Monte-Carlo Methods for Estimating System Reliability

Michael G. Luby

EECS Department, University of California, Berkeley

Technical Report No. UCB/CSD-84-168

, 1984

http://www2.eecs.berkeley.edu/Pubs/TechRpts/1984/CSD-84-168.pdf

An n-component system contains n components, where each component may be either failing or working. Each component i is failing with probability Pi independently of the other components in the system. A system state is a specification of the states of the n components. Let F, the set of failure states, be a specified subset of all system states. In this paper we develop Monte Carlo algorithms to estimate Pr[F], the probability that the system is in a failure state. We now describe the two different formats for the representation of F considered in this paper. In the first format F is represented by m failure sets F1, F2, . . . , Fm. Each failure set is specified by an n-tuple in the input. The set of failure states is then F=m,t=1Fi. We develop a formula for the probability of a union of events. A Monte Carlo algorithm which estimates the failure probability of an n-component system is developed based on this formula. One trial of the algorithm outputs an unbiased estimator of Pr[F]. Let Y be the average of the estimators produced by many trials of the algorithm. We show that, when the algorithm is run for an amount of time proportional to nm, Y is provably close to Pr[F] with high probability. The second format for the representation of F can be described as follows. A network is an undirected graph G, where the edges in the graph correspond to the components in the system. Let G be a planar network and let x1, . . . , xK be K specified nodes in G. For the planar K-terminal problem, the network is in a failing state if there is no path of working edges between some pair of specified nodes. We develop a Monte Carlo algorithm to estimate Pr[F] for the planar K-terminal problem. The algorithm works especially well when the edge failure probabilities are small. In this case the algorithm produces an estimator Y which is provably close to Pr[F] with high probability in time polynomial in the size of the graph. This compares very favorably with the execution times of other methods used for solving this problem.

Advisors: Richard M. Karp

BibTeX citation:

@phdthesis{Luby:CSD-84-168,
    Author= {Luby, Michael G.},
    Title= {Monte-Carlo Methods for Estimating System Reliability},
    School= {EECS Department, University of California, Berkeley},
    Year= {1984},
    Month= {Jun},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/1984/5978.html},
    Number= {UCB/CSD-84-168},
    Abstract= {An <i>n</i>-component system contains <i>n</i> components, where each component may be either failing or working. Each component <i>i</i> is failing with probability <i>Pi</i> independently of the other components in the system. A system state is a specification of the states of the <i>n</i> components. Let <i>F</i>, the set of failure states, be a specified subset of all system states. In this paper we develop Monte Carlo algorithms to estimate <i>Pr</i>[<i>F</i>], the probability that the system is in a failure state.  <p>  We now describe the two different formats for the representation of <i>F</i> considered in this paper. In the first format <i>F</i> is represented by <i>m</i> failure sets <i>F1</i>, <i>F2</i>, . . . , <i>Fm</i>. Each failure set is specified by an <i>n</i>-tuple in the input. The set of failure states is then <i>F</i>=m,t=1<i>Fi</i>. We develop a formula for the probability of a union of events. A Monte Carlo algorithm which estimates the failure probability of an <i>n</i>-component system is developed based on this formula. One trial of the algorithm outputs an unbiased estimator of <i>Pr</i>[<i>F</i>]. Let <i>Y</i> be the average of the estimators produced by many trials of the algorithm. We show that, when the algorithm is run for an amount of time proportional to <i>nm</i>, <i>Y</i> is provably close to <i>Pr</i>[<i>F</i>] with high probability.  <p>  The second format for the representation of <i>F</i> can be described as follows. A network is an undirected graph <i>G</i>, where the edges in the graph correspond to the components in the system. Let <i>G</i> be a planar network and let <i>x1</i>, . . . , <i>xK</i> be <i>K</i> specified nodes in <i>G</i>. For the planar <i>K</i>-terminal problem, the network is in a failing state if there is no path of working edges between some pair of specified nodes. We develop a Monte Carlo algorithm to estimate <i>Pr</i>[<i>F</i>] for the planar <i>K</i>-terminal problem. The algorithm works especially well when the edge failure probabilities are small. In this case the algorithm produces an estimator <i>Y</i> which is provably close to <i>Pr</i>[<i>F</i>] with high probability in time polynomial in the size of the graph. This compares very favorably with the execution times of other methods used for solving this problem.},
}

EndNote citation:

%0 Thesis
%A Luby, Michael G. 
%T Monte-Carlo Methods for Estimating System Reliability
%I EECS Department, University of California, Berkeley
%D 1984
%@ UCB/CSD-84-168
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/1984/5978.html
%F Luby:CSD-84-168