=Paper=
{{Paper
|id=Vol-2763/CPT2020_paper_p-2
|storemode=property
|title=Application of bionic models for situation management
|pdfUrl=https://ceur-ws.org/Vol-2763/CPT2020_paper_p-2.pdf
|volume=Vol-2763
|authors=Olga Gerget,Nataliia Markova
}}
==Application of bionic models for situation management==
https://ceur-ws.org/Vol-2763/CPT2020_paper_p-2.pdf
Application of bionic models for situation management
O.M. Gerget1,2, N.A. Markova1
gerget@tpu.ru, markovana@tpu.ru
1
Tomsk Polytechnic University, Tomsk, Russia;
2
Siberian State Medical University, Tomsk, Russia
The article discusses the concept of choosing the sequence of control actions in order to minimize the possibility of the system state
transition to an adverse one. For this purpose, the bionic model based on the synthesis of information approach, neural networks and
a genetic algorithm is developed. The functionality of each of the model elements and their interaction are presented in this paper.
Special attention is paid to neuroevolutionary interaction. At the same time, information about control actions is encapsulated in the
gene, which allowed increasing the functionality of the algorithm due to multidimensional data representation. The article describes
the principle of data representation in bionic models, which differs from the existing ones by the possibility of explicit or implicit
representation of the control action in the chromosome. In the explicit representation one neural network is formed, it describes the
effect of any of the control actions involved in the training. An implicit view creates a set of models, each of which describes the effect
of only one control action. A brief description of the software implemented in the Python programming language is provided.
Keywords: information approach, neural networks, genetic algorithm, bionic model, choice of control actions.
1. Introduction 2. Information method for calculating the
generalized indicator
The development of information technologies for
monitoring, prediction and situation modeling based on The application of the information approach in the
the application of bionic principles is one of the most structure of the bionic model makes it possible to
promising scientific directions [1-5]. Evolving over evaluate the state of the object of the study at the
millions of years in biosystems, structures have been considered moments of time and to reveal deviations
formed, in particular, genetic, immune, neural, providing from the stationary state. Of interest is the formation of a
balanced development and availability of necessary single quantitative indicator [7, 8], which would allow to
information means of control and adaptive control in a assess the reaction of the object of the study to changes
changing environment. Attempts are currently being taking place under the influence of the internal and
made to integrate artificial information processing external environment.
systems that structurally reflect the functioning of On the basis of implementation and analysis of
dynamic systems. Particular attention is paid to the methods for construction of generalized estimates in the
development of models and methods that information system, the approach [9] which considers the
comprehensively take into account the specifics of the information measure as a measure of preference of the
object of the study. In the article we will focus on the behavior of the bioobject is used.
concept of choosing the sequence of control actions in The choice is justified by the analysis of the results
order to prevent the transition of the dynamic system to obtained by:
an adverse state. For this purpose, we will consider the 1. The integral criterion based on the evaluation of
bionic model proposed by the authors [6]. similarity measure of observed and preferred state
By the bionic model we will mean the mathematical areas in feature space, where similarity measure is
model, as well as its software implementation, built on normalized in Mahalanobis metric by intramultiple
the principle of functioning and organization of distance of reference (steady) state area [10].
biosystems. 2. The integral criterion, in the basis of which Kulbak
In order to implement situation management, namely, information measure is considered as a measure of
to choose the best solution, evaluate and predict the effect bioobject behavior preference [9, 11].
of the application of control actions it is necessary to 3. The entropy method for detecting body response to
have an adequate mathematical model. This role is impact [12].
performed by the bionic model of choosing the sequence
of control actions of the following type: 3. Prediction of the state of the research object
< ππππ, πΊπΊπΊπΊ, πΌπΌ, π΄π΄ > based on bionic models
where ππππ - neural network models, πΊπΊπΊπΊ - genetic Neural networks have been chosen as the basic
algorithms, πΌπΌ - information method for calculating the technology in the bionic model structure in order to
generalized indicator of a biosystem, π΄π΄ - model setting predict the values of features characterizing the state of
algorithms. the system in dynamics, as well as when choosing a
Synthesis of information approaches, neural networks particular control action. This is due to the high
and genetic algorithms in bionic models allows systems efficiency of prediction; the possibility of implementing a
to exchange information and transfer the values of their model ensemble and using recursive neural networks for
characteristics as input to another subsystem which a variable structure vector [13, 14].
improves the quality of functioning and interpretability. Let us consider an object whose state is characterized
by informative features xi β¦ and the ways to influence
Copyright Β© 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY
4.0)
the state of the object U = (u1,...,ul), where l is the object and generalized indicators obtained on the basis of
number of control actions. the information method.
It is believed that the value of the i-th feature at a The implementation of the prediction algorithms is
point of time t +1 is determined according to the formula: based on recurrent neural networks [4, 15, 16], through
π₯π₯ππ (π‘π‘ + 1) = ππ(π₯π₯ππ (π‘π‘), π₯π₯ππ (π‘π‘ β 1), β¦ , π₯π₯ππ (π‘π‘ β ππ), β¦ , π’π’ππ ). which it is possible to obtain an accurate identification of
The problem comes down to solving the task of behavior. The main advantage of using this approach is
predicting the time series. For this purpose, n neural the implementation of the "sequence-to-sequence
networks are formed, each of which describes the effect learning" idea, namely the input and output vectors are
of only one control action. Each of the neural networks not limited in size. You can put a vector of any length at
takes only the vector of indicators of the study object as the input and get the corresponding vector at the output.
an input, and the information about the type of the It should be noted that by analogy with the concept of
simulated impact is presented in all weights of the model. time representation in dynamic neural networks a new
The output of neural networks is the predicted values of principle of data representation in bionic models in the
both variables characterizing the state of the research form of explicit (Fig. 1) and implicit representation of
control actions (Fig. 2) is proposed in this case.
Fig. 1. Machine learning with explicit representation of control actions
Fig. 2. Machine learning with implicit representation of control actions
When using an approach with explicit representation 4. Evolutionary approach in bionic model
of control actions, the result of machine learning is the structure
only model presented (singleton) describing the effect of
The application of the evolutionary approach in the
any of the control actions involved in learning. Such a
structure of the bionic model has made it possible both to
model takes as input not only the values of the indicators
search for the most effective control actions and to
describing the object of the study, but also the logical
optimize hyperparameters of the models. In this case,
variables in which the control action is encoded (0 - no
control actions are presented in two forms:
impact; 1 - impact is performed).
- in case of explicit representation of control actions,
In implicit representation, the result of machine
the chromosome is constructed from genes, each of
learning is a set of models, each describing the effect of
which determines the set of control actions applied to
only one control action. Each of the models accepts only
the object of study at some point of time;
the vector of indicators of the study object, and the
- in case of implicit representation, the chromosome is
information about the type of simulated impact is
determined by genes characterizing certain
implicitly presented (i.e. distributed) in all weights of the
properties of the model (number of layers, neurons
model.
per a layer, short-term memory capacity, degree of
The need to present information in two types is
network connectivity, type of activation function,
justified by the fact that the explicit representation allows
type of dynamic network deployment, etc.).
to regulate the "intensity" of the impact, but requires
Fig. 3 shows a diagram for the general case where
careful presetting of weights and regularization
there are several dependent control actions on the test
coefficients, which would determine the informativity of
object at each time point. In this case, each chromosome
the input neurons that take the values describing the
involved in the formation of the population is a matrix
control action to the input.
(M)ij, where i is the type of action; j is the time point.
When learning models with implicit representation of
Two ways of presenting priori data were reflected
control actions, there is no problem of loss of
both in the construction of the chromosome and in the
informativity of input neurons values. Each model
defining of criterion function.
contains fewer configurable parameters. However, the
With explicit representation, a gene is a vector of
type of impact and its βintensityβ remain unchanged after
logical variables, each of which determines the presence
learning, and a large number of models result in huge
or absence of the control action. When the function
investment of time, compared to a model using explicit
fitting is calculated, each gene is concatenated with a
representation.
vector of values of variables describing the state of the
The approaches described above are applicable to a
study object and it is input to some model. The output
wide class of models, from linear regression to dynamic
value of the model by means of a generalized assessment
and deep neural networks. Optimization of deep neural
reflects the effectiveness of situation management.
network parameters is a computationally more
complicated task as "layered" architectures contain
combinations of nonlinearities defining a wide range of
learning problems described by the general term βdeep
learningβ.
Fig. 3. Chromosome type determined by the way control actions are presented in the model. Note: U = (u(1,) u2,...ul) is a vector of
control actions; W = (w1,w2,...,wn) is a vector of neural model parameters; T = (t1,t2,...,tΟ) are the time points
In the implicit setting of control actions, each gene is developed which implemented the following procedure
a vector of neural network model parameters describing a of actions:
certain action. Thus, some "bank" of learnt models is 1. Loading of the table with integral criteria on tic
formed, vectors of weights remain unchanged in the marks;
course of the evolutionary process, but they participate in 2. Data separation into training and test samples
chromosome formation. according to cross-validation conditions;
Each gene, since it is a fully functional model, takes a 3. Neural network training in order to obtain a
vector of variables describing the object of the study as predictive generalized assessment of the state of the study
input, and its output is the predicted value of the object at the next point of time;
generalized (integral) assessment of the dynamic system 4. Calculation of a prediction error;
state depending on the control action. 5. Optimal selection of neural network
hyperparameters using a genetic algorithm.
5. Methodological basis for choosing the The results of the prototype work made it possible to
sequence of control actions during situation assess the accuracy of this approach and to conclude on
management the need to develop a software complex based on the
We summarize and note the functionality of each chosen algorithms.
element of the bionic model of choosing the sequence of The prototype was developed in Matlab R2008b.
control actions during situation management. Python, a high-level programming language relating
Let us have a database that stores the values of the to freely distributed software, has been chosen as the
properties of the study objects at the following moments main development tool.
of time: prior and after some control action is applied. The program on Python consists of a main module
1. For each object at each moment of time, we and secondary modules that can be connected when
calculate the values of the generalized indicator and running the main module code. Secondary modules are:
evaluate their deviations from the norm. β’ NumPy - data interpolation, statistical functions,
2. The control action on each object is modeled by a optimized calculations;
neural network. As input data, values of variables β’ Pandas - dataframe, loading and saving *.csv-files;
characterizing the state of the research object before β’ PyBrain - artificial neural networks, optimization
applying the control action are used. As the desired methods, including genetic algorithm;
responses, both the variable values and the generalized β’ PyQt β components of the graphic interface.
indicator (I) values are used after the control action is The selected tools allow you to move from the stage
applied. The objective function to be minimized is of prototype development to the implementation of the
defined as the deviation of the generalized indicator from finished software project quite quickly.
the norm. The whole process will be divided into several stages:
3. The obtained set of neural network models forms a collection of primary data and secondary data obtained
plurality of genes that can be specified by a vector of from the results of additional studies. After the
neural network parameters (for an implicit form of calculation of integral indicators based on the
control action) and a vector of logical variables in which information approach, a graph of changes in the state of
the control action is encoded (in an explicit form). the research object in time is displayed. The cubic spline
4. The sequence of control actions is determined by interpolates the data and calculates the rate of change of
known previous states (state variables) and predicted the variables characterizing the state of the research
values of object state variables using genetic algorithm. object.
In the process of genetic algorithm work the request to Further, the calculation of predicted values of each of
neural networks for obtaining the predicted values of the the indicators is performed separately, and a predicted
object state for all future time intervals is performed. integrated assessment of the state during the situation
5. The genetic algorithm forms and chooses new control period is given. The deviations between the
generation by means of selection, crossover and prediction and the stationary (normal) values at each time
mutation. point are determined and displayed on the graph. A
6. The values of the fitting function of the genetic sequence of control actions at which deviation is minimal
algorithm are determined as deviations (from the norm) is formed.
of the value of the generalized indicator predicted by the
7. Results
neural network (by a gene). Among the set of control
actions, defined by chromosomes, the one at which the Development and testing of the bionic model, which
value of the fitting function will be minimal is chosen is based on the interaction of the information method,
(the range is set [0,1; 1]). genetic algorithm and neural networks, allow for
situation control based on correct choice of the sequence
6. Software implementation of control actions. The information method forms a
At the first stage of development of information single integral indicator characterizing the state of the
system in order to test the possibility of using artificial object of the study. Artificial neural networks enable to
neural networks for predicting the state of dynamic obtain prediction of integral assessment of the object
systems (the object of the study) a prototype was state depending on the chosen control actions. The
genetic algorithm chooses a sequence of control actions [8] Alifirova V.M., Brasovsky K.S., Zhukova I.A.,
that reduce the possibility of transition to adverse states. Pecker J.S., Tolmachev I.V., Fokin V.A. Method of
The principle of data representation in bionic models quantitative evaluation of results of postural tests/
presented in the article differs from the existing ones by Bulletin of Experimental Biology and Medicine,
the possibility of explicit or implicit representation of 2016. β V. 161. β β 3. β pp. 421-424.
control action in the chromosome. In the explicit [9] Fokin V.A., Pecker Y.S., Berestneva O.G., Gerget
representation one neural network is formed, it describes O.M. Integrated methods of assessment of the state
the effect of any of the control actions involved in the of complex systems/ Izvestia of Tomsk Polytechnic
training. Information about the type of control action is University, 2012. β V. 321. β β 5. β pp.120-124.
supplied to the input of neural network. In the implicit [10] Fokin V.A. Statistical data modeling at biological
representation a set of models is created, each of which systems state estimation // Izvestia of Tomsk
describes the effect of only one control action, Polytechnic University. Engineering of georesources,
information about the type of simulated impact is 2014. β V. 307. β β5. β pp. 136-138.
distributed in all weights of the model. With a large [11] Devjatykh D.V., Gerget O.M., Mikhalenko I.V. Use
number of control actions, training models with implicit of artificial neural networks to predict the
representation of the actions is time consuming, in development of perinatal damage of the nervous
comparison with a model using explicit representation. system / Izvestia of Volgograd State Technical
However, when the information is encapsulated in a University, 2013. β V. 16. β β 8 (111). β pp. 77-80.
gene, the possibility of n-dimensional representation of [12] Wilson, A.J. Entropy modeling methods for
data appears, which extends the functionality of the Complex Systems. βΠ.: Science, 1978. β 248p.
algorithm. [13] Gerget O.M., Devjatykh D.V. Bionic Model for
Identification of Biological Systems/ Information
Acknowledgments and Mathematical Technologies in Science and
The work was supported by RFBR, Grant β 19-07- Management. 2017. β β 2 (6). β p. 21-29.
00351. [14] Hastie T., Tibshirani R., Friedman J. The Elements
of Statistical Learning: Data Mining, Inference, and
References Prediction, 2009. β 745p.
[15] Goodfellow I., Bengio Y., Courville A. Deep
[1] Pyatkova N. I., Massel L. V., Massel A. G. Situation Learning. Book in preparation for MIT Press.
management methods in research of energy security [16] Wang, H. and Song, G. Innovative NARX recurrent
problems / Izvestia akademii nauk. Energetika, 2016. neural network model for ultra-thin shape memory
β β4. β pp. 156-163. alloy wire. Neurocomputin, 2014 β 134. β pp.289-
[2] Shcherbakov M., Kamaev V., Shcherbakova N. 295.
Automated electric energy consumption prediction
system based on decision tree approach / 7th IFAC About the authors
Conference on Manufacturing Modelling,
Gerget Olga M., professor of the Department Β«Information
Management, and Control, 2013. β V. 46. β Issue 9.
TechnologyΒ», School of Computer Science & Robotics, doctor
β pp. 1027-1032. of Engineering Sciences, National Research Tomsk Polytechnic
[3] Yusupov R. M., Ronzhin A.L., From smart devices University, Tomsk, Russia. E-mail: gerget@tpu.ru
to intelligent space / Bulletin of the Russian Markova Nataliia A., Senior Lecture, School of Core
Academy of Sciences, 2010. β 80:1. β pp. 45β51. Engineering Education, National Research Tomsk Polytechnic
[4] Massel L.V. Problems of the semiotic type University, Tomsk, Russia. E-mail: markovana@tpu.ru
intellectual systems creation for the strategic action
control in critical infrastructures / Information and
mathematical technologies in science and
management. Scientific journal. Irkutsk: MESI SB
RAS, 2016 β β1. β pp. 7-27
[5] Yurchenko M. A., Kochegurova E. A., Fadeev A. S.,
Piletskaya A.Y. Calculation of performance
indicators for passenger transport based on telemetry
information / Engineering Technology, Engineering
Education and Engineering Management :
proceedings of the International Conference on
Engineering Technologies, Engineering Education
and Engineering Management, 2015. β pp. 847-851.
[6] Gerget O.M. Bionic models for identification of
biological systems // Journal of Physics: Conference
Series. β 2017. β V. 803. β β 1. β pp. 12- 24.
[7] Maksimov I.B., Kolyar V.P., Bogomolov A.V.
Applied theory of information support in medical
and biological research. - Π. Binom. - 2013. β 312 p.