Vol 26, No 1 (2020)

A method for automatic finding of a program (in the FORT language) that realizes the given algorithm is developed. The algorithm is specified by a set of tests of the form (input_data) ® (output_data). Both input and output data are represented as a sets of 4-byte integers. Genetic methods have made it possible to find the implementation of even relatively complex algorithms: decimal and binary digits of numbers, GCD, LCL, factorial, simple divisors, binomial coefficients, sorting of short sequences, highs, lows, calculation of polynomial values and others. The genetic approach allows you to build a program from separate blocks, "genes", which turned out to be suitable for at least some part of the test elements. Genetic methods made it possible to find the implementation of even relatively complex algorithms: decimal and binary digits of a number,, NOC, factorial, simple divisors, binomial coefficients, sorting of short sequences, maxima, minima, calculation of polynomial values and others. Our method starts by randomly sorting through short programs, extracting blocks ("genes") from them at least slightly suitable for the problem being solved. And then he builds a program using the found "genes". The set of used "genes" in the process of the algorithm is constantly being adjusted, improved. The complexity of direct enumeration grows exponentially with increasing program length. The genetic method we propose allows us in many cases to drastically reduce the volume of search. The FORT language is chosen because of its compactness: all listed algorithms are placed in no more than 10—15 commands. Although, if we take into account the genes found, the total length of the program will be significantly longer. In addition, the mechanism of embedding "genes" already, in fact, is in the language. Our method is configured to work with integers, but it can be applied to data containing real numbers, strings, etc. In the case of working with real numbers, the method can be considered as an alternative to the methods used in neural networks.

One of the important problems of building modern radio systems is the fulfillment of the requirements for the stability of their functioning in difficult conditions of signal and interference conditions. The greatest relevance of the task of this subject area is acquired when it is necessary to take into account the parameters of deliberate destructive influences. Known approaches are based on expanding the base of signals based on the resources of radio engineering systems. Such approaches are quite effective in conditions of homogeneous models of radio channels, which are characterized by the same parameters of the functions of the distribution of signal characteristics and interference at operating frequencies. Under the conditions of inhomogeneous models of radio channels, the selected class of solutions is rather limited due to the uncertainty of the parameters of the statistical models of radio systems. The aim of the work is to study the properties of the statistical model of radio systems that implement the threshold method of controlling the frequency resource. This method is based on the implementation of procedures for sounding working channels, determining the signal level exceeding the noise level and eliminating working frequencies with unacceptable values of the measured parameter. To achieve the stated goal, a simulation model of noise-free radio engineering systems was developed based on the threshold frequency resource control method. A hypothesis is formulated, based on the dependence of statistical laws on the parameters of the array of excluded frequencies. To prove the hypothesis, an experiment was conducted using the developed simulation model. A set of statistics was carried out, on the basis of which the characteristics of the studied radio engineering systems were obtained. Analysis of the results confirms the validity of the formulated hypothesis. The main conclusions of the work confirm and justify the nature of changes in the statistical characteristics of noise-free radio engineering systems with a threshold method of frequency resource control. At the same time, the distribution functions of the exponential type characterize the largest decrease in the specified performance indicators. In conclusion, shows the approximate boundaries of the array of excluded frequencies, defining the nature of changes in statistical parameters. (не менее 200 слов).

The paper presents the implementation of a dynamic routing algorithm intended for use in networks-on-chip with a three-dimensional circulant topology of type C(N; s1, s2, s3). Compared with the classical algorithms A* or Dijkstra, the proposed algorithm does not require to calculate the entire path of the packet, but calculates the port number to which the packet should be sent so that it can reach the destination node. This makes it possible to significantly simplify the structure of the NoC router. The algorithm can be implemented as RTL state machine in routers for NoCs for finding the shortest routes between any two network nodes. The proposed algorithm was tested on sets of optimal circulants. For this set, it is equal to 0.973 which is a sufficient indicator of efficiency of the algorithm for the task of implementing the algorithm in HDL at the level of a network-on-chip router, since this algorithm has linear complexity and can be easily implemented. In 95 % of the cases considered, for a set of graphs with the number of vertices less than 300, the algorithm showed a result similar to that obtained using the Dijkstra algorithm. The computational time complexity of the created algorithm is similar to the complexity of the Dijkstra algorithm which is considered to be the reference when finding routes in networks-on-chip. When the number of vertices is more than 300, the algorithm becomes inefficient, since the diameters, calculated by "Sequent" algorithm, can be ten times higher than the optimal diameters calculated by the Dijkstra algorithm.

The amount of information provided in text form is increasing every year. Thus, the task of automatic natural language text processing (NLP), and especially the reduction of its volume, is becoming increasingly important. But tools which produce abstracts in the Russian language mainly use the statistical method only. So, there is a need to research the summarization algorithms and create new algorithms and tools that will use several automatic summarization methods simultaneously to improve the results. The article shows the results of extraction algorithms, basing of which an integrated summarization algorithm, a library and service of automatic text summarization in the Russian languages are proposed to enable the use of various reference methods.

Proposed the method of coding and decoding video data based on adaptive three-dimensional discrete cosine transform (3D-DCT), provide increased efficiency of operation of the transfer device and reduction of information quality, systems: distortion levelsf, bit rate, and complexity encoders. To improve the performance of the image compression algorithm, the classification of transport subjects by type of motion is used. The results of testing the algorithm and a comparative analysis of the proposed method with the known methods MPEG2 and MPEG4 are obtained.

This study is devoted to the deep investigation of the main atherosclerosis markers and their predictors using the different machine learning methods. The main goal of the investigation is to find the most efficient markers and predictors, which would allow to identify the presence of this disease with the high accuracy. The dataset under consideration is based on the real patients characteristics, including blood pressure values, clinical, anthropometrical and social features. Such methods as classification trees, multilayered neural networks and Kohonen maps are used for solving the considered problem of binary classification. Accuracy, precision, recall (sensitivity) and true negative rate (specifity) are the main metrics, which are used to evaluate the quality of the resulting classification models. Classification trees method allowed not only to generate the most interpretable models, but also to order the main predictors by their importance, and the most important predictors were then passed as inputs to the multilayered neural networks. The base backpropagation algorithm, used by default for learning the network, has been modified to use the misclassification costs, which allowed to increase the quality of the classification for the unbalanced dataset. In order to show the common rules found in the dataset, the Kohonen maps were build for each of the main atherosclerosis markers. All the developed classification models can be considered as instruments of the non-invasive atherosclerosis diagnostics.