=Paper=
{{Paper
|id=Vol-2098/paper15
|storemode=property
|title=Neural Network and Agent Technologies in the
Structural-Parametric Modeling of Technological
Systems
|pdfUrl=https://ceur-ws.org/Vol-2098/paper15.pdf
|volume=Vol-2098
|authors=Yuri A. Ivashkin,Ivan G. Blagoveschensky,Marina A. Nikitina
}}
==Neural Network and Agent Technologies in the
Structural-Parametric Modeling of Technological
Systems==
https://ceur-ws.org/Vol-2098/paper15.pdf
Neural Network and Agent Technologies in the
Structural-Parametric Modeling of Technological
Systems
Yuri A. Ivashkin1 , Ivan G. Blagoveschensky2 , and Marina A. Nikitina3
1
Moscow Tecnical University Communication and Informatics, Moscow, Russia,
ivashkin@nextmail.ru
2
Bauman Moscow State Technical University, Moscow, Russia,
igblagov@gmail.com
3
V.M. Gorbatov Federal Research Center for Food Systems of Russian Academy of
Sciences, Moscow, Russia
nikitinama@yandex.ru
Abstract. It is offer information technology of identification and fore-
casting of a complex technological system based on structural and para-
metric modeling in combination with neural network and agent technolo-
gies. The function of the neural network module or intelligent agent is to
refine the initially specified coupling coefficients between the monitored
state and target parameters and to recognize abnormal situations in the
system in order to make optimal decisions. The task of recognizing situa-
tions consisted in classifying them based on real-time presentation of the
current states of the system by belonging to the areas of decision-making.
It is offer variant of the architecture of the Hamming neural network
with a multilayer recurrent structure, as a specialized heteroassociative
memory device with pairs of interconnected input and output vectors.
Proposed information technology used in problems of identifying the
anomalous state of technological systems of food production and making
optimal decisions in the management of the quality of products of agro-
processing enterprises.
Keywords: Technological system · Structural-parametric analysis · Sit-
uational analysis · Information technologies · Neural networks · Multi-
agent modeling
1 Introduction
Information technologies for the structurally-parametric and situational analy-
sis of complex chemical-technological and biotechnological systems based on the
Copyright c by the paper’s authors. Copying permitted for private and academic purposes.
In: S. Belim et al. (eds.): OPTA-SCL 2018, Omsk, Russia, published at http://ceur-ws.org
170 Yu. A. Ivashkin et al.
processing of statistical data on a managed object, in conditions of uncertainty
and risk, require the development and inclusion of intelligent modules for rec-
ognizing complex situations for computer support for the adoption of optimal
adequate solutions [1].
The existing direction of the structural-parametric and situational analysis of
the state of the technological system [2] is related to their structuring according
to the functional principle and the description of the functional relationships
between the state and target parameters in the matrix form.
The structural-parametric model (SPM) of the system is represented in the
form of a cellular matrix with blocks of indicators placed along the main diagonal
and off-diagonal blocks, communication operators between the parameters and
their functional groups. The absence of connections is described by the zero-
operators k0k, which determine the non-working domain of interaction.
Initially, the characteristics of the links are determined expertly path and
refined in the presence of statistical data with the determination of the corre-
lation coefficients and linear multiple regression Pij of the current deviations
∆x1 , ..., ∆xn of the system state variables xi from the given norms x0j , depending
on the deviation of the factors of the controlled set ∆xj , j = 1, n j 6= i
mi
X
∆xi = Pij ∆xj , i = 1, n; j = 1, n; j 6= i (1)
j=1
Then follows the transition to the cognitive matrix kCij kn of relative compa-
rable characteristics of the relationships between the different physical quantities
xi and xj according to the formula:
∆x0j
Cij = Pij ; i, j = 1, n; j 6= i (2)
∆x0i
where ∆x0i , ∆x0j - are the admissible deviations of the variables.
The development of information technology for identifying and forecasting
the state of a complex technological system in real time is associated with a
rational combination of applied mathematical statistics with the analysis of fuzzy
data and self-learning based on methods of artificial intelligence, neural network
and agent technologies.
2 Situational Model of Technological System
On the basis of SPM, a situational matrix model of the system Cij ∆xj , i, j = 1, n
by multiplying kCij kn by the diagonal matrix of the normalized deviation vector
∆x1 , ... , ∆xn :
1 c12 . . . c1n ∆x1 ∆x1 c12 ∆x2 . . . c1n ∆xn
c21 c22 . . . c2n ∆x2 c21 ∆x1 ∆x2 . . . c2n ∆xn
· = (3)
... ... ... ... ... ... ... ... ... ... ... ...
cn1 cn2 . . . 1 ∆xn cn1 ∆x1 cn2 ∆x2 . . . ∆xn
Neural Network and Agent Technologies 171
0
i −xi |
where ∆xi = |x∆x i
0 ; i = 1, n - normalized deviations of the state parameters
from the range of permissible deviations ∆x0i .
As a result, the elements of the main diagonal of the situational matrix reflect
the current normalized deviations of the xi controlled factors from the given
values xi 0 , and the off-diagonal Cij · ∆xj ; i, j = 1, n; (i 6= j) - the contributions
of the deviations ∆xj ; j = 1, n, to the deviation ∆xi ; i = 1, n in accordance
with the system of equations
N
X
∆xi = Cij · ∆xj ; i, j = 1, n; i 6= j (4)
j
with ordering by rows of all a priori known causes of the deviation of ∆xi , and
by columns - of the possible investigative effects of the deviation of ∆xi on other
parameters.
In the general case, the situational matrix kcij · ∆xj kn with a multitude
of functional elements {x1 ...xn } and the links between them kcij kn describes
a structurally complex situation of cause-effect interaction of elements in the
current state of the system, by combining an a priori knowledge base on the
structure of links to current information ∆x.
A formalized algorithm for identifying an abnormal situation in a technolog-
ical system is as follows.
In the line of the maximum, diagonal element corresponding to the maxi-
mum deviation from the norm ∆xi 0 in the observed set of state parameters, the
maximum nondiagonal element corresponding to the main cause that caused
this deviation. Then, on the found column, need to go to the new element of the
main diagonal, after which in the new line founds the main cause of the anomaly
on this cause-and-effect step. The search continues until a diagonal deviation are
founds, in the line of which all nondiagonal elements will be zero, which means
finding the original cause of the anomalous situation.
The registration of current situations in real time complements the original
database with the subsequent recalculation of regression coefficients.
However, SPM in the mode of passive observation and accumulation does not
always ensure the necessary speed and accuracy of decision making in problems
of identification and forecasting due to the inadequacy of statistics and the in-
adequacy of regression bounds. The methods, uses for the passive accumulation
of data and active experiment in real time for an operating technological system
are practically not realizable, because require a long period of observation or
an active experiment with a sufficient number of repetitions and verification of
reproducibility.
In this case, the IT technology of situational modeling of technological sys-
tems in real time under conditions of uncertainty and fuzzy data requires further
intellectualization based on neural network and agent technologies.
172 Yu. A. Ivashkin et al.
3 Neural Network Situation Model
The intellectual function of a neural network module or self-learning agent is in
refine and correct the originally defined coupling coefficients between the states
and target parameters, as well as in recognize and classify anomaly situations in
the system as belonging to decision-making classes based on present real-time
conditions [2].
For situation analysis in conditions of fuzzy and inadequate information, a
variant of architecture of the artificial neural network (ANN) Hamming with
a multilayer recurrent structure is proposed as a specialized heteroassociative
memory device with a predefined training sample of reference situations and
associated pairs of input and output layer vectors (figure 1).
The inputs of the network receive values, n components of the current situa-
tion vector x1 , ..., xn and the network problem is to find the minimum hemming
distance between the input vector and the reference vectors training samples,
coded in the network structure.
First ANN layer (neurons 1-3) with unidirect propagation of its output signals
to the neurons of the output layer (11-13) has fixed values of weights correspond-
ing to the components of the vector of the observed situation (image) so that
(1) (i)
wij = xj for i = 1, p (p is the number of neurons of the first layer).
Fig. 1. The Architecture of the ANN of Hamming
Similarly, the weights of the output layer (neurons 11-13) correspond to the
next vectors of reference situations y (i) , related to x(i) :
(2) (i)
wij = yj (5)
The hidden layer, MAXNET, consists of neurons with feedbacks on the prin-
ciple of “everyone with each”. In this case, with a proper output, the neuron
is connected by a positive (exciting) feedback with a weight equal to +1, and
Neural Network and Agent Technologies 173
with other neurons - negative (overwhelming) feedback with a weight inversely
proportional to the number of neurons p.
Neurons of the MAXNET layer (1-3) function in WTA (Winner Takes All)
mode so that the network weights should amplify the neuron’s own signal and
m 1 (m)
weaken the others. To achieve this effect - wii = 1, and − p−1 < wij < 0 for
i 6= j.
m
To ensure absolute convergence of the weight algorithm wii should differ
from each other:
(m) 1
wij = − +ξ (6)
p−1
where ξ - random variable with a sufficiently small amplitude.
Neurons of different layers of ANN are function differently. Neurons of the
first layer calculate the Hamming distances between fed on input N - dimensional
vectors x and the vectors of the weights w(i) = x(i) individual neurons of this
layer (i = 1, 2, ..., p), applied to the input of the network. The values of the output
signals of these neurons are determined by the formula:
dh (xi , x)
ybi = 1 − (7)
N
where dH (xi , x) denotes the Hamming distance between the input vectors x and
x(i) , i.e. the number of bits by which these two vectors differ.
The output signals ybi of the neurons of the first layer become the initial states
of the MAXNET layer neurons in the second phase of the network functioning.
The task of the neurons of this layer is to determine the winner, i.e. a neuron,
whose excitation level is closest to 1 by the recurrence formula:
X (n)
yi (k) = f (yi (k − 1) + wi yj (k − 1)) (8)
j6=i
at the initial value yj (0) = ybi .
Such a neuron points to an image vector with a minimum Hamming distance
to the input vector x.
The activation function f (y) of neurons in the MAXNET layer is given by
expression:
y for y ≥ 0
f (y) = (9)
0 for y < 0
The iterative process terminates at a time when the state of the neurons
is stabilized and the activity continues to manifest only one neuron, while the
rest are in the zero state. The active neuron becomes the winner and, through
(2)
the weights wij of the linear neurons of the output layer, represents the vector
y (i) , which corresponds to the vector x(i) , recognized by the MAXNET layer as
nearest to the input vector x.
In the process of network operation, we can distinguished three phases. In the
first of them an N -element vector x is fed to its input. After the presentation
of this vector, the signals that define the initial states of the neurons of the
174 Yu. A. Ivashkin et al.
second layer are generates at the outputs of the neurons of the hidden layer, i.e.
MAXNET.
In the second phase, the MAXNET-initiated signals are deletes, and the
iterative process (8) within this layer are starts from the initial state formed by
them. The iterative process terminates at a time when all the neurons, except
for the winner with an output signal equal to 1, goes to zero state. A neuron-
winner with a non-zero output signal becomes a representative of the data class
to which the input vector belongs.
In the third phase, the same neuron, by means of weights connecting it with
the neurons of the output layer, forms a response at the output of the network
in the form of a vector y, corresponding to the exciting vector x.
The data, receives from the real-time monitoring system and fed to the input
of the neural network should be normalized by delta coding, with a pixelated
calculation of the difference in the values of the object’s state parameter in the
current and previous control cycles, which can significantly reduce the dynamic
range of the data.
For n parameters, the difference frame are represents as a column vector x
dimensionality n:
X = (x1 , x2 , ..., xj , ..., xn )T
at which on the output of the ANN will form a column-vector Y , are formed as:
Y = (y1 , y2 , ..., yj , ..., yn )T
Trained network before the start of functioning with working technological
parameters need to check for the quality of training and the ability to generalize
the acquired knowledge. In this case, are found out, whether the results that the
network gives out at the outputs are within the permissible error when the sample
is fed into the network with predetermined values of the output parameters, but
different from the training sample.
In the test operation of the network in the food industry of the agro-industrial
complex (AIC), the minimum error in training was 1.04%, which corresponds
to an allowable error of 1.5%, agreed with the technologists responsible for the
quality of food products.
In a particular test implementation, the INS learning algorithm was reduced
to the following sequence of actions:
1. Formation of a matrix of reference samples X of the size k × n of the
Hamming network (Table 1):
Neural Network and Agent Technologies 175
Table 1. Matrix of reference samples X of the Hamming network
Input binary variables
Number image
1 2 ... i ... n
1 x11 x12 ... x1i ... x1n
2 x11 x12 ... x1i ... x1n
... ... ... ... ... ... ...
j xj1 xj2 ... xji ... xjn
... ... ... ... ... ... ...
p xp1 xp2 ... xpi ... xpn
with weight coefficients of neurons of the first layer wij = 0.5xij .
2. Determine the setting of the activation function of the linear threshold
function:
0, for s ≤ 0
f (s) = s, for 0 < s ≤ T (10)
T, for s ≥ T
where T = 0.5p so that the outputs of the neural network can take values within
[0, T ].
3. Entering the synapse values of feedbacks of neurons of the hidden layer in
the form of elements of a square matrix of size p × p:
1 for j = p
εjp = (11)
−ε for j 6= p
where ε ∈ [0, p1 ], or in the matrix form:
1 f or diagonal elements
E=
−ε exclude diagonal elements
Synapses of feedbacks of a Hamming neural network with negative weights
are inhibitory.
4. Setting an allowable difference of output vectors on two consecutive iter-
ations, Emax = 0.1, for estimating the stabilization of the solution found.
The neural network algorithm for classifying situations in the observable sys-
tem to support the making of managerial decisions in real time under conditions
of certainty reduces to the following.
An unknown binary vector are feds on the network inputs → −
x signals of the
current state of the system parameters:
1 if the parameter is within the norm
xij =
−1 if case of deviation f rom the norm
176 Yu. A. Ivashkin et al.
In the case of deviation of the state and output values of the neurons of the
first layer are calculated by the formula:
n
X
s1j = wji x∗i + T (12)
i=1
The activation linear-threshold function (10) uses to calculate the outputs
of the neurons of the first layer - →
−
y 1 . The outputs of the neurons of the second
layer are assigned the values of the outputs of the neurons of the first layer,
obtained at the previous step: →−y2 =→ −
(0)
y 1 , after which the first layer of neurons
is practically not involved.
For each q-th iteration in hidden layer calculates new values of states and
outputs of neurons by recurrent ratio:
n
(q+1) (q) (q)
X
s2j = y2j − ε y2i (13)
i=1,i6=j
The new output values →− (j+1)
y2 are determines using the linear threshold acti-
vation function for processing the corresponding states of the neurons - →− (j+1)
s2 .
This cycle repeats until the output vector stabilizes in accordance with the con-
dition:
k→
−
y j(q+1) − →
−
y j(q) k ≤ Emax (14)
In the ideal case, after stabilization, there is an output vector with a single
positive element with the remaining zero elements, the index of which indicates
the class of the unknown input image of the situation.
If the input image data is very noisy or there is no suitable standard in the
training sample, several positive outputs corresponding to the accuracy condition
(14) can be obtained as a result. In this case, it is concluded that the input image
can not be assigned to a certain class, but the positive output indices indicate
the most similar standards.
On an example of classification according to three reference situations (Table
2) in the technological system of confectionery production [8] the neural network
includes 9 input variables and 3 neurons in the first and second (output) layers
(figure 1).
In accordance with the learning stage algorithm, a matrix (3 × 9) is formed
to configure the Hamming neural network for 3 reference images with 9 inputs
(Table 2).
Neural Network and Agent Technologies 177
Table 2. Matrix of the Hamming network
Number input variable
Number image
1 2 3 4 5 6 7 8 9
1 1 −1 1 −1 1 −1 1 −1 1
2 −1 1 −1 1 1 1 −1 1 −1
3 1 1 1 1 −1 1 1 1 1
Based on the template of the reference images, calculates weighting coeffi-
cients of neurons of the first layer (Table 3).
Table 3. Matrix the weighting coefficients of neurons of the first layer
Number input variable
Number neuron
of first layer 1 2 3 4 5 6 7 8 9
1 0.5 −0.5 0.5 −0.5 0.5 −0.5 0.5 −0.5 0.5
2 −0.5 0.5 −0.5 0.5 0.5 0.5 −0.5 0.5 −0.5
3 0.5 0.5 0.5 0.5 −0.5 0.5 0.5 0.5 0.5
If use T = p2 , we determine the threshold of the activation function T = 1.5.
With restriction are ε ∈ (0, 13 ), the absolute value of the weight of each
inhibitory synapse is ε = 0.3 and Emax = 0.1 and the matrix of the inverse
synapse weighting coefficients (11).
1 for j = i
εjp =
−0.3 for j 6= i
→
−T
The test vector feds to the network inputs: X = [1, 1, −1, −1, 1 − 1, 1, 1, 1]
and the condition (14) determines the column vector of the states of the neurons
of the first layer, and at the output of the activation function of the state (10)
is the vector-column of the output values of the neurons of the first layer:
8.00 4.50
→
−
s1 = 2.00 →
−
y1 = 2.00
3.00 3.00
The ANN outputs assigned the corresponding output values of neurons of
the first layer. Then, using the ratio (13), a series of output vectors calculates
178 Yu. A. Ivashkin et al.
iteratively until the stabilization condition is satisfied. ANN signals obtained in
iteration cycle q when the test situation feds to its inputs, represents in Table 4.
Table 4. ANN signals
number State vector Vector of outputs
ky (q+1) − y (q) k
of iteration s21 (q) s22 (q) s23 (q) y21 (q) y22 (q) y23 (q)
1 8.00 2.00 3.00 4.50 2.00 3.00 −
2 3.00 -0.25 1.05 3.00 0.00 1.05 10.5
3 2.69 -1.22 0.15 2.69 0.00 0.15 0.91
4 2.64 -0.85 -0.66 2.64 0.00 0.00 0.02
As we can see from the table, the criterion for stopping the feedback loop
of the signal after feedback mades after the 4-th iteration. The positive output
value of the i-st neuron indicates that the input vector should be assigned to the
i-st class.
4 Agent-based Situational Modeling of Systems
The presented neural network technology for recognizing and classifying situa-
tions in real time are suggest to use in describing the dynamics of agent behav-
ior in complex situations. An intelligent agent is understood as [4] an imitation
model of an active element capable of performing the functions assigned to it
by a certain living or cybernetic organism, depending on the behavior of other
agents and environmental influences.
Self-learning, purposeful agents are able to accumulate knowledge based on
large amount of data and ontology of events in the process of interaction with
other agents and the environment, adapt to the situation, choose a strategy for
achieving the chosen goal and assess the degree of its achievement.
The general algorithm of the behavior of the intellectual agent [6] includes
the identification of the situation, the assessment of one’s own state and the
correction of the goal, followed by a reflexive reaction or intelligent (intelligent)
decision-making towards the goal. The criterion of the agent’s intelligence is the
degree of completeness and depth of a priori knowledge, learning strategies and
decision-making algorithms under conditions of uncertainty, risk and conflict.
The parametric description of an agent includes a set of goals and a knowledge
base in a specific area, a vector of characteristics of its state; bank of models
and strategies of behavior, description of external relations with agents and
the environment. Practical implementation of agent technologies is associated
with the development of simulation systems that provide an experimentation
environment, an agent-oriented language for describing models and software for
Neural Network and Agent Technologies 179
organizing experiments [5,6]. The methodology of agent modeling of the learning
agent reduces the following stages.
1. Parametric description of the external environment of the agent’s activity
with the formalization of a set of factors of influence on the functional state and
objective function of the agent in situational decision-making conditions.
2. Parametric description of functional blocks of the technological system in
the form of a set of vectors of input and output factors, state parameters and
objective function.
3. Description of the autonomous intelligent agent with a set of state vari-
ables, input and sensory variables that communicate with other agents and the
environment, as well as dynamics of agent behavior with procedures for learning
and identifying current situations and making decisions in the form of discrete-
event descriptions and decision-making strategies in conditions of sufficient, in-
complete and fuzzy information.
4. Creation of agent-oriented model of real-time management of the techno-
logical system, which includes, in accordance with the functional scheme of the
system:
- components describing the state and dynamics of agent behavior;
- organizational components that define the structure of interrelations be-
tween agents and functional blocks of the system;
- mobile components - to describe messages transmitted through a commu-
nication channel between agents and moving objects.
5. Software description of the components of the model of the system under
study in a high-level algorithmic language or agent-oriented modeling language
in a universal simulation system [5,6].
Agent technologies with neural network algorithms of behavior of learning
agents with recognition of current situations open up new possibilities of virtual
research of the influence of various technological factors on the abnormal states
of the system and the adoption of optimal solutions in the control system.
5 Conclusion
The proposed direction of intellectualization of situational modeling of systems
is the basis for constructing intellectual expert systems (IES) for making optimal
decisions and operative management of the quality of food products at all stages
of its production at processing enterprises of the agro-industrial complex [7,8].
The outlined approach to the development of IT technologies for the iden-
tification of multi-factor and weakly formalized technological systems based on
artificial intelligence and agent modeling opens new possibilities for computer
support for making optimal decisions in conditions of fuzzy information, uncer-
tainty and risk.
References
1. Trakhtengerts, E.A.: Computer Support for Decision-Making. SYNTHEG (1998)
180 Yu. A. Ivashkin et al.
2. Ivashkin, Yu. A.: Structural-parametric modeling and identification of abnormal
situations in complex technological systems. Problems of management 3, 39–43
(2004)
3. Savostin, S.D., Blagoveshchenskay, M.M., Blagoveshchensky, I.G.: Automation of
Control of Flour Quality Indicators in the Grinding Process Using Intelligent Tech-
nologies: The Monograph. Publishing house of Frantera (2016)
4. Ivashkin, Yu.A.: Multiagent Modeling in the Simplex 3 Simulation System: A text-
book for higher education. Binom, Laboratory of Knowledge (2016)
5. Schmidt, B.: The Art of Modeling and Simulation: The Textbook. SCS-Europe
BVBA, Chent (2001)
6. Karpov, Yu.G.: Simulation of Systems. Introduction to Modeling with AnyLogic 5:
The Textbook. BHV-S.-Petersburg (2005)
7. Ivashkin, Yu.A., Nikitina, M.A.: Logistics and structural optimization of material
flows in network distribution systems. In: International Scientific and Practical
Conference on Logistics and Economics of Resource Saving and Energy Saving in
Industry (ISPC “LEREP-11-2017”). pp. 240–244. Russian Federation, Tula (2017)
8. Blagoveshchensky, I.G., Blagoveshchenskaya, M.M., Apanasenko, S.I.: Creation of
virtual sensors based on the neural network for determining the main characteris-
tics of confectionery masses. Confectionery and bakery production 11 (154), 37–41
(2014)