Quality of Information Maximization
  in Two-Hop Wireless Networks

  Sucha Supittayapornpong, Michael J. Neely

                 IEEE ICC 2012


                  May 12, 2012


            Electrical Engineering
       University of Southern California
Motivation

Rate optimization problems and algorithms have long been
studied. (Chiang, 2007) (Neely, 2006)
   However, application-layer utility, which affects directly to users, is
   not considered.




                                                                            2/19
Motivation

Rate optimization problems and algorithms have long been
studied. (Chiang, 2007) (Neely, 2006)
    However, application-layer utility, which affects directly to users, is
    not considered.
Quality of Information (QoI) is the usefulness of
information (Kang, 2010) (Johnson, 2005)
    Its value depends on how valuable the information is to users.
    The value is not necessarily proportional to a number of bits.
    Example: QoI may depend on
    - Formats (ex: video, audio, text),
    - Quality (ex: resolution, simpling rate)




                                                                             2/19
Motivation

Rate optimization problems and algorithms have long been
studied. (Chiang, 2007) (Neely, 2006)
    However, application-layer utility, which affects directly to users, is
    not considered.
Quality of Information (QoI) is the usefulness of
information (Kang, 2010) (Johnson, 2005)
    Its value depends on how valuable the information is to users.
    The value is not necessarily proportional to a number of bits.
    Example: QoI may depend on
    - Formats (ex: video, audio, text),
    - Quality (ex: resolution, simpling rate)
A static system of QoI maximization was proposed by Liu
in 2011.
    The system optimizes total quality of information obtained from an
    event.
    An integer programming was proposed.



                                                                             2/19
A System in Consideration

A system consisting of observers maximizes average QoI
obtained from random events.
- This is a more general version of a previous static system (Liu,
2011)




                                                                     3/19
A System in Consideration

A system consisting of observers maximizes average QoI
obtained from random events.
- This is a more general version of a previous static system (Liu,
2011)
Observers select a format to record an event.
    Different formats have different QoI values
    Always selecting highest quality format can overload a network.
    implication: Intelligent format selection is needed.




                                                                      3/19
A System in Consideration

A system consisting of observers maximizes average QoI
obtained from random events.
- This is a more general version of a previous static system (Liu,
2011)
Observers select a format to record an event.
    Different formats have different QoI values
    Always selecting highest quality format can overload a network.
    implication: Intelligent format selection is needed.

Data is transmitted over time-varying channels to a base
station in two modes.
    Direct transmission (3G)
    Relay to neighbors (Wi-Fi)
    in order to utilize better channels (3G) of neighbors




                                                                      3/19
A System in Consideration

A system consisting of observers maximizes average QoI
obtained from random events.
- This is a more general version of a previous static system (Liu,
2011)
Observers select a format to record an event.
    Different formats have different QoI values
    Always selecting highest quality format can overload a network.
    implication: Intelligent format selection is needed.

Data is transmitted over time-varying channels to a base
station in two modes.
    Direct transmission (3G)
    Relay to neighbors (Wi-Fi)
    in order to utilize better channels (3G) of neighbors

Maximum of two hops is allowed to reduce queuing delay.

                                                                      3/19
Contributions

A two-hop system maximizing QoI has been
modeled in such a way that
   Randomness of events and transmission rates are
   considered.
   Loops in Routing are avoided.
   The number of hops is at most 2 to reduce delay.




                                                      4/19
Contributions

A two-hop system maximizing QoI has been
modeled in such a way that
   Randomness of events and transmission rates are
   considered.
   Loops in Routing are avoided.
   The number of hops is at most 2 to reduce delay.

A novel quadratic policy has been proposed.
   The policy reduces significant number of backlogs in
   the system.
   It can also be applied to the general Lyapunov
   optimization technique.



                                                         4/19
Model: A 2-Hop Network

         N    set of nodes (observers)
         0    base station
         Hn   set of neighbors of node n




                                           5/19
Model: A 2-Hop Network

                             N     set of nodes (observers)
                             0     base station
                             Hn    set of neighbors of node n




Time is slotted, t ∈ {0, 1, 2, . . . }
An event occurs at each slot with probability θ.




                                                                5/19
Model: A 2-Hop Network

                             N     set of nodes (observers)
                             0     base station
                             Hn    set of neighbors of node n




Time is slotted, t ∈ {0, 1, 2, . . . }
An event occurs at each slot with probability θ.

 un (t)    uplink transmission rate of node n in slot t
 anm (t)   relay transmission rate from node n to node m in slot t
un (t) and anm (t) depend on time-varying channel conditions
which is fixed in slot t but can change between slots.

                                                                5/19
Model: QoI and Queues

                                   Format
                                    F      set of formats
                                     (f )
                                    rn (t) QoI, node n, format f
                                     (f )
                                    dn (t) Data size, node n, format f

                                   At node n, selecting format f yields
                                     (f )    (f )
event                              (rn (t), dn (t)) in slot t.
        format selection




                                                                          6/19
Model: QoI and Queues

                                   Format
                                    F      set of formats
                                     (f )
                                    rn (t) QoI, node n, format f
                                     (f )
                                    dn (t) Data size, node n, format f

                                   At node n, selecting format f yields
                                     (f )    (f )
event                              (rn (t), dn (t)) in slot t.
        format selection

                                   Queues at node n, at slot t:
                                    Kn (t) input queue
                                    Qn (t) uplink queue
                                              Direct transmission (3G)
                                    Jn (t)    relay queue
                                              Relay transmission (Wi-Fi)


                                                                           6/19
Model: Routing
   Queues at a Node                    One & Two Hops




event
        format selection




                                                        7/19
Model: Routing
   Queues at a Node                         One & Two Hops




event
        format selection



Kn (t + 1) = max Kn (t) − s(q) (t) − s(j) (t), 0 + dn (t)
                           n          n


 Jn (t + 1) ≤ max Jn (t) −             anm (t) + s(j) (t), 0
                                                  n
                                m∈Hn

Qn (t + 1) ≤ max Qn (t) + s(q) (t) − un (t), 0 +
                           n                                   amn (t)
                                                        m∈Hn

                                                                         7/19
Problem Formulation

Received QoI at time t

                   y0 (t) =         rn (t)
                              n∈N




                                             8/19
Problem Formulation

Received QoI at time t

                          y0 (t) =          rn (t)
                                      n∈N



Optimization problem:
                 t−1
             1
max      lim            E {y0 (τ )}
         t→∞ t
                 τ =0
 s. t. all queues Kn (t), Qn (t), Jn (t) are mean rate stable
This problem is solved by the Lyapunov optimization.
(Neely, 2010)



                                                                8/19
Lyapunov Optimization Technique
Lyapunov function         (Tassiulas, 1992)   :
                     1
           L(t)            Kn (t) + Q2 (t) + Jn (t)
                            2
                                     n
                                              2
                     2 n∈N

- All queue lengths at time t are cast to a 1-dim value.




                                                           9/19
Lyapunov Optimization Technique
Lyapunov function         (Tassiulas, 1992)   :
                     1
           L(t)            Kn (t) + Q2 (t) + Jn (t)
                            2
                                     n
                                              2
                     2 n∈N

- All queue lengths at time t are cast to a 1-dim value.

Lyapunov drift: ∆(t)           L(t + 1) − L(t)
- The drift represents the difference of queues in consecutive slots.
- Minimizing the drift lead to mean rate stability of all queues.




                                                                       9/19
Lyapunov Optimization Technique
Lyapunov function         (Tassiulas, 1992)   :
                     1
           L(t)            Kn (t) + Q2 (t) + Jn (t)
                            2
                                     n
                                              2
                     2 n∈N

- All queue lengths at time t are cast to a 1-dim value.

Lyapunov drift: ∆(t)           L(t + 1) − L(t)
- The drift represents the difference of queues in consecutive slots.
- Minimizing the drift lead to mean rate stability of all queues.

Drift-plus-penalty function with variable V                (Neely, 2010)

                        ∆(t) + V (−y0 (t))
where −y0 (t) is a penalty value at time t.
- Minimizing this function every slot will stabilize queues and
optimize the objective function.
                                                                           9/19
Lyapunov Drift Minimization
Pure Lyapunov optimization has quadratic nature of ∆(t).
                   1                             2
          min        (max[Q(t) − b(t), 0] + a(t)) − Q2 (t)
         a(t),b(t) 2


Reduce delay, Non-separable decisions (centralized algorithm)




                                                                10/19
Lyapunov Drift Minimization
Pure Lyapunov optimization has quadratic nature of ∆(t).
                    1                             2
            min       (max[Q(t) − b(t), 0] + a(t)) − Q2 (t)
          a(t),b(t) 2


Reduce delay, Non-separable decisions (centralized algorithm)

Standard Lyapunov optimization optimizes a linearized ∆(t).

     min Q(t) [a(t) − b(t)]      (T assiulas, 1992)(N eely, 2010)
    a(t),b(t)


Large delay, Separable decisions (distributed algorithm)




                                                                    10/19
Lyapunov Drift Minimization
Pure Lyapunov optimization has quadratic nature of ∆(t).
                    1                             2
            min       (max[Q(t) − b(t), 0] + a(t)) − Q2 (t)
          a(t),b(t) 2


Reduce delay, Non-separable decisions (centralized algorithm)

Standard Lyapunov optimization optimizes a linearized ∆(t).

     min Q(t) [a(t) − b(t)]      (T assiulas, 1992)(N eely, 2010)
    a(t),b(t)


Large delay, Separable decisions (distributed algorithm)

Novel Quadratic Lyapunov Optimization preserves the quadratic
nature of ∆(t).
                                    2                2
                   min [Q(t) + a(t)] + [Q(t) − b(t)]
                  a(t),b(t)


Reduce delay, Separable decisions (distributed algorithm)
                                                                    10/19
Quadratic Policy

   min
                          2                    2
                                                                                  
     K (t) − s(q) (t) + K (t) − s(j) (t) +
                  n                      n
                                                                                   
    
       n                        n                                                 
                                                                                   
                                                                                  
                                                               2                  
                                                          (q)
      [Kn (t) + dn (t)]2 + [Qn (t) − un (t)]2 + Qn (t) + sn (t) +
    
                                                                                  
                                                                                   
                                   2                           2
n∈N  Qn (t) +
    
                   m∈Hn amn (t)     + Jn (t) − m∈Hn anm (t) +                     
                                                                                   
                                                                                   
                        2                                                         
     Jn (t) + s(j) (t) − 2V rn (t)
    
                                                                                  
                                                                                   
                 n
                                                                                   

   s. t.
     s(q) (t) ∈ {0, 1, 2, . . . , s(q)(max) }, s(j) (t) ∈ {0, 1, 2, . . . , s(j)(max) } ,
      n                            n            n                            n

     fn (t) ∈ F, dn (t) = d(fn (t)) (t), rn (t) = rn n (t)) (t) , n ∈ N
                           n
                                                   (f

     a(t) ∈ Aγ(t) , u(t) ∈ Uγ(t)


                                                                                  11/19
Separable Problems (1)

Admission-Control problem:
                                                 2
            min       Kn (t) + d(fn (t)) (t)
                                n                    − 2V rn n (t)) (t)
                                                           (f
          fn (t)∈F



Uplink-Routing problem:
                                                     2                         2
          min                    Kn (t) − s(q) (t)
                                           n             + Qn (t) + s(q) (t)
                                                                     n
 (q)             (q)(max)
sn (t)∈{0,1,...,sn        }



Relay-Routing problem:
                                                     2                         2
           min                   Kn (t) − s(j) (t)
                                           n             + Jn (t) + s(j) (t)
                                                                     n
 (j)              (j)(max)
sn (t)∈{0,1,...,sn           }



                                                                                   12/19
Separable Problems (2)


Uplink-Allocation problem:

                     min              [Qn (t) − un (t)]2
                   u(t)∈Uγ(t)
                                n∈N



Relay-Allocation problem:
                                                2                              2

  min              Qn (t) −            amn (t) + Jn (t) −          anm (t)
a(t)∈Aγ(t)
             n∈N                m∈Hn                        m∈Hn




                                                                       13/19
Performance Bounds
QoI vs. V
The avg. QoI approaches optimality with O(1/V )
                     t−1
                 1                            A    (opt)
         lim inf            E {y0 (τ )} ≥ −     + y0
           t→∞ t                              V
                     τ =0




                                                           14/19
Performance Bounds
QoI vs. V
The avg. QoI approaches optimality with O(1/V )
                       t−1
                  1                              A    (opt)
          lim inf             E {y0 (τ )} ≥ −      + y0
            t→∞ t                                V
                       τ =0




Total queue backlog vs. V
The avg. queue size grows with order O(V )
                 t−1
             1
   lim sup                  E {Kn (τ ) + Qn (τ ) + Jn (τ )}
    t→∞      t   τ =0 n∈N
                                         A       V    (max)      ( )
                                     ≤       +       y0       − y0

                                                                       14/19
Simulation: a small network




                                              Quality of Information vs. V
                              8

                              7

                              6
Avg. quality of information




                              5

                              4

                              3

                              2

                              1                                                     MW y0
                                                                                        ¯
                                                                                    QD y0
                                                                                       ¯
                              00        500              1000                1500           2000
                                                          V
                                                                                                   15/19
Simulation: a small network




                                Input queue vs. V                                                           Uplink queue vs. V
                        250    QD K1
                                   ¯                                                       300              QD Q1
                                                                                                                ¯
Time-averaged backlog




                                                            Time-averaged backlog
                               MW K1¯                                                      250              MW Q1¯
                        200
                                                                                           200
                        150
                                                                                           150
                        100                                                                100
                         50                                                                 50
                          00   500    1000    1500   2000                                    00            500     1000     1500    2000
                                       V                                                                            V
                               Relay queue vs. V                                         System backlog vs. Quality of information
                                                              Time-averaged information quality
                        300                                                                       6.0
                               QD J1
                                  ¯
                                                                                                  5.8
Time-averaged backlog




                        250    MW J1
                                   ¯
                        200                                                                       5.6
                        150                                                                       5.4
                        100                                                                       5.2
                         50                                                                       5.0                               QD
                                                                                                  4.8                               MW
                          00   500    1000    1500   2000                                            0 200 400 600 800 1000120014001600
                                       V                                                                 Time-averaged total backlog       16/19
Simulation: a larger network




                               17/19
Simulation: a larger network

                       Time-averaged quality of information vs. Time
               30
               25
               20
Avg. quality


               15
               10
                                                           Time average
                5                                          Moving average
                00      20000       40000          60000     80000     100000
               30                           Time
               25
               20
Avg. quality




               15
               10
                                                           Time average
                5                                          Moving average
                00       1000        2000          3000      4000       5000
                                            Time

                                                                                18/19
Conclusion



We have formulated a more realistic QoI maximization
system.
We have proposed the novel Quadratic Lyapunov
Optimization technique.
   The technique reduces significantly numbers of backlogs.
   The technique is general for Lyapunov Optimization technique.

We have derived the distributed algorithm which
approaches optimality.




                                                                   19/19

Sucha_ICC_2012

  • 1.
    Quality of InformationMaximization in Two-Hop Wireless Networks Sucha Supittayapornpong, Michael J. Neely IEEE ICC 2012 May 12, 2012 Electrical Engineering University of Southern California
  • 2.
    Motivation Rate optimization problemsand algorithms have long been studied. (Chiang, 2007) (Neely, 2006) However, application-layer utility, which affects directly to users, is not considered. 2/19
  • 3.
    Motivation Rate optimization problemsand algorithms have long been studied. (Chiang, 2007) (Neely, 2006) However, application-layer utility, which affects directly to users, is not considered. Quality of Information (QoI) is the usefulness of information (Kang, 2010) (Johnson, 2005) Its value depends on how valuable the information is to users. The value is not necessarily proportional to a number of bits. Example: QoI may depend on - Formats (ex: video, audio, text), - Quality (ex: resolution, simpling rate) 2/19
  • 4.
    Motivation Rate optimization problemsand algorithms have long been studied. (Chiang, 2007) (Neely, 2006) However, application-layer utility, which affects directly to users, is not considered. Quality of Information (QoI) is the usefulness of information (Kang, 2010) (Johnson, 2005) Its value depends on how valuable the information is to users. The value is not necessarily proportional to a number of bits. Example: QoI may depend on - Formats (ex: video, audio, text), - Quality (ex: resolution, simpling rate) A static system of QoI maximization was proposed by Liu in 2011. The system optimizes total quality of information obtained from an event. An integer programming was proposed. 2/19
  • 5.
    A System inConsideration A system consisting of observers maximizes average QoI obtained from random events. - This is a more general version of a previous static system (Liu, 2011) 3/19
  • 6.
    A System inConsideration A system consisting of observers maximizes average QoI obtained from random events. - This is a more general version of a previous static system (Liu, 2011) Observers select a format to record an event. Different formats have different QoI values Always selecting highest quality format can overload a network. implication: Intelligent format selection is needed. 3/19
  • 7.
    A System inConsideration A system consisting of observers maximizes average QoI obtained from random events. - This is a more general version of a previous static system (Liu, 2011) Observers select a format to record an event. Different formats have different QoI values Always selecting highest quality format can overload a network. implication: Intelligent format selection is needed. Data is transmitted over time-varying channels to a base station in two modes. Direct transmission (3G) Relay to neighbors (Wi-Fi) in order to utilize better channels (3G) of neighbors 3/19
  • 8.
    A System inConsideration A system consisting of observers maximizes average QoI obtained from random events. - This is a more general version of a previous static system (Liu, 2011) Observers select a format to record an event. Different formats have different QoI values Always selecting highest quality format can overload a network. implication: Intelligent format selection is needed. Data is transmitted over time-varying channels to a base station in two modes. Direct transmission (3G) Relay to neighbors (Wi-Fi) in order to utilize better channels (3G) of neighbors Maximum of two hops is allowed to reduce queuing delay. 3/19
  • 9.
    Contributions A two-hop systemmaximizing QoI has been modeled in such a way that Randomness of events and transmission rates are considered. Loops in Routing are avoided. The number of hops is at most 2 to reduce delay. 4/19
  • 10.
    Contributions A two-hop systemmaximizing QoI has been modeled in such a way that Randomness of events and transmission rates are considered. Loops in Routing are avoided. The number of hops is at most 2 to reduce delay. A novel quadratic policy has been proposed. The policy reduces significant number of backlogs in the system. It can also be applied to the general Lyapunov optimization technique. 4/19
  • 11.
    Model: A 2-HopNetwork N set of nodes (observers) 0 base station Hn set of neighbors of node n 5/19
  • 12.
    Model: A 2-HopNetwork N set of nodes (observers) 0 base station Hn set of neighbors of node n Time is slotted, t ∈ {0, 1, 2, . . . } An event occurs at each slot with probability θ. 5/19
  • 13.
    Model: A 2-HopNetwork N set of nodes (observers) 0 base station Hn set of neighbors of node n Time is slotted, t ∈ {0, 1, 2, . . . } An event occurs at each slot with probability θ. un (t) uplink transmission rate of node n in slot t anm (t) relay transmission rate from node n to node m in slot t un (t) and anm (t) depend on time-varying channel conditions which is fixed in slot t but can change between slots. 5/19
  • 14.
    Model: QoI andQueues Format F set of formats (f ) rn (t) QoI, node n, format f (f ) dn (t) Data size, node n, format f At node n, selecting format f yields (f ) (f ) event (rn (t), dn (t)) in slot t. format selection 6/19
  • 15.
    Model: QoI andQueues Format F set of formats (f ) rn (t) QoI, node n, format f (f ) dn (t) Data size, node n, format f At node n, selecting format f yields (f ) (f ) event (rn (t), dn (t)) in slot t. format selection Queues at node n, at slot t: Kn (t) input queue Qn (t) uplink queue Direct transmission (3G) Jn (t) relay queue Relay transmission (Wi-Fi) 6/19
  • 16.
    Model: Routing Queues at a Node One & Two Hops event format selection 7/19
  • 17.
    Model: Routing Queues at a Node One & Two Hops event format selection Kn (t + 1) = max Kn (t) − s(q) (t) − s(j) (t), 0 + dn (t) n n Jn (t + 1) ≤ max Jn (t) − anm (t) + s(j) (t), 0 n m∈Hn Qn (t + 1) ≤ max Qn (t) + s(q) (t) − un (t), 0 + n amn (t) m∈Hn 7/19
  • 18.
    Problem Formulation Received QoIat time t y0 (t) = rn (t) n∈N 8/19
  • 19.
    Problem Formulation Received QoIat time t y0 (t) = rn (t) n∈N Optimization problem: t−1 1 max lim E {y0 (τ )} t→∞ t τ =0 s. t. all queues Kn (t), Qn (t), Jn (t) are mean rate stable This problem is solved by the Lyapunov optimization. (Neely, 2010) 8/19
  • 20.
    Lyapunov Optimization Technique Lyapunovfunction (Tassiulas, 1992) : 1 L(t) Kn (t) + Q2 (t) + Jn (t) 2 n 2 2 n∈N - All queue lengths at time t are cast to a 1-dim value. 9/19
  • 21.
    Lyapunov Optimization Technique Lyapunovfunction (Tassiulas, 1992) : 1 L(t) Kn (t) + Q2 (t) + Jn (t) 2 n 2 2 n∈N - All queue lengths at time t are cast to a 1-dim value. Lyapunov drift: ∆(t) L(t + 1) − L(t) - The drift represents the difference of queues in consecutive slots. - Minimizing the drift lead to mean rate stability of all queues. 9/19
  • 22.
    Lyapunov Optimization Technique Lyapunovfunction (Tassiulas, 1992) : 1 L(t) Kn (t) + Q2 (t) + Jn (t) 2 n 2 2 n∈N - All queue lengths at time t are cast to a 1-dim value. Lyapunov drift: ∆(t) L(t + 1) − L(t) - The drift represents the difference of queues in consecutive slots. - Minimizing the drift lead to mean rate stability of all queues. Drift-plus-penalty function with variable V (Neely, 2010) ∆(t) + V (−y0 (t)) where −y0 (t) is a penalty value at time t. - Minimizing this function every slot will stabilize queues and optimize the objective function. 9/19
  • 23.
    Lyapunov Drift Minimization PureLyapunov optimization has quadratic nature of ∆(t). 1 2 min (max[Q(t) − b(t), 0] + a(t)) − Q2 (t) a(t),b(t) 2 Reduce delay, Non-separable decisions (centralized algorithm) 10/19
  • 24.
    Lyapunov Drift Minimization PureLyapunov optimization has quadratic nature of ∆(t). 1 2 min (max[Q(t) − b(t), 0] + a(t)) − Q2 (t) a(t),b(t) 2 Reduce delay, Non-separable decisions (centralized algorithm) Standard Lyapunov optimization optimizes a linearized ∆(t). min Q(t) [a(t) − b(t)] (T assiulas, 1992)(N eely, 2010) a(t),b(t) Large delay, Separable decisions (distributed algorithm) 10/19
  • 25.
    Lyapunov Drift Minimization PureLyapunov optimization has quadratic nature of ∆(t). 1 2 min (max[Q(t) − b(t), 0] + a(t)) − Q2 (t) a(t),b(t) 2 Reduce delay, Non-separable decisions (centralized algorithm) Standard Lyapunov optimization optimizes a linearized ∆(t). min Q(t) [a(t) − b(t)] (T assiulas, 1992)(N eely, 2010) a(t),b(t) Large delay, Separable decisions (distributed algorithm) Novel Quadratic Lyapunov Optimization preserves the quadratic nature of ∆(t). 2 2 min [Q(t) + a(t)] + [Q(t) − b(t)] a(t),b(t) Reduce delay, Separable decisions (distributed algorithm) 10/19
  • 26.
    Quadratic Policy min 2 2    K (t) − s(q) (t) + K (t) − s(j) (t) + n n    n n      2  (q) [Kn (t) + dn (t)]2 + [Qn (t) − un (t)]2 + Qn (t) + sn (t) +     2 2 n∈N  Qn (t) +   m∈Hn amn (t) + Jn (t) − m∈Hn anm (t) +     2   Jn (t) + s(j) (t) − 2V rn (t)     n  s. t. s(q) (t) ∈ {0, 1, 2, . . . , s(q)(max) }, s(j) (t) ∈ {0, 1, 2, . . . , s(j)(max) } , n n n n fn (t) ∈ F, dn (t) = d(fn (t)) (t), rn (t) = rn n (t)) (t) , n ∈ N n (f a(t) ∈ Aγ(t) , u(t) ∈ Uγ(t) 11/19
  • 27.
    Separable Problems (1) Admission-Controlproblem: 2 min Kn (t) + d(fn (t)) (t) n − 2V rn n (t)) (t) (f fn (t)∈F Uplink-Routing problem: 2 2 min Kn (t) − s(q) (t) n + Qn (t) + s(q) (t) n (q) (q)(max) sn (t)∈{0,1,...,sn } Relay-Routing problem: 2 2 min Kn (t) − s(j) (t) n + Jn (t) + s(j) (t) n (j) (j)(max) sn (t)∈{0,1,...,sn } 12/19
  • 28.
    Separable Problems (2) Uplink-Allocationproblem: min [Qn (t) − un (t)]2 u(t)∈Uγ(t) n∈N Relay-Allocation problem: 2 2 min Qn (t) − amn (t) + Jn (t) − anm (t) a(t)∈Aγ(t) n∈N m∈Hn m∈Hn 13/19
  • 29.
    Performance Bounds QoI vs.V The avg. QoI approaches optimality with O(1/V ) t−1 1 A (opt) lim inf E {y0 (τ )} ≥ − + y0 t→∞ t V τ =0 14/19
  • 30.
    Performance Bounds QoI vs.V The avg. QoI approaches optimality with O(1/V ) t−1 1 A (opt) lim inf E {y0 (τ )} ≥ − + y0 t→∞ t V τ =0 Total queue backlog vs. V The avg. queue size grows with order O(V ) t−1 1 lim sup E {Kn (τ ) + Qn (τ ) + Jn (τ )} t→∞ t τ =0 n∈N A V (max) ( ) ≤ + y0 − y0 14/19
  • 31.
    Simulation: a smallnetwork Quality of Information vs. V 8 7 6 Avg. quality of information 5 4 3 2 1 MW y0 ¯ QD y0 ¯ 00 500 1000 1500 2000 V 15/19
  • 32.
    Simulation: a smallnetwork Input queue vs. V Uplink queue vs. V 250 QD K1 ¯ 300 QD Q1 ¯ Time-averaged backlog Time-averaged backlog MW K1¯ 250 MW Q1¯ 200 200 150 150 100 100 50 50 00 500 1000 1500 2000 00 500 1000 1500 2000 V V Relay queue vs. V System backlog vs. Quality of information Time-averaged information quality 300 6.0 QD J1 ¯ 5.8 Time-averaged backlog 250 MW J1 ¯ 200 5.6 150 5.4 100 5.2 50 5.0 QD 4.8 MW 00 500 1000 1500 2000 0 200 400 600 800 1000120014001600 V Time-averaged total backlog 16/19
  • 33.
    Simulation: a largernetwork 17/19
  • 34.
    Simulation: a largernetwork Time-averaged quality of information vs. Time 30 25 20 Avg. quality 15 10 Time average 5 Moving average 00 20000 40000 60000 80000 100000 30 Time 25 20 Avg. quality 15 10 Time average 5 Moving average 00 1000 2000 3000 4000 5000 Time 18/19
  • 35.
    Conclusion We have formulateda more realistic QoI maximization system. We have proposed the novel Quadratic Lyapunov Optimization technique. The technique reduces significantly numbers of backlogs. The technique is general for Lyapunov Optimization technique. We have derived the distributed algorithm which approaches optimality. 19/19