You are on page 1of 27

AI Technologies for the

Cognitive Radio
Dipti Deodhare
Centre for Artificial Intelligence and
Robotics
dipti@cair.drdo.in

Cognitive Radio =
Software Defined Radio
+
Cognition Engine

Cognitive Capabilities
Awareness: should be aware of its
own abilities, the regulating
policies that govern it, its
neighbours and their abilities etc.

Cognitive Capabilities
Perception: should be able to
sense its environment

Cognitive Capabilities
Learning: consequent to
perception it should be able to
learn about the general
characteristics of its environment
and trends

Cognitive Capabilities
Reasoning: relationships between
the various entities should be
understood and sound decisions
inferred, obviating the
overwhelming, and perhaps
impossible, task of enumerating
every single alternative

Cognitive Capabilities
Memory: should demonstrate
improved performance after
operating in the same environment
over an extended period of time

Architecture
Software Radio

Cognition Engine

(perception)

Cognition Engine
Learning Engine
(learning)

Reasoning Engine
(reasoning)

Knowledge Base
(awareness, memory)

Software
Radio

Reasoning
Engine
Knowledge
Base
Learning
Engine

Meters and Knobs


Cognitive capabilities are
SDR
Cognition Engine
predicated on availability
Application Read
Reasoning
meters
of suitable meters and knobs
Engine
Security
Knowledge
in the SDR.
Transport
Base
Set
Meters to continuously
knobs
Network
Learning
Data Link
monitor the performance
Engine
(MAC+LLC)
of the radio.
PHY
Knobs to reconfigure the radio so that the performance of the radio is
maintained at an acceptable level satisfying various context driven
constraints.
The mandate of the cognitive engine is to read the meters and
appropriately tune the knobs.

Layer-wise distribution of Meters


and Knobs Contd.
Application
Security
Transport
Network
Data Link Layer
(MAC + LLC)

Physical Layer

METERS
Interference,
BER, received
signal power,
noise power,
SNR, fading
statistics,
doppler spread,
delay spread,
angle of arrival,
dynamic range

KNOBS
Power, frequency band
of operation, carrier
modulation type,
baseband modulation
type, pulse shaping,
data rate, number of
channels, bandwidth,
equalization, antenna
tuning, antenna
steering, antenna height
adjustment, type of
antenna (if more than
one type of antenna
available)

Layer-wise distribution of Meters


and Knobs Contd.
Application
Security
Transport
Network
Data Link Layer
(MAC + LLC)

Physical Layer

METER
Frame error rate

KNOBS
Frame format,
Frame size,
multiple access,
duplexing, FEC,
ARQ
(enable/disable)

Learning Engine
The Learning Engine
Collection of classifiers
developed using supervised and unsupervised
learning techniques.

Numeric data
Modulation,
Carrier Freq.,
BER, SNR,
Pwr, Coding

Symbolic data

Learning
(collection of
classifiers)

Classify into
qualitative classes
such as good, bad
high, average, low

MTU Size vs Throughput


Ref: Anna Calveras Auge, Josep Paradells Aspas. Performance Optimisation Evaluation of
TCP/IP over Wireless Networks., IEEE Performance, Computing and Communications, 1998

For each BER value,


an optimal MTU
Figure is a graph for Point to Point Protocol (PPP) and
exists!
deterministic errors.
be different
Graphs
BER:would
meter
at the for other protocols like Frame
Relay, Asynchronous Transfer Mode (ATM), Ethernet etc.
physical
layer.
They
would also
vary depending on the error type, namely,
and burst.
deterministic
MTU: knob
at the
It is proposed that the graphs canBER
be learnt by
a MultiNeural
network
layer.
Layer
Perceptron
(MLP) or some
other suitable
mechanism. MTU
Network
Throughput
Reinforcement Learning mechanisms are also
potentially
Classifier
relevant.

OFDM Transmitter
IFFT
C1

incoming
bits

Modulator
(QAM
QPSK
BPSK)

C2
C3
Cn

f1
f2
f3

Parallel
to
serial

fn

OFDM transmission offers opportunity


for adaptation!

OFDM Receiver
IFFT
f1
f2
incoming
bits

Serial to
parallel

f3
fn

C1
C2
C3
Cn

Demodulator

Capacity maximization in a non-AWGN channel


can be modeled as a Reinforcement Learning
problem.
M modulation types: M1, M2, , MM.
N coding types: C1, C2, , CN.
Mi can transmit di data bits per symbol and has a
probability of bit error ei(S) for signal to noise
ratio S.
Coding type Cj has rate rj and can correct cj bit
errors per block of size bj bits.
TS: symbol time, a constant.
B: corrected bit error rate coming out of the
decoder; measured by the radio.

For modulation type Mi and coding type Cj,


the resulting capacity is given by:
Ci,j = ((dirj)/TS ) (1-B)
Ci,j can act as a performance measure.
Objective function:
max f(Mi,Cj,B)= dirj (1-B)

Knowledge Representation
Build Awareness
The Knowledge Base (KB) will be built on
the semantic web framework.
Knowledge is captured through graphs
consisting of nodes interconnected by
relations.

Knowledge Representation
Each node or relation is associated to an
ontology that defines the concept. An
ontology consists of slots representing
various attributes of that concept. For e.g.
here is the Device ontology created in
OWL (Web Ontology Language). (We use
the SWOOP editor to create and edit ontologies
and RDF Gravity to visualize them.)

Device Ontology

Reasoning
Reasoning involves traversing the semantic
graphs to obtain relevant conclusions.
Some ontologies for the cognitive radio: radio,
channel, spectrum, power, coding, modulation,
etc.
Some inferences: frequency fc is sparsely used
from time t1 to time t2; for channel c, capacity is
maximized with modulation type mi and coding
method cj and so on.

A Simple Example using


Predicate Logic (papers include authors from US DoD)
Software Radio (SR) exports predicates to the
knowledge-base regarding detected signals s1, s2,
, sN, of the form
signalFreq(si,fi) signalBW(si,Wi)

Goal: To find some fc and W that does not overlap


any detected signal, while maximizing W and
hence the radios capacity.
Define, notOverlap(fc,W,si)
= (fi+Wi/2<fc-W/2) V (fi-Wi/2>fc+W/2)

A Simple Example using


Predicate Logic
Define predicate:
action: moveBand(fold,Wold, fnew, Wnew)
precond: Vi <= N: notOverlap( fnew,Wnew,si)
postcond: -(centerFreq(fold) centerFreq(W
old))

(centerFreq(fnew) centerFreq(Wnew))

Define the Reasoning objective function:


fR(centerFreq(fc) bandwidth(W)) = W

This will give a policy-based cognitive radio that will


search out the largest continuous piece of bandwidth for
communication.

Memory
Case-based Reasoning Tools are useful
Swarm Intelligence technologies
HMMs
Reinforcement Learning
Can all play a role in searching large knowledge
banks to enable quick responses.

Preliminary Analysis for


Frequency Hopping Solutions
using DEAL Inputs
% of frequencies
jammed (this is
less than the
threshold)

Hop rate

2%
3%

FEC

7/8
Increases to reduce
jamming

Reduces to meet
BER requirements
More robust to
tolerate errors

Max < threshold

Data Rate

Establishing a Test-bed
Simulation in Matlab
GNU radio local vendors available

Thank You

You might also like