Тип публикации: статья из журнала
Год издания: 2025
Идентификатор DOI: 10.22190/fumi241010061e
Аннотация: <jats:p>The widespread introduction of artificial neural networks leads to the need to create a variety of architectures for solving various problems. Optimization methods are actively used to automate this process. This process is called Neural architecture search. Among them the greatest variety of models is given by multi-objectПоказать полностьюive optimization algorithms based on simultaneous optimization of all criteria. An important part of the algorithms is the way the neural networks are encoded. A flexible scheme of encoding and decoding of architectures-individuals by binary strings was proposed: each part of a binary string is responsible for a parameter or a whole block (layer, layer set) of a neural network. This approach allows to create neural networks with different architectures within one run of the algorithm. The use of the notion of strict dominance has also been investigated, computational experiments have shown that the use of strict dominance gives a statistically significant increase in hypervolume metric.</jats:p>
Журнал: Facta Universitatis, Series: Mathematics and Informatics
Выпуск журнала: Т. 40, № 4
Номера страниц: 873
ISSN журнала: 03529665
Место издания: Белград