You are on page 1of 6

A Survey of Fuzzy Cognitive Maps Forecasting Methods

Mohammed A. Al-Gunaid, Maxim V. Shcherbakov, Konstantin S. Zadiran


Aleksej V. Melikov
Volgograd State Technical University, Volgograd, Russia
Volgograd State Agrarian University, Volgograd, Russia
mohammadalgunaid@gmail.com
July 14, 2017

Abstract studied from data mining point of view. A lot of in-


dustries, like energetics [14] to energy consumption
There are a lot of forecasting methods developed till patterns recognition [15], etc. were already studied,
current time. All of them have their own advantages but a lot of industries still needed to be investigated.
and disadvantages. To decide, which method fits into One of the interesting industries is a rental karting
the certain situation, a comparison is required. business. Usually rental karting centres collects big
One of the highly developed trends now is a fuzzy amount of data about their visitors - personal infor-
cognitive maps forecasting. There are a lot of new mation, race statistics, payments etc, which can be
methods were recently proposed. Among them there investigated using data analysis methods.
are multi-step gradient method, method which is us- One of the latest data analysis methodologies is a
ing real-coded genetic algorithm and method which fuzzy cognitive map based approach [2]. Rental kart-
is using structure optimization genetic algorithm. ing business can be described as a set of dependent
In this paper a comparison of several fuzzy cog- factors, which is easily can be translated to terms
nitive maps multidimensional forecasting methods is of fuzzy cognitive map model. Fuzzy cognitive map
provided. (FCM) is a combination of fuzzy logic and cogni-
Main purpose of current paper is to compare fuzzy tive mapping for knowledge representation [2]. It de-
cognitive maps forecasting methods to each other and scribes the problem as a set of nodes (concepts)and
to compare all of them to popular and widely spread links (relations). The popularity of fuzzy cognitive
forecasting methods: linear regression and random maps can be explained by the universality of this tool
forest. for modeling complex dynamic systems [8].
Keywords — karting business, forecasting, mul- The main purpose of current article is to make
tidimensional forecasting, fuzzy cognitive map, com- a comparison of some fuzzy cognitive learning algo-
parison, random forest, linear regression rithms and common and highly used forecasting al-
gorithms - random forest and linear regression.
In section 4 of this paper compared methods are
1 Introduction described, in section 5 software which was developed
to make a comparison is described. Data, parame-
Data mining is become more and more popular nowa- ters, error measurements and results are present in
days because of developing of hardware and meth- section 6. In last section 7 the conclusion for usage
ods of data analysis. Data is aggregated in different fuzzy cognitive maps forecasting algorithms in frame-
industries and not every one industry was already work of karting business has made.

Authorized licensed use limited to: Universidad Pais Vasco. Downloaded on November 03,2022 at 13:29:33 UTC from IEEE Xplore. Restrictions apply.
2 Related Works and sticked to blackboard. Also connections between
these factors was drawn.
FCM is a general universal solution for different fields
of application by concept because of capability to rep- 3.3 Cooperation with karting centers
resent complex structures in convenient form of cog-
nitive map and relations in terms of fuzzy logic [2].
representatives
Researches with Fuzzy Cognitive maps were made in This cooperation was on stage, when real data was
different fields of activity: business [9], energy con- acquired, filtered and classified. The main purpose of
sumption forecasting [10, 11, 13], etc. this cooperation was to eliminate gaps in knowledge
There are a lot of algorithms for FCM learning ex- about time series. Time series contained a lot of un-
ist. Approaches for solving learning task are different: classified anomalies, which was determined, and ques-
some of them are based on genetic algorithms - real- tions about events in these days were asked to rep-
coded genetic algorithm [4], structure-optimization resentatives and answers helped to detect unknown
genetic algorithm [5], gradient methods - multi-step factors and remove unrelated information from time
gradient method [6], decision trees [11], fuzzy neural series.
networks [13], decision rules [12], etc.

4 Methodology
3 Information gathering
The construction of FCM requires a large amount
The research, on which this paper based, was started of concepts and connections that need to be estab-
from scratch. That is why the process of gathering in- lished, which substantially add to the difficulty of
formation was important in this research and should manual development process. There are two big chal-
be described. lenges: first, how the FCM could be drawn, if there
Salespeople, subject area specialists and karting is no expert to construct the aforementioned mental
centers representatives were involved into data col- model and second, how it could be analyzed, if there
lection process. is a way to construct that diagram with more com-
ponents. Therefore, learning mechanisms for FCMs
must be investigated to bridge these two gaps and
3.1 Cooperation with research area solve these problems [1]. The framework of proposed
specialists linguistic forecasting method is shown in Fig. 2.
This cooperation was intended to get the basic knowl-
edge and understanding of research area. Relying on 4.1 SOGA - Structure optimization
it information about business processes, customers genetic algorithm
flows and karting centers work was acquired. Main
concepts and factors were determined. This method is based on usage of modified type of ge-
netic algorithm to learn fuzzy cognitive maps. It de-
fines a new learning error function with an additional
3.2 Cooperation with sales people penalty for high complexity of FCM understood as a
large number of non-zero connections
For obtaining complete knowledge about business
processes and influencing factors salespeople were in- nr
J ∗ (l) = J(l) + b · 2 J(l) (1)
volved into research. Meeting with them was orga- n
nized. Where t - discrete time of learning, T - the number
For fixing knowledge blackboard and sticky notes of the learning records, b - parameter, b > 0, nr - the
was used. Before start of meeting, basic factors were number of the non-zero relations, n - the number of
written on sticky notes (one factor on single note) the concept.

Authorized licensed use limited to: Universidad Pais Vasco. Downloaded on November 03,2022 at 13:29:33 UTC from IEEE Xplore. Restrictions apply.
of the learning records, n - the number of the output
concepts, Ci (t) is the value of the i-th concept, Zi (t)
is the reference value of the i-th concept.

5 Software
Methods RCGA, SOGA, MGM were developed using
programming language R and development environ-
ment R-Studio.
For the linear regression standard library functions
are used. For the random forest regression method
the randomForest library is used.
All methods were run from a Shiny application.
For displaying charts the HighCharts framework
JavaScript library was used.
Developed application allows to specify datafiles,
select output and input concepts, specify share of
records to use in training in testing datasets.
Figure 1: The framework of proposed linguistic fore-
casting method for time series
6 Results
4.2 MGM - Multistep gradient 6.1 Testing data
method
The learning sample is a daily data from karting cen-
m1 m2
X X ter located in USA. Total number of records is 722.
Wji (t+1) = p[−1,1] ( αk Wji (t−k)− βl ηl (t)Jji (t−l))
Each record contains follow columns:
k=0 l=0
(2)
1. Revenue - daily revenue for karting center
The forecasting accuracy of population-based
learning algorithms was compared with multi-step su- 2. IsWeekend - indicates is current day weekend or
pervised learning based on gradient. Multi-step gra- not. Logical value - 1 if true, 0 if false.
dient method (MGM) [6] is described by the equation
(2). 3. IsHoliday - indicates is current day public holi-
day or not. Logical value - 0 if false, 1 if true.
4.3 RCGA - Real-coded genetic algo- 4. DayOfWeek - value of day of week for current
rithm day. Represented as integer number in range
from 1 to 7, where 1 stands for Monday, 7 - Sun-
One of the most well-known population-based learn- day, etc.
ing algorithm for FCM is RCGA [3, 4]. RCGA defines
the learning error function as follows:
6.2 Fuzzification procedure
T −1 X
n The following formula was used for fuzzification pro-
1 X
J(l) = (Zi (t) − Ci (t))2 (3) cedure:
n(T − 1) t=1 i=1
Xi − min(X)
Where: t - discrete time of learning, T - the number Xi = (4)
max(X) − min(X)

Authorized licensed use limited to: Universidad Pais Vasco. Downloaded on November 03,2022 at 13:29:33 UTC from IEEE Xplore. Restrictions apply.
where Xi - value of time series to fuzzify, X - the 1. Iterations: 10, population size: 10
whole time series.
2. Iterations: 25, population size: 25

6.3 Error measurements 3. Iterations: 100, population size: 50


The suitable forecasting methods and were chosen by
6.4.2 SOGA
considering the smallest value of MSE (Mean Square
Error) (5), RMSE (Root Mean Square Error) (6), and For structure-optimization genetic algorithm based
MAPE (Mean absolute percentage error) (7) [7]. method follow parameters used: mutation probabil-
n
ity - 0.5, keep best of current generation - true, b -
1X 2 0.2. Parameters are same to RCGA algorithm.
M SE = (Fi − Ri ) (5)
n i=1 Algorithm was run under different numbers of pop-
ulation size and iterations:
Where n - number of records in testing set, Fi - fore-
casted value, Ri - real value. 1. Iterations: 10, population size: 10
v
u n
u1 X 2. Iterations: 25, population size: 25
RM SE = t (Fi − Ri )2 (6)
n 3. Iterations: 100, population size: 50
i=1

Where n - number of records in testing set, Fi - fore- 6.4.3 MGM


casted value, Ri - real value.
n Multi-step gradient method learning algorithm was
1 X Ri − Fi started with parameters:
M AP E = ( ) · 100% (7)
n i=1 Ri 1. m1 = 3
2. m2 = 2
6.4 Methods parameters
3. ak : −0.2, 0.4, −0.8
Methods were run under follow sets of parameters:
4. βl : 2, 4
6.4.1 RCGA 5. λ = 10
For real-coged genetic algorithm based method follow and different number of iterations: 10, 100, 250
parameters used: mutation probability - 0.5, keep
best of current generation - true.
Parameters (same to SOGA algorithm):
6.5 Results

1. Mutation rate = 0.1


2. Keep best = True
3. Single point crossover
4. Random resetting mutation Figure 2: Testing set for Random forest algorithm
5. Reward-based selection
Algorithms were compared by values of MSE,
Algoritm was run under different numbers of pop- RMSE, MAPE and amount of time, spent on exe-
ulation size and iterations. cution.

Authorized licensed use limited to: Universidad Pais Vasco. Downloaded on November 03,2022 at 13:29:33 UTC from IEEE Xplore. Restrictions apply.
Error measurements Execution time,
Method name
MSE RMSE MAPE seconds
RCGA, 10 iter., 10 pop. size 0.05 0.23 0.95 7.17
RCGA, 25 iter., 25 pop. size 0.04 0.21 0.85 14.918
RCGA, 100 iter., 50 pop. size 0.04 0.20 0.81 85.819
SOGA, 10 iter., 10 pop. size 0.05 0.23 0.93 5.89
SOGA, 25 iter., 25 pop. size 0.05 0.23 0.92 15.569
SOGA, 100 iter., 50 pop. size 0.04 0.20 0.89 80.95
MGM, 10 iterations 0.06 0.24 0.92 9.813
MGM, 100 iterations 0.04 0.21 0.92 98.134
MGM, 250 iterations 0.04 0.21 0.92 242.57
Random forest 0.004 0.06 0.18 0.46
Linear regression 0.009 0.09 0.29 0.002

Table 1: Results of methods evaluation

ing of iterations. E.g. in comparison it can be seen


that on 100 and 250 iterations results of error mea-
surements are the same.
Random forest and linear regression in comparison
Figure 3: Testing set for RCGA algorithm run with to FCM-based algorithms have better performance
parameters - 10 iterations, 10 - size of population and accuracy in all error measurements. Random for-
est shows better accuracy than linear regression, but
lower speed of execution. But despite it, difference
RCGA and SOGA algorithms were run in parallel is not essential because of both algorithms provide
mode in 8 threads. results in less than 1 second.
Results of methods evaluations are presented in the
In karting business classic linear regression based
table 1.
methods now work better than fuzzy cognitive learn-
ing forecasting methods. Random forest and linear
7 Conclusion regression offer better accuracy with better perfor-
mance, than tested FCM based methods: RCGA,
Fuzzy cognitive maps based forecasting is highly de- SOGA, MGM.
veloped and very perspective sphere of researches
now.
As seen from results of running methods with dif-
ferent parameters, most precise algorithm among all
FCM algorithms is RCGA. On the second place there 8 Acknowledgments
is SOGA. It shows worse result in MAPE measure-
ment, but has better speed. Also RCGA and SOGA
can be run in parallel mode to increase performance The reported study was partially supported by RFBR
of calculation. research projects 16-37-60066 mol a dk, and project
MGM almost equals to RCGA and SOGA, but has MD-6964.2016.9. Also authors would like to thank
a slightly worse result. On big number of iterations Professor Peter P. Groumpos for fruitful discussion
results of it go stable and less changing with increas- and remarks.

Authorized licensed use limited to: Universidad Pais Vasco. Downloaded on November 03,2022 at 13:29:33 UTC from IEEE Xplore. Restrictions apply.
References [9] A. K. Tsadiras, ”Using fuzzy cognitive maps for
e-commerce strategic planning,” Proc. 9th Pan-
[1] E. Papageorgiou, ”Learning Algorithms for Fuzzy hellenic conference on Informatics, Thessaloniki,
Cognitive Mapsx2014;A Review Study”, IEEE Greece, pp. 142–151, 2003.
Transactions on Systems, Man, and Cybernetics,
Part C (Applications and Reviews), vol. 42, no. [10] M. Amer, A. J. Jetter, and T. U. Daim, ”Sce-
2, pp. 150-163, 2012. nario planning for the national wind energy sector
through Fuzzy Cognitive Maps,” Proceedings of
[2] B. Kosko, ”Fuzzy cognitive maps”, Int. J. Man- PICMET’13, pp. 2153 – 2162, 2013.
Machine Stud., vol. 24, no. 1, pp. 65-75, 1986.
[11] M. Al-Gunaid, M. Shcherbakov, D. Skorobo-
[3] W. Stach, L. Kurgan, W. Pedrycz and M. Refor- gatchenko, A. Kravets and V. Kamaev, ”Forecast-
mat, ”Genetic learning of fuzzy cognitive maps”, ing energy consumption with the data reliability
Fuzzy Sets and Systems, vol. 153, no. 3, pp. 371- estimation in the management of hybrid energy
401, 2005. system using fuzzy decision trees”, 2016 7th Inter-
national Conference on Information, Intelligence,
[4] W. Stach, L. Kurgan, W. Pedrycz and M. Re- Systems Applications (IISA), 2016.
format, ”Learning fuzzy cognitive maps with
required precision using genetic algorithm ap- [12] M. Al-Gunaid, M. Shcherbakov, V. Kamaev, O.
proach”, Electronics Letters, vol. 40, no. 24, p. Gerget and A. Tjukov, ”Decision Trees based
1519, 2004. Fuzzy Rules”, Information Technologies in Sci-
ence, Management, Social Sphere and Medicine,
[5] K. Poczeta, A. Yastrebov and E. Papageorgiou, pp. 502-508, 2016.
”Learning Fuzzy Cognitive Maps using Structure
Optimization Genetic Algorithm”, Proceedings of [13] M. Al-Gunaid, ”Neurofuzzy model of short-term
the 2015 Federated Conference on Computer Sci- energy consumption forecasting”, Caspian jour-
ence and Information Systems, 2015. nal: management and high technologies, vol. 2,
pp. 47-56, 2013.
[6] A. Yastrebov and K. Piotrowska, ”Simulation
analysis of multistep algorithms of relational [14] M. Shcherbakov, T. Yanovskiy, A. Brebels and
cognitive maps learning” in: A. Yastrebov, N. Shcherbakova, ”Methodology for identifying
B.Kuzminska-Solsnia and M. Raczynska (Eds.) the potential for energy savings based on data
Computer Technologiesin Science, Technology mining”, Caspian Journal: Management and
and Education, Institute for Sustainable Tech- High Technologies, vol. 2, pp. 51-55, 2011.
nologies - National Research Institute, Radom,
[15] T. Janovskiy, M. Shcherbakov, V. Kamaev and
pp. 126–137, 2012.
A. Janovskiy, ”Automatic Identification of Re-
[7] Shcherbakov M.V., Brebel’s A., Shcherbakova peating Energy Consumption Profiles”, World
N.L., Tyukov A.P., Yanovskiy T.A., Kamaev Applied Sciences Journal, vol. 24, no. 24, pp. 68-
V.A. ”A Survey of Forecast Error Measures”, 73, 2013.
World Applied Sciences Journal (WASJ), vol. 24,
spec. issue 24 : Information Technologies in Mod-
ern Industry, Education & Society, pp. 171-176,
2013.

[8] ”Fuzzy Cognitive Maps for Applied Sciences and


Engineering”, Intelligent Systems Reference Li-
brary, 2014.

Authorized licensed use limited to: Universidad Pais Vasco. Downloaded on November 03,2022 at 13:29:33 UTC from IEEE Xplore. Restrictions apply.

You might also like