Optimizing the power of action learning