LANE CHANGING MODEL WITH EARLY

COMMUNICATION OF INTENTIONS

Tomoki Takasago

Graduate School of Science & Technology, Kwansei Gakuin University, 2-1, Gakuen, Sanda, 669-1337, Japan

Kazuko Takahashi

School of Science & Technology, Kwansei Gakuin University, 2-1, Gakuen, Sanda, 669-1337, Japan

Keywords:

Changing lanes, Multi-agents, Intelligent transportation systems, Cooperative driving, Agent simulation.

Abstract:

This paper describes the modeling and simulation of trafﬁc ﬂow with lane changing based on inter-vehicle

communications. We regard vehicles as autonomous agents, and construct a simple trafﬁc model in which

some agents change lanes. We propose a new method for an agent to change lanes in which it tells the neigh-

boring agents its intention beforehand, shares information with them, and determines the lane change time in

cooperation with the other agents. The result of simulating this model showed that the early communication

of intentions is effective in avoiding trafﬁc jams and supporting smooth transportation.

1 INTRODUCTION

Research into alleviating trafﬁc jams and smoothing

trafﬁc ﬂow is receiving increased attention due to the

advances in Global Positioning System (GPS) devices

and car navigation systems.

One reason for trafﬁc jams and collisions is a lack

of understanding between vehicles, meaning that one

vehicle does not know what other vehicles intend to

do. In this research, we focus on lane changing in

which mutual understanding and cooperation is es-

sential.

Lane changing occurs frequently on roads with

multiple lanes, crossings, or highway entrances and

exits, and changing lanes smoothly becomes more

difﬁcult if the road is congested. Signal ﬂashers are

one method of indicating lane-changing intention ex-

plicitly, but they are not always easily noticed on a

congested road. This results in aggressive lane chang-

ing or vehicles making mutual concessions to lose a

good timing.

In the ﬁeld of intelligent transportation systems

(ITS), research on intelligent driving with inter-

vehicle communication using trafﬁc simulation tools

has been a topic of some interest (Kato et al., 2002;

Ammoun et al., 2007; Kanaris et al., 2001). The aim

of such research is safe driving without trafﬁc acci-

dents, and the mechanism for cooperation and inter-

action between vehicles is generally not considered.

In the ﬁeld of artiﬁcial intelligence (AI), several

trafﬁc models that consider vehicles as agents have

been proposed, but few of them examine lane chang-

ing with communication. Ehlert and Rothkranz mod-

eled a vehicle using an agent that determines its be-

havior reactively based on the environment, but no

communication occurs between agents (Ehlert and

Rothkrantz, 2001). Dresner and Stone developed

a model in which vehicles as well as intersections

are modeled as agents that communicate with each

other (Dresner and Stone, 2005).

Nagel and Schreckenberg proposed the cell au-

tomaton model, which simulates trafﬁc ﬂow (Nagel

and Schrechenberg, 1992). Since then, many studies

have used cell automatons to model trafﬁc ﬂows that

include lane changing (Su et al., 2005; Jin et al., 1999;

Knospe et al., 1999). The main goal of this series of

works was to create a trafﬁc model that simulates a

realistic trafﬁc ﬂow, not to model the cooperation and

interaction between agents.

Changing lanes is usually performed as follows:

ﬁnding a gap of sufﬁcient size between vehicles in the

target lane, activating the appropriate signal ﬂasher,

and changing lanes. Gipps analyzed the procedure of

lane changing and proposed a basic model (Gipps,

1986). According to Gipps, lane changes occur for

several reasons:

455

Takasago T. and Takahashi K. (2009).

LANE CHANGING MODEL WITH EARLY COMMUNICATION OF INTENTIONS.

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence, pages 455-461

DOI: 10.5220/0001653804550461

 SciTePress

• the physical possibility of safely changing lanes

without an unacceptable risk of collision

• the location of permanent obstructions

• the presence of special-purposelanes such as tran-

sit lanes

• the driver’s intended turning direction

• the presence of heavy vehicles

• the possibility of gaining a speed advantage

Hidas showed that Gipps’s model is not appli-

cable on congested roads, and proposed a new one.

He noted that lane changing is either a cooperative

or a forced maneuver on congested roads (Hidas,

2002). He assumed inter-vehicle communication, and

showed that the results of a simulation based on his

model coincided with observed trafﬁc data (Hidas,

2005).

Inter-vehicle communication related to lane

changing (i.e., the indication of the intent to change

lanes), however, occurs immediately before the vehi-

cle starts the lane-change action in the models pro-

posed so far. Therefore, lane changing is unsafe

and requires time when the target lane is congested.

Moreover, only the cooperation of the vehicles in the

target lane after the communication is considered, not

the behavior of the subject vehicle before the com-

munication. In fact, the subject vehicle may change

speed to ﬁnd an appropriate space to move into be-

fore actually changing lanes. In addition, when two

vehicles in almost the same position in neighboring

lanes wish to change lanes at the same time, deadlock

will occur, and this issue is not considered so much.

Consider a speciﬁc example. A vehicle in the left

lane wishes to move to the right lane, while another

vehicle right next to it in the right lane wishes to move

to the left lane at exactly the same time as shown in

Fig. 1. This is considered to be a type of deadlock.

Both vehicles might take exactly the same action, that

is, seek enough space in the adjacent lane only to dis-

cover that this does not exist, then decrease speed and

wait for another chance. As a result, neither vehicle

may ﬁnd a suitable opportunity for a long time.

Consider another situation. One vehicle intends

to change lanes, and sufﬁcient space exists in the tar-

get lane, but the vehicle ahead decelerates (Fig. 2),

possibly in preparation for a lane change of its own.

However, the subject vehicle does not know the rea-

son for the deceleration and cannot judge whether it

should change lanes immediately or continue at re-

duced speed for reasons of safety.

Consider yet one more situation. The subject ve-

hicle intends to change lanes. As it is just about to do

so, the preceding vehicle in the adjacent lane acceler-

ates as shown in Fig. 3, and the vehicle behind it also

Figure 1: Deadlock case.

Figure 2: Indeterminate case.

Figure 3: Blocked case.

accelerates without knowing why the vehicle ahead

has sped up. This reduces the gap, and the subject

vehicle loses its chance to change lanes.

All these problems occur because the vehicles do

not know the reason for the actions of the others.

In this paper, we propose a model in which the in-

tentions of vehicles are shared, not just before chang-

ing lanes, but even earlier so that both the subject ve-

ICAART 2009 - International Conference on Agents and Artificial Intelligence

456

hicle and the vehicles in the target lane have sufﬁcient

time to prepare. Thus, lane change can take place

smoothly at the appropriate time. The vehicles all un-

derstand the intentions of others because they share

information and cooperate with each other to allevi-

ate trafﬁc jams and collisions.

This paper is organized as follows. In section 2,

we present the basic lane-changing model without

inter-vehicle communications, and in section 3, we

propose the model with early communication. In sec-

tion 4, we describe and evaluate the experimental re-

sults. In section 5, we compare our method with re-

lated work. Finally, in section 6, we present our con-

clusions.

2 BASIC MODEL

2.1 Road

A road is deﬁned as a sector between signals that con-

sists of two straight lanes. Each sector is divided into

three zones as shown in Fig. 4: entrance, middle and

exit. The entrance is the ﬁrst zone from the start line,

the exit is the ﬁnal zone before the end line, and the

middle is the zone in-between. Lane changing occurs

only in the middle and exit zones, so that the model

can reﬂect the effect of lane changing on the cars be-

hind. (If lane changing were permitted to occur in

the entrance zone, the start time of vehicles would be

delayed, and congestion would not occur.)

middle

exit

entrance

end

start

Figure 4: Road model.

2.2 Vehicle

Each vehicle is modeled as an autonomous agent

that has its own intention and determines its behav-

ior by communicating with other vehicles. The ac-

tion of changing lanes affects the behavior of the other

agents.

Each agent has two goals: to accomplish its own

intention and to contribute to achieving the goals of

the others. These goals, however, sometimes oppose

each other. Even if no such opposition exists, a behav-

ior may be good for one agent but not for the others.

The vehicles immediately ahead and behind the

subject vehicle in the current lane are called the cur-

rent preceder and the current follower, respectively.

The vehicle just ahead and just behind the subject ve-

hicle in the other lane are called the new preceder and

the new follower, respectively, as shown in Fig. 5.

1: current preceder

2: current follower

3: new preceder

4: new follower

0: subject agent

Figure 5: Agents involved in a lane change.

Each agent starts when a sufﬁcient distance is

made between it and the immediately preceding

agent. It has its own goal lane, and if the goal lane

is different from the initial lane, it intends to change

lanes. The velocity of each agent is one of stop, slow,

normal, and fast. This is determined depending on

the distance from the immediate preceding agent and

the distance that will avoid a collision while changing

lanes.

Each agent should satisfy the following condi-

tions.

• Sufﬁcient distance should exist between it and

the current preceder. When the current preceder

stops, the subject agent decelerates gradually to

stop just behind the current preceder.

• When an agent that intends to change lanes en-

ters the middle zone and checks the environment,

if sufﬁcient distance to the new preceder and the

new follower exists, it then activates its signal

ﬂasher; otherwise, it decelerates and waits for an-

other chance to change lanes safely.

• When an agent notices the blinking signal ﬂasher

of an agent in the adjacent lane, it decelerates

(Fig. 6(a)) or accelerates (Fig. 6(b)) depending on

the distance between it and the current preceder

and current follower. In either case, it cooperates

so that the agent with the blinking signal ﬂasher

can change lanes smoothly.

LANE CHANGING MODEL WITH EARLY COMMUNICATION OF INTENTIONS

457

• If an agent cannot ﬁnd an opportunity to change

lanes safely while it is in the middle zone, then the

change of lanes takes place in the exit zone at the

highest priority. If several agents intend to change

lanes, then their lane changes are accomplished on

a ﬁrst come, ﬁrst served basis.

• When two agents are at almost the same level in

neighboring lanes, then both of them decelerate

and wait for a better time to exit the state of dead-

lock. If such a state continues in a designated time

interval, then one of them is given the priority for

changing lanes.

Note that the signal ﬂasher is not activated until

safe lane changing is guaranteed.

(a) (b)

Figure 6: Behavior of the agent upon noticing a blinking

signal ﬂasher.

3 MODEL WITH EARLY

COMMUNICATION OF

INTENTIONS

Lane changing in the basic model has two main draw-

backs. One is that an agent may miss a chance to

change lanes due to the behavior of other agents.

The other is the possibility of deadlock between two

agents in the same position in different lanes. To

eliminate these problems caused by a lack of mutual

understanding, we propose a model in which agents

communicate with each other to convey their inten-

tions in advance.

We call the agent in the target lane nearest the

subject agent the partner. The partner is either a

new preceder or a new follower of the subject agent.

The request message for lane changing is sent in the

entrance zone. The partner replies with a message

of acceptance or rejection. When the request is ac-

cepted, the agent in the target lane tries to facilitate

lane changing by the subject agent. Otherwise, the

subject agent itself manages to control its speed to get

a chance of safe lane changing. Thus both agents co-

operate. Moreover, when the partner receiving the

lane change request decides to accelerate, it sends

a message to its current follower to warn of a new-

comer. We call this message a newcomer notiﬁcation.

It is this message that gives the reason for the acceler-

ation. If this message were not sent, then the follower

would also accelerate without knowing why the pre-

ceder was accelerating, which would leave no space

for a newcomer.

The subject agent sends the lane-change request

message only once. Agents do not engage in a com-

plicated negotiation. In this model, communication

of intention is performed in the entrance zone, lane

changing is usually conducted in the middle zone, and

ﬁnal lane changing occurs in the exit zone.

We describe the protocols for changing lanes be-

low.

The Agent that Intends to Change Lanes

1. If it does not receivea newcomer notiﬁcation from

its current preceder, it maintains speed and sends

its request message to the partner.

2. If it receives a newcomer notiﬁcation from its cur-

rent preceder and if it does not receive a lane-

change request from the adjacent lane, it decel-

erates without sending a request.

3. If it receives a newcomer notiﬁcation from its cur-

rent preceder and also receives a lane-change re-

quest from the adjacent lane, it decelerates and

sends a request message to the partner.

The Agent Receiving the Lane-change Request

1. If it has no intention of changing lanes

(a) If it receives a newcomer notiﬁcation from its

current preceder, it decelerates.

(b) If sufﬁcient distance exists behind the current

preceder, it accelerates, sends an acceptance

message to the subject agent, and also sends

a newcomer notiﬁcation message to its current

follower.

follower, it decelerates and sends an acceptance

message to the subject agent.

(d) Otherwise, it maintains its speed and sends a

rejection message to the subject agent.

2. If it intends to change lanes

(a) If it receives a newcomer notiﬁcation from its

current preceder, it maintains its speed without

sending a request.

(b) Otherwise, it decelerates without sending a re-

quest.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

458

The Agent Receiving the Acceptance

1. It maintains the current speed and changes lanes

according to the basic model when it enters the

middle zone.

The Agent Receiving the Rejection

1. If sufﬁcient distance exists behind the current pre-

ceder, it accelerates.

2. If sufﬁcient distance exists ahead of the current

follower, it decelerates.

3. Otherwise, it maintains its speed.

4 EXPERIMENT

4.1 Simulation

We implemented the model using Java and performed

simulations under different conditions to demonstrate

the effectiveness of early communication of inten-

tions. Figure 7 shows a screenshot of our system.

Figure 7: System screenshot.

We assume that all agents are equivalent; i.e., they

are of the same size, the same performance, and ex-

hibit behavior based on the same maneuvers. The pa-

rameters used in the simulation are shown in Table 1.

4.2 Experimental Results

We performed simulations for three data patterns.

Data patterns 1 and 2 are the cases in which the agents

in both lanes intend to move to adjacent lanes, where

Table 1: Parameters. (px = pixels).

length of the road 700 px

length of the entrance zone 350 px

length of the exit zone 40 px

size of the agent 10 px

smallest safe gap between agents 20 px

total number of agents 50

the times of simulation 100

the rates of the agents that change lane are 30% (data

pattern 1) and 50% (data pattern 2), respectively. Data

pattern 3 is the case in which only the 50% of agents

in the left lane intend to move to the right lane.

We investigated how smoothly lane changing oc-

curs depending on the existence of early communica-

tion.

Tables 2– 4 show the results. In these tables, as

shown in Fig. 8, (a) is the case with no communi-

cation, (b) is the case in which the communication

is performed 234–350 pixels (px) after the start line,

and (c) is the case in which the communication is

performed 117–234 (px) after the start line. DL is

the number of deadlocks, and Final is the number

of agents that change lanes in the exit zone. All the

agents are ready to start when the time is zero, and

each agent starts when a sufﬁcient space is made be-

tween it and the immediately preceding agent. Start

is the mean number of starting time (step) of all the

agents and start SD is its standard deviation. Travel

is the mean number of time (steps) required for each

agent to travel from the start line to the end line, and

travel SD denotes its standard deviation.

middle

exit

entrance

350

700

communication

area in (b)

communication

area in (c)

117

234

Figure 8: Areas in which communication is performed.

4.3 Evaluation

Two Goal Lanes (Data Patterns 1 and 2). We

compare the result with communication (b)(c) and

that without communication (a). Both the number

LANE CHANGING MODEL WITH EARLY COMMUNICATION OF INTENTIONS

459

Table 2: Pattern1.

DL ﬁnal start start SD travel travel SD

(a) 0.16 4.09 200.71 127.59 331.41 29.53

(b) 0.00 0.85 202.02 129.83 318.25 24.95

Table 3: Pattern 2.

DL ﬁnal start start SD travel travel SD

(a) 0.74 8.93 212.71 141.54 405.75 65.23

(b) 0.01 1.33 217.15 149.11 365.28 45.14

of deadlocks and that of the agents ﬁnally changing

lanes are smaller in the former case. Both starting de-

lay and traveling time are smaller, and their standard

deviations are also smaller in the former case. It fol-

lows that early communication is effective in achiev-

ing smooth lane changes. It is much effective in the

case the rates of the agents that change lane is high.

What about the timing of communication? Com-

paring the results (b) and (c), there is no big differ-

ence. If communication is performed just before en-

tering the middle zone, there may not be enough time

to prepare for a smooth lane change. On the other

hand, if communication is performed earlier, the sit-

uation may change after the preparation is accom-

plished. It is necessary to determine the timing of

communication depending on the trafﬁc situation.

One Goal Lane (Data Pattern 3). The delay of the

start was large since one lane was too congested to

create a sufﬁcient gap behind the preceder. In this

case, the agents starting earlier succeeded in changing

lanes smoothly, while those starting later could not

ﬁnd a safe gap immediately and arrived at the end line

rather late. That is why the standard deviations of the

starting time and travelling time are larger in (b)(c).

In this case, early communication is not so effective,

comparing with the case of two goal lanes.

5 DISCUSSION

Numerous studies have been done on trafﬁc con-

trol for autonomous driving, including lane chang-

ing. While most of these models assume inter-vehicle

communications, these occur just before the action

and do not convey the intention in advance. More-

over, the case of deadlock is not correctly modeled.

Hidas proposed a lane-changing model for con-

gested roads (Hidas, 2002) and divided the pattern

of lane changing into three, free, cooperative, and

forced, insisting that the last two patterns occur on

congested roads. When an agent intends to change

lanes but doing so is not feasible, it then sends a cour-

tesy message. In cooperative maneuvering, the sub-

ject agent waits until agents in the target lane make a

space for it before changing lanes, while in the forced

maneuver, the subject agent changes lanes even if not

enough space exists; after the change, the agents fol-

lowing the subject agent decelerate to create a safe

distance. Hidas showed the result of his evaluation

when incidents occur and changing lanes is essential.

The model with communication is more advantageous

than the one without, both in terms of travel time and

trafﬁc ﬂow. Later he analyzed video recordings of

real trafﬁc ﬂow and showed that the proposed simula-

tion model correctly represented the observed trafﬁc

pattern (Hidas, 2005).

The difference between the Hidas model and ours

is in the behavior of the subject agent when lane

changing is not feasible. In the Hidas model, the sub-

ject agent has a plan that is shared with the agents

in the target lanes. The subject agent waits for the

chance without taking any positive action. In our

model, the subject agent itself changes speed. It fol-

lows that our model is more cooperativesince both the

subject agent and the responders cooperate to achieve

the same goal. In addition, we use early communica-

tion of intentions while the Hidas model does not.

6 CONCLUSIONS

We proposed a model with early communication of

intention to change lanes that creates more coopera-

tive behavior among the agents involved. We created

a simulator based on this model and showed that the

model has the following characteristics:

• collision-free

• almost deadlock-free

ICAART 2009 - International Conference on Agents and Artificial Intelligence

460

Table 4: Pattern 3.

DL ﬁnal start start SD travel travel SD

(a) 0.00 22.06 201.93 132.63 540.95 156.00

(b) 0.00 18.72 214.10 145.07 505.95 152.54

• produces a short travel time for each agent

In this paper, we put the assumption that all agents

are cooperative and follow the protocol. It is be-

cause our target is an autonomous driving environ-

ment, which is considered to be an advanced form of

the automated highway systems. However, it is inter-

esting to simulate the case in which some agents do

not follow the protocol as a more realistic situation.

Moreover, we plan to extend the model to cover

the following cases:

• more than two lanes

• multiple sectors

• sudden arrival of a vehicle from a structure beside

the road

From the theoretical point of view, we are consid-

ering a more reﬁned model of lane changing based on

the inner state of agents. The inner state of agents can

be suitably modeled using the Belief-Desire-Intention

(BDI) model (Wooldridge, 2000). The environment

perceivedby an agent and messages sent by the neigh-

boring agents are regarded as beliefs of an agent, the

request conveyed earlier is regarded as a desire, and

the request just before taking the action is considered

an intention. In this way, we hope to create a more

reﬁned model and simulate the agent behaviors from

another perspective.

REFERENCES

Ammoun, S., Nashashibi, F., and Laurgeau, C. (2007). An

analysis of the lane changing manoeuvre on roads :

the contribution of inter-vehicle cooperation via com-

munication. In Intelligent Vehicles Symposium, 2007

IEEE, pages 1095–1100.

Dresner, K. and Stone, P. (2005). Multiagent trafﬁc manage-

ment: An improved intersection control mechanism.

In AAMAS05, pages 25–29.

Ehlert, P. and Rothkrantz, L. (2001). Microscopic trafﬁc

simulation with reactive drving agents. In Proc. of

Intelligent Transportation Systems, pages 860–865.

Gipps, P. (1986). A model for the structure of lane-changing

decisions. Transportation Research, 20B:107–120.

Hidas, P. (2002). Modelling lane changing and merging

in microscopic trafﬁc simulation. Transportation Re-

search Part C: Emerging Technologies, 10(5):351–

371.

Hidas, P. (2005). Modelling vehicle interactions in micro-

scopic simulation of merging and weaving. Trans-

portation Research Part C: Emerging Technologies,

13(1):37–62.

Jin, W., Zheng, Y., and Li, J. (1999). Microscopic simula-

tion of trafﬁc ﬂow at signalized intersection based on

cellular automata. In Proc. of the IEEE International

Vehicle Electronics Conference, IVEC99, pages 106–

109.

Kanaris, A., Kosmatopoulos, E., and Ioannou, P. (2001).

Strategies and spacing requirements for lane changing

and merginging in automated highway systems. IEEE

Transactions on Vehicular Technology, 50(6):1568–

1581.

Kato, S., Tsugawa, S., Tokuda, K., Matsui, T., and Fujii,

H. (2002). Vehicle control algorithm for cooperative

driving with automated vehicles and intervehicle com-

munications. IEEE Trans. on Intelligent Transporta-

tion Systems, 3(3):155–161.

Knospe, W., Santen, L., Schadschneider, A., and Schreck-

enberg, M. (1999). Disorder effects in cellular au-

tomata for two-lane trafﬁc. Physica A, (265):614–633.

Nagel, K. and Schrechenberg, M. (1992). A cellulear

automaton model for freeway trafﬁc. Journal of

Physique, I(2):2221–2229.

Su, X., Jin, X., Min, Y., and Peng, B. (2005). Study on

asymmetric two-lane trafﬁc model based on cellular

automata. In 5th International Conference on Compu-

tational Science (ICCS05), pages 599–606. Springer.

Wooldridge, M. (2000). Reasoning about Rational Agents.

The MIT Press.

LANE CHANGING MODEL WITH EARLY COMMUNICATION OF INTENTIONS

461