Automating The Selection of a Simulation Warm-up Period Stewart Robinson, Katy Hoad, Ruth Davies Warwick Business School Cardiff University - October 2008.

Automating The Selection of a Simulation Warm-up Period

Stewart Robinson, Katy Hoad, Ruth DaviesWarwick Business School

Cardiff University - October 2008

The AutoSimOA Project

A 3 year, EPSRC funded project in collaboration with SIMUL8 Corporation.

http://www.wbs.ac.uk/go/autosimoa

Research Aim

• To create an automated system for advising a non-expert user on how to obtain accurate measures of model performance

• i.e. warm-up, run-length and number of replications

• For implementation into simulation software

AutoSimOA = Automated SIMulation Output Analysis

Simulationmodel

Warm-upanalysis

Run-lengthanalysis

Replicationsanalysis

Use replicationsor long-run?

Recommendationpossible?

Recommend-ation

Output data

AnalyserO

btai

n m

ore

outp

ut d

ata

The Initial Bias Problem

• Model may not start in a “typical” state.• Can cause initial bias in the output.• Many methods proposed for dealing with

initial bias, e.g.: – Initial steady-state conditions– Run model for ‘long’ time– Deletion of initial transient data (‘warm-up

period’)

The Initial Bias Problem

0

20

40

60

80

100

120

140

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Day

Dai

ly t

hrou

ghpu

t

Initial transient Steady-state

f (steady-state)

Steady-statemean

This project uses: Deletion of the initial transient data by specifying a warm-

up period (or truncation point).

The question is:

How do you estimate the length of the warm-up period required?

Methods fall into 5 main types :

1. Graphical Methods.

2. Heuristic Approaches.

3. Statistical Methods.

4. Initialisation Bias Tests.

5. Hybrid Methods.

Literature search – 44 methods

Summary of methods and

literature references on project

web site:

http://www.wbs.ac.uk/go/autosimoa

Short-listing warm-up methods for automation using literature

Short-listing Criteria

» Accuracy & robustness

» Ease of automation

» Generality

» Computer running time

Short-listing results: reasons for rejection of methods

64

3

3

52

2

1

1

0

2

4

6

8

10

12

Graphical Statistical Heuristic

Warm-up method type

Nu

mb

er

of m

eth

od

s

Ease of Automation Accuracy

Generality Excessive running time

Other

• Statistical methods:– Goodness of Fit (GoF) test– Algorithm for a static data set (ASD)– Algorithm for a Dynamic data set (ADD)

• Heuristics:– MSER-5– Kimbler’s Double Exponential Smoothing– Euclidean Distance Method (ED)

Short-listing results: 6 Methods taken forward to testing

Preliminary testing of shortlisted methods

• Rejected methods:– ASD & ADD required a prohibitively large

number of replications– GoF & Kimbler’s method consistently and

severely underestimated truncation point.– ED failed to give any result on majority of

occasions

• MSER-5 most accurate and robust method

MSER-5 warm-up method

0

0.002

0.004

0.006

0.008

0.01

0.012

0.014

0.016

0.018

0.02

0 50 100 150 200 250 300 350 400

Truncation Point

Test

Sta

tistic

0

1

2

3

4

5

6

Batc

h M

eans

MSER-5 test statistic

Output data (batched means values)

Rejection zone

Estimated warm-up period

Estimated truncation point

Further Testing of MSER-5

1. Artificial data – controllable & comparable initial bias functions steady state functions

2. Full factorial design.

3. Set of performance criteria.

Parameters Levels

Data Type Single run

Data averaged over 5 reps

Error type N(1,1), Exp(1)

Auto-correlation

None, AR(1), AR(2), MA(2), AR(4), ARMA(5,5)

Bias Severity 1, 2, 4

Bias Length 0%, 10%, 40%, 100%

Bias direction Positive, Negative

Bias shape 7 shapes

Artificial Data Parameters

• Mean Shift:

• Linear:

• Quadratic:

• Exponential:

• Oscillating (decreasing):

Quadratic ExponentialLinear

Add Initial Bias to Steady state:

Superpostion: Bias Fn, a(t), added onto end of steady state function:

e.g.

2. Full factorial design: 3048 types of artificial data set

MSER-5 run with each type 100 times

...

)(1

etc

taXY

XX

tt

ttt

i. Coverage of true mean.

ii. Closeness of estimated truncation point (Lsol) to true truncation point (L).

iii. Percentage bias removed by truncation.

iv. Analysis of the pattern & frequency of rejections of Lsol (i.e. Lsol > n/2).

3. Performance Criteria

MSER-5 Results

Does the true mean fall into the 95% CI for the estimated mean?

Non-truncated data sets

Truncated data sets

% of cases

yes yes 7.7%

yes no 0%

no no 19.8%

no yes 72.5%

i. Coverage of true mean.

-70

-50

-30

-10

10

30

50

0 20 40 60 80 100run

Lsol -

L

Quadratic bias Mean-shift bias

ii. Closeness of Lsol to L:

• Wide range of Lsol values.

e.g.

(Positive bias functions, single run data, N(1,1) errors, MA(2) auto-correlation, bias severity value of 2 and true L = 100.)

iii. Percentage bias removed by truncation.

0

5

10

15

20

25

300-4

0

40-5

0

50-6

0

60-7

0

70-8

0

80-9

0

90-9

5

95-9

9

99-1

00

100+

% bias removed

% o

f to

tal v

alid

runs

All valid runs

Effect of data parameters on bias removal

No significant effect: Error type

Bias direction

Significant effect: Data type

Auto-correlation type

Bias shape

Bias severity

Bias length

0

50

100

0-4

0

40

-50

50

-60

60

-70

70

-80

80

-90

90

-95

95

-99

99

-10

0

10

0+

% of bias removed

cum

ula

tive

% o

f va

lid c

ase

s Single run

Averaged replications

More bias removed by using averaged replications rather than a single run.

0

50

1000

-40

40

-50

50

-60

60

-70

70

-80

80

-90

90

-95

95

-99

99

-10

0

10

0+

% of bias removed

cu

mu

lative

% o

f va

lid

ca

se

s no a-c AR(1)

AR(2) AR(4)

MA(2) ARMA(5,5)

The stronger the auto-correlation, the less accurate the bias removal.

Effect greatly reduced by using averaged data.

0

50

100

0-4

0

40

-50

50

-60

60

-70

70

-80

80

-90

90

-95

95

-99

99

-10

0

10

0+

% of bias removed

cu

mu

lative

% o

f va

lid

ca

se

s

mean-shift Linear

Quad Exp

OscL OscQ

OscE

The more sharply the initial bias declines, the more likely MSER-5 is to underestimate the warm-up period and to remove increasingly less bias.

0

50

1000

-40

40

-50

50

-60

60

-70

70

-80

80

-90

90

-95

95

-99

99

-10

0

10

0+

% of bias removed

cum

ula

tive

% o

f va

lid c

ase

s 1

2

4

As the bias severity increases, MSER-5 removes an increasingly higher percentage of the bias.

0

50

100

0-4

0

40

-50

50

-60

60

-70

70

-80

80

-90

90

-95

95

-99

99

-10

0

10

0+

reje

ctio

ns

% of bias removed

cum

ula

tive

% o

f va

lid c

ase

s

10%

40%

Longer bias removed slightly more efficiently than shorter bias.

Shorter bias - more overestimations - partly due to longer bias overestimations being more likely to be rejected.

0

100

200

300

400

500

600

700

800

900

x=

0

0<

x≤1

1<

x≤5

5<

x≤1

0

10

<x≤2

0

20

<x≤4

0

40

<x≤6

0

60

<x≤8

0

80

<x≤1

00

x = no. of Lsol rejections

no

. o

f ca

se

s

ARMA(5,5)

MA(2)

AR(4)

AR(2)

AR(1)

No auto-correlation

Rejections caused by: high auto-correlation, bias close to or over n/2, smooth end to data = ‘end point’ rejection.

Averaged data slightly increases probability of getting ‘end point’ rejection but increases probability of more accurate L estimates.

iv. Lsol rejections

0

10

20

30

40

50

1000 1100 1200 1300 1400 1500 1600 1700 1800n

Lso

l re

ject

ion

co

un

t

+ meanshift

+ linear

+ quadratic

+ exp

+ osclinear

+ oscquad

+ oscexp

Giving more data to MSER-5 in an iterative fashion produces a valid Lsol value where previously the Lsol value had been rejected.

e.g. ARMA(5,5)

Lsol values Percentage of cases

Lsol = 0 71%

Lsol ≤ 50 93%

Testing MSER-5 with data that has no initial bias.

Want Lsol = 0

Lsol > 50 mainly due to highest auto-correlated data sets - AR(1) & ARMA(5,5).

Rejected Lsol values: 5.6% of the 2400 Lsol values produced. 93% from the highest auto-correlated data ARMA(5,5).

Testing MSER-5 with data that has 100% bias.

Want 100% rejection rate: Actual rate = 61%

0

1020

30

4050

60

70

8090

100

Line

ar

Qua

d

Exp

Osc

Line

ar

Osc

Qua

d

Osc

Exp

Bias shape

Per

cent

age

of L

sol

reje

ctio

ns

0

10

20

30

40

50

60

70

80

90

M1 M2 M4

Bias severity

Per

cent

age

of L

sol

rej

ectio

ns

Single data Averaged data

Summary

• MSER-5 most promising method for automation– Not model or data type specific. – No estimation of parameters needed. – Can function without user intervention. – Shown to perform robustly and effectively

for the majority of data sets tested. – Quick to run. – Fairly simple to understand.

Heuristic framework around MSER-5

Run k (= 5) replications of length, n ≥ 100

Create averaged

data

Batch data into b batches of length m, where number of

batches = bmn and n* =

b×m ≤ n

MSER-5 returns Lsol value

Produce more data to create

batches of no. orig of %10 or a user specified

number.

Dynamic graph of batched data; single reps, or

MSER-5 statistic

Graph of batched data; single reps,

or MSER-5 statistic with valid Lsol value shown.

Input data into MSER-5 algorithm.

Yes

Yes

No

No

Does User wish to keep running with more data? END

Lsol valid.

Lsol invalid.

Is Lsol ≤ (n* - (m × 5))/2

?

Yes

Have there been 10 invalid Lsol

values in a row?

No

Yes No

Does User wish to keep running with more data?

Produce more data to create

batches of no. orig of %10

Iterative procedure for procuring more data when required.

‘Failsafe’ mechanism - to deal with possibility of data not in steady state; insufficient data provided when highly auto-correlated.

Being implemented in SIMUL8.