Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
13/03/2012
1
Cognitive Radios: A Network Perspective
Luiz DaSilva
Stokes Professor in TelecommunicationsTrinity College, DublinIreland
University of York12.03.2012
About me...
13/03/2012
2
CTVR
• CTVR is a national centre for telecommunications research.
• It is head‐quartered in Trinity College and spans six institutions around Ireland
• Our mission is to carry out industry‐informed research in telecommunications to the highest of standards
• We work both in the opticaland wireless domains
Spectrum crunch• Smartphones use 24x more
data than regular phones• Tablets use 122x more data
than smartphones
How to deal with this?• More efficient utilization of
spectrum• New wireless network
architectures (small cells, seamless integration of different wireless technologies)
13/03/2012
3
‘traditional’ radio
software-defined radio
cognitive radio
+ reconfigurability =
+ awareness + adaptation (+ learning) =
evolvable:Elegant paths to the future. Adaptive and flexible where it should be. Understanding of points of inflexibility.
submissive:Not biased towards any one model of ownership. No unnecessary rule‐setting or resource labelling.
sustainable:Persistent awareness of resource constraints e.g. man‐power, energy, bandwidth, processing power, space, storage etc.
13/03/2012
4
interests
transient ownership of resources
ability to learn
distributed and autonomous
decision making
fundamental principles that will allow the wireless network of the future to evolve into new architectures characterized by increasing autonomy and ubiquity of wireless services
Sustainable, evolvable, submissive: cognitive networks
Cognitive networks – perceive their environment, then decide, learn and act from the results, with network‐wide goals
Cognitiveelement
element goal
cognitive element
element goal
end-to-end goalend-to-end goalend-to-end goal
cognitive element
element goal
cognitiveelement
element goal(s)cognitiveelement
element goal
cognitiveelement
element goal
CSL
RequirementsLayer
CognitiveProcess
SoftwareAdaptableNetworkconfigurable
elementconfigurable
element
configurable element
networkstatussensor
networkstatussensor
cognitive element
element goal(s)
elementstatus
transfer
SAN API
R. W. Thomas, D. H. Friend, L. A. DaSilva, and A. B. MacKenzie, “Cognitive Networks: Adaptation and Learning to Achieve End‐to‐end Performance Objectives,” IEEE Communications Magazine, Dec. 2006
Organization of wireless access networks, including spectrum management, and distributed adaptations by cognitive radios to meet end‐to‐end objectives
Interaction between wireless access and wired networks, including assignment of users to access points, energy efficiency, flexibility to user mobility
13/03/2012
5
This talk
Channel Selection by Autonomous, Frequency‐Agile Radios
Coalition Formation by Wireless Providers and UEs
The Role of Learning in Dynamic Spectrum Access
Incorporating Dynamic Spectrum Access into LTE+ Networks and Beyond
Channel Selection by Autonomous, Frequency‐Agile Radios
The problem – M autonomous transmitter/receiver pairs search for a channel to be used opportunistically, among N possible channels, when the primary user is not active
• For each pair of radios, throughput is affected by its own selection of channels as well as by other radios’ selections and by primary user activity
• There is the potential to learn from past choices and observations
Z. Khan, J. J. Lehtomaki, L. A. DaSilva, and M. Latva‐aho, “Autonomous Sensing Order Selection Strategies Exploiting Channel Access Information,” IEEE Trans. on Mobile Computing (in press, 2012)
13/03/2012
6
Channel Selection by Autonomous, Frequency‐Agile Radios
Objective – A mechanism to enable CRs to autonomously arrive at collision‐free sensing orders• Problem is complicated by the
presence of the PU and by the possibility of false alarms
Channel Selection by Autonomous, Frequency‐Agile Radios
Proposed approach (somewhat over‐simplified)
• Sensing orders selected from a Latin square• Initially, each CR selects a sensing order with equal probability• Whenever successful, or if it finds all channels busy, the CR
selects the same sensing order in the next slot• In the case of a collision, the CR multiplicatively decreases the
probability of picking the same sensing order, by a factor γ, with all other sensing orders equally likely
All permutations of 4 channels
A Latin square
13/03/2012
7
Channel Selection by Autonomous, Frequency‐Agile Radios
N = 10P[PU present] = 0.3TTC = time to arrive at collision‐free sensing orders
P[false alarm] = 0.2
M = 10
Channel Selection by Autonomous, Frequency‐Agile Radios
Convergence:
[Thm] When N = M, and 0 ≤ P[false alarm] < 0.5, for any 0 < γ <1 the network converges to collision‐free sensing orders
[Note] When N > M, convergence to collision‐free sensing orders is even faster (N = M is the ‘worst case’ for convergence)
[Proposition] When P[PU present] = 0, the expected number of slots until collision‐free sensing orders are obtained is O(N)
13/03/2012
8
Channel Selection by Autonomous, Frequency‐Agile Radios
Analytical results:
• An (ugly) analytical expression for M=2
• A bound for M > 2
Channel Selection by Autonomous, Frequency‐Agile Radios
Comparison:
RPS: random sensing order selection from all permutations of channels
LS: random sensing order selection from a Latin square
rand‐AP/LS: randomize after collision
MxP[N,M,θ]: average # of successful transmissions in a time slot
13/03/2012
9
Channel Selection by Autonomous, Frequency‐Agile Radios
Comparison:
RPS: random sensing order selection from all permutations of channels
LS: random sensing order selection from a Latin square
rand‐AP/LS: randomize after collision
MxP[N,M,θ]: average # of successful transmissions in a time slot
The Role of Learning in Dynamic Spectrum Access
• Interested in determining under what circumstances learning is beneficial in DSA problems
• Machine learning algorithms rely on identifying patterns – but is there enough pattern in the observed wireless environment to learn from?
• And what metric can we use to quantify the presence of patterns?
• Lempel‐Ziv complexity – approaches the entropy rate of the system
I. Macaluso, T. Forde, L. A. DaSilva, and L. E. Doyle, ‘Recognition and Informed Exploitation of Grey Spectrum Opportunities,’ IEEE Vehicular Technology Magazine (in press, 2012).
13/03/2012
10
The Role of Learning in Dynamic Spectrum Access
Lempel‐Ziv complexity – synthetic data
• Single‐agent learning
• A secondary user looking for an available channel
• Primary user activity modelled as a two‐state Markov chain, with independent channels
• Q learning (full observability)
• For the secondary user, changing channels incurs a cost
Probability of successfully selecting a channel, as a function of LZ complexity and probability that a free channel exists
How much better does Q‐learning perform, as compared to random channel selection?
The Role of Learning in Dynamic Spectrum Access
Lempel‐Ziv complexity – real data
• Spectrum occupancy measurements conducted by RWTH Aachen and by us at Trinity College
• ISM band (2.4 GHz) shownProbability of successfully selecting a channel, as a function of LZ complexity and probability that a free channel exists
How much better does Q‐learning perform, as compared to random channel selection?
13/03/2012
11
The Role of Learning in Dynamic Spectrum Access
GSM1800
DECT
The Role of Learning in Dynamic Spectrum AccessMachine learning + Game Theory
• Game theory is multi‐agent decision theory• Can concepts from the two fields be brought together to help understand adaptations in a cognitive network?• Initial application: autonomous channel selection• Each secondary user is modeled as a learning automaton (linear reward‐inaction scheme)• We can prove convergence to a Nash equilibrium (when reward parameter b is small enough)
Interference graph – colours represent channels that can be selected by a secondary user for exclusive use or (blue) for shared use
Probability of convergence, averaged over all possible 7‐node interference graphs
I. Macaluso and L. DaSilva, “On the Convergence of Learning Automata to a Nash Equilibrium in Dynamic Channel Selection,” work in progress, 2012.
13/03/2012
12
The Role of Learning in Dynamic Spectrum Access
• Effect of the structure of the graph on the ability to learn
Shown as a function of the number of links in the interference graph
Coalition Formation by Wireless Providers and UEs• Distributed spectrum sharing for
multi‐hop topologies and HetNets(relays, coexistence between small and large cells)
• Adaptations: channel selection, transmit power
• Goals: network‐wide spectrum efficiency, fairness, network connectivity, coverage
• Cooperative game theory, coalition formation
Z. Khan, S. Glisic, L. A. DaSilva, and J. Lehtomaki, “Modeling the Dynamics of Coalition Formation Games for Cooperative Spectrum Sharing in an Interference Channel,” IEEE Trans. on Computational Intelligence and AI in Games, 2011
Types of coalition in equilibrium as a function of link range
13/03/2012
13
Coalition Formation by Wireless Providers and UEs
• Dynamic pricing – providers form coalitions to charge UEs for access to the spectrum
• UEs set their transmit power to maximize their payoff
• Hierarchical game: looking for the Stackelberg equilibrium
• Proposed a distributed mechanism to set prices/transmit power, proved that it is guaranteed to converge to within an ε‐distance of the SE
Y. Xiao and L. A. Dasilva, “Dynamic Pricing Coalitional Game for Cognitive Radio Networks,” Networking 2012 Workshops, May 2012
Incorporating DSA into LTE+ Networks and Beyond
• What architectural changes are needed for LTE+ to be able to augment licensed spectrum with additional spectrum that is available opportunistically?
• Proposed an evolution of the architecture itself and described signal and control planes for DSA‐related functions (e.g., spectrum leases, coordinated sensing, etc.)
• Also considered the deployment of HetNets
SAS
cBS
SpectrumAuction
Licensed CarrierBCCHCCCHDCCHDTCHPCCH
DCCHDTCHPCCH
CONNCONNCONNCONN IDLEIDLEIDLEIDLE
CONNCONNCONNCONN
J. Deaton, R. Irwin, and L. DaSilva, “The Effects of a Dynamic Spectrum Access Overlay in LTE+ Networks,” IEEE DySPAN 2011.
13/03/2012
14
Incorporating DSA into LTE+ Networks and Beyond
J. Deaton, R. Irwin, and L. DaSilva, “Supporting Dynamic Spectrum Access in Heterogeneous LTE+ Networks,” under review, 2012.
Is ‘cognitive radio’ here to stay?
• If not the buzzword, then at least the concepts behind it
• Runtime reconfigurability in radios
• Autonomous adaptations in response to the environment (including the network)
• More flexible regimes of utilization of spectrum
• Virtual networks built on multi‐provider infrastructure and heterogeneous access technologies
13/03/2012
15
On the interwebs
About CTVR… www.ctvr.ie
On email… [email protected]
Papers… luizdasilva.wordpress.com