Kindzerskyi O. Identification of Interval System Models by Bee Colony Software Agents in the NVIDIA CUDA Environment

The dissertation addresses the scientific and technological development of methods for the structural and parametric identification of interval system models by software agents of the metaheuristic artificial bee colony algorithm in the NVIDIA CUDA environment. The aim of the dissertation is to reduce the time complexity of structural and parametric identification methods for interval system models implemented by software agents of the metaheuristic artificial bee colony algorithm in the NVIDIA CUDA environment. The object of the research is the identification processes of mathematical models of dynamic and static systems. The subject of the research is computational methods for identifying interval systems using software agents of the artificial bee colony algorithm in the NVIDIA CUDA environment. Within the dissertation, for the first time, computational methods for the structural and parametric identification of interval system models were developed using software agents that perform the functions of a bee colony in parallel. Each “bee” is interpreted as a software agent that explores the parameter space and the model structure space. Objective functions were formulated in the context of interval data, and the phases of the artificial bee colony algorithm were adapted for parallel execution, which ensured robust global search without loss of interval semantics while simultaneously reducing the time complexity of the method implementation. A technology was developed for applying the NVIDIA CUDA environment to massively parallel evaluation of the objective function, including the generation of CUDA kernels tailored to a specific model structure and their dynamic compilation. It is shown that, depending on model complexity and the dimensionality of experimental data, the GPU implementation provides a reduction in the iteration time of the artificial bee colony algorithm compared with the CPU implementation. The choice of the optimal bee colony population size depending on the problem dimensionality was proposed and justified. Using examples of constructing interval models for environmental monitoring, a regularity was established: for low-dimensional models (2–3 coefficients), the use of the proposed technology and computational methods is impractical. When the dimensionality of the interval model increases to 10 coefficients and the optimal bee colony population size of 4096 is selected, the time complexity of applying the proposed technology and computational methods decreases by more than 45 times. It was found that the higher the problem dimensionality, the higher the efficiency of applying the proposed technology and computational methods. An agent-oriented architecture of a software system for identifying interval system models was designed, combining an object-oriented component structure with a modular organization of CUDA computational kernels, which ensures flexibility, scalability, and efficient use of GPU resources. The developed architecture forms the basis of a computer environment for mathematical modeling of systems based on interval data analysis. The practical significance of the obtained results lies in the development of a computer environment for mathematical modeling of systems based on interval data analysis which, unlike existing solutions, implements an interpreter of basis functions for model construction and combines parallel operation of software agents with their dynamic compilation. Taken together, this simplifies user access to interval modeling modules and reduces the time complexity of model development. The theoretical and applied results of the dissertation have been implemented in the educational process and research activities of the Department of Computer Science at the West Ukrainian National University.