Implementing Cognitive Radio

Cognitive Radio Technologies, 2008

1

Implementing Cognitive Radio

How does a radio become cognitive?


2/77

Presentation Overview• Architectural Approaches • Observing the Environment

– Autonomous Sensing– Collaborative Sensing– Radio Environment Maps and Observation Databases

• Recognizing Patterns– Neural Nets– Hidden Markov Models

• Making Decisions– Common Heuristic Approaches– Case-based Reasoning

• Representing Information• A Case Study


3

Architectural OverviewWhat are the components of a cognitive radio and how do they relate to each other?


4/77

Strong Artificial Intelligence• Concept: Make a machine aware

(conscious) of its environment and self aware

• A complete failure (probably a good thing)


5/77

Weak Artificial Intelligence• Concept: Develop powerful (but limited)

algorithms that intelligently respond to sensory stimuli

• Applications– Machine Translation– Voice Recognition– Intrusion Detection– Computer Vision– Music Composition


6/77

Implementation Classes

• Weak cognitive radio– Radio’s adaptations

determined by hard coded algorithms and informed by observations

– Many may not consider this to be cognitive (see discussion related to Fig 6 in 1900.1 draft)

• Strong cognitive radio– Radio’s adaptations

determined by conscious reasoning

– Closest approximation is the ontology reasoning cognitive radios

l In general, strong cognitive radios have potential to achieve both much better and much worse behavior in a network.


7/77

Weak/Procedural Cognitive Radios

• Radio’s adaptations determined by hard coded algorithms and informed by observations

• Many may not consider this to be cognitive (see discussion related to Fig 6 in 1900.1 draft)– A function of the fuzzy definition

• Implementations:– CWT Genetic Algorithm Radio– MPRG Neural Net Radio– Multi-dimensional hill climbing DoD LTS (Clancy)– Grambling Genetic Algorithm (Grambling)– Simulated Annealing/GA (Twente University)– Existing RRM Algorithms?


8/77

Strong Cognitive Radios• Radio’s adaptations determined by

some reasoning engine which is guided by its ontological knowledge base (which is informed by observations)

• Proposed Implementations:– CR One Model based reasoning (Mitola) – Prolog reasoning engine (Kokar)– Policy reasoning (DARPA xG)


9/77

DFS in 802.16h• Drafts of 802.16h

defined a generic DFS algorithm which implements observation, decision, action, and learning processes

• Very simple implementation

Modified from Figure h1 IEEE 802.16h-06/010 Draft IEEE Standard for Local and metropolitan area networks Part 16: Air Interface for Fixed Broadband Wireless Access Systems Amendment for Improved Coexistence Mechanisms for License-Exempt Operation, 2006-03-29

Channel AvailabilityCheck on next channel

Available?

Choose Different Channel

Log of Channel Availability

Stop Transmission

Detection?

Select and change to new available channel in a defined time with a max. transmission time

In service monitoring of operating channel

Channel unavailable for Channel Exclusion time

Available?

Background In service monitoring (on non-

operational channels)

Service in function

No

No

No

Yes

Yes

Start Channel Exclusion timer

Yes

Learning

Observation

Decision, Action

Decision, Action

Observation


10/77

Security

User Model

Policy Model

WS

GA

Evolver

|(Simulated Meters) – (Actual Meters)| Simulated Meters

Actual Meters

Cognitive System Module

Cognitive System ControllerChob

Uob

Reg

Knowledge BaseShort Term MemoryLong Term Memory

WSGA Parameter SetRegulatory Information

Initial ChromosomesWSGA Parameters

Objectives and weights

System Chromosome

}max{}max{

UUU

CHCHCH

USDUSD

Decision Maker

CE

-user interface

User Domain

User preferenceLocal service facility

Policy Domain

User preferenceLocal service facility

Security

User data securitySystem/Network security

X86/UnixTerminal

Radio-domain cognitionRadio

Resource Monitor

Performance API Hardware/platform API

Radio

Radio Performance

Monitor

WMS

CE-Radio Interface

Search SpaceConfig

ChannelIdentifier

WaveformRecognizer

ObservationOrientation

Action

Example Architecture from CWT

Decision

Learning

Models


11/77

Architecture Summary• Two basic approaches

– Implement a specific algorithm or specific collection of algorithms which provide the cognitive capabilities

• Specific Algorithms– Implement a framework which permits algorithms to be changed based

on needs• Cognitive engine

• Both implement following processes– Observation, Decision, Action

• Either approach could implement– Learning, Orientation– Negotiation, policy engines, models

• Process boundaries may blur based on the implementation– Signal classification could be orientation or observation

• Some processes are very complementary– Orientation and learning

• Some processes make most intuitive sense with specific instantiations– Learning and case-based-reasoning


12

Observations

How does the radio find out about its environment?


13/77

The Cognitive Radio and its Environment

• Spectrum information is provided by the network

• Spectrum information is shared by other cognitive radios

• Observes user's applications, incoming/ outgoing data streams

• Performs speech analysis

User

• Passively "listens" to the spectrum

• Performs channel quality estimation

Spectrum

(communication opportunities)

• Receives GPS signals to determine position

• Parses short-range wireless broadcasts in buildings or urban areas for mapped environment

• Observes the network for e.g. weather forecast, reported traffic jams, …etc.

•Measures temperature, light level, humidity, …

Environment

(physical quantities, position, situations)

Other opportunities to get information

How the cognitive radio gets the information?

Information is about


14/77

Signal Detection• Optimal technique is matched filter• While sometimes useful, matched filter may not be

practical for cognitive radio applications as the signals may not be known

• Frequency domain analysis often required• Periodogram

– Fourier transform of autocorrelation function of received signal– More commonly implemented as magnitude squared of FFT of

signal

0

00

22

0

1lim

2

T j Ftxx TT

P F x t e dT


15/77

Comments on Periodogram• Spectral leaking can mask weak signals• Resolution a function of number of data points• Significant variance in samples

– Can be improved by averaging, e.g., Bartlett, Welch– Less resolution for the complexity

• Significant bias in estimations (due to finite length)– Can be improved by windowing autocorrelation, e.g., Blackman-Tukey

Estimation Quality Factor Complexity

Periodogram 1

Bartlett 1.11 N f

Welch (50% overlap)

1.39 N f

Blackman-Tukey

2.34 N f

2

0.9log

2

N

f

2

1.28logN

f

2

5.12logN

f

2

var

xx

xx

E P fQ

P f

Quality Factor


16/77

Other Detection Techniques• Nonparametric

– Goertzel – evaluates Fourier Transform for a small band of frequencies

• Parametric Approaches– Need some general characterization (perhaps as

general as sum of sinusoids)– Yule-Walker (Autoregressive)– Burg (Autoregressive)

• Eigenanalysis– Pisarenko Harmonic Decomposition– MUSIC– ESPRIT


17/77

Sub noise floor Detection• Detecting narrowband signals with negative SNRs is

actually easy and can be performed with preceding techniques

• Problem arises when signal PSD is close to or below noise floor

• Pointers to techniques:– (white noise) C. L. Nikias and J. M. Mendel, “Signal processing

with higher-order spectrum,” Signal Processing, July 1993.– (Works with colored noise and time-varying frequencies) K.

Hock, “Narrowband Weak Signal Detection by Higher Order Spectrum,” Signal Processing, April 1996

– C.T. Zhou, C. Ting, “Detection of weak signals hidden beneath the noise floor with a modified principal components analysis,” AS-SPCC 2000, pp. 236-240.


18/77

Signal Classification• Detection and frequency identification alone is

often insufficient as different policies are applied to different signals– Radar vs 802.11 in 802.11h,y– TV vs 802.22

• However, would prefer to not have to implement processing to recover every possible signal

• Spectral Correlation permits feature extraction for classification


19/77

Cyclic Autocorrelation• Cyclic Autocorrelation

• Quicky terminology:– Purely Stationary– Purely Cyclostationary– Exhibiting Cyclostationarity

• Meaning: periods of cyclostationarity correspond to:– Carrier frequencies, pulse rates, spreading code

repetition rates, frame rates

• Classify by periods exhibited in R

/ 2 2

/ 2

1lim / 2 / 2

t j tx

tR x t x t e dt

t

0 0, 0x xR R

00 only for / ,xR n T n 0xR


20/77

Spectral Correlation• Estimation of Spectral Correlation Density (SCD)

– For =0, above is periodogram and in the limit the PSD

• SCD is equivalent to Fourier Transform of Cyclic Autocorrelation

/ 2

*

/ 2

1 1, ,

2 2T

t

X T Ttt

S f X t f X t f dtt T

2

T

j fX xS f R e d


21/77

Spectral Coherence Function• Spectral Coherence Function

• Normalized, i.e.,• Terminology:

= cycle frequency– f = spectrum frequency

• Utility: Peaks of C correspond to the underlying periodicities of the signal that may be obscured in the PSD

• Like periodogram, variance is reduced by averaging

0 0/ 2 / 2

xx

x x

S fC

S f S f

1xC f


22/77

From Figure 4.1 in I. Akbar, “Statistical Analysis of Wireless Systems Using Markov Models,” PhD Dissertation, Virginia Tech, January 2007

Practical Implementation of Spectral Coherence Function

( )TX tS f

N point FFT

Shift X(f - / 2) and X(f + / 2)

Correlation of

X*(f - / 2) and X(f + / 2)

Get Spectral Coherence Function

Block averageover t

X( f )x(t)

X(f - / 2)

X(f + / 2)

( )XC f


23/77

Example Magnitude Plots

DSB-SC AM BPSK

MSKFSK


24/77

- Profile -profile of SCF• Reduces data set size, but captures most

periodicities

ˆ( ) max | ( ) |xf

I C f

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1DSBSC AM Cycle Freq. Profile with SNR=0 dB Obs. Length=100

Cycle frequency /Fs

Max. A

mp

ltit

ud

e o

f S

pectr

al C

oh

ere

nce

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1MSK Cycle Freq. Profile with SNR=0 dB Obs. Length=100

Cycle frequency /Fs

Max

. Am

pltit

ude

of S

pect

ral C

oher

ence

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1FSK Cycle Freq. Profile with SNR=0 dB Obs. Length=100

Cycle frequency /Fs

Max

. Am

pltit

ude

of S

pect

ral C

oher

ence

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1BPSK Cycle Freq. Profile with SNR=0 dB Obs. Length=100

Cycle frequency /Fs

Max

. Am

plt

itu

de

of

Sp

ectr

al C

oh

eren

ce

DSB-SC AM

BPSK

MSKFSK


25/77

Combination of Signals

BPSK MSK

BPSK + MSK


26/77

Impact of Signal Strength

BPSK with SNR=-9dB

BPSK with SNR=9dBMain signature remains


27/77

Resolution• High resolution may be

needed to capture feature space– High computational burden

• Lower resolution possible if there are expected features– Legacy radios should be

predictable– CR may not be predictable– Also implies an LPI strategy

Plots from A. Fehske, J. Gaeddert, J. Reed, “A new approach to signal classification using spectral correlation and neural networks,”DySPAN 05, pp. 144-150.

AM

BPSK 200x200

BPSK 100x100


28/77

Additional comments on Spectral Correlation• Even though PSDs may overlap, spectral correlation

functions for many signals are quite distinct, e.g., BPSK, QPSK, AM, PAM

• Uncorrelated noise is theoretically zeroed in the SCF– Technique for subnoise floor detection

• Permits extraction of information in addition to classification– Phase, frequency, timing

• Higher order techniques sometimes required– Some signals will not be very distinct, e.g., QPSK, QAM, PSK– Some signals do not exhibit requisite second order periodicity


29/77

Collaborative Observation• Possible to combine estimations• Reduces variance, improves PD

vs PFA

• Should be able to improve resolution

• Proposed for use in 802.22– Partition cell into disjoint regions– CPE feeds back what it finds

• Number of incumbents• Occupied bands

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100

Grid Index X

Grid

Ind

ex Y

CPE Number = 400, IT Number = 4

Source: IEEE 802.22-06/0048r0


30/77

More Expansive Collaboration: Radio Environment Map (REM)• “Integrated database consisting of multi-domain information, which

supports global cross-layer optimization by enabling CR to “look” through various layers.”

• Conceptually, all the information a radio might need to make its decisions.– Shared observations, reported actions, learned techniques

• Significant overhead to set up, but simplifies a lot of applications• Conceptually not just cognitive radio, omniscient radio

From: Y. Zhao, J. Gaeddert, K. Bae, J. Reed, “Radio Environment Map Enabled Situation-Aware Cognitive Radio Learning Algorithms,” SDR Forum Technical Conference 2006.


31/77

Example Application: • Overlay network of secondary

users (SU) free to adapt power, transmit time, and channel

• Without REM:– Decisions solely based on link

SINR• With REM

– Radios effectively know everythingUpshot: A little gain for the secondary users; big gain for primary users

From: Y. Zhao, J. Gaeddert, K. Bae, J. Reed, “Radio Environment Map Enabled Situation-Aware Cognitive Radio Learning Algorithms,” SDR Forum Technical Conference 2006.


32/77

Observation Summary• Numerous sources of information available• Tradeoff in collection time and spectral resolution• Finite run-length introduces bias

– Can be managed with windowing

• Averaging reduces variance in estimations• Several techniques exist for negative SNR detection and

classification• Cyclostationarity analysis yields hidden “features” related

to periodic signal components such as baud rate, frame rate and can vary by modulation type

• Collaboration improves detection and classification• REM is logical extreme of collaborative observation.


33

Pattern Recognition

Hidden Markov Models, Neural Networks, Ontological Reasoning


34/77

Hidden Markov Model (HMM)• A model of a system which behaves like a Markov chain

except we cannot directly observe the states, transition probabilities, or initial state.

• Instead we only observe random variables with distributions that vary by the hidden state

• To build an HMM, must estimate:– Number of states– State transition probabilities– Initial state distribution– Observations available for each state– Probability of each observation for each state

• Model can be built from observations using Baum-Welch algorithm

• With a specified model, output sequences can be predicted using the forward-backward algorithm

• With a specified model, a sequence of states can be estimated from observations using the Viterbi algorithm.


35/77

Example• A hidden machine selects balls from an

unknown number of bins.• Bin selection is driven by a Markov chain.• You can only observe the sequence of balls

delivered to you and want to be able to predict future deliveries

Hidden States (bins)

Observation Sequence


36/77

HMM for Classification• Suppose several different HMMs have been calculated

with Baum Welch for different processes• A sequence of observations could then be classified as

being most like one of the different models• Techniques:

– Apply Viterbi to find most likely sequence of state transitions through each HMM and classify as the one with the smallest residual error.

– Build a new HMM based on the observations and apply an approximation of Kullback-Leibler divergence to measure “distance” between new and existing HMMs. See M. Mohammed, “Cellular Diagnostic Systems Using Hidden Markov Models,” PhD Dissertation, Virginia Tech, October 2006.


37/77

System Model for Signal Classification


38/77

Signal Classification Results


39/77

Effect of SNR and Observation Length

0 5 10 15 20 25 30 35 400

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Observation Length

Det

ecti

on R

ate

BPSK Signal Classification Performance for various SNRs and Observation Lengths (BPSK HMM with 9dB)

-12dB

-9dB

-6dB

-3dB

0dB

3dB

6dB

9dB

• BPSK signal detection rate of various SNR and observation length(BPSK HMM is trained with 9dB)

• Decreasing SNR increases observation time to obtain a good detection rate

Observation Length (One block is 100 symbols)

0%

50%

100%

Dete

cti

on

R

ate

- 9dB

- 12dB

0 5 10 15 20 25 30 35 40

- 6dB


40/77

Location Classifier Design• Designing a classifier requires two fundamental steps

– Extraction of a set of features that ensures highly discriminatory attributes between locations

– Select a suitable classification model

• Features are extracted based on received power delay profile which includes information regarding the surrounding environment (NLoS/LoS, multipath strength, delay etc.).

• The selection of hidden Markov model (HMM) as a classification tool was motivated by its success in other applications i.e., speech recognition.

Collect statistics to form signature

features

Use pattern matching algorithm to classify features

Location of interest


41/77

Determining Location by Comparing HMM Sequences

Featureextraction

Vectorquantization

HMM forposition 2Candidate received

power profile

v*

Observationsequence,O

HMM forposition 1

HMM forposition n

Selectmaximum

p(O|1 )

p(O|n )

p(O|2 )

v*=argmax [p(O|v ]

1 v n

Index ofrecognizedposition,

• In the testing phase, the candidate power profile is compared against all the HMMs previously trained and stored in the data base.

• The HMM with the closest match identifies the corresponding position.


42/77

Feature Vector Generation

• Each location of interest was characterized by its channel characteristics i.e., power delay profile.

• Three dimensional feature vectors were derived from the power delay profile with excess time, magnitude and phase of the Fourier transform (FT) of the power delay profile in each direction.

-100 0 100 200 300 400 500-95

-90

-85

-80

-75

-70

-65

-60

Excess delay (ns)

Receiv

ed p

ow

er

(dB

m)

Power delay profile

Noise threshold = -80 dBm


43/77

Measurement Setup Cont.

RX 1.1

RX 1.2

RX1.4

RX 1.3

• Transmitter location 1 represents NLOS propagation from a room to another room, and from a room to a hallway. The transmitter and receivers were separated by drywall containing metal studs.

• The transmitter was located in a small laboratory. Receiver locations 1.1 – 1.3 were in adjacent rooms, whereas receiver location 1.4 was in an adjacent hallway. Additionally, for locations 1.1 – 1.3, standard office dry-

erase “whiteboard” was located on the wall separating the transmitter and receiver.

Measurement Locations 1.1 – 1.4, 4th Floor, Durham Hall, Virginia Tech. The transmitter is located in Room 475, Receivers 1.1 and 1.2 are located in Room 471; Receiver 1.3 is in the conference room in the 476 computer lab, and Receiver 1.4 is located in the hallway adjacent to 475.

~17’

~58

’


44/77

Vector Quantization (VQ)• Since the discrete observation

density is required to train HMMs, a quantization step is required to map the “continuous” vectors into discrete observation sequence.

• Vector quantization (VQ) is an efficient way of representing multi-dimensional signals. Features are represented by a small set of vectors, called codebook, based on minimum distance criteria.

• The entire space is partitioned into disjointed regions, known as Voronoi region. Example vector quantization in

a two-dimensional space.http://www.geocities.com/mohamedqasem/vectorquantization/vq.htm

*

*


45/77

Classification Result• A four-state HMM was used to represent each location (Rx

1.1-1.4). • Codebook size was 32 • Confusion matrix for Rx location 1.1-1.4

Position Rx 1.1 Rx 1.2 Rx 1.3 Rx 1.4

Rx 1.1 95% 5% 0% 0%

Rx 1.2 5% 95% 0% 0%

Rx 1.3 0% 0% 100% 0%

Rx 1.4 0% 10% 0% 90%

Correct classification

HMM based on Rx location (estimated)

Ca

ndi

dat

e r

ece

ive

d p

ower

pro

file

(tru

e)

Overall accuracy 95%


46/77

Some Applications of HMMs to CR from VT

• Signal Detection and Classification

• Position Location from a Single Site

• Traffic Prediction

• Fault Detection

• Data Fusion


47/77

The Neuron and Threshold Logic Unit

• Several inputs are weighted, summed, and passed through a transfer function

• Output passed onto other layers or forms an output itself

• Common transfer (activation) functions

– Step

– Linear Threshold

– Sigmoid

– tanh

1

0

af a

a

Neuron

Image from: http://en.wikipedia.org/wiki/Neuron

a

Threshold Logic Unit

f (a)

w1

w2

wn

x1

x2

xn

1

0

af a

a

/

1

1 af a

e

tanha

f a

1nf a a w


48/77

Neuron as Classifier• Threshold of

multilinear neuron defines a hyperplane decision boundary

• Number of inputs defines defines dimensionality of hyperplane

• Sigmoid or tanh activation functions permit soft decisions

x1

x2

Inputs Weights ActivationActivationFunction

x1 x2

0 0

0 1

1 0

1 1

w1 w2

-0.5 0.5

a

0

0.5

-0.5

0

>0.25?

1

1

0

1

w3

0.5


49/77

Training Algorithm• Perceptron (linear

transfer function)• Basically an LMS training

algorithm• Steps:

– Given sequence of input vectors v and correct output t

– For each (v,t) update weights as

– where y is the actual output (thus t-y is the error)

• Delta (differentiable transfer function)

• Adjusts based on the slope of the transfer function

• Originally used with sigmoid as derivative is easy to implement 1k k t y w w v

1k k df at y

da w w v

/

1

1 af a

e

11

df af a f a

da


50/77

The Perceptron• More sophisticated version of TLU• Prior to weighting, inputs are processed with Boolean

logic blocks• Boolean logic is fixed during training

a

Threshold Logic Unit

f (a)

w1

w2

wn

x1

x2

xn

Boolean Logic Blocks


51/77

More Complex Decision Rules• Frequently, it is impossible

to correctly classify with just a single hyperplane

• Solution: Define several hyperplanes via several neurons and combine the results (perhaps in another neuron)

• This combination is called a neural net

• Size of hidden layer is number of hyperplanes in decision rules

x1

x2

x1

x2

Input Layer Hidden Layer

Output Layer


52/77

Backpropagation Algorithm• Just using outputs and inputs doesn’t tell us how

to adjust hidden layer weights• Trick is figuring out how much of the error can be

ascribed to each hidden neuron

o

oo

df at y

da

x1

x2

Input Layer Hidden Layer

Output Layer

m

m

m j jmm

j I

df aw

da

1k k k w w v


53/77

Example Application• Each signal class is a

multilayer linear perceptron network with 4 neurons in the hidden layer

• Trained with 199 point -profile, back propagation

• Activation function tanh• MAXNET chooses one

with largest value

295 Trials unknown Carrier, BW, 15 dB SNR

460 Trials Known Carrier, BW, -9 dB SNR

Results from: A. Fehske, J. Gaeddert, J. Reed, “A new approach to signal classification using spectral correlation and neural networks,”DySPAN 2005. pp. 144 - 150.


54/77

Comments on Orientation• By itself ontological reasoning is likely inappropriate for

dealing with signals

• HMM and Neural Nets somewhat limited in how much they can scale up arbitrarily

• Implementations should probably feature both classes of techniques where – HMMs and NNs identify presence of objects, locations, or

scenarios, and reasoning engine combines.

– Meaning of the presence of these objects is then inferred by ontological reasoning.


55

Decision Processes

Genetic algorithms, case-based reasoning, and more


56/77

Decision Processes• Goal: choose the actions

that maximize the radio’s goal

• Very large number of nonlinearly related parameters tends to make solving for optimal solution quite time consuming.


57/77

Case Based Reasoning• An elaborate switch (or if-

then-else) statement informed by “cases” defined by orientation (or context)

• “Case” identified by orientation, decision specified in database for the case

• Database can be built up over time

• Problem of what to do when new case is identified

A. Aamodt, E. Plaza (1994); Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches. AI Communications. IOS Press, Vol. 7: 1, pp. 39-59.


58/77

Local Search • Steps:

1. Search a “neighborhood” of solution, sk to find s* that that improves performance the most.

2. sk+1=s*

3. Repeat 1,2 until sk+1= sk

• Variant: Gradient search, fixed number of iterations, minimal improvement

• Issues: Gets trapped in local maxima

Figure from Fig 2.6 in I. Akbar, “Statistical Analysis of Wireless Systems Using Markov Models,” PhD Dissertation, Virginia Tech, January 2007


59/77

Genetic Algorithms• Concept: Apply concept of evolution to

searching complex spaces• Really random search with some structure• Successive populations (or generations) of

solutions are evaluated for their fitness.• Least fit solutions are removed from the

population• Most fit survive to breed replacement

members of the population• Breeding introduces mutation and cross-

overs so that new population is not identical to original population– Like parents and kids

• Lots of variants– Parents die off– Niches– Tabu for looping


60/77

Genetic Algorithm Example

PopulationFitness Breeding

PWR F MAC NET

PWR F MAC NET

PWR F MAC NET

PWR F MAC NET

7

5

9

1

PWR F MAC NET

PWR F MAC NETMAC NET

PWR FPWR F MAC NET

Cross over

Mutation


61/77

Comments on GA• Tends to result in good solution very quickly• Long time (perhaps no better than a random search) to find optimum

– Often paired with a local search• Low mutation rates can cause “genetic drift”• High mutation rates can limit convergence• Cross over is like high mutation, but without damaging convergence,

but can get stuck on local maxima• In theory, reaches global optimum, but requires more time to

guarantee than an exhaustive search• Lots of freedom in the design

– Mutation rate, cross over rate, chromosome size, number of generations, population size, number of survivors, breeding rules, surviving rules

– Even more variation used when fitness function or data sets are changing over time (e.g., set mutation rate or population as a function of fitness)

– Theoretically, best combination of parameters is a function of the characteristics of the solution space

– In practice, empirically setting parameters tends to be better (GA to program a GA?)


62/77

Simulated Annealing• Steps:

1. Generate a random solution, s*

2. If s* is better than sk, then sk, then sk+1=s*; else generate random variable r. If r is less than some function of temperature and the difference in value of sk and s* and T, then sk+1=s*.

3. From time to time decrease T so that f(sk – s*,T) decreases over time.

4. Repeat steps 1-3 until stopping criterion

• Comments:– Important to store best result– In theory, reaches global

optimum, but requires more time to guarantee than an exhaustive search

– Often finished with a local search applied to best solution

• Freedom in algorithm– Distributions for generating

s*, schedules for T, change in distributions with T

• Threshold trading can be less costly


63/77

Comments on Decision Processes• Execution time

– Case-based reasoning < Searches

• Good architectural decision is to combine approaches:– CBR except when unknown case– GA for a quick good solution– Refine with local search– Can revisit searches later when excess cycles are available

• CBR can provide initial solution(s) to search algorithms

• Sometimes simpler algorithms are all that are required and will run much faster than any of these– Adjust power level for a target SINR


64

Representing Information

How can a radio store and manipulate knowledge?


65/77

Types of Knowledge• Conceptual Knowledge

– Analytic or axiomatic– Analytic if it expresses or follows from the meaning of objects

• E.g., a mobile radio is a radio with the property of mobility

– Axiomatic – fundamental conceptual relationships not based on meaning alone

• Rules– Relationships or theorems committed to memory– Some authors draw a distinction between rules and conceptual

knowledge, but it could be argued that a rule is just an axiom (or property)

• Can be expressed symbolically (e.g., UML), ontologically, or behaviorally (e.g., GA)


66/77

Why languages to represent information?• Negotiation

– Heterogeneous devices can exchange information• Sharing learned information between devices• Permits reasoning and learning to be abstracted

away from specific platforms and algorithms– Portability, maintainability

• Permits appearance of intelligence by reasoning in a manner that appears familiar to a human

• Note: much of the preceding could also be done with behavioral knowledge (e.g., sharing GA states) but it is somewhat clumsier


67/77

Proposed Languages• UML• Radio Knowledge Representation Language

– Describes environment and radio capabilities– Part of “radioOne”

• Resource Description Language• Web-based Ontology Language (OWL)

– Proposed for facilitate queries between radios

• DAML and (used by BBN)• Issues of language interoperability, testability,

actual “thought” processes


68/77

Language Capabilities and Complexity• Increasing capabilities significantly increases complexity

Modified from Table 13.1 in M. Kokar, The Role of Ontologies in Cognitive Radio in Cognitive Radio Technology, ed., B. Fette, 2006.

Language Features Reasoning Complexity

XTM Higher order relationships None O(N)

RDF Binary Relationships None O(N)

RDFSRDF plus subclass, subproperty, domain, and range

Subsumption O(Nm)

OWL LiteRDFS plus some class constructors; no crossing of metalevels

Limited form of description logic

O(eN)

OWL-DLAll class constructors; no crossing of metalevels

General description logic

<

OWL Full No restrictionsLimited form of first order predicate logic

?


69/77

Comments on Knowledge Representation• Ontologies are conceptually very appealing for realizing

thinking machines• Personal concern that goals of very high level abstraction,

platform independence, lack of a detailed specification, and automated interoperability will lead to JTRS-like implementation difficulties (see theoretically unbounded complexity, JTRS is at least bounded)– However these are really the benefits of using ontologies…

• Building an ontology is a time-intensive and complex task• Combining ontologies will frequently lead to logical

inconsistencies– Makes code validation hard

• Encourage development of domain standardized ontologies– Policy, radio, network


70

Virginia Tech Cognitive Radio Testbed - CORTEKS -

Researchers:Joseph Gaeddert, Kyouwoong Kim, Kyung Bae, Lizdabel Morales, and Jeffrey H. Reed

Modulation schemes

Transmit spectral power mask

Policy (4 ch.)

Modulation schemes


Policy (4 ch.)

Mod. Type

Transmit Power

Symbol Rate

Center Freq.

Waveform

Mod. Type

Transmit Power

Symbol Rate

Center Freq.

Waveform

Received Image

Transmitted Image

Image Display

Received Image

Transmitted Image

Image Display

Spectral Efficiency

Bit error Rate

Packet History Display

Spectral Efficiency

Bit error Rate


Detected Interference

Current CR spectrum usage

Available Spectrum

Spectrum Display



Available Spectrum

Spectrum Display

Personal ComputerMPRG’s OSSIE (Open Source

SCA Implementation Embedded) platform

CoRTekS Software Application(Rudimentary Neural

Network based Cognitive Engine)





Arbitrary Waveform Generator AWG430

Multi-mode transmitter



Arbitrary Waveform Generator AWG710B

Signal upconverter


Signal upconverter

Real Time Spectrum Analyzer (RSA3408A)

Receiver & spectrum sensor



GPIB

GPIB

GPIB


71/77

Current Setup (CORTEKS)








Signal upconverter



GPIB

GPIB

GPIB


72/77

Current Waveform Architecture

Wireless Microphone(Interferer)

OSSIEOpen Source SCA Implementation Embedded

APPLICATION

AWG Component

RSAComponent

AWG710BRSA3408A

Cognitive Engine

Component

Assembly Controller

AWG430

Transmitter ReceiverRF Interface


73/77

CoRTekS Screenshot

Policy (4 ch.)


Modulation schemes

Waveform

Center Freq.

Symbol Rate

Mod. Type

Transmit Power

Image Display

Transmitted Image

Received Image


Bit error Rate

Spectral Efficiency

Spectrum Display

Available Spectrum




74/77

CoRTekS Decision Process

Utility Function

Policy

Acceptable BERLatencyTx Power

Required QoS

Available Spectrum

Neural NtwkMemory

Tx PowerFrequencySymbol RateModulation Type

ParametersUtility Function

Policy

Acceptable BERLatencyTx Power

Required QoS

Available Spectrum

Neural NtwkNeural NtwkMemoryMemory


Parameters


Parameters


75/77

Demonstration of CORTEKs


76/77

Implementation Summary• Broad differences in architectural approaches to implementing cognitive

radio– Engines vs algorithms– Procedural vs ontological

• Numerous different techniques available to implement cognitive functionalities

– Some tradeoffs in efficiencies– Likely need a meta-cognitive radio to find optimal parameters

• Process boundaries are sometimes blurred– Observation/Orientation– Orientation/Learning– Learning/Decision– Implies need for pooled memory

• Good internal models will be important for success of many processes• Lots of research going on all over the world; lots of low hanging fruit

– See DySPAN, CrownCom, SDR Forum, Milcom for papers; upcoming JSACs• No clue as to how to make a radio conscious or if we even should

Documents

Implementing Cognitive Radio