DSpace at VNU: An agent-based model for simulation of traffic network status: Applied to Hanoi city

Simulation: Transactions of the Society for Modeling and Simulation International 1–14 Ó The Authors 2016 DOI: 10.1177/0037549716668259 sim.sagepub.com An agent-based model for simulatio

Simulation: Transactions of the Society for Modeling and Simulation International 1–14

Ó The Author(s) 2016 DOI: 10.1177/0037549716668259 sim.sagepub.com

An agent-based model for simulation

of traffic network status: Applied to

Hanoi city

Abstract

In recent times there have been many agent-based simulation models proposed for transportation network simulation Intuitively, these models are applicable for the transportation simulation of modern cities in developed countries where the transportation network is very well organized: Roads are separated into lanes; most transportation occurs using vehicles; and almost all drivers respect the circulation rule However, these models are not suitable for developing coun-tries where the transportation network is unorganized The reasons for this are as follows: (1) there is no lane on the road; (2) most vehicles are motorbikes; and (3) not many drivers respect the circulation rule This paper introduces an agent-based model for transportation network simulation in which each form of transport is modeled as an agent exhi-biting full features of unorganized circulation behavior The model is designed for a large scale and instead of displaying the circulation for all individual modes of transport, the model displays only the status of the traffic network The simula-tion of circulasimula-tion is therefore considered as a background process The simulasimula-tion is launched and the results are obtained before being displayed This model is applied to the traffic network of Hanoi to analyze the hot or bottle neck points on the transportation network during rush hours in the city

Keywords

Traffic network modeling, transportation network simulation, simulation model, multiagent system

1 Introduction

Transportation network simulation is an active research

field that has attracted many researchers in recent times

Consequently, many models and tools have been proposed

to date Most of them are agent-based models: Each form

of transport could be modeled as an intelligent agent

which is able to observe other forms of transport and

obstacles to avoid and/or change its own speed and

direc-tion in order to reach its destinadirec-tion as fast as possible

The transportation network therefore could be modeled as

a multiagent system where each agent has some personal

attributes such as: physical size, goal (its destination to

go), destination route plan, etc Circulation behaviors also

include: go to the destination, avoid obstacles, find the

shortest or the fastest path, share traffic information to

other agents, etc There are many agent-based models

pro-posed in the literature.1–21

Technically, most of these models are intentionally

designed for an ideal circulation situation for developed

countries: roads are partitioned into lanes; the majority of

modes of transport are cars, buses, trains, etc., and the

drivers mostly respect the circulation rule (see our

sum-mary in Table 1)

However, these basic assumptions make these models unsuitable for developing countries where the transporta-tion network is unorganized because (see the Figure 1 of the traffic situation and the circulation culture of the Vietnamese):

 most of the streets have no lanes and no one respects the rule that drivers have to follow only one lane on a street;

 the majority of vehicles are motorbikes;

 drivers do not always respect the circulation rule: they will move their vehicle anywhere as long as there is enough space for them

1

Posts and Telecommunications Institute of Technology, Hanoi, Vietnam

2

Vietnam National University in Hanoi, Hanoi, Vietnam

3

UMI UMMISCO 209 (IRD/UPMC), Paris, France

Corresponding author:

Manh Hung Nguyen, Faculty of Information Technology, Posts and Telecommunications Institute of Technology, Km10, Nguyen Trai, Ha Dong, Ha Noi, Vietnam.

Email: mhnguyen@ptit.edu.vn

Following the theme of our previous work,22,23this paper presents a simulation model and tool applicable for devel-oping countries where the transportation network is unor-ganized, and it is then applied to the traffic situation and the circulation culture of Vietnam The model is also based

on multiagent systems, in which each mode of transport can be modeled as an intelligent agent which can observe other vehicles and obstacles to avoid and/or change its own speed as well as direction to reach its destination as fast as possible However, in order to increase the ability

to consider a large number of agents in the system, instead

of displaying all the vehicles and their trajectories, the pro-posed model shows only the status of the streets which represent the actual level of congestion for each street and for the overall network All of the remaining calculations for the vehicle agents are carried out in the lower level of the model

This paper is organized as follows: Section 2 presents our model Section 3 presents the modeling of the agents

in the model Section 4 presents the application of the model to simulate the traffic network of Hanoi Finally, Section 5 gives a conclusion and draws some perspectives for future work

2 Agent-based model for simulation of traffic network status

The model is divided into two main levels (Figure 2): First, the low level which contains mostly processing and calculations This simulates the real transportation network

Table 1 Summary of recent proposed models regarding the current Vietnamese circulation situation.

Figure 1 Differences in the circulation situation in developed

countries and Vietnam (a) Circulation in Europe (source:

www.pressoffice.pl) (b) Circulation in Vietnam (source:

www.sgtt.vn).

throughout the day and then calculates the level of

conges-tion in the streets of the network This data could be saved

into storage and then displayed at a later time Second, the

high level, which plays only the role of displaying the

simulation data that is already calculated at the low level

and stored in storage This section presents the main steps

in both of the levels The modeling of vehicles will be

pre-sented in the next section

Step 1: Load GIS files

This step reads the input GIS (Geographic Information

System) data to create the road network The GIS data

rep-resents the realistic data from the real road network In the

input GIS data, each road is characterized by its:

 ID: road identification;

 name: road name;

 direction: one way or dual ways;

 permitted vehicles: types of vehicle which could

circulate on it;

 capacity: the width or through put of road;

 lanes: number of lanes

Step 2: Initiate agent position

This initiates the start place of the vehicles The

loca-tion of a vehicle is determined by a point within a zone

Thus the number of agents in a specific zone z is

deter-mined as follows:

n(z) = a I(z) s(z)

where a is the simulation ratio regarding the real size of

the population; I (z) is the density of the population in the

zone z; s(z) is the surface of the zone z; S is the overall

surface of the simulation zone (entry city); N is the total population of the simulation city

Step 3: Generate agent plans

This step creates plans for the vehicle agents A plan contains the following information:

 start time: the start time to begin movement;

 departure: the start place (longitude (x), latitude (y)) for the movement path;

 destination: the finish place for the movement path;

 max speed: the maximum speed of the vehicle;

 type of vehicle: the type of vehicle on the road

A vehicle could have several plans with different desti-nations and start times

Step 4: Simulation

A vehicle is simulated by an agent which circulates based on its daily plans The actual position of an agent on the road is determined by two factors:

 its movement path: this is dynamically determined

by an optimal algorithm (modeled in Section 3.1);

 its speed: the actual speed of the agent, which could

be changed by the following rules:

– accelerate: when there is neither an obstacle nor a red light in front of it and its speed is not maximal yet;

– no change: when there are some obstacles in front of it such that it can not pass, or its speed

is already maximal;

– slow down: when there are many obstacles in front of it with decreasing speed

Step 5: Observation and sampling This step observes the actual throughput of streets in the network at some moments in time (called sampling time), and then saves the data into storage In our case study of the Hanoi traffic network, we categorise four kinds of road status: blocked, very slow, slow, normal speed

The threshold speed for each status can be estimated by using a Regular Increasing Monotone (RIM) linguistic quantifier Q (Zadeh24) In our case study we use Q(x) = x3 In Vietnam, the circulation policy and rule limit the speed in urban zones to 40 km/h for motorbikes, and 50 km/h for cars, taxis, trunks and buses The actual speed of a crowd is estimated as the speed of the slowest member in the crowd So we use the limiting speed of a motorbike as the reference speed (max speed in the fourth row of the Table 2) for our estimations Consequently, the threshold speed for each status is estimated as presented in Table 2:

 blocked road: the speed is less than 1 km/h;

Figure 2 The two levels, including their main steps in the

proposed model.

 very slow road: the speed is between 1 km/h and 5

km/h;

 slow road: the speed is between 5 km/h and 17 km/h;

 normal road: the speed is higher than 17 km/h

Step 6: Display road/network status This step reads the

traffic status data for all of the streets from the storage

carried out in step 5 and then displays the data on the

net-work Each kind of traffic status is represented by a color:

 red: the road (intersection) is currently blocked;

 orange: the road is almost full, the movement of

vehicles is very slow;

 yellow: the road contains many vehicles, the

move-ment of vehicles is slow;

 gray or green: the road has a normal traffic, the

movement is normal

3 Agent modeling

This section presents the modeling of two kinds of agents

in the system: the vehicle agent and the traffic light agent

Each agent is classified using two parts: attributes, and

behaviors

3.1 Vehicle agent

This agent represents a transport such as a trunk, a bus, a

car (including taxi) or a motorbike All of these vehicles

are controlled by a driver Therefore, we do not need to

separately model drivers and can instead consider the

driver and his vehicle as a unique entity, called a vehicle

agent

3.1.1 Attributes A vehicle agent has the following

attributes

1 Name: name or type of vehicle

2 Length (denoted as l): the physical length of the

vehicle

3 Width (denoted as d): the physical width of the

vehicle

4 Max speed (denoted as ymax): the maximal speed which is allowed by law

5 Current speed (denoted as v): the actual speed of the vehicle at a moment in time

6 Max technical speed (denoted as ytech): the maxi-mal technical speed that a vehicle engine could reach

7 Safety front distance (denoted as df): the minimal distance to the nearest vehicle in front (or behind)

of it such that the movement is still safe This dis-tance is calculated as follows:

df=l y

where u is the minimal distance allowed between two consecutive vehicles when at rest

8 Safety beside distance (denoted as db): the mini-mal distance to the nearest left/right side of the vehicle such that the movement is still safe This distance is calculated as follows:

db=d y

9 Acceleration factor (denoted as a): the increase

of speed in a unit of time

10 Deceleration factor (denoted as b): the decrease

of speed in a unit of time

11 A set of circulation plans which represents the route plan of the vehicle in a daytime

The formulas for df and dbsay that the higher the actual speed of the vehicle, the greater the safety distance must

be between vehicles

3.1.2 Behavior In order to simulate the unorganized circu-lation activities such as that in the situation of Vietnam,

we need to distinguish two kinds of vehicle agent: obedi-ent and disobediobedi-ent The obediobedi-ent vehicle agobedi-ent is the one who respects completely the circulation laws This agent has the following activities

1 Find a path: finds a path to go to its destination This activity occurs when: either the agent starts a new plan; or it has to change the path when its cur-rent path is blocked somewhere

2 Observation: captures the changes in its environ-ment such as observing traffic lights and detecting other obstacles to avoid during circulation

3 Stop: stops at an intersection when the traffic light

is red or at the destination of its plan

4 Accelerate: increases its speed This activity occurs when: (1) there is enough space in front of

it to accelerate; and (2) its current speed could be

Table 2 Threshold speed for road status.

Threshold

speed (km/h)

increased in a low speed regime The new speed

will accelerate to:

where yt, yt + 1 are the speed of the vehicle agent

at the simulation steps t, t + 1, respectively

5 Decelerate: decreases its speed This occurs when:

(1) there is some obstacles in front of it; or (2) it

wants to stop somewhere The new speed will

decelerate to:

The disobedient vehicle agent is the one who usually

breaks the circulation law This agent has some differences

in the following activities:

1 Accelerate: it increases its speed as long as there is

enough space in front of it and its speed is still lower

than the ytech The new speed will accelerate to:

where yt, yt + 1 are the speeds of the vehicle at the

simulation steps t, t + 1, respectively

2 Decelerate: the same principle as with the obedient

agent

The transition among activities of a vehicle agent is

pre-sented in Figure 3: When the time is the same as the start

time of a plan, the vehicle agent starts to move Firstly, it

finds a path to go its destination During movement, it

cap-tures the occurrence or the change of status of three kinds

of object: the traffic light, the destination, and the obstacle

At an intersection, it observes the change of traffic light color: if the light is red it will stop; otherwise, it will con-tinue to move At any point on the path, it will accelerate

if there is no obstacle in the street and its actual speed y is still lower than the allowed speed ymax In contrast, it will decelerate if there are some obstacles or its speed y is already higher than the allowed speed ymax(for the case of

an obedient vehicle agent) Otherwise, it continues to travel with the same speed It could re-find the path if it was blocked somewhere If it is already at the destination,

it stops

3.1.3 Dynamic optimal path finding There have been many models proposed recently to optimize the routing of the whole traffic network by searching for the best path for each individual such as the models by Deng et al.,25Hua and Pei,26Nannicini et al.,27Efentakis et al.,28This paper introduces an algorithm to dynamically optimize the rout-ing at the system level of a transportation network In this algorithm, each individual will be recommended to follow

a path which depends on the current traffic capacity and the user preferences such that the overall throughput of the traffic network is maximized as well as the user prefer-ences being satisfied

Construction of the graph for a traffic network The graph for a given traffic network will be created as follows:

1 Each intersection forms a node of the graph

2 Each road forms an arc with the same direction If

a long road has many segment points where the drivers could change direction, then each segment

Figure 3 Behavior of the transport agent.

of road (between two consecutive segment points)

will form an arc with the same direction

3 Let yavg be the average speed of circulation

(yavg= average(yk), for8k on the same road from

node i to node j), lengthijbe the length of the road

from node i to node j, then the necessary time tijto

pass through the road from node i to node j will be

estimated as follows:

tij=lengthij

In fact, the weight of an arc could be a value which is

pro-positional with one of two options: (i) the length of the

corresponding road; (ii) the necessary time to pass the road

corresponding to the arc We use the second approach

because it takes into account the dynamic nature of the

traffic network whereas the first one is always static

Therefore, we call the best path from an original point to

an end point the fastest path instead of shortest path

Static fastest path

The objective of this step is to estimate the fastest path

for the ideal circulation conditions: we can drive with

maximum allowed speed on all roads in the network:

y= ymax, so yavg= ymax With regards to Equation (7),

when yavg= ymax, then tij becomes constant That is why

we call this the SFP - Static Fastest Path

User preference

We aim to take into account the variations in time

con-straints among drivers by modeling the user preference

Thus, the preference on the time constraint of a driver is

represented by a number K, K 1: The maximal

accepta-ble time of this driver is K SFP (K times the static fastest

path in the ideal conditions) The bigger K, the less

impor-tant the user places on time; and vice versa, the smaller K,

the more strictly time is constrained for the user In

respecting the user preferences, the dynamic fastest path

has to be shorter than K SFP This idea also helps us to

prevent too many people from being recommended to

fol-low the same road because their preferences are different

regardless of it they have the same origin and end point

Dynamic path optimization algorithm

Let o2 G be the original node, e 2 G be the destination

node, H G be the set of visited nodes, U  G be the set

of un-visited nodes, d(i, j) be the fastest time from node i

to node j and P(i) = j represents the node j that is the node

just before node i on the fastest path from the original node

o to node j The algorithm is as follows:

1 Initiation Put node o into H, d(i, j) = tij with tij

estimated from Equation (7), P(i) = 1

2 Repeat Considering all nodes j2 U where there is

at least one node l2 H which has a direct arc from

k to j and d(o, l) + tlj4 K SFP:

 if d(o, j) d(o, l) + tlj then: P(j) = l and d(o, j) = d(o, l) + tlj;

 put j into G, remove j from U;

 if e is already in G or U is empty then stop

3 Constructing of the fastest path From P(e), find P(P(e)) until o is reached Reverse this series and we will obtain the track of the fastest path from

o to e

3.2 Traffic light

This agent represents a traffic controller at an intersection

of the traffic network This agent can not move It only has the ability to change the color of the traffic light to route the traffic

3.2.1 Attributes A traffic light has the following attributes:

1 A set of control directions: These are all directions which are designed at an intersection (e.g., Figure 4 presents four control directions for an intersection)

2 Each control direction has the following:

 green duration (denoted dg): This is an inter-val of time During this interinter-val, the light is green and the vehicles can go through the intersection

 red duration (denoted dr): This is an interval

of time During this interval, the light is red and the vehicles have to stop

 minimal green duration (denoted MinT): the green duration can not be lower than this threshold to avoid the case that the light color changes too quickly to follow

 maximal green duration (denoted MaxT): the green duration could not be longer than this threshold to avoid the case that the remaining control directions have to wait too long

 time counter (denoted counter): this agent changes the color of the light when the value of this variable reaches the green or red duration 3.2.2 Behavior A traffic light has three behaviors:

1 Change to green: this activity occurs when the cur-rent light is red and the counter value is equal

to dr After changing to green, the counter value is reset to zero

2 Change to red: this activity occurs when the light

is green and the counter value is equal to dg After changing to red, the counter value is also reset to zero

3 Optimization of green/red time: this agent dynami-cally updates the dg and dr to optimize the traffic routing at its intersection The algorithm will be presented in the next section

The transition among activities of a traffic light agent is

presented in Figure 5 Firstly, it initiates the value of dr,

dg, and the color for each of the control directions Then

the transition between change to red and change to green

is controlled by the value of counter, by comparing it to

the value of drand dg During this transition loop, the

func-tion of Optimizafunc-tion of green/red time is also evaluated

3.2.3 Dynamic optimal routing for traffic light There are

many models proposed in recent times to optimize the

routing at an intersection of a transportation network, such

as the models by Burguillo-Rial et al.,29 Royani et al.,30

Mehan and Sharma,31 Popescu et al.,32 Gershenson and

Rosenblueth.33This paper introduces an algorithm to

dyna-mically optimize the routing at an intersection of a

transportation network In this algorithm, the next green time of a direction is estimated based on the rate of modes

of transport which pass through the intersection during the last green time

In the case of an independent light direction, the main steps in the algorithm are the following:

1 The green light time must not be lower than a min-imal threshold MinT , and not higher than a maxi-mal threshold MaxT

2 In all cases, the light is set to green if the destina-tion direcdestina-tion still remains for the mode of transport

3 Let nt1, nt be the number of modes of transport which passed the green light during the (t 1)th, tth green light times, respectively

4 Let dtbe the green light time at t

5 The green light time at (t+1) will then be:

dt + 1= min max (dt nt

ð8Þ

In the case of several light directions, the main steps in the algorithm are as follows:

1 The green light time must not be lower than a min-imal threshold MinT , and not higher than a maxi-mal threshold MaxT

2 In all cases, the light is set to green if the destina-tion direcdestina-tion remains for the mode of transport

3 Let k be the number of control directions at the intersection

4 Let nt(i) be the number of modes of transport which passed the green light, for the control direc-tion i, during the tth green light time

5 Let dt(i) be the green light time at t of the control direction i

6 The green light time at (t + 1) will be:

minj = 1(nt(j)), MaxT

ð9Þ

7 The red light time at (t + 1) will be:

dt + 10 (i) = Xk

j = 1, j6¼i

4 A case study: Simulation of the traffic network of Hanoi city

This section presents a case study in which the proposed model is applied to simulate the traffic network of Hanoi (the capital of Vietnam)

Figure 4 Four control directions at an intersection.

Figure 5 Behavior of the traffic light agent.

4.1 Simulation setup

This section presents the initiation of the agent population

and the agent parameters and the construction of the agent

plan

4.1.1 Initiation of agent population The position of agents is

initiated based on the realistic population distribution of

Hanoi, grouped by districts (Table 3; the statistical

num-bers were collected from multi sources in 2013) For

instance, if we take the simulation rate as 1:100, there will

be about 23,000 agents in the nine central districts In

which there will be about 2200 agents living in Ha Dong

district, and 2500 other agents living in Thanh Xuan

dis-trict, etc

4.1.2 Initiation of vehicle parameters The initial conditions

for the vehicle parameters is presented in Table 4 The

val-ues of size, max technical speed, accelerate factor, and

decelerate factor for each kind of vehicle are assigned

based on the majority of vehicles which circulate in

Vietnam The value of max speed, and minimal distance

allowed are defined based on the Vietnamese circulation

law for urban zones

4.1.3 Construction of agent plans The data about the

distri-bution of buildings in the city such as offices, hospitals,

schools, universities, tourist sites, commercial centers,

manufactures, etc is loaded from a GIS file (Figure 6)

This is the input data used to build the vehicle agent plan:

its home position, its jobs will determine an office or a

school For instance, consider a woman officer who lives

in Thanh Xuan district She uses her car to take her son to

school before 7 a.m at Kim Lien quarter Then, she goes

to her office in Ba Trieu street before 8 a.m At the end of

day, she leaves her office at 4:30 p.m and gets to her son’s

school to collect him at 5 p.m after which the mother and

son go back home together (Figure 7)

4.2 Results

This section presents some results of the simulation con-sidering several aspects: runtime, validation of Vietnamese circulation behavior, displaying the traffic network at dif-ferent levels, analyzing the statistics, and the validation of the dynamic optimum of path finding

4.2.1 Simulation runtime The simulation runtime depends

on the configuration of the running machine, as depicted

in Table 5 The chosen simulation rate is 1:100 This means that the number of vehicle agents is about 23,000 Consequently, the simulation time on a normal PC, for a

Figure 7 Representation of an individual’s plans in XML format.

Table 3 Population distribution by central districts.

quarters

Surface (km2) Population

(1000 people)

Figure 6 Distribution of buildings in Hanoi.

Table 4 Initial value of vehicle agent’s parameters.

Max speed (max) 40 km/h 50 km/h 50 km/h Max technical speed 120 km/h 150 km/h 150 km/h

Accelerate factor (α) 0.5 m/s2 0.5 m/s2 0.5 m/s2 Decelerate factor (β) 0.5 m/s2 0.5 m/s2 0.5 m/s2

whole day (24 h) of simulation, is about 15 hours.

Meanwhile, the simulation time on a server with high

con-figuration and performance is about 2 hours Independent

of the machine configuration, the display time in both

cases is only about 3 minutes, including the delay time

after each sample status display The display time is much

less than the simulation time because it simply reads the

simulation results (which are obtained from a long

simula-tion time and already saved into a file) and then displays

them on the interface

These results confirm the advantage of our approach by

separating into the two levels: simulation and display

Despite the length of the simulation time and the

complex-ity of the simulation system, the display time could be

short enough to display as a real-time simulation

4.2.2 Validation of Vietnamese circulation behavior In order to

validate the Vietnamese circulation behavior in the model,

we consider the following scenario:

 track the circulation path of an obedient driver at a

low traffic moment;

 track the circulation path of a disobedient driver at

a high traffic moment;

 display the two tracked paths on the passed streets

to see their differences

The results are presented in Figure 8 There is a big

dif-ference between the two tracked paths The path of the

obedient driver is much more smooth than that of the

dis-obedient driver The path of the disdis-obedient driver is less

structured because the driver adopted the Vietnamese

cir-culation behavior during high traffic moments: he did not

respect the circulation law, the signal of the traffic lights,

as well as the circulation lane He moved forward as long

as there was enough space in front of him to do so

This path illustrates the circulation behavior of many

Vietnamese drivers in the high traffic moments which is

modeled in the proposed model

4.2.3 Displaying of traffic at the local level This model also

enables us to show the traffic situation on a smaller scale

for the traffic network, such as the traffic in a particular street, at an intersection, or at any hot point of the traffic network For instance, the traffic at an intersection is pre-sented in Figure 9 for several situations: the traffic in low traffic (at 6 a.m.) or a high traffic moment (at 7:20 a.m.) and the traffic when the traffic lights work or do not work

It is easy to see that, with the same traffic conditions, the circulation when the traffic lights work is better than that when the traffic lights do not work The circulation in

a low traffic moment is better than that in rush hour The circulation in rush hour without traffic lights (Figure 9(d) simulates completely the real traffic situation depicted in Figure 1(b)

4.2.4 Displaying of traffic status at the overall level The results of traffic network status are presented in Figure 10

At 6 a.m., there is no red or orange street and there are only some streets in yellow because it is not a rush hour yet At 7:20 a.m., it is a rush hour as there are at least five

Table 5 Simulation configuration and runtime.

1 day simulation time ∼ 15 hours ∼ 2 hours

1 day display time ∼ 3 minutes ∼ 3 minutes

Figure 8 Circulation path for the obedient and disobedient drivers (a) Obedient driver’s path at 6:00 a.m (b) Disobedient driver’s path at 7:20 a.m.

Figure 9 Traffic, in local view, at different moments during the

day (a) 6:00 a.m., with traffic lights (b) 6:00 a.m., without traffic

lights (c) 7:20 a.m., with traffic lights (d) 7:20 a.m., without

traffic lights.

Figure 10 Traffic status, in global view, at different moments during the day (a) 6:00 a.m (b) 7:20 a.m (c) 12:00 p.m.

(d) 06:00 p.m.