The paper considers the merits of rules and discretion for monetary policy when the structure of the macroeconomic model and the probability distributions of disturbances are not well defined. It is argued that when it is costly to delay policy reactions to seldom-experienced shocks until formal algorithmic learning has been accomplished, and when time consistency problems are significant, a mixed strategy that combines a simple verifiable rule with discretion is attractive. The paper also discusses mechanisms for mitigating credibility problems and emphasizes that arguments against various types of simple rules lose their force under a mixed strategy.