Chapter 2 System Model
2.5 Source Model
The conversational class traffic is modeled as the ON-OFF model [22] shown in Fig. 2.3. During ON period, voice packets are generated with rate Dv bps. During OFF period, there is no packet generated. This model has a transition rate with value y in the ON state and a transition rate with value z in the OFF state.
Fig. 2.4 depicts the source model of streaming class, which is composed of a sequence of video frames generated regularly with a constant interval Tf [19]. Each video frame consists of a fixed number of slices Ns, where each slice corresponds to a single packet. The size of packet is denoted by Ps, and the inter-arrival time between each packet is Tp.
Fig. 2.5 shows the source model of HTTP interactive class. The interactive class traffic can be modeled as a sequence of packet calls (pages), and each packet call consists of a sequence of packet arrivals, which is composed of a main object and several embedded objects [19]. Four parameters, including the inter-arrival time Treading (reading time), main object size Sm, embedded object size Se, the number of embedded objects per packet call Ne, and the packet inter-arrival time Tp are used in this model.
The background class traffic is modeled as a sequence of file downloads [19] and is shown in Fig. 2.6. Denote the size of each file by Sf, and the inter-arrival time between each file by Tf.
File 1
Sf
File 2
Sf
Tf
File k
Sf
Fig. 2.6 : FTP source model
…
Chapter 3
Game Theory
Game theory can study strategic interactions between agents, where the agent is an actor or player in a model that solves an optimization problem [9]. In strategic games, agents choose strategies which will maximize their payoff, given the strategies of the other agents. There are cooperative and non-cooperative games. In a cooperative game, the players choose their strategy jointly. In the non-cooperative game, each player selects his strategy individually, without any joint-action agreements among players. If a configuration of strategies (one for each player) such that each player’s strategy is best for him, given those of the other players, the set of strategies is called be in Nash equilibrium.
A game consists of a set of players, a set of strategies available to these players, and a specification of payoffs for each combination of strategies. From the definition of the payoff function, there are zero-sum games and non-zero-sum games [9]. In zero-sum games the total payoffs to all players always adds to zero. Poker is an example of zero-sum game, because one wins and the other loses. On the contrary, in non-zero-sum games, the summation of payoffs for all players may be larger or less than zero according to different combination of strategies. This means a set of strategies may cause the result of that all win or even all lose!
As shown at table 3.1, the typical example, which is called prisoner’s dilemma,
is used to explain what the difference between cooperative and non-cooperative game is. For non-cooperative game, prisoner A and B all do not know whether the other side will choose ‘stays silent’ or ‘betrays’. At this situation, they will find ‘betrays’ is the best strategy no matter what the other side’s strategy is. If these two guys are smart, they all need serve 5 years finally. For cooperative game, however, they can know the other side’s strategy. At this situation, they both will choose ‘stays silent’ to get a compromise. That is they just need to serve 6 months respectively.
Table 3.1 : Prisoner’s dilemma
Prisoner B Stays Silent Prisoner B Betrays Prisoner A Stays Silent Each serves 6 months Prisoner A: 10 years
Prisoner B: goes free Prisoner A Betrays Prisoner A: goes free
Prisoner B: 10 years
Each serves 5 years
For the network access selection problem in heterogeneous networks, a cooperative and non-zero-sum game is defined. Based on the defined game, the goal is to find a set of strategies which satisfy the Nash equilibrium to help the access selection scheme.
Chapter 4
Utility Function and Game-Theory Based Network Selection Scheme
When a new call or a handoff call arrives, the system must determine the candidate networks first. The candidate networks selection will find which RANs are usable for the call request by checking some thresholds. Note that those thresholds are just used to check the ‘available’ networks for the call request. After getting the set of candidate networks, whose number is n, the proposed scheme will find the values of NUi and NPi , for i =1,2,…,n. Note that NUi and NPi are gained from utility function for QoS satisfaction and cooperative game for network preference, respectively. For each candidate network, Utility function for QoS satisfaction will compute the utility value from the satisfaction of QoS requirement of the call request for each candidate network. Then, for each candidate networks, the cooperative game for network preference will compute the preference value from the network point of view. The main goal of this part is to decrease the number of handoff and achieve load balance for high system utilization. Finally, by chosen the maximum linear combination of utility values and preference values, the most suitable RAN for the call request can be obtained.
4.1 Candidate Networks Selection
Two constraints are proposed to select the candidate networks: the signal strength constraint and network loading constraint. An access network must fulfill these constraints and then becomes the candidate network for a call request.
4.1.1 Signal Strength Constrain
Define the pilot signal strength from access network i received at the MS as PWi. If the value of PWi exceeds a given power threshold PWth, that is
, (4.1) PWi≥PWth
then the network i will be classified as a candidate network. Otherwise, the network will be neglected from the candidate networks. Notice that the predefined signal strength threshold may be different for different access networks.
4.1.2 Network Loading Constraint
The constraint is used to guarantee that the admittance of a call request will not affect the quality of the ongoing connections. Assume that a call request is required to report its traffic characteristic parameters when it asks to access the network. The traffic characteristic parameters of a call include peak rate, utilization (fraction of time the source is active), and mean peak rate duration of the packets. Then the equivalent capacity for the call request, say a, denoted by Ca, can be obtained [23]. If infinite buffer size is assumed, the equivalent capacity can be derived to be equal to mean rate.
This thesis will use the mean rate of a new call request as its equivalent capacity to do the network loading increment/decrement estimation when the call enters/leaves the network.
Define the current existing network loading intensity before accepting the new
call is ( ), and the new loading intensity increment for the call request a is
Η 0E ≤ Η ≤E 1
ηa. Then the loading intensity of a candidate network after accepting the call request must be under a predefined threshold loading ηth. That is
E ηa ηth.
Η + ≤ (4.2)
On the contrary, this network will not be considered as the candidate network.
In the WCDMA network, the loading intensity increment for a call request a, can be estimated as [15]
0
where f is the factor representing for interference from other cells and is defined as the ratio of inter-cell interference to the total interference in the referenced cell, W is the chip rate of the WCDMA system, and is the required bit-energy to noise-density figure corresponding to the desired link quality of the call request a.
Clearly,
The IEEE 802.16 WMAN uses the OFDMA technique. From chapter 2, the mean capacity of WMAN can be estimated as 4 K L q T× × × / (bps). Then the loading intensity increment for a call request a can be estimated as
/(4ηa =Ca × × ×K L q T/ ), (4.4)
In WLAN network, the measurement-based network load intensity estimation is used. Assume Ts is the total busy occupation transmission time, including successful transmission time and collision time, in the latest observation duration Td. Define the loading intensity as . Only when following equation is satisfied, then this network will be still in the set of candidate networks.
E s/
H =T Td
, ,
E ηth WLAN
Η ≤ (4.5)
where ηth WLAN, is predefined load intensity threshold for WLAN.
4.2 Utility Function for QoS Satisfaction
A utility function Ui for each candidate network i is defined to represent the preference of call request. It is a product of three QoS-related evaluation functions, which is given by
, , ,
i B i D i R i
U = f × f ×f , (4.6)
where fB i, , fD i, , and fR i, are the evaluation functions of data rate, packet delay, and packet dropping rate for access network i, respectively. Note that these three evaluation functions of QoS provisioning are designed from the preference of users.
For each RAN, if QoS measures that the network can provide is better than the QoS requirements of the call request, then the evaluation functions will get higher evaluation value, denoting more preference of users. As shown in Fig. 4.1, if the QoS measure value for the network can satisfy the QoS requirement of call request (in the QoS satisfaction region), then the evaluation value will increase gently (linearly). On the contrary, if the QoS measure value is in the QoS violation region, the evaluation value will decrease sharply (exponentially).
Therefore, fB i, is defined as
,
where Bi is the measured allowed data rate in access network i, Breq is the data rate requirement of the call request, and Bth is a threshold used to represent whether the QoS is highly satisfied. Note that the measured allowed data rate in WCDMA can be obtained by (2.1), and the achievable modulation order in WMAN can be estimated by (2.2), (2.3). Finally, the measured allowed data rate in WLAN is gotten by measurement-based network loading intensity estimation.
Moreover, fD i, is defined as maximum delay tolerance of the call request, and Dth is a threshold used to represent whether the QoS is highly satisfied. Note that the values of average packet delay for different traffic classes are computed separately. Different traffic class has different average packet delay in the same access network.
Similarly, the evaluation functions of packet dropping rate are formulated by
,
where Ri is the measured average packet dropping rate in access network i, Rreq is the maximum allowable dropping rate of the call request, and Rth is a threshold used to represent whether the QoS is highly satisfied. Similar to above case, the values of average packet dropping rate for different traffic classes are computed separately.
Noted that only real-time traffic classes have delay bound, so the non-real time traffic classes need not take fD i, and fR i, into consideration. After getting Ui for each candidate network I, the normalized utility value for each candidate access network i, denoted by NUi, is defined as
where n is the number of candidate network. Clearly,
1
4.3 Cooperative Game for Network Preference
Consider the load balance among the RANs. It is more preferable to choose the network with low load to serve the call request. The load balance can help to achieve the goal of maximizing the total system utilization. Moreover, consider to decrease the number of handoff, including horizontal and vertical handoff, it is more preferable to decrease the number of handoff to reduce the forced termination probability of call request and signaling overhand for handoffs.
Game theory is here adopted to solve the problem. A network preference cooperative game is defined as follows
Players: The players of this game are the candidate access networks. Assume there are n players : {N1, N2 , …,Nn}.
Strategies: n strategies : {NP1, NP2 , …,NPn}. NPi is the preference value for Ni from network provider view point. Note that
1 Payoffs: The payoff for the total candidate networks is defined as
(4.11) accepting the call request,
,
ΗE i ,
ηi th is predefined threshold load intensity of network i , and is penalty weight of network i. The meaning of the payoff function and how to define the penalty weight are described as follows.
wi
4.3.1 Meaning of the Payoff Function
Finding the best set of strategies for each candidate network to maximize the payoff function is the main goal of the cooperative game. First, consider the load balance in the heterogeneous networks system. It is more suitable for the call request to choose the access network with low traffic load. The value represents the remaining resource (in the ratio form) available before allocating resource to network Ni for the call request. The more remaining resource of one RAN, the more likely that the call request will choose this RAN. However, assigning more resource (preference value) to a network means other networks will get less resource. If taking the suitability in each access network for a call request into consideration, the situation that lots of resources are allocated to some unsuitable networks must be avoided. The penalty weight is used to achieve this goal. When is less, this means that network Ni is more suitable to the call request.
Ai
wi wi
Generally, if the remaining available resource of one network is higher, then it
will get higher preference value (strategy). On the other hand, if the penalty weight of one network is higher, it will obtain lower preference value. That is, each network can get compromising strategy (preference value) in order to maximize the payoff function. The one with the highest preference value means that it is most suitable.
4.3.2 Determining the Penalty Weight
As known, if an access network i is more suitable for the call request, the corresponding penalty weight would be lower. Here two factors are considered:
dwell time Tdwell,i in the network i and relative position in the network i for high mobility MSs and low mobility MSs, respectively.
wi
Assume that the estimated holding time, which is gotten from statistics of call request, is Tholding. Furthermore, for high mobility MSs, Tdwell,i can be obtained from the information of radius of network coverage, velocity, position, and direction of motion of MSs in (2.7) and (2.8). Define x=Tholding/Tdwell i, . When x > 1, this means that the call request has high probability to handoff if it chooses the candidate network i. Therefore, this network is considered as an unsuitable candidate for the call request in order to decrease the handoff rate. In this situation, the penalty weight is large.
On the contrary, when x < 1, the call request has high probability to finish the transmission of data in the network i. The handoff can be avoided in this case. Then, the penalty weight would be low. Therefore, can be defined as
However, for low mobility MSs, if an MS is closer to the base station of candidate networki, the network i will be more suitable for the call request coming from the MS. The penalty weight wi will be smaller. On the other hand, if it is far
from the BS and at the edge of one cell, the ping-pong effect will happen with large probability. In this situation, the penalty weight for the candidate network will get larger. By this way, the number of handoff can be decreased and the ping-pong effect can be avoided. Define dbm is the distance between BS of network i and MS, cri is the radius of network i’s coverage, and cri,th is a predefined value. Then is defined as follow
4.3.3 Nash Equilibrium and Optimization problem
After getting , the goal is to find the set of strategy which satisfies the Nash equilibrium for the above network preference game. From the definition of Nash equilibrium, the pure strategy is in a Nash equilibrium if
wi
In fact, the above game can be formulated as an optimization problem expressed as
(4.15)
Maximize PO NP NP NP
subject NP
where POtotal is a quadratic function. For the above problem which subjects to equality and inequality constraints, the KKT condition [24] can be used to find the solution. With the KKT condition, the solution of (4.15) can be obtained efficiently.
See appendix A.
4.4 Candidate Networks Decision
Finally, an access network with the maximum compromised evaluative value is expected to obtain. This network decision issue is formulated as an optimization problem given by
(4.16)
*
Arg [ i (1 ) i],
i
i = Max αNU + −α NP
where i is the ith candidate network, α is a constant whose value is between 0 and 1, NUi is the normalized utility value of candidate network i, NPi is the normalized network preference value of candidate network i, and i* is the chosen access network for the call request.
Chapter 5
Simulation Results and Discussions
5.1 Simulation Environment
As shown in Fig. 2.1, there are 7 WCDMA cells, 7 WMAN networks, and 28 WLAN networks in the simulation environment. The system parameters in the heterogeneous network are listed in Table 5.1. The channel model and the characteristic of MSs have been introduced in chapter 2.
Table 5.1: System parameters for WCDMA, WMAN, and WLAN
Parameters WCDMA WMAN WLAN
Cell radius 1.5 Km 2 Km 0.1 Km
Frame duration (time slot duration) 10 ms 5 ms 9 us Carrier frequency 2 GHz 2.5 GHz 2.4 GHz
load intensity threshold ηth 0.75 1 0.75
Number of cells 7 7 28
Chip rate (W) 3.84M bps
Ratio of inter-cell interference to the total interference in the referenced cell (f)
0.55
Number of subchannels (K) 4
Number of data subcarriers per subchannel (q)
48
Number of slots per frame (L) 16
Capacity 2 M bps
5.2 Source Model and QoS Requirements
As described at chapter 2, there are four traffic classes considered. The source model parameters for conversational, streaming, interactive, and background traffic classes are shown in Table 5.2, 5.3, 5.4, and 5.5, respectively.
Table 5.2: Source model parameters for conversational class traffic
Component Distribution Parameters
ON time Exponential Mean=1 sec
OFF time Exponential Mean=1.35 sec Packets per second Deterministic 50
Packet size Deterministic 28 bytes Call holding time Normal Mean=90 sec,
variance=20 sec Data rate during active period 11.2 Kbps
Active rate 0.426
Mean data rate 4.77 Kbps
Table 5.3: Source model parameters for streaming class traffic
Component Distribution Parameters
Inter-arrival time between each video frame (Tf)
Deterministic 100 ms
Number of packets in each video frame (Ns)
Deterministic 8 Packet size (Ps) Truncated Pareto Min.=40 bytes, Max.=250 bytes
Mean=100 bytes, α=1.2 Inter-arrival time between
packets in a frame (Tp)
Truncated Pareto Min.=2.5 ms, Max.=12.5ms Mean=6 ms, α=1.2 Call holding time Normal Mean =120 sec, variance =30 sec Data rate during active
period
133.33 Kbps
Active rate 0.48
Mean data rate 64 Kbps
Table 5.4: Source model parameters for interactive class traffic
Component Distribution Parameters
Main object size (Sm) Truncated Lognormal
Min.=100 bytes, Max.=2 Mbytes Mean=10710 bytes, std. dev.=25032bytes Embedded object size (Se) Truncated
Lognormal
Min.=50 bytes, Max.=2 Mbytes Mean=7758 bytes, std. dev.=126168 bytes Number of embedded objects
per page (Ne)
Packet size Deterministic Chop from objects with size 1500 bytes
Packet inter-arrival time (Tp) Exponential Mean=0.13 sec Call holding time Normal Mean =120 sec, variance=30 sec
Data rate during active period 92.3 Kbps
Active rate 0.136
Mean data rate 12.55 Kbps
Table 5.5: Source model parameters for background class traffic
Component Distribution Parameters
file size (Sf) Truncated
Lognormal
Min.=50 bytes, Max.=5 Mbytes Mean=2 Mbytes, std. dev.=722 Kbytes Inter-arrival time between
each file (Tf)
Exponential Mean = 180 sec Packet size Deterministic 3000 bytes
Call holding time Normal Mean =180 sec, variance =40 sec
Data rate during active period 88.9 Kbps
Active rate 1
Mean data rate 88.9 Kbps
As mentioned, the calls with different traffic classes have different QoS requirements. The QoS requirements of each traffic class call are listed in Table 5.6.
Table 5.6: The QoS Requirements of each traffic class
Traffic class Requirement Value
Conversational (voice)
Required BER 10-3
Required Eb/No 4 dB Max. delay tolerance 40 ms Max. allowable packet dropping rate 1%
Streaming (video)
Required BER 10-4
Required Eb/No 3 dB Max. delay tolerance 100 ms Max. allowable packet dropping rate 1%
Interactive
5.3 Iterative TOPSIS Algorithm
The proposed UGT algorithm is compared with the iterative TOPSIS algorithm
The proposed UGT algorithm is compared with the iterative TOPSIS algorithm