Professional Documents
Culture Documents
30-35
UDC 681.586
Informational methods for analyzing and managing systems under uncertainty are studied.
The expediency of use of information uncertainty in tasks of identification of control
objects and synthesis of regulatory systems is justified. For the numerical evaluation of
information uncertainty, an amount of disinformation is used as Bongard’s negative useful
information. Such information uncertainty can serve as a criterion for the adequacy of a
mathematical model of a control object. Information assessment of the modeling accuracy
is applicable to any method of identification and allows a researcher to compare the
accuracy of models that differ from each other by the method of obtaining (analytical or
experimental), nature (deterministic or stochastic models), specific implementation (physical
or mathematical models). In the course of parametric identification, an optimization
problem of finding a minimum of information uncertainty in the parameter space of a
mathematical model can be solved. The information uncertainty criterion provides
verification of the statistical hypothesis about the adequacy of a particular model. If the
criterion value exceeds a certain critical value, the adequacy hypothesis must be rejected.
To calculate the critical values of the information adequacy criterion, a statistical experiment
was performed. Using the Monte Carlo method, the probability distribution of the
information criterion was investigated. A sufficiently smooth empirical criterion distribution
function was constructed. The distribution of the information criterion has a pronounced
asymmetry and a small positive kurtosis. It is revealed that this distribution is best
approximated by the Generalized Extreme Value Distribution law. The critical value can
be defined as a quantile of the level of 0.01 or 0.05 for this distribution.
DOI: 10.32434/2521-6406-2019-5-1-30-35
Statement of the problem the obtained model. To solve this problem, additional
As known, the synthesis of control systems researches are carried out, listed in the report [1].
begins with the process of obtaining a mathematical Verification is the process of determining that
model of a controllable object. Meanwhile, a model implementation accurately represents the
developers are encountering the problem of developer’s conceptual description of the model and
incomplete correspondence of models with an its solution.
identified object. Thus, during system synthesis, the Validation is the process of determining the
models remain uncertain. degree to which a model is an accurate representation
The sources of model uncertainty are: of the real world from the perspective of the intended
– incomplete knowledge of the physical and uses of a model.
chemical processes occurring in the object; Both verification and validation are processes
– assumptions and simplifications, taken during that accumulate evidence of a model’s correctness
a model formation; or accuracy for a specific scenario; thus, they cannot
– noises, interferences and measurement errors prove that a model is correct and accurate for all
taking place in experimental researches of objects; possible scenarios, but, rather, it can provide evidence
– parameters changes in the object over time. that a model is sufficiently accurate for its intended
Ànother problem is the quality assessment of use.
performed by the existence factor of its occurring in Thus, the uncertainty estimate brought by the
relation to a certain functioning of the system model is disorganization according to formula (3),
indicator; pj, si, are the weights of j-th situation, in which Уj =pj/qj. We will call the amount of
i-th element and -th time interval, respectively. misinformation (7) the information uncertainty.
Taking formula (2) as a basis, we can derive a Obviously, the largest value of misinformation
set of relations that can be used as an estimate of occurs when H=0. Dividing the information
uncertainty. To simplify, let’s take d=1 and l=1. We uncertainty (7) by its maximum possible value, we
perform the operation of generalization by m obtain the relative information uncertainty:
situations with weighting by probabilities included
in the ensemble (1): pj
m D
p qj
j logj p j log p j
О p j log Уj ,
j
(3) 1 . (8)
j1
sup D p j log q j p j log q j
j j
where Уj is some numerical characteristic of
from X
differences X et , called the disorder Ways to use informational uncertainty in
parameter. identification tasks
Taking a j-th situation disorder probability Information uncertainty in the form (7) can
be used as a criterion for the adequacy of
Уj p j as a parameter, we obtain Shannon’s
mathematical models. Accurate assessment of
informational entropy formula: modeling accuracy is possible with any method that
makes it possible to compare models with different
characteristics (analytical or experimental), character
O H p j log p j . (4) (deterministic or stochastic models), implementation
j
of realizations (physical or mathematical models).
A mathematical model cannot be defined as minimal
Mathematical model of a system under uncertainty.
investigation due to inevitable inaccuracies gives Relative information uncertainty in the form
distribution Q={qj} instead of real distribution P={pj}. (8) can be used as a criterion for the adequacy of
Taking Уj q j , we obtain Bongard’s uncertainty models obtained by other methods. Moreover, there
is both the possibility of comparing several models
formula [10]: and choosing the most adequate, taking the least
amount of misinformation, as well as examining the
adequacy of a particular model’s statistical hypothesis.
O N(p / q) p j log q j . ( 5) Here is the way to determine whether the model
j
is adequate. Let us assume that there is some object,
a state of which is characterized by the variable y
According to Bongard, useful information (output value of the object). As a result of
stored in the distribution Q={qj}, regarding the experimental researches, which used the method of
problem with a probability of response P={pj}, is: a passive experiment, for instance, a statistical
sampling of the object’s values was obtained. The
qj estimate of probabilities pj is a ratio of a sample’s
Iп H N(p / q) p j log . (6) number of vj implementations, falling within a j-th
j pj
separation interval of a range of y value, to the sample
size n.
Unlike the amount of information according Let there also be a mathematical model of such
to Shannon’s classical theory, useful information can an object, for which model value of the output
have both positive and negative meaning. Due to variable ŷ can be calculated, and by methods of
inaccuracies of a system’s mathematical description,
information theory or mathematical statistics, a
we should talk about the misinformation:
distribution Q={qj} can be found.
pj We hypothesize that a mathematical model is
D N(p / q) H p j log . (7)
adequate to an object. To examine the hypothesis,
j qj we need to calculate a criterion of information
uncertainty value for model in the form (7), for which
7. Fang X., Wang L. Feature selection based on fisher êîíòðîëüîâàíèé îá’ºêò, ñòàòèñòè÷íà ã³ïîòåçà, íåâèçíà÷åí³ñòü.
criterion and sequential forward selection for intrusion detection
// Revista de la Facultad de Ingeniería U.C.V. – 2017. – Vol.32. ÈÑÏÎËÜÇÎÂÀÍÈÅ ÈÍÔÎÐÌÀÖÈÎÍÍÎÉ
ÍÅÎÏÐÅÄÅËÅÍÍÎÑÒÈ Â ÇÀÄÀ×ÀÕ
– No 1. – P.498-503.
ÈÄÅÍÒÈÔÈÊÀÖÈÈ
8. Äîðîæêèí À.Ì., Ñîêîëîâà Î.È. Ïîíÿòèå «íåîïðåäå-
ëåííîñòü» â ñîâðåìåííîé íàóêå è ôèëîñîôèè // Âåñòíèê Ìàíêî Ã.È., ×èñòîêëåòîâ Å.Ï.
Âÿòñêîãî ãîñóäàðñòâåííîãî ãóìàíèòàðíîãî óíèâåðñèòåòà. – Èçó÷àþòñÿ èíôîðìàöèîííûå ìåòîäû àíàëèçà è óïðàâ-
ëåíèÿ ñèñòåìàìè â óñëîâèÿõ íåîïðåäåëåííîñòè. Îáîñíîâàíà
2015. – C.5-12.
öåëåñîîáðàçíîñòü èñïîëüçîâàíèÿ èíôîðìàöèîííîé íåîïðåäåëåí-
9. Ýëåêòðè÷åñêèå ñèñòåìû. Êèáåðíåòèêà ýëåêòðè÷åñ- íîñòè â çàäà÷àõ èäåíòèôèêàöèè îáúåêòîâ óïðàâëåíèÿ è ñèí-
êèõ ñèñòåì : ó÷åá. ïîñ. äëÿ ýëåêòðîýíåðã. âóçîâ / Ïîä ðåä. òåçà ñèñòåì ðåãóëèðîâàíèÿ. Äëÿ ÷èñëîâîé îöåíêè èíôîðìàöè-
Â.À. Âåíèêîâà. – Ì.: Âûñø. øêîëà, 1974. – 328 ñ. îííîé íåîïðåäåëåííîñòè èñïîëüçóåòñÿ êîëè÷åñòâî äåçèíôîðìà-
öèè êàê îòðèöàòåëüíîé ïîëåçíîé èíôîðìàöèè Áîíãàðäà. Òàêàÿ
10. Áîíãàðä Ì.Ì. Ïðîáëåìû óçíàâàíèÿ. – Ì.: Íàóêà,
èíôîðìàöèîííàÿ íåîïðåäåëåííîñòü ìîæåò ñëóæèòü â êà÷å-
1967. – 320 ñ. ñòâå êðèòåðèÿ àäåêâàòíîñòè ìàòåìàòè÷åñêîé ìîäåëè îáúåê-
11. Ìàíêî Ã.È., Äîâãîïîëàÿ A.A. Îá èñïîëüçîâàíèè èí- òà óïðàâëåíèÿ. Èíôîðìàöèîííàÿ îöåíêà òî÷íîñòè ìîäåëèðî-
ôîðìàöèîííîãî êðèòåðèÿ ïðîâåðêè ãèïîòåç î çàêîíå ðàñ- âàíèÿ ïðèìåíèìà ïðè ëþáîì ìåòîäå èäåíòèôèêàöèè è ïîçâî-
ëÿåò ñðàâíèâàòü òî÷íîñòü ìîäåëåé, îòëè÷àþùèõñÿ äðóã îò
ïðåäåëåíèÿ îøèáîê èçìåðåíèÿ // Âîïðîñû õèìèè è õèìè-
äðóãà ìåòîäîì ïîëó÷åíèÿ (àíàëèòè÷åñêèì èëè ýêñïåðèìåíòàëü-
÷åñêîé òåõíîëîãèè. – 2013. – ¹ 1. – Ñ.181-184. íûì), õàðàêòåðîì (äåòåðìèíèðîâàííûå èëè ñòîõàñòè÷åñêèå
12. Ð 50.1.037-2002. Ðåêîìåíäàöèè ïî ñòàíäàðòèçàöèè. ìîäåëè), êîíêðåòíîé ðåàëèçàöèåé (ôèçè÷åñêèå èëè ìàòåìàòè-
Ïðàâèëà ïðîâåðêè ñîãëàñèÿ îïûòíîãî ðàñïðåäåëåíèÿ ñ òåî- ÷åñêèå ìîäåëè).  õîäå ïàðàìåòðè÷åñêîé èäåíòèôèêàöèè ìî-
æåò ðåøàòüñÿ îïòèìèçàöèîííàÿ çàäà÷à ïîèñêà ìèíèìóìà èí-
ðåòè÷åñêèì: ìåòîäè÷åñêèå ðåêîìåíäàöèè. – Íîâîñèáèðñê:
ôîðìàöèîííîé íåîïðåäåëåííîñòè â ïðîñòðàíñòâå ïàðàìåòðîâ
Èçä-âî ÍÃÒÓ, 2002. – 153 ñ. ìàòåìàòè÷åñêîé ìîäåëè. Êðèòåðèé èíôîðìàöèîííîé íåîïðå-
äåëåííîñòè îáåñïå÷èâàåò ïðîâåðêó ñòàòèñòè÷åñêîé ãèïîòåçû
Received 20.05.2019 îá àäåêâàòíîñòè êîíêðåòíîé ìîäåëè. Åñëè çíà÷åíèå êðèòåðèÿ
ïðåâûøàåò íåêîòîðîå êðèòè÷åñêîå çíà÷åíèå, ãèïîòåçà îá àäåê-
âàòíîñòè äîëæíà áûòü îòâåðãíóòà. Äëÿ ðàñ÷åòà êðèòè÷åñ-
ÂÈÊÎÐÈÑÒÀÍÍß ²ÍÔÎÐÌÀÖ²ÉÍί
êèõ çíà÷åíèé èíôîðìàöèîííîãî êðèòåðèÿ àäåêâàòíîñòè áûë
ÍÅÂÈÇÍÀ×ÅÍÎÑÒ²  ÇÀÄÀ×ÀÕ ²ÄÅÍÒÈÔ²ÊÀÖ²¯
ïðîâåäåí ñòàòèñòè÷åñêèé ýêñïåðèìåíò. Ñ èñïîëüçîâàíèåì ìå-
Ìàíêî Ã.²., ×èñòîêëåòîâ ª.Ï. òîäà Ìîíòå-Êàðëî áûëî èññëåäîâàíî ðàñïðåäåëåíèå âåðîÿòíî-
Âèâ÷àþòüñÿ ³íôîðìàö³éí³ ìåòîäè àíàë³çó òà óïðàâë³ííÿ ñòåé èíôîðìàöèîííîãî êðèòåðèÿ. Áûëà ïîñòðîåíà äîñòàòî÷-
ñèñòåìàìè â óìîâàõ íåâèçíà÷åíîñò³. Îá´ðóíòîâàíî äîö³ëüí³ñòü íî ãëàäêàÿ ýìïèðè÷åñêàÿ ôóíêöèÿ ðàñïðåäåëåíèÿ êðèòåðèÿ.
âèêîðèñòàííÿ ³íôîðìàö³éíî¿ íåâèçíà÷åíîñò³ â çàäà÷àõ ³äåíòè- Ðàñïðåäåëåíèå èíôîðìàöèîííîãî êðèòåðèÿ èìååò âûðàæåííóþ
ô³êàö³¿ îá’ºêò³â óïðàâë³ííÿ ³ ñèíòåçó ñèñòåì ðåãóëþâàííÿ. Äëÿ àñèììåòðèþ è íåáîëüøîé ïîëîæèòåëüíûé ýêñöåññ. Âûÿâëåíî,
÷èñëîâîãî îö³íþâàííÿ ³íôîðìàö³éíî¿ íåâèçíà÷åíîñò³ âèêîðèñ- ÷òî ýòî ðàñïðåäåëåíèå íàèëó÷øèì îáðàçîì àïïðîêñèìèðóåòñÿ
òîâóºòüñÿ ê³ëüê³ñòü äåç³íôîðìàö³¿ ÿê íåãàòèâíî¿ êîðèñíî¿ ³íôîð- çàêîíîì Generalized Extreme Value Distribution. Êðèòè÷åñêîå
ìàö³¿ Áîíãàðäà. Òàêà ³íôîðìàö³éíà íåâèçíà÷åí³ñòü ìîæå ñëó- çíà÷åíèå ìîæåò áûòü îïðåäåëåíî êàê êâàíòèëü óðîâíÿ 0,01
ãóâàòè ÿê êðèòåð³é àäåêâàòíîñò³ ìàòåìàòè÷íî¿ ìîäåë³ èëè 0,05 ýòîãî ðàñïðåäåëåíèÿ.
îá’ºêòà óïðàâë³ííÿ. ²íôîðìàö³éíà îö³íþâàííÿ òî÷íîñò³ ìîäå-
ëþâàííÿ ìîæå áóòè çàñòîñîâàíà ïðè áóäü-ÿêîìó ìåòîä³ ³äåí- Êëþ÷åâûå ñëîâà: àäåêâàòíîñòü, êîìïüþòåðíîå
òèô³êàö³¿ ³ äîçâîëÿº ïîð³âíþâàòè òî÷í³ñòü ìîäåëåé, ùî â³äð³çíÿ- ìîäåëèðîâàíèå, çàêîí ðàñïðåäåëåíèÿ, èäåíòèôèêàöèÿ,
þòüñÿ îäíà â³ä îäíî¿ ìåòîäîì îòðèìàííÿ (àíàë³òè÷íèì àáî èíôîðìàöèîííûé êðèòå ðèé, ìåòîä Ìîíòå -Êàðë î,
åêñïåðèìåíòàëüíèì), õàðàêòåðîì (äåòåðì³íîâàí³ àáî ñòîõàñ- êîíòðîëèðóåìûé îáúåêò, ñòàòèñòè÷åñêàÿ ãèïîòåçà,
òè÷í³ ìîäåë³), êîíêðåòíîþ ðåàë³çàö³ºþ (ô³çè÷í³ àáî ìàòåìà- íåîïðåäåëåííîñòü.
òè÷í³ ìîäåë³).  õîä³ ïàðàìåòðè÷íî¿ ³äåíòèô³êàö³¿ ìîæå âèð³-
øóâàòèñÿ îïòèì³çàö³éíà çàäà÷à ïîøóêó ì³í³ìóìó ³íôîðìàö³é-
íî¿ íåâèçíà÷åíîñò³ â ïðîñòîð³ ïàðàìåòð³â ìàòåìàòè÷íî¿ ìî-
äåë³. Êðèòåð³é ³íôîðìàö³éíî¿ íåâèçíà÷åíîñò³ çàáåçïå÷óº ïåðå-
â³ðêó ñòàòèñòè÷íî¿ ã³ïîòåçè ïðî àäåêâàòí³ñòü êîíêðåòíî¿
ìîäåë³. ßêùî çíà÷åííÿ êðèòåð³þ ïåðåâèùóº äåÿêå êðèòè÷íå
çíà÷åííÿ, ã³ïîòåçà ïðî àäåêâàòí³ñòü ïîâèííà áóòè â³äêèíóòà.
Äëÿ ðîçðàõóíêó êðèòè÷íèõ çíà÷åíü ³íôîðìàö³éíîãî êðèòåð³þ
àäåêâàòíîñò³ áóâ çä³éñíåíèé ñòàòèñòè÷íèé åêñïåðèìåíò. Ç
âèêîðèñòàííÿì ìåòîäó Ìîíòå-Êàðëî áóëî äîñë³äæåíî ðîçïîä³ë
éìîâ³ðíîñòåé ³íôîðìàö³éíîãî êðèòåð³þ. Áóëà ïîáóäîâàíà äî-
ñèòü ãëàäêà åìï³ðè÷íà ôóíêö³ÿ ðîçïîä³ëó êðèòåð³þ. Ðîçïîä³ë
³íôîðìàö³éíîãî êðèòåð³þ ìຠâèðàæåíó àñèìåòð³þ ³ íåâåëè-
êèé ïîçèòèâíèé åêñöåñ. Âèÿâëåíî, ùî öåé ðîçïîä³ë íàéêðàùèì
÷èíîì àïðîêñèìóºòüñÿ çàêîíîì Generalized Extreme Value
Distribution. Êðèòè÷íå çíà÷åííÿ ìîæå áóòè âèçíà÷åíî ÿê êâàí-
òèëü ð³âíÿ 0,01 àáî 0,05 öüîãî ðîçïîä³ëó.
Êëþ÷îâ³ ñëîâà: àäåêâàòí³ñòü, êîìï’þòåðíå
ìîäåëþâàííÿ, çàêîí ðîçïîä³ëó, ³äåíòèô³êàö³ÿ,
³íôîðìàö³éíèé êðèòåð³é, ìå òîä Ìîíòå-Êàðë î,
Keywords: adequacy, computer modeling, distribution law, 10. Bongard M.M. Problemy uznavaniya [Problems of Rec-
identification, information criterion, Monte Carlo method, con- ognition]. Moscow: Science, 1967, 320 p. (in Russian).
trol object, statistical hypothesis, uncertainty.
11. Manko G.I., Dovgopolaya A.A. Ob ispol’zovanii infor-
REFERENCES macionnogo kriteriya proverki gipotez o zakone raspredeleniya oshi-
bok izmereniya [About using the information criteria for testing
1. Thacker B.H., Doebling S.W., Hemez F.M. Concepts
the hypotheses of measurement error distribution law]. Ques-
of Model Verification and Validation. Los Alamos: Los Alamos
tions of chemistry and chemical technology, 2013, no. 1, pp. 181-
National Lab., 2004. 41 p.
184. (in Russian).
2. Sargent R.G. Verification and validation of simulation 12. R 50.1.037-2002. Rekomendacii po standartizacii [Rec-
models. WSC: Proceedings of the Winter Simulation Confer- ommendations for standardization]. Pravila proverki soglasiya
ence. Syracuse, NY 13244, U.S.A., 2011, pp. 183-198. opytnogo raspredeleniya s teoreticheskim: metodicheskie rekomen-
dacii [Rules for checking the accordance of an experimental and
3. Strizhnev A.G., Markov A.V., Rusakovich A.N. Identi- theoretical distribution: methodological guidelines] Novosibirsk:
fikaciya ob#ekta upravleniya po perekhodnoj harakteristike zamk- NGTU publ., 2002, 153 p. (in Russian).
nutoj sistemy [Ñontrol object identification by a closed system
transition characteristic]. Doklady BGUIR [Reports of BSUIRE],
2012, no. 5(67), pp. 65-72. (in Russian).