Group Formation and Knowledge Sharing in Pedestrian

Egress Simulation

Kyle D. Feuz

and Vicki H. Allan

Electrical Engineering and Computer Science, Washington State University, Pullman, Washington, U.S.A.

Computer Science, Utah State University, Logan, Utah, U.S.A.

Keywords:

Coalition Formation, Pedestrian Simulation, Multi-agent Systems.

Abstract:

Pedestrian simulation has been a topic of research for several decades, especially in regards to pedestrian

egress. Only recently, though, have researchers begun to consider the effects that groups have upon pedestrian

egress. Both empirical studies and simulation models predict a decrease in pedestrian speeds when pedestrians

travel in groups. In this study, we show that this decrease in speed does not necessarily correspond to an

increase in egress time as additional factors such as the amount of knowledge gained through the formation of

groups must be considered. The sharing of route costs helps pedestrians maintain proximity to each other and

under certain circumstances, pedestrian egress times are actually improved by the formation of groups. We

also show that the inclusion of communication costs, sharing knowledge, and group decision-making all have

a strong impact on predicted egress times.

1 INTRODUCTION

In recent years, accurate pedestrian simulation has

become an important research topic (Santos and

Aguirre, 2004; Hoogendoorn and Bovy, 2004; Pan,

2006; Helbing and Johansson, 2009). Pedestrian sim-

ulation models can be employed in the design of safe

facilities, validation of ﬁre codes, and the automatic

tracking and surveillance of pedestrians in live video

feeds (Antonini et al., 2006). Real-world experiments

are too dangerous and too expensive to be a practi-

cal way of learning about egress efﬁciency. For this

reason, simulation models have been developed to

demonstrate crowd behavior in an emergency. Unfor-

tunately, these systems currently fail to capture many

important characteristics of pedestrian behavior such

as group formation and information sharing.

In coalition formation theory, a coalition will only

form if the utility achieved by the agents in the coali-

tion is greater than the utility each agent could achieve

alone (Shehory and Kraus, 1998). This assump-

tion of individual rationality is common for multi-

agent systems (Russell and Norvig, 2010). In pedes-

trian egress, pedestrians frequently move together in

groups. From the above assumptions, the utility of

pedestrians should be greater by joining a group than

if they were to travel as individuals, yet most of the

literature indicates that group formation has a nega-

tive effect on ﬂow rates, average speed, and egress

times (Moussa¨ıd et al., 2010; Qiu and Hu, 2010; Yang

et al., 2005; Zhao et al., 2008). From this data, two

logical conclusions can be drawn, given that individ-

uals do form coalitions. First, the utility of pedes-

trians during egress is affected by more factors than

just egress time. Such factors may include emo-

tions, altruism, social inﬂuences or stress. Several

researchers have considered this perspective (Bosse

et al., 2011; Hoogendoorn et al., 2010; Kułakowski

and Gawro´nski, 2009; Ozel, 2001). Second, other

beneﬁts are gained so that the overall egress time is

not always negatively affected. Such beneﬁts may

include information sharing or stress reduction. In

this paper, we describe a new pedestrian simulation

model with special consideration of group formation,

group decision-making, and information sharing. We

show that these factors can have a signiﬁcant impact

on egress times and, in certain situations, egress time

can be improved through the formation of groups.

2 RELATED WORK

Several researchers have found negative effects on

ﬂow rates, and average speed when group movement

is considered. Moussaid et al. conduct empirical

357

D. Feuz K. and H. Allan V..

Group Formation and Knowledge Sharing in Pedestrian Egress Simulation.

DOI: 10.5220/0004197003570364

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 357-364

ISBN: 978-989-8565-38-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

studies to determine several group parameters includ-

ing size, structure formation, and speed (Moussa¨ıd

et al., 2010). The average speed is found to decrease

with increasing group size. Similarly, Qiu develops a

framework for group modeling in (Qiu, 2010) which

predicts decreased ﬂow rates for pedestrian groups.

Initially, the simulation model predicts an increase in

ﬂow rate as group size increases. As group size con-

tinues to increase, however, the ﬂow rates decrease.

Both Moussaid and Qiu use a similar idea of repre-

senting group interactions through the use of a social

cohesion force which simulates group members’ de-

sire to maintain a close proximity to each other. Nei-

ther model, however, addresses the issues of route se-

lection and information sharing in a group setting or

the cost of communication.

Ji and Gao consider the effect of multiple lead-

ers with perfect evacuation route knowledge (Qingge

and Can, 2007). In their simulation model, they ﬁnd

that including more leaders increases the egress ef-

ﬁciency, in terms of total egress time, up to a cer-

tain saturation point, after which including more lead-

ers decreases the egress efﬁciency (Qingge and Can,

2007). This occurs because, as the number of leaders

increase, more pedestrians receive conﬂicting direc-

tions from multiple leaders which in turn hinders their

ability to quickly egress from the building. Murakami

et al. conduct a similar test using ﬁre drills in a simu-

lated model (Murakami et al., 2002). Leaders can in-

struct evacuees to either follow them or to take a cer-

tain route. In their experiments, the dedicated leaders

are known beforehand and given additional informa-

tion regarding which exits to take. We study the more

general case where leaders are not known nor are they

given special training prior to the simulation.

Yang and Zhao et al. use a simulation model

to measure the effect of grouping upon egress time

(Yang et al., 2005; Zhao et al., 2008). They classify

grouping as either spatial or directional (Zhao et al.,

2008). Spatial grouping relates to individuals’ desire

to be close physically. Directional grouping relates to

the desire of individuals to move in the same direction

as others. Their model indicates that spatial grouping

is detrimental to egress efﬁciency. However, direc-

tional grouping is found to increase the egress efﬁ-

ciency. This is one of the few papers that show any

beneﬁt to grouping. We are interested in showing that

even spatial grouping can lead to additional beneﬁts.

Tsai et al. have developed a sophisticated model

for simulating pedestrian egress with family groups

and authority ﬁgures (Tsai et al., 2011). Their model

includes pedestrians with imperfect knowledge of the

layout, children and parent pedestrians, and trained

authority pedestrians. They point out that most pre-

vious simulators do not accurately model pedestrian

egress because the agents are omniscient and do not

suffer from the effects of fear or stress. Their model

shows that egress times are signiﬁcantly impacted

when pedestrians have imperfect knowledge, seek to

maintain a group formation, and experience the ef-

fect of stress or fear. Their model is still unrealis-

tic, however, because they do not include any time

penalty for the sharing of knowledge. Additionally,

the group model used by Tsai et al. is for families

where the children do not participate in the decision-

making process.

3 OUR MODEL

Our research study is performed using the Pedestrian

Leadership and Egress Assistance Simulation Envi-

ronment (PLEASE), which we developed for this pur-

pose. PLEASE is built upon the multi-agent modeling

paradigm where each pedestrian is represented as an

individually rational agent capable of perceiving the

environment and reacting to it. In PLEASE, pedes-

trian agents can perceive obstacles, hazards, routes,

and other agents. The agents are capable of basic

communication to allow for the formation and disso-

lution of coalitions and the sharing of knowledge. The

agents use a two tier navigational module to control

their movement within the simulation environment.

The high-level tier evaluates available routes and se-

lects a destination goal. The low-level tier, based

on the social force model (Helbing and Johansson,

2009), performs basic navigation and collision avoid-

ance. The social force model is selected for the low-

level tactical navigation for four main reasons: it is

simple to understand and implement, it is widely used

in many simulation models, it successfully reproduces

many crowd phenomena, and it has been validated us-

ing actual pedestrian data (Kretz et al., 2008; Helbing

and Johansson, 2009; Moussa¨ıd et al., 2010; Luber

et al., 2010). Due to space constraints, we do not go

into the details of the social force model here but refer

the readers to (Feuz, 2011; Helbing and Johansson,

2009; Moussa¨ıd et al., 2010)

3.1 Route Selection

PLEASE uses the concept of decision points to fa-

cilitate pedestrian route navigation. A decision point

is deﬁned as a point in the building at which an agent

must decide upon the next location in the route. These

points may be placed at arbitrary locations, but typi-

cally decision points are placed at doorways and in-

tersections. When exiting from a building, pedestri-

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

358

ans navigate from one decision point to another. By

only placing decision points at doorways and inter-

sections, the placement of the decision point does not

require the pedestrian to pass through an area which

they would not normally pass through when navigat-

ing from one area of the building to another.

Several different route selection algorithms are

implemented in PLEASE. For this paper, we focus

on two different route selection algorithms, a local

route selection algorithm (Feuz and Allan, 2012a) and

a trained route selection algorithm (Feuz and Allan,

2012b) which are explained below. We use these two

route selection techniques to compare differences in

egress times when pedestrians have different amounts

of knowledge about the building. The local route se-

lection algorithm does not require any prior knowl-

edge of the building as it uses only locally observable

information. The training algorithm allows pedestri-

ans to know the route costs for decision points with

which they are familiar. This knowledge can then po-

tentially be shared with other pedestrians.

3.1.1 Local Route Selection

The local route selection algorithm estimates the cost

of exiting via a given decision point based upon sev-

eral locally observable characteristics of the point. In

this paper, we use the distance, corridor width, room

signs, and congestion characteristics when estimating

costs. These characteristics have been found to work

well for a variety of building layouts (Feuz, 2011).

3.1.2 Trained Route Selection

The training algorithm allows agents to experience

multiple simulation runs in a building during which

time the agents may learn the expected costs to dif-

ferent decision points. This information is stored in

the agent’s model of the building. More training pro-

vides agents with more knowledge about the building,

allowing for more effective route planning. In this

study, we restrict the learning to distance information.

This is done to facilitate the sharing of knowledge be-

tween group members by enforcing a common cost

metric, as discussed in section 3.4.

3.2 Group Formation

PLEASE allows for pedestrians to walk in groups.

These groups may be formed to share knowledge,

request aid, relieve stress, or interact socially. In

PLEASE, groups are more than just individuals mov-

ing in the same direction. Being a member of a group

implies communication, agreement, and a desire to

remain close together. Pedestrian groups may be of a

static or dynamic nature.

Similar to (Qiu, 2010; Moussa¨ıd et al., 2010), we

represent groups using an additional cohesive force

applied between group members. Pedestrians in a

group seek to maintain a certain proximity to visi-

ble group members. This makes sense because one

pedestrian cannot maintain a certain proximity to

another pedestrian unless the location of the other

pedestrian is known. When turning the corner pedes-

trians might temporarily lose sight of one another.

When this happens, they continue in the same direc-

tion and as the other pedestrians also turn the corner

they are able to reconnect.

3.2.1 Static Group Formation

Static groups can be formed at the beginning of the

pedestrian simulation. These groups represent re-

lationships, which are deﬁned outside of the simu-

lation such as family, friend, or business relation-

ships. Static groups do not change throughout the

simulation: new members cannot be added and cur-

rent members are only removed when they exit the

building. Rather than require the user to deﬁne each

group manually, PLEASE uses user-deﬁned parame-

ters to automatically create groups at the start of the

simulation. Empirical studies have found that pedes-

trian group sizes tend to be small and follow a zero-

truncated Poisson distribution(Moussa¨ıd et al., 2010).

This distribution can be approximated by adjusting

the parameters controlling group formation.

3.2.2 Dynamic Group Formation

Dynamic groups can be formed throughout the sim-

ulation. An agent may seek to join or leave a group

at any time during the simulation, but joining a group

requires the consensus of the group members. Pedes-

trian agents use utility theory when deciding whether

to join a group and whether to accept newgroup mem-

bers. The two actions each have separate utility func-

tions. We refer to the utility of joining a group as the

agent’sindividualutility function. We referto the util-

ity of accepting new group members as the group util-

ity function. PLEASE is built to be extensible, so the

exact utility functions used may be easily changed.

In describing the group formation process, we will

use the following notation.

• A - The set of agents in the simulation

• G - The set of groups of agents in the simulation.

• S - The set of agents in a group.

• S

- The group of which agent x is a member.

GroupFormationandKnowledgeSharinginPedestrianEgressSimulation

359

• L

- The leader of S

Initially ∀x ∈ A, S

= {x} and L

= x. This states

that at the start of the simulation, each agent in the

simulation is the leader of a group consisting only of

the individual agent. The group formation process is

dividedinto four steps, do nothing, request admission,

extend invitation, accept/reject invitation. Any agent

may request admission into any nearby group. Any

agent may accept any received invitation. Only group

leaders may extend invitations to other agents. At ev-

ery time step t, ∀x ∈ A, agent x evaluates its utility

function and then either does nothing, requests ad-

mission or accepts/rejects invitations to nearby groups

based upon the expected utility. ∀S ∈ G, L

evaluates

its group utility function and can then choose whether

or not to extend invitations based upon the expected

utility.

3.3 Route Consensus

The PLEASE model allows groups to use three differ-

ent route consensus mechanisms, which incorporate

suggestions from group members to different degrees.

The route consensus mechanisms are least-cost route

(LC), most-common route (MC), or dictator. The dic-

tator mechanism simply chooses the decision point

proposed by the group leader. This method serves as

a benchmark to measure the effectiveness of groups.

Most other simulators do not consider route consen-

sus for groups and are thus using a dictator-like mech-

anism (Qiu and Hu, 2010; Tsai et al., 2011).

LC and MC require each pedestrian in the group

to submit a preferred decision point and an associated

route cost estimate. From the proposed routes, the LC

mechanism then selects the route with the least cost

(as identiﬁed by group members). In order for the

LC mechanism to work successfully, group members

must use the same scale to measure cost. The group

members are assumed to be reliable, that is to say,

they report their true perception of route costs and do

not lie.

The MC mechanism selects the route which was

proposed most frequently. Ties are broken by average

route cost. The MC mechanism mitigates the prob-

lems inherent in comparing costs computed by vari-

ous means as it chooses the most commonly proposed

route. The group members can have completely dif-

ferent route cost functions and a minority of group

members could be unreliable without affecting the

route selected by the group. However, this mecha-

nism is unable to take full advantage of the special

knowledge which any particular agent may have.

3.4 Information Sharing

When using the training route selection algorithm,

agents are allowed to share expected route cost in-

formation. Current simulation models assume that

pedestrian knowledge is not shared among group

members. This is a valid assumption for many sit-

uations. It represents pedestrian groups choosing an

egress route without prior discussion as to which route

is the most efﬁcient or effective. However, pedes-

trian groups might also ﬁrst discuss the beneﬁts and

drawbacks to a particular route before deciding on an

egress route. PLEASE allows for either scenario, and

in this paper, we consider the effects of both.

In PLEASE, expected route-cost information may

either be public or private. If the information is pub-

lic, then, at the beginning of the simulation, group

members may share all the route cost information for

each decision point learned during the training runs.

Each agent has access to the models of other group

members to integrate into its own model. Currently,

all model information is treated equally so costs are

integrated as an average of the other agents’ costs. A

more complex model might allow for issues such as

trust and reliability to affect the weight that each agent

applies to other agents’ models while integrating the

costs into their own model.

Sharing information requires communication

costs. In an actual situation, sharing information may

take anywhere from a few seconds to a few minutes.

To account for this fact, PLEASE has a sharing cost

parameter which is the time, in seconds, an agent

spends sharing the route information with other group

members. Larger groups will thus require more time

to share route information than smaller groups. While

route information is being shared, no member of the

group moves towards any goal location. If group in-

formation is private, then the group members do not

share complete route-cost information. Hence there is

no associated communication cost.

As the route consensus techniques are combined

with different information sharing techniques, un-

derstanding exactly what information is shared can

be confusing. To help clarify, we state explicitly

what information is shared for the various combina-

tions. Using the LC consensus mechanism with pub-

lic knowledge results in agents sharing all their re-

spective knowledge for each decision point at the be-

ginning of the simulation. As the simulation pro-

ceeds, at every decision point, each agent will propose

their preferred next decision point and estimate of the

total route cost via that decision point. The cheapest

proposed decision point will be selected. Using the

MC consensus mechanism with public knowledge re-

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

360

sults in agents sharing all their respective knowledge

for each decision point at the beginning of the simu-

lation. As the simulation proceeds, at every decision

point, each agent will propose their preferred next de-

cision point and estimate of the total route cost via

that decision point. The most commonly proposed

decision point will be selected. Because all the in-

formation has been shared previously, the only differ-

ence between these mechanisms is in the individual

perspectives of the agents. One agent might have a

clearer view of congestion then another agent, or one

route might be closer to one agent but further away

for another agent so the best next point varies.

With the LC consensus mechanism and private

knowledge, no knowledge is shared between agents

at the beginning of the simulation. As the simula-

tion proceeds, at every decision point, each agent will

propose their preferred next decision point and esti-

mate of the total route cost via that decision point.

The cheapest proposed decision point will be se-

lected. With the MC consensus mechanism and pri-

vate knowledge, no knowledge is shared between

agents at the beginning of the simulation. As the sim-

ulation proceeds, at every decision point, each agent

will propose their preferred next decision point and

estimate of the total route cost via that decision point.

The most commonly proposed decision point will be

selected. In this case, considerable differences exist

between these mechanisms as each agent has unique

knowledge.

4 EXPERIMENTAL RESULTS

In this research, we consider the effects of static

grouping on pedestrian egress times. Various experi-

ments control the route selection algorithm used, the

route consensus mechanism, and the knowledge shar-

ing available to pedestrians. Egress times are cal-

culated with 100 agents per simulation. Tests are

repeated 20 times to put error bars into acceptable

ranges. To quantify differences between the perfor-

mances of the different mechanisms, we deﬁne efﬁ-

ciency as the amount of time taken to evacuate a given

percentage of pedestrians. We will use this deﬁnition

of efﬁciency throughout our discussion of these ex-

periments.

4.1 Static Groups

In the ﬁrst experiment, we compare the results of

static group formation when pedestrians use the

heuristic route selection with and without group for-

mation. The group consensus mechanism and knowl-

edge sharing mechanism have little effect on the

egress time, because in this test, none of the pedes-

trians have prior knowledge of the building and they

all use the same heuristic function. The purpose of

this experiment is to verify that pedestrian groups

have a negative impact on egress time and to quan-

tify that impact when no knowledge is shared between

pedestrians. As can be seen in Figure 1, forming

static pedestrian groups without sharing knowledge

has a negative impact on egress times. The average

time taken to evacuate 50% of the pedestrians is 29%

greater when pedestrians form groups than when no

groups are formed. Group formation is 35% less ef-

ﬁcient at the 70% mark and 69% less efﬁcient at the

90% mark.

0.2

0.4

0.6

0.8

0 50 100 150 200 250 300 350

Percent Evacuated

Time (seconds)

Effect of Group Formation

No Knowledge Sharing

no groups

least-cost

Figure 1: Comparison of static group formation on egress

times when pedestrians have no prior knowledge of route

costs. Groups use the least-cost route consensus mecha-

nism.

In the second experiment, we show that group for-

mation has a negative impact on egress times when

pedestrians have individualized knowledge of the

building (acquired through training) but do not share

that knowledge. To do this we compare the egress

times using no groups versus using a dictatorship for

pedestrians, which have learned route-distances over

the period of 15 training runs. Due to space con-

straints, the results are not shown here but they are

similar to the results shown in Figure 1.

The last experiment for static groups compares the

egress times of pedestrian egress when group forma-

tion occurs and knowledge sharing is allowed. The

route consensus mechanisms are tested pair-wise with

the knowledge sharing mechanisms so we have the

following combinations: 1) No Groups, 2) Dicta-

tor, 3) Least-cost, Private information (cost-prvt), 4)

Least-cost, Public Information (cost-pblc, 5) Most-

common, Private Information (common-prvt), and 6)

Most-common, Public Information (common-pblc).

As with the previoustest, the pedestrians use the train-

ing route selection algorithm and have been trained

GroupFormationandKnowledgeSharinginPedestrianEgressSimulation

361

15 times in the building. This means that most agents

will know one or two exits and several ways to get

there. As the cost of communication is likely to vary

depending upon the circumstances, we consider two

different communication costs, free (0 seconds per

agent) and cheap (10 seconds per agent). As will

be seen from the experiments, communication costs

which are much greater than 10 seconds per agent are

no longer effective, so we do not consider them.

Figure 2 compares the resulting egress times when

knowledge sharing is free. When group members

have their route-cost knowledge public, group forma-

tion leads to decreased egress times compared to no

groups. Because all the route information is pub-

lic between group members, the consensus mecha-

nism has little effect on the egress time and only the

cost-pblc mechanism is shown in the results. When

group members keep their knowledge private, then

the consensus mechanism has a greater effect upon

egress times. If the group uses the least-cost consen-

sus mechanism, then the egress performance is nearly

as good as if the group had route-costs public among

them and outperforms the egress time of individuals

who do not form groups. If the group uses the most

common consensus mechanism, then egress perfor-

mance is actually worse than not forming groups, as

the common consensus mechanism is unable to cap-

italize on the information that may be had by only a

minority of the group members. These results show

that when knowledge is shared for free among group

members, group formation is transformed from hav-

ing a negative effect on egress times to having a posi-

tive effect on egress times.

0.2

0.4

0.6

0.8

0 50 100 150 200

Percent Evacuated

Time (seconds)

Effect of Group Formation

Knowledge Sharing Cost is Free

cost-prvt

cost-pblc

common-prvt

dictator

no groups

Figure 2: Static group formation and egress times using: 1)

No Groups, 2) Dictator, 3) Least-cost, Private information

(cost-prvt), 4) Least-cost, Public Information (cost-pblc),

and 5) Most-common, Private Information (common-prvt).

Sharing is free.

We also consider the case when sharing knowl-

edge is associated with a communication time penalty

of 10 seconds per group member (see Figure 3). Un-

der this assumption, having knowledge public among

group members is no longer the most efﬁcient solu-

tion. However, it can still improve the efﬁciency of

egress as compared to no knowledge sharing. One

might expect that a dictatorship would be more ef-

fective than taking the time to ﬁrst share knowledge

as this avoids the cost of communication. In the be-

ginning of the simulation, this is true, but as the sim-

ulation progresses, the knowledge sharing technique

begins to overtake the dictator method in terms of

percent of pedestrians evacuated. Sharing knowledge

improves the efﬁciency with which pedestrians exit

as compared to the dictator mechanism after the 70%

evacuated mark, and it improves efﬁciency for the

last 6% of pedestrians as compared to the common-

consensus mechanism without knowledge sharing.

Under the assumption of communication costs,

using the least-cost route consensus mechanism with

private knowledge (cost-prvt) is the most efﬁcient

technique for pedestrian egress. It does not incur

the communication costs of publicly sharing all route

knowledge but is still able to beneﬁt from the indi-

vidual knowledge of each pedestrian. In these experi-

ments, cost-prvt is even more efﬁcient than no groups.

This is signiﬁcant because it indicates that group for-

mation does not always have to have a negative effect

on egress times. If the knowledge can be used with-

out explicitly being shared, then the whole group can

beneﬁt.

0.2

0.4

0.6

0.8

0 50 100 150 200

Percent Evacuated

Time (seconds)

Effect of Group Formation

Knowledge Sharing Cost is Size Dependent

cost-prvt

cost-pblc

common-prvt

dictator

no groups

Figure 3: Static group formation and egress times using: 1)

No Groups, 2) Dictator, 3) Least-cost, Private information

(cost-prvt), 4) Least-cost, Public Information (cost-pblc),

and 5) Most-common, Private Information (common-prvt).

Sharing costs 10 seconds per group member.

Besides the efﬁciencyof each mechanism, we also

consider several other statistics which indicate how

well static groups maintain a close proximity (see Ta-

ble 1). The two most relevant statistics deal with the

spatial and temporal proximity maintained by groups

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

362

Table 1: Ranking of route consensus and knowledge sharing mechanisms by spatial proximity.

Rank Mechanism Temporal Proximity Spatial Proximity

of Finish Point of Finish Point

1 cost-pblc 7.08 s 2.33 m

2 common-pblc 7.52 s 2.74 m

3 cost-prvt 7.95 s 5.11 m

4 common-prvt 12.04 s 5.43 m

5 dicator 11.95 s 6.00 m

6 no group 26.82 s 26.63 m

while exiting the building. Spatial proximity is calcu-

lated as the average pair-wise distance between the lo-

cations at which each group member ﬁnishes. Tempo-

ral proximity is deﬁned as the amount of time elapsed

between the successful egress of the ﬁrst pedestrian of

the group and the successful egress of the last pedes-

trian of the group. These two measures reﬂect how

well group members maintain a close formation in

both time and space.

Table 1 shows the average spatial and temporal

proximities for each combination of mechanisms con-

sidered. Actively maintaining the group formation

leads to closer proximities than no groups. Having

knowledge information public among group members

further increases the proximities group members are

able to maintain because each member has the same

knowledge, after the initial sharing has occurred. This

helps groups maintain proximity even if they are tem-

porarily split apart. Without the shared knowledge,

group members are more likely to become lost after

being split apart from the others. This might indicate

one reason for sharing route information even with an

increased communication cost.

4.2 Dynamic Groups

We perform similar experiments with dynamic groups

to evaluate the effect of dynamic group formation on

pedestrian egress time. The results obtained during

these experiments indicate that, like static group for-

mation, dynamic group formation tends to lead to

slower egress times. The degree to which egress time

is affected is dependent upon the number and size

of the groups which form. When only a few small

groups form, egress time is not signiﬁcantly affected.

However, as more groups are formed and as group

size increases, the negative impact on egress times

also increase. When the egress times for the stress

and knowledge utility functions are compared, there

did not seem to be a signiﬁcant difference. The factor

with the largest impact on egress times is the number

and size of groups formed.

It is important to note here that this does not mean

it is not beneﬁcial for an individual to join a group as

a means of compensating for a lack of knowledge or

as a means of relieving stress. Indeed, some individ-

ual agents experience improved egress times by join-

ing a group. However, the overall effect on the sys-

tem is that when too many groups form or the groups

become too large congestion ensues and egress times

slow down.

5 CONCLUSIONS AND FUTURE

WORK

Pedestrian simulation is an important area of research

with many applications. Until recently, group for-

mation in pedestrian egress has largely been ignored.

However, recent work has begun to address the issues

that arise with group formation. In this paper, we have

implemented a novel dynamic group formation tech-

nique which allows pedestrian groups to communi-

cate, share knowledge and reach a consensus regard-

ing route selection. To our knowledge, this is the ﬁrst

such simulation model to address the issues of knowl-

edge sharing with time penalties and group consensus

in pedestrian egress. As future work, we suggest that

the communication costs for sharing route knowledge

be investigated further and that issues of trust and re-

liability be incorporated into the simulation model.

We have shown that, although recent literature

emphasizes the negative impacts group formation can

have upon egress times, positive incentives to group

formation exist. Our simulation model predicts that

sharing knowledge in pedestrian groups can help

pedestrian maintain closer temporal and spatial prox-

imity with greater ease as well as improve egress

times, compared to group formation without the shar-

ing of knowledge. This is true even when the in-

creased communication costs of sharing knowledge

is considered.

Dynamic group formation has impacts on egress

times similar to those found with static group forma-

tion. Pedestrians can form groups to compensate for

a lack of knowledge or as a means to reduce stress,

GroupFormationandKnowledgeSharinginPedestrianEgressSimulation

363

but both reasons have similar impacts on total egress

times. The type of consensus mechanism used, the

amount of knowledge shared, and the cost of sharing

knowledge are all shown to have a signiﬁcant impact

on the overallegress times predicted by the simulation

and are therefore important factors to include when

designing a realistic pedestrian simulator.

This work is supported by NSF research grant

#0812039 entitled “Coalition Formation with Agent

Leadership.”

REFERENCES

Antonini, G., Bierlaire, M., and Weber, M. (2006). Dis-

crete choice models of pedestrian walking behav-

ior. Transportation Research Part B: Methodological,

40(8):667–687.

Bosse, T., Hoogendoorn, M., Klein, M. C. A., Treur, J.,

and Van Der Wal, C. N. (2011). Agent-based analysis

of patterns in crowd behaviour involving contagion of

mental states. In Proceedings of the 24th international

conference on Industrial engineering and other appli-

cations of applied intelligent systems conference on

Modern approaches in applied intelligence - Volume

Part II, pages 566–577, Berlin, Heidelberg. Springer-

Verlag.

Feuz, K. (2011). Pedestrian leadership and egress assistance

simulation environment. Master’s thesis, Utah State

University.

Feuz, K. and Allan, V. (2012a). Simulating knowledge

and information in pedestrain egress. In 4th Inter-

national Conference on Agents and Artiﬁcial Intelli-

gence, pages 246–253.

Feuz, K. and Allan, V. (2012b). Simulating pedestrain

route selection with imperfect knowledge. In 4th In-

ternational Conference on Agents and Artiﬁcial Intel-

ligence, pages 146–153.

Helbing, D. and Johansson, A. (2009). Pedestrian,

crowd and evacuation dynamics. In Encyclopedia of

Complexity and Systems Science, pages 6476–6495.

Springer.

Hoogendoorn, M., Treur, J., van der Wal, C., and van Wis-

sen, A. (2010). Modelling the interplay of emotions,

beliefs and intentions within collective decision mak-

ing based on insights from social neuroscience. In

Wong, K., Mendis, B., and Bouzerdoum, A., edi-

tors, Neural Information Processing. Theory and Al-

gorithms, volume 6443 of Lecture Notes in Computer

Science, pages 196–206. Springer Berlin / Heidelberg.

Hoogendoorn, S. and Bovy, P. (2004). Pedestrian route-

choice and activity scheduling theory and models.

Transportation Research Part B: Methodological,

38(2):169–190.

Kretz, T., Hengst, S., and Vortisch, P. (2008). Pedes-

trian ﬂow at bottlenecks - validation and calibration

of vissim’s social force model of pedestrian traf-

ﬁc and its empirical foundations. Technical Report

arXiv:0805.1788. Comments: Accepted for presenta-

tion at International Symposium of Transport Simula-

tion 2008 (ISTS08) in Gold Coast, Australia.

Kułakowski, K. and Gawro´nski, P. (2009). To cooperate or

to defect? altruism and reputation. Physica A: Statis-

tical Mechanics and its Applications, 388(17):3581–

3584.

Luber, M., Stork, J., Tipaldi, G., and Arras, K. (2010). Peo-

ple tracking with human motion predictions from so-

cial forces. In 2010 IEEE International Conference

on Robotics and Automation (ICRA), pages 464–469.

IEEE.

Moussa¨ıd, M., Perozo, N., Garnier, S., Helbing, D., and

Theraulaz, G. (2010). The walking behaviour of

pedestrian social groups and its impact on crowd dy-

namics. PLoS One, 5(4):e10047.

Murakami, Y., Minami, K., Kawasoe, T., and Ishida, T.

(2002). Multi-agent simulation for crisis management.

In 2002. Proceedings. IEEE Workshop on Knowledge

Media Networking, pages 135–139. IEEE.

Ozel, F. (2001). Time pressure and stress as a factor during

emergency egress. Safety Science, 38(2):95–107.

Pan, X. (2006). Computational Modeling of Human and

Social Behaviors for Emergency Egress Analysis. PhD

thesis, Stanford University, Stanford, California.

Qingge, J. and Can, G. (2007). Simulating crowd evacua-

tion with a leader-follower model. IJCSES, 1(4):250.

Qiu, F. (2010). A Framework for Group Modeling in Agent-

Based Pedestrian Crowd Simulations. PhD thesis,

Georgia State University, Atlanta, Georgia.

Qiu, F. and Hu, X. (2010). Modeling group structures in

pedestrian crowd simulation. Simulation Modelling

Practice and Theory, 18(2):190–205.

Russell, S. and Norvig, P. (2010). Artiﬁcial Intelligence: A

Modern Approach. Prentice Hall.

Santos, G. and Aguirre, B. E. (2004). A critical review of

emergency evacuation simulation models. In Build-

ing Occupant Movement during Fire Emergencies,

Gaithersburg, Maryland.

Shehory, O. and Kraus, S. (1998). Methods for task allo-

cation via agent coalition formation. Artiﬁcial Intelli-

gence, 101(1-2):165 – 200.

Tsai, J., Fridman, N., Bowring, E., Brown, M., Epstein, S.,

Kaminka, G., Marsella, S., Ogden, A., Rika, I., Sheel,

A., Taylor, M. E., Wang, X., Zilka, A., and Tambe,

M. (2011). Escapes: Evacuation simulation with chil-

dren, authorities, parents, emotions, and social com-

parison. In Proceedings of the International Confer-

ence on Autonomous Agents and Multiagent Systems

(AAMAS).

Yang, L., Zhao, D., Li, J., and Fang, T. (2005). Simulation

of the kin behavior in building occupant evacuation

based on cellular automaton. Building and Environ-

ment, 40(3):411–415.

Zhao, D., Yang, L., and Li, J. (2008). Occupants’ behavior

of going with the crowd based on cellular automata

occupant evacuation model. Physica A: Statistical

Mechanics and its Applications, 387(14):3708–3718.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

364