MINIMAX APPROXIMATION AS A METHOD FOR OVERCOMING THE DOUBLE DESCENT PHENOMENON

This paper explores minimax approximation methods as a tool for overcoming the double descent phenomenon in machine learning. Double descent is an important issue that arises when models begin to demonstrate improved results after overfitting. This phenomenon is observed when training complex models with a large number of parameters, which increases the risk of overfitting. Minimax approximation allows for the reduction of maximum losses that occur during training and stabilizes model results even in situations where data may be noisy or unpredictable. The paper describes the main strategies for minimizing maximum losses, optimizing model complexity, controlling overfitting, and implementing regularization. Special attention is given to identifying the optimal point of the second descent, which enhances the generalization capabilities of models and reduces their susceptibility to overfitting. The research shows that using minimax approximation can significantly improve the effectiveness of model training, especially in conditions where data has a high level of noise or when models have a large number of parameters. The obtained results may be beneficial for researchers and practitioners in the field of machine learning, as they demonstrate new approaches to improving overall model performance and enhancing their reliability.