Software & Systems

Articles of journal № 4 at 2022 year.

Order result by: Public date | Title | Authors

1. Algorithms for generating training sets in a system with case-based inference based on example situations [№4 за 2022 год]
Authors: Glukhikh I.N., Glukhikh D.I.
Visitors: 1243
The paper considers the issue of creating training sets and their scaling in machine learning problems. The subject of the study is the process of generating training sets based on examples in order to augment them. To implement the idea of expansion, it is proposed to use the transformation of existing examples of sit-uations. The transformation of examples is based on a well-known optimization method - the method of coordinate descent. The paper describes the statement of the problem of transformations of example situations in terms of the introduced representation model. There are proposed algorithms that make it possible to obtain an ex-tended set from the initial set of example situations specified using formal representations, which will include situations that meet the similarity criteria with these examples. The paper presents the testing of the proposed algorithms for expanding a set of example situations, car-ried out in order to form a data set for the studying artificial neural networks. The obtained results are of practical importance for training artificial neural networks used in intelligent decision support systems. The proposed algorithms make it possible to automate the formation of datasets using the available prepared and approved examples of typical situations and solving the transformation problem as the problem of finding the optimum of the similarity objective function.

2. A software platform demonstrator for configuring ANFIS neural network hyperparameters in fuzzy systems [№4 за 2022 год]
Authors: Ivanov V.K., Palyukh B.V.
Visitors: 1825
This article describes the research demonstrator for experimental verification and evaluation of fuzzy algo-rithms and neural networks in an expert system for complex multi-stage technological processes. The de-monstrator development purpose is to create a scientific and technical foundation for the ready-to-implement solutions transfer to the next project stages. The demonstrator allows assessing the readiness level of the components being developed, conducting re-search tests, checking the operability and efficiency of the software implementations functioning proposed at various parameter values and their combinations. A complex multi-stage technological process state di-agnostics involves the joint primary data processing to obtain probabilistic abnormal critical events or inci-dents characteristics under conditions of uncertainty. The authors propose a way of using a fuzzy neural network, which is trained with data generated by be-lief functions. The approach makes it possible to significantly speed up calculations and to minimize the re-source base. The article focuses on describing the neural network models and training datasets management, neural network training and quality control, the technological process diagnostics in various modes. The con-figurable hyper-parameters of the neural network are described in detail. There are examples of the diagnos-tic procedures implementation in various modes. It is shown that with the software diagnostic system func-tioning in conditions close to real, the initial assumptions concerning the time reduction for detecting and predicting incidents can be verified and experimentally substantiated. In addition, the technological chains sets that are the incidents causes can be more accurately determined.

3. A GraphQL dynamic schema in integrated information system implementation [№4 за 2022 год]
Authors: Chernysh, B.A., Murygin, A.V.
Visitors: 1922
The paper discusses the possibility of using the GraphQL toolkit with a dynamically changing data schema. The standard way to define data types and operations in GraphQL is a static schema. When using it, the en-tire data structure is determined in advance and cannot be changed dynamically while the application serv-ing requests is running. This circumstance does not allow using GraphQL in applications where the data struc-ture can change dynamically. To solve this problem, there is an approach that consists in storing the data schema in the application memory, and regenerating this schema in case of metadata changes. This paper presents a method for implementing this approach using the SciCMS software platform de-veloped by the authors as an example. A feature of the system is its focus on the requirements for working with data on technically complex products. These requirements include data storage in a tree view with op-timizing retrieval mechanisms, product version control, the ability to use different language representations of products, product lifecycle management, and advanced integration with multiple data sources. The paper outlines the techniques and technologies involved in building the system, provides UML dia-grams and diagrams of the main structures and processes of the application core. It also describes the im-plementation details of individual platform subsystems. Experimental data sampling was carried out in order to evaluate the efficiency of executing queries using joins. The most effective tools for optimizing the selection of hierarchical data were selected based on the data obtained. The paper presents the possibilities of the platform for its integration with other systems with-in a single information space.

4. DIY DDoS Protection: operational development and implementation of the service in the National Research Computer Network of Russia [№4 за 2022 год]
Author: Abramov A.G.
Visitors: 2525
Nowadays, the protection of digital infrastructures of organizations and end users from constantly growing in number and becoming more sophisticated cybersecurity threats is receiving increased attention at various levels. An extremely important task is to ensure reliable and effective protection of critical infrastructures of large telecommunications companies. One of the most common types of cybersecurity threats is Distributed Denial of Service (DDoS) performed at different levels of network interaction, from infrastructure to applica-tions, and aimed at different resources and services. This paper provides an overview of modern methods and technologies to prevent and mitigate DDoS at-tacks with an emphasis on protecting the networks of telecom operators and their users. It also discusses such methods as BGP Blackhole and BGP FlowSpec based on dynamic routing mechanisms and protocols, as well as the methods based on network traffic intelligent analysis and filtering by specialized cleaning sys-tems. The main technical requirements, quality criteria and some quantitative characteristics of DDoS pro-tection solutions are outlined. There are examples of commercial and freely distributed systems. A separate section of the paper is devoted to a detailed description of a relatively simple service for pro-tecting against DDoS attacks. The service is developed and put into operation by specialists of the National Research Computer Network of Russia (NIKS) based on real-time processing and analysis of NetFlow data collected from boundary routers and on the BGP FlowSpec protocol. The is also general information about the hardware and software complex, architecture and main components of the service, involved software packages and technologies along with some statistical data on the results of detecting DDoS attacks in the NIKS network infrastructure.

5. Aspect extraction from scientific paper texts [№4 за 2022 год]
Authors: Marshalova A.E., E.P. Bruches , Batura T.V.
Visitors: 1770
The paper focuses on the problem of automatic aspect extraction from the texts of Russian scientific pa-pers. This problem is relevant due to the increase in the number of scientific publications and the growing need for automated extraction and structuring of key information from them. The study involved the creation of a corpus consisting of 291 abstracts of Russian scientific papers an-notated with the following aspects: task, goal, contribution, method, tool, use, advantage, example, and conclusion. The paper provides descriptions and examples for each aspect. As a result of the corpus annota-tion, 1494 aspects were identified with 44 % of them were the contribution aspect. In addition, the paper proposes an algorithm for automatic aspect extraction. The paper considers the aspect extraction problem as a sequence-labeling problem. The BERT neural network is used to implement the algorithm. The authors have conducted a number of experiments related to the use of vectors obtained from various language models, as well as to freezing the weights of the model. A multilingual model fine-tuned on our data, that is, trained without freezing of the weights, has shown the best result. To improve the quality of aspect extraction, some heuristics, which are listed in the paper, have been developed, and the model has been further trained on the new data obtained from automatic labeling followed by manual edit-ing. The developed system can be useful to other researchers, as it simplifies selection of publications on a particular topic, review of methods for solving a particular problem, and analysis of results obtained in other works.

6. Terms extraction from texts of scientific papers [№4 за 2022 год]
Authors: Dementeva Ya.Yu., E.P. Bruches , Batura T.V.
Visitors: 1855
The relevance of the task of extracting terms from the texts of scientific articles is due to the need for auto-matic annotation and extracting keywords in an ever-increasing flow of scientific and technical documents. This paper explores the influence of various language models on the quality of extracting scientific terms from Russian texts. We compare two models: the mBERT model that was pretrained on texts of different languages, and the ruBERT model pretrained only on Russian data. Two training sets of annotated texts were prepared. The au-thors carried out fine-tuning and further comparison of the performance indicators of the two models using these training sets. They also studied the influence of the choice of the language model on the quality of ex-tracting the terminology contained in the texts of scientific articles. The results have become the base for modernizing the algorithm for extracting terminology from texts applied by the Terminator tool, developed at the A.P. Ershov Institute of Informatics Systems. The obtained results showed that within the framework of the task of extracting terminology from the texts of Russian scientific articles, the ruBERT model, which gave the best performance in an ensemble with a dictionary and heuristics, can be considered as the most applicable model. In addition, the difference in the results of models on full and partial match can be stated due to the problem of defining the boundaries of terms in the texts described in the paper. The results obtained also allow concluding that the quality of the training set markup affects the quality of terminology extraction.

7. Modelling a supercomputer job bundling system based on the Alea simulator [№4 за 2022 год]
Authors: Baranov, A.V., D.S. Lyakhovets
Visitors: 2673
Modern supercomputer job management systems (JMS) are complex software using many different sched-uling algorithms with various parameters. We cannot predict or calculate the impact of changing these pa-rameters on JMS quality metrics. For this reason, researchers use simulation modelling to determine the op-timal JMS parameters. This article discusses the problem of developing a supercomputer job management system model based on the well-known Alea simulator. The object of study is our scheduling algorithm used for developing the supercomputer job bundling system. The algorithm bundles jobs with a long initialization time into groups (packets) according to job types. Initialization is performed once for each group, and then the jobs of the group are executed one after the other. By using a bundling system, it is possible to reduce the initialization overhead and increase the job scheduling efficiency. We implemented the bundling algorithm as a part of the Alea simulator. We have done comparative simulation of implemented algorithm for various workloads. The comparison involved the FCFS and Backfill scheduling algorithms built into Alea. Several workloads with different intensities were generated for the simulation. The minimum job initialization share thresholds for these workloads were determined based on the simulation results. The bundling system noticeably im-proves the scheduling efficiency compared to the FCFS and Backfill algorithms starting from these thresh-olds. The study results showed that the developed simulation model could be used as a software tool for a comparative analysis of various algorithms for supercomputer job scheduling.

8. Informational and algorithmic support of an environmental air monitoring intelligent system based on neural networks [№4 за 2022 год]
Authors: Yarygin G.A., Bayukin M.V., Kornyushko V.F., Shmakova E.G., Sadekov L.V.
Visitors: 2033
The article discusses algorithmic and informational support of an intelligent control system for modern gas analyzers used in environmental air monitoring systems called the Electronic nose. Neural networks form the base of information support. The paper describes a modern automatic odor recognition system based on measurements using low-selective sensors in multi-sensor systems for detecting components of gas mixtures in ambient air. It also shows the advantage of the proposed system compared with traditional systems with highly selective sensing elements. There is a library of smell images based on a series of prerecorded respons-es from the sensor matrix. It is stored in the intelligent system database. Then the responses of an analyzed gas are compared with the responses of individual substances from the image library. The authors propose a two-stage data clustering method for information processing. First, observational data is normalized so that each input parameter equally affects the system. Then the data are assembled in-to clusters using self-organizing Kohonen maps and the k-means algorithm. Each cluster represents an odor with a similar smell. Specific assessments are based on experimental data collected in the environmental monitoring system in the area of the waste incineration plant in Kozhukhovo. The paper considers the choice of an odor identification criteria, which will be used by experts in deciding on odor identification. There is a substantiation of choosing the proximity metric of analytical samples as the norm of the distance between the odor vectors in each sample as a criterion. The authors have developed an algorithm for identifying a substance’s gas analytical sample using neu-ral networks and the selected criterion for decision-making support. There is also a developed (using R pro-gramming language) software product that allows assessing data membership obtained from a device to a certain smell followed by providing visual results of a odors’ spread dynamics in real-time. The paper pre-sents the application results of the developed algorithm in the eco-monitoring system of the incinerator plant in the Kosino-Ukhtomsky district of the Moscow region.

9. Classification of common design patterns for multi-agent systems [№4 за 2022 год]
Author: S.A. Chernyshev
Visitors: 1854
Typically, developing multi-agent systems (MAS) involves using special frameworks or simulation model development environments. They provide the developer with the necessary functionality of an agent launch-ing environment, communication between agents, organization of access to resources and much more. However, there are cases when a stakeholder stipulates that it is necessary to avoid dependencies in the form of these toolkits. The lack of a unified database of MAS design patterns without their binding to specific domain in this case is a significant problem. Therefore, developers are coming up with solutions that have been already described earlier. The purpose of this work is to review and analyze the existing classifications of MAS design patterns, to identify common design patterns without their binding to specific domain, which can be used in the design of multi-agent systems and their classification. From more than 200 MAS design patterns in the public domain, the author has formed a base of 60 pat-terns that are not related to a specific domain. He also proposed the following classification of common MAS design patterns: structural, behavioral, migration, communication, architectural (system), protective and cognitive. Some of the classes of patterns allow introducing additional elements that extend functionali-ty of the system, while others aim to implement different aspects of both the agent and MAS. The most prominent class of all proposed patterns is architectural (systemic) patterns, as they specify different types of agent architectures, multi-agent systems or elements that lay down rigid software constraints on the functioning of the developed system or its parts.

10. Modelling of deformation of elastic objects using perturbation functions [№4 за 2022 год]
Authors: Vyatkin, S.I., Dolgovesov, B.S.
Visitors: 1320
The paper presents a method for modelling the deformation of elastic objects using perturbation functions. It describes deformations of elastic materials capable of stretching in such a way as to return to their original shape and size when releasing force. The method uses second-order differential equations and operator functions in exponential integration. As a result, the calculation time decreases and the overall accuracy in-creases. The method is easily parallelized and allows visualizing complex realistic models. Due to parallel processing and the absence of the need to transfer a large amount of data from the shared memory to the GPU memory, the visualization speed increases compared to the option that uses CPU only. The second paragraph considers a way of defining objects that is different from the polygonal descrip-tion. A basic shape and a set of perturbations are used to define an object. This approach allows reducing memory costs and improving image quality. The third paragraph lists the tasks that solved when modelling animation and deforming bodies using the elastodynamics equations. The paper describes the adaptation of the elastodynamics equations for expo-nential integration. Exponential methods are well suited for rigid systems when solving complex problems. For a rigid system, the authors use a time integrator on the scale of the object general movement with suffi-cient accuracy. There is a description of the exponential processing when sampling a time variable over a certain interval. Exponential integration is constructed using quadrature for a nonlinear integral, which leads to a rigidly accu-rate method necessary to save computational resources compared to classical methods. The authors pro-pose a rigidly accurate method using an adapted scheme with a constant time step. For large systems, they use Newton's square root iteration in order to avoid explicit precomputation of the square root. The fourth paragraph gives the results of testing the method and the comparison with classical and mod-ern approaches for rigid systems. To determine the accuracy of specifying functionally specified objects, the depth buffer of the models (functional and polygonal) is calculated and points are compared to estimate the average difference in depth. Thus, the average deviation relative to the entire model is estimated. In conclu-sion, the authors briefly summarize the results and describe the approaches used in the work.

| 1 | 2 | 3 | Next → ►