Abstract

A psycho-physical vehicle-following model tries to capture both the physical and human
components of congested-traffic simulations. These models determine the reactions of the
vehicle driver depending on the vehicle's state. The state of a vehicle can be determined
by the distance and by the difference in speed, in comparison to the leading vehicle. If the
vehicle state changes, the driver must react. The reaction of the driver is to calculate a
new value for acceleration. There are two different classical methods of time advance to
calculate the state changes. This paper examines the effect of these methods on
simulation run times and simulation results. Empirical experiments were done on a SLX-
based simulation model

Introduction

Traffic simulation has been a wide application area for modeling and simulation for many years.
Both analytical and simulation models are used depending on the aims of traffic modeling.
Simulation models have been frequently applied in the area of urban street traffic. Existing
modeling approaches can be classified depending on the level of detail for the simulated process.
Macroscopic approaches describe the traffic flow in the entirety of all vehicles (Kuehne 1995).
Significant parameters are traffic density, flow of traffic and so on. The main application area for
this kind of approach is the simulation of wide-area traffic systems like highway networks and
interregional road networks. Microscopic approaches are characterized by the description of
single vehicles and relations between them (Lieberman 1991). Models from this class show the
behavior of single traffic participants. Valid traffic rules and regulations form the basis for the
applied behavior rules. The traffic rules define, for example, the maximum of speed and the right
of way. The behavior rules contain additional strategies for controlling speed and acceleration.
Currently, microscopic models are used to simulate urban traffic processes on single intersections
or on networks of urban intersections. For many years analytical models have been used in the
field of microscopic traffic modeling. The application of analytical models is, however, subject to
certain conditions such as homogeneity of the arrival process Rough estimates of the performance
parameters can be done with this model class Intelligent load-dependent control approaches for
traffic lights are used by traffic engineers to manage the increasing urban volume of traffic (Hoyer
1993). Simulation models are necessary tools for evaluating the quality of load-dependent controls
in project phases. Traffic engineers demand from the simulation side:
The needed level of detail is focused on modelling the vehicle-following process and rules for
resolving conflicts between vehicles, like who has the right of way Queuing models approaches are
widely used for modelling vehicle- following processes.

Queuing models (Schlothauer, Schulze 1996) describe parts of a road as single- or multiple-server
systems with queue capacity of zero. Arriving entities (vehicles) can be served only if at least one
necessary server is idle. The service time for a new entity depends on the speed of the preceding
vehicle. Passing the predecessors is not allowed. The simple structure of queuing models does not
permit modelling the vehicle-acceleration process. The speed of a vehicle is used as a discrete
parameter in the model. The use of psycho-physical vehicle-following models leads to a higher
level of model accuracy.

Psycho-physical vehicle-following models determine reactions of vehicles in traffic processes with

regard to changing the acceleration of vehicles. Reactions happen if differences in speed and
distance to the predecessor occur (Wiedemann 1974). Different simulators have been developed
and implemented, based on the ideas of Wiedemann. A new SLX-based (Heriksen 1996) psycho-
physical vehicle-following simulation model has been developed and integrated within a new
microscopic traffic simulator. Development has been done at the University of Magdeburg (Fliess
1996). Vehicles are the main objects in this modeling approach. The class of vehicles contains the
following default attributes:

There are only discrete values for the current state attribute. The state of a vehicle describes six
different behavioral patterns of vehicles. The state will be changed if defined limits for distances to
the predecessor or differences in speed are crossed. The acceleration of the vehicle will be newly
recalculated and shall then be constant until the driver has to react to new traffic conditions. Table
1 lists the different limits for the distance to the predecessor (leading vehicle). ILimits for distances
depend on the current speed of the vehicle and a vehicle specific time gap.

where A denotes the limit for the desired distance, PP T denotes the individual desired time gap, v,
denotes the current speed of vehicle i and Do denotes the distance if two vehicles are standing
one after the other. Limits for differences in speed to the leading vehicle can be classified into two
cases. These cases are described in Table 2
where Ar denotes the distance to the leading vehicle and k.k,k denote special constants.

The different limits used can be shown in an observation-decision diagram. Difference in speed
scales the abscissa and ordinate of the diagram. Such a diagram shows the possible states and
crossings over the states for vehicles. Figure 1 presents an observation-decision diagram with
possible values. The vehicle states used are presented in Table 3

A new value for the acceleration attribute must be calculated if the vehicle has changed its state.
For example, the acceleration in the state approximating will calculated by

where Ar, denotes the distance between the leading and following vehicle, v, denotes the current
speed of the following vehicle, and a, is the acceleration of the leading vehicle.

for the following vehicle until the modeling time reaches 100 sec. The values used for the vehicle
attributes are presented in Table 4

3 IMPLEMENTING TIME ADVANCES

There are two methods of implementing time advances in simulation models. These methods are
called the time- sliced or the event-oriented methods. Using the time- sliced-oriented approach,
the attribute values of vehicles will be updated at constant time slices. At a constant rate the
simulator checks for all vehicles if a limit has been crossed and the vehicle must selected value in
existing traffic simulators for the time slice is often 1.0 sec. This value was selected according get a
new state. The to the mean reaction time of a driver. The time-slice oriented method can be
implemented very easily
Using the event-oriented method, the simulator has to compute the time when a vehicle will
change its state. The time for this event can be calculated from the intersection of vehicle's
trajectory with corresponding limits. The necessary condition is: There is no change in the state of
the leading vehicle Determination of intersections between the trajectory and limits for the
distance can be reduced to an intersection between a parabola and a line. To do this, the
simulator has to solve a second-degree equation. The intersection between the trajectory and
limits for differences in speed will be reduced to solving a fourth- degree equation. Suitable
approximation methods like.

Newton iteration can be used for determining the event time. These events are called expected
events. Table 5 presents a list of state events for the following vehicle using the attribute values of
Table 4.

Using the event oriented method, the calculation for new attribute values of involved vehicles has
been done eight times. Ideal conditions lead to these small number of calculations: There are only
two vehicles in the system and the state of the leading vehicle is constant. Using the time sliced
approach with a time slice of one sec., the calculation of new attributes has been done 22 times
for both vehicles.

Realistic traffic processes are characterized by a larger number of vehicles and frequent state
changes of the leading vehicle. So the simulator has to process both expected and unexpected
events. Unexpected events occur if the leading vehicle changes its state and the following vehicle
has to react to the new values of attributes of the leading vehicle. A reaction of the following
vehicle will take place after passing a reaction time.

State changes of the leading vehicle have several consequences. The simulator has to proceed
with the following steps:

Steps 1 and 4 intersection between the trajectory of the leading vehicle with corresponding limits.
The approximation of the intersection between the trajectory and the limits for differences in
speed is a compute-intensive process. Steps 2, 3 and 5 concern operations wth the event list of the
simulator. These actions involve basic operations for removing and inserting events. The run times
for these operations depend on the number of events in the event list, the internal scheduling
algorithms and the implementation level of these algorithms (Schulze and Preuß 1997) The
number of events will be determined by the number of moving vehicles in the traffic system. Event
lists of urban traffic simulators contain not only events for vehicles. For example, they contain
events for street cars, bicycles, pedestrians and traffic lights. An average content of more than 200
events is typical for urban traffic systems.

4 REMARKS ON SIMULATION RUN TIMES Run times for simulation models are affects by different
factors. On one hand, the use of constantly increasing processor power reduces the run time. On
the other hand, the desired higher level of modeling detail leads to increasing run times for
simulation experiments. A lot of runs have to be executed for constructing acceptably small
confidential intervals. The input data for arriving vehicles are characterized by large dispersion.
Our experience in typical traffic simulations is that the number of necessary runs is often greater
than 20. Results have to describe the steady-state of the traffic system The influence of the
transient phase should be rejected. Experiences from real traffic-simulation projects show that the
length of the transient phase can be as much as 60 minutes of simulated time. This value was used
for urban-traffic systems with three networked intersections. The minimum length of one
simulation run was defined to be 120 simulated minutes. The need for a lot of simulation runs and
the necessary simulation length require fast computer nun time for executing simulation
experiments.

The computer run time for traffic simulation models can be divided into two parts:

Basic demand for static objects and

Additional demand for dynamic objects.

The basic demand includes the time for initialization of the static model components like
paramęters of roads and lanes, or timing for

traffic lights. The additional demand includes the time for:

the run time for slice-oriented or event-oriented methods depends on different factors. The
dependencies are shown in Table 6.

The number of dynamic objects affects both methods of time advance. An increased number of
dynamic objects leads to longer run time because more attribute values of these objects have to
be updated. Reduction of the time slice value results in increased run time in the time-slice
approach The computer time in the event-oriented approach is very strongly influenced by the
algorithm used for event list handling and its implementation (McConmack 1979, Schulze and
Preuß 1997). Computer-time efficient algorithms are necessary for processing a large number of
events. The number of events to process depends on the number of dynamic modeling objects, as
well as the interval of time between two successive events for a dynamic object. Short intervals
between occurrence of events leads to a growth of both the number of state changes and event-
list updates. The frequency of state changes depends on the functions describing the limits and on
the operations during a state change.

5 THE SIMULATION SYSTEM

A driver has to make decisions constantly during his trip about his direction and acceleration. The
selection of the direction is determined by the current position and destination of the driver. For
choosing his acceleration.

the driver has to take the conditions on the road network and other road users into consideration.
The leading vehicle has an essential influence on the acceleration. The simulation system (Fliess
1996) used consists logically of three main components driver, vehicle- following, and lane model.
In addition there are minor components, e.g. for describing pedestrians. The driver model reflects
the behavior of the driver. He is influenced by other road users and the circumstances of the road
network, like topology and lane performance. The vehicle-following model is based on the psycho-
physical vehicle-following model described in section two. The circumstances of the road network
are described in the lane-model. The road network can be characterized for example by the
following elements:

lanes

, traffic

lights

, conflict points

, sources and destinations, and

routing tables.

A special feature of this lane model is that the length of each lane is independent of the length of
the vehicles or dynamic objects. The vehicles can also be on several lanes at the same time. In
many other simulation systems, there are restrictions on the length of lanes.

The essential task of the driver-model is to calculate the new acceleration of the vehicle.
Correction of the acceleration is caused by the influence of the leading vehicle and the
circumstances of the road network. Another special feature of this simulation model is the
dynamic selection of the leading vehicle and a predictive estimation of the gap between two
vehicles in conflict situations The simulation model was implemented in SLX (Henriksen 1996) and
uses the method of the layout- based model generation LBMG (Lorenz and Schulze 1995) for
generating the essential parts of the main component lane model.

6 EXAMPLE An urban road traffic system with three intersections was used for an empirical
investigation. Objectives were investigation of different time-advance methods, in relation to run
time and to simulation output in microscopic traffic simulation models. The investigated
simulation models use the approach of psycho-physical vehicle-following modeling. Required data
for describing the example were taken from real traffic projects. Figure 3 contains a schematic
picture of the investigated road network. It has an extension of approximately 800 meters in the
east-west and 250 meters in the north-south direction

Left intersection 1 is a simple four-way intersection. There is just one lane for each direction.
Middle intersection 2 is a three-way intersection, while intersection 3 is a four-way multiple-lane
intersection Traffic lights are intersection 2 Significant parameters to describe the individual
vehicles and drivers are non-deterministic. Within the at intersection and 3, but not at given limits
for these parameters an uniform distribution is used. Table 7 shows some specific vehicle
parameter for the type 'car

7 EMPIRICAL RESULTS

The objective of the investigation was to determine the influence of the time-advance methods on
the run time and simulation results. The simulation runs executed on a PC with a 120 Mhz Pentium
processor Different experiments were carried out with the example described above. The
simulation model used is described were in section three. The two ime advance methods
described, were applied to the vehicle-following and driver model components. Experiments
varied methods of time advance and the traffic load. The values used for the experimental
parameter method of time advance are:

Event oriented, and

Time-slice oriented with time slices of 1.0, 0.5 and 0.2 sec.

The traffic load varied among three levels. Starting with a basic load (Level 1) the load was
increased by 50 percent. That means:
Level2=1.5*level 1 and

Level3=2.0*level1

Each possible combination of the values of input parameters is called a scenario. The total number
of scenarios is 4*3-12. Twenty simulation runs with different seeds for random number generators
were executed for each scenario

7.2 Influence on simulation results The lost time for a vehicle is a typical result of simulated traffic
systems. This time is the difference between the realised and the ideal travel time for a vehicle,
where the ideal time would be achieved in a system with no other road users. These lost times will
be calculated for every possible path from source to destination. There are 45 different paths in
the example model described above.

The lost time increases with growing traffic load. This trend is independent of the method of time
advance. There are no significant differences in the values of lost time for traffic load level one and
two. Differences occur in load level three. The values for the event-oriented approach and for the
time-slice value equal 0.2 are not characterized by significant differences. The calculated
confidence intervals are overlapping The mean for the lost time increases with increasing time
slice. A growing time slice coverts a higher degree of inaccuracy of model results. Figure 5 presents
90 percent confidence intervals for lost times using traffic load level 3

8 CONCLUSIONS

This paper presents results of investigations on the influence of various time-advance methods in
microscopic urban traffic models. Experiments were focused on simulation run times and results.
The implemented effective algorithms for event-list handling, leads to significantly smaller run
times than the time-slice approach. The use of event-oriented methods can be recommended in
microscopic psycho-physical vehicle- following models. The kind of time-advances approach
influences the accuracy of simulation results. event-oriented approach, including

The simulation result lost time is a random variable. Twenty different simulation runs were
executed to estimate the steady-state mean for this random variable. A 90 percent confidence
interval was also calculated additionally. Table 9 shows the confidential interval for mean lost time
on one path