On a Phase Transition of Regret in Linear Quadratic Control: The Memoryless Case

Ingvar Ziemann (1) , Henrik Sandberg (1)

1 KTH - KTH Royal Institute of Technology [Stockholm]

We consider an idealized version of adaptive control of a MIMO system without state. We demonstrate how rank deficient Fisher information in this simple memoryless problem leads to the impossibility of logarithmic rates of regret. This to some extent resolves an open issue concerning the attainability of logarithmic regret rates in linear quadratic adaptive control. Our analysis rests on a version of the Cramér-Rao inequality that takes into account possible ill-conditioning of Fisher information and a pertubation result on the corresponding singular subspaces. This is used to define a sufficient condition, which we term uniformativeness, for regret to be at least order square root in the samples.

Domaines: Optimisation et contrôle [math.OC]; Machine Learning [stat.ML]

Ingvar Ziemann, Henrik Sandberg. On a Phase Transition of Regret in Linear Quadratic Control: The Memoryless Case. 2020. ⟨hal-02546670v5⟩