Accelerating Derivative-Free Optimization with Dimension Reduction and Hyperparameter Learning
share
We consider convex, black-box objective functions with additive or multiplicative noise with a high-dimensional parameter space and a data space of lower dimension, where gradients of the map exist, but may be inaccessible. We investigate Derivative-Free Optimization (DFO) in this setting and propose a novel method, Active STARS (ASTARS), blending the DFO algorithm STARS (Chen and Wild, 2015) and dimension reduction in parameter space via Active Subspace (AS) methods (Constantine, 2015). STARS hyperparmeters are inversely proportional to the known dimension of parameter space, resulting in heavy smoothing and small step sizes for large dimensions. When possible, ASTARS leverages a lower-dimensional AS, defining a set of directions in parameter space causing the majority of the variance in function values. ASTARS iterates are updated with steps only taken in the AS, reducing the value of the objective function more efficiently than STARS, which updates iterates in the full parameter space. Computational costs may be reduced further by estimating ASTARS hyperparameters and the AS, reducing the total evaluations of the objective function and eliminating the requirement that the user specify hyperparameters and AS's, which may be unknown. We call this method Fully Automated ASTARS (FAASTARS). We show that STARS and ASTARS will both converge -- with a certain complexity -- even with inexact, estimated hyperparemters; we also find that FAASTARS converges with the use of estimated AS's and hyperparameters. We explore the effectiveness of ASTARS and FAASTARS in numerical examples which compare ASTARS and FAASTARS to STARS.
READ FULL TEXT