following form g1: g1ðt;xðtÞÞ¼ðaðtÞ0:2Þð3aðtÞÞ ð31Þ Dynamic programming leads to the strategy uðÞ that optimizes reliabilityat any horizon, and the only approximation is that of the discretization. This is showcased for T¼100 by Fig.1, which shows that only initial states grouped around b0 ¼0 have a good reliability. There is, however, a sizable reliability kernel Rel (0.95,100), which shows which initial states – or which system designs – lead to the lowest probability of failure. Such reliability kernels can also be computed at any horizon, which allows for observing the evolution of Rel(0.95,T) asT increases. Its size decreases very little until it abruptly ceases to exist when T tops 254 (Fig. 2). This stability of the reliability kernel Rel(0.95,T) as the horizon increases is matched by that of the optimal strategy. Whatever the horizon, the backward sequence of feedback maps x↦uðt;xÞ from the final date T to the initial date is the same, and what is more, the map becomes constant for trT10. It is noted un and represented in Fig. 3. For a given value of a, the value of un increases as b increases. Yet the relationship between un and b is different for each single value of a, so that the map is very complex. This map has been obtained through the use of dynamic programming, and it is important to recall that usual time-variant reliability kernel are not devised to yield such complex main- tenance strategies. Outcrossing rates as approximated by Pðtout ¼tÞ=Δt can be estimated using this feedback map. SinceΔt¼1, using Eq. (28) the outcrossing rate takes the approximate value of ðλð1pfðt;x0; unðÞÞÞ after less than 10 time steps, where λ2104 is the probability of leaving at t conditional on staying in the survival set up to t1. λ is independent on the initial state, so that the differences in reliability displayed in Fig. 1 account for the probability of leaving the survival set within these first ten time steps. After t¼10, the probability of failure increases very slowly.