ARTIFICIAL INTELLIGENCE AND DECISION MAKING

The possibilities to apply artificial intelligence methods and systems to ensure situational awareness of decision makers (DM) are discussed. The task of maintaining situational awareness is considered in the context of Big Data intelligent analysis, which must be completed in a limited time. Examples of relevant applications of this type in several subject areas are given. The special requirements for IT systems and solutions that should ensure situational awareness of the DM are presented. Some features of project management for the development of large IT systems of the type under discussion are considered.

The relationship between the average characteristics of a fuzzy signal at the input and output of a linear dynamic system is established. Numerical and interval averages of fuzzy numbers are considered as average characteristics. In this paper, in accordance with the system parameters and averages at the input, ordinary differential equations are obtained for determining the numerical averages of a fuzzy-valued solution, i.e. for determining the averages of a fuzzy signal at the output of a linear dynamic system (the defuzzification problem is essentially solved). In addition, ordinary differential equations are established for finding the boundary functions of the interval averages of a fuzzy-valued solution that are optimal in the root-mean-square. An interval approximation of a fuzzy output signal is also obtained using an interval approximation of a fuzzy input signal by the method of ranking fuzzy numbers. Applications for a harmonic oscillator model with fuzzy-valued heterogeneity are considered.

The current problem of constructing a new artificial neural network adapted to the limited resources of onboard computing systems of unmanned aerial vehicles is investigated. The purpose of the study is to develop an artificial neural network architecture that uses the ideas of group consideration of Ivakhnenko’s arguments and the Kolmogorov-Amold method of representing a function of many variables by a composition of functions of lesser complexity. Instead of a difficult-to-implement activation function, the sum of the elements of the Kolmogorov-Gabor polynomial is used here. The result is the overall architecture of the hybrid network and the individual neuron within it. To train such a network, a combination of limited iteration and backpropagation methods is used. The adjustment algorithm is implemented in relation to the problem of binary classification of aircraft by invariant moments of their silhouettes. A comparison with the results of solving a similar problem based on artificial neural networks with activation functions such as sigmoid and “s-parabola” showed their comparability in quality.

This paper proposes an improved method for analytically computing the weights of a Convolutional Neural Network (CNN) and presents a comparative analysis of CNN training with precomputed weights versus randomly initialized weights. Experiments were conducted using the MN1ST handwritten digit dataset. The results demonstrated the advantages of using precomputed weights. A CNN initialized with such weights, based on only 10 randomly selected MNIST images, was able to achieve over 50% recognition accuracy on the test set even before training began. Moreover, the time required to compute the weights was minimal. A comparative analysis of multiple training runs using the same CNN architecture showed that pre-initializing weights improves final recognition performance and reduces overall training time.

The article presents a comprehensive approach to precision control of fixed-wing unmanned aerial vehicle (UAV) rudders for autonomous guidance based on visual tracking. The study aims to develop and validate a control system that ensures high accuracy and reliability in autonomously guiding the UAV to visually identified targets. The proposed system integrates the CSRT visual tracking algorithm, Kalman filtering to enhance the stability of object position estimation, and a developed mathematical control model based on combined data from visual tracking and a laser rangefinder. To address the issue of overshooting in the guidance system, the parameters of the PID controller were optimized using the Zieg-ler-Nichols method. Experimental results demonstrate high targeting accuracy and stable target tracking in real flight conditions, confirming the effectiveness of the proposed approach. The findings have practical significance for various UAV applications requiring high precision and autonomous control.

The paper is devoted to the creation of methods for automatic assessment of the consistency of scientific and technological development priorities and the passports of science majors of the Higher Attestation Commission. The paper proposes a multi-step approach, including the construction of collections of topically similar documents for the priorities and the passports, assessment of the proximity of the found documents using cross-language neural network encoders, and formation of a categorical assessment of the consistency of priorities and passports based on the obtained indicators of document proximity. Experimental studies on a specially created corpus confirm the high quality of the results obtained with the proposed method.