Yisong Yue, California Institute of Technology

Learning to Optimize as Policy Learning
Oct 23, 2020, 11:00 am12:30 pm
Event Description

The design of optimization procedures is typically done via careful analysis (both theoretical and empirical) of the underlying structure of the optimization problem. While effective, this design philosophy is labor intensive and difficult to deploy efficiently to a broad range of domain-specific optimization problems. In this talk, I will describe ongoing efforts to use learning approaches to automatically design solvers. The basic idea is to train the solvers on a pre-collected batch of representative problem instances with the goal of optimizing some performance metric (e.g., total running time, or objective value under bounded time, etc.). I will show this line of inquiry can both motivate novel fundamental research policy learning (which includes reinforcement and imitation learning as special cases), as well as lead to practical impact.