2009 American Control Conference ThB09.

5
Hyatt Regency Riverfront, St. Louis, MO, USA
June 10-12, 2009

Freeway traffic flow simulation using the Link Node Cell transmission
model
Ajith Muralidharan, Gunes Dervisoglu and Roberto Horowitz

Abstract— This paper illustrates the calibration and impu- the process of extracting the fundamental diagram parameters
tation procedure implemented to specify the inputs to the from the flow and density measurements. The onramp flows
Link-Node Cell Transmission model used for simulating traffic need to be specified as an input, while the offramp flows are
flow in freeways. Traffic flow and occupancy data from loop
detectors is used for calibrating these models and specifying needed to extract the mainline split ratios. It is frequently
the inputs to the simulation. In addition, flow data from observed that ramp flow data is either missing or incorrect,
ramps are often found to be missing or incorrect. A model which makes imputation of these flows essential to specify
based iterative learning technique is used to impute these the model completely.
ramp flows by minimizing the error between simulated and This paper illustrates the modeling and simulation of a
measured densities. The simulation results using the calibrated
parameters and imputed flows indicate good conformation with freeway network. Section II reviews the Link-Node Cell
loop detector measurements. Transmission model used for traffic flow simulations and
states a simple four-state switching model approximation
I. INTRODUCTION
used for imputation. Section III illustrates the steps in
Traffic flow simulation tools are essential for re-creating calibrating freeway section fundamental diagrams. Section
flow and speed characteristics of freeways. Operations plan- IV explains the imputation procedure used for determining
ning, which include ramp metering, demand and incident ramp flows. Finally, section V illustrates an example where
management and its benefit assessment depend on the tools the calibration and imputation procedures are used to specify
which successfully simulate the traffic flows in agreement inputs for the simulation of a 23-mile long Interstate-210
with empirical data. The Tools for Operations planning West freeway in the Los Angeles area.
(TOPL) is a set of tools that simulate traffic flows and
control strategies. This forms an integral component of the II. LINK NODE CELL TRANSMISSION MODEL
California Department of Transportation (Caltrans) “corridor The Link-Node Cell transmission model (LN-CTM) is
management program” - which was introduced to reduce the an extension of the CTM, which can be used to simulate
congestion in 2025 by 40 percent [1]. traffic in any road network. Aurora, a simulation tool in
Traffic flow simulations have been frequently based on TOPL, is based on this CTM implementation [4]. TOPL was
microscopic models, which simulate individual driver be- initially based on the Asymmetric Cell Transmission Model
havior to observe freeway network characteristics. While (ACTM) [5], which is specifically used for freeway traffic
this would be ideal, extensive data collection requirements simulation. In comparison, the LN-CTM has the capability to
and extravagant calibration efforts make these models less simulate traffic networks which include freeways and arterial
lucrative for quick results. In comparison, Cell Transmission networks.
Models (CTM) simulate macroscopic traffic behavior which The traffic network is represented as a directed graph
are specified by volume (flow), density and speed [2]. Also, of links in the LN-CTM. Links represent road segments
the data required for simulation is available for California and nodes are formed at the junctions of links. A time-
Freeways via loop detector based vehicle detector stations varying split-ratio matrix is used to specify the portion of
(vds). PeMS [3] routinely archives the flow, occupancy and traffic moving from a particular input link to an output
speed data from these vds. TOPL is based on a modified link. While a normal link connects two Nodes, a “source”
version of the CTM - the Link-Node Cell Transmission link is used to introduce traffic whereas a “sink” is used
Model (LN-CTM), which simulates traffic flow in networks. to accept traffic moving out of the network. A source link
Simulation of traffic flow in freeways requires fundamental implements a queue model. A fundamental diagram (which
diagram parameters for road sections, as well as input vol- specifies the flow-speed-density characteristics) is specified
umes (flow) from the onramps/freeway entry. Calibration is for each link, while the source links are also specified with
A. Muralidharan is with the Department of Mechanical Engineering, an input demand profile. Figure 1 shows the directed graph
University of California, Berkeley, [email protected] representation of a freeway. The nodes specify the location
G. Dervisoglu is with the Department of Mechanical Engineering, Uni- of a merge between ramps and the mainline (freeway road
versity of California, Berkeley, [email protected]
R. Horowitz is a Professor at the Department of Mechanical Engineering, segment). Each node contains a maximum of one on- and one
University of California, Berkeley. [email protected] off-ramp. In California freeways, the onramps are preceded
This work is supported by the California Department of Transportation by the offramps, therefore the split ratio matrix is specified
through the California PATH Program. The contents of this paper reflect the
views of the author and not necessarily the official views or policy of the to block any flow from the onramp to the offramp. Freeflow
California Department of Transportation. is assumed to prevail in both boundaries of the freeway.

978-1-4244-4524-0/09/$25.00 ©2009 AACC 2916

 and the onramp. w̄i (k)(nJi − ni (k)) represents the available
output capacity. Thus, in freeflow, the flow into the link
is not limited by the available space, thereby allowing for
the freeway to cater to the actual demand. However, in the
case of congestion, the flow is limited by the capacity in
the destination link, and the flow from the onramp and the
previous link is scaled accordingly to limit the flow to the
Fig. 1. Freeway with N links. Each Node contains a maximum of one on- capacity. The density update equations for the links of the
and one off-ramp freeway can be summarized as
(a) FF Mode
A “source” node attached to the upstream cell is used to ni (k + 1) = ni (k) + ci−1(k) − ni (k)v̄i (k) (1)
introduce traffic flow into the network. The onramps are (b) FC Mode
also represented as source links, while the offramps are
represented as sinks. It is also assumed that the off-ramps are ni (k + 1) = ni (k) + ci−1(k)
in freeflow. Table I lists the model variables and parameters. w̄i+1 (nJi+1 − ni+1 (k))
− ni (k)v̄i (k) (2)
ci (k)
Symbol Name Unit (c) CC Mode
section length miles
period hours ni (k + 1) = ni (k) + w̄i (nJi − ni (k))
Fi capacity veh/period
vi free flow speed section/period
w̄i+1 (nJi+1 − ni+1 (k))
− ni (k)v̄i (k) (3)
wi congestion wave speed section/period ci (k)
nJi jam density veh/section
βi split ratio dimensionless (d) CF Mode
k period number dimensionless
fiin (k) flow into section i period k veh/period ni (k + 1) = ni (k) + w̄i (nJi − ni (k)) − ni (k)v̄i (k) (4)
fiout (k) flow out of section i period k veh/period Fi
si (k),ri (k) off-ramp, on-ramp flow in node i veh/period where w̄i (k) = min(wi , (nJ −n
and ) v̄i (k) = min(vi , niF(k)
i
).
i i (k))
in period k In addition, the mainline flows can be determined by
di (k) on-ramp demand for Link i+ 1 in veh/period
period k min(ci (k), w̄i+1 (k)(nJi+1 − ni+1(k)))
ci (k) Total demand for Link i + 1 in veh/period fiout (k) = ni (k)v̄i (k)
period k ci (k)
TABLE I fiin (k) = min(ci−1 (k), w̄i (k)(nJi − ni(k))) (5)
M ODEL VARIABLES AND PARAMETERS . while the ramp flows are determined by
si (k) = βi (k) fiout (k)
min(ci (k), w̄i+1 (k)(nJi+1 − ni+1(k)))
The LN-CTM is explained in [4]. The general algorithm ri (k) = di (k)
implements density updates (Link Updates) and Flow up- ci (k)
dates (Node Updates) in separate steps. The conditional di (k + 1) = di (k) + f liin (k + 1) − ri(k) (6)
structure of the general equations are not easily amenable where f liin is the input flow for the onramp i.
for ramp flow imputation. Specifically for traffic flow simu- The above set of equations correctly represent the LN-
lations, the general equations can be simplified to derive a CTM for freeways under a few assumptions. It is assumed
four mode switching model for each link. This incorporates that the offramp flows are not restricted by the flow ca-
the flow updates directly into the density update equations. pacity/congestion. Similarly onramps are also assumed to
For a section i (Figure 1), the density update equations have no flow capacity restrictions. These assumptions are
belong to the following four modes - FF, CF, CC, and FC, not restrictive for modeling purposes, since the offramps are
where F denotes freeflow and C denotes congestion. These usually specified to be in freeflow, while the capacity of the
modes are decided based on the flow conditions existing onramps will not affect the flow calculations if the capacity
at the input and output node of each link. The CF mode is defined to be the maximum flow observed in the ramps.
is said to be active in Link i , if the input flow into
Link i is in congestion and the output flow from Link III. CALIBRATION
i (into Link i + 1 and the offramp) is in freeflow. Other Like most macroscopic models of vehicular traffic flow,
modes can be interpreted similarly. Here, the input into The LN-CTM makes use of the fundamental diagram, an
Link i is classified to be in congestion if the flow into it empirical curve relating observed densities to observed flows
is determined (limited) by its available capacity rather than at a particular point on the road. A calibrated fundamental
total input demand, i.e. ci−1 (k) > w̄i (k)(nJi − ni (k)), where diagram provides freeflow speed, congestion wave speed,
ci−1 (k) = ni−1 (k)vi−1 (k)(1 − βi−1 (k)) + di−1 (k) is the input critical density, jam density and capacity (Figure 2). The
demand into Link i, which consists of flow from Link i − 1 calibration procedure involves the following.

2917
 210W VDS: 717669 Calibrated Fundamental Diagram
2500
VDS: 717669
v = 62.8 mph
w = 10.4 mph
2000 Jam Density = 224.2 vpmpl
Capacity = 2007.0 vphpl

Flow (vehicles per hour per lane)

1500

1000

500

0
0 50 100 150 200 250
Density (vehicles per mile per lane)

Fig. 2. Fundamental diagram for a freeway section.

Fig. 3. Calibrated Flow vs. Density Scatter Plot of vds 717669 on I-210W
over 15 days’ data
A. Freeway Representation
The first step is to define the geometrical characteristics of
the site- the locations of onramps and offramps, number of can reasonably be expected to traverse the cross-section of
lanes, existence of HOV lanes etc. As the LN-CTM dictates, a road segment. This deterministic notion of capacity has
the freeway should be represented in the form of successive been challenged lately by stochastic approaches [7]. A study
cells. Therefore, the freeway network is divided into cells over the sections of the freeway yields that there is indeed a
each with at most one on- and/or off-ramp and one mainline significant variation in observed maximum flows (Figure 4).
vehicle detector station. The cells must be longer than the
freeflow travel distance, i.e. vi ≤ 1; so that the algorithm Variation of Daily Maximum Flows on 210W
converges. Each cell is assumed to be homogeneous in
2300
terms of number of lanes, grade and geometrical features
2200
so that each cell can be represented by a single fundamental
Flow (vehicles per hour per lane)

2100
diagram. The 23 mile stretch of I-210W (extending between
2000
the Fruit Street onramp and the Lake Avenue onramp)
1900
analyzed in this study consists of 33 such cells.
1800
B. Data Acquisition and Selection 1700

Vehicle detector stations(VDS) contain loop detectors that 1600

provide flow and occupancy data. PeMS [3] processes and 1500

archives these data in form of time series over different 1400

days of operation. PeMS also reports detector performance 1300
for each day of operation. The data for calibration were 768028 767940 717694 717678 761374 717669 717653 717634
chosen from days for which PeMS reported over 80% Freeway Section ID (traffic flows left to right)

functionality for all detectors. While PeMS imputes missing

mainline data using data from adjacent detectors, an 80% Fig. 4. Box plots showing the distribution of observed daily maximum
flows for each section of I-210W.
detector health/functionality is preferred to ensure accuracy
of data used. Further, for each detector station, only days on
In Figure 4, the horizontal axis is the detector IDs placed
which that particular freeway section became congested (with
on the freeway in upstream to downstream (left to right)
speeds below 40mph) were chosen, in an effort to observe
succession. The vertical axis reflects the normalized flow
capacity and congested flow characteristics.
values across these sections of the freeway. The horizontal
C. Calibration of the free-flow speed, v lines inside the boxes correspond to the median of the
The free-flow speed, v, is estimated by performing a least- observed daily maximums among days. The lower and upper
squares fit on the flow vs. density data at the time instants box boundaries represent 1st and 3rd quartiles, or 25th and
where the speed was reported to be above 55 mph. The 75th percentiles, respectively. The whiskers span from either
regression line for the free-flow speed can be seen in Figure end of the box to the smallest and largest data points that
3. are non-outliers, i.e points within 1.5 interquartile range away
from box boundaries. The figure reflects significant temporal
D. Estimation of section capacity and spatial variation of section capacities.
In the fundamental diagram, the apex of the triangle The stochastic approach to capacity is based on the notion
corresponds to the section capacity and it is the highest of breakdown, which describes the operation of a freeway
observed flow. In fact, the definition and choice of capacity near a bottleneck at a time instance where there is a change
is a rather delicate point. The Highway Capacity Manual from free-flow to congestion [8]. Numerous studies on the
[6] defines capacity as the maximum amount of flow that stochastic nature of capacity [7] suggest that the breakdown

2918
occurs randomly, affected by various external factors such as ramp metering. This maximum value of flow across the
as driver behavior, road and weather conditions, incidents, section is then projected horizontally to the free-flow line,
etc. and capacity can be defined as a random variable with a to establish the tip of the triangular fundamental diagram
specific probability distribution depending on the probability (Figure 3). The intersection is defined as the critical density
of breakdown. This phenomenon was also investigated in for the section, above which the flow is congested.
this study. Figure 5 reflects the breakdown flows, capacity
and observed daily maximum flow values for section 717644 E. Calibration of the Congestion Speed Parameter, w
on I-210W on Apr 3rd , 2008. In the figure, the speed plots The last parameter to be calibrated is the congestion speed
for section 717644 and 717642, which is right downstream, parameter, w, which also defines the jam density for the
are plotted to observe the breakdown phenomenon. The section. Similar to capacity, this parameter shows significant
horizontal axis is the time of day and the vertical axis is the diversity. Therefore, an approximate quantile regression [10]
speed. The breakdown flows are recorded at instances when was adopted to estimate this parameter at the higher end of
there is a switch in the flow regime in the upstream section its distribution.
(speeds less than 55mph imply dense flow and speeds below After the critical density is determined, the flow-density
40 mph imply congested flow) whereas the downstream points with density values higher than the critical density
section is in free flow (speeds above 55 mph); in other words, (the data points to the right of the tip) are partitioned along
when the upstream section is operating at active bottleneck the horizontal axis (density axis) into non-overlapping bins
conditions. These instances are labeled with 1,2,3,... in the of 10 data points each. Horizontally, each bin is summarized
figure and the corresponding flow values before and after by ”BinDensity,” the mean of the 10 density values in the
breakdown are listed to the left of the figure, among with bin. Vertically, each bin is summarized by ”BinFlow,” the
the daily maximum flow and the capacity observed over the largest non-outlier flow values among the the 10 flow values
stretch of all investigated days. in the bin. Formally, this largest non-outlier is determined as
follows:
210W Breakdown Flows (vds: 717644, day: 03−Apr−2008)
Bin = { f1 , f2 , ..., f10 }
80
Daily Max
BinFlow = max( fi | fi ∈ Bin, fi < Q3 + 1.5IQR) (7)
70 fi

60 where, f1 through f10 are the flow values inside one such bin,
1
Q3 is the 75th percentile of the data points in the bin and
Speed (mph)

50

45 IQR is defined as the difference between the 25th percentile

2 3 7
40
Bef Aft 6
and the 75th percentile of the data.
1) 1512 1546 vphpl
2) 1370 1284 vphpl
30 3) 1466 1591 vphpl
A constrained least-squares regression is performed on
4) 1385 1375 vphpl
5) 1704 1680 vphpl
6) 1582 1675 vphpl
these BinDensity - BinFlow pairs to obtain the congested
20 7) 1699 1606 vphpl
Daily Max: 1886 vphpl
Capacity: 1980 vphpl upstream (vds: 717644)
flow line and complete the fundamental diagram picture
10
downstream (vds: 717642)
(Figure 3). It is required that the regression line passes
0:00 4:00 8:00 12:00 16:00 20:00 0:00
Time of Day (5 min average)
through the tip of the fundamental diagram, so the regression
is constrained accordingly. The point where the regression
Fig. 5. Speed plots of sections 717644 and 717642 on Apr 3rd , 2008 line crosses zero flow is chosen as the jam density of the
section.
Figure 5 (and many others pertaining to other sections
of the freeway not included here) suggests that although IV. IMPUTATION OF RAMP FLOWS
related to the capacity of a section, the breakdown analysis The LN-CTM model is utilized to impute the missing
is not suitable to the purposes of the capacity estimation in onramp input flows as well as the off-ramp split ratios for
the fundamental diagram framework, since the breakdown one day (24-hour) traffic flow simulation on a large freeway
flow values differ substantially from the observed maximum (eg. 40 miles) segment. The imputation procedure involves
flows. A more detailed analysis on the variation of capacity two stages. First, the total demands ci are determined using
and its comparison to breakdown can be found in [9]. an adaptive learning procedure that minimizes the error
The capacity estimate for model calibration is thus chosen between the model calculated densities and the observed
deterministically to be the highest observed flow throughout PeMS density profile. Then the demands and split-ratios are
all investigated days. The choice of this largest observed flow extracted from the total demand, using a linear program that
as the estimate of capacity is based on the assumption that minimizes the error between the model calculated flows and
external factors such as driver behavior, incidents, weather the observed flow profile [11].
and road conditions always affect the capacity adversely and The imputation procedure employs the adaptive iterative
the actual capacities of freeway sections are rarely observed, learning procedure described in [12]. The freeway traffic flow
if ever. This capacity estimate enables the model to replicate process is assumed to be 24-hour periodic, with respect to
the ideal operating conditions of the freeway and is also the flow and density profiles. This assumption is valid, since
essential to the testing of hypothetical control strategies, such the freeway is always found to be in freeflow (with low

2919
densities, flows) at midnight. The LN-CTM algorithm is run (d) CC Mode
multiple times, and at each run, the algorithm adapts the 
unknown demand estimates to minimize the error between ñoi (k + 1) = n̂i (k + 1) − n̂i (k) − añi (k)
the density generated by the model at each link and the data
from the corresponding PeMS measurement. The procedure ŵi+1 (nJi+1 − n̂i+1(k))

is repeated until the density error reduces to a sufficiently + ŵi (nJi − n̂i (k)) − n̂i (k)v̂i (k)
ĉi (k)
small value or stops decreasing. ñoi (k + 1)
As detailed in [12],because of the 24 hour periodicity, ñi (k + 1) =
(1 + G′K T (k)K(k)
the demand vector can be represented as a convolution of
K(k)
a kernel on a constant influence vector, i.e ci (k) = K(k)T Ci Ĉi (k + 1) = Ĉi (k) − ′′ ×
where K(k) represents a 24 hour periodic time dependent  G 
kernel vector, and Ci is the influence vector. Some typical 1
ĉi (k) −
kernel functions (K(k)) include a unit-impulse or a Gaussian 1/ĉi (k) − G′ K(k)ñi (k + 1)
window centered at time k. n̂i (k + 1) = n̂i (k) − añi (k) + ŵi (nJi − n̂i (k))
The imputation procedure assumes initial estimates for ŵi+1 (nJi+1 − n̂i+1(k))
the influence vectors Ĉi . Typical initial estimates incorporate − n̂i (k)v̂i (k) (11)
ĉi (k + 1)
zero onramp and offramp flows. These estimates are then
dynamically adapted at each time step, so that the model
calculated densities for the whole freeway match with the where G′′ = K T (k)K(k) , ĉi (k) = K(k)T Ĉi (k) and G, G′
density profiles obtained from PeMS. At each time step, the are positive gains. The parameter a is chosen so that the
mode for each cell is determined, and the corresponding error equation is asymptotically stable. As the adaptation
learning update equations are used to adapt the influence procedure is carried out, the ‘error’ in the density profile,
vectors. In the following parameter update equations n̂( k) given by ∑ |ni (k) − n̂i (k)| decreases. Since the CF mode does
represents density estimates, ñ( k) = n(k) − n̂(k) represents not involve adaptation equations, the error may converge to
the density error, and ño (k) represents the a-priori error a non-zero value for when this mode is in effect, while
estimate. other modes shows negligible error. This occurs due to
incorrect mode identification at that time instant. In this case,
(a) FF Mode the corresponding estimates are “triggered” automatically so
that the correct modes are identified. After the trigger, the
ñoi (k + 1) = n̂i (k + 1) − adaptation procedure is continued, till the error becomes
(n̂i (k) + ĉi−1 (k) − n̂i (k)v̂i (k) − añi(k)) negligible or stops decreasing.
ñoi (k + 1) The above procedure identifies the Total demand vector,
ñi (k + 1) = from with the on-ramp demand and off-ramp split ratios are
1 + GK T (k)K(k)
decoupled using a linear program. Figure 6 illustrates the
Ĉi−1 (k + 1) = Ĉi−1 (k) + GK(k)ñi (k + 1)
position of the mainline detector, from which flow data is
n̂i (k + 1) = n̂i (k) + ĉi−1(k + 1) − n̂i(k)v̂i (k) − añi (k) (8) available. Depending on the existing flow conditions, the
(b) FC Mode flows preceding the offramp and following the onramp can
 be described by the equations presented in Figure 6. A linear
program that minimizes |( fi+1 in (k) − f meas (k)) − r
ñoi (k + 1) = n̂i (k + 1) − n̂i (k) − añi(k) i+1 i+1 (k)| +
out meas
|( fi (k) − fi+1 (k)) − si+1 (k)| can be used to identify the
ŵi+1 (nJi+1 − n̂i+1(k)) onramp and offramp flows that best match the observed

+ ĉi−1 (k) − n̂i (k)v̂i (k) mainline flow. Once the onramp flows and demands are
ĉi (k)
ñoi (k + 1) obtained, the onramp input flows can be back-calculated
ñi (k + 1) = ′ using the equations in II.
(1 + G K (k)K(k) + GK T (k)K(k))
T

Ĉi−1 (k + 1) = Ĉi−1 (k) + GK(k)ñi (k + 1)

K(k)
Ĉi (k + 1) = Ĉi (k) − ′′ ×
 G 
1
ĉi (k) −
1/ĉi (k) − G′ K(k)ñi (k + 1)
n̂i (k + 1) = n̂i (k) − añi(k) + ĉi−1 (k + 1)
ŵi+1 (nJi+1 − n̂i+1(k))
− n̂i (k)v̂i (k) (9)
ĉi (k + 1)
(c) CF Mode
Fig. 6. Decouple on-ramp and off-ramp flows.
n̂i (k + 1) = n̂i (k) + w̄i (nJi − n̂i(k)) − n̂i (k)v̂i (k) (10)

2920
 PeMS
V. APPLICATION 2
5
x 10 Error is 6.9177 % Simulated

VMT [hr]
1

The calibration and imputation procedure detailed in the 0

0 5 10 15 20 25
Time [hr]
previous sections were used to specify simulation parameters Error is −0.34078%
5000
for a 23 mile section of I-210W freeway in Pasadena,

VHT [hr]
California. After identifying days with good detector health, 0
0 5 10 15 20 25
the data for these days were downloaded and processed to Time [hr]
Error is 6.2312%
obtain the fundamental diagram parameters. The results have 4000

Delay [hr]
been indicated in Section III. The freeway section had a total 2000
0
of 32 onramps and 26 offramps of which a total of 8 onramps 0 5 10 15 20 25
Time [hr]
and 9 offramps were identified to have incorrect/missing
data. The imputation procedure was carried out for these
ramps. The final density error in the imputation was 4.92%. Fig. 9. Performance measures - Vehicle Hours Travelled (VHT), Vehicle
Miles Travelled (VMT) and Delay.
Figure 7 shows that the density estimates have converged to
their true values without appreciable error.
VI. CONCLUSION
Model calculated Density [veh/mile] Measured Density [veh/mile]
0 0 This paper specifies and elaborates the modeling procedure
5 5
used for traffic flow simulation using the macroscopic Link-
Node Cell Transmission Model. The calibration and impu-
Time [hr]

Time [hr]

10 10 tation procedures are used for specifying the inputs to the

15 15 LN-CTM. The calibration procedure implements linear least
squares and approximate quantile regression methods on the
20 20
available Flow vs. Density data to estimate the fundamental
30 35 40 45 30 35 40 45 diagrams for each section of 23 mile long segment of I-210W
PostMile PostMile
freeway in California. The imputation procedure employs
Fig. 7. Final density contours obtained after imputation. an adaptive identification technique and a linear program to
determine the missing/incorrect onramp flows and offramp
split ratios in this network. The simulations, using the
The calibrated and imputed data are used to run the
calibrated fundamental diagram data as well as the imputed
simulation. Figure 8 shows the simulated and the measured
on-ramp flows and off-ramp split ratios, agree closely with
velocity contours, which indicate good conformation of the
the measurements, as shown by the speed contours and
simulation with the actual PeMS detector data. The simu-
performance measures plots.
lated contour plots not only reproduce the locations of the
major bottlenecks, but also accurately capture the congestion R EFERENCES
present in the freeway network. The simulated and measured [1] California Department of Transportation, “Strategic Growth Plan,”
performance measures are compared in Figure 9, which 2006. www.dot.ca.gov/docs/strategicgrowth.pdf - 2006-10-10.
also show good agreement. The simulated data had 4.95 [2] C. Daganzo, “The cell transmission model: A dynamic representation
of highway traffic consistent with the hydrodynamic theory,” Trans-
% and 8.2 % density and flow errors (as compared to the portation Research, Part B, vol. 28, no. 4, pp. 269–287, 1994.
PeMS measurements) respectively. This procedure was also [3] PeMS, “PeMS website,” 2007. http://pems.eecs.berkeley.edu, accessed
implemented successfully for various freeways like I-880N/S 8/28/2007.
[4] A. Kurzhanskiy, Modeling and Software Tools for Freeway Opera-
and I-210E over different days of data. The results of the tional Planning. PhD thesis, University of California, Berkeley, 2007.
imputation algorithm and simulation for I-210E are given in [5] A. C. et al., “Topl: Tools for operational planning of transportation
[11]. networks.” Submitted to DSCC 2008, 2008.
[6] T. R. Board, Highway Capacity Manual 2000, December 2000.
[7] W. Brilon, J. Geistefeldt, and M. Regler, “Reliability of freeway
Simulated Speed Measured Speed traffic flow: A stochastic concept of capacity,” Proceedings of the 16th
0 0
Symposium on Transportation and Traffic Theory, 2005.
[8] B. Persaud, S. Yagar, and R. Brownlee, “Exploration of the breakdown
5 5
phenomenon in freeway traffic,” Transportation Research Record,
vol. 1634, pp. 64–69, 1998.
Time [hr]

Time [hr]

10 10
[9] G. Dervisoglu, G. Gomes, J. Kwon, R. Horowitz, and P. Varaiya,
15 15 “Automatic calibration of the fundamental diagram and empirical
observations on capacity.” Submitted to TRB 88th Annual Meeting,
20 20 2009.
[10] R. W. Koenker, Quantile Regression. Cambridge U. Press, 2005.
30 35 40 45 30 35 40 45 [11] A. Muralidharan and R. Horowitz, “Imputation of ramp flow data for
PostMile PostMile
freeway traffic simulation.” Submitted to TRB 88th Annual Meeting,
2009.
Fig. 8. Velocity Contours obtained from the I-210W simulation using [12] W. Messner, R. Horowitz, W.-W. Kao, and M. Boals, “A new adaptive
calibrated parameters and the imputed ramp flows and split ratios. learning rule,” IEEE Transactions on Automatic Control, vol. 36-2,
pp. 188–197, 1991.

2921