Multiobjective Bayesian optimization on HPGR

Sammanfattning

When applying Bayesian optimization (BO) algorithms which utilizes Gaussian processes (GP) models, several components can be adapted to suit different systems and optimization objectives. In this paper, we explore and evaluate a set of such modifications to tailor the algorithm for a high-pressure grinding rolls (HPGR) system. The system is represented by a steady-state simulator modeling the HPGR machine, used in ore comminution. The primary goal of the optimization is to evaluate the possibility of using Bayesian optimization on the HPGR-system to generate recommendations for its control settings, and how the BO should be set up and tuned for this purpose. We investigate the effects of different covariance functions, mean functions, and acquisition functions on optimization performance by testing them on the default system parameters, and validating the results on time series of varying feed material. Among the tested configurations, the final recommendation for the model settings were using the radial basis function (RBF) covariance and a constant mean function for all objectives and constraints. The recommended acquisition function was a version of Expected Hypervolume Improvement (EHVI).