Перегляд за Автор "Zakovorotnyi, A. Yu."
Зараз показуємо 1 - 3 з 3
Результатів на сторінці
Налаштування сортування
Документ Generating currency exchange rate data based on Quant-GAN model(Національний технічний університет "Харківський політехнічний інститут", 2023) Bao, Dun; Zakovorotnyi, A. Yu.; Kuchuk, N. G.This paper discusses the use of machine learning algorithms to generate data that meets the demands of academia and industry in the context of exchange rate fluctuations. Research results. The paper builds a Quant-GAN model using temporal convolutional neural networks (CNN) and trains it on end-of-day and intraday high-frequency rates of currency pairs in the global market. The generated data is evaluated using various statistical methods and is found to effectively simulate the real dataset. Experimental results show that data generated by the model effectively fits statistical characteristics and typical facts of real training datasets with good overall fit. The results provide effective means for global FX market participants to carry out various tasks such as stress tests and scenario simulations. Future work includes accumulating data and increasing computing power, optimizing and improving GAN models, and establishing evaluation standards for generating exchange rate price data. As computing power continues to grow, the GAN model’s ability to process ultra-large-scale datasets is expected to improve.Документ The neural network art which uses the Hamming distance to measure an image similarity score(Asian Research Publishing Networ, 2019) Dmitrienko, V. D.; Zakovorotnyi, A. Yu.; Leonov, S. Yu.This study reports a new discrete neural network of Adaptive Resonance Theory (ART-1H) in which the Hamming distance is used for the first time to estimate the measure of binary images (vectors) proximity. For the development of a new neural network of adaptive resonance theory, architectures and operational algorithms of discrete neural networks ART-1 and discrete Hamming neural networks are used. Unlike the discrete neural network adaptive resonance theory ART-1 in which the similarity parameter which takes into account single images components only is used as a measure of images (vectors) proximity in the new network in the Hamming distance all the components of black and white images are taken into account. In contrast to the Hamming network, the new network allows the formation of typical vector classes representatives in the learning process not using information from the teacher which is not always reliable. New neural network can combine the advantages of the Hamming neural network and ART-1 by setting a part of source information in the form of reference images (distinctive feature and advantage of the Hamming neural network) and obtaining some of typical image classes representatives using learning algorithms of the neural network ART-1 (the dignity of the neural network ART-1). The architecture and functional algorithms of the new neural network ART which has the properties of both neural network ART-1 and the Hamming network were proposed and investigated. The network can use three methods to get information about typical image classes representatives: teacher information, neural network learning process, third method uses a combination of first two methods. Property of neural network ART-1 and ART-1H, related to the dependence of network learning outcomes or classification of input information to the order of the vectors (images) can be considered not as a disadvantage of the networks but as a virtue. This property allows to receive various types of input information classification which cannot be obtained using other neural networks.Документ Neural networks art: solving problems with multiple solutions and new teaching algorithm(2014) Dmitrienko, V. D.; Zakovorotnyi, A. Yu.; Leonov, S. Yu.; Khavina, I. P.A new discrete neural networks adaptive resonance theory (ART), which allows solving problems with multiple solutions, is developed. New algorithms neural networks teaching ART to prevent degradation and reproduction classes at training noisy input data is developed. Proposed learning algorithms discrete ART networks, allowing obtaining different classification methods of input.