This article studies the design of optimal monetary policy under uncertainty using a Markov jump-linear-quadratic (MJLQ) approach. To approximate the uncertainty that policymakers face, the authors use different discrete modes in a Markov chain and take mode-dependent linear-quadratic approximations of the underlying model. This allows the authors to apply a powerful methodology with convenient solution algorithms that they have developed. They apply their methods to analyze the effects of uncertainty and potential gains from experimentation for two sources of uncertainty in the New Keynesian Phillips curve. The examples highlight that learning may have sizable effects on losses and, although it is generally beneficial, it need not always be so. The experimentation component typically has little effect and in some cases it can lead to attenuation of policy.