Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract—We propose an autonomous and decentralized environment, and switch among different networks by
optimization method for heterogeneous wireless networks, in using handover technologies for macro-mobility. Since such
which a huge number of mobile terminals with multiple dif- an open network architecture like the Internet may include
ferent air interfaces autonomously utilize the most appropriate
infrastructure wireless networks, by sensing available wireless huge number of wireless access networks, it would be a
networks, selecting the most appropriate one, and reconfiguring large scale complex system. To optimize such a complex
themselves with seamless handover to the target networks. For network with various constraints and requirements, we
optimum selection of radio resources, we introduce dynamics need autonomous and decentralized algorithms to keep
of the mutually connected neural network which converges the entire network in an optimal state.
to an optimal state since its property is to autonomously
decrease its energy function. In this paper, we apply such In order to analyze distributed dynamical systems of
neurodynamics to user-centric optimum radio access selection, the cognitive radio networks, the game theory [6] has been
which maximizes the total throughput of the entire networks, applied and a recent hot research topic. It was used for
and minimizes the power consumption of the user terminal analyzing stability and convergence of dynamical systems
and the communication cost. We compose a neural network consisting of distributed and autonomous mobile terminal
that solves such a problem, and we show that they can be
optimized by distributed decisions based on autonomous and behaviors. Although it can deal with various rules-based
simple neuron updates on the terminal side. dynamics, game theory itself is not a tool for globally
optimizing some given objective function.
I. Introduction As more natural optimization method based on an
Various wireless systems, such as cellular phone systems autonomous and decentralized dynamics, we have applied
and wireless LANs, have been deployed. Moreover, mobile the energy function of mutually connected neural network
broadband wireless access systems such as WiMAX will dynamics for load balancing of the heterogeneous wireless
be launched in the next few years. Since these wireless networks [7], [8]. The dynamics of the mutually connected
systems have different QoSs, capacity, costs, and so on, networks is based on the always-decreasing property of
seamless handover technologies among different networks its energy function, and it converges to a stable state
have been important research topics for always using the corresponding to a minimum of the energy function [9][10].
best wireless network, and these technologies include IP This property is useful for solving various minimum search
mobility, software defined radio, cognitive radio [1][2] and problems and does not change even for large-scale neural
common signaling technologies [3][4]. By extending such networks. When we want to solve some minimum search
adaptive heterogeneous wireless networking and enabling problem, we can construct a neural network autonomously
adaptive seamless handover among any existing wireless converges to the minimum of the target objective function
access networks, the combined network becomes scalable, by mapping the objective optimization function to the
and this improves availability for the user’s service re- energy function. The energy function is defined by fixed
quirements. Not only media-independent handover among parameters, i.e., the connection weight between neurons
different wireless systems but also carrier-independent and the threshold of neurons, and variables, i.e., the states
handover among different network operators greatly in- of the neurons. Namely, it is an autonomous dynamical
creases the number of available wireless systems for each system, which is based on a simple neuronal update
mobile terminal user. We call such wireless network equation. Since it is based on decentralized updates of
environments Cognitive Wireless Clouds [5], and they neurons, we consider that such a property is suitable as
are customer-centric and scalable heterogeneous wireless part of a mechanism for deciding the optimum access point
networks. In the concept of the Cognitive Wireless Clouds, selection on the terminal side. A drawback of this approach
mobile terminals utilize any wireless systems in their is that it is not trivial, and it is sometimes difficult
2747
1525-3511/08/$25.00 ©2008 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.
to introduce various objective functions that should be energy function to minimize the tour length. To satisfy the
minimized (or maximized), since they have to be expressed constraint of the TSP, that is, visiting each city exactly
as quadratic functions of the neuronal states. once, they also introduced a constraint term in the energy
In this paper, we apply the dynamics of the mutually function. The connection weights and thresholds were
connected neural network to user-centric optimization of obtained by transforming the objective energy function
radio access selection. This selection includes not only into a form whose variables are neuronal states as in
load-balancing of the heterogeneous wireless networks but Eq. (2), and then mapping the obtained coefficients
also the reduction of power consumption on users’ ter- to a neural network energy whose coefficients are the
minals and communication costs, as a typical example of connection weights and the thresholds. The neural network
optimization problems from the users’ point of view. The can autonomously solve the given problem simply using
neural network dynamics has already been applied to the these obtained connection weights and threshold.
dynamic channel assignment problems in the field of radio One issue of this approach is that it sticks at a local
resource management [11][12], and its usefulness has been minimum. To overcome this weakness, stochastic noises,
shown in those cases. In this paper, we apply the dynamics such as Boltzmann machines [14], Gaussian Machines [15],
to optimize a heterogeneous wireless network environment. or simulated annealing [16], have been used to shake the
First, we show how to construct the neural network as neural network state out of a local minimum so it would
autonomous optimization dynamics for optimum access possibly fall into a better state. Chaotic dynamics has also
point selection in heterogeneous wireless networks, and been used to shake the state of the neural network, and
then we discuss simulation results. various theoretical studies have shown that the solving
performance of chaos is much better than stochastic noise
II. Optimization by Mutually Connected Neural [17]–[20].
Networks Another drawback of the mutual connection neural
The state of the mutual connection neural networks network approach is that the constraint of the problem
converges [9][10] when we use a typical 0-or-1 output should be merged with the energy function as a constraint
neuron whose update equation is given by, term. Therefore, when the neural network converges to a
n
suboptimal state, it becomes difficult to obtain a feasible
1 for wij xj (t) > θi , solution. In the case of easier optimization problems
xi (t + 1) = j=1 (1) with a one-dimensional constraint such as selection or
0 otherwise, classification, it is possible to satisfy the constraint by
modifying the neuron update equation to decide firing by
where xi (t) is the output of the i th neuron at time t which
taking the maximum firing neuron as follows [21],
takes only 0 or 1, wij is the connection weight between
the i th and j th neurons, θi is the threshold of the i th
1 if yi (t + 1) = max[y1 (t + 1),
neuron, and n is the number of neurons in the network. xi (t + 1) = . . . , yn (t + 1)], (3)
A mutual connection neural network composed of such
0 otherwise,
very simple neurons converges to a state corresponding to
n
a local minimum of the energy function Enn (t), which is
defined by fixed parameters, connection weights wij and where, yi (t + 1) = wij xj (t) − θi . The radio access
j=1
threshold θi , and variables, the state of the neurons xi (t).
technology selection problem at each mobile terminal can
The energy function is described as follow,
be included in the group of such selection problems.
n n n
1 Therefore, we use this type of neuron update equation,
Enn (t) = − wij xi (t)xj (t) + θi xi (t). (2) since it can satisfy the constraint without any additional
2 i=1 j=1 i=1
constraint term to the energy function.
This energy function always decreases after each neuron
III. Design of Neural Network for User-Centric Radio
update by Eq. (1) [9][10]. This property can be easily
Access Selection in Heterogeneous Wireless Networks
shown by calculating difference in Enn (t) for before and
after a neuron asynchronously updates. This decreasing A. Optimization problem of user-centric selection and
property has been applied to various optimization prob- objective function
lems to search for the minimum point in high-dimensional As an example of optimization problems in large-scale
state spaces, by using distributed and autonomous pro- heterogeneous wireless networks, we have constructed a
cessing of individual neurons. neural network that optimizes load balancing [7], [8]. The
Hopfield and Tank showed the usefulness of this prop- objective function was defined as maximization of the
erty of the energy function in solving NP-hard combina- average throughput over all the mobile terminals. In this
torial optimization problems such as Traveling Salesman paper, we additionally introduce objective functions for
Problems (TSPs) [13]. They devised a scheme of mapping minimization of power consumption and communication
travel routes into the firing patterns of neurons and an cost, from the user’s point of view. For simplicity, we
2748
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.
In order to apply the mutual connection neural network + θij xij (t). (9)
to optimization problems, we have to map its firing pattern i=1 j=1
to a state of the given problem. First, we have to express ENN (t) always decreases through autonomous and dis-
the state of the objective problem using neurons that tributed updates of neurons according to Eq. (8). There-
2749
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.
link aggregation can be taken into account. As the second objective function for F2 (t) corresponding
to the power consumption, the energy function can be
C. Obtaining the connection weights in the neural net- defined as follows, with a variable xij (t) which is the state
work of the (i, j)th neuron,
The mutual connection neural network for optimization Nm N
AP
2
can be realized by calculating the connection weights EOBJ (t) = Pj xij (t). (15)
Wijkl from the objective functions in Eqs. (5)–(7) and i=1 j=1
the energy function in Eq. (9). Since we decided to use For the third objective function for the communication
the maximum detection for determining firing (Eq. (10)), costs, the energy function can be obtained as follows from
we do not need to obtain the threshold for each neuron. the Eq. (7),
By transforming the objective function F1 (t) in Eq. (5)
1 Nm N
to the energy function EOBJ whose variables are only the 3
AP
2750
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.
neuron and omit such a constraint term from the energy Figure 3 shows the rate of the changes by the neural net-
function of the neural network. work optimization, which is (the results of the neural op-
By comparing EOBJ (t) in Eq. (17) and ENN (t) in Eq. timization)/(the results without the neural optimization),
(9), the connection weight Wijkl and the threshold θij can for throughput, power consumption and the communica-
be obtained as follows, tion cost, respectively. It shows that the total throughput
could be increased, and the power consumption and the
−V1 C1j for j = l
Wijkl = (18) communication cost could be decreased. Improvement for
0 otherwise the power consumption was small, because the parameters
for the power consumption Pj of wireless LAN type
θij = −V2 Pj − V3 Qj (19) systems were set equal for all. From these results, it is
The state of the neural network defined by Eqs. (18) and clarified that the designed neural network works correctly,
(10) converges to an optimal one for the load balancing only by autonomous neuron updates in Eq. (10) with Wijkl
problem through autonomous and distributed neuron in Eq. (18) and θij in Eq. (19).
updates on the terminal side. Since this is a simple model,
it is applicable to large-scale problems in heterogeneous 2.5e+07
Highest capacity AP selection
access point is 300, 300, 300 and 100, respectively. All the Throughput
1.5 Power consumption
access points are available for all users. We also assumed Cost
that the radio resources are shared equally among users. 1.4
The field sizes for the simulations are 1 km × 1 km to 50 1.3
km × 50 km, the number of wireless LAN access points is 1.2
1000, and the number of mobile terminals varies between
1000 to 50000. The wireless LAN access points and mobile 1.1
terminals are distributed in the test field with uniform 1
random numbers. The access points of the cellular system 0.9
are also distributed randomly but with a constraint that
0.8
the distances between them have to be longer than 800m. 40 50 60 70 80 90 100 110
We evaluated average available throughput per user with Average available APs at each location
and without a neural network based optimization scheme,
with changing the size of the test field. Fig. 3. Rate of the change of throughput, power consumption
and communication cost, by the proposed neural-network-based
Figure 2 shows the average throughput versus average optimization.
number of available access points at each location. The
number of channel for wireless LAN type networks was
limited to 4 for the results. The weights of the objective V. Conclusions
functions, V1 , V2 and V3 are all 1 in this simulation. In the We proposed a neurodynamical approach to optimize a
evaluation of the total throughput, the neural-network- large-scale heterogeneous wireless network environment.
based search utilizes the radio resources more efficiently To optimize scalable wireless networks, we need au-
and improves average throughput per user. tonomous and decentralized network selection on the
2751
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.
terminal side. Game theory has in the past been used to Clouds–, Proc. of IEEE Global Communications Conference
analyze the stability of distributed rule-based dynamics. (GLOBECOM), Nov. 2007.
[6] J. Neel, J. Reed and R. Gilles, Game models for Cognitive
We proposed an optimization scheme for large-scale wire- Radio Algorithm Analysis,SDR Forum Technical Conference,
less networks that has natural optimization dynamics of November 2004.
the mutual connection neural network. Since this approach [7] M. Hasegawa, H. Tran, G. Miyamoto, Y. Murata, S. Kato,
DISTRIBUTED OPTIMIZATION BASED ON NEURODY-
is based on clear properties defined by a mathematical NAMICS FOR COGNITIVE WIRELESS CLOUDS, Proc. of
model that decreases the energy function with autonomous International Symposium on Personal, Indoor and Mobile Radio
and distributed dynamics, it can be applied to various Communications (PIMRC), 2007.
[8] M. Hasegawa, H. Tran, G. Miyamoto, Y. Murata, H. Harada, S.
optimization problems. Kato, Autonomous and Decentralized Optimization of Large-
We can apply this optimization approach to general Scale Heterogeneous Wireless Networks by Neural Network
problems involving cognitive wireless networks. It is pos- Dynamics, IEICE Trans. of Communications, Vol. 91-B, No.
1, pp. 110-118, 2008.
sible to introduce various parameters and more practical [9] J. J. Hopfield, Neural Networks and Physical Systems with
objective functions, including the application’s require- Emergent Collective Computational Abilities, Proc. Natl. Acad.
ments, fairness, and so on, by modifying the objective Sci. USA, Vol. 79, pp. 2554-2558, 1982.
[10] J. J. Hopfield, Neurons with Graded Response Have Collective
energy function and calculating the connection weights in Computational Properties like Those of Two-State Neurons,
a similar way as shown in this paper. We can also deal Proc. Natl. Acad. Sci. USA, Vol. 81, pp. 3088-3092, 1984.
with multi-link access selection for link aggregation, by [11] A. Smith and M. Palaniswami, Static and Dynamic Channel
Assignment Using Neural Networks, IEEE Journal on Selected
changing the constraints and neuron update equations. It Areas in Communications, 15,2,238-249, 1997.
is also possible to adapt to wireless environment changes [12] Z. He et al., A Multistage Self-Organizing Algorithm Combined
as the mobile terminals’ move. Even if some parameters or Transiently Chaotic Neural Network for Cellular Channel As-
signment, IEEE Trans. Vehicular Technology, 51, 6, 1386-1396,
available wireless access points change we do not have to 2002.
change the autonomous neural network dynamics acting [13] J. J. Hopfield and D. W. Tank, Neural Computation of Decisions
as the global optimization tool. in Optimization Problems, Biological Cybernetics., Vol. 52, No.
3, pp. 141-152, 1985.
The neurodynamical approach has been applied to [14] E. H. L. Aarts and J. H. M. Korst, Boltzmann Machines for
dynamic channel assignment [11][12], which is a technique Traveling Salesman Problems, European Journal of Operational
on the network side that dynamically assigns channels Research, Vol. 39, no. 1, pp. 79-95, 1989.
[15] Y. Akiyama, A. Yamashita, M. Kajiura and H. Aiso, Combi-
according to the traffic load. Load balancing, which we natorial Optimization with Gaussian Machines, Proc, of IEEE
dealt with in this paper, is also categorized as radio International Joint Conference on Neural Networks, pp. 533-
resource management, but is on the terminal side and 540, 1989.
[16] D. E. Van den Bout and T. K. Miller, Improving the Perfor-
its objective is to share the available radio resources mance of the Hopfield-Tank Neural Network through Normal-
among the terminals, by utilizing any usable candidate ization and Annealing, Biological Cybernetics, Vol. 62, No. 2,
wireless network to improve the average throughput per pp. 129-139, 1989.
[17] H. Nozawa, Solution of the Optimization Problem using the
user. We considered the heterogeneous wireless network Neural Network Model as a Globally Coupled Map, Physica D,
environment, and applied neurodynamics to it, which is Vol. 75, pp. 179-189, 1994.
something that wasn’t dealt with in the previous research. [18] M. Hasegawa et al., An Analysis on Additive Effects of Non-
linear Dynamics for Combinatorial Optimization, IEICE Trans.
In the field of optimization using the neural networks, Fundamentals, Vol. E80-A, No. 1, pp. 206-213, 1997.
various modification methods have been proposed to [19] M. Hasegawa, T. Ikeguchi and K. Aihara, Combination of
improve the performance. Heterogeneous wireless network Chaotic Neurodynamics with the 2-opt Algorithm to Solve
Traveling Salesman Problems, Physical Review Letters, Vol. 79,
optimization could be also further improved by intro- pp. 2344-2347, 1997.
ducing such improvements. Since the mutual connection [20] M. Hasegawa et. al., A Novel Chaotic Search for Quadratic As-
neural network is basically suitable for scalable network signment Problems, European Journal of Operational Research,
Vol. 139, No. 3, pp. 543-556, 2002; M. Hasegawa, T. Ikeguchi and
optimization, the neural network can be an effective tool K. Aihara, Solving Large Scale Traveling Salesman Problems by
for solving various optimization problems in heterogeneous Chaotic Neurodynamics, Neural Networks, Vol. 15, No. 2, pp.
wireless network. 111-123, 2002.
[21] Y. Takefuji, K. Lee and H. Aiso, An Artificial Maximum Neural
Network: A Winner-Take-All Neuron Model Forcing the State
References of the System in a Solution Domain, Biological Cybernetics,
[1] H. Harada, Software defined radio prototype toward Cognitive Vol. 67, No. 3, 1992.
Radio Communication Systems, IEEE Dyspan 2005, Vol. 1, pp.
539-547, 2005.
[2] ed. B. Fette, Cognitive Radio Technology, Newnes, 2006.
[3] G. Wu, P. Havinga and M. Mizuno, MIRAI Architecture for
Heterogeneous Networks, IEEE Comm. Mag., pp. 126-134, 2002.
[4] M. Inoue, M. Khaled, H. Murakami, M. Hasegawa and H.
Morikawa, Novel Out-Of-Band Signaling for Seamless Inter-
working between Heterogeneous Networks, IEEE Wireless Com-
mun., Vol. 11, No. 2, pp. 56-63, 2004.
[5] H. Harada, H. Murakami, K. Ishizu, S. Filin, Y. Saito, H.
Tran, G. Miyamoto, M. Hasegawa, Y. Murata, S. Kato, A
Software Defined Cognitive Radio System –Cognitive Wireless
2752