A0367
Title: Penalized additive Gaussian process for auto-tuning of quantitative and qualitative factors in Black-Box systems
Authors: Yongxiang Li - Shanghai Jiao Tong University (China) [presenting]
Abstract: Optimizing black-box systems with both quantitative and qualitative (QQ) factors is critical in various applications where resource-intensive or time-consuming evaluations make factor-level selection critical. Traditional sensitivity analysis lacks a unified framework for simultaneously screening important QQ factors and struggles to select important qualitative levels. To address this, a penalized additive Gaussian process (PAGP) model is introduced, featuring an interpretable additive (IA) covariance function for QQ factors. This allows sparsity penalties that enable the identification of critical qualitative levels. The three-step model fitting approach utilizes derivatives for acceleration, and a tailored ADMM is proposed to optimize the L1 regularized likelihood. Then, qualitative level screening is proposed utilizing sparse regularization and quantitative factor selection leveraging Shapley value. Finally, Bayesian optimization is introduced to PAGP for the optimization of black-box systems with QQ factors, and the sparse covariance function will guide Bayesian optimization in efficiently searching the important qualitative levels. PAGP distinguishes itself by enabling sparse regularization and efficient screening of qualitative levels. Simulation studies validate the outperformance of PAGP, and it is also applied to the design of paper pilots and neural networks.
