Abstract: First-order methods for quadratic optimization such as OSQP are widely used for large-scale machine learning and embedded optimal control, where many related problems must be rapidly solved. These methods face two persistent challenges: manual hyperparameter tuning and convergence time to high-accuracy solutions. To address these, we explore how Reinforcement Learning (RL) can learn a policy to tune parameters to accelerate convergence. In experiments with well-known QP benchmarks we find that our RL policy, RLQP, significantly outperforms state-of-the-art QP solvers by up to 3x. RLQP generalizes surprisingly well to previously unseen problems with varying dimension and structure from different applications, including the QPLIB, Netlib LP and Maros-Mészáros problems.

Short Bio: Jeffrey Ichnowski is a post-doctoral researcher in the RISE lab and AUTOLAB at the University of California at Berkeley. He researches algorithms and systems for high-speed motion, task, and grasp planning for robots, using cloud-based high-performance computing, optimization, and deep learning. Jeff has a Ph.D. in computational robotics from the University of North Carolina at Chapel Hill. Before returning to academia, he founded startups and was an engineering director and the principal architect at SuccessFactors, one of the world’s largest cloud-based software-as-a-service companies.

This will be a joint talk with Mahan Tajrobehkar. This talk will take place from approximately 3:40 p.m. – 4:05 p.m.