A new optimal stochastic control problem is posed. The criterion is to maximize the mean-time-between-failures criterion given a certain reference value. It is shown that this control problem is closely related both to the problem of minimizing the variance of the signal—minimum variance control — and to the problem of minimizing the so-called upcrossing probability—minimum upcrossing control. It is made plausible that the upcrossing probability is a better approximating criterion to minimize than the variance criterion. The problem of minimizing the upcrossing probability can be thought of as finding optimal weighting-matrices in an LQG-problem. The new controller is compared with the minimum variance controller for a first-order process. It is seen that the new controller causes a lower upcrossing probability and a larger mean time between failures. The improvement in the example is up to about 25%. This makes it possible to choose the reference value closer to the critical level without causing smaller...