Section 8.

3: FiniteHorizon and InniteHorizon Statistics

Discrete-Event Simulation: A First Course
c 2006 Pearson Ed., Inc. 0-13-142917-5

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

1/ 20

Section 8.3: Finite-Horizon and Innite-Horizon Statistics

Denition 8.3.1: Steady-state statistics Produced by simulating the operation of a stationary discrete-event system for an innite length of time Steady-state statistics may not exist

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

2/ 20

Example 8.3.1

Use sis4 to simulate dierent numbers of time intervals, compute average inventory level l for each Output from sis4 for an increasing number of time intervals n n: 20 40 80 160 320 640 1280 2560 l : 25.98 26.09 25.49 27.24 26.96 26.36 27.19 26.75 As n , l seems to converge to 26.75

A better approach is to use interval estimation for the steady-state l , using large but nite n One way is to use batch means Section 8.4

Can only approximate n convergence

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

3/ 20

Example 8.3.2

Use ssq4 to simulate dierent numbers of jobs n

Compute w , l

Use an arrival rate of 1/2 Change Erlang parameters so service rate is 1/2 or less As n increases, w and l will tend to increase, without limit Why? The server cannot keep up with demand, so the queue will grow as more jobs arrive In this case, the steady-state average wait and the average number in the node are both innite

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

4/ 20

Finite-Horizon Versus Innite-Horizon Statistics

Finite horizon A nite-horizon discrete-event simulation is one for which the simulated operational time is nite Innite horizon An innite-horizon discrete-event simulation is one for which the simulated operational time is eectively innite Transient system statistics are those statistics that are produced by a nite-horizon discrete-event simulation Steady-state statistics are produced by an innite-horizon simulation The initial conditions aect nite-horizon statistics The initial conditions do not aect innite-horizon statistics: after enough time, the system loses memory of its initial state
Discrete-Event Simulation: A First Course Section 8.3: FiniteHorizon and InniteHorizon Statistics 5/ 20

Example 8.3.3
With minor modications, program ssq4 can be used to simulate an initially idle M /M /1 service node processing a small number of jobs (say 100) and with relatively high trac intensity (say 0.8) If the program is executed multiple times varying only the rngs initial seed from replication to replication,
the average wait in the node will vary signicantly for most replications, the average wait will not be close to the steady-state average wait

If a relatively large number of jobs (say 10000) are used, variability of the average wait will be much less signicant and be close to the steady-state average number When the number of jobs becomes innite, the initial condition bias disappears

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

6/ 20

Another Important Distinction

In an innite-horizon simulation, the system environment is assumed to remain static If the system is a single-server service node, both the arrival rate and the service rate are assumed to remain constant in time In a nite-horizon simulation, no need to assume a static environment

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

7/ 20

Relative Importance of Two Statistics

The traditional view: steady-state statistics are most important

Steady-state statistics are better understood because they are much more easy to analyze mathematically It is frequently dicult to accurately model initial conditions and non-stationary system parameters

The pragmatic view: transient statistics are most important because steady-state is just a convenient ction Depending on the application, both transient and steady-state statistics may be important Important to decide which statistics best characterize the systems performance

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

8/ 20

Initial and Terminal Conditions

Finite-horizon discrete-event simulations are also known as terminating simulations

In program ssq4, the system state is idle at the beginning and at the end of the simulation The terminal condition is specied by the close the door time The system state of sis4 is the current and on-order inventory levels; these states are the same at the beginning and at the end of the simulation The terminal condition is specied by the number of time intervals

Innite-horizon discrete-event simulations must be terminated; typically done using whatever stopping conditions are most convenient

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

9/ 20

Formal Representation
The state variable X () is known formally as a stochastic process
The typical objective of a nite-horizon simulation of this system would be to estimate the time-averaged transient statistic ( ) = 1 X X (t )dt 0 where > 0 is the terminal time The typical objective of an innite-horizon simulation of this sytem would be to estimate the time-averaged steady-state statistics ( ) = lim 1 x = lim X

X (t )dt
0

( ) is a random variable; x Note that X is not

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

10/ 20

Denition 8.3.3

If a discrete-event simulation is repeated, varying only the rngs initial states from run to run, each run of the simulation program is a replication and the totality of replications is said to be ensemble Replications are used to generate estimates of the same transient statistic The initial seed for each replication should be chosen to be no replication-to-replication overlap The standard way is to use the nal state of each rngs stream from one replication as the initial state for the next replication accomplished by calling PlantSeeds once outside the main replication loop

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

11/ 20

Independent Replications and Interval Estimation

Suppose the nite-horizon simulation is replicated n times, each time generating a state time history xi (t ) x i ( ) = 1
0

xi (t )dt

where i = 1, 2, . . . , n is the replication index Each data point x i ( ) is an independent observation of the ( ) random variable X ( ) can be estimated from a If n is large enough, the pdf of X histogram of the x i ( )

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

12/ 20

Independent Replications and Interval Estimation (2)

( )] point estimate is available as an ensemble Want E [X average n 1 x i ( ) n

i =1

( )] can be calculated An interval estimate for E [X

Use the interval estimation technique from Section 8.1 Requires ensemble mean and standard deviation of x i ( )

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

13/ 20

Example 8.3.5

A modied version of sis4 was used to produce 21 replications 100 time intervals of operation were simulated Inventory parameters were (s , S ) = (20, 80). ( ), the time-averaged inventory level Measured L ( ) Each of these 21 numbers is a realization of L

23
| | | | | | |

27.06 0.82
| | | || ||

31
| | || | | |

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

14/ 20

Example 8.3.5, continued

The mean and standard deviation are 27.06 and 1.76 respectively If a 95% condence interval estimate is desired, t = idfStudent(20, 0.975) = 2.086 The 95% condence interval estimate is 27.06 (2.086)(1.76) = 27.06 0.82 20

( )] is within interval We are 95% condent that E [L 27.06 0.82

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

15/ 20

Example 8.3.6

A modied version of ssq2 was used to produce 20 replications 100 jobs processed through M /M /1 service node
Node is initially idle Arrival rate is = 1.0 Service rate is = 1.25

The resulting 20 observations of the average wait in the node:

1
| || | | || |

3.33 1.04 | | | | | |

12
| | | |

From Algorithm 8.1.1, the resulting 95% condence interval estimate is 3.33 1.04

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

16/ 20

Example 8.3.6, continued

Steady-state average wait for M /M /1 is 1 (See Section 8.5) In our example

Arrival rate is = 1.00 Service rate is = 1.25 1 Average wait is 1.25 1.00 = 4.0

It is possible that 100 jobs are enough to produce steady-state statistics

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

17/ 20

Example 8.3.7

The modied version of program ssq2 was used to produce 60 more replications Consistent with n rule, expect two-fold decrease in the width of the interval estimate Based on 80 replications, the resulting 95% condence interval estimate was 3.25 0.39
1
|| || |||| | || ||| | || ||

3.25 0.39 | | | || ||||| | | || || || || | || ||| | | | |

12
| | || | | | ||| | | | | |

In this case 100 jobs are not enough to produce steady-state statistics

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

18/ 20

Example 8.3.8

As a continuation of Example 8.3.6, the number of jobs per replication was increased from 100 to 1000 20 replications were used to produce 20 observations of the average wait in the node
1
||

3.82 0.37 | | | | | | | | | | || | |

12
| |

Relative to Example 8.3.6, much more symmetric sample mean in Example 8.3.8
1
| || | | || |

3.33 1.04 | | | | | |

12
| | | |

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

19/ 20

Example 8.3.8, continued

The 1000-jobs per replication results are more consistent with the underlying theory of interval estimation
Requires a sample mean distribution that is approximately Normal(, / n) Sample mean distribution is centered on (unknown) population mean

1000 jobs may achieve steady-state; 100 jobs cannot More on this issue in Section 8.4

Discrete-Event Simulation: A First Course

Section 8.3: FiniteHorizon and InniteHorizon Statistics

20/ 20