DTA2012 Symposium: Combining Disaggregate

Route Choice Estimation with Aggregate

Calibration of a Dynamic Traffic Assignment
Model
Moshe Ben-Akiva Song Gao Lu Lu Yang Wen

September 27, 2013

Abstract
Dynamic Traffic Assignment (DTA) models are important decision sup-
port tools for transportation planning and real-time traffic management. One
of the biggest obstacles of applying DTA in large-scale networks is the cal-
ibration of model parameters, which is essential for the realistic replication
of the traffic condition. This paper proposes a methodology for the simul-
taneous demand-supply DTA calibration based on both aggregate measure-
ments and disaggregate route choice observations to improve the calibration
accuracy. The calibration problem is formulated as a bi-level constrained op-
timization problem and an iterative solution algorithm is proposed. A case
study in a highly congested urban area of Beijing using DynaMIT-P is con-
ducted and the combined calibration method improves the fits to surveillance
data compared to the calibration based on aggregate measurements only.

Edmund K. Turner Professor of Civil and Environmental Engineering, Massachusetts Institute of
Technology, Room 1-181, 77 Massachusetts Avenue, Cambridge, MA 02139. Tel: +1 617-253-5324.
Email: [email protected]

Associate Professor, Department of Civil and Environmental Engineering, University of Mas-
sachusetts, 214C Marston Hall, 130 Natural Resources Rd, Amherst, MA 01003, USA. Tel: +1
413-545-2688. Email: [email protected]

Research Assistant, Department of Civil and Environmental Engineering, Massachusetts In-
stitute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139. Tel: +1 617-842-5797.
Email: [email protected]

Google Inc., 76 9th Avenue, 4th Floor, New York, NY 10011. Email: [email protected]

1
1 Introduction and Literature Overview
With the advance of mobile, sensor, and surveillance technology, high quality
traffic data has become increasingly available. Trajectories from cell phones or
GPS-equipped vehicles, for example, are able to continuously provide more accu-
rate travel time and route choice information for large scale transportation network
than ever before.
The extensive deployment of Intelligent Transportation Systems (ITS) in the
past few years has substantially increased the amount of dynamic traffic data. The
abundance of such information and the advances in computational power have
brought new opportunities and challenges to improve transportation planning and
traffic management.
Dynamic Traffic Assignment (DTA) is one of the many promising areas that
would significantly benefit from the availability of new data. A DTA model in
general consists of an integration of models that can be divided into two major
categories (Florian et al.; 2001; Cascetta; 2001): a set of demand models that
capture the time-dependent flow rates on the paths of the network based on trav-
eler behavior (such as travel mode, route choice, and departure time choice), and a
set of supply models for network loading and moving vehicles. Advanced DTA
models, especially those simulation-based, are capable of modeling drivers be-
haviors (including their response to information), utilizing the dynamic estimated
origin-destination (OD) flow, and capturing the complex interactions between de-
mand and supply. They have been increasingly adopted in transportation planning
(see, e.g., (Ben-Akiva et al.; 2007; Rathi et al.; 2008; Balakrishna et al.; 2008;
Sundaram et al.; 2011; Florian et al.; 2001; Barcelo and Casas; 2006; Ziliaskopou-
los et al.; 2004; Balakrishna et al.; 2009)), and have also been applied by many
in real-time traffic managements, with great potential in providing consistent traf-
fic predictions for various situations even when non-recurrent incidents occur (see,
e.g., Ben-Akiva et al. (1997), Mahmassani (2001), Antoniou (2004), Wen et al.
(2006), and Wen (2009)).
To realistically replicate the real traffic condition, however, lots of parameters
in the DTA model need to be calibrated before using the model on a new network.
The calibration is essentially the process of systematically tuning the input pa-
rameters to ensure a DTA model could generate output that matches the historical
observations. Except for extremely simple networks, a good calibration is usually
a prerequisite for the model to reliably reproduce and predict traffic conditions.
The calibration of a DTA model for a new network is a non-trivial task. It is
arguably the biggest obstacle besides the computational tractability for applying
DTA in large-scale networks. It requires not only a plethora of data over time, but
also methodologies that could effectively combine the data in a coherent way, as the

2
data would often come from various sources and could be sometimes inconsistent
or imperfect.
Researchers have come up with various strategies to calibrate DTA models.
For example, Peeta and Ziliaskopoulos (2001), Antoniou (2004), Balakrishna et al.
(2005), and Balakrishna (2006) have reviewed and summarized many early stud-
ies in the area. Particularly, Balakrishna (2006) provided a comprehensive review
of the subject of calibrations by looking at related topics in three broad classes:
(1) demand-supply calibration of DTA models, (2) estimation of supply models,
and (3) estimation of demand models. He concluded that, in prior research, de-
mand and supply models were calibrated independently (sequentially); in addition,
OD flows and route choice model parameters were estimated sequentially, with the
route choice parameters being estimated through manual line or grid searches. He
proposed a methodology for the simultaneous demand-supply calibration of gen-
eral DTA models, and argued that such approach could lead to better results as it
did not ignore the effect of the interactions between demand and supply models.
The simultaneous demand-supply calibration approach has been the state-of-
the-art of aggregate calibration since then, and it has been adopted and extended
by others. Vaze et al. (2009), for example, extended the work by Balakrishna et al.
(2007) to use multiple sources of data (including link counts and point-to-point
travel times) for the calibration. Their study also found that the joint demand and
supply calibration led to more accurate results than the demand-only calibration.
An important challenge that the existing studies have yet to address is how
to effectively use disaggregate information, such as the trajectories of individual
vehicles, in the context of calibrations of DTA models. At the time when those
studies were done, the quantity and quality of disaggregate data were rarely good
enough to be used directly and make a positive impact in the final calibration result.
Usually, the limited amount of disaggregate data would be converted into aggre-
gate form (e.g., computing average travel time from individual measurements or
summing up the number of vehicles passing through a road segment into counts)
before they could be applied in the existing calibration framework, where they are
typically used to measure the goodness-of-fit of the DTA models output (which is
also converted to aggregate form for comparison) (Ben-Akiva et al.; 2012). Such
coversions are useful in dealing with the noisy and incomplete nature and other
deficiencies of disaggregate data, but they also lead to loss of information and fail
to fully utilize the data. As more and more sources of accurate disaggregate data
become available, a new approach should be adopted to take advantage of them.
In simulation-based DTA models, disaggregate data can be used to estimate
parameters that control the behavior of individual travelers at microscopic level.
Parameters used by the route choice model, for example, are potential beneficia-
ries of such data. Route choice captures travelers preferences in selecting a route

3
from an origin to a destination (OD) in a road network. By itself an interesting re-
search topic, route choice is also an important part of the demand models used by
simulation-based DTA systems. With sufficient disaggregate data, whether from
survey by mail, telephone, and the Internet (Ben-Akiva et al.; 1984; Prato; 2004),
or from the more and more widely used GPS trajectories (Frejinger; 2007; Hou;
2010), route choice parameters can be estimated using discrete choice analysis
(Ben-Akiva and Lerman; 1985; Train; 2003), where a single route is selected from
a set of candidates (i.e., the choice set). In a real network, the number of pos-
sible paths between a pair of OD can be large, and for computational tractabil-
ity researchers may choose to use a smaller subset created by choice set gener-
ation algorithms, including the deterministic algorithms such as link elimination
(Azevedo et al.; 1993), link penalty (de la Barra et al.; 1993), and labeling (Ben-
Akiva et al.; 1984), etc., and stochastic path generation algorithms such as sim-
ulation (Ramming; 2002) and doubly stochastic choice set generation (Bovy and
Fiorenzo-Catalano; 2006).
Once the choice set and the attributes about the alternative routes are available,
a route choice model can be developed to predict how travelers decide which path
to take. The Multinomial Logit (MNL) model is one of the most popular for real
applications due to its attractive features such as a closed-form formula to compute
the probability of choosing a path in the choice set. Its simplifying assumption that
the error terms must be identically and independently distributed, however, limits
its use in networks where overlapping paths are common, and the C-Logit model
(Cascetta et al.; 1996) and Path Size Logit model (Ben-Akiva and Bierlaire; 1999)
are proposed to solve this problem. The latter, for instance, has been successfully
implemented in the DTA model of a congested area in the city of Beijing (Ben-
Akiva et al.; 2012).
Researchers focusing route choice have also developed more sophisticated mod-
els such as Multinomial Probit (Yai et al.; 1997), Error Component model (Bolduc
and Ben-Akiva; 1991), subnetwork (Frejinger and Bierlaire; 2007), sampling of
alternatives (Frejinger et al.; 2009). Gao (2005) developed a routing policy choice
model to capture the inherently uncertain nature of traffic dynamics in a stochastic
time-dependent network. Bierlaire and Frejinger (2008) developed a latent choice
model to directly use network-free data. Fosgerau et al. (2012) proposed a logit
model for the choice among infinitely many route in a network. Due to their com-
plexity, those models have yet to be widely adopted in the context of DTA.
This paper proposes an innovative methodology that takes advantage of state-
of-the-art methodologies in both aggregate DTA calibration and disaggregate route
choice estimation and for the first time integrates them in a consistent framework
to improve the accuracy of the DTA modeling system. The contributions are two-
folded. Methodologically, a bi-level optimization problem is formulated for the

4
combined calibration problem, and an iterative solution algorithm is designed. Em-
pirically, a real life case study is conducted to demonstrate the practicality of the
method in highly congested networks.
In the remainder of the paper, section 2 illustrates the problem formulation and
solution methodology. Section 3 provides a case study in the City of Beijing and
Section 4 concludes.

2 Problem Formulation and Solution Methodologies

2.1 Framework for Combined Route Choice Model Estimation and
DTA Calibration
We extend the framework of simultaneous demand-supply DTA calibration
based on aggregate observations introduced in Balakrishna (2006), and incorporate
the disaggregate route choice observations to improve the calibration accuracy.
Let the time period of interest be divided into intervals h = 1, 2, ..., H. All
variables are indexed by time, and the same variable without the time index rep-
resent a vector of the variables over all time periods. The calibration variables at
the upper level include xn - the vector of OD flows departing from their respective
origins during time interval h, h - the vector of simulation supply model parame-
ters and h - the vector of route choice parameters. Note that even though the route
choice parameters are indexed by time for the sake of notational uniformity, they
are in fact invariant over time of the day, as travel behavior is generally viewed as
stable within a day. The calibration problem is formulated as a bi-level constrained
optimization problem.

5
 Aggregate Calibration and Disaggregate Estimation Problem P
Input : G, xa , a , a , F m , w,
Output : x, ,

min w1 ||F s F m ||2 + w2 ||x xa ||2 + w3 || a ||2 + w4 || a ||2 (1)

x,,
s.t. {F, F s } = DTA(G, x, , , C) (2)
C = P3 (F, G) (3)
xah (1 ) xh xah (1 + ), h {1, . . . , H} (4)
ha (1 ) h ha (1 + ), h {1, . . . , H} (5)
ha (1 ) h ha (1 + ), h {1, . . . , H} (6)
g1 (h ) = 0, . . . , gn (h ) = 0, h {1, . . . , H} (7)
a = arg max LL(I, C, F, ) (8)

The objective function (1) at the upper level is a weighted sum of distances
between time-dependent location-specific simulated aggregate measurements and
field aggregate measurements (e.g., counts, speeds, and link travel times) and dis-
tances between calibrated variable values and their respective a priori values. F s
and F m are the vectors of simulated and observed aggregate measurements respec-
tively, and xa , a , a the vectors of a priori values of OD, supply and route choice
parameters respectively. A priori OD trips are usually obtained from the planning
agency, who usually maintains a regional static planning model based on which the
dynamic ODs can be generated and/or has access to OD surveys. A priori supply
parameters are generated by experience, and a priori route choice parameters are
from the lower level problem. The weights w depend on the relative confidence one
can attribute to the corresponding measurements and a priori values. For example,
if sensors are not reliable, a lower weight might be put on counts. The weights also
depend on the order of magnitude of the measurement in order to avoid a situation
where a parameter with a bigger magnitude or more observations dominates the
others in the fitting function.
Constraint (2) is a simulation-based equilibrium DTA model that takes as in-
puts the network topology G, OD trips x, supply-demand parameters and and
route choice sets C, and generates network performance measures F , such as time-
dependent counts, speeds, and link travel times. Generally a simulation-based DTA
model has stochastic elements, and generates different outputs with different input
random seeds. In this case, F should be viewed as the average over multiple DTA

6
runs. Also note that the simulated aggregate measurements F s in the objective
function can be derived directly from F .
Constraint (3) is a choice set generation model (P3 ) that takes as inputs the
network topology G and performance measures F , and generates a choice set of
alternative routes between each OD pair. An overview of the methodologies to
generate route choice sets will be provided in Section 2.2.1.
Constraints (4) through (6) impose upper and lower bounds on OD trips and
supply/demand parameters. is a fractional number between 0 and 1, which spec-
ifies how far we allow the calibration variables to deviate from their a priori values.
Constraints (7) specifies the physical relationships between the model param-
eters, e.g., the free flow speed cannot be smaller than the minimum speed at jam
density in a speed-density relationship. n is the number of such physical relation-
ship expressions.
The a priori values of route choice parameters are derived from the lower level
route choice estimation problem (8), where the likelihood of observing the disag-
gregate route observations (e.g. from GPS traces) I is maximized. The likelihood
function LL is based on a discrete choice model with route choice sets C and at-
tributes generated from performance measures F . An overview of the estimation
problem will be provided in Section 2.2.3.

2.2 Solution Algorithm

The bi-level calibration/estimation problem will be solved by an iterative pro-
cess that alternates between three sub-problems: the upper and lower level prob-
lems and the choice set generation model. We further define the three problems
separately. Inputs to these problems are divided into two groups: the first (before
the semicolon) contains inputs to the overall problem P , and the other (after the
semicolon) contains inputs generated by the other two sub-problems.

Route Choice Set Generation Problem P1

Input : G; F
Output : C

C = P3 (G, F )

7
 Aggregate Calibration Problem P2
Input : G, xa , a , F m , w, ; C, a
Output : x, , ; F

min w1 ||F s F m ||2 + w2 ||x xa ||2 + w3 || a ||2 + w4 || a ||2

x,,
s.t. {F, F s } = DTA(G, x, , , C)
xah (1 ) xh xah (1 + ), h {1, . . . , H}
ha (1 ) h ha (1 + ), h {1, . . . , H}
ha (1 ) h ha (1 + ), h {1, . . . , H}
g1 (h ) = 0, . . . , gn (h ) = 0, h {1, . . . , H}

Compared to the combined problem P , in the aggregate calibration problem P2

the lower level problem (8) and the choice set generation model (3) are removed.
Route choice sets C and parameters a are instead used as inputs to the problem.

Disaggregate Route Choice Estimation Problem P3

Input : I; F, C
Output : a

max LL(I, C, F, )

Figure 1 gives a flow chart of the process. Note that all variable subscripts are
for iteration numbers, as the variables are already treated as vectors covering all
time periods and the time indices are omitted. Note also that inputs to the overall
problem P are omitted from the diagram to more clearly present the interactions
between the three sub-problems.
To initialize, choice sets C0 are generated based on free flow or static traffic as-
signment link travel times F0 . A base route choice model is assumed with a simple
utility function specification, e.g., one that only includes the travel time as the ex-
planatory variable. The a priori parameter values 0a are assumed based on existing
empirical studies in the literature, rather than estimated from the disaggregate route
choice observations. The aggregate calibration problem P2 is then solved, and the
iteration counter k is set to 1. Outputs from P2 include the calibrated OD trips
xk , supply parameters k , route choice parameters k , and network performance

8
Figure 1: Flow Chart of the Solution Algorithm

9
measures Fk . Choice sets are then updated according to Ck = P1 (G; Fk ). The
disaggregate estimation problem P3 is then solved based on the newly generated
choice sets Ck and performance measures Fk . The estimated route choice param-
eters are then used as the a priori values ka for the aggregate calibration problem
P2 in the next iteration k = k + 1. The iteration continues until a convergence is
reached, usually measured as the relative difference between the time-dependent
link travel times from two consecutive iterations.

2.2.1 Choice Set Generation and Evaluation

The route choice set generation problem (P1 ) takes the network topology and
performance measures as inputs and generates a choice set of alternative paths
between each OD pair. The choice set generation algorithms can be classified into
two groups: deterministic and stochastic.
Deterministic approaches include the link elimination and link penalty algo-
rithms. In the link elimination algorithm, the shortest path is first found between
a pair of OD. Then for each link in the shortest path, the algorithm will remove it
from the network, find a new shortest path, test it for uniqueness and store it in the
choice set if it is unique.
As the link elimination algorithm only removes one link at each iteration, it is
possible that the newly generated path only differs from the original one by a short
detour around the removed link, and paths far from the original one are unlikely
to be generated. The link penalty algorithm could potentially resolve the problem,
where the costs of all links included in the choice set are increased in every iteration
until the costs reach a threshold. After the threshold is reached, link costs will be set
to the normal values and increased in future iterations, which ensures the diversity
of the choice set.
Stochastic approaches include the simulation and doubly stochastic algorithm.
The simulation algorithm determines a distribution for the cost of every link in
the network, for example, normal distribution. For each joint draw of link costs,
a shortest path is generated and incorporated in the choice set if it is unique. The
rationale for this method is that travelers might have perception errors of travel
times (Burrell; 1968; Daganzo and Sheffi; 1977). The number of samples is pre-
determined, and can be adjusted empirically depending on the network settings.
The doubly stochastic algorithm is similar to the simulation algorithm. The
cost functions are specified like utilities and both the parameters and the attributes
are randomly generated, and minimum cost paths are calculated based on these
doubly stochastic generalized costs (Bovy and Fiorenzo-Catalano; 2006).
The evaluation of the generated choice set mainly involves two criteria: cov-
erage and computational time. Define overlap as the degree to which a generated

10
route i matches the observed route.
Li,obs overlap distance between generated and observed paths
Overlapi = =
Lobs distance of observed path
(9)
When complete routes are not observable, e.g., those from GPS traces with gaps
due to the limitation of time resolution, we calculate the overlap by dividing the
overlap distance between generated and actually observed traces by the total length
of the observed traces.
Coverage is the percent of observed routes for which a generated route at a
specified overlap threshold exists. It represents the quality of the choice set gener-
ation algorithm, and high coverage is desired.
For any real life application, the choice set generation problem will be solved
for a large number of OD pairs. Furthermore, in the iterative process introduced in
Figure 1, the choice set generation problem for many OD pairs (P3 ) will be solved
multiple times. Therefore the computational efficiency of the algorithm is also an
important consideration in its evaluation.

2.2.2 Aggregate Calibration Problem: SPSA

The aggregate calibration problem P2 is a minimization problem where the
evaluation of the objective function requires a simulation run of the DTA model.
We use the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm
to solve the problem, which is originally developed by Spall (1998), and later ap-
plied to DTA calibration by Balakrishna (2006). The SPSA algorithm is attractive
for large problems because of its efficient gradient approximation by perturbing all
variables at once. It is also designed for stochastic problems and allows for inputs
corrupted by noise, which is usually the case in simulation-based DTA models.
The SPSA algorithm works in an iterative fashion, where at iteration k, a mov-
ing direction from the current solution (the gradient in a gradient-based method)
is determined. Let be the vector of calibration variables, including the OD trips
x, supply parameters and route choice parameters , and the size of is n. To
calculate the gradient numerically, n evaluations of the objective function need to
be carried out, which are prohibitively expensive for a real life DTA calibration
problem where n is usually very large. The SPSA algorithm does not calculate the
gradient exactly; instead an approximation is calculated by two perturbations of
the parameters. The approximate gradient estimate of the ith calibration variable
at iteration k, denoted as gi (k ), is calculated as follows:

z(k + ck k ) z(k ck k )
gi (k ) = (10)
2cki ki

11
where k ={k1 , k2 , ...kn } is generated based on an appropriate random vari-
able distribution, e.g., the Bernoulli distribution, ck ={ck1 , ck2 , ...ckn } is the size
vector for the random perturbation, is the component-by-component multipli-
cation of two vectors and z() is the objective function value with the calibration
variable vector .
The gradient approximation at iteration k is then g(k )={g1 (k ), g2 (k ), ...gn (k )},
which only requires two computations of the objective function.

2.2.3 Disaggregate Route Choice Estimation: Latent Choice

Disaggregate route choice models are usually developed under the framework
of discrete choice analysis, where a decision maker is assumed to choose from a
choice set (see Section 2.2.1) a route with the maximum utility, which is the sum of
a function of explanatory variables with unknown parameters and a random term.
Parameters of the model are obtained by maximizing the likelihood of observing
the chosen routes, namely, solving the problem P3 .
Sometimes the chosen routes cannot be unambiguously identified, e.g., when
there are large gaps between consecutive GPS readings. In the Beijing case study
that will be introduced in detail in Section 3, the GPS readings are at least one
minute apart during which the vehicle most likely has traversed multiple links.
One solution to this problem is to fill the gaps artificially with shortest paths or
other pre-specified types of paths. However, the complete route obtained with this
method is not necessarily the real chosen route and may lead to biased estimation.
For example, the coefficient of the shortest path dummy in the route choice model
would be artificially boosted if we fill these gaps with shortest paths.
Following Bierlaire and Frejinger (2008), we treat the chosen routes as latent
that are not observable. The estimation problem is then based on the observed GPS
traces, defined as a series of links matched from GPS points that are not necessarily
connected. Therefore each GPS trace might correspond to multiple routes, and the
likelihood of observing a GPS trace r for individual n with a given choice set Cn ,
Pn (r|Cn ) can be written as the sum of likelihoods of observing all paths in the
choice set that are consistent with the trace. Formally,
X
Pn (r|Cn ) = Pn (i|Cn )(r|i). (11)
iCn

i is a route in the choice set, Pn (i|Cn ) is the route choice model that predicts the
probability of choosing route i for individual n out of a choice set Cn , and (r|i)
is a binary variable, which equals one if route i passes through the links in trace g
in the same sequence, and zero otherwise. Figure 2 illustrates a situation where a

12
trace r corresponds to multiple paths, where purple links are observed and red ones
are gaps.
A path-size Logit is used for predicting route choice probability, that is,

exp(ln(PSi ) + Vni )
Pn (i|Cn ) = P , (12)
jCn exp(ln(PSj ) + Vnj )

where Vni is the systematic utility of alternative i for individual n, PSi is the path
size of alternative i that describes the level of overlapping of the alternative with
all other alternatives in the choice set Cn (Ben-Akiva and Bierlaire; 1999). PSi is
equal to 1 if alternative i does not overlap with any other alternatives, and 1/J if it
completely overlaps with J 1 other alternatives. This is a deterministic correction
to the IIA problem of a Logit model in predicting choice probabilities of correlated
alternatives (Ramming; 2002).

Figure 2: The Latent Choice Problem

3 Case Study
3.1 Introduction
In this section we discuss a case study in the City of Beijing using the frame-
work proposed above. We first introduce DynaMIT-P, the DTA model used in this
case study, and the network settings in Section 3.2. We then introduce the data pro-
cessing in Section 3.3. Section 3.4 describes the specific models and algorithms
used in the case study under the combined calibration framework, and presents the
results in comparison with a previous study where only aggregate calibration was
conducted.

13
3.2 DynaMIT-P and Network Settings
DynaMIT-P (Dynamic Network Assignment for the Management of Informa-
tion to Travelers-Planning Version) is a state-of-the-art simulation based DTA sys-
tem (Ben-Akiva et al.; 1997, 2001) designed to evaluate Intelligent Transportation
Systems at the planning level. With a built-in microscopic demand simulator, a
mesoscopic supply simulator, and a learning model to capture the complex inter-
actions between traffic demand and supply, it can predict day-to-day evolution of
travel demand, network conditions and within-day traffic patterns.
DynaMIT-P and its corresponding real-time version have been applied suc-
cessfully in major cities in the US, such as Los Angeles, California (Wen et al.;
2006), Lower Westchester County, New York (Rathi et al.; 2008), and Boston,
Massachusetts (Balakrishna et al.; 2008). The Beijing study is, however, the first
highly congested urban network DynaMIT-P was applied to. Severe congestion
was initially observed in the simulation due to the complexity of network and the
large traffic volume. Several enhancements were then done to DynaMIT-P to solve
this problem, including enhancing the route choice model from a simple Logit
model to a Path-size Logit model, introducing lane groups and variable capacity
to the supply model, and doing special treatments to short links to avoid artificial
gridlock (Ben-Akiva et al.; 2012).
As shown in Figure 3, the Beijing network consists of a series of ring roads
connected by arterial roads with frequent on- and off-ramps. Our study area is the
West 2nd Ring Road network and its northern and southern extensions, the area in-
cluded in the rectangle. The computer representation of this study network consists
of 1,698 nodes connected by 3,129 links. Using results from household surveys, a
historical static demand dataset containing 2,927 origin-destination (OD) pairs are
generated. The simulation time period is from 6:00:00 am to 10:00:00 am.

3.3 Data
The aggregate surveillance data and GPS vehicle trajectory data were obtained
from Beijing Transportation Research Center (BTRC).

3.3.1 Surveillance Data for Aggregate Calibration

We used traffic counts and link travel times from six weekdays during Decem-
ber 2007 between 6am and 10am as the surveillance data for aggregate calibration.
The traffic counts were obtained from Remote Traffic Microwave Sensors (RTMS).
There were 154 RTMS detectors deployed in our study area and 140 of them were
functioning normally to provide traffic flow information continuously. Most of

14
 Figure 3: The study area

them (the triangles shown in Figure 4) were on the expressways. The sensor counts
were aggregated with a 15-minute interval by BTRC.
The link travel times were extracted from Floating Car Data (FCD), which were
obtained from Global Positioning Systems (GPS) in taxis. FCD cover nearly 90%
of all the major roads in Beijing, including arterials and local roads where there
is a lack of sensor counts data. The FCD were provided as averages at 5-minute
intervals.

3.3.2 GPS Data for Route Choice Estimation

GPS devices installed in taxis in Beijing record the positions and speeds of
taxis with a time interval of one minute. BTRC matched the GPS points to certain
positions on links. A GPS trace starts when a taxi service begins and ends when
the passenger gets off the taxi. A vacant taxi drivers route choice behavior is
conceivably significantly different from a regular drivers (e.g., circling to look
for customers), and thus excluded from the analysis. In general a taxi driver has
better spatial knowledge than a regular driver, which might be an important factor
in route choice. We focus on the morning peak where the majority of drivers are
commuters, who conceivably have good knowledge of their commuting routes.
Therefore it is reasonable to use taxi drivers data to represent commuters behavior

15
Figure 4: The distribution of detectors and OD points

16
in this particular study. The proposed methodology is not limited and can be easily
applied to regular drivers data once they are available.
Each GPS entry contains the taxi ID, link ID, time, speed, relative traversed
length on the current link, service number and GPS number, which records the
order of GPS points within the same service.
In total, we obtained two sets of GPS data from BTRC which spanned nine
days. The first set of data includes GPS traces 24 hours/day on two days: April 24,
2008 (Thursday) and April 25, 2008 (Friday). The second set of data includes GPS
traces from 6:00am to 10:00am (which matches the DynaMIT-P simulation time)
on seven days, May 20, 2008 (Tuesday) through May 23, 2008 (Friday) and May
26, 2008 (Monday) through May 28, 2008 (Wednesday).
Table 1 shows the overall statistics of the GPS data.
Table 1: Overall statistics of the GPS data

Number of GPS entries Number of taxis Number of traces

8.9 million 10,412 578,857

As the study area is only a sub-network within the Beijing network, we fil-
tered out outside traces and obtained 11,317 traces that were complete in the study
area. As DynaMIT-P simulations are from 6:00am to 10:00am, and the time de-
pendent travel times used for the route choice model estimation are generated from
DynaMIT-P, we only included traces within this time interval in the estimation.
A large number of the traces had very short travel times. Based on practical
experience of the local planners from BTRC, an effective taxi trip in Beijing should
be more than five minutes in most cases. Therefore, we deleted all the traces shorter
than five minutes to ensure a more accurate estimation.
We further eliminated traces that clearly contained mistakes, e.g., Figure 5a
shows a GPS trace that may have a GPS mapping mistake as the link with a yellow
mark in the middle is directed from the destination to the origin. Figure 5b is an
example of those GPS traces that make no sense and for which we cannot find any
convincing explanation.
We finally obtained 1,097 consistent and reasonable traces within the simula-
tion time period for the route choice model estimation. Figure 6 details the spatial
distribution of the traces. From left to right, the first three pictures show the 100,
200, 500 most frequently used links and the fourth one shows all the links that
were included. The traces concentrated in the northern part, which is reasonable
since that is the most congested area. Meanwhile, the traces covered almost the
complete network and were deemed adequate to reflect the route choice behavior
in the whole study area.

17
(a) A possible GPS mismatch (b) A trace with mistakes

Figure 5: Unreasonable GPS traces

Figure 6: Spatial distributions of traces

18
3.4 The Combined Calibration of DynaMIT-P
3.4.1 Initial Aggregate Calibration
In our previous study (Ben-Akiva et al.; 2012), the DynaMIT-P Beijing model
had been calibrated using the SPSA algorithm against the aggregate surveillance
data. The route choice model was a Path-size Logit with only one explanatory vari-
able, the time-dependent travel time. Its parameter was calibrated simultaneously
with other calibration variables against the aggregate data only. The systematic
utility function was not tested or estimated using disaggregate GPS traces, and
likely to be oversimplified.
We use this result as the base case to evaluate the calibration improvements
from combining the disaggregate route choice estimation with aggregate calibra-
tion of DynaMIT-P.

3.4.2 Route choice Set Generation

We simultaneously apply three algorithms in DynaMIT-P to generate the choice
set, namely link elimination, simulation and link penalty. Time-dependent link
travel times are used instead of static link lengths in the calculation of shortest
paths. To capture peoples varying attitudes toward the highway, we implemented
a highway bias, namely, multiplying highway link travel times by a certain weight
in the generation of a choice set. When the weight is greater than 1, the paths are
more likely to include fewer highways. Conversely, when the weight is less than
1, paths including more highways are generated. The link number bias was also
introduced to capture peoples attitudes toward intersections, since oftentimes the
more intersections in a path, the larger the number of links in the path. This was
implemented by adding a constant to each link travel time, and thus a path with a
larger number of links would be penalized more. The constant could be adjusted
to reflect different levels of bias.
Choice sets of all OD pairs consisted of 48,796 paths. The maximum number
of paths in a choice set was 222, with a mean of 27.6 paths per OD pair and a
standard deviation of 35.6 paths. The maximum number of paths consistent with
the GPS trace for an OD pair was 68, the mean was 3.12, and the standard deviation
was 5.8 paths.
The coverage test results are shown in Table 2. The high coverage indicates that
the choice set we generated is of high quality and the algorithms we implemented
can be trusted to generate choice sets for other OD pairs in the DTA simulation.

19
 Table 2: Coverage tests for the choice set generated by DynaMIT-P

Overlap 100% 90% 80%

Coverage 80.0% 85.9% 91.5%

3.4.3 Route Choice Model Specification and Estimation

We specified and compared several models and finally arrived at the utility
function as follows:
Vp =1 T imeDepedentT ravelT imep + ln(pathsizep )
+ 2 ShortestP athp + 3 F astestP athp + 4 M ostHighwayp
(13)

Time-Dependent Travel Time

Based on time-dependent link travel times from the latest DTA run, and con-
sidering the start time of each GPS trace, we computed the time-dependent
travel time for each path with a unit of 1000 seconds.

Path-size
PS is a number between 1/J and 1 where J is the size of the choice set.
When PS is equal to 1/J, all alternatives are completely overlapping. When
PS is equal to 1, a path is not overlapping with any other paths.

Shortest Path Dummy

This is a dummy variable that is 1 for the path with the least total length
among all paths with the same OD pair.

Fastest Path Dummy

This is a dummy variable that is 1 for the path with the lowest average travel
time among all paths with the same OD pair.

Most Highway Dummy

This is a dummy variable that is 1 for the path with the highest ratio of its
length spent on the highway, among all paths with the same OD pair.

The model is estimated with Biogeme and the estimation result is shown in
Table 3.

3.4.4 DTA Re-calibration and Iteration

We implemented the estimated route choice model in DynaMIT-P and ran
SPSA calibration again for this new model. With the newly calibrated output travel

20
 Table 3: The result of route choice model estimation

Parameter Coefficient Robust t-test

Time dependent travel time (1000 seconds) -0.0089 -17.99
Pathsize(fixed) 1 N/A
Shortest path dummy 0.842 5.91
Fastest path dummy 0.467 3.08
Most highway dummy 0.426 2.80
Number of observations 1097
Number of parameters 4
Final log-likelihood -1747.480
Adjusted rho-squared 0.285

times from DynaMIT-P, we generated a new choice set and estimated a new route
choice model based on the latest choice set and travel times. continued carrying
out the iterations as described in Section 2, until the output travel times of the two
consecutive aggregate calibrations are close enough.
Table 4 shows the route choice model in the DynaMIT-P base model and the
route choice model of our final calibrated model.
Table 4: The route choice model in DynaMIT-P base model and final calibrated
model

Parameter Base model Final calibrated model

Time-dependent travel time -0.0183 -0.011
Path-size 1.00(fixed) 1.00(fixed)
Shortest path dummy N/A 0.893
Fastest path dummy N/A 0.504
Most highway dummy N/A 0.345

Figure 7 compares the RMSN (Root mean squared errors normalized) for counts
from the base case and combined calibration. The first (leftmost) group is the over-
all calibration result, and other three groups are links with high flows (more than
1400veh/15min), medium flows (1000-1400 veh/15min) and low flows (0-1000
veh/15min) respectively. We can see more improvements on links with low and
medium flows than high flows. Figure 8 compares the RMSN for observed link
travel times from FCD in the base case and combined calibration. The first group
is the overall calibration result, and there are four groups according to the link
travel time: 0-20 seconds, 20-40 seconds, 40-60 seconds and more than 60 sec-

21
onds. We can see more improvements in links with very short and very long travel
times.

0.5
Base Model
New Model
0.45

0.4

0.35

0.3
RMSN

0.25

0.2

0.15

0.1

0.05

0
Overall 01000 10001400 >1400
Group by Link Flow (veh/15min)

Figure 7: Fit to Counts Statistics of the Base Case (Blue) and Combined Calibra-
tion (Red)

The overall calibration results are also reported in Table 5. The improvement
in RMSN for counts is 7.8% and the improvement in RMSN for floating car travel
time is 8.3%. The improvement could have been larger considering the following
facts:

Compared to the scale of the network, the number of sensors is very limited
(only around 120 sensors). At the same time, the distribution of these sen-
sors is limited to expressways, which leads to a failure in capturing possible
significant improvements in other type of roads in the network.

The route choice model specification is still simple. Only three more dummy
variables are included compared to the base model. A route choice model
that captures more influencing factors could possibly make further improve-
ments, for example, the reliability of travel time. However the calculation
of reliability measures require data to derive travel time probabilistic dis-
tributions, which are not yet available from the project. It also calls for a
potential significant change to the DTA model to explicitly treat travel times
as random variables, which will be included in our future work.

22
 0.5
Base Model
New Model
0.45

0.4

0.35

0.3
RMSN

0.25

0.2

0.15

0.1

0.05

0
Overall 020 2040 4060 >60
Group by Link Travel Time (s)

Figure 8: Fit to FCD Link Travel Time Statistics of the Base Case (Blue) and
Combined Calibration (Red)

Table 5: Comparisons of Overall Calibration Results

No. of Observations RMSE RMSN

Base Case 383.8 0.308
Counts(Veh/15min) 1,680
Combined Calibration 353.1 0.284
Base Case 17.30 0.436
Travel Time(s) 52,545
Combined Calibration 15.85 0.400

23
 For a closer look, Figure 9 gives the fit-to-count comparison between the base
case and the combined calibration during the peak period of 8:30AM-8:45AM for a
specific count station. The x-axis is the observed sensor counts and the y-axis is the
simulated ones. A 45-degree line indicates a perfect match between the observed
and the simulated data, and the closer the dots are to the 45-degree line the better
the fit. We can see that the combined calibration gives better fit than the base case.

Figure 9: Fit-to-count Comparison between the Base Model and the New Model

4 Conclusions and Future Directions

In this paper, we extend on the framework of simultaneous demand-supply
DTA calibration based on aggregate observations, and incorporate the disaggregate
route choice observations to improve the calibration accuracy. We formulate the
calibration problem as a bi-level constrained optimization problem. The objective
function is a weighted sum of distances between time-dependent location-specic
simulated aggregate measurements and eld aggregate measurements (e.g., counts,
speeds, and link travel times) and distances between calibrated variable values and
their respective a priori values. Constraints include (1) a simulation-based equilib-
rium DTA model; (2) a choice set generation model; (3) upper and lower bounds on
OD trips and supply/demand parameters; (4) the physical relationships between the
model parameters; (5) the route choice estimation problem, where the likelihood
of observing the disaggregate route observations (e.g. from GPS traces) is maxi-
mized. A priori values of route choice parameters are derived from the lower level

24
route choice estimation problem. The likelihood function is based on a discrete
choice model with route choice sets and attributes generated from performance
measures.
The bi-level calibration/estimation problem is solved by an iterative process
that alternates between three sub-problems: the upper and lower level problems
and the choice set generation model. A case study is conducted in the City of Bei-
jing using DynaMIT-P, a state-of-the-art simulation-based DTA model, using the
proposed methodology. The SPSA algorithm is used in the aggregate calibration
process. A Path-size Logit route choice model is estimated using the disaggregate
GPS trajectories and a latent choice model is implemented considering the discon-
tinuity of the GPS data. The utility function specification includes time-dependent
travel time, Path Size, shortest path dummy, fastest path dummy and most high-
way dummy. Compared to the base case where only aggregate surveillance data
are used, the combined calibration shows an improved accuracy in terms of fit to
observed link flow and link travel time data. Better data and better designed route
choice model specification may help in achieving more significant enhancement.
In future work, the framework can be extended to incorporate more types of
data other than disaggregate trajectories and aggregate traffic data. For example,
with the development of data mining technologies, online social networking web-
sites could be analyzed and provide information for deriving traffic demand, es-
pecially when special events take place. How to fuse data from different sources
with different forms and provide a consistent calibration of DTA models will be a
challenging, yet meaningful topic.

References
Antoniou, C. (2004). On-line Calibration for Dynamic Traffic Assignment, PhD
thesis, Massachusetts Institute of Technology.

Azevedo, J., Costa, M. S., Madeira, J. S. and Martins, E. V. (1993). An algorithm

for the ranking of shortest paths, European Journal of Operational Research
69: 97106.

Balakrishna, R. (2006). Off-line Calibration of Dynamic Traffic Assignment Mod-

els, PhD thesis, Massachusetts Institute of Technology.

Balakrishna, R., Ben-Akiva, M. and Koutsopoulos, H. N. (2007). Offline calibra-

tion of dynamic traffic assignment: Simultaneous demandand- supply estima-
tion, Transportation Research Record: Journal of the Transportation Research
Board 2003: 5058.

25
Balakrishna, R., Koutsopoulos, H. N. and Ben-Akiva, M. (2005). Calibration and
validation of dynamic traffic assignment systems, in H. S. Mahmassani (ed.),
Transportation and Traffic Theory: Flow, Dynamics and Human Interaction,
Proceedings of the 16th International Symposium on Transportation and Traffic
Theory, Elsevier, University of Maryland, College Park, pp. 407426.

Balakrishna, R., Morgan, D., Slavin, H. and Yang, Q. (2009). Large-scale traffic
simulation tools for planning and operations management, 12th IFAC Sympo-
sium on Transpotaton Systems .

Balakrishna, R., Wen, Y., Ben-Akiva, M. and Antoniou, C. (2008). Simulation-

based framework for transportation network management for emergencies,
Transportation Research Record: Journal of the Transportation Research Board
2041: 8088.

Barcelo, J. and Casas, J. (2006). Stochastic heuristic dynamic assignment based on

aimsun microscopic traffic simulator, Transportation Research Record: Journal
of the Transportation Research Board 1964: 7080.

Ben-Akiva, M., Bergman, M., Daly, A. and Ramaswamy, R. (1984). Modeling in-
ter urban route choice behaviour, Proceeding of the 9th International Symposium
on Transportation and Traffic Theory.

Ben-Akiva, M. and Bierlaire, M. (1999). Discrete choice methods and their appli-
cations to short-term travel decisions, in R. Hall (ed.), Handbook of Transporta-
tion Science, Kluwer, pp. 534.

Ben-Akiva, M., Bierlaire, M., Bottom, J., Koutsopoulos, H. N. and Mishalani,

R. G. (1997). Development of a route guidance generation system for real-
time application, Proceedings of the 8th International Federation of Automatic
Control Symposium on Transportation Systems, IFAC, Chania, Greece.

Ben-Akiva, M., Bierlaire, M., Burton, D., Koutsopoulos, H. N. and Mishalani,

R. (2001). Network state estimation and prediction for real-time transportation
management applications, Networks and Spatial Economics 1: 291318.

Ben-Akiva, M., Bottom, J., Gao, S., Koutsopoulos, H. N. and Wen, Y. (2007).
Towards disaggregate dynamic travel forecasting models, Tsinghua Science and
Technology 12(2): 115130.

Ben-Akiva, M. E., Gao, S., Wei, Z. and Wen, Y. (2012). A dynamic traffic assign-
ment model for highly congested urban networks, Transportation Research Part
C 24: 6282.

26
Ben-Akiva, M. and Lerman, S. (1985). Discrete Choice Analysis, MIT Press.

Bierlaire, M. and Frejinger, E. (2008). Route choice modeling with network-free

data, Transportation Research Part C 16: 187198.

Bolduc, D. and Ben-Akiva, M. (1991). A multinomial probit formulation for

large choice sets, Proceedings of the 6th International Conference on Travel
Behaviour.

Bovy, P. H. L. and Fiorenzo-Catalano, S. (2006). Stochastic route choice set gen-

eration: behavioral and probabilistic foundations, Proceedings of the 11th Inter-
national Conference on Travel Behaviour Research, Kyoto, Japan.

Burrell, J. E. (1968). Multiple route assignment and its application to capacity

restraint, Proceeding of the Fourth International Symposium on the Theory of
Traffic Flow.

Cascetta, E. (2001). Transportation Systems Engineering: Theory and Methods,

Applied optimization, Kluwer Academic Publishers, Dordrecht; Boston, MA.

Cascetta, E., Nuzzolo, A., Russo, F. and Vitetta, A. (1996). A modified logit route
choice model overcoming path overlapping problems: Specification and some
calibration results for interurban networks, in J. B. Lesort (ed.), Proceedings of
the 13th International Symposium on Transportation and Traffic Theory, Lyon,
France.

Daganzo, C. F. and Sheffi, Y. (1977). On stochastic models of traffic assignment,

Transportation Science 11(3): 253274.

de la Barra, T., Perez, B. and Anez, J. (1993). Multidimensional path search and
assignment, Proceedings of the 21st PTRC Summer Meeting, pp. 307319.

Florian, M., Mahut, M. and Tremblay, N. (2001). A hybrid optimization-

mesoscopic simulation dynamic traffic assignment model, Proceeding of the In-
ternational IEEE Conference on Intelligent Transportation Systems, Oakland,
CA, Aug. 25-29, pp. 118121.

Fosgerau, M., Frejinger, E. and Karlstrom, A. (2012). A logit model for the choice
among infinitely many routes in a network, Technical report, Royal Institute of
Technology.

Frejinger, E. (2007). Route choice analysis: Data, models, algorithms and appli-
cations, PhD thesis, Ecole Polytechnique Federale de Lausanne.

27
Frejinger, E. and Bierlaire, M. (2007). Capturing correlation with subnetworks in
route choice models, Transportation Research Part B 41: 363378.
Frejinger, E., Bierlaire, M. and Ben-Akiva, M. (2009). Sampling of alternatives for
route choice modeling, Transportation Research Part B 43(10): 984994.
Gao, S. (2005). Optimal Adaptive Routing and Traffic Assignment in Stochastic
Time-Dependent Networks, PhD thesis, MIT.
Hou, A. (2010). Using gps data in route choice analysis: Case study in boston,
Masters thesis, Massachusetts Institute of Technology.
Mahmassani, H. S. (2001). Dynamic network traffic assignment and simulation
methodology for advanced system management applications, Networks and Spa-
tial Economics 1(3/4): 267292.
Peeta, S. and Ziliaskopoulos, A. K. (2001). Foundations of dynamic traffic as-
signment: The past, the present and the future, Networks and Spatial Economics
1(3/4): 233265.
Prato, C. G. (2004). Latent Factors and Route Choice Behavior, PhD thesis, Po-
litecnico di Torio.
Ramming, S. (2002). Network knowledge and route choice, PhD thesis, Mas-
sachusetts Institute of Technology, Cambridge, MA.
Rathi, V., Antoniou, C., Wen, Y., Ben-Akiva, M. and Cusack, M. (2008). As-
sessment of the impact of dynamic prediction-based route guidance using a
simulation-based, closed-loop framework, the 87th annual meeting of the Trans-
portation Research Board, DVD-ROM, Washington, D.C.
Spall, J. C. (1998). Implementation of the simultaneous perturbation algorithm
for stochastic approximation, IEEE Transactions on Aerospace and Electronic
Systems 34: 817823.
Sundaram, S., Koutsopoulos, H. N., Ben-Akiva, M., Antoniou, C. and Balakrishna,
R. (2011). Simulation-based dynamic traffic assignment for short-term planning
applications, Simulation Modelling Practice and Theory 19: 450462.
Train, K. (2003). Discrete Choice Methods with Simulation, Cambridge University
Press.
Vaze, V., Antoniou, C., Wen, Y. and Ben-Akiva, M. (2009). Calibration of dynamic
traffic assignment models with point-to-point traffic surveillance, Transportation
Research Record: Journal of the Transportation Research Board 2090: 19.

28
Wen, Y. (2009). Scalability of Dynamic Traffic Assignment, PhD thesis, Mas-
sachusetts Institute of Technology.

Wen, Y., Balakrishna, R., Ben-Akiva, M. and Smith, S. (2006). Online deployment
of Dynamic Traffic Assignment: architecture and run-time management, IEE
Proceedings Intelligent Transport Systems 153(1): 7684.

Yai, T., Iwakura, S. and Morichi, S. (1997). Multinomial probit with structured co-
variance for route choice behavior, Transportation Research Part B 31(3): 195
207.

Ziliaskopoulos, A. K., Waller, S. T., Li, Y. and Byram, M. (2004). Large-scale

dynamic traffic assignment: Implementation issues and computational analysis,
Journal of Transportation Engineering 130(5): 585593.

29