1151
though these schemes are promising, they consider only lim- The set of actions is basically the call admission decisions
ited number of wireless network architectures (i.e. UMTS, for the corresponding mobile terminal which can be 1 for
Wireless LAN, Wireless ATM) and do not specifically con- ”accept” or 0 for ”reject”. The set of action can be defined
sider admission control policies as means to provide a unified as
scheme for maximum network utilization, minimum handoff
latency and QoS.
”Set of action → T he decision of the handof f
In this paper, a new CAC scheme, which addresses this
heterogeneity problem and exploits GAs, is proposed. The by the binary action variable, a ∈ A = 0, 1.” (2)
objective of the scheme is to achieve the maximum wireless where
network utilization, mobile terminal’s QoS requirements and
significantly reduced handoff latency. In order to fulfill this a is the binary decision variable
objective, the whole capacities of the channel, the power
A is the set of values that the decision can take (0 for
consumption, the signalling and the switching parameters
”reject”, 1 for ”accept”)
of the wireless architecture and various coverage area archi-
tectures in place are captured via this sophisticated handoff Each parameters in the cost function depends on the wire-
algorithm. These parameters are fed into a cost function less network architecture in the system. Power consumption
that is developed based on the Markov Decision Process, cost, ΨP w is fixed with coefficient such as
which is then optimized using GAs. As a result, the al-
gorithm determines which network architecture the mobile
ΨP w (N ) = ψP w = cP w (3)
terminal (MT) would complete handoff with respect to the
minimum cost value. For the performance evaluation, dif- It is assumed that the bandwidth cost rate ψBw depends
ferent simulation environments have been developed. The linearly on all the capacity of the network architecture and
performance analysis shows that the GAC scheme achieves inversely with the available capacity as follows
the minimum handoff latency and the maximum wireless
network utilization for a variety of wireless network envi- CN
ronments compared to other algorithms. ψBw (N ) = cBw · (4)
C(t)
The remainder of the paper is organized as follows: The
modelling and analysis of the GAC is presented in Section 2. where CN is the total capacity of the channel of the network
The GAC algorithm is described in Section 3. The valida- and C(t) is the available capacity of the channel at time t.
tion of the model, its performance evaluation results along t is the time of the handoff (i.e. if available capacity of the
with the effects of the GAC algorithm on the network perfor- resource is CN then the cost will be only the bandwidth co-
mance and latency are then discussed in Section 4. Finally, efficient). cBw is the bandwidth cost coefficient per capacity
the paper is concluded in Section 5. unit.
Both signalling and rerouting cost coefficients change with
respect to the next network architecture that the handoff
2. MODELLING AND ANALYSIS occurs. Hence, the network decision function can be defined
In the modelling of the problem, all the parameters of the as
wireless network architecture that affect the handoff process
such as signalling, switching, bandwidth and power con- ½
0 , Nc = Nx
sumption are considered in the cost function that have to fN (Nc , Nx ) =
1 , Nc 6= Nx
be optimized. The cost function is formulated using the
Markov Decision Process. Then a final optimality equation where Nc is the current network and Nx is the next net-
is derived for the optimization and handoff decision process work of the cell, which the mobile terminal would probably
which is implemented by genetic algorithms. roams. fN (Nc , Nx ) determines whether the next cell, which
For effective network usage, continuously changing fac- the mobile terminal roams, is at the same network or not.
tors should be considered in handoff decisions. There are
different constraints such as bandwidth, latency, network
ΨSig (N ) = csig [1 + fN (Nc , Nx )] (5)
load, power consumption and capacity of the network. Let
each state defined to determine the final action in Markov
Decision Process [11] represent a different wireless network δ
X
architecture. The collection of these states consists different ψSw (N ) = cSw [1 + fN (Nc , Nx )] (6)
wireless architectures. 1
where cSig and cSw are the signalling and switching cost
Cost f unction ⇒ (Real Cost)N = F(SigN , SwN , P wN , BwN ) (1)
coefficients respectively, and δ is the number of connections.
where The optimization problem is formulated as a Continuous
Sig is the signaling to set-up handoff Time Markov Decision Process (CTMDP). The states in the
theory would represent each different network architectures
Sw is the switching, rerouting of the traffic during handoff for the proposed solution [11]. The cost functions for the
Markov Decision Process (MDP) theory have been defined
P w is the power consumption of the network
above. The optimization objective is to find a policy π ∗ such
Bw is bandwidth of the network that
1152
The optimal decision policy can be found by solving the The state transition probability defined in equation (10),
related optimality equations for each network architecture which is related to our model can then be given by
in the system [11]. We assume that the handoff requests
arrived according to a Poisson process with rate λ and the µ
µ+λ
,j=0
request durations are exponentially distributed with rate µ. λ
pij = , j =i+1 (12)
With the assumptions of a discounted infinite-horizon Con- µ+λ
tinuous Time Markov Decision Process(CTMDP), the opti- 0 , otherwise
mality equations can be written as For the last part of the modelling, the final optimality
equation is to be derived. It is optimized using GAs to de-
( µ ¶X )
λ+µ termine the call admission control policy of the correspond-
ν(N ) = min r(N, a) + p(j|N, a)ν(j) ing MT. As all the required parts are derived as described in
a∈A λ+µ+α j∈N the previous subsections, the cost expectation, r(N, a) in (9)
(8) and the state transition probability in (12) are evaluated in
where r(N, a) is the cost expectation between two decision order to obtain the final optimality equation in Equation (8)
instants and p(j|N, a) is the state transition probability that for j = 0 and j = i + 1 respectively as follows
the system occupies (state) network j at the subsequent de-
cision instant, given that the system is in (state) network N n £ ¤
at the earlier decision instant and action a is chosen. ν(N ) = min cSig 1 + fN (Nc , Ni ) + cP w +
With the Poisson arrival assumption, the time between P £ ¤
h cBw .CN + δ cSw 1 + fN (Nc , Ni ) i
C(t) 1
any two successive events (arrival of requests or departure + ¡ ¢ +
of a request) is exponentially distributed with rate (λ + µ). α+λ+µ
Recall that between two successive events the state of the ¡ λ + µ ¢¡ µ ¢o
system does not change. Hence, r(N, a) can be expressed + f orj = 0, (13)
α+λ+µ µ+λ
by
n £ ¤
h i ν(N ) = min cSig 1 + fN (Nc , Ni ) + cP w +
ψBw (N ) + ψSw (N ) P £ ¤
r(N, a) = ΨSig (N ) + ΨP w (N ) + (9) h cBw .CN + δ cSw 1 + fN (Nc , Ni ) i
C(t) 1
(α + λ + µ) + ¡ ¢ +
α+λ+µ
Since the set of possible actions A which is defined in (2) ¡ λ + µ ¢¡ λ ¢o
is finite and r(N, a) is bounded, it can be proved that the + f orj = i + 1, (14)
optimal policy π ∗ is stationary and deterministic [11]. α+λ+µ µ+λ
The optimality equation consists of two parts. The first The values of ν ∗ (N ) can be found by using either GAs and
part r(N, a) is already derived. In order to conclude the final any heuristic approaches, value iteration or policy iteration
optimality equation, p(j|N, a), state transition probability algorithm which are numerical procedures. The numerical
also needs to be derived. analysis is demonstrated in Section 4.
Let the residency time and call duration time can be rep-
resented as tr and tc , respectively, for any wireless network
architecture. These residency time and call duration time 3. GAC : GENETIC-BASED ADMISSION CON-
are assumed to be independent and identically distributed TROL
random variables. Furthermore, tc is also assumed that it is
exponentially distributed with rate µ and probability den- 3.1 Genetic Algorithms
sity function of tr is fr (t) with mean 1/λ. Genetic Algorithms are directed random search techniques
As the handoff process is modeled as Markov chain, the used to look for parameters that provide the optimal solution
state i in the Markov chain denotes the number of handoffs to a problem. They are based on the principles of evolution
since the last connection is established [10]. When the MT and natural genetics [7]. As an optimization method, GAs
is at a particular state i, a transition to initial state 0 occurs have major differences and advantages over the other opti-
when the ongoing connection is terminated. Hence, the state mization algorithms [7]. The notion of genetic algorithms is
transition probability, pij [10] from state i to state j is given the survival of the fittest of the nature. This implies that
as the ‘fitter’ individuals are more likely to survive and have a
chance of passing their features to the next generation.
pi0 ,j=0 In our scheme, GAs is used to solve the final optimality
pij = 1 − pi0 , j =i+1 (10) function. The next section describes the system and algo-
rithm used in detail.
0 , otherwise
where, pi0 is the probability of the terminated connection 3.2 System and Algorithm for Call Admission
after ith handoff. Since the equilibrium state probability of Control
state i is expressed as pi , then Different wide area type of wireless networks (i.e. 3G-
pico, 3G-micro, 3G-macro, Satellite-LEO, Satellite-GEO)
exist in a typical urban scenario deployed by several service
pi = (1 − pi0 )i p0 (11)
P∞ providers. Besides, large number of WLANs exist in pub-
0 0
where p0 is the state probability of 0 state. As 0 pi = 1 lic and private domain. Both wireless network architectures
we have p0 = pi0 [10]. (3G-pico, 3G-micro, 3G-macro, Satellite-LEO, Satellite-GEO,
1153
NMG Network Mobility Gateway
Wireless Access Point
NG Wireless Terminal
RNC:Radio Network Controller
BSC:Base Station Controller
IP BACKBONE AR:Access Router
Satellite
NMG
WLAN
FES AR
Satellite
Cell RNC BSC AP AP
AP
2G Cell
3G Cell
False
WLAN) must be networked to each other in order to realize as cost coefficients, channel capacities, the number of rerout-
the inter-system handoff between any of the wide area wire- ing connections in the cell, call durations and handoff rates
less systems and WLAN or vice versa in the NGWS. This constitute the total cost for a specific cell. Then, the scheme
can be done in a way that, these wireless systems, hetero- determines which cell and which network architecture has
geneous architectures are connected to one another through the minimum total cost via implementing GAs. GAC incor-
a third party interconnection gateway system such as pro- porates genetic algorithms to achieve a new, fast, accurate
posed in [3, 5, 8]. Since these proposed systems have similar call admission control scheme in order to address the archi-
gateway architecture, we take Network Mobility Gateway tectural heterogeneities of the NGWS. The objective of using
(NMG) proposed in [3] as an example architectural element GAs is to find the minimum cost for the wireless network
of the NGWS in our paper. NMG gateway sits in the Inter- system to provide maximum network utilization and mini-
net as shown in Fig. 1. mum handoff latency and to fulfill QoS requirements. The
NMG is capable of managing the handoffs among these complete GAs flowchart showing the detailed operation of
different wireless network systems. The proposed GAC scheme the GAC scheme are presented in Fig. 2. After determining
takes its place in this step of the system. It is deployed in which network architecture that the mobile terminal should
NMG to manage all handoff management issues. The mobil- handoff, the GAC informs NMG which wireless network ar-
ity management of the mobile terminal in a wireless network chitecture will serve the mobile terminal after all. Accept or
is typically handled by the corresponding base transceiver reject action is then deployed after the propagation of the
station (BTS), access point (AP), radio network controller decision to NMG.
(RNC) or fixed earth station (FES) depending on the wire- The following sub-sections describe the detailed design
less system that the mobile terminal is in. For the inter- and the operations of the GA-based solution for call ad-
system mobility management, all of BTS, AP, RNC and mission control which includes encoding scheme, fitness cal-
FES are involved. These controllers propagate their sys- culation, elitism and crossover and mutation operations.
tem control and signalling messages through NMG. NMG
propagates the messages needed for the functioning of the 3.2.1 Encoding Scheme
algorithm to GAC itself. A set of parameters is sought that will give the best solu-
These messages consist of all the information of the wire- tion in optimization. In order to implement GAs, these set of
less system related to the mobility management. This infor- parameters must be encoded into a string so that crossover
mation is used by GAC to evaluate the minimum total cost and mutation operations can be applied [7].
for each available wireless system at the same area of the cor- Every wireless network architecture that is in the coverage
responding mobile terminal. All the network resources such area of the mobile terminal represents different search areas
1154
for the genetic algorithms. The encoding is not binary, for 10
GAC
Heuristic
the simplicity of the solution and to provide more accurate GAC-Worst
GAC-Best
values in a fast manner, the genes have their actual real 8
values.
The size of the solution space includes total value ranges
of the cost coefficients and the other parameters which con-
1155
10 10
Heuristic Heuristic
GAC(XOver=0.65) GAC(XOver=0.8)
GAC-Worst(XOver=0.65) GAC-Worst(XOver=0.8)
GAC-Best(XOver=0.65) GAC-Best(XOver=0.8)
8 8
CAC Latency (s)
4 4
2 2
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Handoff Events Handoff Events
(a) (b)
Figure 4: The CAC latency experienced with GAC scheme and heuristic algorithm for varying handoff
scenarios and for different crossover probability
10 10
Heuristic Heuristic
GAC(Mut=0.0005) GAC(Mut=0.002)
GAC-Worst(Mut=0.0005) GAC-Worst(Mut=0.002)
GAC-Best(Mut=0.0005) GAC-Best(Mut=0.002)
8 8
CAC Latency (s)
4 4
2 2
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Handoff Events Handoff Events
(a) (b)
Figure 5: The CAC latency experienced with GAC scheme and heuristic algorithm for varying handoff
scenarios and for different mutation probability
Table 1: Simulation Experiments Handoff Scenarios for Roaming Between Heterogeneous Wireless Architec-
tures
Scenario # (N1 , N2 , N3 , N4 ) (AvBW1 , AvBW2 , AvBW3 , AvBW4 )
1 (SatG , SatL , 3Gµ ) (75%,50%,25%)
2 (3Gp , 3Gµ , SatL , −) (50%,50%,80%,-)
3 (SatL , 3Gp , W LAN, −) (100%,50%,10%,-)
4 (3Gp , 3Gm , 3Gµ , −) (50%,50%,50%,-)
5 (W LAN, SatL , SatG , −) (10%,100%,100%,-)
6 (3Gp , 3Gµ , W LAN, −) (50%,80%,10%,-)
7 (SatG , SatL , SatL , −) (100%,50%,25%,-)
8 (3Gm , 3Gµ , SatL , W LAN ) (50%,10%,80%,5%)
9 (W LAN, W LAN, 3Gp , 3Gm ) (25%,30%,50%,75%)
10 (SatL , SatG , 3Gm , 3Gm ) (100%,50%,50%,80%)
Table 2: Cost coefficient parameters deployed for the final optimality equation
Coeff. SatL SatG 3Gp 3Gµ 3Gm WLAN
csig (1,5) (1,5) (1,5) (1,5) (1,5) (1,5)
cpw (W) (2,10) (2,10) (0.01,0.05) (0.01,0.05) (0.01,0.05) (0.04,0.25)
1
cbw ( bw ) (1.5,45) (7.5,45) (0.6,4.3) (4.3,7.5) (7.5,45) (0.01,0.5)
csw (1,5) (1,5) (1,5) (1,5) (1,5) (1,5)
bandwidth and power consumption threshold values [2, 6, 9]. the MT should do handoff.
The handoff cost also demonstrates the network resource us- As shown in Fig. 6, for the majority of the handoff sce-
age while determining which wireless network architecture narios, the total handoff cost experimented using GAC is
1156
6 100
GAC GAC
Bandwidth-Based Bandwidth-Based
Power-Based Power-Based
GAC-Worst
5 GAC-Best
80
Admission Percentage(%)
4
60
Handoff Cost
40
20
1
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Handoff Events Handoff Scenarios
Figure 6: Handoff cost experienced with GAC, Figure 9: Call admission percentage experi-
bandwidth-based and power-based schemes for vary- enced with GAC, bandwidth-based and power-based
ing handoff scenarios. schemes for varying handoff scenarios.
much less than both CAC algorithms. In one or two cases, roam among heterogeneous wireless architectures.
the results are similar with the other schemes. This happens
when there are similar wireless network systems around the 5. CONCLUSION
corresponding MT. Therefore, both algorithms conclude the
Different call admission control (CAC) schemes have been
similar handoff cost results for these cases. For other cases,
developed for different wireless systems in the literature.
GAC performs more efficiently than the other schemes re-
However, these schemes do not provide a single solution to
garding to wireless network utilization and low handoff cost.
address the heterogenous architectures which are revealed
The same experiments are also demonstrated for varying
by the Next Generation Wireless System (NGWS). A new
crossover and mutation rates. It is observed in Fig. 7 and
scheme is needed that can address these heterogeneities.
Fig. 8 that the results depending on these varying probabil-
In this paper, a new unified CAC which is based on genetic
ities. The results are similar for every handoff scenario. The
algorithms (GAs) is presented to realize the NGWS objec-
total handoff cost values are slightly improved when GAC is
tive. GAC incorporates a fast, reliable, accurate artificial
used with increased mutation or crossover rates. When the
intelligence algorithm to provide high network utilization,
mutation and crossover rates are at their highest, the exper-
minimum cost, minimum handoff latency and required QoS
imental results show a greatly improved cost. These settings
level in NGWS. A Markov Decision model is deployed to
are found to give the best performance in our experiments.
derive the necessary equations for the best network perfor-
This makes perfect sense since the GAs has more chance
mance with minimum cost. Genetic algorithms are used to
to determine the better solutions with higher crossover or
optimize this model and determine the final handoff decision
mutation probabilities. Therefore, when the higher the mu-
and accept-reject action for call admission control.
tation or crossover rates are selected, the higher performance
The experimental results also showed that GAC scheme
is observed with respect to the handoff cost. However, the
achieves very efficient resource utilization of very low hand-
higher rates cause higher algorithm delays which affect the
off latencies and costs in different heterogeneous wireless
handoff latency.
architectures. GAC is shown to significantly improve the re-
4.3 QoS Performance source utilization with very low handoff latency over heuris-
tic approaches and very low cost over different previously
Another critical requirement for NGWS is providing a cer-
proposed CAC schemes [4, 6, 14] for a particular wireless ar-
tain QoS level for the MT. This includes a successful call ad-
chitecture. Hence, instead of using CAC schemes developed
mission percentage for a large number of mobile terminals.
for specific architectures, GAC achieves low handoff latency,
These experiments are designed to illustrate how GAC pro-
low cost and certain QoS levels for all of different wireless
vides required QoS level for different types of wireless net-
architectures by adapting its algorithm. As a result, GAC
work architectures and for increasing number of MTs.
scheme addresses the challenges posed by the NGWS and
In order to achieve this, a simulation experiment is per-
significantly improves the performance for reliable, accurate
formed to evaluate the successful number of MTs admitted
and fast handoff management in NGWS.
by the wireless network by implementing GAC and several
other proposed algorithms. In this simulation experiment,
admission percentage is compared with the previously pro- 6. REFERENCES
posed algorithms for varying handoff scenarios. These sce- [1] I. Akyildiz, J. McNair, H. U. J. Ho, and W. Wang.
narios are shown in Table 1. The simulation is tested for Mobility management in next generation wireless
different wireless environments and generated three sets of systems. In Proceedings of the IEEE, volume 87, pages
data. As shown in Fig. 9, GAC maintains its high admission 1347–1384. IEEE, August 1999.
percentage regardless of which different wireless architecture [2] O. Baldo, L. Thong, and A. Aghvami. Performance of
used in the experiments. Therefore, GAC can provide cer- distributed call admission control for multimedia high
tain QoS with high network utilization while mobile users speed wireless/mobile atm networks. IEEE
1157
6 6
GAC(XOver=0.65) GAC(XOver=0.8)
GAC-Worst(XOver=0.65) GAC-Worst(XOver=0.8)
GAC-Best(XOver=0.65) GAC-Best(XOver=0.8)
Bandwidth-Based Bandwidth-Based
5 Power-Based 5 Power-Based
4 4
Handoff Cost
Handoff Cost
3 3
2 2
1 1
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Handoff Events Handoff Events
(a) (b)
Figure 7: Handoff cost experienced with GAC, bandwidth-based and power-based schemes for varying handoff
scenarios and for different crossover probability
6 6
GACMut=0.0005 GAC(Mut=0.002)
GAC-WorstMut=0.0005 GAC-Worst(Mut=0.002)
GAC-BestMut=0.0005 GAC-Best(Mut=0.002)
Bandwidth-Based Bandwidth-Based
5 Power-Based 5 Power-Based
4 4
Handoff Cost
Handoff Cost
3 3
2 2
1 1
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Handoff Events Handoff Events
(a) (b)
Figure 8: Handoff cost experienced with GAC, bandwidth-based and power-based schemes for varying handoff
scenarios and for different mutation probability
International Conf. on Comm., 3:1982–1986, June Vehic. Tech. Conf., 2:1453–1457, May 1998.
1999]. [10] J. McNair. Handoff techniques for next generation
[3] Brain architecture specifications and models, brain wireless multimedia systems. PhD Thesis, 2000.
functionality and protocol specification. [11] M.L.Puterman. Markov Decision Processes: Discrete
D2.2(IST-1999-100050), 1999. Stochastic Dynamic Programming. Wiley, New York,
[4] S. K. H. Chen and C. Kuo. Dynamic call admission 1994.
control scheme for qos priority handoff in multimedia [12] M. L. V.W.S. Wong and V. Leung. Stochastic control
cellular systems. Wireless Comm. Net. Conf., of path optimization for inter-switch handoffs in
1:114–118, March 2002. wireless atm networks. IEEE/ACM Trans.on
[5] L. Dell’Uomo and E. Scarrone. An all-ip solution for Networking, 9(3):336–350, June 2001.
qos mobility management and aaa in the 4g mobile [13] C. C. Y. Xiao and Y. Wang. A near optimal call
networks. Wireless Per. Mult. Comm., 2:591–595, admission control with genetic algorithm for
October 2002. multimedia services in wireless/mobile networks. IEEE
[6] J. Evans and D. Everitt. Effective bandwidth-based Nat. Aero. Elec. Conf., pages 787–792, October 2000.
admission control for multiservice cdma cellular [14] J. D. A. Zhu and J. Hu. Adaptive call admission
networks. IEEE Trans. Vehic. Tech., 48(1):36–46, control for multi-class cdma cellular systems. Fifth
January 1998. Asia-Pacific Conf. Comm. Fourth Optoelect. Comm.
[7] D. Goldberg. Genetic Algorithms in Search, Conf., 1:533–536, 1999.
Optimization, and Machine Learning. Number ISBN [15] A. Zomaya and M. Wright. Observations on using
0-262-13316-4(HB). Seventh printing, 2001. genetic-algorithms for channel allocation in mobile
[8] W. Kellerer and H. Vogel. A communication gateway computing. IEEE Trans. Parall. Dist. Sys.,
for infrastructure-independent 4g wireless access. 13(9):948–962, September 2002.
IEEE Comm. Mag., 40(3):126–131, March 2002. [16] M. Zonoozi and P. Dassanayake. User mobility
[9] M. P. J. Knutsson, P. Butovitsch and R. Yates. modeling and characterization of mobility patterns.
Downlink admission control strategies for cdma IEEE J. Select. Areas Commun.,, 15(7):1239–1252,
systems in a manhattan environment. IEEE 48th September 1997.
1158