Vol 31, No 7 (2025)

The problem of optimal control with free right end of the trajectory is considered. To find its approximate solution, a reduction to a finite-dimensional optimization problem is performed. The control is a bounded piecewise constant function. A real-coded genetic algorithm is proposed to solve the finite-dimensional problem. To maintain the diversity of the population, a dynamic population size is proposed to be introduced into the algorithm. The algorithm finds a solution to the multi-extremal optimal control problem under different initial approximations. The work of the algorithm is tested on the optimal control problem with a non-convex reachability region. The work of the algorithm is compared with the method of variations in the control space and the genetic algorithm without modifications, as a result of which the advantage of using the modified genetic algorithm is shown.

The article discusses various approaches to predicting parameters, in particular the characteristics of the electromagnetic field based on data from a distributed environment. Several classification models were investigated on a simulated sensor data set, covering three different groups of categorization methods: Bayesian methods based on Bayes’ theorem (Naive Bayes and Multinomial Naive Bayes); decision tree methods, which are multi-option methods and the basic elements of the Decision Stump, Hoeffding Tree (Very Fast Decision Trees), Hoeffding Option Tree and Hoeffding Adaptive Tree algorithms; meta/ ensemble methods, which are the combination of a set of classification models that perform the same task, and solutions of individual models that are combined to determine the output. Experimental analysis has shown that the use of models can significantly improve forecasting efficiency. The methods proposed in the article are aimed at efficient classification. The article presented by the authors obtained results related to distributed machine learning. The first is the performance of classifiers with and without regularization in terms of the accuracy metric, the second is the relationship between the size of the dataset and this metric. In this work, in order to avoid overfitting and subsequent reduction in model accuracy, 1 regularization or Lasso regression is used. Regularization or Lasso regression is used. Thus, the results obtained are effectively implemented in a real-time system that measures streaming information about the parameters of the electromagnetic field.

The paper considers the problem of efficient hardware implementation of nonlinear activation functions of neural networks under low-bit computing conditions. Standard activations, such as sigmoid and hyperbolic tangent, require resource-intensive floating-point operations, which limits their use on microcontrollers, FPGAs and other peripheral platforms. As a solution, an approach based on precomputed integer substitution tables (LUTs) is proposed to reduce computational complexity and power consumption. Using the example of the SiLU activation function widely used in popular object detection networks (e.g., YOLO), the quantisation procedure is demonstrated, the principles of constructing and using LUTs are formulated, and a practical algorithm for computing activations using them is described.

The dynamic development of information and communication technologies (ICT), artificial intelligence and digital technologies has contributed to the development of multi-level information systems. In addition to other fields, gesture-human-machine systems have improved and have a more user-friendly interface. The improvement of gesture-human-machine systems not only facilitates the communication and social adaptation of hearing-impaired people, but also expands user opportunities. Automatic recognition of gestures has enabled remote control of devices in the robotics field. The article clarifies the process of recording the parameters of the hand showing the gesture and the working principle of the recognition methods. The gesture recognition algorithm was analyzed using an artificial neural network and a random forest method in a hybrid working environment. Suggestions were made to eliminate the shortcomings that arose in the training process.

A new approach to designing effective software systems is proposed. In this case, efficiency improvement means minimizing any costs associated with a software system at various stages of the life cycle, including the cost of developing, modifying and operating a software system, and effective software systems are those systems that, among other systems designed to achieve a similar result, are associated with minimal costs at all stages the life cycle. The development stage is analyzed in detail, a comparison of the traditional approach to the design of software systems and the proposed new approach is carried out. The proposed approach to the design and modification of software systems is implemented for the task of developing algorithms for computationally complex procedures, which made it possible to reduce the number of iterations of software system development by 2-3 times, while reducing the time required to complete these iterations.