Transportation Research Record 1802 ■ 23

Paper No. 02-2296

Variations in Capacity and

Delay Estimates from
Microscopic Traffic Simulation Models
Zong Z. Tian, Thomas Urbanik II, Roelof Engelbrecht, and Kevin Balke

One of the issues involved in using microscopic simulation models is A number of studies (1–3) have compared the results of various
the variation in the simulation results. This study examined some of traffic simulation models with those of analytical models, such as the
the more popular microscopic traffic simulation models, CORSIM, models in the Highway Capacity Manual (HCM) (4). These studies
SimTraffic, and VISSIM, and investigated the variations in the perfor- usually focus on comparing the results produced by different models
mance measures generated by these models. The study focused on the and make recommendations on which simulation model better repli-
capacity and delay estimates at a signalized intersection. The effects of cates the results of either the HCM or field data. The results from
link length, speed, and vehicle headway generation distribution were simulation models are usually based on the average values from an
also investigated. With regard to variations in performance measures, arbitrary number of runs and simulation times, for example, 10 runs
the study found that CORSIM yields the lowest variations, whereas with a 1-h simulation time. No studies, however, have investigated
SimTraffic yields the highest. The highest variation in each simulation how the performance measures vary among different simulation
model normally occurs when the traffic demand approaches capacity. models. Traffic engineers and researchers need to know when it is
It was also found that delays are affected by the link length and speed in necessary to perform multiple runs and how many runs are needed in
simulation models. Such an impact on delays is closely related to the range order to yield reasonable estimates in the performance measures.
of speed variations. In general, shorter links and higher link speeds result The scope of this paper is limited to signalized intersection oper-
in lower delays. There is no strong evidence that the headway distribution ations. First, a simple case with a single-lane approach and 100%
used to generate vehicles in the simulated network has any effect on through traffic is presented. Another case considers a single-lane
capacity and delay estimates. Multiple simulation runs are necessary to approach with a right-turn pocket, where the variations depend not
achieve an accurate estimate on the true system performance measures. only on general driver and vehicle characteristics but also on the
With a 10% error range in estimated delay, two to five runs may be stochastic process of queue blocking occurrences. Three of the most
enough for under-capacity conditions, but more than 40 multiple runs popular simulation models currently being used in the United States
may be necessary to accurately estimate delay at, near, or over capacity. are examined: CORSIM (5), SimTraffic (6), and VISSIM (7 ). Finally,
a summary and some conclusions are provided, and recommenda-
tions are then made for general applications of simulation models
With advances in computing technology and the ever-increasing
power of personal computers, many sophisticated stochastic micro- in practice.
scopic simulation models have been developed in the area of trans-
portation engineering. Improved user interfaces have significantly SIMULATION MODEL VERSUS
reduced the effort needed to code and interpret the results of these ANALYTICAL MODEL
simulation models. As a result, more traffic engineers are relying on
microscopic simulation models to analyze complex transportation Delay is one of the major performance measures for transportation
problems when analytical methods cannot provide satisfactory solu- systems. However, different methodologies are usually used to cal-
tions. However, the easy-to-use features of many simulation models culate delay in simulation models and analytical models such as those
also present a number of challenges, one of which is that inexperi- of the HCM. Detailed results and discussions on this issue can be
enced users are not aware of the potential variations in the output of found in several studies (1, 2). Although the focus of this paper is on
stochastic simulation models. Although some users may have real- the variations in performance measures, it might be useful to illustrate
ized the importance of reporting results from multiple runs, time and what results are expected when the default parameters of each simu-
budget often constrain them from doing so. An important aspect of lation model are used. The differences between the delays produced
stochastic simulation modeling is that each simulation run produces from simulation and those calculated using the HCM methodology for
different results. Therefore, it is important to understand what fac- signalized intersections are summarized below:
tors contribute to the variations. It is the purpose here to examine
three commonly used traffic simulation models and to investigate • The HCM reports an average control delay, which includes the
the range of results that these models produce and the conditions deceleration, queue moving time, stopped time, and the acceleration.
affecting the variability. However, the length of an approach as well as the speed that may con-
tribute to the acceleration and deceleration portions of control delay
are not specifically considered. For example, higher speeds may
Z. Z. Tian, R. Engelbrecht, and K. Balke, Texas Transportation Institute, Texas
A&M University System, 3135 TAMU, College Station, TX 77843-3135. require longer deceleration and acceleration times.
T. Urbanik II, Department of Civil and Environmental Engineering, University of • Most simulation models report average total delay, which is
Tennessee, Knoxville, TN 37996-2010. measured as the difference in travel time at a lower speed compared
24 Paper No. 02-2296 Transportation Research Record 1802

with the travel time at the free-flow speed. Therefore, total delay period or increasing the number of multiple runs. On the other hand,
includes control delay and other delays resulting from conditions some simulation models, such as CORSIM, may have adopted vari-
such as normal congestion and car following. ance reduction techniques (VRTs) in the model structure (8). The pur-
• CORSIM (Version 4.32 and earlier) and SimTraffic report total pose of VRTs in simulation is to somehow reduce the variance of an
delay on a link basis. Therefore, the delay associated with acceler- output random variable of interest, such as delay, without disturbing
ating to free-flow speed, which typically occurs on the downstream its expectation (the average). Such techniques, if properly used, can
link, is not accounted for in the delay computations. In VISSIM, normally obtain greater precision, for example, smaller confidence
however, the user can define segments on which to collect delay sta- intervals, for the same amount of simulation effort.
tistics. Therefore, delay from VISSIM can be collected to account
for the acceleration by defining the correct travel time segment. A
STUDY CASES
newly released version of CORSIM (5.0) incorporates a methodol-
ogy to take into account the acceleration so that it is consistent with Case 1: Single-Lane Approach with 100%
the HCM’s delay definition (1). Since CORSIM 5.0 was still under
Through Traffic
beta testing at the time of conducting simulation runs in this research,
CORSIM 4.32 was used instead. However, the major conclusions of The first case studied involves a single-lane approach at a signalized
this study remain valid. intersection. The following assumptions are made for further analysis:
• The HCM reports delay for the vehicles arriving during the
analysis period, whereas simulation models only report delay for the • The length of the approach is 760 m (2,500 ft);
vehicles departing during the analysis period. In undersaturated con- • The free-flow speed is 73 km/h (45 mph);
ditions, the input demand equals the throughput flow, and the differ- • The traffic flow consists of through vehicles only, all of which
ence is usually negligible when a relatively long simulation period
are passenger cars with 6.1-m (20-ft) bumper-to-bumper spacing;
(e.g., 15 min) is used. However, the difference can be significant if
• The signal timing has a 90-s cycle and a 50-s effective green; and
oversaturated conditions exist.
• All other default parameters of each simulation model are used,
• Simulation models automatically take into account the residual
including driver characteristics and car-following characteristics.
queues from the previous time period. Although the HCM provides
guidelines on how to consider the residual queue effect, most HCM-
In microscopic simulation models, vehicles are usually generated on
based analytical software packages, for example, the Highway
Capacity Software, do not compute the delay associated with resid- the basis of a certain headway distribution. CORSIM provides three
ual queues. types of vehicle generation distributions: uniform distribution, normal
distribution, and Erlang distribution. Both VISSIM and SimTraffic
Unlike deterministic analytical models, simulation models are use a negative exponential distribution, a special case of the Erlang
driven by samples of random variables from probability distributions. distribution. The default distribution in CORSIM is uniform distri-
These random variables may have large variances. As a result, these bution and is used for the analysis in Case 1. The effect of different
estimates could, in a particular simulation run with a particular ran- vehicle distributions is addressed in a later section of the paper.
dom seed, differ greatly among different runs of the model. The net A range of traffic flow conditions is investigated, and the HCM
effect is that there could be a significant probability of making erro- methodology is used to derive the delay and capacity results for each
neous inferences about the system under study if this variability is not condition, with the HCM default saturation flow rate of 1,900 passen-
taken into account. Several factors may affect the range of variations. ger cars per hour of green. These traffic flow conditions are included
In general, the variation can be reduced by using a longer simulation in Table 1. The capacity of the approach is obtained as follows:

TABLE 1 Traffic Flow Conditions

Traffic Demand, HCM Uniform Delay, Incremental Average Control

veh/h v/c Ratio d1, s/veh Delay, d2, s/veh Delay, d, s/veh
106 0.1 9.4 0.2 9.6
211 0.2 10.0 0.4 10.4
317 0.3 10.7 0.7 11.4
422 0.4 11.4 1.1 12.6
528 0.5 12.3 1.7 14.0
633 0.6 13.3 2.5 15.9
739 0.7 14.5 3.9 18.4
844 0.8 16.0 6.4 22.4
950 0.9 17.8 12.1 29.9
1056 1.0 20.0 27.7 47.7
1108 1.05 20.0 41.8 61.8
1161 1.1 20.0 59.2 79.2
1214 1.15 20.0 78.7 98.7
1267 1.2 20.0 99.3 119.3
Tian et al. Paper No. 02-2296 25

c = ( g C ) × s = (50 90) × 1900 = 1056 (vph) (1) Figures 1 through 3 illustrate the throughput flow rates from the
three simulation models based on different levels of traffic demand
where (simulation input). The results from each individual run, the average
of 30 runs, and the range of variations are depicted. The average
c = approach capacity (vph), throughput flow rates should match the traffic demand input for
g = length of effective green (s), under-capacity conditions. Each simulation model produces a max-
C = cycle length (s), and imum throughput flow rate, which can be used as an estimate of the
s = saturation flow rate (vph). capacity condition for that particular model. It should be noted that
The traffic conditions defined in Table 1 are simulated using each model produces a different capacity value than does the HCM
CORSIM, SimTraffic, and VISSIM. A total of 30 multiple runs are since the default model settings were used in this study. In other
conducted for each traffic flow condition. A 30-min total simulation words, no effort was made to calibrate the models to the capacity
time is used for each run, and the results from the second 15-min predicted by the HCM.
interval are reported. Using a 15-min interval in the simulation is Different levels of variation can be observed for the three models.
consistent with the 15-min analysis period used in the HCM. The CORSIM produces the lowest variations, whereas SimTraffic pro-
same simulation settings are applied for all the study scenarios duces the highest variations. The highest variations for each model
throughout this study. Both the throughput flow rates and the delays are observed when traffic demand is close to the capacity condition.
from the simulation models are analyzed. For example, at the HCM demand level of (v/c)HCM = 1.1 (1,162 vph),

1200 20.0
1100
1000
Throughput Flow Rate, veh/h

900 15.0

Standard Error, veh/h

HCM Capacity
800

700
600 10.0
500
400

300 5.0
200
Standard Error
100
0 0.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
Traffic Demand Based on HCM v/c Ratio

FIGURE 1 Throughput flow rates from CORSIM.

1200 20.0
1100
1000
900 15.0
Throughput Flow Rate, veh/h

Standard Error, veh/h

800 HCM Capacity

700
600 10.0

500
400 Standard Error
300 5.0
200
100
0 0.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio

FIGURE 2 Throughput flow rates from SimTraffic.

26 Paper No. 02-2296 Transportation Research Record 1802

1200 20.0
1100
1000

Throughput Flow Rate, veh/h

900 15.0

Standard Error, veh/h

800 HCM Capacity

700
600 10.0
500
400
300 5.0
Standard Error
200
100
0 0.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio

FIGURE 3 Throughput flow rates from VISSIM.

1300
Demand

1200 VISSIM
Throughput Flow Rate, veh/h

1100

1000
HCM Capacity CORSIM
900 SimTraffic

800

700
0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio

FIGURE 4 Average throughput flow rates.

the throughput flow rates in CORSIM range between 1,088 and approximately 1,145 vph. The estimated capacity from CORSIM is
1,160 vph. For SimTraffic, the highest variation occurs at (v/c)HCM about 1,130 vph.
= 0.9 (950 vph), where the throughput flow rates range between Similarly, the delay results from each individual run are illustrated
780 and 1,028 vph. In VISSIM the highest variation occurs at in Figures 5 through 7. A similar pattern can be observed in the varia-
(v/c)HCM = 1.05 (1,109 vph), where the throughput flow rates range tions from each simulation model. Again, CORSIM yields the lowest
between 1,060 and 1,132 vph. The variations tend to decrease once variation in delays, and SimTraffic yields the highest variation in most
the traffic demand is significantly higher than the capacity. of the simulation runs. The highest variations occur at the capacity
Figure 4 illustrates the average of the throughput flow rates from level and slightly beyond. For example, the delays in CORSIM range
the simulation models along with the input traffic demand curve. For between 23.4 and 71.5 s/veh close to its capacity condition [at (v/c)HCM
under-capacity conditions, the average throughput flow rates from = 1.1, or 1,162 vph], and the delays in SimTraffic range between
the simulation models all match the input traffic demands closely. 27.9 and 248.4 s/veh at the (v/c)HCM ratio of 1.0 (1,056 vph). For
The capacity conditions that each simulation model produces are VISSIM the delays range from 20.8 to 79.7 s/veh at the (v/c)HCM ratio
also demonstrated, which vary from the HCM values depending on of 1.1 (1,162 vph).
the model. SimTraffic yields the lowest capacity value at approxi- It is suspected that the lowest variations in CORSIM may be due
mately 1,025 vph, and VISSIM yields the highest capacity value at to the use of the uniform distribution in generating vehicles. How-
 Tian et al. Paper No. 02-2296 27

1000.0 20.0 1000.0 20.0

15.0 15.0

Standard Error, s/veh

Average Delay, s/veh

Average Delay, s/veh

Standard Error, s/veh

100.0 100.0

10.0 10.0

10.0 10.0

5.0 5.0

Standard Error
Standard Error
1.0 0.0 1.0 0.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio Traffic Demand Based on HCM v/c Ratio

FIGURE 5 Delay results from CORSIM. FIGURE 7 Delay results from VISSIM.

ever, a further investigation to be conducted later in this section indi- not specifically considered in the HCM methodology. With this in
cates that the same level of variation is achieved in CORSIM when mind, another set of simulation runs was conducted using a differ-
other types of vehicle distributions are used. ent link length and travel speed. Figure 9 shows the results from
Figure 8 illustrates the average delay from the three models along these runs with the CORSIM model.
with the results from the HCM. Recognizing the differences among As may be observed from Figure 9, link length and travel speed do
the delays produced by different models and the specific conditions affect the delay results. For the cases studied, lower delays result from
assumed in the simulation models (link length of 760 m and speed using shorter link length, and higher delays result from using lower
of 73 km/h), the results indicate that the delays from CORSIM match speed. However, the capacity values appear to remain the same. Sim-
closely the HCM results for under-capacity conditions. SimTraffic ilar results are also found in SimTraffic. However, the delays are
results in slightly higher delays and VISSIM results in slightly lower identical in VISSIM when different speed and link length are used.
delays than those in the HCM. A dramatic increase in delay can be There are two possible reasons for these results. The first is that a
observed once the capacity condition is reached. It should be kept in portion of the delay occurs while a vehicle is traveling on the link
mind that each model produces slightly different capacity results following a slower leading vehicle. Longer links provide more
when the default parameters are used. opportunity to accrue additional, noncontrol delay in this manner.
Figure 8 indicates that the delays reported from the simulation The second reason is associated with the speed distribution assumed
models deviate somewhat from the HCM results, especially when in each model. Both CORSIM and SimTraffic assume 10 different
the capacity condition is reached. Besides the differences in delay driver types, each with a specified speed relative to the link’s free-
calculations in each model discussed earlier, the results from simu- flow speed. For example, Driver Type 1 has a speed that is 75% of
lation may be specific to the conditions assumed in the models. The the free-flow speed, and Driver Type 10 has a speed that is 127% of
length of the link and the travel speed are the two elements that are the free-flow speed. In VISSIM, however, there is no such a default

1000.0
1000.0 20.0

HCM
CORSIM
15.0
Standard Error, s/veh
Average Delay, s/veh

Average Delay, s/veh

SimTraffic
100.0 100.0
VISSIM

10.0

10.0 10.0

5.0

Standard Error
1.0 0.0 1.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio Traffic Demand Based on HCM v/c Ratio

FIGURE 6 Delay results from SimTraffic. FIGURE 8 Average delay from different models.
 28 Paper No. 02-2296 Transportation Research Record 1802

100.0 by 2.0 s to achieve better correlation with the HCM results. It can be
90.0 seen that different headway distributions yield negligible differences
HCM in the delays. The capacity values are also almost identical. The same
80.0 Base - 760 m/73 km/h
low level of delay variation is obtained under low traffic volume
Shorter Link - 305 m/73 km/h
conditions. When traffic demand approaches capacity, the exponen-
Average Delay, s/veh

70.0
Lower Speed - 760 m/48 km/h
tial distribution does result in different levels of variations compared
60.0
with the normal and uniform distributions, but no consistent conclu-
50.0 sions can be drawn. It is also interesting to note that a nonmonotonic
40.0 trend is observed in the variations for all three models tested. It is sus-
pected that when a link is highly saturated, drivers tend to behave
30.0
more uniformly, resulting in less variation in the capacity and delays.
20.0

10.0
Case 2: Single-Lane Approach with
0.0 Right-Turn Pocket
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio Case 2 is derived from a real-world application. At a signalized inter-
section, the projected traffic volume is 1,150 vph on the eastbound
FIGURE 9 Effect of link length and speed on delay.
approach with 230 right-turn vehicles and 920 through vehicles (all
passenger cars). The current HCM procedure only considers a shared-
lane situation or an exclusive-lane situation; the effect of queue block-
speed distribution, and the user has to define a speed distribution ing on the right-turn pocket is not modeled specifically. On the basis
profile. In the current study, the speed distribution is defined in such of the HCM methodology, the capacities and level of service for the
a way that the 85th-percentile speed is taken as the free-flow speed, approach with a single shared lane and with an exclusive right-turn
and the minimum and maximum speeds are obtained by subtracting lane can be obtained. Using a 90-s cycle length and a 50-s effective
from and adding 10 km/h to the 85th-percentile speed. As a result, green (identical to that in Case 1), the approach would operate at level-
both CORSIM and SimTraffic would show higher speed variations of-service (LOS) F with a v/c ratio of 1.25 if the approach has a sin-
than would VISSIM. Larger speed variations result in larger poten- gle shared-lane configuration. With an exclusive right-turn lane, the
tial delays to following vehicles. Since CORSIM and SimTraffic through movement, the critical movement of the approach, would
define speed variation as a percentage of free-flow speed, higher operate at LOS C with a v/c ratio of 0.87. The result indicates that a
free-flow speed leads to larger speed variations and a larger amount right-turn pocket is needed to handle the projected traffic demand.
of noncontrol delay accrued on a link. Because of the topography and the right-of-way constraints, con-
The effect of different headway distributions is further investigated structing a longer-than-necessary right-turn pocket can be costly. The
using CORSIM, where the three types of headway distributions are purpose of the investigation is to determine the minimum length of
applied to Case 1, presented earlier. Figure 10 illustrates the delay the right-turn pocket to accommodate the projected traffic demand.
results based on the three headway distributions: uniform, normal, and Traffic operations based on different lengths of right-turn pocket
negative exponential. The standard errors of the delays are also (long enough to store N vehicles) are analyzed using CORSIM. Fig-
shown. In the study, the default saturation headway of 1.8 s is replaced ure 11 illustrates the estimated capacity values for the eastbound

100.0 20

90.0 Uniform Delay

Normal Delay
80.0
Exponential Delay 15
Standard Error, s/veh

70.0
Average Delay, s/veh

Uniform SE
Normal SE
60.0
Exponential SE
50.0 10

40.0

30.0
5
20.0

10.0

0.0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2

Traffic Demand Based on HCM v/c Ratio

FIGURE 10 Effect of vehicle-generating distribution, showing delay results

and their standard error (SE).
Tian et al. Paper No. 02-2296 29

1400 20

y= 19.128x + 1067.7
1300 15
R2 = 0.9741

Estimated Capacity, vph

Standard Error, veh/h

1200 10

1100 5
Standard Error

1000 0
0 1 2 3 4 5 6 7 8 9 10 11

Length of Right-Turn Pocket, Cars

FIGURE 11 Capacity estimates with regard to length of right-turn pocket.

approach based on various lengths of right-turn pockets. The capaci- capacity, and the variation in delay is significantly reduced. Although
ties in this case are estimated on the basis of the throughput flow rates, over-capacity conditions also exist with N between 2 and 4, the vari-
in which the input traffic demands in the simulation are set high ation in delay is not as high as that in Case 1, and the average delays
enough to block the right-turn pocket in every cycle at the beginning are still within the range between LOS C and D. From the results of the
of the green phase. It can be seen that the standard error ranges analysis, it may be recommended that the right-turn pocket be designed
between 5 and 8 veh/h, which is higher than that in Case 1. Such an with a length of two or three vehicles to minimize the construction
increase in the variations is primarily due to the added random cost while still maintaining acceptable operations most of the time.
elements in traffic flow. It can be observed that the capacity for However, occasional breakdowns may be expected with this design.
the eastbound approach is reached at approximately N = 4 cars
(the throughput flow rate is approximately 1,150 vph), although each
individual run may still vary significantly. NUMBER OF SIMULATION RUNS
Figure 12 illustrates the delay results associated with the length of
the right-turn pocket. It can also be observed that high variations exist Because of the variation in results of microsimulation models, mul-
for at- or over-capacity conditions (with N fewer than four cars). tiple runs must be conducted in order to provide an accurate estimate
Once N reaches five cars or longer, the approach will function under of the true performance measures. From the theory of probability and

160.0
LOS
140.0

120.0
Average Delay, s/veh

100.0
F
80.0

E
60.0

40.0
D

C
20.0
B
0.0 A
0 1 2 3 4 5 6 7 8 9 10 11

Length of Right-Turn Pocket, Cars

FIGURE 12 Delay estimates with regard to length of right-turn pocket.

30 Paper No. 02-2296 Transportation Research Record 1802

statistics, Equation 2 can be used to estimate the required number of 15-min and 1-h results in Table 2. However, only a 15-min period
runs to provide an estimate of the mean with a specified confidence would be consistent with the HCM procedures. A dilemma situation
interval and an error range: occurs in practical applications, however, in that the capacity con-
dition is usually an unknown parameter. As a general guideline, it is
z σ 2
recommended that at least two runs be conducted under any condi-
n =  α2  (2)
 E  tion. Whenever a capacity condition is suspected (a high delay is
obtained from an individual run), more runs are necessary to achieve
where more accurate estimates.

n = required number of simulation runs;

σ = sample standard deviation (based on 30 multiple runs from
this study); SUMMARY AND CONCLUSIONS
zα/2 = threshold value for 100(1 − α)% confidence interval; with
The results of an investigation into the variation produced by three
95% confidence interval, zα/2 = 1.96; and
commonly used microscopic traffic simulation models—CORSIM,
E = allowed error range.
SimTraffic, and VISSIM—are presented. Variations in both the
Using Equation 2 and the delay results presented in Case 1, the delay and the throughput traffic flow rates are examined on the basis
required number of runs is calculated with a 95% confidence inter- of different traffic flow conditions. The following conclusions are
val and a 10% error range, as presented in Table 2. A 10% error reached this study:
range is considered acceptable for general practice since it provides
sufficient accuracy to predict the correct level of service about 60% • For the three models tested, CORSIM produces the lowest vari-
of the time over a delay range between 0 and 80 s. For the remain- ation in both delay and throughput flow rates, whereas SimTraffic
ing 40% of the time, the predicted level of service would be within produces the highest variations. The highest variations occur when
one level of service of the actual level of service. capacity conditions are reached. The lower variations in CORSIM
As the results in Table 2 indicate, the highest number of runs is may due to the variance reduction techniques applied in the CORSIM
required for at- and over-capacity conditions because of the high vari- model. When the average values are considered from multiple runs,
ations. For example, a total of 45 runs is necessary for CORSIM at the the throughput flow rates from all three models match the input
level of (v/c)HCM = 1.1. An even higher number of runs is necessary demand closely for under-capacity conditions.
for SimTraffic and VISSIM because of the high variation in the result- • The three simulation models tested produce different results
ing delay measures. Both SimTraffic and VISSIM also indicate a when the default traffic flow parameters from each simulation model
higher number of runs at the lower traffic demand levels. Because are used. In general, VISSIM produces the highest capacity and low-
the delay values are usually small for low traffic demand conditions, est delay estimates, and SimTraffic produces the lowest capacity and
a biased estimate based on a fewer number of runs may not be a con- highest delay estimates.
cern in general practice. In general, two or five multiple runs may be • All three simulation models tested in this study illustrate a non-
sufficient for under-capacity conditions. However, when the traffic monotonic trend in the variations of the performance measures. It is
demand is at or over the capacity level, at least 40 runs may be neces- suspected that when a link is highly saturated, drivers tend to behave
sary, depending on the model to be used. The number of runs may be more uniformly, resulting in reduced variations in the performance
reduced by using a longer simulation time, as shown by the VISSIM measures.

TABLE 2 Required Number of Simulation Runs

Traffic Demand VISSIM-
(Based on HCM v/c Ratio) CORSIM SimTraffic 15 min VISSIM-1 h
0.1 2 18 56 6
0.2 3 15 18 2
0.3 4 13 10 2
0.4 2 8 9 1
0.5 2 4 7 1
0.6 1 6 8 2
0.7 1 5 7 1
0.8 1 11 6 1
0.9 1 58 4 1
1.0 2 137 13 2
1.1 45 25 57 18
1.2 14 18 1 1

NOTE: Traffic demand is determined based on the v/c ratio from the HCM. Each simulation has a different
capacity condition. The HCM v/c ratio at 1.0 does not necessarily mean the capacity condition for a
particular model.
Tian et al. Paper No. 02-2296 31

• Although link length and speed are not factors in the HCM REFERENCES
methodology for delay calculation, the study found that link length
and link speed directly affect the delays in microsimulation models. 1. Zhang, Y.L., L. Zhang, and L. Owen. Control Delay Definition, Model-
ing, and Calculation in CORSIM. Presented at the 80th Annual Meeting
In general, the simulation models would produce lower delays with of the Transportation Research Board, Washington, D.C., Jan. 2001.
shorter links or higher speeds. Such an impact is mainly due to the 2. Milam, R. T., and F. R. T. Choa. HCM and CORSIM—Resolving the
speed variability and car-following behavior. The headway distri- Differences. In Compendium of Papers, District 6 Annual Meeting, San
bution used to generate the vehicles does not show a significant Diego, Calif., Institute of Transportation Engineers, Washington, D.C.,
2000.
effect on delay and capacity estimates as well as on the variations, 3. Bloomberg, L., and J. Dale. Comparison of VISSIM and CORSIM
on the basis of a test performed with CORSIM. Traffic Simulation Models on a Congested Network. In Transportation
• Variations can be reduced by either using a longer simulation Research Record: Journal of the Transportation Research Board,
period or conducting more runs. It is also important to realize the dif- No. 1727, TRB, National Research Council, Washington, D.C., 2000,
ferences between simulation and analytical method when the perfor- pp. 52–60.
4. Highway Capacity Manual. TRB, National Research Council, Washing-
mance measures, such as delay, are reported. Although longer runs ton, D.C., 2000.
are desired to reduce the variations, only a 15-min period would be 5. CORSIM User’s Manual, Version 4.32. Office of Safety and Traffic Oper-
consistent with the HCM procedure, especially when varied traffic ations R&D, Intelligent Systems and Technology Division, FHWA, U.S.
demand is modeled over time. Department of Transportation, June 1999.
6. Synchro 4, Traffic Signal Coordination Software. Trafficware Corporation,
• To produce a reasonable estimate of the average conditions, it is
Albany, Calif., 1999.
recommended that at least two runs be conducted for low to medium 7. User Manual, VISSIM 3.50. PTV Planung Transport Verkehr AG,
traffic demand level. For near- and over-capacity conditions, at least Karlsruhe, Germany, 2001.
20 runs are recommended. This recommendation is based on the sim- 8. Rathi, A. K., and M. M. Venigalla. Variance Reduction Applied to Urban
plest case tested in this study. Higher variations are expected for most Network Traffic Simulation. In Transportation Research Record 1365,
TRB, National Research Council, Washington, D.C., 1992, pp. 133–146.
practical applications; therefore, more simulation runs may be neces-
sary. Simulation models should include automated utilities to perform
multiple runs and aggregate the results. Alternatively, simulation Publication of this paper sponsored by Committee on Highway Capacity and
models should be designed to allow users to develop such utilities. Quality of Service.