Özet:

Abstract Hyperparameter optimization is a critical task in enhancing machine learning model performance. This paper introduces a novel approach, for hyperparameter tuning without making assumptions about the underlying hyperparameter distribution or convergence behavior. This approach treats hyperparameter configurations as indices and focuses solely on the associated loss sequences. The objective is to efficiently search for the configuration by minimizing the selected configuration's validation error. The algorithm employs an acquisition function to determine the next configuration to evaluate and leverages a classification model to guide the search process. The proposed methodology is agnostic to the structure of hyperparameter relationships, aiming to strike a balance between resource usage and performance improvement. Experimental results demonstrate the proposed approach’s effectiveness in identifying optimal configurations while being adaptive to various domains and data types.

Anahtar Kelimeler: