Model-based machine diagnostics techniques require the modeled states and machine inputs to be measured. Because measurement of all the states and inputs is not always possible or practical, a simultaneous state and input observer is required. Previous work has developed this type of acausal observer and shown it is susceptible to noise. This paper develops a steady-state optimal observer that minimizes the trace of the steady-state error covariance of the state and input estimates for discrete, linear, time-invariant, stochastic systems with unknown inputs. In addition, a method to distinguish the best measurement set among the available measurement sets is developed. Results from numerical simulations show that the optimal observer can greatly improve estimation results in some cases.