Vol 31, No 12 (2025)

The paper describes a method for solving the problem of coverage based on the hybridization of heuristics, and mechanisms of collective adaptation and swarm intelligence. А modernized agent swarm metaheuristics is proposed, characterized in that adaptive chromosomes serve as agents, and the search process is organized in an affine solution space. An algorithm has been developed for the random formation of an initial population of solutions in the form of a set of legal matrices of boundary requirements. А comparison with known algorithms has shown that with a shorter operating time, the solutions obtained using the developed algorithm have a deviation of the objective function from the optimal value by an average of 6 % less.

In this paper, common types of adversarial attacks (DTA, FGSM, and BIM) are used to generate adversarial samples to test the vulnerability of IDS using the UNSW-NB15 dataset. Then, basic defense mechanisms are developed, including adversarial pattern detection and filtering. Experiments are conducted on Random Forest (RF) and Logistic Regression (LR) machine learning classifier.

The paper considers the problem of a drop in real performance with an increase in the number of computing cores on supercomputer systems. The results of the first three systems from the TOP500 List are analyzed. The main approaches to solving the problem of increasing real performance are given.

A dataflow computing model with a dynamically formed context is considered. The architecture of a parallel dataflow computing system implementing this model is described. The model and architecture are one of the approaches to improving the real performance of computing systems. The principles of functioning of hardware ternary content-addressable memory, which implements the dataflow computing model in the most efficient way, are described. This is due to the fact that the concept of the computing model, such as a token workspace, involves the simultaneous comparison of an incoming token with all tokens present in this space. One of the key problems of content-addressable memory is analyzed — high power consumption when performing matching operations. Methods for optimizing the matching process are proposed, which are divided into three groups — hardware, software, and hardware-software. Optimization methods are aimed at overall reducing the number of "parasitic" comparisons, as well as reducing the number of compared bits. An analysis of its effectiveness was carried out for each method. One of the effective methods of reducing the number of matchings in the content-addressable memory of the keys of the matching processor of the parallel dataflow computing system is the use of special "Double grouped" tokens. These tokens allow not only to reduce the total number of task tokens, but also to reduce the load on the communication network, reduce the number of comparisons, and free up execution units by transferring part of the load to the matching processor.

The research results obtained when performing various tasks on the behavioral block-register model of the system and the emulator are presented. The results demonstrate the effectiveness of the proposed methods.

This study addresses the problem of intelligent anomaly prediction during the early stages of very-large-scale integration (VLSI) design, under the constraints of increasingly complex topologies and modern technological standards. The problem formulation includes the formalization of criteria for anomaly detection, such as congestion zones, topological conflicts, and routing violations. To enhance classification accuracy and adaptability to diverse design data structures, a hybrid architecture is proposed that combines metaheuristic methods with machine learning algorithms. А computational experiment was conducted using both synthetic and publicly available datasets, demonstrating the effectiveness of the stacking model and graph neural networks in achieving high prediction quality within acceptable training times. The study also explores the impact of cross-validation and class balancing on resistance to overfitting. The obtained results confirm the practical applicability of the proposed approach in modern computer-aided design (CAD) systems. This article will be of interest to specialists in design automation and intelligent analysis of VLSI systems, as well as researchers working on hybrid methods in engineering applications.