below ﬁgure.
It is more robust, where optimal weight can be assigned with greater
accuracy.

Learning
As per the application perspective, MLP and RNN may have a wider
application in physical layer communication.
Learning in context of deep neural network
Oﬄine : There is no time bound for estimating the required
parameter. Ex-
Calibration of hardware unit for working in a real-time scenario.
Analysis of call pattern.
Online : Estimating the parameter within a limited time constraints.
Coherent channel estimation and detection
Error detection and correction
Repeat and request operation in signal transmission

Use Case 1
MIMO Detection:
Downlink scenario is taken consideration.
Information bits are passed through N RF chain and it is translated
across M number of antenna.
Through M antenna, information bits are sent to K mobile users.

Use Case 2
From information bits transmission to detection by end-user, multiple
operation can be modeled through feed-forward network and recurrent
network.
Real time modulation classification:
For different application, various modulation techniques are used, such as
BPSK, QPSK, 16-QAM, 32-QAM,...256-QAM.
Note 1:
If signal strength is very poor, then BPSK is more preferable over
higher order modulation for maintaining constant BER.
Higher order modulation quantify the transmission efficiency of
wireless network.
Note 2:
Multiple devices experience different SNR and improved transmission
efficiency is the ultimate goal physical layer communication.

Conclusion
In this context a huge amount of data processing and management is
required. Hence machine learning may be the possible path.
Conclusion
1 Machine learning may be helpful for the physical layer communication.
2 Due to proliferation of larger computational capability, classical
problem can be solved up-to some extent.
3 In current scenario, all problems related to physical layer
communication can not be mathematically modeled.
4 Due to advancement of wireless technology, we have ample amount of
training data under diﬀerent scenario, that can be beneﬁcial for
analytical modeling.
5 Still the usage of machine learning in physical layer communication is
in an infantry stage, but in upcoming future it will lead for solving
physical layer communication problem

References
E. Alpaydin, Introduction to machine learning. MIT press, 2020.
J. Grus, Data science from scratch: ﬁrst principles with python. O’Reilly Media,
2019.
T. M. Mitchell, The discipline of machine learning. Carnegie Mellon University,
School of Computer Science, Machine Learning , 2006, vol. 9.
S. Haykin, Neural Networks and Learning Machines, 3/E. Pearson Education
India, 2010.

Application of ML in physical layer

In this document

More Related Content

What's hot

Similar to Application of ML in physical layer

More from VARUN KUMAR

Recently uploaded

Application of ML in physical layer