Bayesian Optimization for Policy Search in High-Dimensional Systems via Automatic Domain Selection

Bayesian Optimization (BO) is an effective method for optimizing\nexpensive-to-evaluate black-box functions with a wide range of applications for\nexample in robotics, system design and parameter optimization. However, scaling\nBO to problems with large input dimensions (>10) remains an open challenge. In\nthis paper, we propose to leverage results from optimal control to scale BO to\nhigher dimensional control tasks and to reduce the need for manually selecting\nthe optimization domain. The contributions of this paper are twofold: 1) We\nshow how we can make use of a learned dynamics model in combination with a\nmodel-based controller to simplify the BO problem by focusing onto the most\nrelevant regions of the optimization domain. 2) Based on (1) we present a\nmethod to find an embedding in parameter space that reduces the effective\ndimensionality of the optimization problem. To evaluate the effectiveness of\nthe proposed approach, we present an experimental evaluation on real hardware,\nas well as simulated tasks including a 48-dimensional policy for a quadcopter.\n