ARTIFICIAL INTELLIGENCE AND DECISION MAKING

The article considers one of the key problems of container systems related to the detection of attacks and anomalies. The mechanisms of isolation of container systems and attacks on such systems are described. A classification of approaches to the detection of attacks and anomalies is presented. A systematic analysis of the main approaches to the detection of attacks and anomalies in container systems, as well as methods for their implementation, is performed. Traditional approaches based on signatures and rules, their features, advantages and disadvantages are considered in detail.

The results of training and testing of an artificial neural network for recognizing human finger movements based on signals from electromyographic sensors are presented. Special attention is paid to the issues of preliminary processing of initial signals, including digital filtering, setting the optimal level corresponding to the resting state of the muscle, and calculation of signal attributes. In the paper, an envelope of the electromyographic signal was built on the basis of the “average energy” attribute, and the definition of muscle activity areas was carried out using two thresholds: adaptive in level and fixed in time. Three attributes are used directly for training the artificial neural network, which are specified depending on the requirements to the quality of training, either by indicator of distinguishability or by a complete enumeration of combinations of attributes. Optimization of the set of attributes for training the artificial neural network allowed achieving the level of correct answers more than 97%.

The paper considers the problem of selecting linear regression factors using an optimization model that includes characteristics of the relationship of features, as well as the dependence of the feature and the effective indicator. To solve it, it is proposed to reformulate the original problem in the form of an inverse while minimizing the sum of the absolute values of the arguments. The results of computational experiments, including comparison with nonlinear programming methods implemented in mathematical packages and the Python library, demonstrated the high efficiency of the proposed algorithm for solving the modified problem.

Widely used in practice the recommender systems, decision support systems, intelligent control, AI assistants in medicine, and search engines can influence users and properties of the environment in which they are employed. The process of repeated machine learning describes such systems in which continuous improvement of machine learning models is performed over time using training data obtained from the users. In this paper, we study how feature space transformations influence properties of the repeated machine learning process. In particular, we investigate the conditions under which the prediction of the asymptotic behavior of a system over time obtained in the original space can be applied to a similar system in the transformed space. The results of the research indicate the possibility of using simpler systems in spaces of lower dimensionality to study processes in more complex systems.

The article describes a method for lexical simplification of Russian text, using a specially marked base of synonyms and a set of rules, which allows automatic lexical replacement of words and phrases with restoration of the correct syntax and preservation of the text semantics. To form a marked database of synonyms, dictionaries of synonyms that are in the public domain were used. To preserve semantics, an analysis and reduction of synonymous series was carried out, the frequency of members of the synonymous series was analyzed in order to select a dominant, and entries in dictionaries were marked and a mechanism for processing labels was proposed to comply with the syntax rules in a simplified text. A base of production rules has been developed to preserve the correct syntax after lexical adaptation of the text, allowing for the correct replacement of individual words, phrases with one word and phrases with a phrase.

Neural networks and deep learning algorithms are increasingly used in medicine, including image analysis. In surgery, soft tissue wounds assessment remains challenging but necessary issue to assess the course of healing process and treatment effectiveness. Digital wound images are used for noncontact wound analysis. The paper presents the results of pre-trained network models (AlexNet, ResNet50, ResNet152, VGG16) used to classify pressure ulcer images as examples of chronic wounds. The Segment Anything Model (SAM) demonstrated an accuracy of 86.46% in solving the problem of segmenting the edges of a wound defect and tissue types within it. The results can be used to create an expert system for analyzing soft tissue wound images.