You are on page 1of 11

Dynamic Assignment-Simulation

Methodology for Multimodal


Urban Transit Networks
.
I . Ömer Verbas, Hani S. Mahmassani, and Michael F. Hyland

This paper presents an integrated transit assignment-simulation tool. nents (a to c), which provides transportation decision makers with an
Finding least cost hyperpaths in a large-scale network and assigning invaluable tool to assess the effectiveness of strategic-level (e.g., add,
travelers onto these paths are computationally challenging problems. remove, and alter transit lines) and operational-level (e.g., transit line
Moreover, modeling the spatial and temporal complexities in a transit frequency setting) transit decisions.
network that result from the discontinuities in transit events, such as “Assigning” individuals (i.e., travelers) to specific paths in a trans-
missing a connection and not receiving a seat, exacerbates the issue portation network has a number of practical uses, many of which relate
of capturing realism. These challenges are overcome by (a) using a least to transportation planning and forecasting. By using the Beckmann
cost hyperpath algorithm that captures the multimodal, multi­pattern, et al. mathematical programming formulation (2), which mathemati-
time-, and approach-dependent features of a transit network to provide cally describes a transportation network in a state of user equilibrium
realistic optimal strategies; (b) using a gap-based assignment approach (3), and the Frank–Wolfe algorithm (4), transportation planners are
to reach fast convergence; and (c) developing a multiagent particle able to “statically” assign individuals to the road network. Sheffi
simulation platform that is able to capture the heterogeneities and the provides a thorough, stand-alone reference for understanding and
discontinuities in travel. The platform was tested on the Chicago Transit using static road network assignment algorithms (5). Merchant
Authority network of 14,000 nodes and 64,000 links; 1.25 million travel- and Nemhauser introduced a dynamic traffic assignment model and
ers were assigned and simulated, along with 21,000 tran­sit vehicles. The algorithm (6). Dynamically modeling vehicles in a transportation
assignment-simulation framework can be used as a network evaluation network allows modelers to relax an unrealistic assumption made
tool to assist decision making at the strategic and operational levels. in static network assignment models that link travel times and flows
are constant over a given time period (7).
Dial (8) and Le Clercq (9) were early contributors to the modern
Public transit plays an essential role in the transportation of indi-
transit assignment problem. Their work assumes that transit travelers
viduals in every major city in the United States and around the
choose the single best path before arriving at their destination. This
world. It provides affordable mobility to those without access to
assumption was later replaced by a more realistic “optimal strategy”
personal vehicles; it provides improved accessibility and mobility
approach (10), which is also referred to as the “hyperpath” approach
to all urban residents in dense cities with high vehicular traffic lev-
in the literature (11). Rather than selecting a single best path to arrive
els; it can also offer a number of environmental benefits. The impor-
at their destination, travelers can reduce their expected total travel
tance of public transit, along with the budgetary constraints faced by
time by considering multiple transit lines (10, 11). The set of all paths
city and state transportation decision makers, requires these decision
and routes a transit traveler considers is referred to as a “hyperpath.”
makers to effectively allocate the limited resources available to them.
Nguyen and Pallottino (11) apply Wardrop’s user equilibrium prin-
This paper introduces a comprehensive transit assignment-simulation
ciple to the case of hyperpaths (3); hence, their formulation requires
framework. It is composed of (a) a time-dependent least cost hyper-
each traveler to choose the least cost hyperpath. The optimal strategy
path algorithm (1), (b) a gap-based transit assignment algorithm, and
requires travelers to take the first transit vehicle in their attractive set
(c) a multiagent particle simulation platform. A review of the litera-
to arrive at the transit station. The attractive set is the set of all transit
ture related to these three fields (a to c) is presented below. The unique
lines of which excluding or including another line would increase
contributions of a and b are discussed in Verbas and Mahmassani and
the expected travel time. Finding the correct subset of transit lines to
a subsequent work, respectively (1). The main contributions of this
include in one’s attractive set is also referred to in the literature as the
study are, first, a simulation platform that captures the spatial and
“common lines” problem (12).
temporal complexities of a regional transit network, including het-
Spiess and Florian also consider the effects of congestion in their
erogeneity in waiting times, discontinuities in seating and standing,
transit assignment algorithm; however, a cost function that incor-
discontinuities in boarding and being rejected, as well as transfers and
porates demand factors is not included (10). De Cea and Fernandez
missed connections and, second, the integration of the three compo-
(13) and Wu et al. (14) explicitly include the effects of congestion
via passenger loads into their assignment algorithm; however, both
Transportation Center, Northwestern University, 600 Foster Street, Evanston, algorithms include a “soft capacity restriction” whereby travelers are
IL 60208. Corresponding author: H. S. Mahmassani, masmah@northwestern.edu. never explicitly prevented from boarding a vehicle although there is
a waiting cost penalty that tends to infinity as vehicles approach their
Transportation Research Record: Journal of the Transportation Research Board,
No. 2498, Transportation Research Board of the National Academies, Washington, carrying capacity. Cepeda et al. use a method of successive averages to
D.C., 2015, pp. 64–74. solve the transit assignment problem with congestion effects and soft
DOI: 10.3141/2498-08 capacity restrictions (15). Hamdouch and Lawphongpanich show the

64
Verbas, Mahmassani, and Hyland 65

existence of a solution in a mathematical formulation with “hard” will keep her previous path (applies only after the first iteration). In
capacity constraints (16). this case, several rules are applied to determine whether she is to
Hamdouch et al. increase the realism and applicability of the transit be switched to a new hyperpath, a topic that will be discussed in the
assignment problem by distinguishing between seated and standing following sections. Second, if the traveler is to be assigned a new
travelers in their capacity constrained analytical transit assignment hyperpath, the algorithm determines which elementary path she will
model (17). Schmocker et al. propose a formulation that considers follow out of the multiple branches of a hyperpath. The details of
seat capacity in a spatially expanded network, whereby travelers make this assignment will be discussed in the following sections.
route choice decisions based on the probability of finding a seat on After the least cost hyperpath calculation and assignment proce-
different routes (i.e., the cost associated with standing in a transit dures are completed for every destination zone, the simulation sub-
vehicle is different from sitting in a transit vehicle on one’s journey) algorithm is called to move the travelers and vehicles in the network.
(18). However, this formulation assumes unlimited standing capac- The main purpose of the simulation is to obtain the experience of
ity. Cominetti and Correa include walking links in their formulation every traveler and vehicle. Obtaining this experience has two uses:
(19). Walking links allow travelers to walk between transit stops or
even forgo a transit leg if it is faster to walk than to wait for and ride 1. The link and node-pass costs, which include waiting times,
a transit vehicle. seated loads, and total vehicle loads, are updated for the least cost
Other, rule-based assignment techniques have also been proposed hyperpath calculation in the next iteration.
and used primarily in the context of service design procedures (20, 21). 2. The experienced generalized cost of a traveler and its gap
Several transit simulation tools have been developed to assist plan- from the best hyperpath’s cost affect the reassignment evaluation in
ners with operational and strategic decisions, but none of these the next iteration.
are used in the transit assignment framework (22). MILATRAS
is a simulation tool created to solve the transit assignment prob- Once all vehicles and travelers are simulated in the network, a
lem (23). MILATRAS is a microscopic, learning-based, stochastic, convergence check is performed to either terminate or continue to
non-equilibrium-based simulation-assignment tool. MATSIM is an the next iteration.
open source transit simulation tool that seeks to converge a tran-
sit simulation toward user equilibrium (24). An early simulation-
assignment framework for intermodal networks was presented by Multimodal Least Cost Hyperpath Calculation
Abdelghany and Mahmassani, but it only considered congestion on
the road network and did not consider transit capacity constraints; This section describes the properties of the hyperpath algorithm without
its scalability to large-scale networks was also limited (25). More going into technical details, which are described separately (1).
recently, work by Noh et al. (26) and Noh (27) developed a transit In an urban setting, transit travelers have multiple transit options.
assignment logic for large-scale networks intended to interface with While some of these options are available at a traveler’s current
a road highway simulator. Unlike the stochastic non-equilibrium- stop or station, other attractive options may exist within walking or
based MILATRAS, the simulation-assignment model described in biking distance. These nearby options are not properly modeled in
this paper is an equilibrium-based, deterministic simulation model the existing literature. Furthermore, for shorter distances, walk-
to be used in the transit assignment framework. A similar freight ing and biking modes may be a substitute for transit modes. There-
transportation simulation-assignment model was introduced by fore, the proposed hyperpath formulation is considered multimodal,
Mahmassani et al. (28). Their model analyzes the transportation of including transit modes such as bus, rail, and commuter rail as well
multiproduct, intermodal freight over a given network. The model was as walking and biking.
successfully implemented on a large-scale intermodal rail network Another property of this algorithm is that it captures the concept
spanning the continent of Europe (29). of service patterns (30, 31). A service pattern is a subset of stops
served from the entire stop set of a route. A commonly used pat-
tern is the “short-turn” pattern, which serves the higher-demand
Methodological Features segment of a route and overlaps the full-length pattern. The avail-
ability and frequency of different routes and patterns are changing
This section presents the overall transit assignment and simulation over time. Therefore, the hyperpath algorithm used in this study
algorithm followed by three detailed subsections that discuss multi­
is time dependent. Furthermore, the link and node costs are based
modal least cost hyperpath calculation, transit assignment, and tran-
on the frequency-of-service approach rather than a schedule-based
sit simulation. These subalgorithms are integrated into an iterative
approach.
framework to reach equilibrium for large-scale transit networks with
Hyperpath formulations, as well as conventional least cost path
vehicle seating and capacity constraints.
formulations, use labels at the nodes to designate the generalized cost
from a given node to the destination node or from the origin node
Main Algorithm to the given node according to which direction the least cost path is
constructed. Time-dependent algorithms expand this label structure to
Figure 1 presents the flowchart of the assignment and simulation include the time dimension, whereas movement-dependent algorithms
algorithm used in this study. The algorithm starts by reading the add another dimension that factors in the approach used to get from
input data followed by a convergence check, which is skipped at the the previous node to the current node (32). In a transit framework,
first iteration. Next, the algorithm builds least cost hyperpath trees it is important to capture this movement and approach dependency
sequentially for every destination zone in the network. Once the because travelers’ decisions are dependent not only on which node
time-dependent, multimodal least cost hyperpath tree is constructed they are at but also on the way they have arrived at that node. As an
from all nodes to the destination zone in question, the algorithm example, it might be easier for walking travelers who are one block
loops over all travelers going to that destination. Next, two decisions away from their destination to continue walking rather than waiting
are made: first, the algorithm determines whether or not the traveler for a bus.
66 Transportation Research Record 2498

Start Terminate

Y Main
Y

FIGURE 1   Main algorithm (iter 5 iteration; dest 5 destination; N 5 no; Y 5 yes).

Another feature of the algorithm is that it counts the number of to include different modes and service patterns. The cost of each
transfers along the hyperpath, a feature that enables the progres- link–service pair changes over time intervals to capture time depen-
sive penalization of transfers. In this study, the transfer penalty is dency. However, it is assumed to be constant within the time interval.
increasing quadratically with the number of transfers. Finally, the See Figure 2 for the vectorized links and the approach dependency
probability of standing and the probability of being rejected are also of choices.
incorporated into the formulation to deter travelers from congested Figure 3 presents a sample network. There are two types of nodes in
conditions. The inconvenience of standing is modeled by increasing the network: zone centroids (origin and destination centroids included)
the perceived in-vehicle travel time, while an exponentially growing and transit nodes. Origin centroids are used for demand generation,
capacity cost is used to penalize links with highly loaded vehicles. whereas destination centroids serve as destination nodes for the least
cost path calculations. Transit nodes represent the real-world transit
network.
Network Representation There are three types of links in the network: centroid (access and
egress) connectors, transfer links, and transit links. The egress con-
As described in the previous section, the algorithm incorporates dif- nectors serve as links and always have a transit node as the upstream
ferent modes, as well as the different service patterns. The generalized node and a destination centroid as the downstream node. Conversely,
cost of these services is changing over time according to availability an access connector is a source and always has an origin centroid
and frequency. Furthermore, the cost of using these services is also as the upstream node and a transit node as the downstream node.
dependent on how the traveler is reaching this service (movement Transfer (walk and bike) links connect transit nodes in the vicinity
and approach dependency). To address these multi­dimensionalities, of each other if they were not already connected by a transit link.
studies in the literature have expanded networks either spatially by Transit links connect two transit nodes that are on the route of a
adding imaginary nodes for boarding, alighting, and waiting or tem- transit line. As a result, a transit link has at least one transit pattern in
porally by adding nodes for different time points. This study does its vector. If it is served by a bus pattern, walking and biking patterns
not expand the network by adding nodes for either spatial or tem- are also added to its vector. If it is a rail pattern, walking and biking
poral purposes; the only expansion is the vectorization of each link are prevented by default.
Verbas, Mahmassani, and Hyland 67

Reassignment Decision

In a gap-based approach, the probability (SProbtr) of switching a


traveler (tr) is

EGC tr − GCLatqr
SProb tr = (1)
EGC tr
(a)
where EGCtr is the experienced total cost of traveler tr and GCLatqr
is the least cost hyperpath between the origin, destination pair (q, r)
at the assignment interval at. For every traveler, a random number
is generated between zero and one, and if this number is less than
SProbtr, the traveler is reassigned. In this approach, a traveler is very
likely to be reassigned if the experienced cost EGCtr has a high dis-
crepancy from the least cost hyperpath cost GCLatqr. Compared with
the conventionally used method of successive averages approach,
this gap-based formulation increases the convergence rate (33).

(b) Constructing an Elementary Path

FIGURE 2   Network representation: (a) multiple modes and Given the set of nodes N in the network, a traveler is generated at an
patterns and (b) approach dependency. origin centroid node q ∈ N at a certain simulation interval st. This
simulation interval falls into a least cost hyperpath calculation time
interval t. At the origin node q at time t, the traveler has a set of
options depicted by R xiq (t). Each element k ∈ R xiq (t) in the set points
Transit Assignment to a next possible node hk and service pattern yk. Each option has the
k( )
xyk
probability LProb iqh t , which represents the probability of going
Following the least cost hyperpath calculation for a destination to node hk via pattern yk at time interval t given that one has arrived
zone, the transit assignment subalgorithm is called as seen in Fig- at node q from node i via pattern x. Let the cumulative probability
ure 1. For a given least cost hyperpath tree, two levels of decisions of each option be
are made:
k
1. Whether a traveler will be reassigned and xyk
CProb iqh t = ∑ LProb iqh
k( )
xym
m( )
t (2)
2. If yes, which elementary path the traveler will be assigned to. m =1

Transit Stop
Zone Centroid
Transit Link
Transfer Link
Centroid Connector

(b)
(a)

FIGURE 3   Sample network: (a) overview and (b) closer look at intersection with
transfer links.
68 Transportation Research Record 2498

A random number is generated between zero and one, which will Moving Travelers
certainly fall into one of the intervals CProb iqh k −1 ( )
xyk −1 xyk
t , CProb iqh t ,
k ( )

k ∈ [1, |R iq (t)|], where CProb iqh0 (t ) = 0 by definition. This procedure


x xy0 The flowchart of MOVE_TRAVELERS(st) is presented in Figure 4.
assigns a traveler to her next node and pattern. Say she is assigned The algorithm starts by scanning the Traveler_Queue(st). The first
to node 11, pattern A, and she would arrive there at time t ′ on the traveler tr in the queue is selected and then deleted from the queue.
basis of pattern A’s travel time on link (q, 11) at time interval t. Then, If the traveler has arrived at her destination, said traveler is removed
the attractive set of options would be R q,11 A
(t ′) with probabilities from the network and the algorithm moves to the next traveler. If the
LProb qAy,11,k hk (t ) . The procedure continues until the destination node traveler is not at her destination yet, and her upcoming mode is walk-
r ∈ N is reached, which completes the elementary path construction ing or biking, she is transferred to her next node j from her current
of a traveler. node i. Since the in-motion travel time LTT xij (t) for biking or walking
The origin node q, (i, q) is an imaginary loading link, and x is is known for the corresponding link (i, j), pattern x = walk or x = bike
an imaginary loading pattern. The approach dependency is dem- at time interval t, the next node arrival time st′ = st + LTT xij (t). The
traveler is then inserted into Traveler_Queue(st′).
k( )
xyk
onstrated in Figure 2b. The probabilities LProb ijh t are calculated
on the basis of service frequencies; details can be found in Verbas If the traveler tr is already in a vehicle, the traveler’s movement
and Mahmassani (1). is processed in the MOVE_VEHICLES(st) subroutine. If the trav-
eler tr is waiting for a vehicle, the Vehicle_Queue(st) is scanned.
If a vehicle veh is at node i and its pattern is x, the person’s wait-
ing time is updated by taking the difference between the current
Multiagent Particle Simulation
time st and the node arrival time of traveler tr. Then the traveler is
Capturing the disaggregate movement of travelers and vehicles in a inserted into Veh_Traveler_Queue(veh). On the basis of seat avail-
transit network is an important contribution of this study. There are ability, the seating and standing status of the traveler is updated. If
many benefits of moving each agent individually in the network the traveler is rejected to board, she keeps waiting until the next
as opposed to working solely with traveler and vehicle flows. Dis­ vehicle arrives. If there is no vehicle at node i at time interval st
aggregating travelers allows the model to capture (a) heterogeneity in belonging to pattern x, the traveler keeps waiting and is inserted
waiting times, (b) discontinuities in seating and standing, (c) discon- into Traveler_Queue(st + 1).
tinuities in boarding and being rejected, and (d) transfers and missed
connections.
The simulation of travelers and vehicles in the network follows Moving Vehicles
these steps:
The flowchart of MOVE_VEHICLES(st) is presented in Figure 5.
The algorithm starts by scanning the Vehicle_Queue(st). The first
Step 1. Initialize by moving all vehicles and travelers to their
vehicle veh in the queue is selected and then deleted from the queue.
starting time and locations.
If the vehicle has arrived at its destination, said vehicle is removed
Step 2. Simulate.
from the network and the algorithm moves to the next vehicle. If
– for every simulation interval st
the vehicle is not yet at its destination, it is transferred to its next
• Call MOVE_TRAVELERS(st) and node according to its schedule. The seated and total load informa-
• Call MOVE_VEHICLES(st) tion remains the same. Then it is inserted into Vehicle_Queue(st′),
– next. where st′ = st + LTT xij (t).
Step 3. Terminate. With the movement of the vehicle, the travelers in the vehicle are
moved as well. This move is achieved by scanning Veh_Traveler_
The time interval st here is allowed to be different from the hyper- Queue(veh). Every traveler tr in the vehicle is moved to the next node
path calculation time interval, the assignment interval, or both. This of the vehicle, and the travel times and node arrival times of each are
approach allows for a more precise simulation by choosing a small updated. A conditional check is performed on whether the traveler
simulation interval size, as the simulation is computationally not very will stay in the vehicle for the node thereafter. If true, the next traveler
costly in comparison with hyperpath calculation and assignment. tr is called from Veh_Traveler_Queue(veh).
The simulation subroutine is called after the hyperpath calculation If the traveler is alighting from the vehicle, she is removed from
and passenger assignment are over for all destinations and travel- Veh_Traveler_Queue(veh). The vehicle veh’s seated and total loads
ers in a given iteration as can be seen in Figure 1. The subroutines are updated accordingly. After all travelers in the vehicle are scanned
MOVE_TRAVELERS(st) and MOVE_VEHICLES(st) are called at and moved forward and their alighting and staying decisions are pro-
every simulation interval st; however, in a given simulation time cessed, the algorithm moves to the beginning of the Veh_Traveler_
interval st, only a subset of passengers and vehicles is scanned and Queue(veh) to reassign the seats. As the newly boarded travelers
updated. Those scanned are the vehicles and passengers that are in processed by MOVE_TRAVELERS(st) are inserted at the end of
the Traveler_Queue(st) and Vehicle_Queue(st), respectively. Once the Veh_Traveler_Queue(veh), the travelers already in the vehicle
scanned, an agent is removed from her queue at st and inserted into are given seating priority, automatically ordered by how early they
a future queue st′ > st. The future interval st′ is calculated according boarded the vehicle.
to the movement speed of the transit vehicle, biking, or walking. In
the case in which a traveler is going to wait, the traveler is inserted
into Traveler_Queue(st′), where st′ = st + 1. Experiments
All insertions into queues are done at the end of the queue. The
scanning of a queue always starts from the beginning. The deletion Two validation experiments are demonstrated on the Chicago Tran-
from the queue can be done from the middle of the queue granted that sit Authority (CTA) network in Chicago, Illinois, presented in Fig-
the order of the remaining agents in the queue remains unchanged. ure 6. The first experiment runs from 7 a.m. to 12 p.m., whereas
Verbas, Mahmassani, and Hyland 69

FIGURE 4   Flowchart for movement of travelers.


70 Transportation Research Record 2498

FIGURE 5   Flowchart for movement of vehicles.


Verbas, Mahmassani, and Hyland 71

• 63,602 links,
• 134 routes,
• 823 patterns,
• Vehicle trips:
5,975 for 7 a.m. to 12 p.m. and
20,736 for a full day, and
• Transit traveler demand:
438,687 for 7 a.m. to 12 p.m. and
1,261,320 for a full day.

The network and the vehicle schedule were prepared by using the
General Transit Feed Specification data (34); the passenger demand
data were obtained from a forecasting project (35, 36). The experi-
ment was performed to assign and simulate travelers and vehicles
with the use of the following weights and parameters:

• Relative weight of waiting (xw) with respect to seated in-vehi-


cle travel time (IVTT), XW = 2;
• Relative weight of walking (ww) with respect to seated IVTT,
WW = 1;
• Walking speed is set to 4.6 ft/s;
• Link travel times of transit vehicles are obtained from the
General Transit Feed Specification data and are assumed not to be
affected by traffic or vehicle load;
• Relative weight of standing (ws) with respect to seated IVTT,
WS = 2;
• Biking is disabled;
• Base transfer penalty PX = 5 min;
• Simulation interval = 1 s;
• Assignment interval = 30 min; and
• Least cost path calculation interval = 30 min.

Figure 7 presents the progression of the average gap per traveler


over iterations for both scenarios. The 7 a.m. to 12 p.m. experiment
converges to an average gap of 3 min, while the full-day experiment
converges to 4 min.
FIGURE 6   CTA network. Similarly, the 7 a.m. to 12 p.m. experiment converges to an average
cost of 61 min, while the full-day experiment converges to 62 min, as
seen in Figure 8. The average cost is a measure of system optimum,
the second experiment runs for the entire day. The network has the while the average gap is a measure of user equilibrium.
following properties: Introducing vehicle capacity constraints results in the rejection of
some travelers trying to board a vehicle during the simulation and
• 1,072 zones, induces penalties in the hyperpath calculation. A rejected traveler
• 13,754 nodes: would board the next vehicle of the same pattern or route in the best
11,610 stops and stations and case and would be unable to complete the trip in the worst case, result-
2,144 centroids, ing in very high experienced costs. As an implementation strategy, a
Gap per traveler (min)

7 a.m.–12 p.m.
Full Day

Iteration

FIGURE 7   Change in average gap over iterations.


72 Transportation Research Record 2498

Cost per traveler (min)


7 a.m.–12 p.m.
Full Day

Iteration

FIGURE 8   Change in average cost over iterations.

penalty of 3 h is added to the experienced cost of a traveler who was dependent features of a transit network enables finding strategies
not able to finish her trip. This mostly ensures the reassignment of that reflect the real-world behavior of travelers. A gap-based solution
the traveler in the next iteration as the gap between the least cost and approach for assignment, which evaluates the individual experience
the experienced cost becomes very high. of every traveler in the network, provides a good convergence rate for
Figure 9 presents the number of rejection occurrences over the large-scale applications. These individual experiences are obtained
number of travelers as percentages. The 7 a.m. to 12 p.m. and the via the multiagent particle simulation platform, which is able to cap-
full-day scenarios converge to 1.0% and 0.8%, respectively. Simi- ture the heterogeneities and the discontinuities in travel, such as seat-
larly, the 7 a.m. to 12 p.m. scenario converges to 0.01% incomplete ing and standing, boarding and being rejected, and transferring and
travels, whereas the full-day scenario converges to 0.23%, as seen missing a connection. The platform is tested on the CTA network.
in Figure 10. These numbers show that most of the rejected travelers Convergence is reached for a 5-h and for an entire-day scenario on a
find another vehicle to board and are able to reach their destinations. network of 14,000 nodes and 64,000 links in which 1.25 million trav-
With a 64-bit personal computer with a central processing unit elers are assigned and simulated along with 21,000 transit vehicles.
speed of 2.30 GHz, one iteration requires This assignment-simulation tool has been integrated with a road net-
work traffic particle simulator (37, 38) and an activity-based travel
• 50-GB memory and 1.5 h for the 7 a.m. to 12 p.m. scenario and
demand model (39, 40) for use for strategic planning decisions by
• 120-GB memory and 12 h for the full-day scenario.
the Chicago Metropolitan Agency for Planning. Application in con-
The results presented in this section demonstrate that the junction with optimal frequency allocation and service design is also
simulation-assignment framework introduced in this paper is able illustrated in Verbas and Mahmassani (41).
to equilibrate the assignment of travelers in a large-scale transit
network such as the CTA network.
Acknowledgments

Conclusion This paper is based on research funded by the Chicago Metropolitan


Agency for Planning (CMAP) as part of the project Network Micro-
This paper introduces an integrated tool to find least cost hyperpaths, simulation Extension to Activity-Based Travel Model conducted by
assign travelers, and simulate them along with vehicles in a transit the Northwestern University Transportation Center (NUTC), under a
network. Modeling the multimodal, multipattern time- and approach- subcontract to Parsons Brinckerhoff, Inc. (PB). The authors are grate-
Travelers (%)

7 a.m.–12 p.m.
Full Day

Iteration

FIGURE 9   Change in percentage of rejections over iterations.


Verbas, Mahmassani, and Hyland 73

Travelers (%)
7 a.m.–12 p.m.
Full Day

Iteration

FIGURE 10   Change in percentage of incomplete travels over iterations.

ful to Peter Vovsha of PB and Kermit Wies of CMAP for helpful 17. Hamdouch, Y., H. W. Ho, A. Sumalee, and G. Wang. Schedule-Based
comments on the method described in this paper. The authors also Transit Assignment Model with Vehicle Capacity and Seat Availability.
Transportation Research Part B, Vol. 45B, No. 10, 2011, pp. 1805–1830.
acknowledge the contributions of other NUTC researchers, especially 18. Schmocker, J. D., A. Fonzone, H. Shimamoto, F. Kurauchi, and M. G. H.
Ali Zockaie, Hooram Halat, Andreas Frei, and Xiang Alex Xu. Bell. Frequency-Based Transit Assignment Considering Seat Capacities.
Transportation Research Part B, Vol. 45B, No. 2, 2011, pp. 392–408.
19. Cominetti, R., and J. Correa. Common-Lines and Passenger Assignment
References in Congested Transit Networks. Transportation Science, Vol. 35, No. 3,
2001, pp. 250–267.
. 20. Han, F. H., and N. H. M. Wilson. The Allocation of Buses in Heavily
1. Verbas, I. Ö., and H. S. Mahmassani. Least Cost Hyperpaths in Multi- Utilized Networks with Overlapping Routes. Transportation Research
modal Transit Networks: A Time and Approach-Dependent Formulation. Part B, Vol. 16B, No. 3, 1982, pp. 221–232.
Presented at 94th Annual Meeting of the Transportation Research Board, 21. Baaj, M. H., and H. S. Mahmassani. TRUST: A Lisp Program for the
Washington, D.C., 2015. Analysis of Transit Route Configurations. In Transportation Research
2. Beckmann, M. J., C. B. McGuire, and C. B. Winston. Studies in the Eco- Record 1283, TRB, National Research Council, Washington, D.C., 1990,
nomics of Transportation. Yale University Press, Conn., 1956. pp. 125–135.
3. Wardrop, J. G. Some Theoretical Aspects of Road Traffic Research. 22. Toledo, T., O. Cats, W. Burghout, and H. N. Koutsopoulos. Mesoscopic
Proc., Institute of Civil Engineering Part 2, 1952, pp. 352–378. Simulation for Transit Operations. Transportation Research Part C,
4. Frank, M., and P. Wolfe. An Algorithm for Quadratic Programming. Vol. 18C, No. 6, 2010, pp. 896–908.
Naval Research Logistics Quarterly, Vol. 3, No. 1-2, 1956, pp. 95–110. 23. Wahba, M., and A. Shalaby. Multiagent Learning-Based Approach to
5. Sheffi, Y. Urban Traffic Networks: Equilibrium Analysis with Mathemati- Transit Assignment Problem: A Prototype. In Transportation Research
cal Programming Methods. Prentice Hall, Englewood Cliffs, N.J., 1985. Record: Journal of the Transportation Research Board, No. 1926, Trans-
6. Merchant, D. K., and G. L. Nemhauser. A Model and an Algorithm for the portation Research Board of the National Academies, Washington, D.C.,
Dynamic Traffic Assignment Problems. Transportation Science, Vol. 12, 2005, pp. 96–105.
No. 3, 1978, pp. 183–199. 24. Bekhor, S., C. Dobler, and K. W. Axhausen. Integration of Activity-
7. Mahmassani, H. S., and R. Herman. Dynamic User Equilibrium Depar- Based and Agent-Based Models: Case of Tel Aviv, Israel. In Trans-
ture Time and Route Choice on Idealized Traffic Arterials. Transportation portation Research Record: Journal of the Transportation Research
Science, Vol. 18, No. 4, 1984, pp. 362–384. Board, No. 2255, Transportation Research Board of the National
8. Dial, R. B. Transit Pathfinder Algorithm. In Highway Research Record 205, Academies, Washington, D.C., 2011, pp. 38–47.
HRB, National Research Council, Washington, D.C., 1967, pp. 67–85. 25. Abdelghany, K. F., and H. S. Mahmassani. Dynamic Trip Assignment-
9. Le Clercq, F. A Public Transport Assignment Method. Traffic Engineer- Simulation Model for Intermodal Transportation Networks. In Trans-
ing and Control, Vol. 14, 1972, pp. 91–96. portation Research Record: Journal of the Transportation Research
10. Spiess, H., and M. Florian. Optimal Strategies: A New Assignment Model Board, No. 1771, TRB, National Research Council, Washington, D.C.,
for Transit Networks. Transportation Research Part B, Vol. 23B, No. 2, 2001, pp. 52–60.
1989, pp. 83–102. 26. Noh, H., M. Hickman, and A. Khani. Hyperpaths in Network Based
11. Nguyen, S., and S. Pallottino. Equilibrium Traffic Assignment for Large- on Transit Schedules. In Transportation Research Record: Journal of
Scale Transit Networks. European Journal of Operational Research, the Transportation Research Board, No. 2284, Transportation Research
Vol. 37, No. 2, 1988, pp. 176–186. Board of the National Academies, Washington, D.C., 2012, pp. 29–39.
12. Chriqui, C., and P. Robillard. Common Bus Lines. Transportation Science, 27. Noh, H. Capacitated Schedule-Based Transit Assignment Using a
Vol. 9, No. 2, 1975, pp. 115–121. Capacity Penalty Cost. PhD dissertation. University of Arizona, 2013.
13. De Cea, J., and E. Fernandez. Transit Assignment for Congested Pub- 28. Mahmassani, H. S., K. Zhang, J. Dong, C.-C. Lu, V. C. Arcot, and
lic Transport Systems: An Equilibrium Model. Transportation Science, E. Miller-Hooks. Dynamic Network Simulation-Assignment Platform
Vol. 27, No. 2, 1993, pp. 133–147. for Multiproduct Intermodal Freight Transportation Analysis. In Trans-
14. Wu, J. H., M. Florian, and P. Marcotte. Transit Equilibrium Assignment: portation Research Record: Journal of the Transportation Research
A Model and Solution Algorithms. Transportation Science, Vol. 28, No. 3, Board, No. 2032, Transportation Research Board of the National Acad-
1994, pp. 193–203. emies, Washington, D.C., 2007, pp. 9–16.
15. Cepeda, M., R. Cominetti, and M. Florian. A Frequency-Based Assign- 29. Zhang, K., R. Nair, H. S. Mahmassani, E. Miller-Hooks, V. C. Arcot,
ment Model for Congested Transit Networks with Strict Capacity Con- A. Kuo, J. Dong, and C.-C. Lu. Application and Validation of Dynamic
straints: Characterization and Computation of Equilibria. Transportation Freight Simulation–Assignment Model to Large-Scale Intermodal Rail
Research Part B, Vol. 40B, No. 6, 2006, pp. 437–459. Network: Pan-European Case. In Transportation Research Record:
16. Hamdouch, Y., and S. Lawphongpanich. Schedule-Based Transit Assign- Journal of the Transportation Research Board, No. 2066, Transportation
ment Model with Travel Strategies and Capacity Constraints. Transporta- Research Board of the National Academies, Washington, D.C., 2008,
tion Research Part B, Vol. 42B, No. 7-8, 2008, pp. 663–684. pp. 9–20.
74 Transportation Research Record 2498

30. Furth, P. G., and F. B. Day. Transit Routing and Scheduling Strategies Procedures for ATIS. ATMS Applications. Technical Report DTFH61-
for Heavy-Demand Corridors. In Transportation Research Record 1011, 90-R-0074-FG. Center for Transportation Research, University of
TRB, National Research Council, Washington, D.C., 1985, pp. 23–26. Texas at Austin, 1994.
31. Furth, P. G. Zonal Route Design for Transit Corridors. Transportation 38. Mahmassani, H. S. Dynamic Network Traffic Assignment and Simu-
Science, Vol. 20, No. 1, 1986, pp. 1–12. lation Methodology for Advanced System Management Applications.
32. Ziliaskopoulos, A. K., and W. Wardell. An Intermodal Optimum Path Networks and Spatial Economics, Vol. 1, No. 3-4, 2001, pp. 267–292.
Algorithm for Multimodal Networks with Dynamic Arc Travel Times and 39. Vovsha, P., R. Donnelly, M. Bradley, J. Bowman, H. Mahmassani,
Switching Delays. European Journal of Operational Research, Vol. 125, T. Adler, K. Small, and F. Koppelman. Improving Our Understanding
No. 3, 2000, pp. 486–502. of How Highway Congestion and Price Affect Travel Demand. Final
33. Lu, C.-C., H. S. Mahmassani, and X. Zhou. Equivalent Gap Function- report. SHRP 2 Capacity Project C04. Transportation Research Board
Based Reformulation and Solution Algorithm for the Dynamic User of the National Academies, Washington, D.C., 2012.
Equilibrium Problem. Transportation Research Part B, Vol. 43B, No. 3, 40. Zhang, K., H. S. Mahmassani, and P. Vovsha. Integrated Nested Logit
2009, pp. 345–364. Mode Choice and Dynamic Network Micro-Assignment Model Plat-
34. Google Developers. General Transit Feed Specification (GTFS). https:// form to Support Congestion and Pricing Studies: The New York Metro­
developers.google.com/transit/gtfs/reference. Accessed March 1, 2014. politan Case. Presented at 90th Annual Meeting of the Transportation
35. Mahmassani, H. S., L. Jiang, A. Frei, T. Hou, A. Zockaie, and M. Saberi. Research. Board, Washington, D.C., 2011.
Modeling and Forecasting of Toll Revenues. Final report. FHWA, Wash- 41. Verbas, I. Ö., and H. S. Mahmassani. Integrated Frequency Allocation
ington, D.C., 2013.
and User Assignment in Multi-Modal Transit Networks: Methodology
36. Zockaie, A., M. Saberi, H. S. Mahmassani, L. Jiang, A. Frei, and T. Hou.
and Application to Large-Scale Urban Systems. Presented at 94th Annual
Towards Integrating an Activity-Based Model with Dynamic Traffic
Meeting of the Transportation Research Board, Washington, D.C., 2015.
Assignment Considering Heterogeneous User Preferences and Reli-
ability Valuation: Application to Toll Revenue Forecasting in Chicago.
Presented at 94th Annual Meeting of the Transportation Research The authors are responsible for the content of this paper.
Board, Washington, D.C., 2015.
37. Mahmassani, H. S., T. Y. Hu, S. Peeta, and A. Ziliaskopoulos. Devel- The Standing Committee on Transportation Network Modeling peer-reviewed
opment and Testing of Dynamic Traffic Assignment and Simulation this paper.

You might also like