Relevance. Building high-precision predictive models in modern intelligent systems requires automating the search for optimal hyperparameters. Traditional optimization methods demonstrate low efficiency in high-dimensional spaces and with significant computational costs for estimating the objective function, which necessitates the development of new approaches at the interface of heuristic search and statistical approximation. Problem. The main difficulty lies in the need to find the global extremum of the «black box» function with a strict limitation of the computing budget. The high resource intensity of each access to the complete machine learning model requires minimizing the number of iterations without losing the accuracy and robustness of the final solution. Methods. A hybrid EA-SM algorithm is proposed that integrates the mechanisms of evolutionary search and adaptive surrogate modeling based on Gaussian processes. The mathematical apparatus includes the use of a data collection function to balance space exploration and exploit the found minima, as well as Tikhonov regularization to ensure the computational stability of covariance matrices. Results. Experimental verification on the tasks of stochastic object classification and time series forecasting (AutoForecast, Chronos) confirmed the superiority of the method. A reduction in the number of calls to the objective function by 30-70% has been found compared to the DIRECT and Optuna algorithms, while maintaining high approximation accuracy in the vicinity of extremes. Conclusions. The developed approach provides asymptotic convergence to the global optimum and resistance to stochastic noise. The algorithm is suitable for configuring neural network architectures in high-dimensional environments, minimizing time and hardware costs in monitoring and anomaly detection systems.
Relevance. Building high-precision predictive models in modern intelligent systems requires automating the search for optimal hyperparameters. Traditional optimization methods demonstrate low efficiency in high-dimensional spaces and with significant computational costs for estimating the objective function, which necessitates the development of new approaches at the interface of heuristic search and statistical approximation. Problem. The main difficulty lies in the need to find the global extremum of the «black box» function with a strict limitation of the computing budget. The high resource intensity of each access to the complete machine learning model requires minimizing the number of iterations without losing the accuracy and robustness of the final solution. Methods. A hybrid EA-SM algorithm is proposed that integrates the mechanisms of evolutionary search and adaptive surrogate modeling based on Gaussian processes. The mathematical apparatus includes the use of a data collection function to balance space exploration and exploit the found minima, as well as Tikhonov regularization to ensure the computational stability of covariance matrices. Results. Experimental verification on the tasks of stochastic object classification and time series forecasting (AutoForecast, Chronos) confirmed the superiority of the method. A reduction in the number of calls to the objective function by 30-70% has been found compared to the DIRECT and Optuna algorithms, while maintaining high approximation accuracy in the vicinity of extremes. Conclusions. The developed approach provides asymptotic convergence to the global optimum and resistance to stochastic noise. The algorithm is suitable for configuring neural network architectures in high-dimensional environments, minimizing time and hardware costs in monitoring and anomaly detection systems.