Professional Documents
Culture Documents
Rare Events Occur in many realworld systems, may cause serious consequences, Occur with extremely small probability (e.g. p < 109) Difficult to simulate. Reliable statistics require sufficiently large number of observations: for probability 109 on average 109 trials for one observation.
Rare Events in Computer and Communication Systems Bit errors in digital communications Component and system failures (fault tolerant systems, reliability, availability) Queuing Systems
Generate samples X1,X2, . . . ,XN, IID according to distribution P. Calculate sample mean and variance
1 = N
I
i =1
(Xi ) =
2 ( ) =
I2
N
(1 )
N
Confidence interval
z1 / 2
(1 )
N
0
1 (1 ) 1 Relative error: = N N
z1 / 2
(1 )
N
< N
z12 / 2 1
Numerical example:
99% confidence interval 1 = 0.01 and z1/2 = 2.576 maximum relative half-width of 10%; =0.1
N 100 2.576
2
Simulation Speed-Up
Simulation = statistical estimation. Simulation speed-up: significantly reduce the time to get statistical estimates of desired accuracy (confidence interval half-width, relative error) Two ways:
Make more experiments in same time Generate more rare events during simulation, in same time increase the frequency of the rare event of interest (Rare Events Provoking)
Importance Sampling Importance Splitting (RESTART)
Objective: determine the area of region C Generate N IID samples (uniformly distributed over the entire space A) Estimate the area of C as SC = NC/N, where NC is the number of hits within region C The variance of this estimate is inversely proportional to N, while the precision of the estimate is related to the number of hits NC in the important region
Bias the sampling procedure to increase the fraction of samples that result in hits. Double the probability that samples are generated in the quadrant D, containing C The average number of samples in region C is doubled, increasing the estimator precision Each sample which results in a hit in region D must be weighted by a factor of 1/2 to obtain the correct, unbiased result
where L(x) = f(x)/f*(x) is denoted the likelihood ratio. Observe that the expected value of the observations under f is equal to the expected value of the observations under f* corrected for bias by the likelihood ratio,
f ( x1 ) = L( x1 ) ** f f ((x) x1 )
y x1
x x
with variance
Rare event A2
Birth-death process with P(A2) P(A1) 1 Further: P(A2) P(A2|A1) and P(A1|A2) ~ 1
Restarting the simulation in valid states of the system that provoke the events of interest more often than in a straightforward simulation, thus leading to a shorter simulation run time. Determine P(A1) with sufficient accuracy Determine P(A2|A1) with sufficient accuracy by repeated simulation runs starting from the state A1 Target value P(A2) = P(A2|A1) P(A1)
One-level Restart
Random variable X over simulation time Threshold at I, rare event at B P{ x B} = P{x B | x I} P{x I}
Single Step vs. Global Step Single Step approach : complete all the paths starting from a certain stage before we move to the next one. Global Step approach : generate all the offspring originating from a single root (i.e. a state in the upper stage which is reached immediately from the lower state) before we move to the next root.