Abstract

The paper under the name “ A dynamic traffic assignment model for highly congested urban
networks” written by Moshe E. Ben-Akiva , Song Gao, Zheng Wei , Yang Wen proposes a study of
enhancing and calibrating DynaMIT-P, a mesoscopic traffic simulation system for a 18-square-mile
area in Beijing, which included 1698 nodes and 3180 directed links. The model simulated
approximately 630,000 vehicles over 4 hours during the morning peak, using data from 2927 origin-
destination pairs. This paper presents these following sections: the first is the introduction, the
second part is about the modelling challenges and solutions, the third is a case study and finally the
fifth gave conclusion and future directions.

The introduction addresses the complexities and solutions in modelling highly congested urban traffic
networks by taking Beijing, China as an example. The study supports the call for better performing
Dynamic Traffic Assignment (DTA) models that can accurately capture dynamic traffic movements
and demand management in complex urban settings. The paper highlights a sub-problem such as
dealing with huge network scale, multi-directional lanes (sides) at close interchanges and severe
congestion/problem of significant effect from non-motorized transport. The first calibration of a DTA
model for Beijing failed due to impossible queues and wrong sensor data. These include the use of
Path-size Logit (PSL) model to obtaining a more consistent route choice output and an upgraded
lane-group queue in which combine making simulation results closer accurate with calibrations. As
these models embed simulation-based DTAs and transfer existing knowledge, it is argued in the
paper that one main strength of them compared to analytical traffic models lies within their detailed
replication of dynamic's on road network level enabling evaluation with respect to both operational
strategies like individualized real-time traffic information or dynamic tolling. The paper also covers
recent developments in DTA models, with improved data structures and algorithms as well as
distributed computing which should make more realistic global behaviour accessible at higher levels
of computational efficiency. In order to adequately replicate congestion; model hybrid intersections
and short links; deal with interactions between autos, bicycle, pedestrian traffic-calibration clearly
requires historical movement data.

The second section of the paper is about modelling challenges and solutions. Route choice models
are often used in Dynamic Traffic Assignment (DTA) models to convert origin-destination flows into
path and link flows. The study shows that the Multinomial Logit (MNL) model is not perfect as it
assumes independent and identically distributed error terms, resulting in unreasonable traffic jam on
urban network with overlapping paths. Specifically, MNL models for Beijing network overestimate
estimating freeway usage which led to great traffic congestion. As a remedy, this research
investigates alternative approaches like the C-Logit model and Path-size Logit model (PSL). PSL
introduces an amendment which measures the size of paths so that it becomes possible to forecast
people’s route choices more accurately when many courses are superimposed. When PSL was
introduced, there was significant decrease in gridlock and better representation of reality during
simulation experiments within Beijing network. In DTA models, roads are typically represented as
links (connected by nodes) Capacity differences due to geometric changes such as lane modifications
can be modelled downstream through link segmentation. This usually consists of a high-speed "loop"
portion and then a part with slow traffic which queues up. Realistic congestion estimation requires
accurate queuing models. Although simple and calculating queuing methodology based on segments
seems easy, to some extent such models that permit queue formation over multiple lanes
uncontrollably may generate misjudgement of traffic movement within the context as mentioned

1
high complexity intersections. Queues forming simultaneously across all incoming branches in
uncontrolled way causes errors in point for representation. The left lane effect and overpass
phenomenon were mitigated by developing more specific approaches for treating lanes of various
functions (Left Turn, Through or Right Turn) as a 'lane group', which had been prohibited to handle
such situations appropriately. This method eliminates even the speculation of unreasonable
congestion and bottlenecks by making queuing dynamics orthogonal to real lane use, with an
example shown in study performed for Beijing network. Furthermore, complex networks with short
links, like those in Beijing present problems for traffic simulations. Previous studies indicated extreme
congestion on these links created through artificially reduced vehicle speeds and avoidable queuing.
These problems also help to emphasize the limitations of mesoscopic supply models, especially on
short links. To solve these issues study discusses separately segment and lane inaccuracies. Therefore
lane group capacity problems of traffic simulation models are focused in this study. Use of the speed-
density relationship in mesoscopic models results sometimes into unrealistic congestion, especially
on networks like those found in Beijing. In particular, the capacity increases so fast that vehicles on
these short segments may need to decelerate too much. This was mitigated by modifying the
calibration for minimum speeds in short segments to their observed averages. This section also
includes the disadvantages of traditional traffic simulation models in short road segment parts and
lane group capacity. Mesoscopic models often have unrealistic congestion in short segments
especially when using the speed-density relationship as it is applied with a different coefficient in all
network like that seen in Beijing. For instance, the high rate of increase in density on these short
segments can cause vehicles to decelerate too much. Fixed by calibrating minimum speeds for short
segments to match average observed speed. Moreover, the research takes into account lane group
acceptance capacities problems that may distort congestion over-predictions especially when these
occur in highly congested networks. Acceptance capacity is ignored as long there aren't any queues
in a modified formula which provided better acceptance capacities. In addition, the research embeds
varying output capacities for better capturing shocks related to non-motorized vehicles over time
periods in which congestion is of most concern within urban areas.

The third section, a study case, exposes DynaMIT-P system, a DTA tool, impact analysis of traffic
simulation in an extremely congested network of Beijing used for by the research. DynaMIT-P
comprises a demand model based on a microscopic simulation and supply model based on
mesoscopic simulation, thus creating interfaces to simulate traveller behaviour and vehicle activities
under different traffic situations. To do this the study concentrates mainly on West 2nd Ring Road in
Beijing and the surrounding vicinity, which among the worst in the world for traffic congestion. The
network model has 1698 nodes and 3180 links inside the study area of 18 square miles with traffic
flows of about 630,000 vehicles in the morning peak hours. Actual data originating from Remote
Traffic Microwave Sensors (RTMS) and GPS enabled taxis were used to fine tune the model that had
to emulate network behaviour and response to congestion level very close to real traffic. The study
could encounter certain difficulties owing to Beijing’s peculiarities of traffic short link roads and
possible traffic jam. In calibration for Beijing’s highly congested urban networks, DynaMIT-P needed
to fine-tune 69093 parameters of which 46832 are OD flows and speed-density parameters for 3180
segments. A calibration was performed using the Simultaneous Perturbation Stochastic
Approximation (SPSA) algorithm which can be formulated into a constrained minimization problem.
The results of the model were compared with actual counts and travel timings and the values
obtained were tested for RMSE and RMSN for the potential of accuracy. The calibration process was
able to fit the high congestion levels of the network, hence the improvement in the objective
function value as well as the stabilization of the RMSN which depicted the model’s fitness to simulate
traffic in Beijing as witnessed during data collection. That way, once tuned, the DynaMIT-P DTA

2
system can be used to assess traffic management scenarios. This study explains its application by
evaluating the effectiveness of Beijing‘s post 2008 Olympics ‘rotating no driving day’ restriction. This
policy allows private car use only one weekday in a week according to the license plate numbers; this
significantly minimizes the demand. The comparison is made between actual traffic and the
corresponding fully simulated restriction. Main findings consist in its demand decrease by 20%,
leading to a drop in its impact by 8%. 4 percent reduction in vehicles that get to the destinations as
well as a 17 percent cut in the stream of people accessing their places of work. It will be reduced on
an average by 9% of the total time being spent on travel. Further, the queuing time of links, which
had congestion, was reduced to half in this case. So, it becomes clear that similar limitations can
significantly increase road availability and decrease congestion in the short term.

The fourth and final part concludes the study and addresses future direction. It enhances traffic
simulation in congested urban areas by integrating key features into DynaMIT-P: An advanced route
choice model, the explicit representations in the link lane groups, short link treatment, and dynamic
capacities of non-motorized traffic. When applied to Beijing, it has been claimed that the model was
calibrated using traffic sensor and floating car data to show that the author’s management strategy
can be efficiently evaluated by the model. As for the improvement in the future, more attention will
be paid on the refinement of the route choice model by integrating the vehicle trajectory and the
elaboration by the way of network coding and surveillance data.