Safe learning & online control
My lab's interest in safe learning originally grew out of demand response problems in power systems, where even a single exploratory action can be physically unacceptable. Motivated by these stage-wise safety constraints, which cannot be violated at any point during the learning process, we have studied stage-wise safety in bandits, online convex optimization, and learning-based control, with an emphasis on guarantees that are meaningful for safety-critical systems rather than only average-case performance.
