5.3. Decreasing carrying capacity
Let us now suppose that the system performance decreases over time. We choose a simple model of linear decrease from [22] with a diminishing carrying capacity cðtÞ¼30:01t. Let us note
this function g2, Eq. (30) becomes g2ðt;xðtÞÞ¼ðaðtÞ0:2Þð30:01taðtÞÞ ð32Þ As expected, this linear decrease in performance affects relia- bility, so that Rel(0.95,T), even though it assumes a similar shape as for g1, vanishes for T454 (Fig. 4). Besides, unlike for the case of a constant carrying capacity, the optimal control maps change at each time step. This would make them very difficult to find if it were not for dynamic programming. Yet, for trT10 it seems that the map unðt;: Þdoes not depend on the horizon T. Thus, the maps for T¼100 and T¼200 are identical until the date t¼92, while those for T¼150 and T¼200 are identical until t¼145. This makes the computation of the outcrossing rates values computed with the optimal strategy for T¼200 applicable to lower time horizons. No matter the initial state, the outcrossing rate is low after the first ten time steps then gradually increases to peak at t¼124 (Fig. 5). Then, it decreases because the decreasing quantity ð1pfðt;x0;unðÞÞ in Eq. (28) compensates the growth of the probability of leaving the survival set at t conditional on staying in it until t1. Like for the previous case, the amplitude of the outcrossing rate after t¼10 depends on the odds of leaving the survival set within the first few time steps. The cumulative probability of failure through time can be com- puted alongside the outcrossing rate (Fig. 6).
6. Discussion
The limitations of dynamic programming algorithms should be kept in mind. In practice, they can only solve systems which state space has a low dimension. Too high a dimension leads to the so- called “curse of dimensionality” which designates the exponential increase of the needed computational time and memory. There exist approximation and decomposition algorithms that have been used to deal with the dimension problem in dynamic program- ming, such as the Benders decomposition [39] or dual approxima- tions, e.g. [40], but their applicability lies well outside the scope of this work. Yet, the confrontation of reliability with control theory looks very promising. In this paper, it led to the formal definition of coupled design and maintenance problems without restrictive hypotheses on the aging of the system with time, nor on the
5.3. Decreasing carrying capacity
Let us now suppose that the system performance decreases over time. We choose a simple model of linear decrease from [22] with a diminishing carrying capacity cðtÞ¼30:01t. Let us note
this function g2, Eq. (30) becomes g2ðt;xðtÞÞ¼ðaðtÞ0:2Þð30:01taðtÞÞ ð32Þ As expected, this linear decrease in performance affects relia- bility, so that Rel(0.95,T), even though it assumes a similar shape as for g1, vanishes for T454 (Fig. 4). Besides, unlike for the case of a constant carrying capacity, the optimal control maps change at each time step. This would make them very difficult to find if it were not for dynamic programming. Yet, for trT10 it seems that the map unðt;: Þdoes not depend on the horizon T. Thus, the maps for T¼100 and T¼200 are identical until the date t¼92, while those for T¼150 and T¼200 are identical until t¼145. This makes the computation of the outcrossing rates values computed with the optimal strategy for T¼200 applicable to lower time horizons. No matter the initial state, the outcrossing rate is low after the first ten time steps then gradually increases to peak at t¼124 (Fig. 5). Then, it decreases because the decreasing quantity ð1pfðt;x0;unðÞÞ in Eq. (28) compensates the growth of the probability of leaving the survival set at t conditional on staying in it until t1. Like for the previous case, the amplitude of the outcrossing rate after t¼10 depends on the odds of leaving the survival set within the first few time steps. The cumulative probability of failure through time can be com- puted alongside the outcrossing rate (Fig. 6).
6. Discussion
The limitations of dynamic programming algorithms should be kept in mind. In practice, they can only solve systems which state space has a low dimension. Too high a dimension leads to the so- called “curse of dimensionality” which designates the exponential increase of the needed computational time and memory. There exist approximation and decomposition algorithms that have been used to deal with the dimension problem in dynamic program- ming, such as the Benders decomposition [39] or dual approxima- tions, e.g. [40], but their applicability lies well outside the scope of this work. Yet, the confrontation of reliability with control theory looks very promising. In this paper, it led to the formal definition of coupled design and maintenance problems without restrictive hypotheses on the aging of the system with time, nor on the
การแปล กรุณารอสักครู่..
