20.11.2012 Views

Contents Telektronikk - Telenor

Contents Telektronikk - Telenor

Contents Telektronikk - Telenor

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.

this method works. Remember that the<br />

problem of simulating a system containing<br />

rare events is that we get a large<br />

number of simulation cycles that e.g.<br />

does not visit state N. The reason is simply<br />

that after leaving state 0 you will most<br />

likely return to this state (from state 1) if<br />

the difference between the upward and<br />

downward transition probabilities are<br />

large (in favour of a downward transition).<br />

If we manipulate the transition<br />

probabilities we can get another<br />

behaviour. Increasing the upward relative<br />

to downward transition probabilities we<br />

increase the frequency of cycles which<br />

are visiting state N (the state of interest,<br />

e.g. defect or full system probability)<br />

before returning to the starting state 0.<br />

Hence, the number of simulation cycles<br />

needed for observing a full system are<br />

reduced. However, these observations are<br />

not stemming from the original process<br />

and will therefore give biased estimates.<br />

Therefore, we have to rescale our observations<br />

by the likelihood ratio, which is<br />

the accumulated ratio between the transition<br />

probabilities of the recorded path in<br />

the original and the new processes. During<br />

the simulation cycle we have to<br />

record this ratio in addition to the quantity<br />

of interest. In Figure 6 you find an<br />

illustration of 4 simulation cycles of the<br />

reference example with only N = 2.<br />

The results produced by application of<br />

importance sampling is reported to be<br />

very sensitive to how much the underlying<br />

distribution is changed [11,34], e.g.<br />

the transition probabilities. To the author’s<br />

knowledge, optimal biasing only<br />

exist for very limited “classes” of systems,<br />

such as the reference system. For<br />

one-dimensional models like the reference<br />

example, you find in the literature<br />

application of either<br />

- balanced biasing, i.e. the transition<br />

probabilities are similar and equal to<br />

0.5, or<br />

- reversed biasing, i.e. upward and<br />

downward probabilities are interchanged.<br />

In the reference example, the latter is the<br />

optimal biasing [29,35].<br />

In dependability simulation the change in<br />

parameters is concerning an increase in<br />

the probabilities of failure, basically not<br />

very far from the ideas behind traditional<br />

accelerated lifetime testing of HW components,<br />

see e.g. [36]. In the literature we<br />

find a number of approaches where IS is<br />

applied, and studies are done to try to<br />

find an optimal, or at least a “good” way<br />

p 01 =1 p 12 =0.1<br />

0 1 2<br />

p 10 =0.9 p 21 =1<br />

Original system<br />

Figure 6 Simulation cycles with importance sampling<br />

Full model 0 1 2<br />

Estimating the<br />

probability of<br />

intermediate state<br />

(here: state 2)<br />

RESTART subchain, starting from the<br />

intermediate point (here: state 2)<br />

of changing the parameters, such as failure<br />

path estimation [34], minimise<br />

empirical variance [37], balanced failure<br />

biasing [38], failure distance approach<br />

where the parameters are optimised during<br />

simulations [39].<br />

Similarly, in a teletraffic performance<br />

study of e.g. the cell loss rate in an ATM<br />

queue, the parameter change is due to an<br />

increase in the offered load to the queue. 6<br />

Generally, of course we have the same<br />

problems in teletraffic as in dependability<br />

of finding an optimal change in parameters.<br />

But, in the specific case of our<br />

reference example, Parekh and Walrand<br />

[29] reported that reversed biasing is an<br />

optimal biasing where Rare event theory<br />

6 Note that the theory of IS restricts us<br />

to either increase the load per customer,<br />

not the number of customers, or<br />

increase the mean service time per<br />

server, not to decrease the number of<br />

servers.<br />

λ<br />

p* 01 =1 p* 12 =0.5<br />

0 1 2<br />

p* 10 =0.5 p* 21 =1<br />

Biased (stressed) system<br />

{using balanced biasing}<br />

λ λ λ<br />

µ µ µ µ<br />

λ<br />

0 1 2<br />

state I<br />

λ λ λ<br />

µ µ µ µ<br />

Figure 7 The RESTART method on the reference example<br />

2<br />

λ λ<br />

µ µ<br />

is the theoretical fundament [35]. The<br />

results from Section 5.4 show an significant<br />

increase in the speed-up gain using<br />

reversed instead of balanced biasing.<br />

5.3.3 RESTART<br />

Another importance sampling concept is<br />

the RESTART (REpetitive Simulation<br />

Trials After Reaching Thresholds), first<br />

introduced at ITC’13 by Villén-Altamirano<br />

[17]. The basic idea is to sample the<br />

rare event from a reduced state space<br />

where this event is less rare. The probability<br />

of this reduced state space is then<br />

estimated by direct simulation, and by<br />

Bayes formulae the final estimate is<br />

established.<br />

Consider the model of our reference<br />

example, the rare events are visits to the<br />

state of full system of system failure<br />

(state N) before returning to the initial<br />

state 0. This means that during a direct<br />

simulation experiment, a typical<br />

sequence of event is visits to states near<br />

the state 0 (the regenerative state). The<br />

RESTART reasons as follows: Say that<br />

N<br />

N<br />

N<br />

201

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

Saved successfully!

Ooh no, something went wrong!