Using Multiple Runs in the Simulation of Stochastic Systems for
Estimating Equilibrium Expectations
Winfried Grassmann
Department of Computer Science, University of Saskatchewan, Saskatoon, SK S7N 5C9, Canada
Keywords:
Steady-state Simulation, Multiple Runs, Initialization Bias.
Abstract:
In complex stochastic systems, Monte-Carlo simulation is often the only way to estimate equilibrium expec-
tations. The question then arises what is better: a single run of length T, or n runs, each of length T/n. In this
paper, it is argued that if there is a good state to start the simulation in, multiple runs may be advantageous.
To illustrate this, we use numerical examples. These examples are obtained by using deterministic methods,
that is, methods based on probability theory not using Monte-Carlo methods. The results of our numerical
calculations forced us to make a sharp distinction between the time to reach equilibrium and the appropriate
length of the warm-up period, and this distinguishes our study from earlier investigations.
1 INTRODUCTION
Monte Carlo methods are often the only way to es-
timate equilibrium expectations in complex stochas-
tic systems, such as queueing networks. The ques-
tion then arises whether a single long run is preferable
to several shorter runs. Hence, instead of doing one
run of length T, should we do n runs, each of length
T/n? This is a question that has been addressed by
several authors, including (Whitt, 1991), (Alexopou-
los and Goldman, 2004) and (Kelton, 1989). In con-
trast to these studies, we make a sharp distinction be-
tween the time to reach equilibrium within a given
tolerance, and the appropriate length of the warm-up
period. Here, the warm-up period is a period during
which the data obtained by the simulation is ignored.
We were forced to do this because in our numerical
experiments, the time to reach the equilibrium within
±10% was clearly much longer than the warm-up pe-
riod that minimizes the mean squared error of the esti-
mate as shown in (Grassmann, 2011). The problem of
multiple runs is also related to the initialization bias
problem, for details see (Pasupathy and Schmeiser,
2010), (Pawlikowski, 1990) and (Grassmann, 2014),
and references therein.
Our investigation was motivated by a paper by
Madansky (Madansky, 1976), who showed that when
estimating the expected number of elements in an
M/M/1 queue, the best starting state, judging by the
mean squared error (MSE), is the empty state as long
as T is not too short, a result confirmed by (Wilson
and Prisker, 1978) and (Grassmann, 2008). More-
over, in (Grassmann, 2008) it was shown that if start-
ing the M/M/1 queue empty, any warm-up period in-
creases the MSE, that is, the optimal warm-up period
is 0. However, when starting at 0, the expected queue
length is far from its equilibrium value, that is, the
time to reach equilibrium at a tolerance of ± 10%
is far from 0. This observation forces us to make
a sharp distinction between the time to reach equi-
librium within a given tolerance and the appropriate
length of a warm-up period, a distinction that is some-
what hazy in literature. As it turns out, the distinction
between the time to reach equilibrium and the optimal
length of the warm-up period forces us to re-evaluate
the issue of multiple runs.
People may argue that the M/M/1 queue is spe-
cial, and results derived from its study are invalid.
This opinion leads to the philosophical issue as to
how many counter-examples one needs in order to re-
fute strongly held beliefs. Even though in mathemat-
ics, a single counter-example is sufficient, we believe
more than one example is needed. However, (Grass-
mann, 2014) provides additional cases where it is cru-
cial to distinguish between the time to reach equilib-
rium within a giventolerance and the warm-up period.
Moreover, I feel that before making any conclusions
about the generality of an observation, we have to un-
derstand the reasons why the time to reach equilib-
rium within a given tolerance is not necessarily a good
indicator for the length of the warm-up period. For-
tunately, relevant reasons are given in (Grassmann,
138
Grassmann W..
Using Multiple Runs in the Simulation of Stochastic Systems for Estimating Equilibrium Expectations.
DOI: 10.5220/0005535501380144
In Proceedings of the 5th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH-2015),
pages 138-144
ISBN: 978-989-758-120-5
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
2014), a paper that provides a good understanding of
the main issues, and without such an understanding,
the question as to what models are exceptions, and
why, cannot be addressed in any scientific way.
The setup we use is like the one of (Whitt, 1991):
The system is described by d state variables X
1
(t),
X
2
(t), ..., X
d
(t), and we define X(t) = [X
k
(t),k =
1,2,...,d]. We are interested in the process R(t) =
f(X(t)), and the problem is to find the expectation
lim
t
E(R(t) = E(R). If d = 1, we use X(t) instead
of [X
1
(t)]. We assume that the process is ergodic, in
which case E(R) exists and it is unique. R(t) could
be, for instance, the number of jobs in a queueing net-
work with 3 queues. In this case f(X(t)) becomes
f([X
1
(t),X
2
(t),X
3
(t)]] = X
1
(t) + X
2
(t) + X
3
(t).
We assume that we know f(.). The question is how
many runs should be made to find the estimate for
E(R) with the lowest possible MSE. We assume that
the runs are of equal length, and that the total time
for all runs is T. Hence, if there are n runs, each run
has a length of T/n. We use time averages to estimate
E(R), that is:
¯
R
ν
(T/n) =
1
T
Z
T/n
0
R(t)dt (1)
and we take the overall average
¯
¯
R(T) =
n
ν=1
¯
R
ν
(T/n). (2)
If there is only 1 run, we use
¯
R(T) instead of
¯
¯
R(T).
All runs start in the same state, which implies that all
R
ν
(T/n) have the same distribution. We will there-
fore sometimes omit the subscript. The common ex-
pectation will be denoted by E(
¯
R(T/n)), and the com-
mon variance by Var(
¯
R(T/n)). Equation (1) assumes
continuous sampling, but similar formulas can easily
derived if sampling is only done at certain epochs, or
if the data sampled is discrete, as is the case when
considering successive waiting times. Also note that
for reasons discussed later, we use no warm-up pe-
riod. If a warm-up period of length w is used, then the
lower bound of the integral of (1) must be replaced by
w, and
1
T
by
1
Tw
. To judge the quality of
¯
¯
R(T), we
use the mean squared error, that is
MSE(
¯
¯
R(T)) = E((
¯
¯
R(T) E(R))
2
)
= Var(
¯
¯
R(T)) + Bias
2
(
¯
¯
R(T)) (3)
where
Bias(
¯
¯
R(T)) = E(
¯
¯
R(T)) E(R). (4)
When we use the term MSE, we always refer to
MSE(
¯
¯
R(T)) or, if there is only one run, MSE(
¯
R(T)).
Our attention is restricted to point estimations, as
opposed to interval estimations and confidence inter-
vals (Alexopoulos and Goldman, 2004). Of course,
multiple runs have the additional benefit to provide
some idea about the reliability of an estimate.
The models we use are rather simple, such as the
M/M/1 queue or a simple tandem queue. However,
from such simple models, insights can be gained that
allow us to form, and potentially prove, interesting
conjectures when analyzing more complex models. It
would be much more difficult to gain these insights
by looking at complex systems. For simple models,
there are very effective methods to find expectations
and variances of time averages (Grassmann, 1987),
methods that are deterministic in the sense that they
do not use Monte Carlo techniques, but probability
theory. Incidentally, finding variances by simulation
with acceptable precision requires a large number of
observations, and this leads to long execution time,
making these method non-competitive for small mod-
els. However, since deterministic methods increase
exponentially with the dimensionality of the model,
simulation is often the only way to analyze models
with many state variables.
2 WHY MULTIPLE RUNS CAN
DECREASE THE MSE
Though it may sound contradictory, warm-up peri-
ods can increase the MSE. As it turns out, these are
exactly the cases were multiple runs can be advan-
tageous. In (Grassmann, 2014), a number of rea-
sons have been given why warm-up periods can in-
crease the MSE when initializing the system in cer-
tain states. We list these reason here to make the paper
self-contained.
1. Consider the case where sampling is not done
continuously, but only at time 0, τ, 2τ, 3τ, ...,
mτ = T. If τ is large enough, then the read-
ings at these sample points are almost indepen-
dent random variables. Now, if the state at time
0 is close to E(R), then it would make sense to
include it in the time average, that is, instead of
¯
R(T) =
1
m
m
ν=1
R(ντ), one uses
1
m+1
m
ν=0
R(ντ).
In fact, this would amount to a weighted average
with a weight of
1
m+1
for R(0) and a weight of
m
m+1
for
¯
R(T), which makes sense, provided R(0)
is a reasonable estimate for E(R).
2. In many cases, E(R(t)) is subject to a drift, that is,
E(R(t)) changes in a predictable fashion. In this
case, starting in a state with R(0) close to E(R)
may not be a good starting state. Consider, for in-
UsingMultipleRunsintheSimulationofStochasticSystemsforEstimatingEquilibriumExpectations
139
stance, an M/M/1 queue with arrival rate λ and
service rate µ, with λ < µ. If the system starts
in a fixed non-empty state, then in a small inter-
val of length h, the expected number of arrivals is
λh+ o(h), and the expected number of departures
is µh+o(h), which means that the expected queue
length decreases at a rate λ µ, a rate that is nega-
tive. Similar effects were observed by (Kelton and
Law, 1985), using numerical experiments. Hence,
before the M/M/1 queue reaches the empty state,
the expectation will always decrease, precluding
thus to be in an equilibrium. In other words, no
reasonable equilibrium estimate can be obtained
before reaching 0. This may explain the findings
of Mandansky (Madansky, 1976) that the smallest
MSE for the expected number of elements in an
M/M/1 queue is obtained when starting the sys-
tem empty, provided the simulation is of a reason-
able length.
3. Most textbooks (Tocher, 1963), (Banks et al.,
2005), (Law and Kelton, 2000) suggest to start the
system in a typical state. This allows for different
interpretations: One could call a state typical if it
has a high equilibrium probability, or if it is fre-
quently visited while the process is stationary. If
the system starts in such a typical state, warm-up
periods may be detrimental.
Choosing an initial state according items 1 to 3 is only
possible if some prior information about the system
is known. However, if prior information is known,
its weight can be increased by doing multiple runs.
Moreover, the better the information, the more runs
should be done.
To find the states where a warm-up period can
reduce the MSE, consider the integral given by (1),
with the lower bound replaced by w, and with
1
T
changed accordingly. Suppose there is only 1 run.
Let
¯
R(T, w) be the resulting expression. We now
form the derivatives of Bias(
¯
R(T,w)), Var(
¯
R(T,w))
and MSE(
¯
R(T,w)) with respect to w for w = 0 to get
according to (Grassmann, 2014):
Bias
(
¯
R(w,T))
w=0
= E
(
¯
R(w,T))
w=0
=
1
T
(E(
¯
R(T)) f(a)) (5)
Var
(
¯
R(w,T))
w=0
=
2
T
Var(
¯
R(T)) (6)
MSE
(
¯
R(w,T))
w=0
=
2
T
(Var
¯
R(T) +
Bias(
¯
R(T))(E(
¯
R(t)) f (a))).(7)
Here, f(a) = f(X(0)) = R(0) as defined in the in-
troduction. If the derivative of the MSE is negative,
any warm-up period decreases the MSE, that is, a
warm-up period is beneficial. On the other hand, if
the derivative of the MSE is positive, a warm-up pe-
riod of zero is a local minimum, and it is reasonable to
expect that this minimum is also global. In this case,
any warm-up period is detrimental.
States with a positive derivative are obviously
good states to start a simulation in, and the benefits of
starting in these states could potentially be increased
by doing multiple runs. In fact, we observed that for
most cases with positive derivatives with respect to w,
multiple runs proved to be beneficial. Surprisingly,
we even found some cases with a negative derivative
for the MSE where multiple runs are optimal.
Next, we investigate as to what happens if T in-
creases by a factor of k. Consider first the case where
there is only one run. If T is large enough, then ac-
cording to (Asmussen and Glynn, 2007), Var(
¯
R(T))
and Bias(
¯
R(T)) both decrease by a factor of k as T
increases by a factor of k. This relation was orig-
inally suggested in (Grassmann, 1982), and (Grass-
mann, 2008) shows numerically that convergence to
the limiting form can be quite fast. Because of (3),
this implies that in the limit, the contribution of the
bias decreases by k
2
, whereas the variance decreases
by k. Hence, as T increases, and T is large enough,
the contribution of the bias diminishes.
If there are multiple runs, then we conclude from
equations (2) and (1):
MSE(
¯
¯
R(T)) =
1
n
Var(
¯
R(T)) + Bias
2
(
¯
R(T)). (8)
It follows that compared to a single run, the impor-
tance of the bias increases.
If n, the number of runs, is too large, T/n is very
small, and
¯
¯
R(T) is essentially an average of the initial
conditions, making the contribution of the simulation
almost meaningless. This follows from (8), which im-
plies that the relative importance of the variance de-
creases with n, approaching 0 as n . This means
that as n increases, starting states close to E(R) be-
come increasingly competitive. If R(0) is very close
to E(R), then it is optimal to make a large number of
very short runs. For very short runs, E(
¯
R(T)) R(0),
meaning that we essentially take the average of the
initial conditions, practically ignoring the results of
the simulation. The simulation runs are then almost
useless. In order to avoid such meaningless results,
we have to require that the run length T/n is long
enough such that the simulation can make a reason-
able contribution.
SIMULTECH2015-5thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
140
3 DRIFT AND FLOW MODELS
In this section, we discuss the drift in the context of
flow models (Newell, 1982). These models allow us
to find good initial states. Flow models are also help-
ful for deciding when the number of runs is so large,
and the individual runs are so short as to make the
simulation meaningless.
Note that every state i = [i
1
,i
2
,. .. ,i
d
] is associ-
ated with a drift for each state variable X
k
(t). The
drift is the derivative of E(X
k
(t)) with respect to t,
given X
k
(t) = s
k
, k = 1,2,. .., d. In a typical system,
almost all states have a non-zero drift, that is, when-
ever one knows the state, there tends to be a drift of at
least some state variables. Stochastic system in equi-
librium have no drift because they can be in many
different states, but we have no knowledge in which
state they actually are. Of the potential states the sys-
tem could be in, state variables of some states have an
up-drift, which is compensated by other states where
these variables have a down-drift. This compensation
vanishes if we know the state we are in.
If all states visited from 0 to t either have all an up-
drift for some relevant state variable, or they all have a
down-drift for this variable, a simulation run extend-
ing only from 0 to t cannot be expected to provide
a meaningful estimate of the equilibrium expectation.
The minimum time for a simulation is therefore the
time until all relevant state variables have changed the
sign of the drift at least once. Hence, we must visit a
pair of states (i, j) at least once where the sign of the
drift changes for a particular state variable when go-
ing from i to j. Theoretically, such pairs would have
to be found for every state variable, but for reasons to
be discussed, there are often pairs of states where all
state variables either change the sign of the drift, or
where the drift goes at least to zero when going from
one state of the pair to the other. To find such pairs
of states, we use a flow analysis (Newell, 1982), with
the flows being equated to the drifts.
To make the conversion of our stochastic model
to the flow approximation, we have to make all state
variables continuous. Typically, we have simple for-
mulas for the drift, and though these formulas possi-
bly apply only to integers, in a flow model, we dis-
regard this restriction. For instance, in the M/M/c
queue with arrival rate λ and service rate µ, the
drift when i elements are in the system is λ for
0 < i c, and this remains the formula for the flow
model, except that i is now continuous.
A flow model is in a state of equilibrium if no
state variable changes as the time t increases. In other
words, the drifts of all state variables in the flow ap-
proximation are zero. Let s = [s
1
,s
2
,. .. ,s
d
] be this
equilibrium state. If the equilibrium state s is stable,
and we assume that this is the case for now, then from
any state close to s there is a drift towards s. For
instance, in the flow approximation of the M/M/c
queue, the equilibrium state is s = λ/µ, because for
the M/M/c queue, we have a drift λ , which is
zero for i = λ/µ.
If there are two states, i = [i
1
,i
2
,, ...,i
d
] and j =
[ j
1
, j
2
,. .. , j
d
], with i
k
> s
k
and j
k
< s
k
for some k,
then being attracted to s, i must have a down-drift, and
j an up-drift. Similar results would have to be true for
all k. If the model is stochastic, then it may be possi-
ble to go from i to j, and if the path goes through s,
the sign of the drift changes for all state variables, or
at least it falls to 0. In a stochastic system, the path
may not exactly move through s, but we still can use
this result as a heuristic. If a state variable is integer,
we have to round it either up or down when convert-
ing the equilibrium state in the flow approximation to
a state in the stochastic model. If i and j are two states
obtained from s through rounding, and if one can go
in one step from i to j, then the drift typically changes
the sign for most state variables when going from i to
j. In the M/M/c model, for instance, the flow equilib-
rium is r = λ/µ. If r is not integer, it is easily verified
that the sign of the drift changes if one goes from r
(r rounded down) to r (r rounded up), or from r
to r. If all state variables of s that are supposed to
be integer are integer already, we only have one state
rather than a pair of states. To get a pair in this case,
we change one state variable by 1. This only makes
the drift change to zero when moving from the modi-
fied state to s, but a sign change is very likely to occur
soon. For instance, if r in the M/M/c queue is in-
teger, r = r. In this case, we have to extend the
simulation to a time long enough to move beyond this
no-drift state.
In some flow models, there is no state with a flow
of zero, and in this case, there is either no equilibrium,
and consequently no meaningfulequilibrium expecta-
tion E(R) to estimate, or the state variables hit some
boundary. For instance, in the M/M/1 queue, the drift
for all non-zero states is λ µ, and if λ < µ, the flow
model hits a boundary at zero. It is easily verified that
when hitting the boundary in an M/M/1 queue, or
even in an open network of M/M/1 queues, the sign
of the drift changes at the boundary points. The same
can be expected for most other systems when they hit
a boundary.
In summary, the minimum length of the simula-
tion should be such that we either go beyond a sign
change of the drift of all relevant state variables, or at
least a state where the drift is zero. Hence, we have
to find pairs (i, j) where the drift of all relevant state
UsingMultipleRunsintheSimulationofStochasticSystemsforEstimatingEquilibriumExpectations
141
variables changes as we go from i to j, or where the
drift of all state variables goes to zero. Before j is
reached, the drift is always in the same direction, and
the results of the simulation are close to meaningless.
Drift-change pairs (i, j) can be unstable, that is,
though the drift changes sign when going from i to
j, the drift can move us away from j. For example,
consider a birth-death process with a maximum pop-
ulation of N, and with birth rate λ
i
, 0 i < N and
death rate µ
i
, 0 < i N, where i is the size of the pop-
ulation. If there is a value k such that λ
i
µ
i
> 0 for
k i < N and λ
i
µ
i
< 0 for 0 < i < k, then there is
a drift toward N if i k, and toward 0 otherwise. In
situations where such unstable drift-change pairs are
possible, multiple runs are suggested, starting with
a drift-change pair, with each run closely monitored.
However, we will not consider this case further.
Stable drift-change pairs will be called pairs of at-
traction, or, if there is only one point states of attrac-
tion. For many models, starting a simulation with a
member of a pair of attraction can be recommended
because one can expect that a sign change in the drift
will occur in the near future, giving the simulation
some validity. Often, such starting states are also
states close to the state with the highest probability
or the highest frequency. For instance, in the M/M/1
queue, the pair of attraction is (0,1), with 0 being the
state with the highest probability, and 1 the state with
the highest frequency.
Since it is usually easy to formulate flow mod-
els and find pairs of attraction, starting states from
such pairs can be recommended. The fact that they
are often close to the maximal equilibrium probabil-
ity is a further advantage. The state variables of pairs
of attraction normally have values that lie below their
expected values. Consequently, choosing states with
the state variable having slightly higher values than
the ones in pairs of attraction may be advantageous in
some cases, especially if multiple runs are planed.
4 EXAMPLES
In this section, we present a number of examples to
highlight the different issues discussed earlier. In our
tables and discussions, the derivative of MSE is al-
ways multiplied by T because otherwise, it is very
small number,and representing small numbers in dec-
imal form requires a lot of space.
First, we show, using the M/M/1 queue with
X(t) = R(t) representing the number of elements in
the system as an example, that multiple runs can re-
duce MSE(
¯
¯
R(T)). The arrival rate is λ = 0.8 and
the service rate is µ = 1. Since the pair of attrac-
tion is (0, 1), we start with X(0) = 0, and we use
T = 1000. Since our programs cannot handle infinite-
state queues, we restrict the number in the system to
50. For this problem, TMSE
(
¯
R(T)) = 2.340, indi-
cating that the MSE is increasing when introducing a
warm-up period. As one can see from Table 1, the
MSE is smallest when 6 runs are made.
Table 1: MSE for M/M/1 queue, number of runs varying
from 1 to 8, λ = 0.8, µ = 1, T = 1000, X(0) = 0.
Runs 1 2 3 4
MSE 1.5691 1.3748 1.2154 1.1045
Runs 5 6 7 8
MSE 1.0412 1.0178 1.0259 1.0583
In the case of the M/M/1/N queue, one may also
be interested how the buffer size N and ρ = λ/µ af-
fect the optimal number of runs. Hence, in Table 2,
we do the calculation for a buffer size of N = 10 and
N = 50, and for ρ = 0.8 and 0.9. Table 2 shows in
column “Opt. runs” that more runs are optimal if the
buffer size has the higher value, and that fewer runs
are optimal if ρ has the higher value.
Table 2: Optimal number of runs for M/M/1/N queue with
varying N and λ when µ = 1, T = 1000 and X(0) = 0.
1 Run Multiple Runs
N λ MSE TMSE
Opt. runs MSE
10 0.8 0.171 0.180 3 0.168
50 0.8 1.569 2.340 6 1.018
10 0.9 0.228 0.124 2 0.227
50 0.9 14.856 16.770 3 9.555
To demonstrate that the derivative of MSE with re-
spect to w is a good indicator for using multiple runs,
consider the M/M/20/50 queue with λ = 0.8, µ =
1/c = 0.05, and T = 1000. The derivatives of MSE
in this case are negative for X(0) < 9 and X(0) > 25,
and they are positive for X(0) ranging from 9 to 25.
Table 3 gives the MSE for the number of runs from
1 to 3, and the optimal number of runs (Runs
) for
X(0) ranging from 6 to 10. As one can see, for both
X(0) = 7 and X(0) = 8, the results of Table 3 suggest
multiple runs, even though the derivative in question
is negative. In such cases, multiple runs may be com-
bined with warm-up periods. This, however, seems to
be an exception. For X(0) = 10, three runs are opti-
mal because the MSE for 4 runs is 2.077.
The next question we need to address is whether
or not states obtained from pairs of attraction will lead
to low values of the derivative of the MSE. To resolve
this question, consider first the M/M/20/50 queue
discussed earlier. In this case, λ/µ = 16, that is, for
X(0) = 16, there is no drift. Hence, this should be a
good starting state. However, the starting state with
SIMULTECH2015-5thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
142
Table 3: MSE(
¯
¯
R(T)) for the M/M/20/50 queue with λ =
0.8, µ = 0.05, and T = 1000.
X(0) TMSE
MSE Runs
1 run 2 runs 3 runs
6 -2.207 2.209 2.191 2.364 1
7 -1.184 2.200 2.147 2.260 2
8 -0.243 2.192 2.107 2.165 2
9 0.615 2.186 2.072 2.079 2
10 1.389 2.180 2.040 2.002 3
the lowest MSE when using only 1 run is X(0) = 13,
with MSE = 2.177. For X(0) = 16, the MSE is 2.181,
which is not much higher. However, when starting
with X(0) = 16, and doing 12 runs, one finds an MSE
of only 1.069, which is significantly lower. The best
one can obtain with X(0) = 13 is 4 runs with an MSE
of 1.746.
We now increase T by a factor of 10 form 1000 to
10000. In this case, the derivatives of of the MSE
with respect to the warm-up period is positive for
8 X(0) 24, and negative otherwise, which is very
close to the result obtained for T = 1000. Also, the
optimal number of runs does not change significantly
for X(0) between 6 to 10, as seen in Table 4. Also,
the starting state with the lowest MSE is state number
13, as before. Hence, in this model, the best starting
state, and the best number of runs is insensitive to T.
Table 4: MSE(
¯
¯
R(T)) for the M/M/20/50 queue with λ =
0.8, µ = 0.05, and T = 1000.
X(0) TMSE
MSE Runs
1 run 2 runs 3 runs
6 -0.185 0.239 0.239 0.240 1
7 -0.081 0.239 0.238 0.239 2
8 0.015 0.239 0.238 0.238 2
9 0.102 0.239 0.238 0.238 3
10 0.180 0.239 0.237 0.237 3
As our final example, consider a three station tan-
dem queue with stations 1, 2 and 3. Each station has a
buffer size of 5, including the place for the part being
served. All jobs arrive at station 1, and must proceed
to station 2, then 3, in that order, after which they
depart. If the first buffer is full, arrivals are lost. If
the buffer of station 2 or 3 is full, the previous sta-
tion is blocked, that is, departures are delayed until
there is a place in the buffer they go to. Arrivals are
Poisson with a rate λ = 0.75, and service is exponen-
tial, with a rate of µ
i
= 1 for station i, i = 1, 2, 3.
We use T = 1000. Our R(t) is given by the total
number in the system. The number in station i will
be denoted by X
i
(t), i = 1,2,3, and the state where
X
i
(t) = x
i
, i = 1,2,3 will be denoted by [x
1
,x
2
,x
3
].
With this notation, one pair of attraction is the state
([0,0,0],[1, 0, 0]). Unfortunately, both states of this
pair show that increasing the warm-up period from
0 to some positive value would decrease the MSE.
Accordingly, using two runs for state [0, 0, 0] would
increase the MSE form originally 0.1710 to 0.1736.
For state [1,0,0], the corresponding values are 0.1705
and 0.1715. The state with the lowest MSE is state
[5,0,0], with an MSE of 0.1694. Note that in our
model, E(R) = 5.44, and for state [5,0,0], R(t) = 5,
which is close to E(R). When starting in this state, the
MSE is reduced from 0.1715 to 0.1661 if two runs are
made. Also, the derivative of the MSE is positive. In
fact, out of the 216 states that can be used for starting
the simulation, 153 states have a positive derivative of
the MSE, and in 112 starting states, 2 runs are better
than 1. There are thus better states to start the simu-
lation in than [0, 0, 0] or [1, 0, 0]. On the other hand,
the potential improvement in the MSE by using states
other than [0,0,0] or [1,0,0] is limited. However,
since pairs of attraction are so easy to find, and since
the possible improvements obtainable by using other
starting states is small, using pairs of attraction is still
no mistake. Improvements can also be made by in-
creasing the state variables slightly. For instance, for
state [1,1,1], the MSE is 0.1702, its derivative is pos-
itive, that is, no warm-up period should be used, and
if two runs are made, the MSE decreases to 0.1696.
5 CONCLUSIONS
In this paper, we demonstrated that multiple runs may
be optimal if the simulation starts in a well-chosen
state. One method we presented for finding well cho-
sen starting states uses flow analysis to find pairs of
attraction: The elements of these pairs typically yield
good starting states. Sometimes, it is best not to use
these starting states directly, but to change the val-
ues of the state variable to bring them closer to their
equilibrium expectations. In particular,state variables
representing queues and obtained from pairs of attrac-
tion typically have values below their equilibrium ex-
pectation. Once a good starting state is at hand, its
effect would be watered down by any warm-up pe-
riod. The beneficial effect of a well chosen starting
state can be increased by making multiple runs. In
this case, the effect of the bias is strengthened, which
implies that starting states with R(0) close to E(R) be-
come increasingly advantageous.
We also presented a number of numerical exper-
iments. Though in the models used, pairs of attrac-
tion did not necessarily provide the minimal mean
squared error, they led to a reasonable performance
at all times, especially after a slight increase of the
state variables.
UsingMultipleRunsintheSimulationofStochasticSystemsforEstimatingEquilibriumExpectations
143
ACKNOWLEDGEMENTS
This research was partially supported by NSERC of
Canada, Grant 8112.
REFERENCES
Alexopoulos, C. and Goldman, D. (2004). To batch or not to
batch? ACM Transaction on Modeling and Computer
Simulation, 14(4):76–214.
Asmussen, S. and Glynn, P. W. (2007). Stochastic Simula-
tion: Algorithms and Analysis, volume 57 of Stochas-
tic Modelling and Applied Probability. Springer Ver-
lag, New York.
Banks, J., Carson, J. S., Nelson, B. L., and Nicol, D. (2005).
Discrete Event Simulation. Prentice Hall, Englewood
Cliffs, NJ.
Grassmann, W. K. (1982). Initial bias and estimation error
in discrete event simulation. In Highland, H. G., Chao,
Y. W., and Madrigal, O., editors, Proceedings of the
1982 Winter Simulation Conference, pages 377–384,
Piscataway, NY. The Institute of Electrical and Elec-
tronics Engineers, Inc.
Grassmann, W. K. (1987). Means and variances of time av-
erages in Markovian environments. European Journal
of Operationonal Research, 31:132–139.
Grassmann, W. K. (2008). Warm-up periods in simulation
can be detrimental. Probabilitiy in Engineering and
Informational Sciences, 22:415–429.
Grassmann, W. K. (2011). Rethinking the initialization
bias problem in steady-state discrete event simulation.
In Jain, S., Creasey, R. R., Himmelspach, J., White,
K. P., and Fu, M., editors, Proceedings of the 2011
Winter Simulation Conference, pages 593–599, Pis-
cataway, NY. The Institute of Electrical and Electron-
ics Engineers, Inc.
Grassmann, W. K. (2014). Factors affecting warm-up
periods in discrete event simulation. Simulation,
90(1):11–23.
Kelton, W. D. (1989). Random initialization methods in
simulation. IIE Transactions, 21(4):355–367.
Kelton, W. D. and Law, A. M. (1985). The transient behav-
ior of the M/M/c queue, with implication for steady-
state simulation. Operations Research, 33:378–396.
Law, A. M. and Kelton, W. D. (2000). Simulation Modelling
and Analysis. McGraw Hill, New York, third edition.
Madansky, A. (1976). Optimal conditions for a simulation
problem. Operations Research, 24:572–577.
Newell, G. F. (1982). Applications of Queueing Theory.
Chapman and Hall, London, second edition.
Pasupathy, R. and Schmeiser, B. (2010). The initial tran-
sient in steady state point estimation: Context, a bi-
ography, the MSE criterium, and the MSER statistic.
In Johansson, B., Jain, S., Montaya-Torres, J., Hugan,
J., and Y¨ucesan, E., editors, Proceedings of the 2010
Winter Simulation Conference, pages 184–197, Pis-
cataway, NY. The Institute of Electrical and Electron-
ics Engineers, Inc.
Pawlikowski, K. (1990). Steady-state simulation of queue-
ing processes: a survey of problems and solutions.
ACM Computing Surveys, 22(2):123–170.
Tocher, K. D. (1963). The Art of Simulation. English Uni-
versity Press, London.
Whitt, W. (1991). The efficiency of one long run ver-
sus independent replications in steady-state simula-
tion. Management Science, 37(6):645–666.
Wilson, J. R. and Prisker, A. A. B. (1978). Evaluation of
startup policies in simulation experiments. Simula-
tion, 31:79–88.
SIMULTECH2015-5thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
144