Hyper-Parameter Optimization Using MARS Surrogate for Machine-Learning Algorithms

摘要

Automatically searching for optimal hyper parameters is of crucial importance for applying machine learning algorithms in practice. However, there are concerns regarding the tradeoff between efficiency and effectiveness of current approaches when faced with the expensive function evaluations. In this paper, a novel efficient hyper-parameter optimization algorithm is proposed (called MARSAOP), in which multivariate spline functions are used as surrogate and dynamic coordinate search approach is employed to generate the candidate points. Empirical studies on benchmark problems and machine-learning models (e.g., SVM, RF, and NN) demonstrate that the proposed algorithm is able to find relatively high-quality solutions for benchmark problems and excellent hyper-parameter configurations for machine-learning models using a limited computational budget (few function evaluations).