Talk: IEEE CDC 2024 — Using Riemannian Optimization to Learn Optimal Controllers
Published:
I presented this work at IEEE CDC 2024 in Milan, Italy.
What problem was this about?
Direct policy optimization for LQG can be slow and unstable when using standard Euclidean parameterizations and first-order methods.
What was the main contribution?
The work formulates policy optimization on a geometric quotient/orbit structure and develops Riemannian optimization methods with strong convergence guarantees.
Why it matters
This leads to faster and more reliable policy learning for an important benchmark control problem, while preserving theoretical structure.