Large-scale Agent-based Multi-modal Modeling of Transportation

Networks

System Model and Preliminary Results

Ahmed Elbery

, Filip Devorak

, Jianhe Du

, Hesham A. Rakha

and Matthew Klenk

Department of Computer Science, Virginia Tech, Blacksburg, VA 24061, U.S.A.

Palo Alto Research Center (PARC), A Xerox Company, 3333 Coyote Hill Road, Palo Alto, CA 94304, U.S.A.

Virginia Tech Transportation Institute, 3500 Transportation Research Plaza, Blacksburg, VA 24061, U.S.A.

Keywords:

Large-scale Modeling, Agent-based Modeling, Multi-modal Systems.

Abstract:

The performance of urban transportation systems can be improved if travelers make better-informed decisi-

ons using advanced modeling techniques. However, modeling city-level transportation systems is challenging

not only because of the network scale but also because they encompass multiple transportation modes. This

paper introduces a novel simulation framework that efﬁciently supports large-scale agent-based multi-modal

transportation system modeling. The proposed framework utilizes both microscopic and mesoscopic modeling

techniques to take advantage of the strengths of each modeling approach. In order to increase the model sca-

lability, decrease the complexity and achieve a reasonable simulation speed, the proposed framework utilizes

parallel simulation through two partitioning techniques: spatial partitioning by separating the network geo-

graphically and vertical partitioning by separating the network by transportation mode for modes that interact

minimally. The proposed framework creates multi-modal plans for each trip and tracks the travelers trips on

a second-by-second basis across the different modes. We instantiate this framework in a system model of

Los Angeles (LA) supporting our study of the impact on transportation decisions over a 5 hour period of the

morning commute (7am-12pm). The results show that by modifying travel choices of only 10% of the trips a

signiﬁcant reduction in trafﬁc congestion is achievable that results in better trafﬁc ﬂow and lower travel times.

1 INTRODUCTION

The performance of transportation systems is a criti-

cal factor that affects the human standard of life. The

environmental impact of the transportation sector has

major effects on human health (Levy et al., 2010).

Trafﬁc congestion not only increases fuel consump-

tion and emission levels, but also wastes traveler ti-

mes. Moreover, the congestion experienced by trave-

lers increases the stress and affects the individual so-

cial interactions (Boniface et al., 2015). As a result of

all these economic, social, psychological and health

impacts, the academic community has devoted sig-

niﬁcant research efforts to improving transportation

system performance. While the majority of these stu-

dies use simulation (Osorio and Selvam, 2015), (Zehe

et al., 2015), (Zhang et al., 2017), there are signiﬁ-

cant modeling challenges including scaling, calibra-

ting, and validation issues that impact the accuracy of

the results. In this paper, we present a novel agent-

based framework for modeling of large-scale trans-

portation systems. The presented framework supports

city-level networks with different modes of transpor-

tation (cars, buses, railways, walking, biking, and car-

pooling). The proposed framework utilizes both mi-

croscopic and mesoscopic simulation to leverage their

respective strengths of accuracy and scalability. The

framework spatially partitions the network enabling

distinct portions of the region to micro-simulated in

parallel, and vertically partitions the network into lay-

ers represented loosely interacting modes. In this

way, we can utilize the available processing resour-

ces either using single or multiple machines. The

framework is capable of tracking individual trave-

lers on a second-by-second basis from their origin to

their destination across transportation modes. To the

best of our knowledge, the proposed framework is the

ﬁrst tool that supports an agent-based city-level trans-

portation system, combining both microscopic with

mesoscopic simulations, tracking individual travelers

and vehicles on a second-by-second basis, and sup-

porting multi-modal mobility. We instantiate this fra-

Elbery, A., Dvorak, F., Du, J., A. Rakha, H. and Klenk, M.

Large-scale Agent-based Multi-modal Modeling of Transportation Networks.

DOI: 10.5220/0006690301030112

In Proceedings of the 4th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2018), pages 103-112

ISBN: 978-989-758-293-6

103

mework into a system to study the impact of routing

on travel time and fuel consumption in the Greater

LA city from 7am to 12pm. In terms of the paper la-

yout, the paper ﬁrst introduces the related literature.

Because of space limitations, the system is described

brieﬂy. The section following the literature review

provides an overview of the system architecture, com-

ponents, and the high-level operations. Subsequently,

the last two sections demonstrate the case study on the

Greater LA network along with preliminary results.

2 PREVIOUS WORK

The beneﬁts of modeling large-scale transportation

networks have attracted attention over the last three

decades. In 1997, the TRANSIMS simulation tool

(Nagel et al., 1996) was used to simulate the traf-

ﬁc in large areas for trafﬁc planning purposes. The

research work in (Nagel et al., 1996) uses discrete

space modeling for the trafﬁc micro-simulation based

on the cellular automaton approach (White and Enge-

len, 1993), where the road is separated into cells (of

length 7.5 meters) which are either empty or occupied

by one car. It uses a simple algorithms for car follo-

wing and lane changing. The use of cellular automa-

ton makes this system fast, however, it cannot accu-

rately capture observed transportation phenomena in-

cluding car following, lane changing, and gap accep-

tance. In 2002, TRANSIMS was updated to better in-

clude the impact of the congestion on the system per-

formance and it was run on a parallel cluster for ﬁfty

iterations to achieve better trip planning (Cetin et al.,

2002). TRANSIMS has been used to model the Swit-

zerland network in the morning peak hours using pa-

rallel computation (Raney et al., 2003), (Balmer et al.,

2004). Then, in 2012, TRANSIMS was used in (Zhao

and Sadek, 2012) to evaluate the performance of the

transportation network of the Buffalo-Niagara metro-

politan area during signiﬁcant snow events. However,

the authors mentioned that extensive efforts are requi-

red to make the simulated network realistic in terms of

network conﬁguration, lane connectivity, pocket lane

and signal locations. In (Guo et al., 2013) the same

modeler was used to evaluate the impact of dynamic

routing on the fuel consumption. Similar to TRAN-

SIMS, our proposed framework supports the paral-

lel computation either on multi-core or even multi-

ple machines. However, in TRANSIMS the deﬁni-

tion of the microscopic simulation is limited to the

demand, that is, each trip is simulated individually

as an agent. But, the links and the mobility of the

vehicles on these links are modeled using a parallel

queuing approach (Cetin and Nagel, 2002). These

queuing models are inaccurate in estimating the link

travel time especially in congestion situations such as

the LA morning commute. Furthermore, it cannot

capture the accelerations/decelerations events of each

vehicle that have a signiﬁcant impact on the fuel con-

sumption and emissions. In contrast to TRANSIMS,

the proposed framework uses continuous space mo-

del for the micro-simulation, which is the enabler to

capture the many of the mobility parameters. A hy-

brid trafﬁc modeler was presented in (Burghout et al.,

2005), (Burghout and Wahlstedt, 2007), (Yang and

Morgan, 2006), (Balakrishna et al., 2009) to model

large-scale trafﬁc networks. The hybrid modeler si-

mulates different network links with different ﬁdelity

levels (microscopic, or mesoscopic levels), where mi-

croscopic simulation was applied to areas of speciﬁc

interest, while simulating a large surrounding network

in lesser detail with a mesoscopic model. In this way,

it can provide a customized performance and simula-

tion speed. In our proposed system, we also utilize

microscopic-mesoscopic hybrid modeling. However,

in our proposed model, we do not have this spatial

separation between the microscopic and mesoscopic

simulations. In the proposed system, links are as-

signed to the simulator based on their importance in

the network In (Ahn et al., 2012), (Ahn and Rakha,

2013), the authors used the INTEGRATION software

to fully microscopically model the dynamic routing

on the fuel consumption in the downtown Cleveland

and Columbus, Ohio, USA, in the case of different sy-

stem market penetration rates and congestion levels.

The network has about 3,000 links with a trafﬁc de-

mand of 65,000 vehicles per hour during the morning

peak hour. Our proposed framework uses parallel IN-

TEGRATION instances enabling our system to cap-

ture the morning commute of 1.2M vehicles. In 2015,

the authors of (Zehe et al., 2015) proposed the Sca-

lable Electro-Mobility Simulation (SEMSim), an ar-

chitecture for a cloud-based platform, as a proof of

concept to use the cloud for simulation of large-scale

transportation systems. The authors used this model

to simulate the network of Singapore that has about

500,000 private owned vehicle. However, the mo-

del uses simple vehicle characteristics (e.g., kinema-

tic model) and driving behavior models. In contrast

(Zehe et al., 2015), our proposed framework is based

on mature models that have been validated against ob-

served transportation phenomena and supports travel

across different transportation modes. Compared to

the MATSIM (Balmer et al., 2009), which is conside-

red the state of the art in simulating large-scale trans-

portation system, our proposed model is not only an

agent-based simulation. In addition to that, it utilizes

a hybrid simulation approach, it is also capable of mi-

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

104

croscopically simulating all the transportation aspects

including demand, mobility, trafﬁc signals, and road

network aspects, as will be described in next section

3 THE PROPOSED MODEL

In this paper, we redeﬁne the state-of-the-art of mo-

deling and simulation of large-scale transportation sy-

stems by introducing a new framework that is capable

of modeling large city level transportation systems.

To achieve both the required accuracy and scalability,

the proposed model utilizes both microscopic and me-

soscopic modeling techniques. The microscopic si-

mulation deﬁned in this paper includes all the aspects

of simulation: demand, mobility, and network. From

the demand perspective, our framework models each

individual vehicle as an agent in the network that in-

teracts with other vehicles as well as with the trafﬁc

signals and road control signs. It also provides dyna-

mic demand modeling, that is, trafﬁc demand changes

throughout the simulation. From the mobility stand-

point, the proposed framework tracks every indivi-

dual vehicle at a time resolution of deci-second (0.1

seconds). These features are gained basically from

the microscopic nature of the INTEGRATION traf-

ﬁc simulator (Rakha et al., 2012) utilized in propo-

sed model. Based on this time resolution, it captu-

res all the driving events by using validated models

for car following, lane changing and gap acceptance.

The model also can simulate different stochastic mo-

bility phenomena such as stochasticity in speed cal-

culation, route selection and driver aggressiveness in

acceleration/deceleration events. From the network

standpoint, many network topological details such as

link control methods (stop sign, yield sign, and traf-

ﬁc signals), lane striping, lane prohibition, and high

occupancy vehicles (HOV) lanes were modeled in this

framework. To the best of our knowledge, none of the

current trafﬁc simulators support all these features for

large-scale networks. The proposed framework also

incorporates other simulators for the modeling of the

railway, pedestrian and biking travel modes in addi-

tion to buses and carpooling. However, the details of

these simulators are beyond the scope of this paper,

but will be described in more detail in separate papers.

An important advantage of the proposed framework

is its ability to track each trip on a second-by-second

basis across different modes. Because of the compu-

tational cost required for the above-mentioned simu-

lations, the proposed framework uses two partitioning

techniques: vertical and spatial. Vertical partitioning

combines mesoscopic and microscopic road vehicle

simulation along with mode speciﬁc simulations for

walking, biking, and trains. Spatial partitioning divi-

des the microscopic network into smaller geographic

regions. A simulation controller divides each trip into

sub-trips to be simulated in different processes and

monitors each sub-trip to ensure consistency. Before

describing the model, the following subsection gives

some deﬁnitions that will be used in the model.

3.1 Deﬁnitions

Global-network: The global road network includes

all the road links in the area of interest. Each link is

marked to be in either the micro-network, the meso-

network or the train and pedestrian networks.

Meso-network: The meso-network is the connected

subset of the links in the global-network that is simu-

lated mesoscopically.

Micro-network: The micro-network is the connected

subset of links in the global-network that is simulated

microscopically.

Sub-network: A sub-network is a spatial partition of

the micro-network. Sub-networks are simulated mi-

croscopically using INTEGRATION software.

Zones: Nodes that can act as an origin or destina-

tion of the trafﬁc. To have a fully connected network,

each zone in the micro-network is mapped to a corre-

sponding zone in the meso-network. However, some

zones in the meso-network do not exist in the micro-

networks.

Interconnection Zones (IZones): Interconnection

zones are correspondences between zones in different

networks. For example, for the micro-meso network

connectivity, the zones exist in both micro and meso

networks are IZones.

Trip: A trip is a travelers planned path from origin

zone to destination zone in the global-network. A

single trip can go through multiple network layers

(multi-modal trips) and/or multiple sub-networks. For

example, in the trip shown in Figure 1, a person can

drive his/her car on the local road (in meso-network)

from his/her home to the main road. Then, he/she

continues driving on the main road in the micro-

network (where he/she travels through two micro sub-

networks). Then he/she parks his/her car and walks

(on the pedestrian network layer) to the nearest rai-

lway station (rail network) from which he/she takes

the train. Then he/she walks again to his/her work.

3.2 Vertical Partitioning

The main roads, the arterial roads and highways are

the most inﬂuential roadway segments of the city

transportation network. These roads can signiﬁcantly

Large-scale Agent-based Multi-modal Modeling of Transportation Networks

105

Figure 1: Multimode trip example.

affect the network performance. For example, conge-

stion on a main road can affect thousands of vehicles.

On the other hand, the local roads that connect the

main roads to residential areas are of low importance

because of their low trafﬁc ﬂow rates. However, we

cannot totally ignore these local roads because they

contribute to the travel time and fuel consumption of

the vehicles. In the proposed framework, the criti-

cal links (main roads, arterial roads, and highways)

are simulated microscopically which gives the hig-

hest possible ﬁdelity for this portion of the network.

Alternatively, the local roads are modeled mesosco-

pically to capture their impacts while reducing the

modeling and computational requirements. Conse-

quently, the framework has two mandatory layers: the

meso-network and the micro-network layers. In addi-

tion, the framework supports layers for other transpor-

tation modes such as railways and pedestrians. Figure

2 demonstrates the layering concept. A traveler uses

more than one transportation mode means moving

him/her from one network to another, consequently

from one simulator to another. These interactions be-

tween different simulators are managed by a simula-

tion controller (SC). Simulations notify the SC when

a traveler ﬁnishes a sub-trip at an IZone. The SC ﬁnds

the next sub-trip for this traveler and sends him/her to

the appropriate simulator. The IZone must exist in the

next network to guarantee the connectivity of the trip.

Figure 2: Partitioning the network into different layers.

Based on the scale of the area of interest, the

micro-network can be divided into a set of sub-

networks. Each sub-network can run individually. In

this way, we can utilize the available resources, and

increase the simulation speed. There is a trade-off

between modeling speed and accuracy. The larger

the sub-network size, the slower the simulation speed,

but the higher the accuracy of the simulation results.

Within a subnetwork, the simulator models impacts

between all connected road links (e.g., spilling the

vehicles queue back from one link to its upstream

link, and the interactions between vehicles in the in-

tersection). Selecting the appropriate sub-network si-

zes depends on the total network size and the available

computational resources.

3.3 Traveler Types

In our framework, there are two types of travelers;

background and controlled.

Background Travelers: background travelers create

the network trafﬁc conditions in each network, such

as congestion levels in the micro-network and meso-

networks; and vehicle loading on public transit vehi-

cles, etc. In the micro-network, INTEGRATION

tracks every individual traveler (both background and

controlled). In the other simulators, the travelers of

the background trafﬁc are not tracked individually,

instead they are used to estimate the network state

(e.g., congestion levels, train loads) in order accura-

tely calculate the travel time and fuel consumption.

Controlled Travelers: Each controlled traveler repre-

sents a person traveling from an origin to destination

at a particular time. A planner creates a trip for each

controlled traveler and the simulation controller ensu-

res the traveler traverses the networks in the appro-

priate simulators. Each controlled trip is tracked on a

second-by-second basis in all the transportation mo-

des. Moreover, the controlled trip can be re-routed or

re-planned, while the person is traveling.

3.4 System Architecture and

Components

Figure 3 shows the general architecture of the pro-

posed framework. A basic idea is separating the sy-

stem software components from the hardware com-

ponents. The communication layer is the enabler to

transparently run this system on different infrastructu-

res with minimal conﬁguration changes. The commu-

nication layer utilizes the RabbitMQ implementation

(Videla A., 2012) of the Advanced Message Queuing

Protocol (AMQP) (Fernandes et al., 2013). The exe-

cution layer of the system consists of two plans: (1)

the planning and simulation plan which is responsible

for simulating trips and creating the multimodal rou-

tes for the controlled trips (2) the control plan which is

responsible for controlling and managing the different

system components. The framework has the compo-

nents shown in in Figure 4. The input data repository

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

106

Figure 3: System architecture.

Figure 4: System components.

contains all the required input data to be used by the

system components. For example, the road maps for

both the meso-network, micro-network and the sub-

networks are stored in this database along with the

OD inputs that represent the trafﬁc demand, transit

schedules, energy models and transit loading. The

SC manages the different simulations. When star-

ted, each simulation module imports the correspon-

ding input ﬁles from the input database then it initi-

alizes its environment and starts its internal synchro-

nization procedure that communicates to the SC. Due

to space limitation, we will give a brief overview of

the operation for only the basic components including

micro-simulator, meso-simulator, planner and the SC

focused on additions not reported in previous rese-

arch.

3.5 INTEGRATION and Micro-Models

Micro-simulation using INTEGRATION software is

the focus our framework. INTEGRATION is a

discrete-time continuous-space trip-based microsco-

pic trafﬁc simulation and optimization model which is

capable of modeling networks with thousands of cars.

It is characterized by its accuracy that comes from its

microscopic nature and its small time granularity. IN-

TEGRATION provides 10 trafﬁc assignment/routing

options with a full support of ﬁve vehicle classes, each

class has its own parameters and routing trees.

3.5.1 INTEGRATION Car Following Model

INTEGRATION updates the vehicle speed and lo-

cation every decisecond based on a user-speciﬁed

steady-state speed-spacing relationship along with the

speed differential between the subject vehicle and

the heading vehicle. INTEGRATION uses the vari-

able power vehicle dynamics model to estimate the

vehicle’s tractive force. Consequently, it implicitly

accounts for gear-shifting on vehicle acceleration,

which ensures a realistic estimation of the vehicle

acceleration. More speciﬁcally, the model compu-

tes the vehicle’s tractive effort, aerodynamic, rolling,

and grade-resistance forces, as described in details in

the literature (Rakha et al., 2001), (Rakha and Lucic,

2002). In INTEGRATION, the car-following model

computes the speedu

(t +t) of the following vehicle

(n) at the new time step t+t as (Rakha et al., 2012):

(t + ∆t) = min

(

(t) + a

(t)∆t ,

−c

+ c

(t + ∆t) −

√

(n−1)

(t + ∆t)

+ d

max

(

(t + ∆t) −

)

(1)

where

A =



−c

¯s

(t + ∆t)



−4c



¯s

(t + ∆t)u

−c



(2)

and c

, and c

are the model constants which are

computed as:

(2u

−u

) (3)

−u

)

(4)

−

(5)

and the vehicle ¯s

(t + ∆t)spacing is computed as:

¯s

(t + ∆t) = x

(n−1)

(t) −x

(t)+



(n−1)

(t) −u

(t)



∆t + 0.5a

(n−1)

(t + ∆t)∆t

(6)

Here a

(t) is the acceleration of the vehicle n; u

is the free-ﬂow speed of the roadway; u

is the ro-

adway speed-at-capacity; q

is the roadway capacity;

is the roadway jam density; x

(t) and x

(n−1)

(t) are

the positions of the subject vehicle the lead vehicle at

time t; d

max

is the maximum acceptable deceleration

level (m/s

Large-scale Agent-based Multi-modal Modeling of Transportation Networks

107

3.5.2 Delay Computation

Within INTEGRATION, the delay D

experienced by

the vehicle n is computed for each traveled link l ,

as the difference between the vehicles simulated tra-

vel time and the free ﬂow speed speed travel time for

this link (Dion et al., 2004). And the total delay D

experienced by the subject vehicles is computed as:

∑

(l ∈ the vehicle path)

∑

(l ∈ the vehicle path)

−

u(t)

)dt

(7)

where t

and t

are the times at which the vehicle en-

ters and exits the link l respectively.

3.5.3 Fuel Consumption and Emissions

Computing the fuel consumption and emission levels

is important to capture the travel costs and environ-

mental effects of transportation decisions. The IN-

TEGRATION software is capable of computing the

second-by-second fuel consumed; vehicle emissions

of carbon dioxide (CO2); carbon monoxide (CO);

hydrocarbons (HC); oxides of nitrogen (NOx); and

particulate matter (PM). The micro-simulator uses the

VT-Micro model (Rakha et al., 2004) to calculate

the second-by-second fuel consumption and emissi-

ons for each vehicle in the micro-network.

3.6 Meso-simulator

The meso-simulator is implemented as a discrete

event simulation (M.H., 1989). The events represent

the instant of reaching a network node through some

link, at which moment a new event is generated for

the next link and is added to the discrete event queue.

The discrete event queue is an ascending sequence of

events ordered by the time of their occurrence. The

meso-simulator is given paths in the meso-network to

be simulated together with the initial start time. The

ﬁrst event for each path then consists of the ﬁrst node

in the path and the start time, while all other events

are generated as a consequence of the initial event.

In the meso-simulator, each road link has its conﬁgu-

ration parameters such as free-ﬂow speed, speed-at-

capacity, and jam density. In addition, each road link

has state information which includes the number of

vehicles on this link, and a queue that has these vehi-

cles. This state information is updated by the events

happening on the subject link, such as a vehicle enters

the link or a vehicle exits the link. The average speed

and travel time for each individual vehicle are calcu-

lated based on the current state of the link at the time

the vehicle enters that link. The arrival of a vehicle

to a given link triggers the meso simulator to calcu-

late its average speed and travel time, subsequently,

to schedule another event at the time in which vehi-

cle expected to exit that link. At the exit time, the

meso-simulator estimates the fuel consumption of the

vehicle on this link and adds it up to the vehicles total

fuel consumption.

3.7 Planner

Each controlled traveler has an origin, destination and

a travel window. The main task of the planner is the

planning of these multi-modal routes for the control-

led trip. The planner also is responsible for updating

or changing these routes whenever needed. During

a window that begins 30 minutes before the earliest

possible departure time for the controlled traveler, the

planner starts planning the trip by using the up-to-date

cost and timing information reported from each in-

dividual simulator. It also uses the connectivity in-

formation between the different sub-networks and/or

layers in order to create the optimal route for the sub-

ject trip. The trip can be replanned or rerouted after

the trip starts. For example, if the traveler can not ca-

tch the train at the scheduled time, or he/she can not

board the scheduled bus because the bus is full, the

responsible simulator notiﬁes the SC which requests

the planner to ﬁnd an alternative route for the traveler.

3.8 Simulation Controller

The SC is the core component of the model which is

responsible for:

• Initializing the simulation,

• Synchronizing the different simulators,

• Moving travelers between layers/sub-networks,

• Tracking the individual controlled trips.

In the initialization process, the SC reads parame-

ters such as the simulation duration and the locati-

ons of the input ﬁles for each simulation component.

Then it reads in the network ﬁles, builds the requi-

red graphs for the networks, and checks for the ap-

propriate connectivity among the different layers/sub-

networks. It also builds a list of all the controlled

trips. Then, it starts the different simulators (INTE-

GRATION, meso-simulator, bike and pedestrian si-

mulator (BPSim), and railway simulator (RailSim))

and waits for all of them to initialize. When a simula-

tor starts, and initializes its own environment, it must

send the ﬁrst synchronization request to the SC and

wait for the simulation start messages from the SC.

When all the simulators are ready, the SC allows them

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

108

to start the simulation. During the simulation, all the

simulators must be synchronized at pre-speciﬁed in-

tervals. This period is deﬁned as the maximum sy-

nchronization interval, which is a system-wide vari-

able., its defualt value is 1 second. After this time

interval, the simulator can not progress the simula-

tion process until permitted by the SC. When the SC

identiﬁes that all the simulators reached the same si-

mulation time, it allows them to run the next interval.

During the simulation, the SC receives the state in-

formation about each controlled trip or sub-trip from

each simulator. Consequently, it can track every in-

dividual controlled trip in different networks/layers

and is responsible for moving the traveler from one

sub-network/layer to another. By doing so, the SC

establishes the connectivity between different sub-

networks/layers. For example, when a driver ﬁnishes

his/her sub-trip on the meso-network (say, IZone1)

and needs to be moved to the micro network, the meso

simulator informs the SC to 1) update the trip infor-

mation (travel time, fuel consumption, current loca-

tion, , etc.); 2) pull the trip information from its da-

tabase and ﬁnd the destination of the next sub-trip

on the micro sub-network (say, Z2); and 3) request

the corresponding INTEGRATION instance to start a

new sub-trip in its network from IZone1 to Z2 and

passes the initial route for this sub-trip ot INTEGRA-

TION. In this case, INTEGRATION may defer the

start time of this vehicle if the link to which the vehi-

cle should enter is at jam density.

4 CASE STUDY: LA NETWORK

AND PRELIMINARY RESULTS

We use this system to model the overall city of LA

in the peak hours. This section describes the network

and the preliminary simulation results.

4.1 LA Networks

To build the micro-network and meso-network, we

used three different data sources: (1) NavTeq is used

for generating nodes and links, (2) OpenStreetMap

is used for intersection trafﬁc control information,

and (3) Google Maps are used for validating road

attributes including the number of lanes, one-way

streets, speed limits, bus lane locations, etc. The

global-network has 62,984 nodes and 181,840 links,

as shown in Figure 5-a. The LA area is divided into

ﬁve sub-networks shown in Figures 5-b through 5-

f. The walking and biking simulators use the meso-

network as input. Our system model includes the lar-

gest operator of public rail and buses in LA, LA Me-

tro. LA Metro bus service includes 170 lines, 15,967

bus stops, and 854,693 boardings/day. LA Metro rail

service includes 6 passenger rail lines, 93 stations,

and 359,861 boardings/day. Station level boarding

data were provided by LA metro along with speci-

ﬁcations of the vehicle ﬂeet.

4.2 Trafﬁc Calibration

The trafﬁc is created based on real data from Perfor-

mance Measurement System (PEMS) database. The

count and speed data from PEMS database are ag-

gregated and the trafﬁc demand between each Origin-

Destination (OD) pair is estimated using the Queen-

sOD (Aerde et al., 2003) software which utilizes the

Maximum Likelihood Least Relative Error (LRE) ap-

proach. A portion of these trips is used as control-

led travelers, while the remaining are modeled as the

background. The background travelers are modeled in

each network separately based on the calibrated traf-

ﬁc for each sub-network as shown in Table 1. The

vehicle count in Table 1 is the total trafﬁc on each

sub-network that includes both the controlled and the

background trafﬁc.

5 SIMULATION RESULTS

We created the system model to enable assessing po-

tential system-wide effects of individual transporta-

tion decisions across the LA region. We ran two

scenarios for the LA area. In the base scenario, all

the travelers make travel decisions by themselves. In

the controlled scenario, the controlled travelers (10%

of the driving population) are given directions regar-

ding modes and routes by the planner. Our hypot-

heses are that the controlled case will result in an

energy reduction and a reduced network congestion

level. So, in the micro-network routing conﬁgura-

tion, the Sub-population Feedback Eco-Assignment

(ECO) (Ahn and Rakha, 2013) is used for the control-

led trafﬁc, while the Time-Dependent Sub-Population

Feedback Assignment (SFA) (Rakha et al., 2012) is

used for the background trafﬁc. In the base scena-

rio, since it does not have controlled trafﬁc, only the

Table 1: Sub-network sizes.

Sub-net Nodes Links Signals Vehicles

1 743 1691 256 404,191

2 940 2251 361 447,948

3 1625 3561 459 592,343

4 741 1724 237 445,857

5 647 1507 203 362,415

Sum 4696 10734 1516 2,252,754

Large-scale Agent-based Multi-modal Modeling of Transportation Networks

109

(a) LA network (b) Subnet 1 (c) Subnet 2

(d) Subnet 3 (e) Subnet 4 (f) Subnet 5

Figure 5: LA total network and micro-subnetworks.

SFA trafﬁc assignment was run. All the other simu-

lator use the energy as the routing metric. Further-

more, we expect a controlled travel mode distribution

to be dominated by driving in the micro and meso-

networks. The global network trafﬁc calibration sho-

wed that there were approximately 1.3 million vehicle

trips in Greater LA area. When calibrating these ODs

for the micro-networks, it generated 2.25 million trips

as shown in Table 1. The reason for this large diffe-

rence is that a portion of the global trips pass through

multiple sub-networks, being divided into multiple

sub-trips across the sub-networks. Table 2 shows the

system-wide comparison for the traveled distance, we

can notice that the traveled distance decreases for the

micro-network, while it increases for other modes as

the 10% of controlled trips are planned over multi-

ple transportation modalities. Combining the results

in Tables 2 and 3 together, we can notice that in the

micro-network the vehicle’s average traveled distance

Table 2: System wide traveled distance comparison.

Transportation Traveled Distance (Km)

Mode Base Controlled

Walking 0 1,418.6

Cycling 0 147,246.8

Riding Bus 0 3,886.4

Riding Train 0 9,230.5

Driving on Micro 18,298,072.8 17,760,530.8

Driving on Mesok 616,105.2 791,424.0

Carpooling 0 356,549.1

Total 18,914,178.0 19,070,286.2

is about 8 km and the vehicle average travel time is

about 39 minutes in the base case. In the controlled

case, the vehicle’s average traveled distance remains

approximately the same while the vehicle average tra-

vel time is reduced to approximately 23 minutes de-

monstrating that by controlling 10% of the trafﬁc,

the vehicles moving on the main roads and highways

(micro-network) achieved a 40% saving in the total

travel time. Table 3 also shows that those vehicles

achieved about an 18% saving in the fuel consump-

tion, and their average delay is reduced by about 46%.

Shifting the controlled trips to the public transit sy-

stem is energy efﬁcient because the increases in the

energy consumption by the buses and trains due to the

extra passenger loads is less than the savings accrued

as a result of a reduction in the trafﬁc congestion. Ho-

wever, some of these savings come at the cost of an

increment of energy consumption in other transpor-

tation modes. Speciﬁcally, the energy consumed in

both the meso-network and the public transit increa-

sed by 110% from 270028 to 568737 KW-hr. We have

to mention that the system wide comparison of the

fuel/energy is not possible in the current version be-

cause some of energy/fuel consumption models have

not been implemented in the various sub-models.

6 CONCLUSION

The paper proposes and describes a novel multi-

modal large-scale agent-based transportation network

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

110

Table 3: Micro-network results for base and controlled sce-

narios.

Subnet Number of vehicle Trips Total Travel Time (s) Total Delay (s) Fuel Consumed (Liter)

Number Base Controlled Base Controlled Base Controlled Base Controlled

1 404,191 399,298 1,425,241,984 737,314,176 725,780,800 223,516,944 839,149 580,246

2 447,948 399,301 716,592,000 370,059,488 374,176,480 201,135,536 580,276 448,206

3 592,343 539,949 1,466,868,736 907,472,448 702,194,240 437,199,680 901,559 714,016

4 445,857 402,510 1,126,294,656 602,945,728 544,548,864 318,766,464 704,458 544,593

5 362,415 337,392 588,823,424 306,889,216 295,441,600 126,577,376 499,629 383,526

Sum 2,252,754 2,078,450 5,323,820,800 2,924,681,056 2,642,141,984 1,307,196,000 3,525,071 2,670,588

Sub-network Average/veh 2363.25 1407.15 1172.85 628.93 1.56 1.28

modeling system that has a wide spectrum of appli-

cation. The proposed system is capable of modeling

large urban cities including different transportation

modes of travel (driving, biking, walking, riding a

bus, riding a train, and carpooling). This system is

tested by modeling the Greater LA Area during the

morning peak period. The preliminary results show

that the network is currently very congested, with an

average speed of approximately 12.3 km/hr. The re-

sults also show that by re-planning 10% of the trips,

the performance of the network can be signiﬁcantly

improved. An important future effort is to improve

the system to achieve faster simulation speeds. Cur-

rently, the average simulation speed is approximately

half real-time, i.e. every virtual second is simula-

ted in 2 actual seconds. We also plan to improve

the mesoscopic trafﬁc simulator to achieve better esti-

mates of energy consumption, delay, and travel time.

It is also important to study the complexity of the

system by quantifying its simulation speed and me-

mory usage for different demand levels. An advan-

tage of a detailed system model, like the one propo-

sed in this paper, is that it enables modeling mode

changes for different scenarios including trafﬁc inci-

dents, construction, and special events. We intend to

explore potential savings that could result from infor-

med decision-making by groups of travelers in these

scenarios.

REFERENCES

Aerde, M., Rakha, H., and Paramahamsan, H. (2003).

Estimation of origin-destination matrices: Relati-

onship between practical and theoretical considerati-

ons. Transportation Research Record: Journal of the

Transportation Research Board, (1831):122–130.

Ahn, K. and Rakha, H. A. (2013). Network-wide impacts of

eco-routing strategies: a large-scale case study. Trans-

portation Research Part D: Transport and Environ-

ment, 25:119–130.

Ahn, K., Rakha, H. A., and Moran, K. (2012). System-

wide impacts of eco-routing strategies on large-scale

networks. In Transportation Research Board 91st An-

nual Meeting, number 12-1638.

Balakrishna, R., Morgan, D., Slavin, H., and Yang, Q.

(2009). Large-scale trafﬁc simulation tools for plan-

ning and operations management. IFAC Proceedings

Volumes, 42(15):117–122.

Balmer, M., Nagel, K., and Raney, B. (2004). Large-scale

multi-agent simulations for transportation applicati-

ons. In Intelligent Transportation Systems, volume 8,

pages 205–221. Taylor & Francis.

Balmer, M., Rieser, M., Meister, K., Charypar, D., Le-

febvre, N., and Nagel, K. (2009). Matsim-t: Archi-

tecture and simulation times. In Multi-agent systems

for trafﬁc and transportation engineering, pages 57–

78. IGI Global.

Boniface, S., Scantlebury, R., Watkins, S., and Mindell, J.

(2015). Health implications of transport: evidence of

Large-scale Agent-based Multi-modal Modeling of Transportation Networks

111

effects of transport on social interactions. Journal of

Transport & Health, 2(3):441–446.

Burghout, W., Koutsopoulos, H., and Andreasson, I.

(2005). Hybrid mesoscopic-microscopic trafﬁc simu-

lation. Transportation Research Record: Journal of

the Transportation Research Board, (1934):218–255.

Burghout, W. and Wahlstedt, J. (2007). Hybrid trafﬁc si-

mulation with adaptive signal control. Transportation

Research Record: Journal of the Transportation Re-

search Board, (1999):191–197.

Cetin, N. and Nagel, K. (2002). Parallel queue model ap-

proach to trafﬁc microsimulations. In Proceedings of

Swiss Transportation Research Conference.

Cetin, N., Nagel, K., Raney, B., and Voellmy, A. (2002).

Large-scale multi-agent transportation simulations.

Computer Physics Communications, 147(1-2):559–

564.

Dion, F., Rakha, H., and Kang, Y.-S. (2004). Comparison of

delay estimates at under-saturated and over-saturated

pre-timed signalized intersections. Transportation Re-

search Part B: Methodological, 38(2):99–122.

Fernandes, J. L., Lopes, I. C., Rodrigues, J. J., and Ullah, S.

(2013). Performance evaluation of restful web servi-

ces and amqp protocol. In Ubiquitous and Future Net-

works (ICUFN), 2013 Fifth International Conference

on, pages 810–815. IEEE.

Guo, L., Huang, S., and Sadek, A. W. (2013). An evaluation

of environmental beneﬁts of time-dependent green

routing in the greater buffalo–niagara region. Journal

of Intelligent Transportation Systems, 17(1):18–30.

Levy, J. I., Buonocore, J. J., and von Stackelberg, K. (2010).

Evaluation of the public health impacts of trafﬁc con-

gestion: a health risk assessment. Environmental He-

alth, 9(1):65.

M.H., M. (1989). Simulating Computer Systems: Techni-

ques and Tools. MIT press.

Nagel, K., Rickert, M., and Barrett, C. L. (1996). Large

scale trafﬁc simulations. In International Conference

on Vector and Parallel Processing, pages 380–402.

Springer.

Osorio, C. and Selvam, K. K. (2015). Solving large-scale

urban transportation problems by combining the use

of multiple trafﬁc simulation models. Transportation

Research Procedia, 6:272–284.

Rakha, H., Ahn, K., and Trani, A. (2004). Development of

vt-micro model for estimating hot stabilized light duty

vehicle and truck emissions. Transportation Research

Part D: Transport and Environment, 9(1):49–74.

Rakha, H. and Lucic, I. (2002). Variable power vehicle dyn-

amics model for estimating truck accelerations. Jour-

nal of transportation engineering, 128(5):412–419.

Rakha, H., Lucic, I., Demarchi, S. H., Setti, J. R., and

Aerde, M. V. (2001). Vehicle dynamics model for pre-

dicting maximum truck acceleration levels. Journal of

transportation engineering, 127(5):418–425.

Rakha, H. A., Ahn, K., and Moran, K. (2012). Integra-

tion framework for modeling eco-routing strategies:

Logic and preliminary results. International Journal

of Transportation Science and Technology, 1(3):259–

274.

Raney, B., Cetin, N., V

ollmy, A., Vrtic, M., Axhausen, K.,

and Nagel, K. (2003). An agent-based microsimula-

tion model of swiss travel: First results. Networks and

Spatial Economics, 3(1):23–41.

Videla A., J. W. (2012). RabbitMQ in action: distributed

messaging for everyone. MANNING.

White, R. and Engelen, G. (1993). Cellular automata and

fractal urban form: a cellular modelling approach to

the evolution of urban land-use patterns. Environment

and planning A, 25(8):1175–1199.

Yang, Q. and Morgan, D. (2006). Hybrid trafﬁc simula-

tion model. In Transportation Research Board 85th

Annual Meeting, number 06-2582.

Zehe, D., Knoll, A., Cai, W., and Aydt, H. (2015). Semsim

cloud service: Large-scale urban systems simulation

in the cloud. Simulation Modelling Practice and The-

ory, 58:157–171.

Zhang, C., Osorio, C., and Fl

otter

od, G. (2017). Efﬁcient

calibration techniques for large-scale trafﬁc simula-

tors. Transportation Research Part B: Methodologi-

cal, 97:214–239.

Zhao, Y. and Sadek, A. W. (2012). Large-scale agent-based

trafﬁc micro-simulation: Experiences with model re-

ﬁnement, calibration, validation and application. Pro-

cedia Computer Science, 10:815–820.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

112