j o u r n a l o f t r a f fi c a n d t r a n s p o r t a t i o n e n g i n e e r i n g ( e n g l i s h e d i t i o n ) 2 0 2 0 ; 7 ( 4 ) : 5 0 7 e5 1 9

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.keaipublishing.com/jtte

Original Research Paper

A methodology for calibration of traffic micro-simulator

for urban heterogeneous traffic operations

Palak Maheshwary a, Kinjal Bhattacharyya b, Bhargab Maitra b,*,

Manfred Boltze c
a
ARCADIS, Bengaluru 560017, India
b
Department of Civil Engineering, Indian Institute of Technology Kharagpur, Kharagpur 721302, India
c
Department of Civil and Environmental Engineering, Technische Universitat Darmstadt, Darmstadt 64287, Germany

highlights

A traffic micro-simulation calibration methodology for heterogeneous traffic is proposed.

Linear regression models using sensitive parameters were developed.

Genetic algorithm based optimization was adopted to obtain optimal parameter values.

Different sets of optimal driving behavior parameter for different vehicle classes are suggested.

The optimal parameter sets were found to vary significantly across vehicle classes.

article info abstract

Article history: Traffic micro-simulation is a widely accepted tool in many countries for the evaluation
Received 5 January 2018 and assessment of alternative design schemes. However, for several developing coun-
Received in revised form tries, replicating heterogeneous, non-lane based traffic in a micro-simulation framework
5 June 2018 is gaining increased importance and still remains a challenge due to its complexity. The
Accepted 21 June 2018 present study demonstrates a methodology to calibrate a traffic micro-simulation model
Available online 10 July 2019 giving due consideration to vehicle-class specific driver behavior in an urban Indian
scenario for a midblock section and an intersection approach in Kolkata. The sensitive
Keywords: parameters affecting the driver behavior were identified for every vehicle type using
Traffic operation Latin Hypercube design, taking vehicle class specific travel time as a performance
Micro-simulation measure. Linear regression models were developed for each vehicle class considering the
Heterogeneous traffic sensitive driving behavior parameters. The models highlight that the dependency of
Vehicle-class measure of effectiveness (MOE) of one vehicle type is not only limited to its own driver
Driver behavior behavior but also on parameters of other vehicle classes. A genetic algorithm based
VISSIM optimization was adopted to obtain optimal parameter sets for different vehicle classes.
The optimum values were found to vary significantly across all vehicle classes at a 95%
confidence level. Single and multi-criteria calibration principles are also implemented to
yield much more realistic results and subsequently minimizing weighted error for all
vehicle classes.

* Corresponding author. Tel.: þ 91 94 3404 0738.

E-mail addresses: palakmaheshwary1@gmail.com (P. Maheshwary), kinjalb.ce@gmail.com (K. Bhattacharyya), bhargab@civil.iitkgp.ac.in
(B. Maitra), boltze@verkehr.tu-darmstadt.de (M. Boltze).
Peer review under responsibility of Periodical Offices of Chang'an University.
https://doi.org/10.1016/j.jtte.2018.06.007
2095-7564/© 2019 Periodical Offices of Chang'an University. Publishing services by Elsevier B.V. on behalf of Owner. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
508 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519

© 2019 Periodical Offices of Chang'an University. Publishing services by Elsevier B.V. on

behalf of Owner. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).

traffic. As different microscopic simulation platforms have

1. Introduction different parameters and/or variables to explain the inbuilt
mechanism of driver behavior, the calibration methodology
With the advancements in computational technology over developed in this study particularly focuses on the VISSIM
the last few decades, microscopic traffic simulation has microscopic simulator (Fellendorf and Vortisch, 2010).
become a popular tool among researchers for the evaluation The paper has been structured in seven sections. After
of different alternative design and management strategies portraying a background and need for the study in the intro-
for a road network before adopting full-scale real-life duction, a brief review of the works and identification of gaps
implementations. Field implementations generally have related to calibration of driving behavior in heterogeneous
legal and financial constraints associated to them while traffic conditions is discussed in section 2 with reference to
traffic simulation is a cost-effective, objective, and flexible VISSIM microscopic simulator. A methodology of calibration
approach to analyze and evaluate transportation systems of driver behavior is explained in section 3, followed by a
(Rakha et al., 1996). When compared to analytical models, brief description of the heterogeneous traffic corridor
microscopic simulation models can keep a track of selected for the study and the database developed for a
individual vehicle movements and interactions on a second demonstration of the calibration methodology for the
to sub second basis (FHWA, 2004) and they are quite selected study corridor in sections 4 and 5 respectively.
helpful in realistic representation of complex traffic Discussions based on statistical analysis of results at
behaviors. different stages of calibration are presented in section 6. The
In developing countries such as India, absence of lane paper is concluded with a summary of the work highlighting
marking, non-lane based traffic, disorganized movement of a the major findings and contributions in section 7.
varied mix of vehicle types, is a commonly observed scenario
which is difficult to replicate using traditional analytical
models thus creating a necessity for microscopic simulation 2. Previous works on VISSIM calibration
models (Mathew and Radhakrishnan, 2010). However, the
efficiency and acceptability of a traffic simulation model in In the early stages of traffic flow micro-simulation, re-
evaluating different scenarios lies in its ability to reflect the searchers calibrated simulation models on a trial-and-error
local study area's network, infrastructure, and driver basis and often used default parameters for replicating
characteristics. This can be achieved by model calibration diverse traffic scenarios resulting in large errors in the model
which may be defined as the process to address the most outputs as reported by Park and Schneeberger (2003). In the
important differences between the model's default recent years, in order to improve the reliability of traffic
assumptions, and those conditions actually observed locally micro-simulation models, several studies have projected the
in the field by selecting the best set of model input need of an accurate calibration of micro-simulation models,
parameters and modifying or fine-tuning their default values and they have proposed general requirements and
in order to accurately reflect the field-measured and guidelines on calibration (Dowling et al., 2004; Hourdakis
observed local traffic conditions (Park and Schneeberger, et al., 2003; Park and Qi, 2005; Park and Schneeberger, 2003).
2003; Rakha et al., 1996). In the process, procedures have evolved from simple time-
Microscopic traffic flow simulation software contains several consuming manual calibration to more efficient automated
independent parameters to describe the traffic control opera- processes with an objective to develop a realistic and
tion, traffic flow characteristics, and drivers' behavior (Park and acceptable calibrated model for end users. Table 1 is a
Schneeberger, 2003). It is difficult, cumbersome and expensive comprehensive database of a majority of the available
to estimate some of these parameters directly from the field. techniques that have been used in different stages of
Hence, in general, these parameters are kept at their default calibration of a microscopic traffic flow simulation model. It
values which reflect a well-disciplined and homogeneous may be mentioned that these studies focused primarily on
traffic environment. Thus, in order to replicate different homogeneous traffic conditions and related aspects. Only a
complex traffic scenarios, the values of these parameters need few studies have been reported in literature with regard to
to be modified and adjusted by the user by following a defined the calibration of heterogeneous traffic conditions
procedure of calibration. However, adequate research has not (Manjunatha et al., 2013; Mathew and Radhakrishnan, 2010;
been conducted to account for the heterogeneous traffic Siddharth and Ramadurai, 2013).
environment and subsequently incorporate them while Hellinga's (1998) eight-step procedure set forth a general
undergoing calibration. Therefore, the aim of this study is to set of guidelines for the calibration and validation process
evolve a well-defined methodology through a step-by-step which was further elaborated by Park and Schneeberger
approach for calibration of driving behavior for heterogeneous (2003) explaining calibration and validation of VISSIM for
 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519 509

Table 1 e Different stages and techniques for calibration of traffic flow micro-simulation models.
Source Selection of Number of Optimization MOE for
calibration parameter parameters procedure calibration
calibrated
Park and Schneeberger (2003) Engineering judgment, Latin 6 Linear regression model Travel time
hypercube (LH) design
Dowling et al. (2004) Engineering judgment e Non-linear search Travel time, flow, speed
algorithm
Kim et al. (2005) Moses' Distribution Free Rank- 6 Genetic algorithm Travel time
Like Test, Wilcoxon RankeSum
Test, KolmogoroveSmirnov
Test
Park and Qi (2005) Engineering judgment, LH 9 Genetic algorithm Travel time
design, one-way ANOVA
Yu et al. (2006) Engineering judgment 10 Genetic algorithm Speed
Menneni et al. (2008) Engineering judgment 5 Evolutionary algorithm Speed-flow curve
Cunto and Saccomanno (2008) Engineering judgment, Plackett 6 Genetic algorithm CPI*
eBurnman with foldover
fractional design
Ishaque and Noland (2009) Engineering judgment 15 e Travel time
Huang and Sun (2009) Unknown 5 NSGA-II Speed, flow
Duong et al. (2010) Factorial design 7 Genetic algorithm Flow, CPI*
Mathew and Radhakrishnan (2010) Literature review, sensitivity 4 Genetic algorithm Stopped delay
analysis
Manjunatha et al. (2013) 2-way ANOVA 5 Genetic algorithm Delay
Siddharth and Ramadurai (2013) ANOVA, quasi-OTEE 9 Genetic algorithm Flow
Li et al. (2011) Sensitivity analysis 8 Manual pairwise Desired speed,
iteration acceleration -deceleration,
saturation flow
Mehar et al. (2014) Literature review, 2-way 2 Trial and error Capacity
ANOVA

Note: * CPI: crash potential index.

signalized intersections in the U.S.A. based on a linear bounds on link capacity. Siddharth and Ramadurai (2013)
regression model. Dowling et al. (2004) presented a also developed a methodology on calibrating heterogeneous
calibration of key global capacity parameters followed by traffic conditions in India using VISSIM to minimize the
parameters affecting the route choice. Additionally, various differences between simulated and field-observed flow
parameter optimization algorithms have been applied in the values with reference to two intersections in Chennai City.
process of calibration; among these, the genetic algorithm The driver behavior forms a significant part of calibration
(GA) has emerged to be a very popular technique. For process when heterogeneous traffic is involved. In a country
example, Wu et al. (2005) applied GA to find a suitable like India, driver behavior is not the same for all vehicle types;
combination of VISSIM parameters for a study conducted on it is largely vehicle class dependent which has not been taken
an Expressway in Shanghai. Park and Qi (2005) proposed a into account in the previous studies. In heterogeneous traffic
general methodology to calibrate VISSIM using an actuated conditions prevalent in urban India, the driver aggressiveness
signalized intersection as a case study. tends to vary with respect to the vehicle class (Maini and
Calibration using GA has also proven to be effective in the Khan, 2000). In general, motorized two-wheeler and three-
limited number of studies involving heterogeneous traffic. wheelers drivers are more aggressive than the car and bus
Mathew and Radhakrishnan (2010) suggested a methodology drivers. Motorized two-wheelers and three-wheelers tend to
for calibrating parameters for heterogeneous traffic utilize the available gaps between vehicles to make their
operations prevalent in developing countries like India. The way at a faster speed even in congested traffic situations.
calibration procedure included the identification of sensitive Hence, the output/MOE is yet another factor which is
parameters, setting the selected parameter ranges generally expected to be different based on vehicle category.
heuristically, and the tuning of selected parameters using In the previous works related to calibration of VISSIM,
GA. Three four-legged signalized intersections with calibration was generally done for the traffic stream as a
heterogeneous traffic operations were modelled and the whole. For instance, if speed is the MOE, calibration was
models were calibrated and validated using stopped delay as done in such a way that average stream speed matches with
the measure of effectiveness (MOE). The approach was later the observed values instead of calibrating speed for every
modified by Manjunatha et al. (2013) who introduced the vehicle separately. However, research in mixed traffic
concept of multi-parameter sensitivity analysis along with streams has indicated that the speed distribution is not
GA for calibration. Additionally, a brief idea of multi-criteria uniform and is influenced by the proportion of slow moving
calibration was also presented by taking delay as MOE with vehicles (Dey et al., 2006). Moreover, in most of the studies,
510 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519

intersection approaches have been the only focus for required from the field. All the three sub-tasks are interde-
calibration in heterogeneous traffic conditions until now. No pendent in a way that the MOE is identified such that it is
attempts have been made to calibrate the simulation models consistent with the study goals, available field data and
with reference to different sections of the roadway. This simulation model capabilities. Among different available
study attempts to overcome these gaps by proposing a simulation models, VISSIM is a widely used micro-simulation
detailed methodology to calibrate the VISSIM micro- platform, and its varied features and parameters make it quite
simulator based on vehicle-class wise driver behavior. The flexible to model heterogeneous traffic conditions (Man-
proposed methodology is demonstrated with reference to a junatha et al., 2013; Mathew and Radhakrishnan, 2010). Hence,
relatively heterogeneous traffic corridor in Kolkata city as a VISSIM was selected for the purpose of this study. The sub-
case study. sequent stages are discussed considering VISSIM as the
selected micro-simulation platform.

3. Methodology 3.2. Initial modelling

The proposed procedure for the calibration of microscopic The second stage deals with network coding and once the
traffic flow simulation models broadly includes 4 stages as simulation model was set up, simulation runs were performed
demonstrated in Fig. 1: pre-modelling (i.e., definition of study with default parameter values followed by a comparison of
objectives, selection of study area, determination of MOEs, the model output with the observed field values. If the dif-
field data collection and extraction), initial modelling ferences are significant, it shall justify the need for calibration
(network coding, comparison of simulation with default and subsequent steps would then be followed.
values and field results), calibration (identifying calibration
parameters and ranges, experimental design for calibration 3.3. Calibration
and finally optimization, e.g., using GA), validation and
visualization (using a different set of flow data for validating There are some parameters that cannot be measured directly
the calibrated model, visualization check for realistic on the field. These include several driver behavior parameters
animations to identify the best parameter set). that affect the output significantly in case of heterogeneous
traffic and hence need to be calibrated so that the on-field
3.1. Pre-modelling conditions can be replicated. This is further explained in five
sub-steps as shown in Fig. 2.
This is the first stage dealing with initial steps before actually
modelling the network. This includes the selection of a traffic 3.3.1. Identifying sensitive parameters and range setting
micro-simulation platform, the identification of a suitable VISSIM micro-simulation platform provides a number of pa-
MOE, and accordingly collecting and processing the data rameters which may be modified to replicate a particular

Fig. 1 e Flowchart for proposed methodology.

 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519 511

Fig. 2 e Detailed calibration methodology.

traffic scenario. However, all parameters are not equally number of combinations for many parameters at a time is
sensitive to all kinds of traffic scenarios. Hence, it is essential enormous making it impossible to examine all possible com-
to first identify the set of sensitive parameters for the partic- binations of a parameter set. Latin hypercube design (LHD), a
ular traffic scenario under consideration. Such an initial set of stratified-random sampling procedure, provides an efficient
sensitive parameters was identified by varying each param- way of sampling variables from their distributions (Iman and
eter by small amounts of 10% and keeping all others at their Conover, 1980). A LHD algorithm reduces the number of
defaults and observing the effect on the MOE. The parameters combinations to a reasonable level, still covering the
which, on small increments, give a relatively higher change in entire parameter space. Hence, in this case, parameter
the MOE were pre-selected as sensitive parameters. To restrict combinations were designed based on LHD algorithm.
the wide range of values of these parameters, a suitable range All possible combinations of parameter values obtained
needs to be decided within which these parameters can vary. from LHD were provided as input to the model one by one, and
The range was identified in a way that it covered a range of the results were recorded. Multiple runs were conducted for
0.5e2 times of the MOE value observed in the field. It would every combination to address the stochastic variability.
not be logical to consider any parameter value which gives an Finally, one-way ANOVA test, a method of determining the
MOE beyond this range as the final objective is to find suitable statistical significance of the parameters with respect to the
parameter values for which the MOE obtained from simula- output, was performed to further reduce the sensitive pa-
tion match the field observed MOE. Thus, every sensitive rameters to get the “predictor variables” or the most sensitive
parameter was incremented or decremented individually parameters.
until the desired lower or upper limit of the MOE is reached.
3.3.3. Multiple linear regressions
3.3.2. Latin hypercube design and ANOVA Once the final set of driver behavior parameters were known,
Once an initial set of parameters that may be sensitive has it was necessary to modify their values within the defined
been identified as discussed in the previous step, it is further range to obtain the parameter values which best describe the
necessary to screen them to select the most sensitive pa- particular traffic scenario. In order to simplify and make the
rameters for the particular traffic scenario. This shall enable process less cumbersome, a linear regression model was
to efficiently reduce the processing time for calibration. The developed with driver behavior parameters as independent
512 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519

variables (X) and the desired output as dependent variable (Y) against the calibrated model. It is also important to check the
from LHD samples wherein 100 parameter value sets were visual proximity of the simulation to the traffic flow observed
generated and 3 runs was taken for every set. Multiple re- in the field thus ensuring the plausibility of the calibrated
gressions allow determining the overall fitness of the model parameter set.
and the relative contribution of each of the predictors to the
total explained variance. Moreover, the equations obtained
will be specific only to the traffic scenario under consideration 4. Study area and database development
and a new regression model must be created in case of other
traffic scenarios. The study corridor as shown in Fig. 3 is a part of the 3 km long
stretch called Dr. Meghnad Saha Sarani (or, Southern Avenue)
3.3.4. Optimization using a genetic algorithm in South Kolkata, connecting Golpark with Deshapriya Park.
This is the final step of calibration where the equations The study area covers about 500 m with an intersection
generated would act as inputs to the genetic algorithm toolbox approach, and the road network is coded in VISSIM 7.0.
in MATLAB. Genetic algorithm is an extensively used search Traffic operations in Indian cities are characterized by
technique that inherits ideas from natural evolution to varying degrees of heterogeneity with a mix of several types
effectively find good solutions for combinational parametric of fast and slow moving vehicles. The corridor selected for
optimization problems (Wu et al., 2005). The GA uses a specific the present study is located in one of the major central
number of digits, called a chromosome, which is generated at business district (CBD) areas in the Kolkata city. In order to
random. The chromosome's specific digits correspond to add some relief to the already congested traffic operations,
parameter values. By generating a single chromosome, several traffic management measures are presently
values for each parameter are generated and with them a implemented in the CBD areas during the regular working
completely randomized simulation run is done. Three basic hours, such as, restriction of entry to large commercial
operators of a GA analysis are the reproduction, crossover, vehicles and manually operated slow moving vehicles. The
and mutation. The optimization can be done in 2 ways: present study corridor and other corridors in the vicinity of
the study area, therefore, majorly comprise of three vehicle

Single criteria e wherein parameter values are obtained in types, i.e., car, bus and bike (or motorized two-wheelers).
a way that the equation for every vehicle type is optimized The study corridor has also been carefully selected to
individually (one at a time). The objective function is evaluate the effectiveness of the adopted calibration
specified in a way that the difference between simulated methodology in replicating vehicleedriver interaction in a
and observed travel time is not significantly high. heterogeneous traffic environment by minimizing the influ-
ence of any other external factors, such as, side friction,
ðrTTobs  TTsim rÞ interaction with opposing traffic, public transport stops and
%absolute error ¼  100
TTobs pedestrian interactions. The data collected on the field formed
a part of the input data given to VISSIM while modelling the

Multi-criteria e wherein optimized parameter values are network.
obtained for every vehicle type is optimized by a single The data comprises of geometrical measurements, free flow
objective function. The objective function is defined as the speed, vehicle compositions, travel time, turning movement,
percentage weighted error. The weightage may depend on and signal timing studies. Also, longitudinal and lateral dis-
the share of a vehicle type in the traffic stream. tances were measured for all possible combinations of vehicles
at intersections. Average delay per vehicle is an aggregate
ðW1 ,error1 ÞþðW2 ,error2 Þþ…þðWn ,errorn Þ measure and VISSIM is a microscopic tool and, therefore, a
%weighted error¼
W1 þW2 þ …þWn better approach of calibration should be a comparison of dis-
100 tribution of travel time between field and micro-simulation
instead of a discrete aggregate measure as average vehicle
where W1, W2, …, Wn is the share (%) of a particular vehicle delay. Moreover, there is a lot of approximation associated with
type in the traffic stream, error1, error2, …, errorn are the ab- measuring vehicle delays in comparison to obtaining travel
solute errors of output for vehicles 1, 2, …, n, respectively. times directly from the field. VISSIM measures vehicular delay
by “… subtracting the theoretical (ideal) travel time from the
3.3.5. Constraint insertion actual travel time.” This shows that travel time is more generic
Optimization of the objective function does not necessarily for comparing field and simulation outputs. It may also be
guarantee the minimization of individual vehicle errors. This mentioned that instead of a single free flow speed, a desired
necessitates moving a step further in the process of calibra- speed distribution for each vehicle class is considered. The
tion by putting constraints on individual errors in addition to implications are that two vehicles with similar travel times may
the % weighted error. This will ensure a high degree of accu- have different delays as their desired speeds are also different.
racy and robustness of the calibrated model. This is true even for real-life scenarios. However, due to ap-
proximations, such realistic delay estimates can't be made
3.3.6. Validation and visual check directly from the field. Therefore, travel time was chosen as the
To perform validation of the microscopic simulation model, MOE for the study. The parameters were calibrated for peak
a new set of field data for different hours of the day or hour and then validated for non-peak hour for the same study
different sections of the corridor was collected and tested area. Additionally, calibration was done for a midblock section
 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519 513

Fig. 3 e Video screenshot and map of study corridor. (a) Screenshot. (b) Map.

and an intersection approach to check for any statistically sig- seen from Table 3, the lateral and longitudinal distances
nificant difference in the optimized parameter value set. vary with respect to the participating vehicles. In general, at
The extracted data revealed that the traffic stream consti- 0 speed, buses maintain higher distances and bikes
tutes 70% cars, 23% bikes and 7% buses. Lateral distances and maintain lower distances with reference to cars.
speed-acceleration profiles of the vehicle types contribute
towards the simulation of heterogeneous traffic (Li et al.,
2011). These data have also been extracted from the field as
shown in Tables 2 and 3 and were provided as input to the 5. Application
VISSIM micro-simulator in the form of speed distribution
graphs, acceleration distribution graphs and lateral The following section presents the stepwise application of the
distances between vehicles at 0 km/h. There is a variation in calibration procedure on the selected study corridor. The
desired speed of vehicles near intersections and at midblock methodology was carried out with reference to the VISSIM
sections away from intersections. Therefore, two travel time traffic flow simulator. The car following model in VISSIM is
sections have been selected to capture the desired speed based on the continued works of Wiedemann (PTV AG, 2013)
distributions, one for a distance of 50 m from the STOP line with the assumption that a driver can be in one of four
and another a distance of 70 m starting from 150 m away driving modes: free driving, approaching, following, and
from the STOP line. Table 3 shows the lateral and braking. Wiedemann-99 model is used in the present study
longitudinal gaps between different vehicle types. It may be given that it offers more flexibility over Wiedemann-74
mentioned that, in the case of longitudinal gaps, the vehicle model. The following section presents the stepwise
type in the first row is the leading vehicle and the vehicle application of the proposed procedure on the study area
type in the second row is the following vehicle. As it may be taken.

Table 2 e Observed vehicle speed and acceleration characteristics.

Type of Length Width Acceleration (m/s2) at speed ranges* (kmph) 85th percentile free flow speed** (kmph)
vehicle (m) (m)
0e20 20e40 40e60 60e80 Near intersection Away
Car 4.0 1.60 1.5 1.2 0.9 0.7 51 62
Bus 9.8 2.20 0.8 0.7 0.4 0.3 40 48
Bike 1.6 0.55 1.3 1.0 0.6 0.5 50 60

Note: * Source: Manjunatha et al. (2013).

** Source: primary traffic studies by the authors.

Table 3 e Lateral and longitudinal distance between vehicles at intersection.

Average gap (m) Car Bus Bike
Car Bus Bike Car Bus Bike Car Bus Bike
Lateral ** 0.60 0.90 0.40 0.90 1.10 0.60 0.40 0.60 0.30
Longitudinal ** 1.20 1.60 0.70 1.92 1.59 0.36 0.92 0.80 0.49

Note: ** Source: primary traffic studies by the authors.

514 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519

5.1. Initial modelling reducing the number of parameters to eight as shown in

Table 5. Their ranges were obtained such that the lower and
5.1.1. Network coding upper bound values covered an interval of 0.5e2 times the
With the data gathered from the field, the study area was observed value of travel times on field.
modelled in VISSIM. The geometrics of the area were first
coded with the help of Google maps, and then the data 5.2.2. Latin hypercube design and ANOVA
extracted from the field was given as input into the model. ANOVA analysis was further carried out on the 8 parameters
based on which six parameters were finalized as predictor
5.1.2. Simulation with default parameter values variables for intersection approach as shown in Table 5.
Once the study area was modelled, ten simulation runs were ANOVA analysis was conducted at 95% confidence interval
carried out keeping the driver behavior parameters at their based on the samples obtained from Latin hypercube design
default values. and the parameters having p-value less than 0.05 were
considered as the ones sensitive to the travel time.
5.1.3. Comparison of the simulated result with field Firstly, ANOVA analysis was considered for the intersection.
observations CC0, i.e., the standstill distance between vehicles, is a major
The average difference between simulated and observed contributing parameter to the travel time as vehicles are
travel times for all vehicle types was then calculated for both, stacked closely at intersections. However, these CC0 values are
intersection and midblock section. The difference being sig- not of much significance to the midblock section because ve-
nificant, calibration was necessary and the following steps hicles don't normally stop there. So, fixing the CC0 values ob-
were adopted. tained for intersection, ANOVA was done for midblock with the
assumption that the CC0 for car, bus and bike is invariant
5.2. Calibration throughout the road stretch being modelled. Table 6
demonstrates the most significant parameters for the
5.2.1. Identifying sensitive parameters and range setting intersection and the midblock section that are further used in
The process started with 30 driver following behavior pa- the subsequent steps.
rameters in total, i.e., CC0 to CC9 for car, bus and bike. A brief Here, CC0 is the average standstill distance between 2 ve-
description of the parameters is provided in Table 4. These hicles. It has no variation. CC1 is the following distance. It is
parameters were subjected to 10% increment/decrement, the distance that a driver wants to maintain at a certain speed.
Higher the speed, more cautious the driver is. CC3 controls the
deceleration process, i.e., how much time before safe dis-
Table 4 e A brief description of driver following behavior tance, driver recognizes a preceding slower vehicle. CC7 is the
parameters (PTV AG, 2013). oscillation during acceleration (PTV AG, 2013).
Parameter Description
5.2.3. Multiple linear regressions
CC0 It is the average desired standstill
distance between two vehicles. It With the help of Latin hypercube samples, three regression
has no variation. equations were developed for intersection and midblock each.
CC1 It is the distance in seconds which a The travel time for every vehicle type was considered as the
driver wants to maintain at a dependent variable and the driver behavior parameters as the
certain speed. The higher the value, independent variables or predictor variables.
the more cautious the driver is.
Initially, three equations were developed (for car, bus and
CC2 It restricts the distance difference
bike, respectively) for peak hour at the intersection. Once the
(longitudinal oscillation) or how
much more distance than the optimized values for driver behavior parameters were ob-
desired safety distance a driver tained using the genetic algorithm procedure (explained in
allows before he intentionally section 5.2.4), only then calibration was carried out for the
moves closer to the car in front. midblock as it required a fixed value for CC0. The optimal
CC3 It controls the start of the
deceleration process, i.e., the
number of seconds before reaching
the safety distance. At this stage the
driver recognizes a preceding Table 5 e Sensitive parameter ranges for car, bus and
slower vehicle. bike.
CC4 It defines negative speed difference
Vehicle Parameter Default Lower Upper
during the following process.
type value bound bound
CC5 It defines positive speed difference
during the following process. Car CC0 1.50 0.00 2.00
CC6 Influence of distance on speed CC1 0.90 0.90 4.20
oscillation while in following CC3 8.00 13.00 8.00
process. Bus CC0 1.50 0.00 1.60
CC7 Oscillation during acceleration CC1 0.90 0.85 3.85
CC8 Desired acceleration when starting Bike CC0 1.50 0.00 0.80
from stand still. CC1 0.90 0.80 2.50
CC9 Desired acceleration at 80 km/h. CC7 0.25 0.25 2.00
 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519 515

CC0 values obtained for car, bus and bike were changed from 5.2.4. Optimization using genetic algorithm
default to those obtained from the intersection calibration. The equations were used in the GA process to get the optimal
Keeping these values fixed, ANOVA analysis was repeated values of the driver behavior parameters such that the
for the remaining parameters. Additionally, Latin hypercube objective function was reduced down to a minimum possible
samples were generated for the sensitive parameters to value. The GA parameters like crossover and mutation prob-
develop separate regression models for midblock travel abilities were chosen heuristically. A general roulette wheel
times. Eqs. (1)e(3) are for intersection, and Eqs. (4)e(6) are technique was used in the selection of parameter values. The
for midblock section. From the equations, it may be calibration principle may be of two types: (i) single criteria
observed that, in several cases, the travel time (TT) of a where travel time of car, bus and bike are calibrated separately
particular vehicle-class is not only dependent on driver taking only those parameters that are sensitive to a particular
behavior parameters of that class but also on that of the vehicle type under consideration, and, (ii) multi criteria where
other vehicle-classes as well. all three travel times are considered together and weighted
For intersection error is minimized. Single criteria calibration focuses only on
the absolute error in the output for a particular vehicle type. In
TTCar ¼ 28.48 þ 3.96 CC0Car þ 3.49 CC1Car þ 0.44 CC1Bus that process, the errors in the other vehicle types are not taken
þ 0.73 CC1Bike (1) care of. Multi criteria calibration proved to yield much more
realistic results as the travel time for all vehicles were
TTBus ¼ 32.23 þ 2.92 CC0Car þ 2.85 CC1Car þ 0.76 CC1Bus considered and was preferred. Table 7 demonstrates the
þ 1.08 CC1Bike þ 1.25 CC0Bus (2) optimum value of parameters, weighted error and individual
error after calibration considering the midblock section
TTBike ¼ 25.57 þ 2.82 CC0Car þ 1.77 CC1Car þ 0.58CC1Bus during the peak hour. It may be observed that although the
þ 1.08 CC1Bike þ 0.96 CC0Bus þ 1.18 CC0Bike (3) weighted error was within acceptable limits, the individual
errors for buses and bikes were really high.
For midblock
5.2.5. Constraint insertion
TTCar ¼ 0.075 þ 4.64 CC1Car þ 0.52 CC1Bus þ 1.425 CC1Bike (4) Only multi criteria calibration did not provide an optimal set
of parameters. Although, the objective to get low weighted
TTBus ¼ 4.79 þ 5.22 CC1Car þ 0.86 CC1Bus  0.66 CC3Car (5) error was achieved, however, the individual vehicle errors
(esp. bus and bike) continue to have larger values. For
TTBike ¼ 10.52 þ 2.93 CC1Bike (6) example, the weighted errors as demonstrated in Table 7 lie

Table 6 e ANOVA analysis for intersection and midblock section.

Vehicle type Parameters Intersection Midblock
p-value Significant parameter p-value Significant parameter
Car CC0Car 0.00 CC0Car, CC1Car, CC1Bus, CC1Bike Fixed CC1Car, CC1Bus, CC1Bike
CC1Car 0.00 0.00
CC3Car 0.37 0.08
CC0Bus 0.27 Fixed
CC1Bus 0.05 0.05
CC0Bike 0.59 Fixed
CC1Bike 0.05 0.03
CC7Bike 0.17 0.54
Bus CC0Car 0.00 CC0Car, CC1Car, CC0Bus, CC1Bus, CC1Bike Fixed CC1Car, CC3Car, CC1Bus
CC1Car 0.00 0.00
CC3Car 0.43 0.02
CC0Bus 0.00 Fixed
CC1Bus 0.00 0.02
CC0Bike 0.52 Fixed
CC1Bike 0.00 0.24
CC7Bike 0.21 0.54
Bike CC0Car 0.00 CC0Car, CC1Car, CC0Bus, CC1Bus, CC0Bike, CC1Bike Fixed CC1Bike
CC1Car 0.00 0.62
CC3Car 0.42 0.25
CC0Bus 0.02 Fixed
CC1Bus 0.01 0.84
CC0Bike 0.05 Fixed
CC1Bike 0.00 0.00
CC7Bike 0.54 0.32
516 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519

Table 7 e Individual and weighted error for midblock in the peak hour.
Type of Parameter Default Vehicle type Optimized Individual error (%) Weighted
calibration value value error (%)
Multi criteria CC0Car 1.5 Car, bus, bike 1.70 Car: 0.04, bus: 18.26, bike: 23.07 6.61
CC1Car 0.9 3.10
CC3Car 8.0 12.25
CC0Bus 1.5 1.35
CC1Bus 0.9 3.21
CC0Bike 1.5 0.45
CC1Bike 0.9 1.88

well within limits even though the error in travel time of bus improved optimized parameter set was obtained by putting
and bike are on a higher side. To avoid such a situation and an additional constraint that the individual errors were also
get more realistic results, insertion of constraints on within specified limits. The required changes were made to
individual travel time error of vehicles was necessary with the optimization function and the final set of optimal
simultaneous minimization of weighted error. Thus, an parameters was obtained as shown in Table 8. From the

Table 8 e Optimal parameter values for intersection and midblock in the peak hour.
Flow Parameter Default Optimized value Optimized value Weighted error Weighted error
condition value for intersection for midblock for intersection (%) for midblock (%)
Peak hour CC0Car 1.5 1.09 1.09 3.87 with individual 2.24 with individual
CC1Car 0.9 2.85 3.75 constraints <15% constraints <10%
CC3Car 8.0 8.00 9.37
CC0Bus 1.5 1.25 1.25
CC1Bus 0.9 2.45 2.72
CC0Bike 1.5 0.50 0.50
CC1Bike 0.9 1.22 2.30

Fig. 4 e Travel time plots in peak hour. (a) Car for midblock section. (b) Bus for midblock section. (c) Bike for midblock section.
(d) All vehicles for midblock section.
 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519 517

Table 9 e Statistical significance test results for intersection and midblock.

Section of corridor p-value Remark
Intersection Midblock section
CC1Car ¼ 2.73 CC1Car ¼ 3.85 0.000 Null hypothesis rejected
CC3Car ¼ 8.00 CC3Car ¼ 10.48 0.005 Null hypothesis rejected
CC1Bus ¼ 2.56 CC1Bus ¼ 3.15 0.050 Null hypothesis rejected
CC1Bike ¼ 1.53 CC1Bike ¼ 2.05 0.011 Null hypothesis rejected

Table 10 e Statistical significance test results for vehicle-wise driver behavior parameters.
Vehicle type p-value Remark
Car Bus Bike
CC0Car ¼ 1.06 CC0Bus ¼ 1.16 e 0.368 Null hypothesis accepted
e CC0Bus ¼ 1.16 CC0Bike ¼ 0.384 0.000 Null hypothesis rejected
CC0Car ¼ 1.06 e CC0Bike ¼ 0.384 0.000 Null hypothesis rejected
CC1Car ¼ 2.73 CC1Bus ¼ 2.56 e 0.452 Null hypothesis accepted
CC1Car ¼ 2.73 e CC1Bike ¼ 1.53 0.000 Null hypothesis rejected
e CC1Bus ¼ 2.56 CC1Bike ¼ 1.53 0.000 Null hypothesis rejected

table, it may be highlighted that on inserting constraints in the 5.3.2. Vehicle-wise driver behavior parameters
algorithm, the parameter search space becomes much more These parameter values are the same for every vehicle type by
restricted thus giving more realistic estimates of the default in VISSIM. However, that is not the case in the real
parameter values. It also reduced the weighted error in traffic conditions observed in India. In the present study, the
addition to the individual errors. values of CC0 (stopping distance) and CC1 (headway time)
were observed to be statistically significant in most of the
5.2.6. Plots cases when compared between car, bus and bike (Table 10).
Normal and cumulative frequency distribution graphs were
plotted for travel time of cars, buses, bikes showing how
calibration is effective when compared to field distribution of
travel time rather than default. The calibrated travel time 6. Conclusion
distributions shift very close to field distributions in compar-
ison the distributions with default parameter values as can be A 4-step methodology for calibration of VISSIM micro-simula-
seen in Fig. 4(a)e(d). tion software for heterogeneous traffic environment has been
presented in this paper. Initially a set of parameters sensitive to
travel time as the MOE were identified using Latin Hypercube
5.3. Validation and visual check design and ANOVA and linear regression models were devel-
oped for each vehicle class considering the parameters which
The calibrated model was successfully validated for a have significant influences on the travel times. An interesting
different flow level (non-peak conditions) with a weighted finding from the development of regression models is that the
error of 10.3% at intersection and 3.8% at midblock. travel time of a particular vehicle-class depends not only on the
In order to justify the modified values of the optimal parameters of that vehicle-class but also of other vehicle-clas-
parameter sets by comparing them with the default parameter ses. Such kind of influences have not been duly investigated or
values, statistical significance tests are necessary. To carry out given due attention in the previous studies reported on cali-
these tests, five different parameter set solutions were derived, bration of VISSIM. Optimal values of the parameter sets for
satisfying the constraint conditions. With the help of these each vehicle-class were then obtained by minimizing the
values, statistical significance was tested for (a) intersection weighted error of travel time using genetic algorithm. An
and midblock section and (b) vehicle-wise driver behavior. additional constraint was also inserted so that the error for
individual vehicle-class does not become exceedingly high.
5.3.1. Intersection and midblock section This improved the robustness of the model as the weighted
For peak as well as off-peak, statistically significant (p < 0.05) error was minimized to as low as 2% with individual errors not
differences were obtained in the values of parameters CC1 higher than 10%.The calibrated model was validated with
(headway time), CC3Car (deceleration process controller) when respect to different flow data which also yielded satisfactory
calibrated for intersection and then for midblock as shown in results. Hence, the methodology gives due consideration to the
Table 9. This is because vehicles stack more closely near the vehicle-class specific driving behavior and suggests different
intersection as compared to the midblock sections. sets of optimal driver “following” behavior parameter for
Moreover, due to the influence of traffic signals, the different vehicle classes.
deceleration behavior is also different at intersections as The methodology has been demonstrated with reference to
compared to the midblock sections. a case study of a traffic corridor in the Kolkata metro city,
518 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519

India. Emphasis was given to calibration of both intersection Dowling, R., Skabardonis, A., Halkias, J., et al., 2004. Guidelines for
and midblock area and statistically significant differences calibration of microsimulation models: framework and
were obtained in the calibrated parameter values for both applications. Transportation Research Record 1876, 1e9.
Duong, D., Saccomanno, F., Hellinga, B., 2010. Calibration of
cases. Near the intersection, vehicles tend to be more closely
microscopic traffic model for simulating safety performance.
packed minimizing the safety distances as compared to the In: Transportation Research Board 89th Annual Meeting,
midblock section. Also, the CC0 and CC1 values separately for Washington DC, 2010.
car, bus and bike were found to be significantly different. The Fellendorf, M., Vortisch, P., 2010. Microscopic traffic flow
safety distances maintained by these vehicle classes are simulator VISSIM. In: Fundamentals of Traffic Simulation.
different with the minimum being for bike which tends to Springer, New York, pp. 63e93.
utilize the available gaps between vehicles to move effectively Federal Highway Administration (FHWA), 2004. Traffic Analysis
Toolbox Volume III: Guidelines for Applying Traffic
in congested conditions. This is reflected in the calibrated
Microsimulation Modelling Software. FHWA-HRT-04-040.
model and it supports the primary objective of this study to FHWA, Washington DC.
calibrate the driving behavior with respect to different vehicle Hellinga, B.R., 1998. Requirements for the calibration of traffic
classes. This finding is extremely relevant in case of hetero- simulation models. Proceedings of the Canadian Society for
geneous traffic operations in developing countries. Civil Engineering 4, 211e222.
Although the calibration results are case-specific, the Hourdakis, J., Michalopoulos, P.G., Kottommannil, J., 2003. Practical
procedure for calibrating microscopic traffic simulation models.
suggested methodology can be applied in similar scenarios of
Transportation Research Record 1852, 130e139.
non-lane based heterogeneous traffic where driving behavior
Huang, W., Sun, J., 2009. A ngsa-ii based parameter calibration
is highly vehicle-class dependent. Moreover, the methodology algorithm for traffic microsimulation model. In:
has been successfully applied in a traffic corridor where the International Conference on Measuring Technology and
vehicle composition includes car, bus and motorized two- Mechatronics Automation, Zhangjiajie, 2009.
wheelers. The findings encourage further research to calibrate Iman, R.L., Conover, W.J., 1980. Small sample sensitivity analysis
VISSIM micro-simulation model considering other vehicle- techniques for computer models, with an application to risk
assessment. Communications in Statistics: Theory and
classes as well in different contexts of heterogeneous traffic
Methods 9 (17), 1749e1874.
operations in developing countries. Continued research by Ishaque, M.M., Noland, R.B., 2009. Pedestrian and vehicle flow
application of the calibration methodology will help to calibration in multimodal traffic microsimulation. Journal of
develop optimal driving behavior parameter value sets for Transportation Engineering 135 (6), 338e348.
different vehicle classes and for different traffic scenarios in Kim, S.J., Kim, W., Rilett, L.R., 2005. Calibration of
developing countries. This will benefit practitioners to get microsimulation models using nonparametric statistical
more realistic results while applying traffic simulation models techniques. Transportation Research Record 1935, 111e119.
Li, J., Zheng, F., Henk, V.Z., et al., 2011. Calibration of a traffic
in the context of developing nations.
simulation for a Chinese city. In: 2nd International
Conference on Models and Technologies for Intelligent
Transportation Systems, Leuven, 2011.
Conflict of interest Maini, P., Khan, S., 2000. Discharge characteristics of
heterogeneous traffic at signalized intersections. In:
Transportation Research Circular E-C018: 4th International
The authors do not have any conflict of interest with other Symposium on Highway Capacity, Maui, 2000.
entities or researchers. Manjunatha, P., Vortisch, P., Mathew, T.V., 2013. Methodology for
the calibration of VISSIM in mixed traffic. In: Transportation
Research Board 92nd Annual Meeting, Washington DC, 2013.
Mathew, T.V., Radhakrishnan, P., 2010. Calibration of
Acknowledgment microsimulation models for nonlane-based heterogeneous
traffic at signalized intersections. Journal of Urban Planning
and Development 136 (1), 59e66.
This study was conducted as a project within the Indo-German
Mehar, A., Chandra, S., Velmurugan, S., 2014. Passenger car units
Master Sandwich Program. The authors would like to express
at different levels of service for capacity analysis of multilane
their sincere thanks to the German Academic Exchange Service interurban highways in India. Journal of Transportation
and Science and Engineering Research Board grant number SB/ Engineering 140 (1), 81e88.
S3/CEE/015/2013 Government of India for providing their kind Menneni, S., Sun, C., Vortisch, P., 2008. Microsimulation
support. They also appreciate a lot that PTV AG, Karlsruhe calibration using speed-flow relationships. Transportation
provided a VISSIM license for conducting this study. Research Record 2088, 1e9.
Park, B., Qi, H., 2005. Development and evaluation of a procedure
for the calibration of simulation models. Transportation
Research Record 1934, 208e217.
references Park, B., Schneeberger, J., 2003. Microscopic simulation model
calibration and validation: case study of VISSIM simulation
model for a coordinated actuated signal system.
Cunto, F., Saccomanno, F.F., 2008. Calibration and validation of Transportation Research Record 1856, 185e192.
simulated vehicle safety performance at signalized PTV AG, 2013. Vissim 7.00 User Manual. PTV Karlsruhe, Berlin.
intersections. Accident Analysis & Prevention 40 (3), 1171e1179. Rakha, H., Hellinga, B., Van Aerde, M., et al., 1996. Systematic
Dey, P.P., Chandra, S., Gangopadhaya, S., 2006. Speed distribution verification, validation and calibration of traffic simulation
curves under mixed traffic conditions. Journal of models. In: Transportation Research Board 75th Annual
Transportation Engineering 132 (6), 475e481. Meeting, Washington DC, 1996.
 J. Traffic Transp. Eng. (Engl. Ed.) 2020; 7 (4): 507e519 519

Siddharth, S.M.P., Ramadurai, G., 2013. Calibration of VISSIM for Bhargab Maitra is currently a professor in the
Indian heterogeneous traffic conditions. Procedia e Social Department of Civil Engineering and also
and Behavioral Sciences 104, 380e389. heads the School of Infrastructure Design
Wu, Z., Sun, J., Yang, X., 2005. Calibration of Vissim for Shanghai and Management at the Indian Institute of
expressway using genetic algorithm. In: 2005 Winter Technology Kharagpur. He is an Alexander
Simulation Conference, Orlando, 2005. von Humboldt Fellow and Deutscher Aka-
Yu, L., Yu, L., Chen, X., et al., 2006. Calibration of VISSIM for bus demischer Austausch Dienst (DAAD) Fellow.
rapid transit systems in Beijing using GPS data. Journal of In 2011, he received the Pandit Jawaharlal
Public Transportation 9 (3), 239e257. Nehru Birth Centenary Award from the In-
dian Roads Congress. He is chair of the Spe-
cial Interest Group (SIG) D3 “Applications of
Palak Maheshwary is a M.Tech graduate in
Travel Behavior Analysis and Demand Modelling Approaches” for
transportation engineering from Depart-
the World Conference on Transportation Research Society
ment of Civil Engineering, Indian Institute
(WCTRS) since 2013.
of Technology Kharagpur (2015). She is a
DAAD IIT Masters' Sandwich Scholarships
Fellow (2014e2015) and is currently working Manfred Boltze is a professor and head at the
as a transport planner at ARCADIS, Institute for Transport Planning and Traffic
Bangalore. Engineering in the Department of Civil and
Environmental Engineering, Technische
Universitat Darmstadt. He is a member of
the advisory board “Joint International
Research Laboratory of Transportation
Kinjal Bhattacharyya is a doctoral student in Safety”, Tongji University since 2017. He has
the Department of Civil Engineering at the been a member of the Steering Committee of
Indian Institute of Technology Kharagpur. the World Conference on Transport
He is a DAAD IIT Masters' Sandwich Schol- Research Society (WCTRS) since 2010 and
arships Fellow (2011e2012) and DAAD Bi- also serves the editorial board for the WCTRS Book Series on
nationally supervised Doctoral Degree Transportation Research. As topic area manager of “traffic oper-
Fellow (2015e2016). He is currently pursuing ations, management, and control”, he helped to prepare for WCTR
a PhD degree in the transportation engi- 2013 and as chairman, scientific committee to prepare for WCTR
neering specialization. 2016. In addition, he is also a member of the Heureka foundation
for environment and mobility since 2017.