I.
INTRODUCTION
convergence to the NE. On the other hand, FPC does not need
any iteration to yield the output at all and will respond
immediately which will make it quite suitable for very
dynamic environments.
Fuzzy controllers have the advantages of robustness, ease
of design and flexibility. Besides, in fuzzy control, an exact
mathematical model is not necessary, because linguistic
variables are used to define the system behavior making the
design process simpler. Also, fuzzy controllers have shown a
great ability to map nonlinear and complicated relationships
between input and output spaces [13].
Some examples of the use of fuzzy controllers in wireless
systems include the following. In [15], a fuzzy open-loop
transmission power control for WLANs has been proposed
which aims to minimize node transmission power, while
keeping the transmission errors at an acceptable level. The
authors in [16] proposed a transmission power adjustment
scheme using the fuzzy control in wireless sensor networks
for dynamic topology control. For further study of the
utilization of fuzzy logic for power control in CDMA
networks refer to [17][18].
Fig. 1.
Interference Model
shown in Fig 2.
Fig. 2.
GENETIC LEARNING
Fig. 3.
A.
Chromosome Structure
To codify the fuzzy controller as a chromosome, we
consider that each chromosome is formed by two parts:
parametric genes and rule genes that represent the fuzzy
membership functions and the fuzzy rules, respectively. For
the membership functions, we use a Gaussian function:
( x mij ) 2
ij ( x) = e
2 ij2
(1)
Fig. 4.
Fig. 5.
ijk
B.
Genetic Operators
Since we have real valued genes codifying the membership
function parameters and integer valued genes codifying the
fuzzy rules in the chromosome structure, we do not employ
the usual binary code operators, i.e. N-point crossover and bit
inversion for recombination and mutation respectively.
Instead, for recombination, we use two crossover methods
comprising arithmetic and heuristic crossovers being selected
randomly with equal probabilities. In arithmetic crossover, an
interpolation along the line formed by the corresponding
genes of two parents is performed, while in the heuristic
crossover, an extrapolation is done along the line formed by
the two parents outward in the direction of the better parent.
The mutation operator is uniformly selected from three
methods which are Gaussian, uniform and non-uniform
mutations. In the Gaussian mutation, the gene is changed with
the probability of a normal Gaussian distribution. The other
two methods change the value of the gene based on a uniform
distribution and non-uniform distribution respectively in the
specified range of the variable.
The selection strategy utilized to pick individuals into the
mating pool to produce an offspring for the next generation is
a ranking selection based on the geometric distribution.
C.
Learning Strategy
During the learning process, in addition to the elitism that
takes place at each generation, referred to as over-generation
elitism, we also employ an over-step elitism in which after a
certain number of generations, the WBANs move according to
TABLE I.
hii P i
( P , P ) = log(1 + SINRi ) ci P
(4)
where ci and are pricing parameters which provides the
ability to adjust the tradeoff between capacity and power
consumption.
In [25], the authors proved the existence and uniqueness of
the NE in this game and proposed a best-response strategy for
players to reach the NE.
The second approach is the ADP algorithm [26], which is a
game theory based power allocation approach designed for
sensor networks. The ADP algorithm applies a pricing
mechanism to coordinate interference among nearby nodes so
as to improve the overall system capacity.
Table 1 summarizes the parameters and their values used in
our simulations.
t
Fitness Function
The GA tries to find the individual which maximizes a
given fitness function that conveys the objectives of the
optimization problem at hand. In the present case, the
optimization problem the GA solves is to find a fuzzy
controller which maximizes link capacity using as little power
for transmission as possible in the shortest convergence time.
A candidate fitness function reflecting these objectives can be
as follows:
= C P N
(2)
where C is the link capacity calculated using the Shannon
channel capacity formula: B log 2 (1 + SINR ) , normalized by the
maximum capacity achievable at zero interference; P is the
value of transmission power normalized by the maximum
power Pmax; N is the normalized value for the maximum
number of iterations needed by the FPC to converge to a
steady state output; and , , and are weighting factors
whose dimensions make dimensionless. Their values can be
varied to change the importance of C, P, and N in the fitness
function. In this study we set all the weighting factors to one.
D.
h
j i
ji
P j + B.N 0
t
SINRi =
(3)
Fig. 6.