No 1 (2024)

Artificial intelligence (AI) has become one of the most disruptive approaches to processing huge volumes of heterogeneous data and performing fundamental cyber security tasks such as intrusion detection, vulnerability management, security monitoring, asset prioritization, access control. The article presents the current state of the use of AI methods (primarily machine learning methods) in cyber security. Key areas of focus at the intersection of AI and cyber security are analyzed. The article partially reflects the content of the plenary report given at the XX National Conference on Artificial Intelligence with International Participation (NCAI-2022).

The paper examines the expressive and formal capabilities of the theory of system-object modeling for describing knowledge about the structure, functioning and states of a computer network. Methods for describing distributed information systems using systems calculus as functional objects are considered. The limitations of using the alphabet of elementary node objects are presented. A mechanism for describing the states of a computer network node is proposed.

The paper describes an intelligent recommendation system for restoring and training the human respiratory system using individually selected special exercises and increasing motivation when performing them. Personal recommendations for the exercises’ composition are formed on the basis of interactive intellectual analysis of video information about a person’s physical activity, taking into account his/her experience. Machine learning models and methods are used to select exercises and evaluate the effectiveness of their implementation. The results of testing the recommendation system are presented.

The paper presents an algorithm for anomaly detection in bivalve activity data using the error between the value predicted with SARIMA model and the actual value. Decomposition of time series was carried out to determine the seasonal component of the models. The optimal model for all averaging times of activity data of freshwater bivalve was made. After this, using the developed algorithmic software, the root mean square error metric was calculated for the entire data set, which made it possible to determine the potential threshold for the operation of the algorithm, as well as the algorithm’s response time to anomalies at different data averaging times. The results obtained will be included in the algorithmic software of an automated complex for biomonitoring the state of water quality based on bivalves, which is already functioning and located in the waters of Sevastopol, which will allow faster and more likely to detect anomalies and generate an alarm signal.

The article discusses the limitations of using interference tests to construct a tracer testing program as a list of injection-production wells pairs. The decision tree classifier proposed in earlier works is considered as more preferred method for this task. The disadvantages of the existing tree learning algorithm is that it tends to overfit, especially in conditions of small data sets. In this work, we suggest to use techniques from combinatorial theory of overfitting, namely the complete cross-validation and the expected overfitting, as splitting criteria in decision tree nodes to enhance the algorithm's generalization ability. The approach is tested on two fields in Western Siberia, resulting in a statistically significant improvement in the quality of the decision tree and reduced overfitting, leading to more accurate constructing the plan of tracer testing for assessing the presence of hydraulic connectivity between injection and production wells. The application of combinatorial theory of overfitting to decision tree classifiers offers a promising avenue for enhancing the effectiveness of tracer testing in the oil and gas industry.

The features and capabilities of the PointNet neural network architecture in relation to artificially generated clouds of laser reflection points in the Terra_Maker information system are presented. The results of training by the Paintnet network are analyzed and the accuracy of the obtained models and graphs is evaluated. An approach is proposed to determine the parameters that give maximum accuracy when performing experiments on the example of point clouds obtained from the Terra_Maker information system.

The paper presents a method for reducing the search space for optimal clustering parameters. This is achieved by selecting the most appropriate data transformation methods and dissimilarity measures at the stage prior to performing the clustering itself. To compare the selected methods, it is proposed to use the silhouette coefficient, which considers class labels from a small labeled data set as cluster labels. The results of an experimental test of the proposed approach for clustering news texts are presented.

The XXI National Conference on Artificial Intelligence with International Participation (CAI-2023) was held in Smolensk, Russia, on October 16-20, 2023. The conference was co-organized by the Russian Association of Artificial Intelligence, the Federal Research Center "Computer Scince and Control" of the Russian Academy of Sciences, and a branch of the National Research University "MPEI" in Smolensk. Various areas and applications of artificial intelligence were discussed in plenary reports and section meetings.