Timezone: »

 
Robust online control with model misspecification
Xinyi Chen · Udaya Ghai · Elad Hazan · Alexandre Megretsky

@
in Workshop on Reinforcement Learning Theory »
We study online control of an unknown nonlinear dynamical system that is approximated by a time-invariant linear system with model misspecification. Our study focuses on {\bf robustness}, which measures how much deviation from the assumed linear approximation can be tolerated while maintaining a bounded L2-gain. Some models cannot be stabilized even with perfect knowledge of their coefficients: the robustness is limited by the minimal distance between the assumed dynamics and the set of unstabilizable dynamics. Therefore it is necessary to assume a lower bound on this distance. Under this assumption, and with full observation of the $d$ dimensional state, we describe an efficient controller that attains $\Omega(\frac{1}{\sqrt{d}})$ robustness together with an L2-gain whose dimension dependence is near optimal. We also give an inefficient algorithm that attains constant robustness independent of the dimension, with a finite but sub-optimal L2-gain.

Author Information

Xinyi Chen (Google Research)
Udaya Ghai (Princeton)
Elad Hazan (Princeton University)
Alexandre Megretsky (Massachusetts Institute of Technology)

More from the Same Authors