CC BY
22
https://orcid.org/0000-0002-3931-6729 https://orcid.org/0000-0002-5523-6120
UDC 004.9; 504.064.3
A.S. AVRAMENKO*, S.V. GOLUB*
DECREASE OF TIME OF MODEL SYNTHESIS IN INTELLECTUAL MONITORING SYSTEMS
Bohdan Khmelnytsky Cherkasy National University, Cherkassy, Ukraine **Cherkasy state technological university, Cherkassy, Ukraine
Анотаця. У cmammi дослгджено сучасн Monimopumoei ттелектуальт системи (М1С), яю здатн прогнозувати на^дки прийнятих керуючих ршень систем тдтримки прийняття ршень (СППР) завдяки моделюванню характеристик об 'eктiв монторингу. Продемонстровано недолти iсную-чих реалiзацiй М1С при роботi в умовах кризового монторингу. Так як кризовий монторинг на-кладаеряд обмежень на швидюсть роботи СППР та велику вiрогiднiсть виходу навчених моделей М1С iз строю, то використання iснуючих реалiзацiй М1С е проблематичним. До^джено причини iснування даних недолтв та алгоритми, з якими це пов 'язано. До^джено переваги та недолти iснуючих методiв формування мiжрiвневих зв 'язюв у М1С. Особливу увагу звернено на метод кла-сифтацп масивiв вхiдних даних (МВД) за гх характеристиками до вiдповiдного класу алгоритмiв синтезу моделей (АСМ). Запропоновано вдосконалити вiдомий метод класифiкацiг масивiв вхiдних даних за допомогою використання уткальних адаптивних класифiкаторiв для кожного iз клаав алгоритмiв синтезу моделей iз списку реалiзованих у системi. Проведено тестування запропоно-ваних вдосконалень. Для проведення тестування запропонован вдосконалення було реалiзовано у програмному комплека, побудованому на кафедрi ттелектуальних систем прийняття ршень Черкаського нащонального утверситету iменi БогданаХмельницького. Об 'ектом монторингу для тестування було обрано результати спостережень за захворюватстю населення Черкаськог об-ластi впродовж 2000-2016роюв. Для оцтки роботи методу використовувались показники якостi, отриман при навчанн моделей, та швидюсть навчання. За результатами тестування вдалося досягти збтьшення швидкостi навчання системи в 3-4 рази при незначних втратах в якостi отриманих ктцевих моделей, що не перевищують 4% похибки моделювання.
Ключовi слова: мотторинг кризи, монiторинговi ттелектуалью системи, синтез моделей, кла-сифтащя, багаторiвневе моделювання, сктченний автомат, швидюсть навчання моделей, час реструктуризацИ системи, помилка моделювання.
Аннотация. В статье исследованы современные мониторинговые интеллектуальные системы (МИС), которые способны прогнозировать последствия принимаемых управляющих решений систем поддержки принятия решений (СППР) благодаря моделированию характеристик объектов мониторинга. Продемонстрированы недостатки существующих реализаций МИС при работе в условиях кризисного мониторинга. Так как кризисный мониторинг накладывает ряд ограничений на скорость работы СППР и большую вероятность выхода обученных моделей МИС из строя, то использование существующих реализаций МИС является проблематичным. Исследованы причины существования данных недостатков и алгоритмы, с которыми это связано. Исследованы преимущества и недостатки существующих методов формирования межуровневых связей в МИС. Особое внимание обращено на метод классификации массивов входных данных (МВД) по их характеристикам к соответствующему классу алгоритмов синтеза моделей (АСМ). Предложено усовершенствовать известный метод классификации МВД посредством использования уникальных адаптивных классификаторов для каждого из классов АСМ. Для проведения тестирования предложенные усовершенствования были реализованы в программном комплексе, построенном на кафедре интеллектуальных систем принятия решений Черкасского национального университета имени Богдана Хмельницкого. Объектом мониторинга для тестирования были выбраны результаты наблюдений за заболеваемостью населения Черкасской области в течение 2000-2016 годов. Для оценки работы метода использовались показатели качества полученных при обучении моделей и скорость обучения. По результатам тестирования удалось достичь увеличения скорости обучения системы в 3-4 раза при незначительных потерях в качестве полученных конечных моделей, не превышающих 4% погрешности моделирования.
© Avramenko A.S., Golub S.V., 2019
ISSN 1028-9763. Математичш машини i системи, 2019, № 3
Ключевые слова: кризисный мониторинг, мониторинговые интеллектуальные системы, синтез моделей, классификация, многоуровневое моделирование, конечный автомат, скорость обучения моделей, время реструктуризации системы, ошибка моделирования.
Abstract. The article investigates modern intellectual monitoring systems (IMS), which are able to predict the consequences of the adopted control decisions of decision support systems (DSS), thanks to the modeling of the characteristics of monitored objects. The drawbacks of existing implementations of IMS show when working in crisis monitoring. Since crisis monitoring imposes a number of restrictions on the speed of DSS and the high probability offailure of the trained IMS models, the use of existing implementations of IMS is problematic. The reasons of the existence of these shortcomings, and the algorithms with which it is connected lies in existing methodology. The paper investigates advantages and disadvantages of existing methods for the formation of inter-level relations in the IMS. A particular attention is paid to the method of classification of input data arrays (IDA) according to their characteristics, to the corresponding class of model synthesis algorithm (MSA). This paper proposes to improve the well-known method of classifying MIA by using unique adaptive classifiers for each of the MSA class. For testing, the proposed improvements implemented in a software package built at the department of intelligent decision-making systems at Bohdan Khmelnytsky Cherkasy National University. The monitored object for testing was selected from the results of observations of the disease incidence of the population of Cherkasy region during the years 2000-2016. To evaluate the performance of the method, the quality indicators obtained from the learning models and the learning speed were used. According to the test results, it was possible to achieve an increase in the learning rate of the system by 3-4 times, with insignificant losses in the quality of the final models obtained, not exceeding 4% of the modeling error.
Keywords: crisis monitoring, intellectual monitoring systems, model synthesis, classification, multilevel modeling, finite state machine, model learning speed, system restructuring time, modeling error.
DOI: 10.34121/1028-9763-2019-3-129-134
1. Description of a task
1.1. Introduction
With developing of modern technologies of multilevel modeling [1] it became easier and more effective to build the intellectual monitoring systems (IMS). The main objective of these systems is to provide information for decision-making according to parameters given by client. However, ISM not just gather information they can also simulate reactions to a made control decision by modeling the parameters of monitored objects.
Systems like this is popular in socioecology, medicine, economics, cybernetics and other spheres of study where there are many objects that need to be monitored.
Example of such tasks is crisis monitoring, which is a monitoring of objects in emergency situations. Decision making in this situations demands fast and adequate processing of gathered data with a bigger adaptivity to quick changes to characteristics and structure of input data. Part of objects characteristics can lose their informativity that will demand of finding additional characteristics instead. It increases the chances of errors while the decision making system uses the premade models. Such broken models will then be resynthesized witch will increase the time and cost of using such monitoring systems. In crisis monitoring tasks, such cost is inexcusable.
Therefore, the main task of this article will be to ensure the reduction of time of re-learning the system while maintaining the quality parameters of the models.
1.2. Analysis of monitoring system
Multilevel monitoring systems built as hierarchical combination if multiparameter models [2].
Models like this synthesized by using special inductive algorithms, neural networks, genetic algorithms and others.
In this technology selection of algorithms for model synthesis (AMS) implemented by testing the models synthesized by all of them, and choosing the best model. On Fig. 1 presented a subsystem for information processing in automated hierarchical system for multilevel socio-ecological monitoring.
Models on every level of hierarchy solve their own local tasks of information processing and combination of all levels solves the global task of a system. Such structures can combine many models for example one hundred and more.
In process of monitoring in emergency situations characteristics of input data arrays (IDA) are constantly changing. This means that there is a high chance that one or multiple pre made models could start give inadequate results. To repair such «damaged» models we need to replace them and all models related to them with newly synthesized models.
2. Suggested ways to solve the problem
Obviously, characteristics of IDA are different for different objects. This means that we can choose different AMS individually for each IDA [3]. This way, adaptation of model synthesis to changes in characteristics of IDA is ensured. Today in existing IMS synthesis of models made by sequential testing of algorithms implemented in a system with subsequent choice of the best.
Using this we formulated hypothesis that reducing of time for synthesis of models can be achieved by adaptation of model synthesizer. It proposed to adapt synthesis by solving the problem of classification of IDA. At the same time, the main task becomes building a deciding rule with which we determine affiliation of new IDA to a class of IDA's which best AMS already found experimentally.
Thus, we have the set Q of IDA classes, whose power is determined by the volume of the constructed AMS:
IQ \=r+1, (1)
where y - number of AMS constructed in system. +1 for a class "impossible to classify".
Characteristics of IDA are represented by the set X. Elements of set X is vectors Xi g X, structure of which contains attributes of classification characteristics of the IDA [3]:
X = (x,, x-, ..., x }, i = 1,n, (2)
l ^ il? i27 7 im> > V /
where m - numbers of IDA characteristics, n - numbers of IDA used in building classifier and equal to the power of set X.
To provide the models with specified parameters we tested and found the best AMS for each IDA. This will let us find out which elements of X associated with elements of Q, meaning we found the answer to:
a :X ^ Q, (3)
t.
Mr
FWP/
Figure 1 - Hierarchical structure of subsystem for information conversion
experimentally.
We need to find analytical expression for a deciding rule a (3) which will provide maximal numbers rightly classified elements of X and provide effective work of model synthesizer.
Synthesizer can be described in a form of deterministic finite automaton:
M = (V, Q, qo, F, S), (1)
where V = {0,1,..., k} - input alphabet, k - numbers of AMS known to system, Q - set of automaton states, q0 - starting state of automaton, F - set of finishing automaton states F c Q, S - switch function.
S :Qx (V ^ Q . (5)
Therefore, IDA classifier generates a signal that brings the information about what class
of AMS each IDA belongs to, which will tell us what ASM need to be used to synthesize best models. Model synthesizer receives the signal from classifier automaton switches into state that represents the appropriate AMS. In this state AMS uses IDA and follows the steps of synthesizing, testing and the usage of model. The whole algorithm presented schematically in Fig. 2.
Thus was formed a hypothesis that to build the deciding rule we need to use algorithms for inductive modeling which was already implemented inside of model synthesizer. On input, we need to send an array of characteristics vectors of IDA, which is suggested in [3].
The algorithm represented in Fig. 2 and all required functions was implemented as a part of IMS created on department of intellectual systems for decision making in Bohdan Khmelny-tsky Cherkasy National University. After that, the IMS was used to confirm the hypothesis.
To create a solving rule we used multi GMDH algorithms [2]. To synthase models, we used the Results of monitoring morbidity in Cherkassy region during 2000-2016 years [1]. We write a special application that implemented all algorithms and options needed. AMS's in this application mostly constructed based on GMDH algorithms with different options. For criteria with which we choose, best models and best algorithms that made the model can be used standard deviation and absolute deviation. In addition, to make a quality of classification models better we used the method of adaptive level multiplying for each of them [8].
In Fig. 3, we can see comparison of time spent on finding and synthesizing models with the best AMS between "test all and compare the results" method that is used as standard and our classification algorithm. Comparison shows us that speed of synthesis is 4 times bigger at max and 70% bigger in average.
To test the quality parameters of synthesized models between algorithms we used two cases. First, one (see Tabl. 1) is when models modeled separately and use their own IDA. Second, one (see Tabl. 1) is when models grouped in a strict hierarchy where models of higher levels use data provided by models of lesser levels in their IDA. In both cases, modeling error of classification algorithm did not differ, from standard algorithm more than in 5% on average.
Figure 2 - Functional scheme for process of IDA classification and model synthesis
sec
1 2
5 6 7 8 —Full exsastion
9 10 11 12 13 14 15 ^—Classification №
Figure 3 - Speed of synthesis in comparing with classifier Table 1 - Comparison of the modeling errors
Diseases of Modelling error %
without hierarchy with hierarchy
standard classification standard classification
Breathing 12,4 12,4 12,4 12,4
Blood 8,22 8,22 8,22 8,22
Stomach 111 111 13,6 13,6
Endocrine system 13,6 13,6 37,1 38,4
Nervous system 35 37,1 26 26
Bronchitis 26 26 10,2 10,2
Asthma 10,2 10,2 16,8 16,1
Gastritis 16,8 16,8 11,3 11,3
Diabetes 11,3 11,3 26,6 24,3
Iron-deficiency anemia 26,6 26,6 27,3 27,3
Allergy 180 215 30,5 22,6
Pneumonia 27,3 27,3 40,1 40,1
Genitourinary system 30,5 30,5 19,9 19,9
Glomerulonephritis 36,8 40,1 7,79 7,78
Genetic anomaly 19,9 19,9 3,32 3,2
4
1
0
3
4
3. Conclusions and suggestions
Growth of modeling error is "payment" for reducing the time of model synthesis. Given the fact that the structure of the information system of multilevel data transformation contains 100 models and more, it is possible to achieve a significant reduction in the time structure by adapting to changes in the properties of IDA. Results like this give us hope and possibility to effectively use IMS with multilevel information processing technologies to provide data for a decision-making in situations of crisis monitoring.
In this method the reduction of time for synthesis of models reached by replacing the full testing of AMS for a deciding rule that can classify IDA to a best AMS for it.
REFERENCES
1. Голуб С.В. Багаторiвневе моделювання в технолопях мошторингу оточуючого середовища. Черкаси: ЧНУ, 2007. 218 с.
2. Ивахненко А.Г. Индуктивный метод самоорганизации моделей сложных систем. Киев: Наукова думка, 1981. 296 с.
3. Колос П.О. Визначення множини шформативних параметрiв таблиц первинного опису об'екта моделювання. Bîcuuk Черкаськогоуниверситету. Черкаси: Вид. ЧНУ, 2009. Вип. 173. С. 121-128.
4. Гмурман В.Е. Теория вероятностей и математическая статистика: учеб. пособ. для вузов. 10-е изд. Москва: Высшая школа, 2004. 479 с.
5. Бахрушин В.Е. Методы оценивания характеристик нелинейных статистических связей. Системные технологии. 2011. № 2 (73). С. 9-14.
6. Гантмахер Ф.Р. Теория матриц. М.: Наука, 1988, 548 с.
7. Харебов К.С. Компьютерные методы решения задачи наименьших квадратов и проблемы собственных значений. Владикавказ: Изд-во СОГУ, 1995, 76 с.
8. Голуб С.В., Немченко В.Ю. Адаптивне формування дублюючих рiвнiв в структурi ieрархiчних систем багаторiвневого сощогшешчного мошторингу. 1ндуктивне моделювання складних систем: зб. наук. праць / ред. В.С. Степашко. Кшв: Мiжнар. наук.- навч. центр шформ. технологш та систем НАН та МОН Украши, 2011. Вип. 1. С. 41-48.
Стаття над1йшла до редакцп 08.07.2019
