Since π is fixed beforehand and does not depend on time, it is given by the initial state y0 ¼yð0Þ. As for Eq. (12), for a given initial state y0, a given strategy uðÞ and a given scenario wðÞ, there is a unique trajectory yðy0;uðÞ;wðÞÞ¼ðyð0Þ;yð1Þ;…;yðT1Þ;yðTÞÞ. With these notations, and since probabilities then take into account all the scenarios, we can then write the probability of failure as a function of t, y0 and uðÞ: pfðt;y0;uðÞÞ¼P½(τA½0;t;Xðt;πÞAFðτÞjy0;uðÞ ð19Þ The search of reliable designs can be turned into that of initial states y0 such that there is a control strategy uðÞwhich guarantees that the system is reliable with a confidence level β during the planning period. The corresponding reliability kernel, derived from Eq. (14), is simply noted Relðβ;TÞ (instead of Relπ;uðβ;TÞ). It is the following set: Relðβ;TÞ¼fy0AYj(uðÞAUðTÞ;pfðT;y0;uðÞÞr1βgð 20Þ where the set Y is the state space. In order to compute this kernel, let us now relate it to a mathematical object from stochastic viability theory, the stochastic viability kernel.
4.3. The stochastic viability kernel
Similar to reliability, stochastic viability theory [34,1] focuses on the probability for a system to stay in the survival set during a given time frame. In discrete time, it focuses on the time evolution of a state vector, through a governing equation that is none other than Eq. (16). Stochastic viability assumes that the state vector y(t) is known at each time step, and that given y(t), the probability of being in the survival set at t is also known. Thus, stochastic viability focuses on a very similar problem to that described in Section 4.2. One of its central concepts is the so- called stochastic viability kernel, which importance comes from the original deterministic control framework of viability theory (for a quick overview of viability theory and the viability kernel, see Appendix A). It is defined as the setof all states for which there is a strategy such that the system has a probability β or higher of staying in the survival set S(t) for a given time horizon T. It can be formally defined by the following equation in which it is noted Viab ðβ;TÞ: Viabðβ;TÞ¼fy0AYj(uðÞAUðTÞ;Pð8tA½0;T;Xðt;πÞASðtÞjy0;uðÞÞZβg ð21Þ Stochastic viability is related to the closed-loop reliability problem of Section 4.2 through the remark that: pfðT;y0;uðÞÞ¼1Pð8tA½0;T;Xðt;πÞASðtÞjy0;uðÞÞ ð22Þ Through Eqs. (20) and (21), the stochastic viability kernel is the reliability kernel of Eq. (20): Relðβ;TÞ¼Viabðβ;TÞð 23Þ Yet, by itself, Eq. (23) does not allow for the computation of the reliability kernel Rel ðβ;TÞ. Its interest comes from the fact that there exists a dynamic programming algorithm to compute the stochastic viability kernel.
4.4. A dynamic programming solution
In this section, we are in the Markovian case, meaning that all the w(t) of Eq. (16) are statistically independent from each other. Then, Doyen and De Lara [1] establish that the problem of finding the stochastic viability kernel can be solved by dynamic program- ming, a widespread category of recursive algorithms designed to solve the problem backwards from date T to the initial date. Thus, dynamic programming also allows for solving the reliability– viability problem in the Markovian case.