Optimal agnostic control of unknown linear dynamics in a bounded parameter range (Q7006400)

This paper investigates a control problem where the underlying dynamics depend on an unknown parameter, which is bounded within a known range. The authors explore two control strategies: Bayesian control, which utilizes a prior probability distribution for the unknown parameter, and agnostic control, which does not assume any prior knowledge. The paper presents optimal strategies for both approaches, demonstrating that the agnostic control problem can be reduced to a Bayesian control problem under certain conditions. The authors also discuss the mathematical challenges involved in minimizing regret over a fixed time horizon, particularly in scenarios where the parameter can lead to unstable dynamics.