Rare-Event Simulation Splitting for Variance Reduction IE 680, Spring 2007 Bryan Pearce.

Rare-Event SimulationSplitting for Variance Reduction

IE 680, Spring 2007

Bryan Pearce

What is a Rare Event?

ΩA

B

Formal Problem Definition

AB

B

1A

0

is estimate tolike would that wemeasurey probabilit then the

"reached, is state first time the" :0inf and

reached" is state first time the" and :0inf If

.\ xstate initialan with and

event), rare thefrom(far and event) rare (the subsetsdisjoint with the

, space state on thechain Markov time-discrete a define 0Let

P

BBXj

AAXAXj

B

AB

,jX

j

jj

j

0. as toincreases

ˆVarˆ estimator MC simple The

21

n

nRE

Splitting: the beginning

• Importance function h– Measures “how close” a state is to the rare

event

• Divide the intermediary state space into m ‘levels’ according to the thresholds l0, l1, …, lm

h(x) = l0 = l1 = l2 = l3 = lm = l

More formally:

events rare' sonot ' of seriesA

...0

levels-sub specifieduser by subdivided is interval this

0on values have willor in not states All

)(:

0)(:

:

:space state theof orderingan allows function Importance specified-userA

10 llll

l

,lxhBAx

lxhxB

xhxA

Rh

h

m

k

How to choose h?

• Defining the importance function can be difficult.

• Ideally our h should reflect:– The most likely path to the rare event

– pk(x) = pk (indep. of state)

– pk = p (indep. of level)

• Presumes apriori knowledge of the system.

First sub-interval

time

h

0

l1

MC Sim N0 independent chains. R0 reach l1.

0

11ˆN

Rp

Second sub-interval: Splitting

time

h

0

l1

MC Sim N1 chains, splitting from the previously achieved threshold states.

R1 reach l2.

1

22ˆN

Rp

l2

…and so on for each sub-interval

Notation

estimator ˆ

levelanother achieving ofy probabilit |

on terminatibefore achieved

achieved is imeearliest t :0inf

,,...,1For

1

1

k

kk

kkk

kAkk

kkjk

N

Rp

DDPp

lTD

llXhjT

mk

m

kk

m

kkm ppDP

11

ˆˆ and

Splitting policy – fixed splitting

• Each chain that reaches level k is cloned ck times.

• Nk will be random for each level k > 0

• Stratified sampling from the entrance distribution of level k

Splitting policy – fixed effort

• Fix Nk in advance. Choose the states represented in the entrance distribution by:

Random assignment– Choose these Nk states randomly from the entrance

distribution

Fixed assignment– Choose an equal quantity of each state– Better stratification

Pros & cons of splitting method

• Fixed splitting – – Asymptotically more efficient under optimal

conditions

– Efficiency very sensitive to splitting factor ck

• Fixed effort– Higher memory requirement– More robust

Efficiency

nn

nnn

ˆ compute to timeexpected theis ˆC where

ˆCˆVar

1ˆEff

Our hope is that splitting will allow our variance to shrink faster than our computational time grows. This has indeed been shown to be true in many cases.

Truncation - Motivation

h

0

l1

l2

l3

l4

Simulation time spent reaching l1

Simple (biased) Truncation

Choose β:

• If a chain falls below the level lk-β then terminate.

• Estimator becomes biased, moreso with small β.

• Large β does not reduce workload very much.

• RESTART

h

0

l1

l2

l3

l4

Terminate

} β = 2

Unbiased Truncation

Use the ‘Russian Roulette’ principle:

The first time a chain ‘down-crosses’ a level threshold it dies with probability (1 – 1/rk,j). If it survives then its weight is increased by a factor of rk,j.

(these rk,j are user-defined and determine the ‘strength’ of the truncation)

How to choose the rk,js

• The selection of the rk,js at each level of the process will control the aggressiveness of the truncation policy.

• A tried-and-true value:

jkjk p

r

ˆ1

,

h

0

l1

l2

l3

l4

Dies with prob. (1 – 1/r3,2)

Weight increases by a factor of r3,2 if the chain survives.

Russian Roulette, cont.

• There are various methods by which to use the chain weights can compensate for this truncation bias.

– Probabilistic

– Tag-based

– Periodic

Truncation w/o weights

• Chain weighting truncation methods can inflate the variance of our gamma estimator.

• We can avoid this problem by allowing our chains to probabilistically re-split upon re-achieving previously achieved goals.

Conclusions and notes

• Potential performance– With γ = 10-20,

Var[MC] = 10-23 while Var[split] = 10-41

• Poorly-behaved systems– Inefficient to apply

References

L’Ecuyer, P., V. Demers, B. Tuffin. 2006. Splitting for rare-event simulation.

Glasserman, P., P. Heidelberger,and T. Zajic. 1998. A large deviations perspective on the efficiency of multilevel splitting.

L’Ecuyer, P., V. Demers, B. Tuffin. 2006. Rare-events, splitting, and quasi-Monte Carlo.

Garvels, M. J. J. 2000. The splitting method in rare event simulation.