The Use of Genetic Algorithms in Mobile Applications
Plechawska-Wojcik Malgorzata
Institute of Computer Science, Lublin University of Technology, Nadbystrzycka 36 B, Lublin, Poland
Keywords: Genetic Algorithm, MOGA, Travelling Salesman Problem.
Abstract: The goal of the paper is to present the application of genetic algorithm in practice. The main result is the
mechanism based on genetic algorithm applied in the application dedicated to tourists. The goal of the
mechanism is to propose the most effective route between points - tourist facilities. Those objects are also
chosen automatically based on user’s interest as well as on his and his friends opinions expressed via social
networking services Facebook. Genetic algorithm was implemented to obtain efficient way of solving the
problem of matching the appropriate route regarding requirements concerning time and location. The results
are obtained in short time by the genetic algorithm running on the web server. The paper presents also
results of the application and mechanisms testing, including performance testing.
1 INTRODUCTION
The motivation of the paper is to adapt the multi-
objective genetic algorithms to the problem of
determining route in the mobile application
dedicated to tourists. We developed mobile
application offering tourist routes individually
matched for a user who would like to visit new
locations without prior preparation. Applying
genetic algorithm should help to optimize results.
Genetic algorithms belong to a wide range of
evolutionary algorithms. They were proposed by
J. Holland (Holland, 1992) as a result of research on
strict adaptation processes observed in nature
(Streichert, 2001). All these algorithms search over
possible solutions of the problem. They are inspired
by real processes occurring in biological evolution,
but they differ with use of individual operators
(Dorronsoro, 2004; Goebel, et. al, 1998).
Common features of all evolutionary algorithms
are: high randomness manifested during the initial
population drawing and operators using, and quality
improvement of individuals with successive
generations reflecting the evolution occurring in
nature (Fox and McMahon, 1991).
Unlike the standard analytical algorithms
evolutionary algorithms do not process parameters
of given problem directly, but they use their encoded
form (McCarthy, 2007; Kumar, 2013; Kumar and
Jyotishree, 2012). Crossover operators do not
interpret the data (except heuristic operators), they
modify only their representations (Guanci, et. al,
2006). However, interpretation, known also as data
decoding, is required during the selection of
individuals, when individual individuals must be
evaluated using the objective function (El-Sharkawi,
et. al., 2005; Chakraborttya, 2013).
Evolutionary algorithms might be applied to
problems whose solution is impossible to achieve
using analytical methods. For example, such
problems are present in classes of optimization,
decision or optimization problems (like NP-hard
problems). Examples of evolutionary algorithms are
genetic and evolutionary programming or learning
and classifying systems (Yang, et. al, 2008).
2 PROBLEM STATEMENT
2.1 The Application Requirements
Application user selects initial and target location,
and the time they can spend on travelling. Initial
user's location might be assigned based on GPS built
in mobile device. The application assigns the best
route for user in given time, finding locations based
on user’s interest. Those interests are assigned based
on Facebook portal and likes given to different
locations by the user of his friends. Such locations
are described with different attributes such as
categories or coordinates. The identification of
geographic coordinates of locations and routes
520
Malgorzata P..
The Use of Genetic Algorithms in Mobile Applications.
DOI: 10.5220/0004952805200525
In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 520-525
ISBN: 978-989-758-027-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
between them might be defined based on Google
Maps service.
2.2 Problem Formulization
The main goal of the study was to:
1. build and implement the algorithm automatically
constructing the model of tourist facilities
(objects) adjusted to individual user’s
preferences
2. construct and verify the mechanism of selection
and arrangement of chosen points into the tourist
route meeting the requirements of time as well as
starting and finishing point.
The specificity of the mobile application project
imposed required the application worked fast and
efficient.
The construction of the solution (1) was
performed based on activities of user’s and his
friends on social networking portal – Facebook.
The solution (2) was developed gradually to find the
method meeting the requirement of its speed and
performance. That is why the research question was
“Is it possible to create mechanism of selection and
arrangement of points and present user the travel
map in relatively short time?”. This question was
verified based on efficiency testing of the
constructed mechanism.
3 ALGORITHM SELECTION
AND ADJUSTMENT
The propose of created algorithm is to propose user
a route satisfying its requirements in terms of his
interests, location of starting and destination point
and the time. This task can be formed as selection of
points from a given set of all available locations
maximizing all the attributes imposed by the user.
3.1 Applied Algorithm
The used algorithm is multi-objective genetic
algorithms (MOGA) (Shang and Zhang, 2009;
Yang, et. al., 2008). The main advantage of using an
evolutionary algorithm in this task is the ability to
targeting the solutions in terms of all occurring
dimensions due to appropriate use of the objective
function and selection of parents. Therefore, the
fulfilment of all the conditions determining the route
targeting the solutions in terms of all occurring
dimensions due to appropriate use of the objective
function and selection of parents. Therefore, the
fulfilment of all the conditions determining the route
quality can be obtained after only one run of the
algorithm.
3.2.1 Coding, Operators, Parameters
The problem domain must define the type of data
representation used to encode individuals in the
genetic algorithm. Natural chosen solution is an
ordered list of all the sites considered to present to
the user.
The algorithm implements two-point crossover
which gives good results in solving the travelling
salesman's problem. This choice results in necessity
to use the order encoding (OE). Crossing is done
with probability 80%, so it plays the major role in
genes development through forcing the genes
exchange between individuals. This is consistent
with one of the distinguishing features of genetic
algorithms from other evolutionary algorithms: the
dominant role of the crossover operator during the
reproduction of individuals.
The algorithms uses two mutation operators. The
first one replaces two separate strings within a single
chromosome. It is also responsible for abrupt
changes, and thus fulfils the basic functions of
mutation in the genetic algorithms, independent of
the problem domain. The second mutation operator
is necessary to obtain the correct route when the
target time does not allows to visit all the places.
This operator negates the substring value in the
second array describing the individual that stores
information whether a place a given position is
a part of the route or not.
Therefore, this operator can operate in two ways:
first, removing excess locations for routes with
limited time, and also, in later stages, of restoring
locations excluded from the route. In some cases it
can improve the quality of the route in terms of user
interests.
Of course this is not guaranteed, but in the worst
case individuals altered in this way will be
eliminated over the next generations through the
natural selection process. A mutation is performed
with the probability of 20%, so much less than the
crossover operation, but still quite significant. This
probability cannot be too low, because the algorithm
could not react properly to excluding locations form
chromosomes.
3.2.2 Objective Function
The algorithm used in the application uses three
objective functions responsible for routes
development in terms of their quality, user interests,
TheUseofGeneticAlgorithmsinMobileApplications
521
and time. Values returned by these functions are
used to generate the Pareto front during determining
the ranking of individuals. The values of all
objective functions are converted in such a way that
they can be minimized.
Before running the algorithm, ranking of
categories of available location is created on the
basis of information collected with Facebook.
Ranking is directly proportional to the number of
likes of category by the user or his friends. After that
all values in the ranking are normalized. The
objective function, during the route evaluation,
calculates the average values of categories for
locations assign to them. As the values are always in
the interval (0, 1), the final step is subtracting the
resulting value from one in order that the function
could be minimized.Finally, the objective function
responsible for estimation of route in terms of user
interests has the form (1)
1
(1)
where
∑
where
,
,⋯,
is a chromosome
representing a route, n – number of locations at the
route,
– category value for location i.
The objective function assessing the length of
route should promote routes with the least differing
the set time. To do it, a mapping between locations
lengths and the time domain.
The conversion between lengths and time was
made for medium walking speed, 5 km/h. The
absolute value of the difference between given time
and the route time (2).
|
|
(2)
where X – a chromosome,
- the time set by user,
- the time needed for finishing the route.
is
given with (3).
,1
∗1 (3)
where t(i,i+1) is the time needed to move from
particular locations of the route, n - number of
locations along the route, t_p - average time spent in
each of the locations.
The quality of the route needs to be controlled in
terms of the locations order. Genetic algorithms
randomly search the solutions space. Chromosomes
will be created with good time rates and user
interest, which are, however, not acceptable because
of locations disorder.
The objective function responsible for assessing
the route quality takes into account two factors. The
first is the number of intersections between lines
joining locations appearing on the route, the second
– the sum of angles between successive lines.
Evaluation of these aspects reduces the
probability of reproduction of individuals making
suboptimal route topologies. The algorithm should
be able to reject such solution or create a route,
which reorders locations, or removes or adds them in
order to minimize the difference between the
assumed and real time.
After calculating the total values of objective
functions for the entire population, individuals
ranking is created based on the Pareto front.
Afterwards it is used in selecting individuals to be
reproduced using tournament method. The place in
the ranking is determined by the number of
chromosomes dominated by an individual.
Individuals from all populations that make up the
Pareto front are transferred to a separate file. Before
each stage of reproduction chromosomes from this
collection are attached to the processed population.
This treatment significantly affects performance of
the algorithm, because it makes simultaneous
generating the population ranking and updating
Pareto front of all individuals. When the algorithm is
finished, routes that should be considered to be
presented to the user are located in the Pareto front
collection.
4 RESULTS
The characteristic feature of genetic algorithms is
randomness of generated results, due to their
mechanisms based on probabilistic. What is more,
the task is dependent on two factors. The first is the
same user and associated with it data hosted on
Facebook, determining the ranking of his interests.
The second factor is all the variables specified by the
user before starting the search for the route, such as
the place of start and destination and the total
travel time and waiting time. Of course it is not
possible to check all combinations of parameters that
can be passed to the genetic algorithm.
The following section presents the average
results obtained during the search process of ten
routes, each of them with the same parameters.
4.1 Parameters of the Algorithm
Genetic algorithm was running in following
configuration:
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
522
The number of starts: 10,
The number of chromosomes: 100,
The maximum number of generations: 200,
The probability of crossing: 80%,
The probability of mutation: 20%,
Duration of the tour: 4 hours,
Average time spent at location: 30 minutes.
Each instance of the algorithm worked on a fifty
locations, without the start and destination points.
4.2 Results
Before starting the searching process the ranking of
categories which is used to determine the quality of
location in terms of user interest. This action
requires to use objects marked by Facebook feature
"I like it", either by the user as well as related
friends. All these elements are available in Facebook
Graph API. This solution provides processing of
significant number of tagged objects (Table 1).
Table 1: Example of category ranking.
Category name Number of
objects marked
Ranking value
Sport 14 0.3783783784
Religion 9 0.2432432432
Culture 28 0.7567567568
Health 17 0.4594594595
Gastronomy 23 0.6216216216
Entertainment 37 1
The first column contains the name of the category,
the second the number of all elements in the
category, and the last column - value in the ranking.
Category names from the portal are mapped to the
names used internally by the application.
Table 2 presents data related to the initial
location, destination and all other attributes available
during the route searching process. The first column
contains the identifier for each location
corresponding to the identification of objects in the
Graph API of Facebook service. The second column
contains the ranking of categories value assigned to
the location allocated based on the ranking category.
Ranking values are normalized before running the
algorithm, they are in the interval <0, 1>. The third
and fourth columns, respectively, present width and
length of geographical location.
Figures 1, 2 and 3 show graphs of the best results
of specific criteria for the individuals evaluation,
depending on the single generation, respectively, to
time of the route, interests and its quality. Charts
also considers initial generation, randomly
calculated. Values occurring in each dimension has
been transformed in such a way that the algorithm
can minimize them.
Table 2: Data set of the algorithm.
Identifier Ranking
value
Geographical
width
Geographica
l
length
49598082
0459035
0.45945
945
51.24462 22.558498
50902751
2493449
0.62162
16
51.244976 22.558838
53788746
2921307
0.45945
945
51.242558 22.55469
A lower value means a greater quality of the
individual. All of these graphs present the
phenomenon of convergence of genetic algorithm,
achieving the point where the algorithm is not able
to create better individuals for next generations.
Figures 2-4 present detailed performance results.
Figure 2 is a graph of convergence of the algorithm
in time of the route. Values correspond to the
absolute value of the difference between the actual
time of the route, which also includes the time spent
in each location, and the target time specified by the
Figure 2: Best value for the criterion of the route, depending on the generation.
TheUseofGeneticAlgorithmsinMobileApplications
523
Figure 3: Best value for the criterion of interests, depending on the user.
Figure 4: Best value for the criterion of the quality of the route, depending on the generation.
user. The value obtained from the best individual
form the randomly created initial generation differs
from the average user expectations about 90,000
seconds (26 hours). Then, generation after
generation, the algorithm obtains better results until
it achieves alignment between the hundredth and one
hundred and fiftieth generation, where the difference
between the obtained and the expected length of the
route is reduced almost to zero.
Similar characteristics have graphs presenting
the convergence for criterion of interests and quality
of the route. In developing routes in terms of user
interests values can reach values in the range (0,1),
as a result of the values provided by the rating
category. Values corresponding to the quality of the
route have no upper limit , while the lower limit is
value zero. In practice, however, there is a very
small probability of taking the minimum value for
the fitness function for the route quality dimension,
because the route would have to be a perfect straight
line from the initial to the target point, by all
locations occurring on the path.
Table 3 shows the measured time of several
algorithm runs. Differences are caused by diversified
size of global for all generations Pareto front, which
is included in each stage of the individuals selection
for the reproduction.
Table 3: The timing of the genetic algorithm runs.
Instance
no
1 2 3 4 5
Time [s] 2,20 5,66 2,58 2,21 2,42
Instance
no
6 7 8 9 10 Mean
Time [s] 2,31 2,14 2,12 2,5 2,44 2,66
Increasing the number of Pareto front elements
increases the computational effort, as the calculation
it for n individuals requires n^2 comparisons.
Resizable Pareto front for each run of the algorithm
and for each generation is a result of randomness of
mechanisms existing in genetic algorithms.
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
524
5 CONCLUSIONS
From the obtained results we can draw the following
conclusions. Even the best individuals included in
the initial generation consisting of randomly created
chromosomes are very low quality. This
phenomenon can be seen primarily in an example of
the route criterion. What is more, the algorithm
optimizes all three dimensions equally describing all
individuals.
For all optimized dimensions algorithm achieves
convergence between a hundredth and one hundred
and fiftieth generation. In order to reduce the run-
time of the algorithm one can reduce the maximum
number of generations to, for example, 150 or 100.
This treatment can be combined with parallel
implementation of a dynamic stopping condition of
the algorithm which will stop the algorithm, if the
quality in all dimensions will not be improved
among the last generations.
In spite of the randomness present in the genetic
algorithms and the differentiation of the Pareto front
size, duration of the individual instance of the
algorithm for the same parameters varies slightly.
ACKNOWLEDGEMENTS
This work is supported with the scholarship under
the project "Qualifications for the labour market -
employer friendly university", cofinanced by
European Union from European Social Fund.
REFERENCES
Dorronsoro, B., 2004. Cellular evolutionary algorithms.
Springer-Verlag.
El-Sharkawi, M. A., Ngatchou, P., Zarei, A., 2005. Pareto
multi objective optimization. Intelligent Systems
Application to Power Systems. Proceedings of the 13th
International Conference, 84 – 91.
Fox, B., McMahon, M., 1991. Genetic operators for
sequencing problems, in Foundations of Genetic
Algorithms, G. Rawlins, Ed. Morgan Kaufmann
Publishers, San Mateo, CA, 284-300.
Goebel, R., Poole, D., Mackworth A., 1998. Computional
Intelligence: A Logical Approach. Oxford University
Press.
Guanci, Y., Quingsheng, X., Shaobo, Li., 2006. Studies on
fast Pareto genetic algorithm based on fast fitness
identification and external population updating
scheme. Global Design to Gain a Competitive Edge
An Holistic and Collaborative Design Approach based
on Computational Tools. Ed. Xiu-Tian Yan, Benoit
Eynard, William J. Ion.
Kumar, A., 2013. Encoding schemes in genetic algorithm.
International Journal of Advanced Research in IT and
Engineering, 2(3).
Kumar, R, Jyotishree, 2012, Novel Encoding Scheme in
Genetic Algorithms for Better Fitness, International
Journal of Engineering and Advanced Technology
(IJEAT), 1(6).
Holland, J., 1992. Adaptation in Natural and Artificial
Systems. Cambridge, MA: MIT Press.
McCarthy, J., 2007. What is artificial intelligence?
http://www-
formal.stanford.edu/jmc/whatisai/whatisai.html.
Shang, Y., Zhang, J., 2009. Improved multi-objective
adaptive niche genetic algorithm based on Pareto
front. Advance Computing Conference, 2009. IACC
2009. IEEE International, 300 – 304.
Streichert, F., 2001. Introduction to evolutionary
algorithms. Frankfurt MathFinance Workshop.
Yang, Y., Li, J., Dai, W., 2008. Multi-objective genetic
algorithm based on the correlation coefficient and its
application. Control and Decision Conference, 3898 -
3902.
Chakraborttya, R., Hasinb M, 2013. Solving an aggregate
production planning problem by using multi-objective
genetic algorithm (MOGA) approach. International
Journal of Industrial Engineering Computations, 4, 1–
12.
TheUseofGeneticAlgorithmsinMobileApplications
525
