1

Cognitive Radio: From Spectrum Sharing to Adaptive

Learning and Reconfiguration
Feng Ge, Qinqin Chen, Ying Wang, and Charles W. Bostian
Virginia Polytechnic Institute and State University
Wireless @ Virginia Tech
Center for Wireless Telecommunications, Virginia Tech,
Blacksburg, VA 24061, USA
540-231-2558
{gef, chenq, ywang06, bostian}@vt.edu
Thomas W. Rondeau
Centre for Telecommunications Value-chain Research (CTVR), Trinity College, Dublin, Republic of Ireland
trondeau@mee.tcd.ie
Bin Le
Cognitive Radio Technologies (CRT), LLC
bin.le@crtwireless.com

AbstractThis paper
12
introduces important cognitive radio
developments like spectrum sharing, learning and adaptation
algorithms, and the software and hardware architecture to
support these functions. A cognitive radio is defined here as
a transceiver that is aware of its environment and can
combine this awareness with knowledge of its users
priorities, needs, operational procedures, and governing
regulatory rules. It adapts to its environment and configures
itself in an appropriate fashion. The radio learns through
experience and is capable of generating solutions for
communications problems unforeseen by its designers.
Our spectrum sharing cognitive radio is built upon GNU
Radio and uses the Universal Software Radio Peripheral
(USRP) device as our radio front end platform. We use
cyclostationary feature analysis to detect low SNR
modulated signals because of its ability to distinguish
between modulated signals, interference, and noise in low
signal to noise ratios. A parallel algorithm running on a Cell
Broadband Engine (Cell BE) is used to attack the associated
high computational complexity. A new spectrum sensing
scheme, incorporating spectrum monitoring, data
transmission, and dynamic channel switching, is designed to
fully utilize the idle time of the primary user.
Our work is based on the concept of a Cognitive Engine: an
intelligent software package that reads the meters and
turns the knobs of any attached software defined radio
(SDR) platform. Using an eclectic combination of artificial
intelligence techniques including case-based decision
theory, multi-objective genetic algorithms, and neural
networks, it implements a system of nested cognition loops.
Applied to public safety communications, this technology is
the basis of a working prototype Public Safety Cognitive
Radio that can scan the public safety spectrum (multiple
bands and multiple waveforms, all incompatible) and
configure itself to interoperate with any public safety
1
1
1-4244-1488-1/08/$25.00 2008 IEEE.
2
IEEEAC paper #1668, Revision 2, updated January 14, 2008.
waveform that it finds within 0.1 seconds of determining
that a signal is present.
TABLE OF CONTENTS
1. INTRODUCTION ..................................................... 1
2. SPECTRUM SHARING COGNITIVE RADIO ............. 2
3. THE COGNITIVE ENGINE ...................................... 4
4. PUBLIC SAFETY COGNITIVE RADIO ..................... 7
5. CONCLUSIONS ....................................................... 8
6. ACKNOWLEDGEMENTS ......................................... 9
1. INTRODUCTION
Spectrum occupancy measurements reported in 2005 by
Shared Spectrum Company (SSC) showed that the average
spectrum occupancy over all the radio bands between 30
MHz and 3,000 MHz was 5.2% as observed over multiple
typical geographical locations [1]. This is in sharp contrast
to the overcrowding of wireless communication in the
public safety bands, the WiFi band, and most other
unlicensed industrial, scientific and medical (ISM) radio
bands, etc. To alleviate this disparity, the Federal
Communications Commission (FCC) is exploring the
possibility of a spectrum sharing mechanism, by which a
secondary user can share the spectrum on conditions of non-
interference to the primary users.
Cognitive radio [2] is the most promising technology in
spectrum sharing. A cognitive radio can change its
transmitting or receiving parameters to communicate
efficiently while avoiding interference with licensed or
unlicensed users. We define a full cognitive radio as a
transceiver that is aware of its environment and can
combine this awareness with knowledge of its users
priorities, needs, operational procedures, and governing
regulatory rules. It adapts to its environment and configures
itself in an optimal fashion. The radio learns through
experience and is capable of generating solutions for
communications problems unforeseen by its designers. This
full cognitive radio envisions rapid progress in software

2
radio technology, which has been aided by advances in
processors, RF technology, and software since 1991 [3].
Therefore, a cognitive radio is the evolution of a SDR into
an automatically reconfigurable communications system
that responds to network and user demands.
This paper introduces some important cognitive radio
developments which cover both spectrum sharing and some
full cognitive radio functions. In Section 2, a Spectrum
Sensing Cognitive Radio (SSCR) design based on the SDR
software platform GNU Radio is introduced. Specifically,
we will analyze a signal detection approach based on
cyclostationary feature analysis and use the Cell BE to
attack the computational challenge. In Section 3, we extend
our SSCR closer to a full cognitive radio. We designed the
platform-independent Cognitive Engine to read the meters
and turn the knobs of any specified SDR system. This
Cognitive Engine follows our radio learning and adapting
software core in response to radio environment and user
requirement. The learning and adapting software core builds
upon case-based decision theory, multi-objective genetic
algorithms, radio and user database, and a policy engine. In
Section 4, we apply our Cognitive Engine architecture and
equip specific radio application functions on the GNU
Radio platform to design the prototype Public Safety
Cognitive Radio. It can scan the public safety spectrum
(multiple bands and multiple waveforms, all incompatible)
and configure itself to interoperate with any public safety
waveform that it finds within 0.1 seconds of determining
that a signal is present.
2. SPECTRUM SHARING COGNITIVE RADIO
A Spectrum Sharing Cognitive Radio is defined by IEEE as
a radio frequency transceiver that is designed to
intelligently detect whether a particular segment of the
radio spectrum is currently in use, and to jump into (and out
of, as necessary) the temporarily-unused spectrum very
rapidly without interfering with the transmissions of other
authorized users [22]. The most challenging technology for
SSCR is the spectrum sensing detector which should have
both low SNR sensitivity and high agility for wide band
signal detection. In general, there are two ways to achieve
both requirements: one is mainly based on expensive and
highly sensitive analog radio component for signal
detection, such as DARPAs XG program [4]; the other
procedure uses digital domain techniques by moving as
close to the antenna as possible and relies on the low-price
and fast computing processors to solve radio analog
components inefficiency or shortcomings.
We developed our SSCR following the second way, which
is essentially the development trend for SDR [5]. A
Software (Defined) Radio is defined by the FCC as a radio
that includes a transmitter in which the operating
parameters of the transmitter, including the frequency
range, modulation type or maximum radiated or conducted
output power can be altered by making a change in software
without making any hardware change [23]. Among a
number of commercial platforms and free open source ones,
we have chosen GNU Radio as our SDR software
architecture and the Universal Software Radio Peripheral
(USRP) device as our radio analog platform.
USRP and GNU Radio
The Universal Software Radio Peripheral (USRP) is an
openly designed low-price SDR hardware platform which
implements radio front-end functionality and A/D and D/A
conversion currently using the Universal Serial Bus (USB2)
to connect to the PC that hosts the device. The current
USRP device (Figure 1) consists of a motherboard
containing up to four high speed 12-bit 64 Msps analog to
digital converters (ADC), four high speed 14-bit 128 Msps
digital to analog converters (DAC), an Altera FPGA and a
programmable Cypress FX2 USB 2.0 controller. The ADCs,
DACs and the FPGA together provide support for IF
processing. The FPGA on the board provides four digital up
converters (DUC) and four digital down converters (DDC)
to shift frequencies from the baseband to the required
frequency. The FPGA can be reprogrammed to provide
additional functionality. RF front ends are attached in the
form of daughter cards which can currently cover all the
existing radio bands from 0 Hz to 2.4 GHz.

Figure 1 - Block Diagram of the USRP, adapted from [6])
Figure 2 - A basic SDR system based on GNU Radio and
USRP
GNU Radio is an open source toolkit for building software
radios [7]. It was started in early 2000 by Eric Blossom and
others and has evolved into a mature software infrastructure
used and supported by a large community of developers. It
was originally designed to run on General Purpose
Processors (GPP), combined with minimal analog radio
hardware, and allows software radio development of
waveforms, modulations, protocols, signal processing, and
other communications functions in the digital domain. The
GNU Radio signal processing library includes existing and
developing blocks for most signal processing functions,
such as waveform modulation and filter creation. It also

3
includes I/O operations like file access. Programming in the
GNU Radio platform uses a combination of C++ and
Python, a simple, high-level language: the computationally
intensive processing blocks are implemented in C++ while
the control and coordination of these blocks for applications
that sit on top are developed in Python. The USRP is fully
supported by the GNU Radio library and a combined system
of both is given in Figure 2.
Spectrum Sharing CR Architecture
The first challenge for SSCR is to design a signal detector
that can quickly search through a wide bandwidth for vacant
spectrum in order to establish a new channel. Further, the
cognitive radio has to quickly switch to another channel
when a primary user appears. We designed a SSCR based
on GNU Radio and USRP [8] with the architecture shown in
Figure 3.
Figure 3 - Spectrum sharing CR architecture
The SSCR uses two RF chains: the receiver chain, including
the monitor antenna and the channel monitoring, monitors
the working spectrum, and the transceiver chain, including
the data antenna and the data transceiving, works as the
secondary user. In this example, data transceiving is
supposed to fully utilize the idle time of a primary channel.
The URSP radio analog component control by our SSCR is
through GNU Radio functions. The channel monitor will
detect any active signals in the specified frequency range
and it will also continuously update the spectrum database.
When a primary user appears, the data transceiver will
switch from its previous working spectrum and search for
the available channel from the spectrum database that offers
the best QoS to continue the previous communication. This
process can guarantee an efficient utilization of the primary
signals idle time.
Currently SSCRs spectrum sharing scheme works in an
offline mode from the GNU Radio perspective. It collects
radio signals from USRP and digitizes them and then stores
the digital samples on the host computer through GNU
Radio functions. Next, the samples are analyzed in the
computer by the FFT accumulation method (FAM)
algorithm. The cyclic frequency result is used to determine
spectrum switching following the spectrum scheme shown
in Figure 3. We also want to point out that the FAM
algorithm for signal detection will be incorporated into the
GNU Radio function library on GPPs so that we eliminate
memory reading and writing delay and achieve better agility
in spectrum sharing.
Signal Sensing Approach
A critical requirement for opportunistic spectrum sharing is
to sense the spectrum holes quickly and accurately, so that
non-interference to privileged users is guaranteed. For
individual cognitive radios, cyclostationary feature detection
has advantages for spectrum sensing due to its ability to
differentiate between modulated signals, interference, and
noise in low signal to noise ratios. It is well suited for
signal detection and modulation recognition, signal
parameter estimation, and the design of communication
signals and systems [9].
In this paper, we use the well-known FAM algorithm [10] to
estimate the spectral correlation function (SCF) s for a
signal process x(n) over t seconds duration. The discrete
s value at each point of the two dimensional cyclic
frequency and signal frequency space is shown in equation
(1)
S
x
u
i
+qu
(nI,

)
t
= `X
1
(rI,
k
)
P-1
=0
X
1
-
(rI,
I
)
g(n -r)c
-
2nq
P
(1)
where o

+oo represents discrete cyclic frequency,

is
the discrete signal frequency, g(n) is a data tapering
window, L is a decimation factor in the frequency domain, P
equals ((N - N
i
)/ L+1) where N is the total number of
samples and N
i
is the number of samples used to calculate
each complex demodulator X
1
(rI,
k
). The choice of N
i

must take into consideration that the time-frequency
resolution product (NN
i
) must satisfy NN
i
> 1 for a
statistically reliable measurement [10] and that N
i
is large
enough to obtain the desired frequency resolution. L is
usually chosen to be less than or equal to N
i
4 [11].
We use the crest factor (CF) for signal detection and feature
extraction by exploiting cyclic frequency domain profile
(CDP) shape [12]. To evaluate
any signals presence, we used the following simple model
x(t) = s(t) +(t) (2)
where x(t)is the continuous form of the signal processing
x(n), s(t) denotes any detected signal, and (t) denotes
additive white Gaussian noise (AWGN). A threshold C
1H
is
defined when no signal is present, i.e., when s(t) = (t),
for sampled signals:
C
1H
= max
`

I
i
(o)

I
i
2
(o)
N
u=0
N
/

(S)
where I
i
() = max
I
|S
x
u
() | for x(t) = (t).
Similarly, we have
C
I
= max
`

I(o)

I
2
(o)
N
u=0
N
/

(4)
where I() = max
f
|S
x
o
() | for x(t) = s(t) +(t).
To test signal presence in AWGN, the following binary
hypothesis testing is performed.
E
0
: x(t) = (t)

4
E
1
: x(t) = s(t) +(t)
Based on the threshold C
1H
, we can test the signals
presence as follows:
C
I
< C
1H
: cclorc E
0

C
I
> C
1H
: cclorc E
1

The presence of active signal declares E
1
and its
spectrum frequency is compared to the current secondary
working spectrum. If the comparison result indicates the
primary signals presence, spectrum switching will be
triggered and the secondary user spectrum is switched to
another vacant spectrum. The active signals spectrum will
be always used to update the spectrum database.

Speedup by Cell BE
Cyclostationary feature detection has advantages for
spectrum sensing. However, a key issue with
cyclostationary signal analysis is the high computational
complexity arising from the large number of required
complex convolution operations. In addition, the
computation requirement increases significantly in
proportion to the bandwidth to be covered. These factors
limit receiver agility and sensitivity.
We have studied the well-known FAM algorithm for
cyclostationary feature analysis and have shown that the
computation required to cover a single IEEE 802.11g WiFi
channel bandwidth at a fine resolution is too high for current
GPPs [8]. Future broadband technology will use
significantly larger bandwidths, which inevitably demands
new computation methods and architectures. We have
designed a parallel FAM algorithm on a Cell BE powered
PlayStation 3 with six usable Synergistic Processing
Elements (SPEs) and one Power Processor Element (PPE).
We use two levels of parallelism: the task level, supported
by SPEs and one PPE, and the data level parallelism
supported by SIMD and Vector Multimedia Extension
(VMX) instructions. To fully utilize the Cell BEs capacity,
several available speeding techniques are also used such as
Direct Memory Access (DMA), loop unrolling, pre-
computed sinusoid and cosine arrays for FFT, double-
buffering, etc. Further, GNU Radio is being ported into a
PlayStation 3, after which we plan to run our SSCR fully in
the PlayStation 3.
3. THE COGNITIVE ENGINE
In this part, we extend our SSCR closer to a full cognitive
radio and introduce the Cognitive Engine [13] which is
designed to be independent from any specific SDR
hardware and software architecture. It is an open
architecture for developing and applying cognitive radio
algorithms and deploying cognitive radio functionality.
Specifically, the current Cognitive Engine, as shown in
Figure 4, contains seven components: the cognitive
controller, the sensor, the user interface, the optimizer, the
policy verifier, the radio platform, and the decision maker
attached with the knowledge base. The outermost layer
contains several possible functions that can be deployed in
the cognitive engine. Particularly, our SSCRs spectrum
switching resides in the decision maker, the spectrum
database resides in the knowledge base, and the spectrum
searching resides in the optimizer.
Figure 4 - Cognitive Engine architecture, adapted from
[13]
Cognition Loop
To guide our Cognitive Engine development, we proposed a
new version of the cognition loop [14] shown in Figure 5.
This new version is based on Mitolas original cognition
cycle definition [2] which consists of observe, orient,
plan, decide, learn, and act. Our cognition loop
steps are to (1) collect the radio environment parameters; (2)
synthesize the information into scenario representation; (3)
compare scenarios with the radio and user database and
decide either to use an existing successful case setting or
find another better case setting; (4) find the optimal setting
using some optimization algorithms if a new optimal setting
is needed; (5) update the database using the new optimal
parameter setting; (6) reconfigure the attached radio
platform for this scenario using the optimal parameter
setting. Our cognition loop is more closely directed towards
PHY and MAC level cognition than Mitolas definition
which is more like human cognition.
Figure 5 - Cognition loop, adapted from [14]
The Cognitive Engine Execution
In the Cognitive Engine, the cognitive controller follows the
cognition loop and coordinates and executes attached

5
function components by sending commands and data
through an interface to the components. The learning and
adapting ability is achieved through the cooperation of the
seven components.
The sensor observes external and internal environments. For
example, one signal sensor can collect external surrounding
environment parameters including propagation path, radio
position and location, and the spectrum availability, etc.
Another sensor can collect internal data, or read the
meters, which displays the radios self performance and
operating parameters such as received signal power, noise
power, bit error rate (BER), frame error rate (FER), and
battery life, etc. The decision maker and knowledge base
will first perform scenario synthesizing and case-based
decision making, and then they will estimate performance
from the optimizers result and update the knowledge base.
Our optimizer, the multi-objective based wireless system
genetic algorithm (WSGA), will perform the link configure
optimization. The policy verifier, together with the decision
maker, verifies the optimizers result to follow regulatory
rules and updates the knowledge base. The radio platform,
which can be any SDR system with both hardware and
software architecture, will receive the policy verifiers
results and reconfigure itself, or turn the knobs such as
transmit power, modulation, coding, symbol rate and
spectrum shaping, to achieve a good QoS user setting. In
addition, the human friendly user interface displays some
necessary parameters and creates command buttons bound
to the cognitive controllers control commands.
Each component is launched as a separate process that
interfaces and exchanges data between processes through
some generic interface. This architecture has two
advantages. First, it enables distributed processing, where
different components can reside on one single processor, as
in our current system, or it can reside on different processors
or hosts [13]. The other advantage is that this architecture
adopts a standard interface to enable future changes or
development of components and algorithms. Upgrades of
any component will not affect other components. This
advantage simplifies the developing and testing of the
cognitive radio system.
Application Programmable Interface (API)
The cognitive engines architecture openness is achieved
through the attachable function modules, the cognitive
controller, and the application programmable interface
(API) which interfaces the components. Essentially,
different components need a common language to exchange
domain knowledge and parameters. In addition, a
mechanism is also required to coordinate parameter and
command sending and receiving among different
components without any conflict and intolerant time delay.
To achieve the above function, we created the API which
rides on the eXtensible Markup Language (XML). XML is
used as the common language to convey command, domain
knowledge, and parameters among different components.
The operating system possessed Internet Protocol-based
network socket functions are used for sending and receiving
the aforementioned XML files. For example, the sensed data
such as propagation path, radio position and location,
received signal strength, noise power, and BER are first
formulated as a hierarchical data structure stored in XML
files in the sensor domain. Then the XML files are
transferred to the cognitive controller through the API. The
cognitive controller will parse the XML files and get the
original sensed data.
Wireless System Genetic Algorithm (WSGA)
WSGA is a multi-objective based genetic algorithm
designed to find the optimal parameter setting for
configuring the radio platform by responding to the users
requirement and the QoS. It models the physical radio
system as a biological organism and optimizes its
performance through genetic and evolutionary processes.
Radio resources encompass spectrum, power, time, etc,
which determine QoS for communication. The QoS can be
commonly specified through eight important objectives
which include: BER, signal to interference plus noise ratio
(SINR), bandwidth, spectrum efficiency, throughput, power,
computational complexity, and interference [15]. Finding
the above optimal parameter setting in terms of radio
resources is modeled as a multi-objective problem.
Objective space is defined as the set of objectives that
represent the radio performance with each one modeled as a
function of resources. Multiple objects are usually needed to
fully describe the radio performance and they may not be
independent from each other. The multi-objective problem
is modeled as:
min max (y) = (x ) = |
1
(x ), ,
n
(x )] (5)
subject to:
x = (x
1
, x
2
, , x
m
) e X
y = (y
1
, y
2
, , y
n
) e
where is the objective space, and

X represents the
resources.
The WSGA uses chromosomes to encode input parameters
like payload size, power, coding techniques, encryption,
equalization, number of sub-carriers, network protocol,
retransmission requests, and spreading technique/code. It
uses a Pareto ranking selection method [16] to determine the
chromosomes survival to the next generation in terms of
their fitness to maximize the objectives. During the fitness
evaluation, the WSGA awards points for every objective
that an individual wins. The WSGA uses one crossover and
mutation point operation chosen from uniform random
numbers with a static probability of crossover and mutation
occurring. The use of constraints to a multi-objective
problem gives the WSGA the opportunity to incorporate
regulatory and physical restrictions during chromosome
evolution. If a trait determined by the chromosome exceeds
the limits of the radio's capabilities, like finding a center
frequency outside the tunable range of the radio, or breaks

6
the law by transmitting too much power in a band, then the
WSGA forces random mutations on the gene until it is legal,
thus preserving large portions of the chromosome structure
as well as introducing legal genes into the population. The
WSGA terminates when it reaches a desired level of
objectives or the specified maximum number of generations.
In Figure 6, we show how the WSGA optimizes the
waveform. The XML file which defined the waveform
describes the available knobs and the range SDR is capable
of. Objective functions are received from the controller to
describe functions in a library, and the WSGA calls the
functions to calculate meters from the current set of knobs.
Figure 6 - WSAG waveform optimization, adapted
from [13]
Case Based Decision Maker
Built on case-based theory, the decision maker combined
with the knowledge base uses feedback to aid future
optimization which will heuristically learn from experience
instead of using pure object optimization. The optimized
problem is saved in the data-base with the solution and
performance, and when the new problem is received, the
Cognitive Engine is looking for the similarly optimized
problem [13]. The similarity between the new problem and
old cases is input as one of the augments of objective
function. This is how the Cognitive Engine learns from
experience and knowledge. The multi-objective problem is
now described as

min max (y) = g(x ) = |
1
(x ), ,
n
(x ), s
1
, s
2
, , s
k
] (6)
subject to:
x = (x
1
, x
2
, , x
m
) e X
y = (y
1
, y
2
, , y
n
) e

Where X and

are same as in Equation (5) and
) ,.., 2 , 1 ( k i s
i
= is the similarity between the new problem
and the cases stored in the database. In Figure 7, it shows
how a case-base is applied to the optimization process to
learn from feedback.
This mechanism offers a significant advantage to real time
processing in a cognitive engine where a quick solution
needs to be provided as situations and environments change.
It can find a sub-optimal parameter setting good enough to
support a QoS level within a short time instead of finding
the best parameter setting in a long time. On the other hand,
it can also narrow the searching space for the optimizer
which starts searching at the local variable space including a
similar stored case and does not have to search the whole
variable space. However, the database size and its data
organization will affect the performance in that only a
certain amount of cases can be stored in the database. The
Cognitive Engines decision rule governs the organizing and
manipulation of cases relative to time and prioritization. It
uses the maximum utility forgetfulness [13] which replaces
the case with the lowest utility with the new case. For our
simulation experiment, a maximum of 100 cases can be
stored in the database.

Figure 7 - Case-base application to optimization process,
adapted from [13]
Simulation and Experiment of Cognitive Engine
To verify the Cognitive Engines learning and adapting
ability, Thomas W. Rondeau carried out a sequence of
simulation and over the air experiment [13] which generated
new parameter settings subject to different QoS, here
different objectives. Both simulation and experiment were
performed on the Cognitive Engine with selected function
components of GNU Radio SDR platform with USRP RF
front-ends, power spectral density (PSD) sensor and WSGA
optimizer. During the simulation [13], the Cognitive Engine
collected a set of meters such as transmit power and symbol
rate. It also specified a set of objectives such as BER and
SINR. The optimizer, i.e., WSGA, combined with the
decision maker, was able to find an optimal parameter
setting within 400 generations and tested the performance
subject to the objectives.
For the experiment [13], Thomas W. Rondeau set up a
digital communication link between two Cognitive Engine
nodes with one master and one slave. Given the objective
of high data throughput to the slave with low error, the
master was able to find the vacant spectrum among several
interference nodes and configured the radio to produce a
200 kbps QPSK signal with a 12 dBm transmit power
within a short time. In addition, the waveforms frequency
One XML file describes
the radio capabilities
Another describes
the waveform
Objectives are
located in a DLL

7
and power were verified by the policy engine through a
regulatory spectrum mask.
4. PUBLIC SAFETY COGNITIVE RADIO
As a system level application, we attach a specific set of
components to the Cognitive Engine and design the
prototype PSCR especially for public safety communication
[17], as shown in Figure 8. It can scan the public safety
spectrum (multiple bands and multiple waveforms, all
incompatible) and configure itself to interoperate with any
public safety waveform. The reconfiguration happens within
0.1 seconds after its signal sensor determines that a signal is
present.
The PSCR aims at providing universal interoperable
communication service of voice and data to solve the
interoperability problem. This problem is that various
incompatible public safety waveforms cannot communicate
with each other. This is a widely existing and severe
problem [18]. The PSCR is designed to sense the frequency
band of interest, detect and identify existing public safety
waveforms and networks, and then be reconfigured to talk
to any detected channel. Furthermore, it serves as a gateway
to bridge incompatible waveforms, different frequency
bands, and networks. It can also serve as a multi-mode
multi-band wireless terminal for the user.
Figure 8 PSCR System Block Diagram, adapted from
[19]
PSCR Architecture
Corresponding to the Cognitive Engine architecture, we
have selected GNU Radio as the radio platform and
developed several important functions in order for radio
reconfigurability to be achieved. We have developed a GUI
which will function as both the cognitive controller for
controlling and the user interface for displaying. The sensor
includes a spectrum sweeper, which is an energy detector
for signal detection and a signal classifier for waveform
recognition. The waveform knowledge base stores public
standard waveform parameters which are used to help the
signal classifier and also to determine radio platform
settings through the case-based waveform solution maker.
The API uses XML for data formulation and sends data and
command through Internet Protocol-based network socket
functions. The optimizer and policy verifier are currently
being implemented for possible other functions in the
PSCR. We will introduce most components in detail
together with the functions they provide in the following
section.
PSCR Working Modes
PSCR provides three working modes to achieve its design
goal:
Scan mode is used for a PSCR node to detect any active
signal in a specified spectrum range and recognize its
waveform parameters and modulations. Its function is
achieved through the signal sensor.
Talk mode is used for the PSCR to establish a link with any
detected standard public safety network for voice and data
communication. Once a network is specified by the user,
either selected from our database or detected from our
sensors, the radio platform will reconfigure itself to
establish a communication link between the specified
network and our PSCR. Currently PSCR supports the
analog push-to-talk function and digital link based on
several waveform modulations.
Gateway mode is used to bridge two incompatible public
safety networks.
The above three working modes are displayed and
controlled through our graphical user interface (GUI) which
is designed in the Java. The snapshots are shown in Figure 9
for scan mode, talk mode, and gateway mode respectively
from bottom to top.

Figure 9 PSCR GUI Display, adapted from [17]
PSCR Sensor
Currently, the PSCR uses a specially designed sensor which
includes the spectrum sweeper, the signal classifier, and the
waveform recognizer. The spectrum sweeper is essentially
an energy detector based on FFT. It uses GNU Radio and
USRP to collect signal samples, just the same as our
SSCRs channel monitor. The signal classifier is based on a

8
K-nearest-neighbor (KNN) algorithm and utilizes the fact
that the averages of the standard deviation of the complex
envelope of public safety waveforms are different at the
same SNR [20]. The SNR value for a detected signal is
computed through our energy estimation in the spectrum
sweeper. The signal classifier is trained beforehand through
typical public safety waveforms at different SNR values.
The threshold borders between different waveforms at each
SNR value is identified and will be used to classify any
detected waveforms modulation.
The waveform knowledge base, implemented in a MySQL
database system, stores the training result and the standard
public safety waveform specifications. It is used to
determine the standard parameters for public safety
waveforms after the sensor identifies the spectrum and
modulation of any detected signal.
Radio Platform Reconfiguration at PHY/MAC Layer
The PSCRs reconfigurability is achieved through the radio
platform based on GNU Radio and USRP. It is fully
reconfigurable at the PHY and MAC layers for the above
working modes. We designed a multithreaded control
system to coordinate the reconfiguration at MAC and PHY
layers, as shown in Figure 10.

Figure 10 CWT Waveform Framework Multi-
threading Control, adapted from [19]
The framework thread will receive commands from the GUI
controlling function, and it will get parameters from the
sensor or the knowledge base. All the commands and
parameters are formulated in XML for exchanging.
According to the command, the framework thread initiates
and coordinates the three working modes. Correspondingly,
it parses the XML file and extracts the right parameters for a
specific applications PHY/MAC layer configuration, as
shown in Figure 11.
The flow graph thread creates waveforms and their
modulations, demodulations, filtering, coding, decoding,
etc. Its name comes from GNU Radios concept of flow
graph. The flow graph thread coordinates with the MAC
thread to transmit and receive voice and data. For the data
link example, the flow graph thread receives the necessary
parameters and initiates a new signal flow graph
consisting of filter, gain (control), demod(ulation), FEC
(forward error correction), etc. It is coupled with a new
MAC sensing function, as shown in Figure 12. Most
individual signal processing blocks are available in GNU
Radio and the radio RF front end is based on USRP. Based
on this data link, we are able to build a wireless IP network
through GNU Radios virtual network interface: the
TUN/TAP open source library [21]. The application layer of
an existing Internet Protocol-based network seamlessly
works with the above wireless network through interfacing
with the TUN/TAP function block. We tested text
messaging, web browsing, voice over IP, and audio
streaming.
Figure 11 PSCR PHY/MAC Waveform Framework,
adapted from [19]

Figure 12 PSCR PHY/MAC Waveform Framework,
adapted from [19]
The current Radio Platform can support legacy FM,
narrowband BPSK, QPSK, 8PSD, and GMSK digital
modulations at multiple public safety channels. It can
support analog FM talk in talking mode and the gateway
mode. For example, we bridged the Cobra MicroTalk PR
100-2VP (22 Channels) 2-Way Radio and a Motorola
Radius P110 police hand-held radio.
5. CONCLUSIONS
In this paper, we introduce our cognitive radio development
from spectrum sharing to cognitive radio learning and

9
adapting. We design an open cognitive radio architecture
and implement several functions with simulation results
shown. Furthermore, we develop the PSCR and demonstrate
our cognitive architectures feasibility.
However, we still need to add more functions to accomplish
the above mentioned functions. We also need to add
broadband waveforms such as OFDM.
Furthermore, the full cognitive radio system is a significant
software undertaking, which includes not only a software
defined radio, but also a plethora of modules responsible for
cognition, reasoning, decision-making, etc. The success of
such a complex system depends on the reliability and
quality of the embedded software, the full re-configurability
of the radio hardware, and the high computation speed of
the computing component. In particular, several critical
cognitive radio modules require advanced computation
performance beyond the abilities of todays best General
Purpose Processor (GPP). We are working to port the
whole system into the PlayStation 3 powered by Cell BE.
6. ACKNOWLEDGEMENTS
This work was supported by the National Science
Foundation under grant CNS-0519959 and by the National
Institute of Justice, Office of Justice Programs, US
Department of Justice under Award No. 2005-IJ-CK-K017.
The opinions, findings, and conclusions or
recommendations expressed are those of the authors and do
not necessarily reflect the views of the National Science
Foundation or the Department of Justice.
REFERENCES
[1]"Shared Spectrum Company,"
www.sharedspectrum.com.
[2] J. Mitola, "Cognitive Radio: An Integrated Agent
Architecture for Software Defined Radio," Royal Institute of
Technology (KTH), 2000.
[3] J. Mitola, Software Radio Architecture: Object Oriented
Approaches to Wireless Systems Engineering: John Wiley
and Sons, 2000.
[4] K. N. Steadman, A. D. Rose, and T. T. Nguyen,
"Dynamic Spectrum Sharing Detectors," in New Frontiers
in Dynamic Spectrum Access Networks (DySPAN), Dublin,
Ireland, 2007. .
[5] J. H. Reed, Software Radio: A Modern Approach to
Radio Engineering, Englewood Cliffs, NJ: Prentice-Hall,
2004.
[6] Ettus Research, LLC., The USRP,"
http://www.ettus.com/.
[7] E. Blossom, "Exploring GNU Radio,"
http://www.gnu.org/software/gnuradio/doc/exploring-
gnuradio.html, November 2004.
[8] F. Ge and C. W. Bostian, "A Wideband Spectrum
Sensing Approach with Agility and Low SNR Sensitivity,"
in Software Defined Radio Technical Conference, Denver,
Colorado, 2007.
[9] D. Cabric, S. M. Mishra, and R. W. Brodersen, "
Implementation issues in spectrum sensing for cognitive
radios," in Asilomar Conference on Signals, Systems, and
Computers, 2004, pp. 772 - 776.
[10] R. S. Roberts, W. A. Brown, and J. H. H. Loomis,
"Computationally efficient algorithms for cyclic spectral
analysis," in IEEE Signal Processing Magazine. vol. 8,
1991, pp. 38-49.
[11] N. J. Carter, "Implementation of Cyclic Spectral
Analysis Methods," vol. Master: Naval Postgraduate
School, Monterey, CA, 1992.
[12] K. Kim, I. A. Akbar, K. K. Bae, J. Um, C. M. Spooner,
and J. H. Reed, "Cyclostationary Approaches to Signal
Detection and Classification in Cognitive Radio," in New
Frontiers in Dynamic Spectrum Access Networks
(DySPAN), Dublin, Ireland, 2007, pp. 212-215.
[13] T. W. Rondeau, "Application of Artificial Intelligence
to Wireless Communications." vol. PhD: Blacksburg:
Virginia Polytechnic Institute and State University, 2007, p.
198.
[14] T. W. Rondeau, C. W. Bostian, D. Maldonado, A.
Ferguson, S. Ball, and B. Le, "Cognitive radios in public
safety and spectrum management," in Telecommunications
Policy and Research Conference,. vol. 33, Sept. 2005.
[15] J. G. Proakis, Digital Communications, 4 ed.: New
York: McGraw Hill, 2000.
[16] R. Purshouse and P. Fleming, "Evolutionary many-
objective optimisation: an exploratory analysis," in IEEE
Congress on Evolutionary Computing,, 2003, pp. 2066 -
2073.
[17] B. Le, P. Carcia, Q. Chen, B. Li, F. Ge, M. Y. EI-
Nainay, T. W. Rondeau, and C. W. Bostian, "A Public
Safety Cognitive Radio Node System," in Software Defined
Radio Technical Conference Denver, Colorado, 2007.
[18] J. Brito, "Sending Out an S.O.S.: Public Safety
Communications Interoperability as a Collective Action
Problem," Federal Communications Law Journal, vol. 59,
pp. 457-492, 2007.
[19] B. Le, "Building a Cognitive Radio - From
Architecture Definition to Prototype Implementation," in
Electrical and Computer Engineering. vol. Doctor of
Philosophy: Virginia Polytechnic Institute and State
University, 2007.
[20] B. Le, "Modulation Identification Using Neural
Network for Cognitive Radios," in Software Defined Radio
Forum Technical Conference, Anaheim, CA, 2005.
[21] "Virtual Point-to-Point(TUN) and Ethernet(TAP)
Devices," in http://vtun.sourceforge.net/tun/.
[22] Improving Spectrum Usage through Cognitive Radio
Technology, IEEE USA Position, Nov 13, 2003, Available
online:
http://www.ieeeusa.org/policy/positions/cognitiveradio.asp.
[23] Federal Communications Commission, "NOTICE OF
PROPOSED RULE MAKING,"
http://www.fcc.gov/Bureaus/Engineering_Technology/Notic
es/2000/fcc00430.doc, December 8, 2000.

10
BIOGRAPHY

Fe Ge is currently a PhD student in
the Department of Electrical and
Computer Engineering at Virginia
Tech, Blacksburg, VA. His PhD
research is primarily focusing on
cognitive radio development in the
Cell processor. He received his
masters degree in Computer
Science and Engineering from the
University of South Carolina in
2006. From 2004 to 2006, he worked as a research assistant
in computer vision and image processing during his
masters degree pursuit. He received his Bachelors degree
in Engineering Mechanics from Tsinghua University,
Beijing, in 2002.
Qinqin Chen is currently a PhD
student in the Department of
Electrical and Computer
Engineering at Virginia Tech,
Blacksburg, VA. Her research
interests primarily include digital
gateway design for cognitive radio
and reconfigurable SDR platform.
She received her Masters degree
in Communication and
Information Systems from University of Science &
Technology of China in 2005 and Bachelors degree in
Electrical Information Engineering from Wuhan University,
China in 2002. Before transferring to Virginia Tech in
2006, she spent almost one year on the research of VLSI
Design for Soft-decision Decoding of Reed-Solomon Codes,
as a first-year Ph.D. Student majoring in Computer
Engineering at Oregon State University.
Ying Wang is currently a PhD
student in the Department of
Electrical and Computer
Engineering at Virginia Tech,
Blacksburg, VA. Her research
interests primarily include
wideband cognitive radio and
cognitive OFDM development. She
received her Masters degree from
the University of Cincinnati in
2006 working in Wireless Communication Lab focusing on
wideband and fast fading channel communication research.
She received her bachelors degree in Information
Engineering from Beijing University of Posts and
Telecommunications, China in 2003.
Thomas W. Rondeau is a research engineer with the Centre
for Telecommunications Value-chair Research (CTVR) at
Trinity College, Dublin, Republic of Ireland. He received
his Ph.D. from the Bradley
Department of Electrical and
Computer Engineering at Virginia
Tech in September of 2007 under
Dr. Charles Bostian. His current
research interests are in improving
communications
through software and cognitive
radio. Tom received his B.S.
degree in Electrical Engineering
from Virginia Tech in May 2003, graduating Summa Cum
Laude, and he received his M.S. degree in Electrical
Engineering from Virginia Tech in 2005. Tom is also a
developer in the open source GNU Radio project.
Bin Le received both his M.S. and
Ph.D. degrees in Electrical
Engineering at Virginia Polytechnic
and State University, where he
served as a research assistant in the
Center for Wireless
Telecommunications (CWT). After
graduation, he joined Cognitive
Radio Technologies (CRT), LLC as
a system engineer. He also serves as
a research associate for Wireless @ Virginia Tech at the
same time. His research interests include cognitive radio
and network, software defined radio and machine learning
algorithms. He is a member of IEEE Communications
Society, and a member of cognitive radio working group of
Software Defined Radio Forum (SDRF).
Charles W. Bostian received the
Ph.D., M.S., and B.S. (with highest
honors) in Electrical Engineering, all
from North Carolina State University,
in 1967, 1964, and 1963, respectively.
Dr. Bostian joined the Virginia Tech
faculty in 1969 following a short
period of employment with Corning
Glassworks Electronics Research
Laboratory and service as an officer
in the U.S. Army. From 1972 through 1988, he headed
Virginia Techs Satellite Communication Group. Since 1993
he has directed the Universitys Center for Wireless
Telecommunications (CWT). Bostian is the co-author of two
widely used textbooks, Solid State Radio Engineering and
Satellite Communications, now in its second edition. His
current teaching interests are in RF design and in
undergraduate circuit analysis. Bostian's primary research
interests are in cognitive electronics and radio system
design. Currently he directs National Science Foundation
(NSF), National Institute of Justice (NIJ) and Defense
Advanced Research Projects Agency (DARPA) projects on
cognitive radio. He holds a Virginia Tech Alumni
Distinguished Professorship and is a Fellow of the IEEE.