Ana!

ysis of Wiedemarn n 7 4 and 99 Driver Behavio:r Parameters

Authors:

Sandeep Menneni
Depatiment of Civil and Environmental Engineering
University of Missouri Columbia
E2509 Lafferre Hall
Columbia, MO 65211
Email: sm7b2@mizzou.edu

. Carlos Sun
Department of Civil and Environmental Engineering
University of Missouri Columbia
E2509 Lafferre Hall
Columbia, MO 65211
Email: csun@missouri.edu
Office (573) 884-6330

PTV Project Supervisors:

Peter Vortisch, PTV, A.G.

Thomas Bauer, PTV, America
TABLE OF CONTENTS

LIST OF FIGUR ES .. ...... ..... ... .. ........ ........ .................... .. ............. ...... ... ... ...... ... ... .. ........ ... .... .. ... ...... .4
LIST OFT ABLE S ...... .............. ...... ... ....... ... ... ..... ... ....... .. ......... .... ........ ... ............ ................ ... .... ...... 5
ABSTRACT ... ................. ... ... ........ .... .......... .... .. ....... .. ............. .... ..... ... .. .... ... .... .. .. ... .. ...... .. ....... .. ..... .. 6

1. lntroduction ..... ......... ....... .. ... .. ....... ......... ........ .... .... .. .... ... ......... ... .... ......... ........ ..... ... .. ...... .... ..... .8
2. VlSSIM Car Following Model s ................ .. ..... .. .... ...... .. .. .... .... .. .. ....................... .................. ... 10
2. 1. rntroduction
2.2. Wiedemann 74 Car Following Model s
2.3. Wiedemann 99 Car Fo llowing Mode ls
2.4. Vehicle Following Behavior Paramete rs
3. Microsimulation Calibration .. ......... ....... ... .. ....... .... ................ ..... ... .... ... .... .......... .. .... .. .. ........... l 6
3.1. Calibration
3.2. Literature Review on Calibration
3.3. Classification of Objectives in Calibration Methodologies
3.3.l.Qualitative Methods
3.3.2.Quantitative Methods
3.4. Data Considerations in Calibration
3.4.l.Microscopic Data
3.4.2.Macroscopic Data
4. Traffic Flow Representation in Calibration ....... ........ ........... .. .. ... ........... .. ..... ..... ... .. .. :.............. 27
4.1. Introduction
4.2. Aggregate-Data Based Representation
4.2.l.Capacity
4.2.2.Speed-Fiow Graphs
4.2.3.Fiow-Density, Speed-Dens ity, and Speed-Flow-Density Graphs
4.2.4.Maximum Sustained Flow Time Graphs
4.2.5.Speed, Flow, or Density Contours
4.2.6. Histogra ms : Speed, Headway, and Travel Time
4.2.7.Time Series: Volume, Speed, and Density
4 .3. Disaggregate-Data Based Representation
4.3.1. Vehicle trajectories
4.3 .2. Relative distance vs. relative velocity
5. Applications ofPattern Recognition in Calibration ....... ... .................................... .. ................. 38
5. 1. Introduction
5.2. Pattern Recognition
5.3. Defining a Generic Calibration Objective: Going Beyond
5.4. Literature Review on Dissimilarity Measures
5.4. l.Dissimilarity measures for quantitative data
5.4.2.Dissimilarity meas ures for binary, Boolean, or dichotomous data
5.4.3.Dissimilarity meas ures for popul ations
5.4.4.Dissimilarity measures for sequences
5.4.5.Dissimilarity meas ures for sets
5.5. Applications of Dissimilarity Measu res in Calibration
5.5.1 .Capacity
5.5 .2.Speed-flow or other similar graphs
5.5.3.Speed-flow graphs as partial point sets
5.5.4.Speed-flow-density or higher dimension graphs
5.5.5.Contours
5.5.6 .Speed and headway distributions

2
6. Opt imi zation Re prese ntation in Ca lib ration ..... ... ... .. ..... ........ .. ..... ....... . ...... ........ ....... ...... .... .... 53
6. I. Introduction
6.2. Evolutionary A lgor ithm s
6.3. EA Tool in MATLAB.
6.4. EA Tool Impl eme ntat ion for Ca librati on
7. VISSIM Calibration Methodology ..... .. ........... ...... ...... .... ... ... ..... ....... .. ............ ...... ..... ..... .... .... 60
7. l . Generalized Ca librat io n Methodology
7.2. Sensitivity A nal ys is and Range Definiti on
7.3. Theoretical Analysis ofthe Wiedemann Mode l Parameters
7.3.l.Capacity
7.3.2.Speed-Fiow Graphs
7.3.3 .Shockwaves
7.4. Summary
8. Case Study: US l 0 l NB , San Mateo .. .. .......... ........ .. ........ ...... ... .. ... .. .... .. .. .. .. ................. .. .. .. .. .. .. 67
8. I. Introduction
8. l .l.Network description
8. l .2. VISSIM model
8.2. Pattern Recognition Based Objective Function
8.3. Evolutionary A lgorithm
8.4. A Practitioner Oriented Simu lation Model: Test Network
8.5 . Results
8.5.1.Test network
8.5.2.US10 1 NB network
8.5.3 .Test-ca librated-va lues appl ied to USIOI NB
8.5.4. Categories of traffic ana lysis tool s .
8.6. Case Study Summary and Conclusions
9. Implementation of Integrated Ca libration Methodo logy .......... ...... ..... .. ........ ............. ... ... ....... . 80
9.1. Introduction
9.2. NGSJM Data Analysis
9.3. Disaggregate Data Based Calibration
9.4. Aggregate Data Based Calibration
9.5. Simplified Calibration Method for Practitioneis
I 0. Conclusions ............................ .. ...................... :............. ... .. .......... ... ....... .... .. ........ ..... ............ .. ! 05

3
LIST OF FIGURES
Figure 2.1 Wiedemann 74 car following model thresholds
Figu re 2.2 Car Fo ll owing Logic in Wiedeman n Model [adapted from Ehl ert (2004 )]
Figure 3.1 Solut ion to steady state calibration approach [Tol edo et a l. , 2004]
Figure 3.2 Latin Hypercube Des ig n (LHD) based calibration me thodology [Park and Qi, 2005 ]
Figure 4.1 Traffic flow representati o n of simu latio n and the fi e ld sce nario
Figtue 4.2 Typical VJ SS IM generate speecl-tlow graphs
Figure 4.3 Maximum Flow Sustained Time G rap h (MSFTG) ge nera ted us in g VI SSIM
Figure 5.1 Representations in calibration objective problem definiti on
Figure 5.2 Row concatenation representation of a binary image matri x
Figure 6.1 A prototypical evo lutionary a lgo rithm implementation
Figure 6.2 Three leve ls of representati o n in m icros imutation ca librat ion
Figure 7.1 Calibration based on three level s of representation
Figure 7.2 Proposed VISSIM calibration methodology
Figu re 7.3 Shockwaves representati on in vehicle trajectory data
Figure 8.1 Extent ofsimulation for USIOI NB, San Mateo, CA
Figure 8,2 ll!ustration of the test network
Figure 8.3 Test network speed-flow graph comparisons for different objective functions
Figure 8.4 Speed-flow graph comparisons for USlOl NB using different objective functions
Figure 9.1 Sample NGSIM relative velocity vs . relative distance graph
Figure 9.2 Relative velocity vs. rel ative distance graph: terms and definition s
Figure 9.3 Average following speed vs . minimum following distance
Figure 9.4 Average following speed vs. average following-distance
Figure 9.5 Average following speed vs. following distance variation
Figure 9.6 Average following distance vs . following distance variation .
Figure 9. 7 Histogram offollowing variation on US 101
Figiue 9.8 Histogram offollowing vari at ion on ISO
Figure 9.9 Speed oscillations in rel ative velocity vs. relative distance graphs
Figure 9.10 Maximum positive and negative speed oscillation vs. average following distance
Figure 9.11 Speed oscillations in re lati ve ve locity vs: relative di stance grap hs .
Figure 9.12 Rejected speed oscillations in relative velocity vs. relative distance graphs.
Figure 9.13 Closer look at speed oscillation in relative velocity vs. relative distance graphs.
Figure 9.13 Speed flow based calibration of US I 01 NB with varying numbers of parameters.
Figure 9.14 Flow Contours of maximum 5 m in flows sustained for varying amounts of time.
[CC4 = -0.70 and CC5 = 0.70]
Figure 9.15 Flow Contours of maximum 5 m in flows sustained for varying amounts of time.
[CC4 = -0.35 anfCC5 = 0.35]

4
LIST OF TABLES

Table 8.1 Calibration Parameters for Test Netwo rk

Table 8.2 Calibration Parameters for US IOI NB
Table 8.3 Speed-F low Objective Evaluati ons for US I0 I NB Speed-F low Graphs
Table 9.1 Descriptive Statistics of Following Di stance Variation: US I 0 I
Table 9.2 Descriptive Statistics of Following Distance Variation: 180
Table 9.3 Speed Flow Based Calibrat ion with Va ryi ng Number of Parameters

5
ABSTRACT
The research o n ana lysis of the Wiedemann 74 and 99 mode l parameters was conducted to better
understand the mode ls and their parameters. The resea rch a lso provides use r assista nce for
calibration of model parameters to capacity. In thi s research , a new generic ca libration
methodology based on integration of three di stinct leve ls of representation in ca libration is
developed. Level -! is traffic flow representation ofthe simul at ion model ; Level II is dissimilarity
representation of traffic flow representation; and , Level Ill is di ssi milarity represe ntation as
objective functi o n in optimi zation . In level-! representation , new innovative traffic flow
representations are described and argued to perform better than the state-of-the-art representations
of the simulation model. In le vel-11 representation , new pattern recognition based di ss imilarity
measures are developed for application in micros imul ation calibration. In levei-III
representation, the pattern recognition based methodologies are integrated with search-based
optimization methods. Thereafter, a generalized VISSIM calibration methodology is developed
to integrate all the three representations into the calibration procedure. The developed
methodology also accounts for data considerations in microsimulation calibration. A case
study of a real-world freeway network was performed to apply the benefits of the new traffic flow
representations like speed-flow graphs, pattern recognition based methods, and integration with
Evolutionary Algorithms (a search-based optimi zation method) . The results showed a better
calibrated simulation model capabl e of producing accurate results across a range of traffic flow
conditions. It was also shown that a calibration methodology based on matching speed-flow
graphs performs better than the currently used calibration methodologies based on capacity.
An integrated disaggregate-data and aggregate-data based calibration methodology was
utilized to calibrate the VISSIM driver behavior models. Theoretical analysis combined with bi-
level calibration methodology assisted in intelligent calibration of the VISSIM car following
model. Such calibration methodology provides more accurate estimates than one based on
"blind" calibration of the microsimulation model.
The analysis of the Wiedemann 99 model revealed that CC l and CC2 are the most
important factors while calibrating the model to capacity or speed-flow graphs. In addition, CC 1
has a much more significant effect on capacity than CC2. CCO, CC4, and CCS also have an
affect on capacity, but are not advised for use in calibration to capacity. The combination effect
ofCCO, CCl, and CC2 can result in multiple solutions for a desired capacity value. So instead,
CCO is not involved in capacity calibration, but can calibrated using other microscopic
information. CC4 and CCS also have an effect on capacity, because they control vehicle coupling
in following condition. But the longitudinal variation of the vehicle following is controlled to a

6
greater extent by CC2 . In addition , CC4 and CC S values ha ve been ca librated us in g th e NG SIM
vehicl e traj ectory data. The ca libration revealed symmetric CC4 and CCS thresholds at [-0.70
and 0.70] , whi ch are sli ghtl y hi gher than default values . It is al so identifi ed that CC4 is an
important parameter that control s shockwave propagation speeds more tha n any other parameter
in the Wiedemann 99 mod e l. Th e proposed values ofCC4 and CCS res ult in shockwaves speeds
that are close to real -world s peeds of approximately 12 mph . The other parameters, CC3 , CC6,
CC7, CC8, CC9, are not recomm ended for calibration to capacity or speed flow graphs.
User assi stance is al so provided in this research for the Wi edemann 99 model.
Calibration methodologies based on speed-flow graphs can be directly applied by practitioners
using simple pattern recognition based objectives. In case such information is unavailable, the
practitioners can use contour plots of capacity or maximum flow rates defined on CC 1 and CC2
plane to calibrate their simulation model. The flow contours are developed over varying values
of CC I and CC2. The practitioners with information about maximum flow and sustained time in
the field can select CC 1 and CC2 values to ea! ibrate their simulation model, hence saving
valuable time.

7
 I Introduction

1.0 I INTRODUCTION
The objectives of this study are multiple; first, to provide user assistance for calibration of the
Wiedemann 74 and 99 model parameters to capacity; second, to theoretically analyze Wiedemann
74 and 99 models and their pat'ameters; third, to develop a generic calibration methodology
capable of being applied to a variety of traffic variables . .
The first objective of providing user assistance for calibration of simulation model includes.
four tasks that are listed below.
1. Identify the relationships between driver behavior parameters and capacity.
2. Identify the most critical parameters,
3. Determine the various ranges of values of parameters from real-world data.
4. Provide calibration parameter values to attain a certain capacity

The second objective of theoretically analyzing the Wiedemann car following models and
their parameters is to better understand the model. The understanding will in turn help in
intelligent calibration of the car following model , rather than one based on " blind" calibration. It
is the intent of the study to use real-world traffic data in calibration of the car following model.
The real-world data will also shed light on the Wiedemann perception thresholds. This study also
utilizes state-of-the-art vehicle trajectory data provided by the Federal Highway Administration
(FHW A) through its Next Generation Simulation (NGSIM) community effort.
The third and final objective is to develop calibration methodologies and tool s that are
generic, simple, and robust. Although much of the study effott is directed to calibration to
capacity, the methodologies developed in this research are generic to allow reusability in other
scenarios.

8
 This report is divid ed into l 0 different chapters. Chapter I, Introduction, provides
obj ectives and motivation for thi s research . Chapter 2, VISS IM car following models, provides a
brief explanation of the Wiedemann car followin g models in VISSIM. C hapter 3, Calibration,
provides an extensive literature review on calibration methodologies, tool s, and technologi es .
Chapter 4, Traffi c Flow Represe ntation, in Ca li brati on is the first level in three levels of
representation in calibration . Chapter 5, Appli cations of pattern recognition in calibration, is the
second level in the three levels of representation in calibration. Chapter 6, Evolutionary
Algorithms in Calibration, is the third level of representation in calibration. Chapter 7, VISSIM
Calibration Methodology, introduces the proposed calibration methodology that integrates the
three levels of representation. Chapter 8, Case Study, compares the aggregate calibration
methodology based on speed-flow graphs with methodologies based on capacity. Chapter 9,
Implementation of Integrated Calibration Methodology, demonstrates the integration of
disaggregate and aggregate data based calibration methodology proposed in chapter 7. In
addition, it provides user assistance by providing calibration parameter values to attain a certain
capacity. Chapter I 0, Conclusions, provides a summary of the methodologies and findings of this
study.

9
 II VISSIM Car Following Models

2.11 INTRODUCTION
VISSIM driver .behavior models are classified as psycho-physical car following models, which
are based on modeling human perception and reaction thresholds in a car following process.
Todosiev (1963) and Michaels (1965) are among the first to research human perception and
reaction thresholds in a car following process. Wiedemann ( 1974, 1991) later extended the
research to develop the traffic simulation model MISSION that incorporated perception
thresholds to model car following behavior. Wiedemann ( 1974, 1991) defines four different car
following regimes in both the car following models .

Free-Flow: The vehicle is not influenced by any other vehicle; the vehicle . tries to keep its
desired speed, but fluctuates around its desired speed due to imperfect throttle control.

Approaching: Once the vehicle realizes it is approaching another vehicle, it decelerates to match
lead vehicle's speed as it reaches its desired safety distance

Following: In this driving condition, the following vehicle unconsciously follows the lead vehicle
and keeps the speed difference and acceleration low.

Emergency: If the vehicles following distance falls below desired following distance, it reacts by
applying maximum deceleration (within vehicular capabilities) to avoid collision.

The VISSIM microsimulation software has two different car following models: Wiedemann 74,
and Wiedemann 99. The Wiedemann 74 model is suggested for use in urban conditions, whereas

10
the Wiedemann 99 mode l is suggested for use in inter-urban motorway or freeway conditi ons.
The tvvo car following models are similar in that both are based on si mil ar human perception
thres holds. The difference lies in the ca lcul ation of perception thres holds . The Wiedemann 74
and Wiedemann 99 models were both developed by Wiedemann ( 1974, 1991 ). Descriptions of
the Wiedemann models are provided in the following sub-sections.

2.21 WIEDEMANN 74 CAR FOLLOWING MODEL

The Wiedemann 74 car following model is one of the two car-following models ava ilabl e 111

VISSIM, and is suggested for use in urban traffic conditions. The formulation of the Wiedemann
model is best explained using a relative velocity vs. relative distance graphs. Many of the
thresholds can be represented in these dimensions. Figure 2.1 describes a typical goal-seeking
behavior of a following car in a car-following process. A brief explanation of different
perception thresholds is presented below.

AX: is the minimum distance headway (front-bumper to front-bumper distance) in a standstill

condition

ABX: is the minimum desired followin g distance

SDX: is the maximum desired following distance

SDV: is the threshold at which driver recognizes that he is approaching a slower moving vehicle

OPDV: is the threshold for speed difference in an opening process [following vehicle moving
away from the lead vehicle] during a following condition

CLDV: is the threshold for speed difference in a closing process [following vehicle moving
closer to the lead vehicle] during a following condition

The minimum desired following distance of a vehicle in a following condition is proportional to

the square root of the slower vehicle's speed. The slower vehicle can be either the lead vehicle or
the following vehicle. In addition, the following vehicle behavior is also controlled by number of
observed vehicles [vehicle following behavior parameter] downstream of preceding vehicle.

II
 ABX = !LX+ (bx _ add+ bx _ mull * N [0 .5, 0 . 15]) * v,""' "'

The state upd ate mechanism for each dri ver-veh ic le unit is perform ed in d iscrete-time steps and is
shown in fi g ure 2.2. The acceleration computat ion is based on the ca r fo ll owing regim e.

t::, x d istance headway

perception
no rea ct io n threshold

unconscious
.......--

difference of velocity t::, v

.__ increasing distance decreasing distance __.

Figure 2.1 Wiedemann 74 car following model thres holds

Calculation of D.x and t,.v

t,._y = xn-l -XII

BRAKING

yes

APPROACHING

FOLLOWING
no

FREE FLOW

vn (t + t,.t )

Figure 2.2 Car Following Logic in Wiedemann Model [adapted from Ehl e1t (2004)]

12
 2.21 WIEDEMANN 99 CAR FOL LOWING MO DEL
The Wiedemann 99 car following model is very similar in relati on to the Wiedemann 74 model.
The core execution or logic (figure 2.2) in Wi edemann 99 remained the same; however, some of
the thres holds are calcul ated differentl y from Wiedemann 74 model. The thres holds described in
the Wiedemann 74 model have the same meaning in the Wi edemann 99 model , but are calcul ated
in a different way. In order to explain the difference, ten Wi eclemann 99 calibration parameters
are defined as follows.

CCO: defines the desired rear bumper-to-front bumper distance between stopped cars. This
parameter has no variation .

AX = CCO + L"_,

CCl: defines the time (in seconds) the following driver wishes to keep. The VISSIM manual
(PTV AG, 2007) reports this as time headway, but this term does not have its traditional meaning,
because it does not define front bumper-to-front bumper time differences, rather a time headway
to the standstill distance headway.

ABX = AX + CC J * V , Jm.-er

CC2: restricts the longitudinal oscillation during following condition. In other words, it defines
how much more distance than the desired safety distance (ABX) before the driver intentionally

(
moves closer to the lead vehicle.

SDX = ABX + CC2

CC3: defines the start (in seconds) of the deceleration process; i.e., the time in seconds, when the
driver recognizes a slower moving preceding vehicle, and starts to decelerate.

SD V: 6.,, = CC3 Llv + SDX + CC3 (-CC 4)

13
 CC4 and CCS: define the speed di ffe rence (in m/s ) durin g the fo ll ovv in g process. CC4 contro ls
speed differences d uring c los ing process, and CCS co ntrol s s peed differe nces during an opening
process.

CLDV = -CC4

OPDV = -CC5

CC6: defines the influence of di stance on speed osc illation during following condition.
Increasing CC6 increases the oscillation of speed with increasing dista nce ..

CC7: defines actual acceleration during oscillation in a followin g process .

CC8: defines the desired acceleration when starting from a stand st ill.

CC9: defines the des ired acceleration when at 80km/hr. However, it is limited by maximum
acceleration allowed for the vehicle type.

The ten parameters defined in above are the Wiedemann 99 calibration parameters. Some of
them define perception thresholds , while others define desired driver behavior parameters . .

2.31 VEHICLE FOLLOWING BEHAVIOR PARAMETERS

In addition to the parameters described in Wiedemann models, there are a few parameters that are
common to both the models. Descriptions ofthese parameters are provided below.

Look-Ahead Distance: de fines the distance that a vehicle can see forward in order to react to
vehicles either in front or side of it (within the same link). This parameter is in addition to
number of observed vehicles.

Number of Observed Vehicles: controls the number of vehicles affecting a vehicle's movement.
In addition to normal vehicles, some of the elements in the simulation model are modeled as
vehicles; therefore, increased values are necessary near cross-sections with higher number of
network elements

14
Temporary Lack of Atte ntion (Du ration and Probability): defi nes the probability and durat ion
of lack of attent ion. However, these parameters are not imp lemented in Emergency driv ing
co ndition.

15
 Ill Micrositnulation Calibration

3.1 I CALIBRATION
Calibration is arguably one of the most important steps in the traffi~ simulation model
development process. Calibration has always been an important issue since the advent of traffic
simulation models. The validity of measurements and decisions made using simulation models
are often directly proportional to the time and effo11 devoted to calibration. Development of a
traffic simulation model that can replicate a vari ety of driving behav iors that exist in the real-
world is practically impossible. Instead, traffic simulation models have parameters that can be
tweaked to reproduce such a variation in driving behavior. Calibration is the process of varying
such parameters to match local driving behavior. Without calibration, traffic simulation models
cannot be expected to provide accurate results.
Since this study effort is dedicated to calibration of the Wiedemann car following models,
there is a need for a comprehensive literature review on calibration methodologies. The literature
review on calibration is presented in the next section. The calibration methodologies described
are for the most extent generic and are not microsimulation software specific .

. 3.2 I LITERATURE REVIEW ON CALIBRATION

Calibration of microsimulation models has always been an important iss ue since the sta11 of
development of microsimt:llation models in the early 1950s. Various methodologies have been
proposed for calibration ofmicrosimulation model s. There has been much literature published in
the field of calibration. Research efforts vary from "strategies" or logical steps to perform
calibration to optimizations methods employed in the calibration process. But the goals of all
such research efforts remained the same; i.e., achieving efficiently calibrated models that
reasonably replicate local traffic behavior or conditions. This reasonable replication is achieved

16
when the calibrated simulatio n models meet des ired co nstraints or "targets". The met hodol ogy
for cali bration of simulat ion mode ls is constantly altered to suit research or organi zational goa ls.
In addition, there has also bee n co nsiderable research effott directed towards developing
universa l calibration methodologies.
Calibration ofmicrosimulation process in volves varying a wide variety ofparameters to
match the local traffic behavior. Calibration by itself can be a very time consuming task, due to
the large number of parameters involved . Since parameters can have effects that go beyond a
single measure of performance due to correlation, calibration requires an iterative process to he I p
effectively reach an optimal solution. A variety of methods have been used to calibrate
simulation models. These methods were largely dependent on research objectives and the
simulation models. Traditionally, the type of calibration parameter and specific nature of the
simulation model determined the calibration methodology.
In the initial years of research in traffic simulation, much of the calibration effmt was
directed towards specific car following models. Various car following model s were calibrated to
the field conditions based on experimental data. The experimental data varied from test track
data to vehicular tunnel experiments. A common theme among such experiments is the use of
relative velocities, relative distances, speeds, and accelerations of the leading and following
vehicles. General Motors' and New York Pott Authority's experiments were based on two cars
connected to each other with a cable on a pulley. These experiments were used to determine the
ranges for the GM car following model sensitivity parameters. But most of these results were
limited to specific scenarios, due to the unavailability of large data sets. Hoefs and Leutzbach
( 1972) and Hoefs ( 1972) collected relative speed, relative velocity, and speed data that was used
to calibrate perception thresholds in psycho-physical car following models of Michael s ( 1965)
and Todosiev (1963).
The initial years of calibration were dedicated to calibrating specific car following
models; therefore, lacking a comprehensive generalized methodology. Many of the general
calibration techniques were primitive and only concentrated on modifYing a few of the calibration
parameters. In addition, microscopic vehicle trajectory data used by some of the authors did not
model the whole range of driver behavior and traffic conditions (e.g. Benekohal and Treiterer,
1988). Over the next decade, researchers developed "strategies" for calibration (e.g. Cheu et al.,
1994). Many of the calibration objectives were based on reducing the difference between
observed and field capacities. Many of the researchers (e.g. Aycin and Benekohal, 1998; Payne
et al., 1997; Radwan et al., 1991) calibrated their simulation models based on single or averages
values of traffic variables, such representation usually results in a significant loss of information.

17
A ca libration method that is based on distributi on of traffic variab les wou ld have resu lted in a
better calibrated s imul at ion model. C heu et al. ( 1994) ca librated their s imul ation mode l to
produce 30-second average vo lume and occupancy data si milar to the Ca li fo rnia Department of
T ransportation (Caltrans) detector output. In the ir calibration meth odo logy, scatter plots for th e
vo lume and occupancy data fro m the simul at ion and the field we re used. The corre lation
coeffic ient and the slope of fitted line th at passes through the origin of the scatte r plot was used as
performance meas ures in ca librat ion . A ca libration methodology based on matc hing volume over
time at loop detectors wo uld result in a much better calibrated mode l (e .g. Rakha et al. , 1998);
however, such methodologies need prior OD information.
Cheu et al. ( 1998) used a genetic algorithm as an optimi zation and search tool for
calibrating a FRESIM simulation model. The objective of the calibration process was to reduce
the difference in volume and speed counts over detectors. The sum of the negative exponentiai of
the absolute-average error terms for speed and volume terms over time was used as fitness
function in the genetic al gorithm. Exponential terms were used to reduce incompatibility
between units for speed and volume. Ma and Abdulhai (2002) used a genetic algorithm based
optimization for calibration. The calibration methodology for calibration is based on reducing the
difference between observed and simulated volume counts. The Four different types of measures
that were used in the study are Point Mean Absolute Error (PMAE), G lobal relative Error (GRE),
Theil's Inequality Coefficient, and Point Mean Relative Error (PMRE). Both the studi es, Cheu et
al. (1998) and Ma and Abdhulai (2002), and several other studies demonstrated the usefulness of
genetic algorithms as optimization tools in a calibration process.
Rakha and Crowther (2002) proposed a unique calibration methodology based on steady-
state car following behavior. In this methodology the steady-state behavior of various car
following models such as CORSIM, VISSIM, and Van-Aerde are used to develop a macroscopic
traffic stream model. Then the car following sensitivity parameters are proposed as so lutions to
such macroscopic relationships. This methodology can provide good ball-park numbers, but is
not suitable for accurate calibration. Gardes et al. (2002) describe a general calibration
methodology that involves both qualitative and quantitative aspects of simulation . Qualitative
assessment such as vi sual analysis of vehicle movements is used in calibration . Toledo at a l.
(2003) and Toledo et al. (2004) described a two step calibration procedure as part of the ir
calibration framework . The calibration framework is replicated in Figure 3.1. The calibration
procedure is divided into two levels: disaggregate and aggregate. In the first step, parameters in
individual models such as driver behavior and route-choice models are statistically estimated
from disaggregate data or microscopic data such as vehicle trajectory data. In the second step, the

18
genera l parameters are calibrated using aggregate data s uch as speed, fiow, time headways, etc.
The aggregate cal ibration is formu lated as an optimization o r sea rc h problem . The authors stress
the importance of the inter-relationships between the 0-D matrices, route-cho ice parameters, and
d ri ve r behavior param eters and their effect on ca libration. The authors propose using an iterat ive
approach to calibrate route choice parameters and esti mate 0- D mat ri x, and an overall iterat ion
with driver behavior parameter until convergence. This methodol ogy, although sound, requires
significant amount of time, due to repeated driver behav ior ca libration. Driver behavior
ca libration is perform ed using repeated eva luati ons of the s imul at ion mode l, hence requiring a
significant amount of time.

Initial P<1i'<llll8ter vnluo:;s.

seecl OD m atrix

minZ(ODJi3)

Ca librati on of
route cl1oice ancl
minZI.I3JOD)
driving t' e il<~vior
pnrometers

Figure 3.1 Solution to steady state calibration approach [Toledo et al., 2004]

Hourdakis et al. (2003) described a generic calibration methodology for freeway simulation
models. Initially, the parameters are divided into two categories: global parameters such as
desired speeds, vehicle character istics, maximum acceleration and deceleration, and minimum

19
 headway; local paramete rs s uch as speed limits, and other link specific paramete rs. T he
calibration process is then divided into mainl y two stages: volume, and speed-based calibration.
In addition, an optional third stage involvin g a research objecti ve-based calibration for fine tunin g
of the simulation model is also suggested. Between the first two stages, the authors suggest that
volume-based calibration be followed by speed-based calibration. Dowling et al. (2004) describe
a generic calibration framework for s imulation models . The authors sugges t dividin g the
calibration paramete rs into two categories . One category includes parameters that the use r is
reasonably certain and not willing to adjust , and second category of parameters the user is
unce1tain and willing to adjust. The authors then suggest dividin g the adjustable parameters into
parameters that directly impact capacity and parameters that directly impact route choice. It is
also suggested that each of the parameters be classified into global or li1ik specific parameters.
The authors then propose the use of three-step methodoiogy: capacity calibration, route choice
calibration, and system performance calibration. In addition, the authors propose that calibration
of the model to capacity be one of the first steps in microsimulation calibration. It is also
suggested that queue discharge flow rate be used to estimate a numerical value for capacity. The
calibration procedure is performed until the calibration targets are met, since otherwise
calibration can be a never-ending task. These calibration targets are usually based on
organi zational and research goals.
Brockfield et al. (2004) describe calibration of car following models based on data
obtained from test track in Japan . The calibration of the car following models was based on
reducing the percentage absolute error in gaps measured in simulated vs. the field data, which in
other words is similar to matching headway distributions. Bayarri et al. (2004) describe a unique
calibration methodology based on Bayesian analysis techniques. The authors suggest dividing
calibration parameters into three categories: parameters than can be estimate form field data
(vehicle mix, arrival rate, turning percentages); parameters that are not measurable or based on
choice (discharge headway distribution); parameters that are not "real" but required for tuning by
the simulation model (free-flow speed). The authors dedicated much of the effort to
demonstrating their Bayesian methodology for demand and turning probabilities.
Gomes et al. (2004) describe a unique calibration methodology of matching the location
of bottlenecks, queue lengths, and HOV lane utilization for congested freeway simulation model
in VISSIM. The calibration methodology was based on qualitative as well as quantitative
measures of differences in simulation and field data. Kim et al. (2005) used a non-parametric
statistical method for calibration of VISSIM simulation model. The authors used a genetic
algorithm as an optimization tool in the objective function. The method is a distribution free

20
method where the simulation output is compared to field data usin g Moses ' distributi on free rank-
like test and Wicoxon rank-sum test. T he poss ible so lutions that pass th ese two tests are se lected
to the next ge neration in the ge neti c a lgo rithm. T he comparative fitn ess va lues of possible
so lutions are evaluated us ing a Ko lmogrov-Sm irnov test.
Park and Qi (2005) descri be a multi -stage calibration methodology for s imulation
model s. In the first step, an initi a l eva luation of default paramete rs is performed by comparing
fi e ld and simulation outputs, and default parameters are accepted if it is a close match. In the
second stage, an initial calibration parameter set is identifi ed. A Latin Hypercude Design (LHD)
is used to develop multiple paramete r sets that reasonably cover the parameter surface while
keeping the set size to a manageable number. Within the second stage, additional feasibility tests
and Analysis of variance (ANOV A) tests are performed to identify key parameters and ranges.
This stage is followed by the third stage involving parameter caiibration us ing genetic algorithms.
And finally, the genetic algorithm recommended parameter set is evaluated using statistical
techniques. A flow chart summarizing the following methodology is provided in figure 3.2.
Over the recent years , researchers stmted developing systematic calibration frameworks
for application in microsimulation calibration (e.g. Dowling et a l. , 2004; Toledo et al., 2003,
2004). The concept of calibra tion targets were also introduced to assist in optimization process.
Most of the objective functions for measuring the difference between the field and measured
discharge flows are based on different statistical goodness-of-fit measures (Dowling et al., 2004;
Toledo et al., 2003, 2004; Schultz et al., 2004). Games et al (2004) described qualitative efforts
based on manually eye-balling the differences in the field and simulated graphs. However, these
methods lacked a quantifiable or mathematical basis for accessing the differences . K.im et al.
(2005) repOtted the use of non-parametric statistical methods for comparison of distributions, but
such methods are not applicable in measuring the degree of closeness of graphs . Much of the
research still suffers from poor traffic flow representation of the model. And in cases where a
good representation is used, no quantifiable methods are proposed to measure the differences .

21
 Slimrlatlon Model setup

I.Study-selec
- S t t~
st:.ope
tion
~ncl purpose

- Dcte:rmrn.Jtron of m e~su res of effecavr:1~e~5

- Freld dat<.1 collcctton
- Nel 1.'iOrk co(!rng

Initial Calibration
- ldentifr cation O rcnli.br.Jtion par.Jmeters
- E xpen m ent~ l design for c.:>ltbr,1tron (l,1Ho 1- - - - ,
Hypercube Desrgn)
- ~-lu lt ip l e runs

/ ~ .. No
< feo sibllity Te~t :.-------'
~ .......... ~n sse d ~ ........--
~~--~ .. ~"" ---

,_ End

Figure 3.2 Latin Hypercube Design (LHD) based calibration methodology [Park and Qi, 2005]

3.3 I CLASSIFICATION OF OBJECTIVES IN CALIBRATION METHODLOGIES

After an extensive review of literature on calibration methodologies and frameworks proposed by
many researchers over the decades. The objectives in calibration methodologies can be broadly
classified into two different categories: Qualitative, and Quantitative. Each of the categories can
be classified further into various subcategories.

3.3.1 I Qualitative Methods

These methods, as the name suggests, are methods that lack mathematical or statistical
foundation. It is not that such methods cannot be expressed in mathematical or statistical form ,
rather such representation is often complex, or more appropriately beyond the scope of the study.
In qualitative methods, authors usually rely on their perception of reality to decide if the

22
s imul atio n model is performin g c lose to reality. For exa mple, in test ing car fo ll ow in g model s,
authors subj ect a lead vehi cle in the platoon to speed d isturbances that range from moderate to
extreme, and observations are made on the stability of the car following process of the platoon,
and compared to author 's perception of rea lity; e.g., Za rea n and Nemeth (1988). A similar
example re lates to Gmdes et al. (2002), a nd seve ral oth e r studi es, is the use of visual analysis of
vehicle movements in a traffic network to calibrate s imul at ion models. Such methods a re usuall y
unde1taken to rul e out " bad" calibration parameter va lues form possible parameter so lutions. A
very good example of qualitative meas ures in calibration is Gomes at al. (2004), wherein the
authors matched speed-contour graphs obtained from the field and simul ation . This method is
similar to matching the extent of congestion in a traffic network. In this process, the authors'
primary intent was to vary certain calibration parameters to replicate bottleneck locations, and
extent of queues in the network. These types of qualitative methods are applied all through the
calibration process to tweak the model.

3.3.2 I Quantitative Methods

Quantitative methods are methods that are based on mathematical or statistical foundation . Due
to a wide variety of quantitative methods in calibration, sub categorization of these methods is
necessary to gain a better understanding of the differences. The categorization is loosely based
on different concepts like parameters and their representations , and procedural techniques .

Single Parameter Methods are based on measuring the difference between field and simulated
parameter values. There are many statistical goodness-of-fit measures for measuring the
difference between the field and simulated values (e.g. Rillet et al:, 2001; Ma and Abdulhai,
2002). Examples of parameters that are typically used in calibration include capacity, average or
maximum queue lengths, average travel time, total travel time, many others. Much of the studies
in the current state-of-the-art utilize such simple measures of difference to calibrate their
simulation models. Typical examples of statistical goodness-of-fit measures for single parameter
values include the following:

1. Mean Error (ME)

2. Mean Absolute Error (MAE)
3. Mean Absolute Error Ratio (MAER)
4. Mean Square Error (MSE)
5. Root Mean Squared Error (RMSE)

23
 6 . Mean Perce ntage Error (MPE)
7. Mean Abso lute Percentage Erro r (MAPE)
8. Root Mean Sq uared Percentage Error (RMSPE)
9. Root Relative Sq uare Error (RRSE)
10. Relative Abso lute Error (RAE)

For a detailed expl anation about these stati st ical good ness-of-fit measures , the reader is directed
to any standard stati stica l reference. Each ofthe error has its own advantages and di sadvantages .
Selection among these meas ures is usually based on research objectives, goa ls, and nature of the
problem . Discussions about "noise", outliers, weights for the errors, etc is important to select the
appropriate meas ure.

Multiple Parameter Methods are similar in nature to single parameters methods. In multiple
parameter methods the errors from different parameters should be weighted according the units
and importance. The objective can include penalties to include different constraints. A typical
example is calibrating of simulation models with respect to volume and speed (e.g. Toledo et al.,
2004).

Time-Series Methods include error accrual of multipl e parameters over time. Various multiple
parameters methods like weights, penalties are also applicable to these methods. Examples of
such methods include calibration of simulation model s to microscopic vehicl e trajectory data,
volumes over time, speeds over time, etc. (e.g. Toledo et al., 2003)

Statistical Parametric Methods are based on comparing di stributions of parameters, where it is

assumed that the population is assumed to fit a parameteri zed distribution . Some of the typical
assumptions include either a normal or a uniform distribution. Common methods employ
measures based on mean and variance . Some of the popular tests include the z-test, t-test, F-test,
ANOVA, etc.

Distribution-Free or Non-Parametric Statistical Methods, unlike parametric measures, do not

have an underlying assumption about the nature of the distribution of the population. Typical
examples of non-parametric statistical methods include Kolmogrov-Smirnov test, chi-square test,
etc. (e.g. Dixon et al., 1997; Kim et al., 2005)

24
Other Methods there a re methods that are different from the o nes described abo ve. Some of the
methods are heuristic in nature and spec ific to the problem at hand. So me exa mpl es of these
ca libration methods incl ude methodo logies wherein a s imultaneo us iterative app roach is used.
Bayesian inference techniques have also found use in some of the calibration me thods (Bayarri et
al. , 2004). Steady-state behav ior based car followin g ca libration methodology was shown in
Rakha and Crowther (2002)

3.4 I DATA CONSIDERATION IN MICROSIMULA TION CALIBRARTION

Data Considerations are important in the developm ent of calibration procedures for simulation
models. It is paramount to incorporate data cons iderations and requirements into the
development process. It is the intent of the study to develop a generic calibration methodology
based on such considerations. A brief description of the data collection efforts and types of data
is provided in this section. Comprehensive efforts were directed towards data collection to obtain
a variety of microscopic and macroscopic real -world data. Data was obtained from a variety of
sources including Federal Highway Administration (FHW A), state depa1tments of transportation
(DOTs), and Traffic Management Centers (TMC). Data collected for this project is divided into
two main categories: microsco pic and macroscopic.

3.4.1 I Microscopic Data

Microscopic Data includes detailed sub-second information about vehicle trajectori es over time.
The main source for microsco pic data for this research is provided by Federal Highway
Administration. FHWA collected detailed sub-second vehicle trajectory information as part of
Next Generation Simulation (NGSIM) program. The data is available for free to the research
community through the World Wide Web (WWW). Microscopic data available from the NGSIM
effort includes sub-second vehicle position, speed, acceleration, headway, and spacing
information. The data also includes information about following and leading vehicles for each of
the vehicles. The data collected from this source include detailed vehicle trajectory information
from two freeway sections and one arterial section . The two freeway sections include I-80 and
US 101 freeway sections. NGSIM has put forth data sets from three different locations: I-80,
US101, and New Lankershim Boulevard [arterial dataset]. The length of each of the data sets is
approximately two to three IS minute periods.

25
3.4.2 I Macroscopic Data
Macroscopic Data includes aggregated vehi c le information. Thi s data is ty pically obtained from
state DOTs and TMCs. The aggregation leve l va ri es form 20 seconds to 15 minutes. Fo ur
different so urces of macrosco pic data were ide ntifi ed for this research.
1. Performance Management Sys tems , Ca lifornia Depa11ment ofTranspo11ation (CaiTrans)
2. Traffic.com, St. Louis, Missouri
3. Pot11and Traffic Management Center, Oregon Depat1ment of Transportation
4. NAVIGATOR, Atlanta Traffic Manageme nt Ce nter, Georgia.

Macroscopic data available from the sources mentioned above are based on loop detectors or
side-fire microwave detectors on freeways. The information available ranges from 20-second to
i 5 minute aggregate volume, speed, and occupancy counts. All detectors do not have speed
information, since most of the detectors are single loop detectors. Any s peed information from
single loop detectors is biased information. But, there are some detectors with speed information

coming from dual loops, and microwave detectors.

26
 IV Traffic Flow Representations in
Calibration

4.1 I INTRODUCTION
Since the simulation model and the field scenario cannot be compared directly, traffic flow
variables are used to represent the field scenario and the simulation model during the calibration
process. The assumption being that by reducing the difference between traffic variables, the
difference between the simulation and the field is simultaneously reduced . Hence, traffic flow
variables or representations (henceforth referred to as traffic flow representations) are an
important pmt of calibration of s imulation models (see figure 4.1 ). This chapter concerns itself
with level-1 representation ofthe simulation and field .
The objectives and methods used in calibration methodologies have a great impact on
traffic flow aspects of microsimulation models. Calibration methodologies that do not account
for traffic flow concepts are inherently flawed . A calibration model that only concentrates on
reducing the difference between the field and simulated variables, does not account for flow
process that exist beyond those variables. In addition, since many of the calibration parameters
affect multiple flow processes, it is impotiant to account for such interactions during the
calibration process. Calibration of a microsimulation model involves tweaking a wide variety of
parameters. It is important to only expose calibration parameters that significantly affect the
traffic variables (capacity, average travel time, queue lengths, etc) that are being used for the field
vs. simulation comparison in a calibration methodology. If there is a parameter that significantly
affects multiple traffic variables, it is important to constraint such a parameter so that it doesn ' t
severely affect the variables that are not being used in the calibration process .

27
 In this chapter, vano us traffic flow representations are st udi ed by eva luating their
suitability for use in a microsimulation calibration process. In addition, combinations and
relationships based on traffic var iabl es such as speed-flow, speed-occupancy, and flo w-
occupancy are also studi ed. Since th ere are a vast number of traffi c flow represe ntations , study
of all such variables is beyond th e scope of thi s researc h. On ly the most e leme ntary and
s ignificant variables are addressed in thi s research. Some of the concepts addressed include
capacity, flow/s peed/density time series, speed/flow/density re lationships, speed/flow/density
contours, shockwaves, headway distributions, and relative distance vs . relative ve locity graphs.

cs~~~~~~n Mo~~ ~Id Scenario

c===_l ___]
J J
I"'-------~-. . ,.,_.,. .___. L-~---~-
. ---.. . .~ -
Traffic Flow Variables and
Relationships
Capacity Representation Level 1
' Speed-Flow

l---~-"----~~~~Co:~~"--- u- --"'_J
Figure 4.1 Traffic flow representation of the simulation model and the field scenario.

4.21 AGGREGATE-DATA BASED REPRESENTATION

4.2.1 I Capacity
Capacity is one of the most important variables in traffic flow theory that is often used by both
researchers and practitioners. It is important study the applicability and suitability of capacity in
calibration. As it can be seen in the literature review presented in the previous chapter, capacity
is considered as one of the primary variables of interest, and often part of the first level of
calibration. As it will be seen through the next few sections that there are some unanswered
questions about how capacity should be used in the calibration process. To gain a better
understanding, a state-of-the-art literature review on capacity is presented. Discussions have also
been presented on tradeoffs of using capacity in microsimulation calibration.
The capacity estimation procedures are important, because calibration of the driver
behavior parameters is usually performed by minimizing the difference in simulated and the field
capacity values. In addition, one of the primary objectives of this study is to provide calibration

28
parameter values to attain a certain capacity. And also, it is important that capac ity est imation
procedures are consistent with field es tim atio n procedures. However, the problem ste ms from the
different interpretations of HCM 's definition of capacity. For example, FHWA microsimulation
guidelines for simulation recommend procedures, which are based on the assumption that
capacity is the queue discharge flow rate. Different views exist on the definition of capac ity, for
example, whether capacity is the queue discharge flow rate (QDF) or the pre-queue flow rate
(PQF).

4.2.1.1 1Capacity Concepts in Highway Capacity Manual

The standard source of reference on traffic facility operations is the Highway Capacity Manual
(HCM), which defines freeway capacity as follows:
"The capacity of a facility is the maximum hourly rate at which persons or vehicles
reasonably can be expected to traverse a point or a uniform section of a lane or roadway during
a given time period under prevailing roadway, traffic, and control conditions. "
This definition of capacity raises questions, which are necessary to be answered for
application in simulation and the field data; however, a few ciarifications have been provided in
continuation with the definition (pg.2-2). The first clarification states that capacity is not the
absolute maximum flow rate observed at a facility, and that reasonable expectancy is an
important part of defining capacity. From the definition it can be inferred that it the value of
capacity is selected from a distribution of maximum flow rates observed at a facility over several
days. However, there is no guidance provided for selecting a value representing capacity from
the distribution. Thevarious possibilities include mean, median, mode, 85th percentile, etc. The
concept of reproducibility of capacity is also introduced in this definition. A slightly different
definition is provided for freeway facilities (pg. 13-2). This alternate definition of capacity
introduces the concept of sustainability in capacity. The definition also transforms maximum
hourly flow rate to 15-min flow rate expressed in cars per hour per lane, which introduces
passenger car equivalents . But, HCM accepts vehicles per hour as an acceptable measure of
capacity. The definition also assumes no influence from downstream traffic operations. The
definition does not provide guidance on amount of time the flows need to be sustained. It is not
clear if maximum sustained 15-min flow rate refers to which ofthe following:

1. The maximum sustained count-period (i.e., x-minute, xis less than 15 minutes) flow rate
over a period of 15 minutes. For instance, capacity is maximum 5 minute flow sustained
over a period of 15 minutes.

29
 2. T he maximum s usta ined 15-min flow rate susta ined ove r a peri od of time. For insta nce,
capac ity is maximum 15-mi n flow s ustained fo r 60 minutes .

In summary, some of the important concepts, in HCM 's defi niti on of ca pac ity, whi ch need to be
incorporated into ca pacity estimat ion procedures, in clud e the follo w in g:
1. Reasonabl e expectancy
2. Reproducibility
3. Sustain ability

The Highway Ca pacity Manual ' s (HCM) capacity definitions and its various clarifications (HCM,
2000), pg. 2-2, 13 -2) could be interpreted in several ways, e ven though HCM clearly notes some
of the important concepts underlying the definition of capacity like reasonable expectancy,
reproducibility, and sustainability. All these concepts cannot be easily captured through a single
numerical value of capacity. These concepts are important in order to implement the capacity
calibration process in the calibration procedure. Capacity is an important parameter that defines a
facility and its operational capability. However, defining capacity as a single numerical value
results in loss of information . A di stribution of capac ity values has more infonnation than a
single numerical value. The re have been recent studi es which address the stochastic nature of
capacity (Brilon et al. , 2007). If the ca pacity calibration process is based on a single numerical
value, matching the means of capacity di stribution does not necessaril y match the other important
prope1iies of a di stribution; for example, spread, shape, and median .
This concept can be extended further to include other traffic parameters. like speed . For
instance, is it impmiant to match the speed at which capacity values are observed? There are
many studies that demonstrate that there are two different types of maximum flo ws observed in
the field, one is queue discharge flow (QDF) and the other is pre-queue flow (PQF) (Banks,
1991 ). Each of these flows is sustained for a different amount of time. With-respect-to
calibration, is QDF or PQF more impmiant to be matched? The author suggests the matching of
both values . In summary, it is important to maximi ze the information used during the calibration
process.

4.2.2 I Speed-Flow Graphs

Speed-flow graphs have been used to describe operational capabilities of highways, and in
developing macroscopic relationships for freeways. It is relatively easy to collect speed, flow,
and occupancy information due to wide use of instrumentation on freeways. Speed-flow graphs

30
contain much information as described below. It is the intent of this resea rch to use such
information ava ilable in speed-flo w grap hs. This methodology is based on matchin g speed-flow
graphs obtained from simulation and field. The HCM (2 000) divides speed-flow graphs into
three regions: free-flow, congested, and queue discharge. T he HCM concepts of capacity are well
represented through speed-flow graphs. As a result, capacity information is available in a speed-
flow graph. QDF and PQF information can also be derived from speed-flow graphs. In addition ,
speed-flow graphs also provide information about free -flow and congested reg ions, which is not
present in a si ngle numerical value or distribution of capacities.
A calibration procedure based on speed-flow graphs, which provides information about
all the three regions: free-flow , congested, and queue discharge could replicate the whole range of
traffic behavior and not just peak period. One could also just use a portion of the speed-flow
graph instead of the entire graph for calibration such as the queue discharge region.
It is important to recognize that speed-flow graphs lack information about time.
However, the intensity of points in speed-flow graphs contains pmiial frequency information.
Hence, the concept of sustainability is only pmiially captured in a speed-flow graph. An alternate
graph, maximum flow sustained time graph, based on the concept of maximum flow sustained
over different periods of time provides a better picture of the flow sustainability concept in
capacity definition. However, due to the stochastic nature of both real and simulated traffic
. flows; it would not be possible to exactly replicate a volume-over~time curve. Despite the fact
that fluctuations and breakdowns will not occur at exactly the same time in the real -world as in
simulation, the simulation model is nevertheless correctly calibrated.
The concept of replicating field speed-flow graphs has been used in a number of previous
studies. Wiedemann (I 974, 1991) used speed-flow graphs to demonstrate closeness of field and
simulation model. Fellendorfand Vortisch (2001) demonstrated the ability of a simulation model
to replicate speed-flow graphs from real-world freeways. However, there was no literature found
that used speed-flow graphs in the microsimulation calibration process. Ngoduy et al. (2004)
used an objective function based on speed and flow, which is mathematically close to replicating
speed-flow graphs, for calibration of a macroscopic simulation model. The calibration of speed-
flow graphs is just one step in microsimulation calibration and should be followed by route-
choice calibration and system performance calibration.
Many of the research studies based on comparison of speed-flow graphs were based on
manually eye-balling the differences. However, it is impmiant to quantifY the differences in the
graphs. It is also beneficial to automate the process for use in optimization methods. The research

31
of quanti tying the differe nces, auto mation, and integration with opt imi zat ion methods is
presented in chapter 5 and 6.
In developing a speed-flow graph, the importance of location for co ll ecting speed-flow
graphs has been de monstrated by May (1990). Data can be co llected over different locations and
multiple days and co mbined to show a co mpl ete speed-flo w graph . In terms of fie ld data, speed
and flow information for the section under co ns ideration can be co ll ected for instrumented
highways over many locations and differe nt days. The simulation data can be co llected by
placing detectors at exactly the same locations as detectors in the field. Sa mpl e speed-flow
graphs developed in the microsimulation model are presented in fi gure 4 .2. The methodology
presented here is equally applicable to matching speed-flow graphs at individual detector stations .
The objective would then be stated as a combination of individual detector stations. But to use
individual detector stati ons, prior OD data and route choice information is necessary.

4.2.3 I Flow-Density, Speed-Deilsity, and Speed-Flow-Density Graphs

In continuation with arguments in the previous section, it can be seen that other relationships that
exist between pairs of flow, speed, or density variables can be a valuable source of information.
Speed-density or flow-density relationships are similar in nature to speed-flow graphs. These
relationships offer greater amount of information than capacity. If a calibration procedure is
solely based on capacity, there can be resultant changes in calibration paramete rs that affect the
relationships that exist between speed, flow and occupancy. For instance, we mi ght calibrate our
simulation model to achieve a field capacity value, but end up with two different speed-flow
graphs . The logical thing to say then is to calibrate the model to achieve both the speed-flow
graph and also capacity. Again, intuitively, matching speed-flow graphs should match the
capacity values, since capacity information - to a certain extent - is inherent in speed-flow or
flow-occupancy graphs. It is important to recogni ze that speed-density graphs have no
information related to capacity, due to lack of flow dimensions. However, this concept can be
further extended to a tluee-dimensional speed-flow-density graph, which has information in all
three dimensions, and hence, higher fidelity.
An important dimension that is missing in all the relationships between speed, flow, and
density is time. Although, partial information of time is available in each of the relationships
tluough intensity (or frequency), it is not complete. Temporal information is impotiant and does
relate to the concept of sustainability discussed in the capacity section. To account for time,
concept of flow sustained over time is introduced.

32
 Wiedema nn 99 DCs- 100 to 2900
80 .-------~------~-------,-------,------~-------,-,

60

50
~
E
c
-gV 40
0.
(/) +
Ol
>
<( 30

20
+ 1;.
+ :+

0 500 . 1000 1500 2000 2500 3000

Flow in veMlrAn

Wi edemann 74 DCs - 1 to 29
80 .-------,-------.--------.-------.-------.------~-,

++
60

~ 50
-EO
E
.!:::
-gV 40 +
0.
(/)
;+ +
5"
<( 30

20

10

QL-------~------~-------L------~------~------~_J
0 500 1000 1500 2000 2500 3000
Flow in veMlrAn

Figure 4.2 Typical VISSIM generated speed-flow graphs

4.2.4 I Maximum Sustained Flow Time Graphs

Maximum Sustained Flow Time Graphs (or MSFTGs) contain information about maxtmum
sustained flows over varying time periods, and over many days (or runs). Better and more
complete information is obtained when MSFTG is developed with data from many days or
simulation runs. MSFTGs, a new concept developed in this research, can be valuable m

33
describing the concept of susta in ability in capacity. A maximum sustai ned flow ti me graph is
based on a typical min-max fun ction. T he graph is built by comput in g the maximum of the
minimum flow that was sustained for a period of tim e during the data co ll ect ion project. By
defi nition , MSFTG is a non-increas ing fu nct ion ove r time. A sampl e exa mple of max imum
sustained flow speed t im e gra ph is provided in figure 4.2. However, the maximum sustai ned
flows vary by day (or runs) and demand patterns. Hence it is difficult to use suc h a graph in
ca libration ofmicrosimulation model. However if good 0-0 data is avai lab le, this method can be
used for secondary calibration of the mi cros imul at ion model. Since, maximum flow sustained
time grap hs are depe ndent on both supply (capacity or speed-flow-den s ity) a nd demand (0-0).

4.2.5 I Speed, Flow, or Density Contours

Speed, flow , o r dens ity contou rs have been traditionally used in qualitative calibration of
microsimul ation mode l. These contour graphs have a wealth of information. Information that
can be obtained from contours includes extent (space and time) of congestion, location of
bottlenecks, speed, flow, or den sity inform ation, extent of queues, etc. This kind of information
can be valuable in calibration of microsimulation mode ls. Since these contour graphs at:e a result
of supply and demand, logica lly, there are to be calibrated after suppl y and demand parameters .
It is impotiant to only minima lly alter the global suppl y and demand parameters, but concentrate
on altering the local link spec ific parameters for calibration.

4.2.6 I Histograms: Speed, Headway, and Travel Time

Histograms also form an important part of the aggregate data based cal ibiation procedures in
microsimulation models. Histograms of speed, headway, and travel time have traditionally been
used for calibration of microsimulation models. However, travel time hi stograms can probably
be argued to be more important, because several of the meas ures of performance of traffic flow
systems are based on trave l time data.

4.2.71 Time Series: Volume, Speed, and Density

There have been many studies that adopted a time-series based calibration procedures. Most of
these methodologi es assumed that good 0-D data was available prior to s uch calibration
procedure. Such assumptions are usually not valid in the real-world. Some researchers, e.g.
Toledo et al. (2004), suggest using an iterative calibration methodology to simultaneously
estimate 0-D flows and calibration of driver behavior and route choice parameters, but such

34
 procedures can be exte nsive and time consuming. The maxim um s usta ined flow time graphs are
similar in nature to volume time-series in that both assume, in most cases, 0 -D flo ws as a given .

Wiedemann 99: Flow vs . Tim e Sustained Graph

3000

2900 ~ ~

2800 r
I
J
2700~
i 0
g
?,
j
~

~
i3 ;') ()
i'J ~
c~
2600
~
:,
:-j
~ ~-,
j
?-i ~ " )

"'
-@ 2500
~
)
'
t)
0 ~
~
~0 t:l
;;!
I
"""' ~ ~ ~
~ 0 l~
tJ ~

a
~
~~ 0 0
"
.~ 2400 r~ 8
. : ) ~ ~
~ 2300 (.

2200

2100
(
2000

1900

1800
15 30 45 60 75 90 105 120
Time Sustained in Minutes

Wiedemann 99Fiow Rate vs. Time Sustained Box Plot

2800rr- . - - . - . - - . - , - - , - , - - , - , - , - - , - , - - , - ,- , - - , -, - - , - , - - , - . ,
2750 '
2700 '
2650
2600
2550
2500
' T
- 2450 I I I T

~ 2400
Ru:
.. ~.r. 6. 8:. : : ,
~ 2350
:I "
. .~,. , G
' L
>7 .R'
u\J U
;; 2300
I
r
I
r
I
l
I
1
I
81 2250 .J_ . -4-- .J_ . I
J.
I
l
l
+
~ 2200
u::
2150
2100
2050
2000
1950
1900
1850
1800 ~--2L__L~--~-6L_~7L-~8--~9~1L0~11~1~2~~1L3~14~1~5~1L6~17L-1~8~1~9~2L
0~21~2~2~
Time Sustained (15 min to 2 hrs in steps of 5 min)

Figure 4.3 Maximum Flow Sustained Time Graph (MSFTG) generated using VISSIM

35
4.3 I DISSAGREGATE-DAT A BASED REPRESENTATIONS

4.3.1 1Vehicle Trajectories

Vehicle trajectory information can be valuable in a calibration process. Vehicle traj ecto ri es can
be used to eva luate the stability of car following models. Some researchers (e.g. Brockfield et al.
(2004)) have used vehicle traj ectories for ca libration by minimi zing the diffe rence in gaps (or
headways ) in field and simulated trajectories. Traditionally in such resea rch , the lead vehicle ' s
trajectory in the field is replicated in the simulation by external control. Thereafter, the car
following model parameters are calibrated to minimize the difference in gaps. The microscopic
vehicle trajectory data is obtained either from a freeway or a test track. Microscopic data can
provide valuable information for calibration of microsimulation models. However, such
calibration method suffers from serious flaws. First, the microscopic vehicle trajectory data does
not obtained from a range of traffic conditions. Second, test track data usually model s a limited
number of drivers, thereby only modeling a limited variety of driver behavior. Third, many of the
microsimulation models are developed to capture average conditions; modeling specific driver
behavior is beyond the scope of many microsimulation models. Fowih, many of the
microsimulation models are stochastic in nature. A same lead vehicle trajectory might produce a
different platoon behavior depending on the random seed. For these reasons, microsimul ation
calibration based on vehicle trajectories is not an attractive solution, and is not considered in this
research.

4.3.2 I Relative Distance vs. Relative Velocity

Relative distance vs. relative velocity graph (henceforth referred to as dx-dv graph) is also based .
on microscopic traffic data. The dx-dv graphs were among the popular calibration methodologies
used in microsimulation calibration. The dx-dv graphs describe a typical goal-seeking behavior
in many car following models. The following car always tries to maintain a safe distance from
the lead vehicle, while trying to reduce speed differences. The dx-dv graphs also show a typical
oscillation behavior in following condition. Dx-dv graphs form the core of many psycho-physical
models like Wiedemann 74 and Wiedemann 99, Michaels ( 1963), Todosiev( 1963), etc. In
psycho-physical models, the thresholds are mainly modeled in the dx-dv plane. Therefore, such
information can be valuable in calibrating driver behavior thresholds in psycho-physical car
following models like VISSIM. The calibration of microsimulation models using dx-dv graphs
suffers from some of the issues described in the previous subsection. However, aggregated
information from dx-dv graphs can be a valuable source of information in calibration. An

36
aggregated behavi or can provides ins ights into ave rage fo llo wing va riation, speed osc ill a tion ,
minimum followin g distances , etc. T hi s inform ation is espec ia ll y va luabl e in a psycho-phys ical
car following mode l li ke VISSIM .

4.4 I SUMMARY
In summary, there are various traffic flow representations poss ible for application in
microsimulation. The traffic flow represe ntations presented in this chapter are argued to be of
higher fidelity . The representations in the state-of-the-art are based on low fide lity
representation, which are s usceptibl e to significant error. Representation of simulation and field
based on a single numerical value results in a significant loss of information. The traffic flow
representations presented here contain more information that can help better calibrate the
simulation model s. Ofthe traffic fl ow representations presented in thi s section, speed-flow, flow-
occupancy, speed-flow-occupancy, and dx-dv graphs are more promi sing. Some of the
representations e.g. speed/flow/density contours, maximum flow sustained time graphs are also
promising, but need good OD estimates to be beneficial. However, there is a need to quantify
the differences in these representations. Traditional parametric or non-parametric statistical
methods are not applicable in some of these scenarios. Research was conducted to deve lop robust
measures of degree of closeness for application in ca libration. The research on developing such
measures is presented in the next chapter.

37
 V Applications of Pattern Recognition in
Calibration

5.1 I INTRODUCTION
Based on the discussions provided in the previous chapter, it can be seen that there are several
ways to calibrate a simulation model. New concepts based on relationships between flow, speed,
density/occupancy, and time were argued to be capable of being used in calibration. It is possible
to develop several objectives, each of which is tailored to the nature of the variable, for
calibration. But, it is the intent of this study to develop objectives that are generic in nature, so
they find application in several areas and require minimal adjustment. In addition, currently there
are few existing quantifiable methods available to calibrate simulation models based on new
concepts introduced in the previous chapter. After an extensive literature review on ri1ethods and
tools, it was identified that traditional parametric statistical methods have limited application in
calibrating simulation models based on relationships between flow, speed, density, and time. It
was also identified that non-parametric statistical methods can only be applied for frequency
based distributions; although, no priori distribution is assumed in such methods. These non-
parametric statistical methods can be applied to histograms (speed, headway, flow distributions),
but have limited or no application in 2D or 3D graphs, contours, or time-series. However, there
are certain statistical and structural pattern recognition methods that are quite well suited to solve
such problems.
Traditionally, calibration of simulation models using speed-flow graphs and contours has
been based on qualitative matching by human-eye matching. In such cases, the researcher usually
looks at graphs from the simulation and the field, and decides how closely the graphs match each
other. The human-eye matching technique, although very sophisticated, cannot discern the extent
of differences in close situation. In other words, unless if the differences between two

38
alternatives are s ignificant, the researc her has no way of sayi ng wh ic h o ne is close r to th e fie ld
graph . This direc tly impacts the ca libration process, since it is not automatabl e; i.e., the
researcher has to be in volved in eac h step of the cal ibrat ion making deci sions abo ut the close ness
of the graph, some of whi ch mi ght be wrong.
There are problems in other scientifi c fields where measures of close ness of objects, sets,
or data are needed. Pattern recogn iti on is a sc ience with a vas t amount of literatu re that deals
with measures of similarity (closeness) or di ss imilarity. Pattern recognition has found
applications in a variety of fi elds includin g database management, digital recognition, data
mining, etc. However, pattern recognition has found minimal application in the field of
microsimulation calibration. The methods and tools provided by pattern recognition can be
invaluable in thi s field of calibration. In the following section, a general introduction to pattern
recognition is presented, followed by definition of a generic ca libration objective. Thereafter,
comprehensive resea rch on applications of pattern recognition based di ss imil arity meas ures in
calibration is presented.

5.21 PATTERN RECOGNITION

Pattern recognition is as much a science as it is an art. The science and art of pattern recognition
is a source of solution to many a problems in the scientific field . It as has made poss ibl e so me of
the most sophisticated production systems to autonomous robots . Human behavior, some argue,
is based on recogni zing patterns. Humans store a priori information based on perceived patterns
of shapes, sounds, smells, behavior, characters, images, events, and data. The sc ience of
transforming such ability to recogni ze and classify objects to help in machine learning is called
pattern recognition. Examples of application of patter recognition in the sc ientific field include
automatic sorting of fruit (by quality) in super-fast production lines, modeling learning behavior
in autonomous robots, scene analysis, face recognition, biometric security systems, etc.
The concepts of pattern recognition can be borrowed into calibration, almost as if it was
the most natural way to solve the calibration problem. For example, one of the fundamental
questions in pattern recognition is how to tell the difference between objects. It is inherently an
easier task to tell if two objects are different than to say they are the same. Pekal ska and Duin
(2005) suggest that dissimilarity is more fundamental than similarity, because " .. . only when the
difference has been observed and characterized, similarity starts to play a role." Drawing a
correlation, the calibration objective, which is to reduce the difference . between two objects
(simulation and field), can be stated in pattern recognition terms as reducing the dissimilarity
between simulation and the field.

39
 A di ss imil arity representation provides a va lue to measure the degree of co mmona lity
between objects. For imp lementation of the calibration objective, the re is a need to deve lop
representation at two leve ls. T he first leve l of representatio n captures the obj ects; the seco nd
level captures the variables. For instance, in fig ure 5. 1, th e objects are rep rese nted at level - 1
using capacity, speed-flow, etc. T he leve l-2 represe ntat ion relates to representat ion of the
variables as points, point sets, areas , volumes, distr ibutions, etc. A detailed formulation of the
process is provided in the next secti o n by de finin g a ge neri c calibration obj ecti ve based on such a
concept.

~ation Model Field Scenario

t--------,----J
----~ -- -- --~----

Traffic Flow Variables an d

Relationships
Capacity Representation Level 1
Spee-d-Flow
Speed-Contour

Dissimilarity Measures
Distribu tion
Point Set
Representation level 2
AreaNolume/1 ntensity

Figure 5.1 Representations in calibration objective problem definition

5.3 j DEFINING A GENERIC CALIBRATION OBJECTIVE

A generic calibration objective seeks to minimize the difference between two objects; the
simulation model: S; and the real-world (field) scenario: F. It can be mathematically defined as
follows:

Objective: Z = Minimize (F difference S)

40
Traditional ly, measures of di fference in objects, sets, and data are referred to as dissimilarity
meas ures, whereas measures of c loseness are referred to as similarity measures .

The real world scenario can be considered as an obj ect : F(l ), where I, possibly infinite, is the
input (includes traffic control)

The simulation model can be defined as an object: S(P, Q, 0, A), where in the inputs are as
follows :
I. A finite vector of calibration parameters: P = [p 1 p2 .. Pm] , where for some elements of p
are linearly or non-linearly constrained.
2. A finite vector of simulation inputs: Q=[q 1 q2 qn] , where 'Q' is an estimate of 1,, which
is a subset of I.
3. A finite vector of options: 0 = [o 1 o2 . . . op]
4. A finite list of assumptions: A

Some of the impm1ant things to note in this representation are:

l. It is assumed that all ofthe input that goes into the real-world is expresses as a vector of
quantities: I. This vector is possibly of infinite length, since all of the quantities are not
measurable or expressible in mathematical form. In addition, a subset of 1: 1, is defined,
which includes some of the quantities that are measured in the field for the purpose of
feeding into the simulation model. Since the measurement techniques cannot be
guarantied to accurately estimate the quantity, due to measurement error, assumptions in
methodology, human error, etc, a new vector Q, which is an estimate of I, is defined.
2. A finite vector P of calibration parameters for the simulation model is defined.
3. A finite vector of options 0, which defines the options selected to build the model. For
example, type of distribution for vehicle generation - Poisson, Normal, etc - can be
considered as possible option.
4. Finally, a finite list of assumptions: A, is defined to account for assumptions made while
building a simulation model.

Redefining the objective,

Objective: Z =Minimize: Dissimilarity (F(I), S(P, Q, A, 0))

Or Maximize: Similarity (F(I), S(P, Q, A, 0))

41
This redefi ned objective function minimi zes the dissimilarity between field scenano and the
si mulation using a Dissimila rity ( ) function. For a set of measurabl e traffic variables: V =

{v 1,v2, ... ,v,), where ' v' can be a combination of two or more parameters , is defined. Thi s is the
first-level of representat ion in the ca libration objective definitions. The dissimilarity function ,
which is second-level of representation, operates on a representation of the variables . The
dissimilarity function is defined as follows :

Dissimilarity {(rep(fie/d(V)), rep(sim(V)))} I F(I), S(P, Q, 0, A)

Where,
sim( ), field() are function that return simulated or field values of arguments
rep( ) is a representation of the variables in dissimilarity space. For instance, the speed-flow
graph can be represented as an image or a binary string, instead of point sets.

In summary, an objective function in microsimulation calibration can be stated in general as a

function which measures the dissimilarity between the simulated and the field conditions. An
objective function is one of the most important patis of the calibration process . First, the better
the traffic variables represent the simulation and the field, the better is the calibration process .
Second, the better the representation of variables, the better is the calibration process. In
combination, they form an objective function , and are valuable to the calibration process. The
objective function can range from simple function based on a single parameter to complex
function based on multiple parameters. In addition, much of the research presented here is based
on dissimilarity measures, but the methods can be restated as similarity measures in various ways .
The arguments presented in this section are two fold. In a two level representation of the
calibration objective, each of the representations results in a certain loss of information. But such
representations are ultimately required to solve this complex problem. In chapter 4, various
representations of the simulation and the field were presented. It was argued , for some
representations, that there was less loss of information than the current state-of-the-mi
representations, thus resulting in a better calibration. Traditionally, many of the representations
provided in chapter 4 were used in qualitative calibration, due to the lack of quantifiable methods .
Time-series and histograms are the only notable exceptions in representations where researchers
have previously developed quantifiable methods. Differences in histograms can be quantifiable
in parametric as well as non-parametric statistical methods like kolmogrov-Smirnov tests, chi-

42
square tests, t-test, f-test, etc. But, s uch tests are not applicable to other representations presented
in chapter 4. In the next section, a detailed discussion has been provided for dissimilarity
representation, which is the second leve l of represe ntation.

5.4 I LITERATURE REVIEW ON DISSIMILARITY MEASURES

Dissimilarity measures are relative measures of dissimilarity; the small er the diss imil arity
measure, the more similar are the objects. The di ss imilarity meas ures are not always defined for
objects, rather for features of the object. This is same as representation leve l I discussed in
previous section. These features of objects are also known as variables or attributes.
Dissimilarity measures can be defined for a variety of feature types. Examples of features include
points, set of points, images, symbols, text, shapes, etc. There are different dissimilarity
measures available depending on the feature type. Pekaiska and Duin (2005) define five different
types of features.
1. Binary
2. Categorical
3. Quantitative
4. Ordinal
5. Symbolic ot nominal

There are a variety of measures available for each of the feature types . But, much of the
calibration objectives in microsimulation are related to either quantitative or binary data. A brief
survey of dissimilarity measures for different types of data is provided in the following section.
Much ofliterature pt:esented in this section is based on Pekalska and Duin (2005), Duda and Hart
(1973), and Ftikunaga (2003).

5.4.1 I Dissimilarity Measures for Quantitative Data

There are many dissimilarity measures available for quantitative data. Most of the common
statistical error measures presented in chapter 3 are also dissimilarity measures. Some of the
common statistical error measures include the following:
1. Correlation coefficient
2 . Mean squared error or Root mean squared error
3. Percent mean squared error
4. Mean absolute error
5. Root relative squared error

43
 6. Relative absolute error

Mean squared error (MSE), root mean squared error (RMSE), and percent mean squared e rror are
some of the most commonly used error measurement techniques in the fie ld of traffic
microsimulation . And if more than one variabl e is involved, the errors are accrued over several
variables. But, there are some other dissimilarity measures available for quantitative data . Some
of the popular measures are presented below.
l. Euclidean distance or weighted Euclidean distance
2. Taxicab metric, city block Distance, or Manhattan distance
3. Max norm or Chebychevdistance
4. Lr-distance or Minkowski distance
5. Canberra Distance
6. Correlation-based

Mink.owsky distance is a generic form of Manhattan, Euclidean, and Chebychev distance. All of
the distance measures are for point-to-point distances. For a more comprehensive list of
dissimilarity measures for quantitative data, the reader is directed to Pekalska and Duin (2005)

5.4.2 I Dissimilarity Measures for Binary, Boolean or Dichotomous Data

Binary, Boolean, or dichotomous data represents variables that only accept two distinct values.
Many of the dissimilarity measures on dichotomous data are based on four different counters:
1. a= the number of properties common to both objects (i, j)
2. b =the number of properties which i has but j lacks
3. c = the number of properties which j has but i lacks
4. d =the number of properties that both objects lack

Most of the dissimilarity measures for dichotomous data are based on these four basic counters.
Some of the dissimilarity measures include the following:
1. Russel and Rao: a/(a+b+c+d)
2. Simple matching: (a+d)/(a+b+c+d)
05
3. Binary Euclidean: (b+c)
4. Hamming distance: b+c
5. Variance: (b+c)/{4(a+b+c+d)}
6. Binary Pattern Difference: bc/(a+b+c+d) 2

44
5.4.3 I Dissimila rity Meas ures fo r Populations
There are a number of parametric and non-parametric methods that are ava ilabl e for measuring
diss imilarity betwee n populations. If mean vectors are used to represent popu lations, then
dissimilarity meas ures for quantitative data can be used. If population is represented as a
probability density function (pdf), there are measures available to compute di ssimilarity of two
distributions. Some of the more common dissimilarity measures are
1. Ko lmogrov metric
2. Kolmogrov- Smirnov test
3. Mahalanobis distance
4. Chi-square test
5. t-test

5.4.4 I Dissimilarity Measures for Sequences

Dissimilarity measures for sequences such as the ones based on binary data or finite discrete
elements are used in pattern recognition and machine learning. There are many problems that can
be formulated in terms of sequences, and hence the dissimilarity measures based on such
representation are invaluable. Some of the common dissimilarity measures are listed below.
I. Hamming distance measure is based on counting the number of positions in which the
sequences differ. It is one of the most simple measures for measuring dissimilarity of
sequences
2. Fuzzy Hamming Distance measures the cost editing one distance usmg insettion,
deletion, and shift. The operations are used to transform one sequence into another, and
costs accrued over the operation are used as a dissimilarity measure.
3. Levenshtein Distance measure is one of the most popular edit distance measures. It is
based on costs accrued over operations including insertion, deletion, and substitution. If
the sequences are not of equal length, then a normalized version of the distance can be
used.

5.4.5 I Dissimilarity Measures for Sets

Dissimilarity measures are possible between two sets of points in space. These measures are
valuable in several applications where features of objects can be represented as points in space.
This measure is different from dissimilarity measure of quantitative data. For quantitative data,
each ofthe points is a measure of a feature belonging to the object; whereas for sets, sets of points
are defined as a feature. Hausdorff distance and its several variations are defined for measuring

45
dissimilarity between point sets. T he differe nt dissimilarity measures for po int sets are provided
below.
1. Hausdorff d istance is defined as the maximum of directed Hausdorff distances between
two sets A and B. The directed Hausdorff distance , d1-1 A->B = max.{m inb d(a ,b) }, is the
maximum of collection of minimum di stances from each point a in A to b in B.
2. Variants of Hausdorff di stance are generalizations of the Hausdorff di stances that are
more robust against outliners and noi se. (Pekalska and Duin, 2005)
3. Modifi ed Hausdorff di stance is minor variation of Hausdorff distance wherein a average
of minimum is applied instead of maximum of minimum.

5.4.6 I Dissimilarity Meas ures fo r Images: Template Matching

Template matching is a method usually used in the field of scene analysis in pattern recognition,
where a template pattern or image is matched to an image closest to itself from a reference set of
patterns or images. Template matching utilizes a dissimilarity measure. An image is usually
represented as pixel data and the differences in pixel values are measured and accrued over all
pixels. The distance between pixel values can be computed by using any of the dissimilarity
measures for quantitative data.

5.5 1APPLICATIONS OF DISSIMILARITY MEASURES IN CALIBRATION

Based on the literature review of dissimilarity measures presented in section 5.4, research was
conducted on developing representations for various traffic variables or features of simulation and
field. As discussed earlier, features of simulation or field is the first level of representation in a
calibration objective. In this section, the second level of representation, dissimilarity
representation of the features (i.e. , traffic variables), is discussed in detail. Various dissimilarity
measures are presented for each of the representations presented in chapter 4.

5.5.1 I Capacity
Capacity based calibration methods are one of the most commonly used methods in
microsimulation. These methods are based on minimizing the difference between measured and
simulated values of capacity. There are many methods available to measure the difference, and
the choice depends on the type of application. As per the discussions in chapter 4 section 4.2,
capacity can be either represented as quantitative data or as a distribution of values. There has
been recent research of defining capacity as a random variable. Brilon et al. (2007) argue that
capacity is a random variable, and that it can be represented as a Weibull distribution. But, such

46
research is in its initi al stages of development, and has not found s ignificant practical
implementation. Much of the practitioners and researc hers st ill define capacity as a quantitative
data. In thi s researc h, di ss imil arity represe ntations have been provided fo r both capacity as a
quantitative data, and capacity as a popul ation or distribution.
If capacity is defined as quantitat ive data, any of dissimilarity meas ures for quantitative data
can be used for representation. Some of the dissimilarity measures that can be a pplied include
many of the statistical meas ures of error like root mean square e rror, absolute error, root relative
error, etc. But with-respect-to di ss imilarity measured based on distance metri cs the following can
be used:
1. Euclidean distance
2. City block distance
3. Correlation-based distance
4. Any ofthe LP distances

Implementations of c&pacity as a quantitative data can be found in much of the calibration

research in the state-of-the-art. After an extensive lite!'ature review, applications of calibration
based on distribution or population representation of capacity were not found. In chapter 4 .
section 4.2.1, the use of distribution or population representation of capacity was argued to
perform better than a quantitative representation of capacity. If capacity is represented as a
distribution or population, any of the dissimilarity measures for populations or distributions can
be utilized. Some of the dissimilarity measures are listed below.
1. Kolmogrov Metric
2. Kolmogrov-Smirnov test
3. Chi-square test
4. t-test

But ifthe mean ofthe distribution or population of capacity is used, then dissimilarity measures
for quantitative methods can be used. If capacity is assumed to belong to a parameterized
distribution (e.g. normal distribution), parametric statistical methods like t-test are available for
testing the goodness-of-fit. But if no assumption is made about the distribution of capacity, many
of the distribution-free statistical methods can be used for measuring the goodness-of-fit. In
summary, the dissimilarity representation of capacity can be performed in multiple ways, but the
decision about type of representation for implementation should be based on research problem

47
 and goals. However, it is hypot hes ized that representat ion of capac ity as a dis tri bution or
population wo ul d res ult in a bette r ca librated s imul at io n mode l.

5.5.2 I Speed-Flow or Other Similar Graphs

The concept of speed-fl ow, fl ow-density, or speed-density graphs as leve l-one re prese ntations of
the simul ation mode l and the fi e ld scenari o is an ori gina l contributi on of thi s research . The
arguments for use of speed-flow, flovv-dens ity, or speed-density gra phs in calibration of
simulation mode l have been presented in section 4.2.2 and 4.2 .3. The di ss imil arity representation
of speed-flow graphs or other s imil ar graphs can be done in many ways. Muc h of the di scussions
provided here are described using speed-flow graphs, but all of the representations presented in
thi s section are equally applicabl e other graphs of s imil ar nature.

5.5.2. 11 Speed-flow graphs as point sets

If speed-flow graphs are represented as point sets, the di ss imilarity measures for point sets can be
. utilized for measuring the degree of closeness between speed-flow graphs from the simulation
and the field. In this case, speed-flow measurements are represented as points in a two-
dimensional space. There three different diss imil arity measures defined as follows.
1. Hausdorff di stance
2. Modifi ed Hausdorff distance
3. Variants of Hausdorff distance

The two most popular measures, Hausdorff di stance and modified Hausdorff distance, are
presented in this section. But, there are several other Ha usdorffvariations avail able. For pairs of
flow and speed measurements, the followin g terms are defined .

A is a set of n A pairs from the simulation, and a E A

B is a set of n 8 pairs form the field, and b E B

d(a,b) is the distance measurement between a and b

Hausdorff Distance, d H ( A, B), can be defined as follows:

48
dH(A , B) = max{d -"H(A,B ), d _, H(B,A)} , where d _, H(A,B )i s the directed Hausdorffdistance

definedas, cr'H(A, B) =maxmind(a,b)

ueA heH

Modi fied HausdotffDistance, d ,1m (A,B), ca n be defin ed as follo ws :

dMH(A , B)=max{d--.avr(A,B),d _,mr(B , A)}, where d --.,,!H( A, B)i s the directed Hausdorff

1
distance defined as, d ->,ll'r ( A,B) = - - L mind(a ,b) .
n A aeA beB

The distance d(a,b)can be measured in different ways. It can be a Euclidean distance, weighted

Euclidean distance, or any of the dissimilarity measures for quantitative data. But in this case,
weighted measure is essential due to difference in units between speed and flow.

5.5.2.2 I Speed-flow Graphs as Symmetric Difference

If speed-flow graphs are represented as binary images, gray-level, or intensity based images,
dissimilarity measures developed using syinmetric difference or template matching. Symmetric
difference of sets, A and B , is defined as (A- B)U (B - A). But, there is a cettain loss of

information in the discretization process, due to conversion of vector information to raster

information. The continuous speed-flow measurements in a two-dimensional space are
transformed into discrete cells, also known as pixels in images . The discretization can be at
binary level, where each pixel either represents existence of a data point (I) or non-existence of a
data point (0). The discretization can also be at gray-level or an intensity based measure, where
the value of pixel is proportional or equal to the number of data points within the cell or pixel.
The image can also be normalized, smoothened, or passed through various transformations in
order to make the data more representative of the characteristics of the graph. The discretization
process, although resulting in some loss of information, possesses some advantages. The
discretization smoothens out the data set reducing the "rioise" in the data set. Since simulation
models are stochastic in nature, there is a lot of noise inherent in the measurement.
If speed-flow graphs are defined as an area, then the measure of symmetric difference can
be used as a dissimilarity measure. For practical reasons, it is easier to represent both the field
and simulation graphs at a similar resolution. The mathematical dissimilarity for two graphs is
then formulated as follows:

49
D(.) is the di sc retization function that convet1s vector information to raste r (or pixel) information.

D( A) = A(i, j) is the speed-flow graph from the simulation obtained by transformation

D(B) = B(i, j) is the speed-flow graph from the field

The dissimilarity measured.r{A(i , j) , B(i,J)) is defined as follows (Duda and Hart, 1973):

dr{A(i,j), B(i, J)) = 2>L d(a( i, j) ,b(i, J)))

i J

Again, various dissimilarity measures can be used ford(a,b) . But if Euclidean distance is used ,

the formulation can then be rewritten as.

1/2

~ ~ [a(i, j) -
2
dTE(A(i, j), B(i, J)) = b(i, J)]
{ )

5.5.2.3 I Speed-flow graphs as sequences

This representation is similar to bit1ary, gray-level, or intensity images; except that such images
are transformed and expressed as a sequence (or vector), rather than a matrix. For instance, let 's
assume a 4-by-4mage of 16 pixels; this can be transformed to a 16 element row or column vector
by concatenating the rows or columns. (Figure 5.2)

0 1 1 0
1 0 0 1
0 1 1 0
0 1 1 0

1
IO 1111 Io11 IoIo11 Io11 11 IoIo11 11 IoI
Figure 5.2 Row concatenation representation of a binary image matrix

This process can transform matrix image data to sequences, and dissimilarity measures based on
sequences can be used in measuring the degree of closeness. Some of the measures that can be

50
used are Hammi ng distance, fuzzy ham ming distance, and Levenshtein distance. However, this
method suffers from a serious fl aw; i. e., neighborin g cells in the sequences need not be neighbors
in the image. For examp le, an 8-by-2 matri x, formed by manipulating the rows in the matri x as
[row! row 2; row3 row4], will be reported as similar. Thi s is obviously wrong!
A way to partially cou nteract for such a prob lem is to use add di ss imilarities from vectors
formed by row concatenation and column concatenation . Referring to the previous example, the
8-by-2 matri x would res ult in hi gh di ss imil arity in compari son of co lumn vectors. In addition,
this method can be extended using a diagonal vector.
A probable method that can partially so lve nei ghborhood problem employs alternate
reverse row or column concatenation. For example, referring to the previous example, the
sequence is formed using [row! reverse (row2); row3 reverse (row4)]. This method ensures that
neighbor cell is spatially eq uidistant up to one neighbor. But, such efforts do not lead significant
savings, and hence is not considered for use in dissimilarity measurement in this research.

5.5.3 I Speed-Flow Graphs as Partial Point Sets

This is a discussion . that is the most relevant to this research, because usually the point sets
(speed-flow graphs) are not complete. The field speed-flow graphs obtained are usually not
complete; for that matter, the speed-flow graphs obtained from simulation are also not complete.
Defining dissimilarity meas ures for partial point sets has not been found in the literature . Since in
calibration, the simulation is the one that is being tweaked. It is logically better to consider
differences with-respect-to the field data set. A logical explanation for this is that, to be the best
of our knowledge, the information coming from the field is reality; it is the simulation model that
needs to be tweaked. A symmetric difference method presented in section 5.5 .2.2 is not valid
here, because in such a method, both point sets are considered equally important. Which is not
clearly the case with speed-flow graphs, or methods based on calibration.

The dissimilarity measured for partial point sets is defined as follows:

d(A,B) =B-An B

This dissimilarity measure is a better estimate than d(A, B)= (A- B)U (B - A)= -An B

5.5.4 I Speed-Flow-Density (SFD) or Higher Dimension Graphs

51
Speed- flow-density graphs are representat ions of simul ation and the field. They are hypoth es ized
to contain more inform ation , hence resulting in better ca librated s imulation models. T he
dissimilarity measures delineated for speed-flo w graphs are equa ll y applicable, or otherwise, ca n
be exte nded to capture 30 or higher dimen sional variables.

5.5.4.1 I SFD graphs as point sets

The concepts presented in section 5.5.2.1, di ss imilarity measures for speed-fiow graphs as point
sets, can be extended to cover three dimensional graphs or point sets. The Hausdorff distance
defined earlier is still valid in higher dimension.

5.5.4.2 I SFD graphs as Symmetric Difference

The concepts presented in section 5.5.2.2, dissimilarity measures for speed-flow graphs as
symmetric difference, can also be extended to cover three dimensional graphs or point sets. The
dissimilarity measure for higher dimension can be rewritten as follows:

dr(A(i,j, ... k), B(i,j, ... k)) = L _L ..._Ld(a(i, j ,...k),b(i,j, .. .k))

j k

The problem with such a definition is that as the number of dimensions increase, the amount of
computations require significantly increase, resulting in higher computational cost.

5.5.5 I Contours
Dissimilarity measures for speed/flow/density contours can be defined sim ilarly to measures
presented in section 5.5.2.2 (dissimilarity measures for speed-fiow graphs as symmetric
difference). The contours can be represented over n-levels and the measures previously proposed
can be utilized. In terms of implementation, the levels in the contour are represented in
incremental values.

5.5.6 I Speed and Headway Distributions

As described earlier, speed and headway distributions are one of the most commonly used
calibration methods. The dissimilarity measures presented in section 5.4.3 for populations or
distributions are applicable in this case.

52
 VI Optimization Representation in
Calibration

6.1 I INTRODUCTION
Due to the stochastic nature and complexity of microsimulation model, representation of
microsimulation as closed-form equation is not usually possible. As a result, traditional calculus
based optimization methods cannot be applied (Dowling et al., 2004). The calibration of
simulation models requires the use of search and optimization methods based on multiple
evaluations of the objective function. In chapter 5, various dissimilarity measures were proposed
for different traffic flow representations. The objective all through the development of calibration
objective was to produce measures that are based on evaluations of the simulation model. There
are many derivative-free optimization methods that can be applied to microsimulation calibration.
Most of these methods direct the search path based on multiple evaluations of the objective
function.
For an optimization method that is based on multiple evaluations of the objective
function, the performance of the method is dependent on the quality of the objective function. If
the objective function can give more information about the differences, the optimization method
can use such information to better direct the search path. Since the amount of information
available in an objective function is of utmost impmtance, speed-flow graphs could perform
better since they contain more information. For example, if minimization of the difference of
numerical capacity values is the objective, there always exists a possibility of optimization
method recommending non-optimal parameters, because the degree of freedom is too high. A
possible solution is to expose only the calibration parameters that have a significant effect on the
objective function. However, by using speed-flow graphs, a higher number of parameters can be
exposed to the calibration process, resulting in a better fine-tuned simulation model.

53
A literature rev iew was con ducted to esta bli sh various opt imi zati on methods for use in
microsimulation ca libration . It was determined that there a variety of optimi zation methods
available for use in ca librat ion . The choice of the method usua ll y depend s on a vari ety of factor
including number of variabl es, perform ance, run times , availability, etc . But much of the
methods can be c lass ifi ed as Direct Search Methods (DSM). There are some derivative-free
optimization methods are also applicable in ca libration . Over th e yea rs, the terms direct search
methods and derivative-free optimization methods have been used sy nonymous ly, but there are
few researchers claim differences between these terms (Lewis et al. , 2000).
Direct search methods are usually referred to as heuristic methods lacking sound
mathematical basis. But, direct search methods are often used for s implicity, flexibility, and
reliability (Lewis et al., 2000). Many of the direct search methods are also robust in nature.
Some of the direct search methods that can be used in calibration are:
1. Hooke and Jeeve method
2. Nelder-Mead simplex method
3. Pattern Search Methods
4. Box's complex algorithm

The other category of optimization methods are based on natural or evolutionary concepts.
Examples of such methods include genetic algorithms, evolutionary algorithms, evolutionary
strategies, etc. For the past few years, Evolutionary Algorithms (EA) has been successfully
applied in traffic microsimulation calibration (M a and Abdhulai , 2002; Kim and Rill et, 2001;
Schultz and Rillet, 2004)). Evolutionary algorithms have been chosen for implementation of
calibration objective in this research. It is not the intent of this study to comment or test the
performance differences in different optimization methods. In the following sections, description
and implementation framework of evolutionary algorithms is presented.

6.2 I EVOLUTIONARY ALGORITHMS

Evolutionary algorithm (EA) is a search method wherein some of the mechanisms are based on
concepts of natural selection and evolution. These algorithms work on a population of solutions,
rather than on a single point. EA is one of the many algorithms based on principles of natural
selection and can be considered as an extension to a simple or a canonical genetic algorithm over
types of representation, crossover, mutation, and selection. Genetic algorithm (GA) was
introduced by Holland ( 1975), and later developed by DeJong ( 1975). The algorithms developed

54
by Holland ( 1975) are usually referred to as canonical geneti c algorithm. Genetic algorithms are
different from traditional optimization methods in that they work on a representation (or cod in g)
and operate on probabilistic trans ition rules (Goldberg, 1989) .

Encoding (or representation): defines the way the parameters are represented . Some of the
popular representations include:
I. Binary strings
2. Gray-coded string
3. Real-value

In canonical genetic algorithms, the parameters are encoded as strings, and various operators are
defined on strings. A simple genetic algorithm consists of three operators: Reproduction,
Crossover, and Mutation.

Reproduction (or selection): Is the process of copying strings, based on fitness values (measure
of performance). The process of selection involves two steps. In the first step, the objective
evaluations of individual members are transformed using a fitness assignment. There are various
ways of assigning fitness values. Some of the methods include:
I. Proportional
2. Linear
3. Rank-based

In the second step, the actual selection process is implemented based on fitness assignment. Some
of the selection methods include:
I . Roulette-wheel
2. Stochastic uniform
3. Tournament
4. Remainder

Crossover (or recombination): Is the process of selecting a pair of string and swapping string
bits to form two new strings. The process of recombination depends on representation or
encoding. Some of the methods in recombination include:
I. Single point
2. Two point

55
 3. Scattered
4. Arithmetic

Mutation : Is random alteration of bits in a strin g, often with a very small probability. The type
of mutation ap pli ed usuall y depends on encod ing. Some of mutation methods include
1. Gaussian
2. Uniform
3. Adaptive feas ible

A prototypical flow chart of the evo lutionary algorithm search procedure is shown in figure 6.1

Initial Population

Objective
Function .......

Best Solution
~
True
Check
Optimization
Criteria
I
False

Reproduction

Crossover

Mutation

Figure 6.1 A prototypical evolutionary algorithm implementation

56
6.3 1 EA TOOL IN MATLAB
An EA tool in MATLAB, a scie ntific programming too l, was se lected fo r use in the project. The
EA tool in MATLAB a llows for impl ementation of various operations in vo lved in evo lution ary
algorithms. Some of the features of EA Tool in MA TLAB are as follows :
1. Fitness sea ling
2. Selection
3. Reproduction (including elite individual , crossover fract ion )
4. Mutation
5. Crossover
6. Migration
7. Algorithm settings (penalty settings)
8. Hybrid function (e.g. pattern search)
9. Stopping criteria (tolerance, maximum number of generations, etc)
10. Real-valued , binary, or custom encoding

The EA tool also poses built-in plotting functionality. The tool allows for both unconstrained and
constrained minimi zation, with linear or non-linear constraints, and bounds.

6.4 I EA TOOL IMPLEMENTATION FOR CALIBRATION

The pseudo code of the evolutionary algorithm implementation used in this research is presented
below.

Pseudo-Code for the Evolutionary Algorithm based calibration

BEGIN
INITIALIZE population with random candidate solutions
RUN multiple microsimulation runs for each candidate
EY ALUATE fitness using objective [pattern recognition for speed-flow graphs]
REPEAT UNTIL (TERMINATION CONDITION is satisfied) DO
SELECT parents
RECOMBINE pairs of parents
MUTATE the resulting offspring
CHECK if candidates are already evaluated
RUN and EY ALUATE candidates not previously evaluated
STORE fitness values
SA YE elite individual
SELECT elite individual and remaining individuals for the next generation
END

END

57
An individual candidate is defined as a set of ca lib rat ion parameter va lu es . Parents are a pair of
individual candidate solutions that are transformed to produce a new possi bl e candidate solution
(offspring). Fitness values of candidates are computed by running micros imulation model over
many random seeds. The candidates for the next generation are se lected based on their fitness
values . However, the elite individual (best individual candidate seen yet) is sent to the next
generation without mutation. In this research, real -number representation was used . A roulette-
wheel method using proportional scaling of fitness values (for recombination) was employed for
selection. Gaussian methodology was used for mutation. It is possible to choose more than one
elite individual to send to the next generation, but increasing the number of elite individuals
usually results in a premature convergence due to decreased diversity .
The EA based optimization can be thought of as the third level of representation in the
calibration process. In this research, there are three defined level s of representation in
microsimulation calibration: traffic flow, dissimilarity, and optimization. Figure 6.2 describes the
three levels of representation.

I s;ffiulalion Model Field Scenario

I . ~:=. '-----r~------

r-.-- r,affi:;;lo~~a~able:~"d---
J Relationships
I Capacity Representation level 1
l
L_
Speed-Flow

Sp"~:F"~' ______,_ ___

Dissimilarity Me asures
Distribution
Point Set
Representation level 2
AreaNolume/lntensil)'

~-~~-- . -Op~1izati~n Methods

Evolutionar; ~.lgorilhrns
Simulated Annaallng
Pattern Search Representation level 3
Random Search

Figure 6.2 Three levels of representation in microsimulation calibration

58
 VII I VISSIM calibration Methodology

7.11 GENERALIZED CALIBRATION METHODOLOGY

The methods presented in previous chapters 4 and 5 were integrated to a cettain extent, wherein
the traffic flow representations in chapter 4 are represented in dissimilarity space in chapter 5. In
chapter 6, an evolutionary algorithm was introduced as a search-based optimization tool for
microsimulation calibration. There are many other search-based optimization tools available for
use in microsimulation calibration, and the dissimilarity representation can be applied in most of
the methods. However, evolutionary algorithm is an attractive tool that is flexible and powerful
enough to be applied in a whole range of problems. Evolutionary algorithms, unlike some
traditional search-base a lgorithms, are capable of being used in nonlinearly constrained multi ~
objective optimization. In summary, chapters 4, 5, and 6 can be thought of as three levels of
representation of calibration (figure 7.1 ).
In this chapter, the higher level calibration methodology is provided. In the higher-level
calibration methodology (figure 7.2), there are two levels of calibration . After sensitivity analysis
and range definition, disaggregate data based calibration is performed, and is followed by
aggregate data based calibration. NGSIM freeway data sets are used in di saggregate data based
calibration. For aggregate data based calibration, speed, flow and occupancy information from
data sources such as California PeMS is used.

59
 I Initial Candidate. !
i Solution s i
L-------r~------__J

I1
'
Potential Candidate
Solutions
' - - - - -r-- - -
J'-----
'

Multiple Runs of -l
Microsimulatio~ Mod~
_=r__
C Traffic Flow
Representation

No
Convergence
~
~,/
. /

Figure 7.1 Calibration based on three levels of representation.

60
 Initial Calibration
Parameter Set

Test Networks Sensmvity ,{nalysis and] - - -- - --

F reeway Soqmem wl R<lmps 1--;l
Circular Road Range Definition

- 1
Selected Calibration
Para meter Set

1 - -----
Disaggregate Data NGSIM Freeway
":---
Based Calibration Data Set

1
Aggregate Data California PeMS
iE--
Based Calibration Data Set

t
Final Calibration
Parameter Values

Figure 7.2 Proposed VISSIM calibration methodology.

7.2 I SENSITIVITY ANALYSIS AND RANGE DEFINTION

This is one of the first steps in a calibration process. After a preliminary list of calibration
parameters is selected for the simulation model, the sensitivity of each of the parameter to the
objective at hand is analyzed. In this process, sensitivity analysis is conducted at two levels :
aggregate and disaggregate. Two test networks were created to assist in the sensitivity analysis
process. For.disaggregate data, a two mile one-lane circular track was created. A circular track
was created to ensure constant global density during the data collection process, since many of
the traffic flow processes are dependent on density. For aggregate data, a 6-mile two-lane
freeway network with two on-ramps, one at 2-mile and other at 4-mile marker, was created. This
network was created after significant testing. Initially, a test network without ramps was created,
but such a network was not capable of producing a wide variety of traffic conditions. Testing
followed first with one ramp and then with two ramps, and network with two ramps created the
variety needed for sensitivity analysis. For assisting in the evaluation process, data analysis
module was created using MATLAB . In addition, a VC++ COM module was created to
automate the communication with VISSIM.
The sensitivity analysis was performed on disaggregate data (headways, relative distance
vs. relative velocity) and aggregate data (capacity, speed-flow, flow-occupancy, etc). In addition,
sensitivity analysis included other traffic flow processes like shockwaves . The final result of

61
sensitivity analysis is categorization of se lected parameters to accepted, declined , and external
control parameters. T he declined parameters are removed from the ca libration process. The
accepted parameters and external control parameters continue on to the next step in the
calibration process. The accepted and external control parameters are processed further to obtain
range information for each of the parameters. For range definition , the test networks similar to
the ones used in sensitivity analysis model can be applied ; or otherwise, test networks can be
created to analyze specific parameters. Similar to the sensitivity analysis model , disaggregate and
aggregate data is utili zed . No specific statistical model is utilized here, and much ofthe decision
making is left to the researcher, who utilizes a variety of data obtained from the simulation model.
In addition to the tests performed on simulation models, real -world aggregate and disaggregate
data described in section 3.4 of chapter 3 is utilized. In addition, a thorough theoretical analysis
of the simulation model is performed. Mathematical and physical deductions based on the .
theoretical framework are used in range definition process. This process uses both real-world and
simulated data to develop range definitions for calibration parameters. The theoretical analysis is
Wiedemann model parameters are presented in the next section, whereas empirical analysis of
parameters using real-world data is presented in chapter 9.

7.31 THEORETICAL ANALYSIS OF WIEDEMANN MODEL PARAMETERS

The sensitivity analysis and range definition of VISSIM car following models included a
theoretical analysis of various calibration parameters. Each of the parameters was analyzed to
define the sensitivity to capacity and other important traffic variables. The sensitivity analysis
was conducted primarily on Wiedemann 99 parameters, but such analysis is also applicable to
Wiedemann 74 parameters due to high correlation in threshold definitions. In addition, only three
user adjustable parameters are available in Wiedemann 74 car following model, whereas
Wiedemann 99 model offered ten user adjustable parameters. In this section, results of analyses
of each of these parameters are presented.

7:3.11 Capacity
Capacity is one of the important traffic variables in operational analysis. The sensitivity of each
of the parameter to capacity is useful in selecting parameters that can be exposed in calibration,
since many of the calibration methodologies recommend capacity calibration. As discussed in
chapter 4, there are a variety of definitions that exist for capacity. Capacity can be either defined
as queue discharge flow (QDF) or pre-queue flow (PQF). Queue discharge flows are lower than
pre-queue flows, but are sustained for longer time periods . Therefore, the sensitivity tests are

62
dependent on the defin it ion that is adopted fo r capac ity. Fo r examp le : if QDF is adopted, th e
sensitivity of para meter to s usta ined fl ow is important; if PQF is adopted, the sensiti vity of
parameter to hi gh flow is importa nt.
The flow rates can be described as rec iprocal of average ti me headways. In W iedemann
99 car following model there are three parameters that describe a verage time headways . T he
three parameters are CCO, CC I, and CC2. Similarly, in Wiedemann 74 model , there are three
parameters that describe average time headways; the parameters are ax_add , bx_add, and
bx_m ult. There are parameters in W iede mann 74 mode l that affect headways, but are hard-coded
into the mode l and are not user a lterabl e. The following process of the vehicle is defined in terms
of minimum followin g distance and longitudina l variation during following process.

In Wiedemann 99, safe minimum following distance is defined as follows:

AX99 =L 11
_
1 + CCO

ABX99 = AX99 + CCl * vslower

The maximum following distance is defined using following variation (CC2) as follows :

SDX 99 = ABX99 + CC2

Similarly, in Wiedemann 74 the followin g terms are defined.

AX 14 = L + ax _add+ ax _ mull RAND()

_
11 1

BX 74 = (bx _add+ bx _mull N(0.5,0 .15)) ~V''""'"'

ABX74 = AX 74 + BX74
SDX 74 = AX 74 + EX 74 BX 74

Of the three car-following parameters that affect capacity, CCO is the weakest parameter, because
the contribution to headway is the least compared to CC 1 and CC2. On the contrary, CC 1 is the
strongest parameter influencing capacity. All of the arguments concerning CC 1 are equally
applicable to bx_add and bx_mult in Wiedemann 74 model. In addition, ax_add is similar to
CCO; therefore, arguments concerning CCO are also applicable to ax_add. Since the following
distances are the additions of terms defined by these parameters, there are multiple solutions for

63
the same following distance (or headway). In other words , sim il ar following distance can be
obtained by decreas ing CC2 and slightly increasing CC!. Th is ca uses prob lems in ca librat ion
because it is difficult to ascertain the correct so lution . In order to decrease the combination
effect, CCO can be kept constant through the calibration.
In add iti on to CCO, CC l , and CC2, there are two other parameters that affect capacity.
CC4 and CCS also have an affect on vehicle coupling in following . These thresholds , as
explained in chapter 2, control the speed difference in a following process. CC4 defines speed
thresholds for a closing process; CCS defines speed threshold for an opening process. In other
words, CC4 and CCS also contribute to vehicle coupling in following. The extent of contribution
to capacity of these parameters is certainly lower than CC I, but comparison to CC2 is a difficult
one. However, one important conclusion can be drawn. Although, CC4 and CCS affect
following, the longitudinal variation of following distance is controlled by CC2. Therefore, the
affect ofCC4 and CCS on capacity can be argued to be less significant than CC2. However, CC4
and CCS have a critical effect on stability of the car following process. For this reason, CC4 and
CCS are critical when analyzing the stability and sustainability of flows . CC3 does not have an
effect on capacity, because the vehicles are already in following condition during at-capacity
conditions.
Two other parameters affecting the oscillation process CC6 and CC7 do not have a
significant affect on capacity. CC6 controls the effect of distance on speed oscillation, and CC7
controls the acceleration during following. Both of these parameters do not have a s ignificant
effect on headway, which is in most part is controlled by CC2. In the Wiedemann 99 model , the
exact implementation of the maximum speed oscillation using CC6 is not known.
There are two remaining parameters in the Wiedemann 99 model CC8 and CC9 that
control the acceleration at standstill and acceleration at 50 mi /hr respectively. If capacity is
defined as pre-queue flow rate, these two parameters do not have a significant effect on capacity,
because pre-queue flows occur before breakdown. If capacity is described as queue-discharge
flow, CC8 and CC9 can be argued to not affect capacity. The reason for thi s is that downstream
of a queue, the flow does not change by distance from end of the queue. The reason being that
flows are measured over time, and since there is no congestion (or queue build-up) downstream
of the queue, flow remains same over all the points within the section. For this reason ,
theoretically, CC8 and CC9 can be argued to not affect queue discharge flow . However, CC8
and CC9 do affect other flow processes.
The vehicle following behavior parameters look-ahead distance and number of observed
vehicles also have no significant effect on capacity. The look-ahead distance is not applicable

64
under congested or at-capacity conditions. The number of observed ve hi cles can affect th e
following process of the followin g ve hi cle, but there is no theoretical bas is to change this
parameter during calibration process . Tempormy lack-of-attention parameter was also considered
in the analysis, but was rej ected due to unavailability of data. These parameters cannot be
exposed during the calibration process, because of the combination effect it has with CC I
parameter. Lower CC l values with hi gher lack-of-attention will result in same effect as higher
CCl with no lack-o f-attention. Also, there is no sc ientific data available to differentiate "good"
from "bad" values, if incase used in ca libration .

7.3.2 I Speed-Flow Graphs

CCO, CC I, CC2, CC4, and CC5 contribute to capacity, and hence directl y a ffect speed-flow
graphs. However if these values are calibrated to capacity, it does not assure that these
parameters will perform accurately in flow regions other than at-capacity. This point was central
to the discussion provided in chapter 4. In addition to these five parameters, CC3 also affects
speed-flow graphs. CC3 affects the aggressiveness qf the driver, because lower absolute values
of CC3 indicate slower reaction to the lead vehicle: The other parameters in the Wiedernann 99
. .
model affect sped-flow graphs, but are not as significant as other parameters. Therefore,it is not
advisable to expose the rest of the parameters in the speed-flow calibration process. With respect
to the six parameters CCO to CCS, the comparative sensitivity of each of the parameters is
analyzed in a later chapter.

7.3.3 I Shockwaves
Shockwaves are impmiant flow processes in traffic flow. The speed of propagation of"jam
front" defines congestion patterns and impacts . Some research based on real-world vehicle
trajectmy data has shown that the shockwave speeds are fairly invariant and are not dependent on
traffic flow speed prior to congestion. (Lu and Skabaidonis, 2006). This information can be
invaluable in calibrating simulation models. Empirical analysis of shockwave speeds was
conducted using a one-lane circular road under constant global density. The empirical analysis
showed CC4 was the only parameter that significantly affected shockwave speeds. This can be
explained theoretically. While a vehicle is in a following condition, the first threshold that is
usually broken when the following vehicle starts to encounter a suddenly decelerating lead
vehicle is CC4. Therefore, CC4 significantly affects shockwave speeds. Higher absolute values
of CC4 result in higher shockwave speeds, because the following vehicle waits longer to brake,

65
thus res ulting in higher shockwave (backward propagating) speeds. An example of shockwave
from empir ical analysis is presented in figure 7.3.

Vehicle Trajectory Data: CC4 = -0 .35 and CC5 = 0.35

'
3000 -.------- _ ... - - - - - - - - - _ _ ... -
I I
- - - - - - - - - - J.-- -
I
- - - - - - - - J. - - - - - - - - - - - ... _ - -
I

'
'

2500

1000

2010 2020 2030 2040 2050 2060

Simulation Time (s)

Figure 7.3 Shockwaves representation in vehicle trajectory data

7.4 I SUMMARY
In this chapter a high-level calibration methodology is presented. Theoretical sensitivity analysis
of Wiedemann parameters is also presented. The sensitivity analysis is performed for three
different traffic representations: capacity, speed-flow, and shockwaves. Details of disaggregate
and aggregate data based calibration is presented in chapters 8 and 9. In chapter 8, a case study is
conducted to compare the speed-flow based aggregate calibration methodology with traditional
calibration methods based on capacity. Thereafter, integration of disaggregate and aggregate
calibration methodology is presented in chapter 9.

66
 VI I Case Study : U Sl Ol N B , San Mateo

8.11 INTRODUCTION

8.1.1 I Network Description

To demonstrate the usefulness of the proposed methodology and to provide results of comparison
with current calibration procedures, a real -world freeway was simulated and calibrated using the
described methodology. The case study network is a S-mile 4-lane section on US 101 Northbound
in San Mateo, CA. A picture describing the extent of the simulation model and detector locations
is provided in Figure 8.1. The data source is the California PeMS database (PeMS Group, 2007).
Data was collected over four detectors numbered 401199, 400935, 400007, and 400420; and from
January 2, 2007 to January 4, 2007 from SAM to lOPM over each day. It was verified that all
detectors were in good health and reporting throughout the data collection period.

67
 d(l(tRIJO
Ail S 4 1q . l 57 US 101 NB
SAN MA TEO COUNTY, CA

ANZA fl LVD
4(10007

BROADWAY
ll
400420
A IJ S4 17.4 J)

I -

,\IJ S 41 R. IS7 400q.\5

A HS 4 16 ? i i

DO RE AVE
4011 '>9
A BS 416.557

100720

t
N
AB S 4 16.107

(nt~)
KEHOE
AVE

Figure 8.1 Extent of simulation for US101 NB, San Mateo, CA.

8.1.2 I VISSIM Model

VISSIM 4.30 microsimulation software was used to develop and test the calibration procedure.
VISSIM is based on a psycho-physical car-following model and uses perception thresholds to
model drivers (Wiedemann, 1974 and 1999). The model was coded for the extent provided in
Figure 2 and detectors were place at exactly the same position as repmted by California PeMS
database (PeMS Group, 2007). Speeds and flows are aggregated and averaged over lanes over a
5 minute interval. This aggregation over time and space follows the procedures published by
PeMS Group (2007). In addition, a test network similar to one illustrated in an earlier section
was developed to test the applicability of test networks to assist in calibration of simulation
models.

V/SS/M Driver Behavior Calibration Parameters

The following five calibration parameters were used in the calibration procedure.
1. CC 1: The headway time (in seconds) that the driver wants to keep between vehicles.
2. CC2: Following variation - controls longitudinal oscillation in the car-following process.

68
 3. CC3 : Threshold for entering car-following- control s the start of the dece leration process,
i. e. when a driver recogni zes a precedi ng s lower vehicl e.
4. CC4: Following thres hold - contro ls the speed differences durin g closin g in followin g
process.
5. CC5: Following thres hold - control s the speed differences durin g openin g in followin g
process.
For a more detail ed expl anation of VISSIM microsimulation and the calibration parameters,
please refer to chapter 2.

8.2 I PATTERN RECOGN ITION BASED OBJECTIVE FUNCTION

The objective function used in calibration is based on the concept of reducing the difference
between speed-flow graphs obtained from the field and simulation. Traditionally, for comparing
speed-flow graphs, researchers have tested the closeness of the graphs by visually matching
simulated and field graphs. A fitness funCtion for measuring the closeness of simulated and field
speed-flow graphs is needed. Standard statistical parametric or non-parametric methods (e.g. Ma
and Abdhulai , 2002; Kim and Rillet, 2001 ; Kim et al , 2005) are not applicable for measuring the
closeness of speed-flow graphs , In addition, the fitness function should be automated and
consistent across multiple evaluations. The development of an automated comparison method to
replace manual eyeballing has the following benefits:
I. Multiple simulation runs can be compared without human intervention . .
2. Maintains consistency across multiple comparisons.
3. Produces a quantitative value for the objective function.

The use of a mathematical formulation for testing the closeness of graphs has been used for many
years in the field of pattern recognition for retrieving or matching images. The measures of
closeness are often referred to in the field of pattern recognition as similarity measures .
Alternately, the same problem can also be formulated using dissimilarity measures. A generic
objective function based on minimizing the dissimilarity between speed-flow graphs was
developed.
In order to formulate a fitness function to measure the dissimilarity of speed-flow graphs,
a closer look at important properties like shape, scatter, extent, and intensity is necessary. Since
the speed and flow measurements are stochastic in nature, it is also required to account for such a
noise created in the speed-flow graphs. Dissimilarity measures are easily defined for points, but

69
are harder for point sets . One of the more prominent dissimilarity meas ures for point sets is the
Hausdorff distance (Peka lska and Duin, 2005) .
But, Hausdorff distance does not effectively represent shape created by the point sets and
is not robust aga inst outliers or noise (Pekalska and Duin, 2005). However, a different
dissimilarity measure, which is based on template matching in pattern recognition, is more
appropriate to the problem at hand. An important part of a speed-flow graph is its shape. Shape
can be described in te rms of area. The dissimilarity of two graphs can be measured by calculating
the amount of area that is not covered by th e other. Since speed and flow measurements are
represented as point sets, discretization to conveti point information to area is necessary. The
discretization also has the effect of smoothing thus reducing noise in speed and flow
measurements for matching purposes. Since the information coming from the field and
simulation is often just patiial and not a complete speed-flow graph, the comparison is only made
over the space occupied by the field graph. The objective function (Z) for the calibration process
of the field is as follows:

Min. Z =Sum ofall the speed-flow area in the field data that is not covered by the simulated
data.
=F-FnS

Pseudo-code for fitness function evaluation

BEGIN
CONVERT field and simulation vector data to raster data at a certain resolution
FOR every cell in field data
CHECK IF the same cell has a value in simulation data
IF
TRUE continue
FALSE increment the objective function by a certain value
END
RETURN the final objective function va lue
END

The objective function can also be alternatively defined as symmetric Difference:

FM= F uS- F 11 S. This is the sum of the area covered by either ForS, but not both.
However, this function assumes that both graphs are complete. Since the field speed-flow graph

70
does not change during ca libration, this function essenti all y asks the simulation to only produce
points that are on the field graph . But this is valuable in two cases : one, when acc urate OD data
is available prior to calibration ; two, when the speed-flow graph from the field is complete.
These two assumpt ions were not made in this methodology. However, if good OD data is
ava ilabl e, the objective function F uS- F n S would result in a better solution
thanF- FnS .
This objective function F- F n S does not capture the frequency of occurrence of speed-
flow measurements. This information is purposely left out for reaso ns mentioned in an earlier
section. The objective function is consistent over different evaluations, because the number of
speed-flow points developed by the simulation model across each evaluation is constant. In order
to compare the described methodology with existing methodologies, two add itional commonly
used objective functions that do not require OD data are defined. In order to compare the
described methodology with existing methodologies two additional objective function s are used.
These two objective functions are
1. Minimize the difference between maximum 5 min . flows observed 111 the field and
simulation.
2. Minimize the difference between maximum 5 min. flow sustained over 15 min. observed
in fi eld and sim ulation .

There are ramifications related to the use of different aggregation period for speed and flow data.
If speed and flow information is collected and aggregated over a five minute interval, the
simulation model might need to be run for long periods of time and over many random seeds to
develop a reasonable sample of points. For example, if data is aggregated over five minute
intervals, it would require about 16 simulation-hours to produce 500 points on the speed-flow
graph. The amount of time required to run 16 simulation-hours depends on the size of the
microsimulation model. In addition, most of the optimization methods available are based on
multiple evaluations of the microsimulation model.

8.2.11 Test Network

In order to keep the amount of time required to produce a reasonable number of sample points to
a minimum, it might be advantageous to use a test network, instead of an actual network, to assist
in optimization of calibration parameters by producing speed-flow graphs. Test network is a
small-scale network developed to produce the widest range of speed and flows over the least
amount of time (Figure 8.2). However, the use of test network for calibration is only valid under

71
the assumption th at speed-flow graphs from a tes t network are equi va lent to th e graphs produced
by an actual network . Test netwo rk also provides other advantageo us in ca li bration process. A
simple test network was deve loped to replicate different fl ow co nditions and a compl ete speed-
flow graph. It is important to produce di fferent range of operations to deve lop co mpl ete speed-
flow graphs.

Section I : 2 mil es Section 2: 2 miles Section 3: 2 miles

R2

Figure 8.2 Illustration of the test network.

Figure 8.2 shows an example of a test network that consists of a mainline and two ramps. The
main line is approximately 6 miles long, and two rampsare placed at 2 miles apart. The mainline
has the same number of lanes as the actual network. Detectors are placed every ~ mile and no
detectors are placed near the influence area of on-ramps to avoid di storted results. A total of 9
detectors (3 in each section) are placed on the mainline. An artifici al demand pattern is created to
force demands above capacity in order to produce breakdown conditions on the freeway. The
demands are gradually increased, sustained, and dropped over time. It is not possible to replicate
a complete speed-flow graph without ramps. This test network was intended to represent the
mainline freeway segment as specified in the Highway Capacity Manual, but two ramps had to be
introduced in order to create the diversity in range of operations that would lead to a complete
speed-flow graph.

8.4 I EVOLUTIONARY ALGORITHM

The evolutionary algorithm presented in chapter 6 was used for calibrating the simulation model.
Each simulation model was run for 3 simulation-hours and five different random seeds.
Empirical tests were carried out to achieve a balance between number of runs and algorithm run
times. Since this method is based on comparative evaluation of different possible solutions, a
balance between number of runs and algorithm run times would make the algorithm more
practical. The evolutionary algorithm used 10 members per population. The choice in number of
members and runs per member was tested empirically, and the choice of 10 members and 5 runs
per member resulted in an adequate trade-off among optimality, diversity and run times. Since

72
there was the possibility of some of the individual s being repeated in future generat ions, the
fitness values fo r individual s were stored in a database to reduce computation times. The tota l
run times for the evo lutionary algorithms ranged from 12 to 30 hours based on the obj ective
function.

8.5 I RESULTS
In this section the results from three different applications of the methodology are presented.
I. Calibration of test network to the objective function s
2. Calibration of US 101 NB to the objective functions
3. Application oftest-network-calibrated parameter values to US101NB

The first application is on the test network shown in Figure 3 and the second and third application
on the US 101 freeway network shown in Figure 4. For each scenario, 5 runs were performed,
each run being 3 simulation-hours. Five minute speed-flow data was collected from the mainline
detectors. The data fiom multiple runs was aggregated and the speed-flow graphs were produced.
The three objective functions consisted of comparisons between the field and simulated data for:
a. Maximum five minute flows (Maximum Flow)
b. Maximum five minute flows sustained over 15 minutes (Sustained Flow)
c. Matching speed-flow graphs (Speed-Flow)

8.5.1 I Test-Network
Figures 8.3.a-8.3 .c show the results of the compari sons. The visual inspection of these three
figures shows that the speed-flow objective resulted in the best match between field and
simulated data. This is as expected since the objective function itself involves the minimi zation
of non-overlapping speed-flow points. Figure 8.3.a shows that the maximum flow objective
focuses solely on the highest flow values thus severely missing the congested regions of the
speed-flow graph. Figure 8.3.b shows that while the sustained flow objective captures the
congested region better than maximum flow, it does not produce the realistic scatter of data in the
congested region as shown in Figure 8.3.c. Consequently, the matching of the shape ofthe speed-
flow graph results in the most complete and realistic coverage of the speed-flow graph.
Theoretically, this means that the simulation based on shape-matching can better represent traffic
during longer periods of time and not just the top five or fifteen minutes of the day. The practical
implication is that the use of the new methodology would better capture the entirety of the peak
period which typically lasts several hours in most metropolitan areas.

73
Table 8.1 Calibration Parameters fo r the Test Network
Objective CCl (s) CC2 (m) CC3 (s) CC4 ccs
Default Va lues 0.9 4 -8 -0 .35 0.35
Speed-flovv 1.28 8.\1 - \0.92 -0.4 5 1.06
Sustained flow 1.06 7.79 -5.66 -2.5 0 2.3 2
Maximum flow 0.83 12.02 -14. 19 -1.42 2.3 1

Table 8.1 shows the calibration va lues that resu lted fro m the test network. The default values are
provided in the first row to draw a comparison with the optimization-suggested values. The CC I
column shows that speed-flow-matching produces the longest CCI value. This is expected since
speed-flow-matching involves the whole range of traffic conditions and not just the capacity or
short headway conditions. CC2 and CC3 show similar trends in that the maximum flow produces
the highest value while hi ghest sustained flow produces the smallest value. Speed-flow is
between those two values. For CC4 and CC5, the speed-flow results in a driver that is much
more sensitive to speed differences between the leader and follower than capacity (rows 3 and 4 ).
Table 8.1 shows that parameters calibrated using speed-flow graphs are significantly different
from those calibrated using capacity.

74
 Maximwn Flow based Calibration Res ults for Test Network
~-------.------.-------.------.------,

70 r - - .! ." _..~): .......... L~.!'."~~~~----l

I ..' - ' ~
........ ~-
t#
;f.~~-.~_:, _
I: .: ~ ~~i.:::,;; -
60

1l
~ 30\------1f------1---------t------t---------i

N \------+------1-------+---------i------~

Field
Simulation J

500 l(XXl 1500

Flow (veh/hr/ln)

(a)
Sustained Flow based Calibration Results for Test Network

~~----+-----~---1~~:~

~40f-------+-------+-------h~~
]
~30f-------+-------+-------+~r~---+----~

N\------+------1-------+---------i------~

10 Field
Simulation
0 o~=====500~====~~~
oco
~~--~~500
~----~N00
~----~2500
Flow (veh/hrlln)

(b)

(c)

Figure 8.3 Test network speed-flow graph comparisons for different objective functions.

75
8.5.2 I US101 NB Network
Ana logous to the test network cases, the same three obj ective functions were also applied to the
US I 0 I network. Figures 8.4.a-8.4.c show the res ults of the compari sons. Compared with the test
network, the simulated US 101 network produces a much more scattered pattern just before the
pre-queue flow area. Thi s might be explained due to ex istence of more va ri ed traffic conditions
for US!Ol NB than the test network. Once again the objective based on the matching of the
speed-flow graphs produces the simulated speed-flow graph that matc hes the field data the most
closely. Figure 4.a shows that maximum flow objective results in a speed-flow graph that is
shifted to the left, so flows in the congested region are underestimated significantly. On the other
extreme, thesustained flow objective results (Figure 8.4.b) in a graph shifted to the right so flows
are overestimated. It is only the third objective function that produces the closest match for all
traffic regions of the speed-flow graph (Figure 8.4.c).

Table 8.2 Calibration Parameters for US 10 I NB

Objective CCl (s) CC2 (m) CC3 (s) CC4 ccs
Default V a! ues 0.9 4 -8 -0.35 0.35
Speed-flow 1.09 10.59 -7.91 -2.5 0 0.64
Sustained flow 1.12 2.98 -5.5 -2 .59 2.46
Maximum flow 0.83 12.12 -7.7 -2.20 1.17

Table 8.22 shows the. calibration values that resulted from the simulated USIOI network. The
first column, CC1, shows that maximum flow produces the smallest value of headway which
represents overly aggressive driving. The second column, CC2, show that sustained flow
produces fairly insensitive behavior to longitudinal oscillation that is unrealistic. The third
column, CC3, shows that sustained flow produces a smaller value for the car-following threshold.
The other interesting observation is that CC3 values suggested by both the test network and
USlOl were close to the default values suggested by VISSIM.
CC4 and CC5 are speed thresholds in following for closing and opening processes. For
example, in VISSIM, when a car is in a following condition the driver tries to hold the
acceleration values to a minimum. The driver reacts when one of the following speed thresholds
is broken. In addition, CC2 al so controls the oscillation and results in driver reaction during the
following process. Therefore, in a closing condition, if speed differences exceed absolute value
of CC4, the driver reacts by braking. Analogously, if CC5 is broken, the driver reacts in an
opening condition by accelerating. Intuitively speaking, CC4 and CC5 should exhibit an
asymmetrical behavior since drivers are more sensitive when closing than when opening. The
speed-flow objective values for CC4/CC5 were not in compliance with this and the values

76
suggested by speed-flow objective are very different from the defa ult va lues 111 VISSIM.
However, the res ults of the test network are in comp liance with thi s effect.

Maximum Flow based Calibration Resulls for US!OI NB

00.------.------.-------.------.------,

50 1------+------+-------h~!!-:'!:
~401------+------+-------~!~-,~- ;~----~
1l
~ ~ 1------+------+-------~-, ~---+------~

~~----+------+----~~----+-----~

l(DJ 1500
Flow (veh/hr/ln) Flow (veMrrlln)

(a) (d)
Sustained Flow based Calibration Results for US lOl NB Sustained Flow based TeslNetwork;Calibraled Values on USlOl NB
00,------r----~,_------.------.------, 00,------r------,-------.------.------,

501------+------+------t-'~.'ft ;.,,l.:t''-----1
E
~401------+------+-------r ~; ~~~ t~~---4
:l
~ ~ r------+-------r-----~-- .,~---r----__,

~ 1------+------+-------+-----~------~

10 Field
Simulation
0 o~====~500
~==~-1~
00J~----~1500
~----~~~-----2500~
l(DJ 1500 2500
Flow (velt~lflln) Flow (vel~rrlln)

(b) (e)
Sp eed Flow based TestNelwork-Calibra!ed Values on USlOl NB
00,------r------,-------.------.------,

501------+----~+------~~~

~401------+------+-------ti~~!~~~~----~
1l
~~~-----+------+-------~~~~-4------4

~1------+------+-------+-----~------~ ~ 1------+------+-------+-----~------4

100J 1500 2500 100J 1500 2500

Flow (veMrrlln) Flow (veMrrlln)

(c) (f)
Figure 8.4 Speed-flow graph comparisons for US I 0 I NB using different objective functions.

77
 8.5.3 I Test-Calibrated-Values Applied to USlOl NB
In order to evaluate the applicabi li ty of using the test network to ca librate an actua l network, the
test-network-ea! ibrated va lues for each of the objective functions is fed into the US I 0 I simulation
model. The results of such comparisons are provided in Figures 8.4.d-8.4.f. The dissimilarity
(opposite of closeness or goodness-of-fit) of speed-flow graphs in figure 4 was measured using
dissimilarity : F- F n S. Table 8.3 shows the fitness values of each of the graphs in Figure 4. A
lower dissimilarity value indicates a bette r match . The comparison a mongs t thes e values is fair,
because the speed-flow meas urem ents are made over the same s imul ation model (US I 0 I)
resulting in equal numbers of observations. The comparison amongst graphs produced from the
test network and US 10 I is not fair, because the number of observations on the test network is not
equal to US101. Figures 8.4.c and 8.4.fare reasonably similar to each other. The speed-flow
objective evaluation for these graphs (Table 8.3) shows a 10% variation in objective function
values for speed-flow based calibration, A much higher variation is observed for the other
objective functions . However, the graph achieved using test-network-calibrated values for speed-
. flow objective is quite comparable to the ones achieved using US 101. In addition, the test
network can take advantage oftime again when applied to large-scale simulation models.

Table 8.3 Speed-Flow Objective Evaluations for USIOI NB Speed-Flow Graphs

Objective Speed-Flow Graph Dissimilarity: F- F n S
Calibration using Test-Network-Calibrated
US101 NB values on USlOl NB
Speed-flow 193 210
Sustained flow 244 317
Maximum flow 237 293

8.6 I CASE STUDY SUMMARY AND CONCLUSIONS

A new methodology for calibrating microsimulation was presented. This new methodology
introduced the use of speed-flow graphs as a calibration objective, defined a fitness function for
point sets, and included an automated technique for optimization based on Evolutionary
Algorithm . Encouraging results were obtained from the application on the US I 01 freeway in
California. The results from speed-flow objective function have been shown to perform better
than the objective function based on maximum 5-min. flow and maximum 5-min. flow sustained
for 15 minutes. The calibration parameter values from the speed-flow calibration are shown to
perform well in all three regions of the speed-flow graph: free-flow, congested, and discharge.
The speed-flow objective calibrated parameter values resulted in a scatter in speed-flow graphs
similar to the field graphs. This accounts for the stochastic nature of speed and flow observations

78
on the field. T he res ults s uggested by the Evo lutionary Algorithm based o n speed-flow objective
are plausibl e and sometimes even close to default va lues. The evolutionary a lgorithm suggested
ca libration parameter val ues resulted in speed-flow graphs s imil ar to the one noticed in the field .
It is important to note that the evo luti onary a lgo rithm could have suggested physically
unreas onabl e calibration parameter values, if the degree of freedom was too high. The objective
function can be expected to perform sim il a rly even if a diffe rent sea rch and optimization method
was employed. But the comparative effic ienc ies of the search and optimi zat ion methods are
beyond the scope of this study. It can a lso be seen that tes t networks can be employed to assi st in
calibration of parameter values when the s imulation networks are large. The pattern recognition
based objective function can be extended to measure dissimilarity of other traffic relationships
like flow-density, speed-density, flow-speed-density, and flow/speed-contour graphs. For
example, the pattern recognition objective can account for intensity to evaluate closeness of
speed-contour graphs. The pattern recognition can be defi ned in term s of volume, instead ofarea,
to evaluate closeness of flow-speed -density graphs .

79
 IX Implementation of Integrated
Calibration Methodology

9.1 I INTRODUCTION
In the previous chapter, it was shown that a speed-flow based calibration methodology performed
better than traditional objective functions based on capacity. After demonstrating the
applicability of the pattern recognition based speed-flow calibration methodology, an integrated
calibration methodology is presented in this chapter. There are several questions that were left
unanswered in the previous chapter. In this chapter, a comprehensive calibration methodology
presented in chapter 7 is implemented.

9.2 I NGSIM DATA ANALYSIS

As mentioned in chapter 3, NGSIM data consists of sub-second vehicle trajectory information
over two freeways in California. The two microscopic data sets are available on the World Wide
Web (NGSIM, 2007). The vehicle trajectory data was analyzed to gather information that can be
used in calibrating the VISSIM driver behavior model. Many of the Wiedemann model
parameters and thresholds are represented on a relative velocity vs. relative distance graphs. For
this reason, much of the data analysis concentrated on developing relative velocity vs. relative
distance graphs. These graphs are developed by selecting a pair of vehicles (leader and follower)
and plotting their relative velocity and relative distance. A sample graph from such a process is
presented in figure 9.1.
Many different kinds of information were obtained from the microscopic data set. Much
of the data collected and analyzed from NGSIM data set is related to studying vehicle following
behavior: the opening process and the closing process. Much of the effort was dedicated to
understanding calibration parameters CC2, CC4, CCS, and CC6 . A typical relative velocity vs.

80
relative distance graph consists of distance and speed oscillation in following. In order to extract
usable information that can be related to these four parameters, the researchers extracted
speed/distance osc illation loops from the graph. For exa mple in fi gure 9. 1, three di stinct
oscillation loops: red, blue, and gree n were obtained and studi ed . The oscillation loop was
selected based on the vehicle sta1iing at a point on the graph and approximately returnin g to that
point over time, thus completing a single oscillation process .
For instance, the red loop from fi gure 9.1 is presented in isolation in figure 9.2. Figure
9.2 describes data collected from these isolated loops.

Sample NGSIM Relative Velocity vs. Relative Distance Graph

25 --- - -- ------1---
, ---:-- -----:---- --- ----- ---- ---- -------- --- ----------
I
1
I
I
I
I
I
I

' ' ' ' '

'
'
'
'
' .
.s
~ 20
c
m
u;
0
<l>

~ 15
())

Cl:

10

Relative Speed (m/s)

Figure 9.1 Sample NGSIM relative velocity vs. relative distance graph

81
 Relative Velocity vs . Relative Dist ance Gra ph: Term s and Definiti on s
15
I I I I

14 - - T
'
- - - "' -
I
- - -, -
I
- - - - - - r - - - - - - T -
I
- - - - - ,
I
- - - - ,
I
- - - - - - r - - - - - r - - - -
I
I I I I I I

I I
I
I
I
I
I
I
I
I
I

13

12
g
u
Q)
11
c
m
u;
0 10
<l>
-~
ro 9
Q)
0:::
8

5
-5 -4 -3 -2 -1 . 0 2 3 4 5
Relative Speed (m./s)

Figure 9.2 Relative velocity vs. relative distance graph : terms and definitions

There are six different types of data collected from the oscillation process.
1. Minimum following distance during oscillation
2. Following distance variation during oscillation .
3. Average following distance during oscillation
4. Maximum absolute positive speed oscillation
5. Maximum absolute negative speed oscillation
6. Average following speed during oscillation

It is important to note that these six types of data do not exactly represent CC2, CC4, CCS and
CC6. However, this data can be useful in qualitative calibration of these parameters. For
instance, following distance variation is not exactly CC2, because SDX is defined as ABX+CC2,
where ABX is a speed dependent term. However, following distance variation is still a good
indication of CC2. Similarly, maximum positive and negative speed oscillations are not same as
CC4 and CCS, but are a good representation of symmetry or asymmetry in speed oscillation
thresholds. In this research, following behavior was analyzed from approximately 300 different
oscillations, equally distributed among USlOl and I80.

82
 The first set of observations is related to average speed durin g osc illation and minimum
following distance (figure 9.3). The two pictures in the figure relate to NG SIM data sets from
USlOl and ISO. The observations from USIOI are very similar to observations from ISO. The
observations show a fairl y linear relationship among these two quantiti es with quite a bit of
scatter. The slope of the best linear fit is greater than one in both scenarios. Observations were
also made on average following distance and average following speed relationships (figure 9.4).
A fairly linear relationship can be seen in USIOI data, but a more constant rel at ionship is seen in
ISO data. This might be due to the absence of low speed data in US I0 I dataset. Or otherwise, a
higher following variation might be present at lower speeds. In order to understand the speed
dependence of following distance variation, the relationships are plotted and shown in figure 9.5 ..
However, the data shows constant variation across various average speed levels . This indicates
either missing low speed data in USI01 or a different driving behavior on ISO . In addition, ISO
shows consistently higher following distance variation than US I0 I. The dependence of
following variation on following distance was also analyzed for both data sets and is presented in
figure 9.6. The following variation seems to be fairly invariant of following distance with much
ofthe values below 10 m at all speed levels, except for some outliers. In summary, the following
variation is fairly constant across various speed levels and following distances, which is one of
the primary assumptions in the Wiedemann car following models.
The histogram and descriptive statistics of following variation of US I 01 and ISO are
presented in figures 9.7 and 9.S and tables 9. I and 9.2. This provides a good understanding of
distribution of following variation. The following distance variation in both data sets show an
exponential distribution, with majority of the values falling below 6 m, and highest number below
3 m. The relative closeness of mean and standard deviation in tables 9.1 and 9.2 also indicate an
underlying exponential distribution. The median following distance variations for US 101 and ISO
are 3.4 m and 4.4 m respectively.
The next set of relationships that is analyzed relates to speed oscillations. The relative
velocity vs. relative distance graphs were analyzed to determine changes in relative accelerations.
Sample graphs of USlOl and ISO datasets from different time periods are shown in figure 9.9.
The blue points in the graph indicate relative acceleration that leads to reduction in absolute
difference in relative velocity. The red points indicate relative acceleration that leads to increase
in absolute difference of relative velocity. In other words, blue markers indicate path moving
towards origin and red markers indicate path moving away from origin. These points indicate
change in direction of slope along the relative velocity vs. relative distance graph. The graphs in
figure 9.9 show a remarkable similarity in distribution of blue and red markers. The distribution

83
is fairly uniform across positive and negative relative speeds indicatin g a symmetric behavior.
The distribution does not show an in creasing speed oscillation ove r increasing followin g
distances. However when max imum speed oscillations are plotted for varying speed leve ls as
shown in figure 9. 10, the maximum speed oscillation shows increasing variation over increasing
distances. Again, the graphs from USJOJ and 180 show remarkable similarity. The minimum
maximum -speed-oscillation for both datasets is always greater than - I and I m/s. In addition, the
maximum positive and negative speed oscillations show a distinctive slope. Increasing distances
show increasing maximum speed oscillation. The CC6 parameter describes the speed oscillation
dependency on following distance. No specific recommendation is provided about CC6, since
the exact implementation ofCC6 in the Wiedemann 99 model is not known.

84
 Average Following Speed vs. Minimum Following Distance in Oscillation
NGSIM US101 Data

US101
45 ' - ' : ' : + : +
g : ;
r ' -- ,. --

40 -- ----.Y = 1.1~x + 6 . 6.------~-------l --- ---~------)- - -----~--- ----~- ----

n; : : ' : :+ + : '
0
~ 35 ---- -- .: :
------ ... ---- -- -------- .... ------ .- ----- ... ------ ....:-------
: : :* :
+ :.... ---- -- .. ---- -
+ : : + :
c: : : : : : : : + : :
Q) 30 ------;----- - ; -- ---- -:------ -~--- --- ;--- - --;- _-t_ -- ~-- - ---------- '- ----
u
c:
-t; 25
0
g> 20
~
0

~ 15
E
E iO
I I I I I I

5 ------f------~--------------~------t------~------~-------~------~-----
0 '

'
'
OL----L----L----L----L---~--~L---~--~----~--~

0 2 4 6 8 10 12 . 14 16 18 20
Average Following Speed in Oscillation (m/s)

Average Following Speed vs. Minimum Following Distance in Oscillation

NGSIM 180 Data

45
I I+
,-------.-------'------,------
:
.,. -- I :

,, ' '
' '
c : y = 1 4*x + 5: : : : : : :
0 40 ------f , :-------:-------t--+-~-------: --- ----:------ : ~-----
ro I I 1

~ 15
E
E 10
' ' '
5 I I I I I I I
--1--------------f------i------1------1-------~------r-----

' , '
-*'

Average Following Speed in Oscillation (m/s)

Figure 9.3 Average following speed vs. minimum following distance

85
 Average Following Speed vs_ Distance in Osc ill ation : NGSIM US 10 1 Data

55[ !: !: !: !:
50 ------- ~- ----- - ~------ - ~------- ~--- ---- ~ - ------:-- -- - -- : ------- ; - - - - - -
!: . !: !: !: ~
I ~ : : : : : : . :
c 45 - - - - - - - ~ - - - - - - - ~ - - - - - - - ~ - - - - - - - ~ - - - - - - - +- - - -- - - f ------- i ------ i-------
ro
0 I

l
I

I
l

I
I

1
o
1
I

I I I I I I I I

u
Ul
40 -------:-- -----:- ----- -:-------:- ------:-------: ---- . --: 9.--; . -:----- -
0
c:
Q)
35 1 I I I

----- --~-- -- - -- ~- ------ ~- ----- -~-- ---- -!-- -----:- - - ~--- !-- -----:---- --
I I I I
I

I
'
I

I
I

I
I

I
')

u
c I I :. I I . :

~ 30 ----- --~ -- - - - -- ~ ------ -~------ -~ ~--- -- -~-- _t ___ ~ --.---- ~ . --- --- ~- -----
I I I. I I I

0 I I I : : I :: :

------- ~ ------- ~-- ----- ~--- .-+- ~--- t--!---. _'t- t------
.t +--
I I I I I I I I
C
25
.
: ! t. --
.. :
c: 1 -,.--- -
-~
0

0
LL.
20
: : : : : : +

----- -- :~------- :~------- :~---- -.-~~:-- -.:---::-+/ _.. -..:-- t~- -.;-..----::-;.:- ----- .
:
I
:
.: ..... "1."*:
I
:
:. +
I
: .:
I I I I I

15 ------- ~----!- -~--- ~ - ~-- ----,.--- .... .Y.+y- t- ...- .:._ ~------- i------- ~---- --
I I .: :. : .. I I I

I I : ~ : I +, :
10 L L t .. L L 1 1 1 1
I I I I : : I I

Average Following Speed in Oscillation {m/s)

Average Following Speed vs_ Distance in Oscillation : NGSIM 180 Data

55

,E, 50 ---- ---~ ------ -~- ---- --~- ----- -~- ------! ------ -~-- ----- ~ -------:-------
' ' '
' ' '
' ' '
~
c: 45 --- -- -- i-------- i-------- i-------- i--------f.-------+-- -----+--- ----+----- -- I I I I

-m
--u 40
0
Ul
-------:--.----:-------:--,_---:- -.---- :------:-------:-------:-------
I

I .
I

I
I

I
I

I
I

I
I

I
I

1
I

E
QJ
35 - - - ~ --- ~ - - -- - - - ~ - - - - - - - ~ - - -+_._ - - ~- - - - - - _.~ -!- - - - -! -------!------_;_ ------
I I I I I I I I

u ' ' '

c: : I I 'i I : I I
ro
u; 30 ------- ~----. -- ~----- - *~- ------ ~- ._----- ~-.-.--- i------- i------- i----.--
I I I I I I I I

0 : .. : : :.. + : .. : : :
----.-- -::--- _._~_ - --- ~-.- t-- -t--- -.---;--- --- -~------- t-------!-------
Cl
25
.I. . :... : . :. . .+. :: :: ::
c:
~~
+ : : ~ : : ~ : :
~

: . :. . . . . : . .
0 20 ------ -'1-- ...----..---- -- --...~+.-~_.. ----~--- ---- -------.-- -------- ----
LL. ~ : : . i
Q)
C : : : : :
!I! 15 ------ .t---\--- -f------- ~- ..:.- .~~.----- ~- .t. ---- ~- -.---- i-------;-------
+ + :+ ~. : + :. : : :
"' : I

~
10 ,
:. + .:. :

:
I

I
:
I
. ~ ~
:
I

I
:

:
I
I
:

:
I
: :

:
I

I
:
..... .... .. L .............. L ........ ..... L _ ............ L .+........... J. .............. 1 ............. 1 ............ 1 .. .. . . . . ..
:
I

I
:
I

I
-

5
0 2 4 6 8 10 12 14 16 18
Average Following Speed in Oscillation (m/s)

Figure 9.4 Average following speed vs. average following-distance

86
 NGSIM US 101 Data
30.----.-----,----- ,----,-----,-----, -----,----.

I I I I

,_ 25 -- ------~ ------ - - }- -- ---- -~---- ---- ~ -- - ----- ~- -- ----- ~-- ----- - ~-- ------
c +
0

m
u
0 ------ -- --- --
.. I I I

20 - -~ - - --
I
-- --~-
I I
~-- - --- -- ~ --- -- --- ~-- - --- - -~ - - --- --- ~ --- -----
I
c
c +
m
0

;::: 15 -------- ~-------- ~-------- ~------- - ~ -- - -.---- ~--------.:....-------;--------
ro I I ~ . I I I I

>Q)

u
c ..'
ro
u; 10
I

. . 1 I

-----.--~--------~--------~--------~------ -- ~--------~--------~---.----
: : :
I

: ..
I

: . . I

:
I

:
. ....
0

. : .
C1) I ...

c +
'
-~ : : :. : : :
~ 5 ------- -~--- ----- ~------- ~----- .-- ~---- +- - -}- _t;----- .:..-- ---- ~----~---
0
l.L t : ~+, +
.. ',+
:

: .. : .:

:+ .. i t
I I

, I + ~* +...: ! ..... ,:: : ..

i* + I +
I ::

1
QL---~----~-----L-----L----~----~----L---~
l ( 't
2 4 6 8 10 12 14 16 18
Average Following Speed in Oscillation (m/s)

NGSIM 180 data

30r----.----.-----r----.----.----.-----.----T----.

_E_ , , , , , , , ,
-- 25 . --.. -- ~ -- --- - - ~ ---- - - - ~ - - --- -- ~ ---- -- - ~ -- --- - - ~ -- -- - --; -- - - - - - ~ - -- - - - -
c
0

-ro
u
0
I I I I I I I I

20 -------~-------~-------~-------~-------~-------~-------~-------~-------

.--
.E '
c
0 :
-ro I I I

-.- -- -- -- --- --- -- -- - - - - -- - - - -- -

'
I I I '
I . 1

--
>
;:::

Q)
<..>
15 -- - -- - -} - - - - - ~ - -- - - - +,
I

..
-~ -- I

... .
I

'
~
I
~
I
- ~ -
I
~ - --
I
:' -

c
ro +.: : :. +'
;;; 10 ------
I

~------- ~------ ~ ------ -:------- ~- ....... ----

I I

~------- ~------- ~-------

I

0 :.. + : : : I I + I I

en
c
I
i+
~ +1
: .. * +'
+ :

,.
t
-~
0 ~. .. . : . ~ ..
5 -------.-- ~--: :+"" ~---- !*.-~---- ~-- -- """1" ..... --- ~ ------ ~------- ~-------
:. . : . . i : :
0

... :tt-9 .: i~
I .... I I ....
l.L
i* : * .:
t.++
;+ ; ; #. 1 ;. ; ; ;
I. . , : I I

QL---~----L---~----~--~~--~--~----~--~
0 2 4 6 8 10 12 14 16 18
Average Followng Speed in Oscillation (m/s)

Figure 9.5 Average following speed vs. following distance variation

87
 NGSIM US10 1 Data
30.---.----.---.----.---~---.---.----.----.---.

~ f ----------: __ _:__ _ ~ _; ______ r

25 J --

~ 20
0
- - - - - - : - - - -- -
' ~----
' - - _:_' ---- - .L.---
' -- _ :~--- _-t- j----- - ~- --- - --
' ' '
I I 0 I I

+ +:+:
c
c
0
m I + I I

15 -- -- -- ; - -- --- ~-- --- - -:- ------ ~--- +- :p --- -- ~ --- --- ~-- -----
> : : ~
-rt- : : : :
(l)
u : +
c
ro
1i) 10
0
(J)
c
":
. ' -----
------r
0
LL .
+

10 15 20 25 30 35 40 45 50 55
Average Following Distance in Oscillation (m)

NGSIM 180 Data

30

' '
E I I I I I I I I

c
25 - - - - -- ~ - -- - -- ~ - - - - - - -:- - - - - - -~ - -- - -- ~ --- - -- ~- - ---- ~- - - -- - -;.. - - - - - - ~ - -- - - -
0
ro
u I I I I I I

------ ------ ------:- - - - --+'- +

I I I I I I I I I

0 20
I
~
I
~- ----- - -- -- --
I
~
I
~
I
~ --- --- ~- - - -- - -~ - -- - --~ - -- - - -
I I I

'' '' ''

c
c
+.+
0
ro +i
----t-- ---+~-- ---=fi- ------~------t--- --- ~------ ~- --- -- -~----- -~ - - - - --
I I I I I I I I I

>
~ 15 --
: +: ,:+ : :
' ++ + :' +:'+
I :

~
C : :
' ' '
~ -'f --- ~ -~+- -- ~-t--- *:------ -~- +--- ~--- --~-- ---- ~- ----- -~- ----- ~------
I : : :

' .r-t+ '* ': ++

10
0
-+-? ...D -l-
+ :' :' :' :' :' :'
= :+-t=JH""": + :+ : :+ : : : :
c
-~
.
++.- ~+
5 - -+- - #:!!+~
: + :+ : : : + : :
-., :rJr.::---+~+--- i------ ~------ ~-- - --- -~- --+-~--- -- -
0

+~t\-tt:t + ++- ~
LL

-if-1 + -fr.4t(
-If++ T'
:' + :' :' :' :'
+++i+- ~+ +i + i
QL---~---L--~L---~---L~~----L-__J __ __ L_ _~
i-f i i i i
5 10 15 20 25 30 35 40 45 50 55
Average Following Distance in Oscillation {m)

Figure 9.6 Average following distance vs. following distance variation.

88
 Histogram of Following Variation : NGSIM US101
so
_ _ _ _ _ _ _ ... _ _ _ _ _ _ _ ._ _ _ _ _ _ _ _ .. _ _ _ _ _ _ _ .1 _ _ _ _ _ _ _ ... _ _ _ _ _ _ _ - - ----
70 ' I I I I
I I I I
I I I I
I I I I
I I I 1
I I I I I

60 - - - - - - .L - - - - - - . L - 1 . - - - - l 1 - - - - . - .
I I I I I
I I I I

' '
'' ' '
' ' '
' ' '
50 ------- --------------------------------
I
'
---------------
I
I
I
I
I
I
I
I
I I I I I
I I I I I

"'cu"' '
'
'
'
'
'
'
'
~ 40 -------r-------r-------,- ------y-------r------- ------
I I I I I

rr '
I
I
I
I
I
I
I
I
I
~ I I I I I
I I I I
l..i... I I I I

30 -------r-------r-------y-------,-------,------- ------
I

'
I

I
I

I
I

I
I

I
I I I I I
I I I I I
I I I I I
I I I I I
I I I I

20 r-- r T T T --- --
I I I I I
-

I I I I I
I I I I
I I I I
I I I I
I .1 I I

10 -------~---------
1 I
- --------------
I I
- --- --- -- -
I I I I

' '

0
0 3 6 9 12 15 18 21 24 27
Following Variation (m)

Figure 9.7 Histogram of following variation on US 10 l

Table 9.1 Descriptive Statistics of Following Distance Variation: USIOl

Following Distance Variation Statistics NGSIM USJOJ

Mean 4.966987
Standard Error 0.378934
Median 3.4
Mode 1
Standard Deviation 4.732879
Sample Variance 22.40015
Kurtosis 5.766431
Skewness 2.26217
Range 26.8
Minimum 0.7
Maximum 27.5
Sum 774.85
Count 156
Largest(1) 27.5
Smallest(l) 0.7
Confidence Level(95.0%) 0.748541

89
 Histogram of Following Variation: NGSIM 180

50 ------ -- ' -------y-

--------r- ' ------ -y--------
' --------

_____ ___ ._ _____ ___ ... _____ ___.., ________ --- ----- - ---- --
40 '' '' ''
' ' '
' ' '
' ' '
""
u
c:
'
'
'
'
'
'
~ 30 '
---- - ---r----- - -- y' - -------y-------
' - -- - - - - -- --- - ---
rr ' ' '
~ ' ' '
' ' '
lL ' '

________ ' ______ __... ________ ------ - - -------

20 ...

' '

'
'
'
10 '
-- - -----y------ -- -------- -------

0
0 3 6 9 12 15 18 21 24
Following Variation {m)

Figure 9.8 Histogram of following variation on 180

Table 9.2 Descriptive Statistics of Following Distance Variation: 180
Following Distance Variation Statistics NGSIM 180

Mean 5.494872
Standard Error 0.336752
Median 4.4
Mode 1.9
Standard Deviation 4.206029
Sample Variance 17.69068
Kurtosis 1.837483
Skewness 1.374309
Range 20
Minimum 0.5
Maximum 20.5
Sum 857 .2
Count !56
Largest( I) 20.5
Smallest( 1) 0.5
Confidence Level(95 .0%) 0.665215

90
 N GSIM ISO Data NGSIM US I 0 I Data

f RelatNe Veroc uy vs Reialive Distance G1 apw NGS!M 18() J 00 . d 15 PMJ'

70

60
.. t. ~ i
.s- so f ~~.
4 .. ..

01 ~5---.-c
, o:---~.5,------:':-----:---::to,-----:'15'
Relat ive Velocity (mls) Re lativeVeloc1ty(mls)

(a) (d)

Relative Ve locity vs. Relat ive Distance Graph: NGSIM 100 5 00 5 15 PM

70

60 .. .. . .. ... .. ... . . + ; .. .. t .... 60

+
50 ... . .... . .. . , .. + 50
+
E t
! 40
0 +
...... .+. ...; .. ....+.
.
-~
a:
30

20 ..... . ... ' t + -~- ... 20

+ :+

to .' +: ~ .. : . . ... :t .. tO . ... .. . ... ... ; .~- - -

~1~
5 ---.-:-,o:---~.5~---:':----!,-----::
, o,----:',.5
Relative V el ocity (mls) RelaliveVe locity(mls)

(b) (e)

Relative Velocit y vs. Re lative Distance Graph NGSIM DJ 5:15- S.lJ PM Rei al ive Velocit y vs . Rel ative Distance Graph: NGSIM US 10 1 8:20- 8:35AM
70 7 0.-----~----------~------.-----~-----,

~ ~
60 60 . Lt . .
f t ' . ..
~ .:. +.t
50 . . . .. . , ... t.,!- ., .. .. .. .
~

+
20
+
+:
10 10

~~
5 -----.::
,o;---~.5~~~-:':-----~-----::
, o;-----:',5.
Re lat ive Velocity (mls) RelativeVe lo city (m/s)

(c) (f)
Figure 9.9 Speed oscillations in relative velocity vs. relative di stance graphs

91
 Maximum Positive and Negative Speed Oscillations in Following: NGSifVl US101 Data
55.---~---.----.----,---,,---.----.----~--~--~

r- __ ___!______ ~ ___+___-:-______ ~ _____ _!___ ___ ; ______ ~I+_______ ~ ______ ~ ____ _

t I I I ~ I i I I
i l l I I I I I I

50
:
I
;
I
: -+
I
:
I
:
I
:
I
+: I
:
I
:
I

45 L... ~- - -7 --- - -- -{--- - -- -:-- ----- :- -- -- -- t--- -- - ~--- --- -:---- - -- :- ---- -- i- - ----
'E' ' '
' ''
ID 40
c'-'
m
t> 35
0

-4 -3 -2 -1 0 1 2 3 4 5
Maximum Speed Oscillation (m/s)

Maximum Positive and Negative Speed Oscillations in Following: NGSIM 180 Data
55 .

I I I I I I I I I

50 - - - - -- T -

I
- - - - - ., -

'
*
- - - - - -,- -

I
-- - - - r- - - -- - T -

I
- -- - - ., - - - - - - ., -

:
-

*
- - - - - r - - - - - - r - - - - -

' '
45 ...... l ...... J ------------~----- - - ! ...... J------~----- - -~-- - ---L-----
' I I I I I I .I

g + I I I
: +. I
:
I
: I

(l) 40 - - - - - - .1. -
I
- - - - - .J - - - - - - -- - - - - - - L. -
I I I
- - - - - J. - - - - - - .J - - - - - - ..J- - - - - - -
I
L.. -
I
- - - - - ... -
I
- - - -

'-' I I
'.
I . I
c
m '
t> 35
0
0)
c
~
30
0
0
lL 25
(l)
C
~
ID 20
~
15

10

5
-5 -4 -3 -2 -1 0 1 2 3 4 5
Maximum Speed Osdlation (m/s)

Figure 9.10 Maximum positive and negative speed oscillation vs. average following distance

92
9.3 l DISAGGREGA TE DATA BASED CALIBRATION
After an extensi ve analys is ofNGS IM vehic le traj ector ies and fo ll owing processes using re lat ive
velocity vs. re lative distance graphs, qualitative calibration of relevant Wiedemann car-followin g
parameters is presented. As exp lained in chapter 7 figure 7.2, di saggregate and aggregate
calibration procedures are iterative. Although the procedure was iterative, the material presented
in this section has been presented as a whol e to avoid back-and-f01th between sections. The
disaggregate data based calibration was performed us ing re lative ve loc ity vs . relative distance
graphs. A one-lane circul ar road , similar to the on e described in chapte r 7, was simulated to
collect speed osc illations in re lative velocity vs. rel ative distance graphs. The NGSIM data
showed symmetric speed oscillation, and thus the calibration effort was narrowed to symmetric
speed thresholds (CC4 and CC5) . The various graphs at various s imulation resolutions are
presented in figure 9.11. The two speed threshold pairs tested include [-0.35/0.35] and [-
0.70/0.70]. Other speed thresholds pairs were tested but rejected due significant difference from
NGSIM data. The rejected graphs can be seen in figure 9.12.

After analyzing the graphs, three possible choices that indicate a close match with NGSIM data
include the following:
1. CC4 = -0.35 and CC5 = 0.35 at Simulation Resolution= 3
2. CC4 = -0.35 and CC5 = 0.35 at Simulation Resolution= 5
3. CC4 = -0.70 and CC5 = 0. 70 at Simulation Resolution = 5

In order, to understand the correct choice maximum speed oscillation graphs [figure 9.1 0] from
US 101 and 180 were used. A closer look at the graphs produced using these options are provided
in figure 9.13 . The closer look at the graphs reveals that only option 3: [CC4 = -0.70 and CC5 =
0.70 at Simulation Resolution= 5] provides a minimum maximum-speed oscillation ranging from
[-1 to 1 m/s] like NGSIM data shown in figure 9.10. For this reason, the choice ofCC4 and CC5
of -0.70 and 0.70 m/s respectively was chosen.

93
 CC 4 = -0.35 I CC5 = 0.35 CC 4 = -0.70 I CC5 = 0.70

Retatrve Velocny vs Re tatrte Otstance Graph CC.& =-0 35 J CC5= 0 35 RelatiVe Vetoccty YS RelatiVe OISI<lnce Grapns CCA= .Q 70 f CC5 = 0 70
Stmulatcon Resolulton = J Stmulat10n Resolutcon = 3

.] !
6()

50

~ 40

6"
-~ 30
"
Cl:
20

15
Relative Velctty (m/s) Relatrve Ve loctly (m/s)

(a) (d)

Relat ive Velocily vs. Relative Distance G1aph. CC4= .0.35 I CC5 = 0.35 Relative Ve locit y vs. Relative Distance Graph CC4 = -0 70 I CC5 = 0.70
Simulation Resolution = 5 [Default) Simulation Resolutmn = 5 [Qera ull)
70,------.-----.----~~------------~r----, 70,-----.------.~~.,--r---~----.-------.

6()

50

20

10

0
-15 -10 -5 15
Rel ative .Velocity(m/s) Relat iveVeloctty (m/s)

(b) (e)
Relative Veloci1y vs. Relative Dis tance G1aph: CC4 = -0.35 I CC5 = 0.35 Relative Velocity vs. Relative Distance Graph: CC4 =0.70 I CC5 = 0.70
Simulation Re solut ion= 10 Simulalion Resolu!ion = 10
70

6()

I I
c
"
i5
~

-
"" 20
Cl:

10

0
-15 15
Relative Velocity (mls) Rela! ive Ve locity (mls)

(c) (f)
Figure 9.11 Speed oscillations in relative velocity vs. relative distance graphs.

94
 Rela!lve /eloC !ly V) ~ela 1 r.-e Oslance Grap n cc~= 1 I CC5 =t Rela11ve VelocIY vs Relal rre Ostance Graph CCd = 1 5/CC5= 15
10
701

50 60

.
.s
50
1 I
50

6
~ 40
;;; i! ~ >10
5
.~ 30

"
o;
a:
i ~:ll
1i
" 20
er
20

\Or
j 10

0 0
15 -10 5 \0 15 -15 15
RelatiVe Velocly (mls) Rel aiiVeVeloc ty(mls)

Figure 9.12 Rejected speed oscillations in relative velocity vs. re lative distance graphs.

95
 Retahvt: Vetoc 11y "s Relal tve Otstance Graph - CC 4 = -0 35 / CC5= 0.35
Simul ation Resoi ulion = 3

Relative Velc il y (m/s)

Re lative Ve locit y vs . Re lative Dist ance Graph: CC 4~ -0 35 I CC5 ~ 0.35

Simulation Resolution ~ 5 JDefauii J

Relative V el oc il y (mls)

Rel ative Ve locit y vs . Re lative Dist ance Graph: CC4 ~ -0.70 I CC5 = 0.70
Simulation Res oluti o n~ 5 JD efaultJ

~
u
c
"'
0"'
~

~
Qi
<r

5 ...... .......... ... . .

OL____ L_ __ J_ _ _ _ L __ __ L_ _ _ _L __ _~--~--__J

-4 -3 -2 1 0 2 3
Re lative Ve locity (mls)

Figure 9.13 Closer look at speed oscillation in relative velocity vs. relative distance graphs.

96
9.41 AGGREGATE DATA BASED CALIBRATION
Aggregate data based ca libration was co nducted for US I 0 I simulation models as described in
chapter 8. However. the significance or se nsitivity of eac h of the parameters to fitness function is
analyzed . As mentioned ear lier, an iterative process was conducted along with disaggregate
based calibration to attai n a balance in objectives. In this iterative process, va ryi ng numbers of
parameters are exposed and the best fitness function values from each calibration are noted. The
final parameter values and best fitness va lues are shown in table 9.3 . T hi s process also helps in
determining the degree of freedom of the objective function.
In the first calibration run, all of the five parameters were exposed the res ulting speed-
flow graph provided good coverage over the field speed flow graph. However, the CC4 and CC5
values were not in confirmation with the observations from the NGSIM data set. It was also
determined that CC4 value of -2.50 m/s resulted in an approximate shockwave speed of20 mi/hr,
which is much higher than real-world values of approximately 12 mi/hr (Lu and Skabardonis,
2006). Tests were then conducted to determine CC4 values that resulted in shockwaves speed
that are closer to real-world values. The possible solutions ranged from CC4 values ranging form
-0.35 to -0.70 m/s. A secOnd calibration run was conducted with four calibration parameters
[CC1, CC2, CC3, and CC5], the res ulting CC5 value was close to 2 m/s, which was i1ot in
confirmation with symmetric speed osc illation observation in the NGSiM dataset. In addition,
the resultant CC 1 value was similar to the first calibration run. The final fitness function value
was actually slightly lower than the first calibration run . The speed flow graph from the
calibration process can be seen in figure 9.14(a).
In the third calibration run, the significance ofCC3 parameter on the fitness function was
tested. A calibration run was conducted using three parameters [CC1 , CC2, and CC5]. The
resulting parameter values are shown in table 9.3 and speed flow graph is shown in figure
9.13(b). There was no reduction in the final fitness function value; however, slightly lower CC1
value and slightly higher CC2 values were suggested. This can be due to the combination effect
discussed in chapter 7. In addition, CC5 value suggested by the calibration was still high at 2.30
m/s.
In the fourth calibration run, CC4 and CC5 values from disaggregate data based
calibration were employed. In addition, only two parameters CC 1 and CC2 were used. The
resultant CCl and CC2 values were close to the values suggested in previous runs. However, the
final fitness value increased by 5% from previous runs. The final speed flow graph after
calibration is shown in figure 9.13(c). Although the final fitness function value increased by 5%,
the resultant balance in disaggregate and aggregate data based objectives was more important.

97
A fifth calibration run using two parameters CC I and CC2 was conducted at CC4 = -0.35 and
CC5 = 0.35. T here was no change in the final fitness value, but hi gher CCJ and lower CC2
values were suggested. Again, this mi ght be due to the combination effect di sc ussed in chapter 7.
However the disaggregate data based calibration us ing CC4 = -0.35 and CC5 = 0.35 m/s, were
not in close confirmation with observations from the NGSIM dataset.

Table 9.3 Speed Flow Based Calibration with Varying Number of Parameters
Calibration Parameters Fitness Fixed Parameters
CC! CC2 CC3 CC4 CC5 Value CC4 CC5 Others
1.09 10.59 -7.91 -2.50 0.64 193 default
1.1 7.97 -12.3 2 192 0.7 default
1.0377 9.01 2.31 193 0.7 default
1.16 9.9 204 0.7 0.7 default
1.29 5.65 204 0.35 0.35 default

Speed-flow Calibration of US101 NB using CCl, CC2, CC3, and CC5 Speed- Fl ow Calibration of US101 NB using CC I ; CC2, and CCS
CC4 =-0.70 CC4 = -0.70

Flow (veh/hr) Flow {vehlhr)

(a) (b)
Speed-Flow Calibration of USIOINB using CCt and CC2 Speed Flow Calibration of US101 NB us ing CCI and CC2
CC4 =-0.70 and CC5 =0.70 CC4 = -0.35 and CCS =0.35

60

-=-55 ..... .
5
fso
~ 45

Flow (vehlhr) Flow (vehlhr)

(c) (d)
Figure 9.13 Speed flow based calibration ofUS101 NB with varying numbers of parameters.

98
9.5 I SIMPLIFIED CALIBRATION METHOD FOR THE PRACTITIONERS
After an iterative disaggregate and aggregate based calibration of the simulation model , it was
determined that CC 1 and CC2 parameters along with CC4 and CC5 values of[-0.70 and 0.70
m/s] resulted in a good balance in calibration objectives . In order to assist practitioners in
calibrating a simulation model to capacity, flow contours were developed at varying CC I and
CC2 values. These flow contours were developed using the test network, which was employed in
chapter 8 and shown to be applicable in these cases. The flow contours were deve loped at two
levels: [1] CC4 = -0.70 and CC5 = 0.70 and [2] CC4 = -0.35 and CC5 = 0.35. The graphs for
option 1 are shown in figures 9.14 (a)-(d), and graphs for option 2 are shown in 9.15 (a)-(d). The
following graphs were developed for 0% trucks and FFS = 70 milhr. However, graphs for other
combinations are fairly simple to develop.

99
 Contour oi Maximum 5 min Flow
CC4 =-0_70 and CC5 = 0_70
16

14

12

10
s
N
0
0
8

0.6 OJ 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
CC1{s)

(a)
Contour of Maximum 5 m in Flow Sustained for 15 mins
CC4 = -0.70 and CCS = 0.70 .
16

14

12

10
s
N
0
0 8

0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
CC1(s)

(b)

100
 Contour of Maximum 5 min Flow Sustained for 30 mins
CC4 =-0.70 and CC5 =0.70

14

12

10
E
N
0
0 8

0.6 0.7 0.8 0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
CC1(s)

(c)
Contour of Maximum 5 m in Flow Sustained for 60 m ins
=
CC4 -0.70 and CC5 = 0.70

10
.s
N
0
0 8

0.6 0.7 0.8 0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
CC1(s)

(d)
Figure 9.14 Flow Contours of maximum 5 m in flows sustained for varying amounts of time.
[CC4 = -0.70 and CCS = 0.70]

101
 Contour of Maximum 5 min Flow
CC4 = -0.35 and CC5 = 0.35
16

14

12

10
'E'
N
0
0
8

0.6 0.7 0.8 0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7 :1.8 1.9 2
CC1(s)

Contour of Maximum 5 min Flow Sustained for 15 mins

CC4= -0.35 and CC5 '= 0.35
16

14

12

10
g
N
0
0 8

0.6 0.7 0.8 0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
CC1(s)

102
 Cont our of Maximum 5 m in Flow Sustained for 30 rnins
CC4 =- 0.35 and CC5 = 0.35
16

14

12

10
'E'
N
0
(.) 8

0.6 0.7 0.8 0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
CC1(s)

Contour of Maximum 5 rnin Flow Sustained for 60 mins

CC4 = -0.35 and CC5 =0.35

~ 10
.E.
N
(.)
0 8

0.6 0.7 0.8 0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1_8 1.9 2
CC1(s)

Figure 9.15 Flow Contours of maximum 5 min flows sustained for varying amounts oftime.
[CC4 = -0.35 anfCC5 = 0.35]

103
The combination effect of CC l and CC 2 on maximum flow , maximum flow sustained for l5
minutes and maximum flow sustai ned for 30 minutes can be seen in figures 9. 14 and 9.15. From
the figures it can be seen that CC2 has an increased significance as the sustained time period
increases. In other words, QDF is more sens itive to CC2 than PQF. Similar flow values can be
attained for different combinations of CC I and CC2, but it is difficult to say whi ch one is correct.
For default values [CC 1 = 0.9, CC2 = 4, CC4 =-0.35 and CC5 = 0.35 ], maximum 5 min flow is
approx. 2600 veh/hr; max 15 and 30 min sustained flows are approx. 2400 veh/hr; max 60 min
sustained flow is approx . 23 00 veh/hr.

The practitioners can use the followin g flow contours to calibrate their simulation models. For
example, let's say following is hypothetical field data
1. max. 5 min flow is 2400 veh/hr
2. max. 15 min sustained flow is 2300 vei1/hr
3. max. 30 min sustained flow is 2300 veh/hr
4. max. 60 min sustained flow is 2200 veh/hr.

One possible solution can be derived from the four figures 9.15 (a)-(d) as CC I = 1.1 and CC 2 =
6, but certainly, other solutions are possible.

104
 X CON CLUSION S

10.1 I CONCLUSIONS
A generalized calibration methodology that integrates three levels of representation was
devdoped in this research. The three levels of representation in calibration methodology better
define the calibration process. The representations also provide insight into the flow of
information in the calibration process. Based on the case study of a real-world freeway network
in California, it was demonstrated that a calibration method based on speed-flow graphs
performed better than a calibration method based on capacity. The generalized calibration
methodology was used to calibrate the VI SSIM driver behavior mod el. An integrated calibration
methodology based on iterative di saggregate and aggregate data based calibration is used for
calibrating the VISSIM driver behavior models. The disaggregate data based calibration was
performed using NGSIM vehicle trajectory data. The aggregate data calibration methodology
was developed using fundamental traffic flow variables based on volume, speed, and occupancy.
The case study also demonstrated the usefulness . of a pattern recognition based
dissimilarity representation for measuring the degree-of-closeness of graphs. The pattern
recognition based objective integration with evolutionary algorithm also performed very well.
The automation of the graph matching using pattern recognition provides new way of quantifying
the degree-of-closeness of graphs. It solves various consistency problems that occur in eye-
balling techniques previously applied by researchers. The automated graph matching concept can
be extended to higher dimensions based on similar principles from 20 or 30 graph matching.
Since simulation models cannot be expressed as closed-form equations, many of the optimization
methods are based on multiple evaluations of the simulation model. This is where a pattern
recognition based measure can have the greatest impact by automating the matching process of
complex data set. The other advantages of pattern recognition methods are that they are simple,
flexible, efficient, and robust. These are some of the important qualities that are required for

105
implementation in the real -wo rld . The simplicity of the model he lps practitioners deve lop and
dep loy pattern recognition based ca librat ion that require minimal to no hum an intervention.
The patte rn recognition based objective function is flexible and effic ient. The
dissimilarity meas ures deve loped in chapter 5 can be applied in a variety of traffic flow
representations. The main adva ntage of dissimilarity meas ures is that there are no s pecific
assumptions being made about the system, data, population, or the model. The meas ures are
equally applicabl e in other model s in a non-traffic arena. The dissimilarity measures can also be
applied to macroscopic or mesoscopic simulation mode ls, without any major modifications.
Finally, user ass istance is provided to practitioners by suggesting calibration parameter values to
attain a certain capacity
User assistance Js also provided in this research through theoretical discussion on
Wiedemann 74 and 99 models . Calibration methodologies based on speed-flow graphs can be
directly applied by practitioners using simple pattern recognition based objectives . In case such
information is unavailable, the practitioners can use contour plots of capacity or maximum flow
rates defined on CC I and CC2 plane to calibrate their simulation model. The flow contours are
developed for varying values of CC I and CC2. The practitioners with information about
maximum flow and sustained time in the field can select CC! and CC2 values to calibrate their
simulation model , hence saving valuable time.

106
REFERENCES
l. Aycin, M. F., and B. F. Benekohal. ( 1998). Linear Acceleration Car-Fo ll owing Model
Development and Validation. In Transportation Research Record: Journal of the
Transportation Research Board, No. 1644, TRB, National Research Council, Washington,
D.C., pp. 10-19.
2. Banks, J. H. (1991). The Two-Capacity Phenomenon: Some Theoretical Iss ues. In
Transportation Research Record: Journal of th e transportation Research Board, No. 1320,
TRB. National Research Council, Washington D. C., pp . 234-241.
3. Bayarri, M. J., J. 0 . Berger, G. Molina, N. M. Rouphail, and J.Sacks. (2004). Assessing
Uncertainties in Traffic Simulation: A Key Component in Model Calibration and Validation.
In Transportation Research Record: Journal of the Transportation Research Board, No.
1876, TRB, National Research Council, Washington, D.C., 2004, pp. 32-40.
4. Benekohal, R. F., and J. Treiterer. ( 1988). CARSIM: Car-Following Model for Simulation of
Traffic in Normal and Stop-and-Go Conditions. In Transportation Research Record: Journal
ofthe Transportation Reseaich Board, No. 1194, TRB, National Research Council,
Washington, D.C., pp. 99-111.
5. Brockfeld, E., R.D. Ki.ihne, and P. Wagner. (2004). Calibration and Validation of
Microscopic Traffic Flow Models . In Transportation Research Record: Journal of the
Transportation Research Board No. 1876, TRB, National Research Council, Washington,
D.C., 2004, pp. 62-70.
6. Brilon, W., J. Geistefeldt, and H. Zurlinden. Implementing the Concept of Reliability for
Highway Capacity Analysis. Presented at 87th Annual Meeting of the Transportation
Research Board, Washington, D.C., 2007.
7. Cheu, R. L., W. W. Recker, and S. G. Ritchie. (1994). Calibration ofiNTRAS for Simulation
of 30-sec Loop Detector Output. In Transportation Research Record: Journal of the
Transportation Research Board, No.1457, TRB, National Research Council, Washington,
D.C., pp. 208-215 .
8. Cheu, R. L., X. Jin, K. C. Ng, Y. L. Ng, and D. Srinivasan. (1998). Calibration ofFRESIM
for Singapore Expressway Using Genetic Algorithm. In Journal of Transportation
Engineering, Vol. 124, No. 6, Nov-Dec 1998.
9. Dixon, P. M. (1997). Calibrating and validating TRAF-NETSIM Model of Single-Point
Urban Interchange .. In Transportation Research Record: Journal of the Transportation
Research Board, No . 1591, TRB, National Research Council, Washington, D.C., pp. 38-44.

107
10. Dowling, R., A. Skabardonis, J. Ha lki as, G. McHale, and G. Zammi t. (2004). Guide lines for
Calibration ofMicrosimulation Models: Framework and App li cations. In Transportation
Research Record: Journal of th e Transportation Research Board, No. 18 76. TRB , National
Research Council , Was hington. D. C. . pp. 1- 9.
11. Duda, R. 0., and P. E. Hart. ( 1973). Pattern Class ifi cation and Scene Analys is. A Wiley-
Interscience Publication, John Wiley and Sons, New York.
12. Drew, D. R. (1968). Traffic Flow Th eory and Control. McGraw-Hill Book Company, New
York.
13. Federal Highway Administration. (2004). Traffic Analysis Toolbox Volume III: Guidelines
for Applying Traffic Microsimulation Software. FHWA-HRT-04-040. Washington D. C.
14. Fellendorf, M. and P. Votiisch. (2001). Validation ofthe Microscopic Traffic Flow Model
VISSIM in Different Real-World Situations. Presented at 80th Annual Meeting of the
Transportation Research Board, Was hington, D. C.
15. Gardes, Y., A. D. May, J. Dahlgren, and A. Skarbardonis. (2002). Bay Area Simulation and .
Ramp Metering Study. UCB-ITS-PRR_2002-6, California PATH Research Repoti.
16. Gomes, G., A. May, and R. Horowitz. (2004). Congested Freeway Microsimulation Model
Using VISSIM. In Transportation Research Record. Journal of the Transportation Research
Board, No . 1876, TRB, National Research Council, Washington, D.C., pp. 71 - 81.
17. Highway Capacity Manual. (2000). TRB, National research Council, Washington, D. C.
18. Hourdakis, J., P. G. Michalopoulos, and J. Kottommannil. (2003). Practical Procedure for
Calibrating Microscopic Traffic Simulation Models. In Transportation Research Record:
Journal ofthe Transportation Research Board, No. 1852, TRB, National Research Council,
Washington, D.C., pp. 130- 139.
19. Jha, M. , G. Gopalan, A. Garms, B. P. Mahanti, T. Toledo, and M.E. Ben-Akiva. (2004).
Development and Calibration of a Large-Scale Microscopic Traffic Simulation Model. In
Transportation Research Record: Journ al ofthe Transportation Research Board, No. 1876,
TRB, National Research Council, Washington, D.C., 2004, pp. 121-131.
20. Kim, K., and L. R. Rilett. Genetic-Algorithm-Based Approach for Calibrating Micioscopic
Simulation Model s. Proc. , IEEE Intelligent Transportation Systems Conference, IEEE,
Oakland, Calif., 2001 , pp. 698-704.
21. Kim, S. J. , W. Kim, and L. R. Rilett. (2005). Calibration of Microsimulation Models Using
Nonparametric Statistical Techniques. In Transportation Research Record: Journal ofthe
Transportation Research Board, No. 1935, Transportation Research Board of the National
Academies, Washington, D.C., 2005, pp. 111-119.

108
22 . Ma, T., and B. Abdulbai. (2002). Ge net ic Algo rithm -Based Optimi zation Approach and
Ge neric Tool Fo r Ca librati ng Traffic M icros im ulat ion Parameters. In Transportation
Research Record: Journal of the Transportation Research Board, No . 1800, T RB , National
Research Council , Washington, D.C. , pp.6- 16.
23. May, A. D. ( 1990). Traffic Flow Fundamentals. Prentice-Hall Inc. , New Jersey.
24. Michales, R. M. (1965) . Perceptual Factors in Car Following. Proceedings of 2"d
Intrernational Symposium on Theory of Traffic Flow.
25 . NGSIM. (2007). Home of the Next Generation Simulation Community. Federal Highway
Administration. Washington D.C., http: //www.ngsim.fhwa.dot.gov/. Accessed June, 2007.
26. Ngoduy, D., S. P. Hoogendoorn, and H. J. V. Zuylen. (2004). Comparison ofNumerical
Schemes for Macroscopic Traffic Flow Models. In Transportation Research Record: Journal
ofthe Transportation Research Board, No. 1876, TRB, National Research Council,
Washington, D.C., pp. 52- 61 .
27. Park, B. and H. Qi. (2005). Development and Evaluation of a Procedure foi the Calibration of
Simulation Models. In Transportation Research Record: Jownal of the Transportation
Research Board, No. 1934, Transportation Research Board of the National Academies,
Washington, D.C. , 2005 , pp. 208-217.
28. PeMS. (2007) Freeway Pe1jormance Measurement System. California Partners for Advanced
Transit and Highways. University of California Berkeley and California Depa1iment of
TranspOiiation. http://PeMS.EECS.Berkeley.EDU. Accessed June, 2007.
29. Pekalska, E., and R. P. W. Duin.(2005). The Dissimilarity Representation for Pattern
Recognition. World Scientific Publishing Co. Pte. Ltd., Singapore.
30. PTV (2007). Vissim User Manual- V4.30. Planung Transport Verkehr AG, K.arlsruhe,
Germany, March 2007.
31. Rakha, H., M. Van Arde, L. Bollberg, and X. Huang. (1998). Construction and calibration of
a Large-scale Microsimulation Model of the Salt Lake Area. In Transportation Research
Record: Journal of the Transportation Research Board, No. 1644, TRB, National Research
Council, Washington, D.C., pp. 93-102.
32. Rakha H., and B. Crowther. (2002). Comparison ofGreenshields, Pipes, and Van Aerde Car-
Following and Traffic Stream Models. In Transportation Research Record: Journal of the
Transportation Research Board, No. 1802, TRB, National Research Council, Washington,
D.C. , pp. 248-253.
33. Rillet, L. R., K. 0. Kim, and B. raney. (2001). Comparison of Low-Fidelity TRANSIMS and
High-Fidelity CORSIM Highway Simulation Models with Intelligent Transportation Systems

109
 Data. In Transportation Research Record Journal of the Transportation Research Board,
No. 1739, TRB, National Research Co uncil, Washington , D.C. , pp. 1-8 .
34. Schultz, G. G. , and L. R. Rilett. (2004) Analysis of Distribution and Calibration ofCar-
Following Sensitivity Parameters in Microscopic Traffic Simulation Models. In
Transportation Research Record: Journal of the Tramportation Research Board, No. 1876,
TRB, National Research Council, Washington, D.C., pp. 41-51.
35. Todosiev, E. P. ( 1963 ). The action-point model of the driver-vehicle-system. The Ohio State
University, Report 202 A-3.
36. Toledo, T. M. E. Ben-Akiva, D. Darda, M. Jha, and H. N . Koutsopoulos . (2004). Calibration
of Microscopic Traffic Simulation Models with Aggregate Data. In Transportation Research
Record: Journal of the Transportation Research Board, No. 1876, TRB, National Research
Council, Washington, D.C., pp. 10- 19.
37. Toledo, T., H. N . Koutsopoulos, A. Davol, M. E. Ben-Akiva, W. Burghout, I. Andreasson, T .
Johansson, and C. Lundin. (2004). Calibration and validation of Microscopic Traffic
Simulation Tools: Stockholm Case Study. In Transportation Research Record: Journal of the
Transportation Research Board, No. 1831, TRB, National Research Council; Washington,
D.C., pp. 65-76.
38. Wiedemann, R.(1974). Simulation des StrafJenverkehrsjlusses (in German). Schtiftenreihe
des Instituts fur Verkehrswesen der Universitat Karlsri.the, Heft 8, Germany.
39. Wiedemann, R. ( 1991 ). Modelling of RTI-Elements on Mu lti-Lane Roads . In Advanced
Telematics in Road Transport, Vol. ll, Elsevier, Amsterdam, Holland, pp. I 001-1019.
40. Zarean, M., and Z. A. Nemeth. (1988). WEA VSIM: A Microscopic Simulation Model of
Freeway Weaving Sections. In Transportation Research Record: Journal of the
Transportation Research Board, No. 1194, TRB, National Research Council, Washington,
D.C., pp. 48-54.

110