Optimization of hyperparameters with constraints on time and memory for the classification model of the hard drives states

This article explores the issues of hyperparameters optimization of machine learning classification models under time and memory constraints A number of methods for optimizing hyperparameters of classification models in the learning process are considered: RandomSearch, GridSearch, TPE, CMA-ES. The effectiveness of these optimization methods is tested in the context of the task of classifying the states of hard disks. The construction and creation of classification models is based on two machine learning algorithms: LSTM and Random Forest. The hyperparameters of the proposed classification models are optimized based on the above methods. The models are trained on a public dataset from BackBlaze cloud storage. The article provides estimates of the values of the main indicators of classification quality, a comparative analysis of optimization methods is carried out. The experimental results confirm the feasibility of using optimization methods under time and memory constraints. Of particular note is the TPE method, which outperformed other methods in achieving the task of maximizing classification quality indicators.