17.01.2014 Views

Evaluating dependability metrics of critical systems: Monte ... - iaria

Evaluating dependability metrics of critical systems: Monte ... - iaria

Evaluating dependability metrics of critical systems: Monte ... - iaria

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Adaptive stochastic approximation (ASA)<br />

ASA just uses a single sample path (y 0 ,... ,y n ).<br />

Initial distribution for y 0 , matrix ˜P (0) and guess µ (0) (·).<br />

At step j <strong>of</strong> the path, if y j ∉ ∆,<br />

◮ matrix ˜P (j) used to generate y j+1 .<br />

◮ From yj+1 , update the estimate <strong>of</strong> µ(y j ) by<br />

µ (j+1) (y j ) = (1 − a j (y j ))µ (j) (y j )<br />

+<br />

h<br />

i<br />

a j (y j ) c(y j ,y j+1 ) + µ (j) P(yj ,y j+1 )<br />

(y j+1 )<br />

˜P (j) (y j ,y j+1 ) ,<br />

where {a j (y), j ≥ 0}, sequence <strong>of</strong> step sizes<br />

◮ For δ > 0 constant,<br />

!<br />

ˆc(yj<br />

˜P (j+1) , y j+1 ) + µ (j+1) (y j+1 )˜<br />

(y j , y j+1 ) = max P(y j ,y j+1 )<br />

µ (j+1) , δ .<br />

(y j )<br />

◮ Otherwise µ (j+1) (y) = µ (j) (y), ˜P (j+1) (y, z) = P (j) (y, z).<br />

◮ Normalize: P (j+1) (y j , y) =<br />

˜P (j+1) (y j ,y)<br />

Pz ˜P (j+1) (y j ,z) .<br />

If y j ∈ ∆, y j+1 generated from initial distribution, but estimations <strong>of</strong><br />

P(·, ·) and µ(·) kept.<br />

Batching techniques used to get a confidence interval.<br />

G. Rubino (INRIA) <strong>Monte</strong> Carlo DEPEND 2010 46 / 72

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!