Almost Optimal Agnostic Control of Unknown Linear Dynamics

We consider a simple control problem in which the underlying dynamics depend on a parameter a that is unknown and must be learned. We study three variants of the control problem: Bayesian control, in which we have a prior belief about a; bounded agnostic control, in which we have no prior belief about a but we assume that a belongs to a bounded set; and fully agnostic control, in which a is allowed to be an arbitrary real number about which we have no prior belief. In the Bayesian variant, a control strategy is optimal if it minimizes a certain expected cost. In the agnostic variants, a control strategy is optimal if it minimizes a quantity called the worst-case regret. For the Bayesian and bounded agnostic variants above, we produce optimal control strategies. For the fully agnostic variant, we produce almost optimal control strategies, i.e., for any varepsilon>0 we produce a strategy that minimizes the worst-case regret to within a multiplicative factor of (1+varepsilon).