SIMULATING PEDESTRIAN ROUTE SELECTION WITH

IMPERFECT KNOWLEDGE

Kyle Feuz and Vicki Allan

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

Keywords:

Multi-agent systems, Pedestrian simulation, Heuristic route selection, Emergency egress.

Abstract:

Heuristic evaluation of possible route choices allows pedestrians to make decisions in a timely and efﬁcient

manner. The heuristic function used to evaluate the route and the subsequent route selection has a large impact

on the egress time of the pedestrian. We implement several common heuristic functions using the PLEASE

simulation model and allow these heuristics to be combined using weighted factors. When the total distance

of a route is unknown, using a greedy strategy of selecting the shortest-leg ﬁrst route is shown to be a poor

choice. When combined with other heuristic estimates however, including shortest-leg ﬁrst costs can help to

decrease egress times. We show that for a variety of building layouts using a heuristic function based upon

width, distance, signage and congestion levels leads to better egress times.

1 INTRODUCTION

In everyday life, pedestrians frequently utilize heuris-

tics to make potentially complicated decisions within

an allotted time frame. These heuristics often allow a

person to make a reasonable decision without requir-

ing extensive amounts of time and energy. One ex-

ample arises from the evacuation of a burning build-

ing. Selecting the best route can mean the difference

between life and death, but frequently insufﬁcient in-

formation is used in making this decision. Currently,

detailed information about pedestrian egress from ac-

tual emergencies is not widely available, and thus it

is too costly, too dangerous, and impractical to de-

termine what heuristics pedestrians use to select an

egress route when in emergency situations. Pedestrian

simulation models have been designed to address this

problem.

Current pedestrian simulation models generally

fall into one of two categories when determining

pedestrian route selection. Either the model assumes

pedestrians have perfect knowledge of the building

layout and are thus able to select the best route, or the

model assumes that pedestrians only know about the

immediately visible routes and must make a decision

based upon some heuristic (Pan, 2006; Thompson and

Marchant, 1995; Gwynne et al., 2001). Models which

assume perfect knowledge are clearly unrealistic for

many situations as not all pedestrians will be famil-

iar with any given building layout. Yet models re-

lying upon a heuristic may also be unrealistic in the

amount of information provided to pedestrians. Typi-

cally such heuristic models consider distance, conges-

tion or social comparison for making the route selec-

tion decision. In this paper, we show that using dis-

tance as the sole means of comparing visible routes

leads to poor egress times when the total route dis-

tances are unknown and is actually equivalent to the

shortest-leg ﬁrst heuristic which Golledge found in

(Golledge, 1995) to be less preferred than many other

heuristics. Therefore, if total route distances are as-

sumed to be unknown to the pedestrians, then other

route selection heuristics must be used to supplant the

unknown information.

In this paper, we consider the effect of various

heuristic functions on pedestrian egress time for a

variety of different building layouts. Determining

a comprehensive list of heuristic functions, which

pedestrians might use, is outside of the scope of this

paper. However, several common factors including

distance, signage, corridor width, congestion/usage,

and common consensus are used to produce a variety

of realistic heuristic functions.

2 RELATED WORK

In (Ozel, 2001), Ozel raises several pertinent ques-

tions regarding the issue of stress management and the

decision process. He suggests that pedestrians utilize

146

Feuz K. and Allan V..

SIMULATING PEDESTRIAN ROUTE SELECTION WITH IMPERFECT KNOWLEDGE.

DOI: 10.5220/0003726201460153

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 146-153

ISBN: 978-989-8425-96-6

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

various coping mechanisms (heuristics) to make a de-

cision in a time-pressured environment. In particular,

the familiarity of routes and the negative connotations

of emergency exits are shown to have a large impact

on the route choice of pedestrians in an emergency.

Hoogendoorn and Bovy use distance and con-

gestion heuristics for their route-choice and activity

scheduling model (Hoogendoorn and Bovy, 2004).

Gwynne et. al. indicate that pedestrians maintain so-

cial roles and norms even during emergency situations

(Gwynne et al., 2006). This may lead to pedestri-

ans choosing a different route based on the choices of

their peers. Similarly, Fridman and Kaminka develop

a pedestrian simulation model based upon social com-

parison theory (Fridman and Kaminka, 2007).

Golledge ranks the preference and prevalence of

several different heuristic values pedestrians use in

navigation and route selection in an outdoor environ-

ment such as a college campus through the use of

questionares and observations (Golledge, 1995). He

ﬁnds that pedestrians prefer routes which are direct,

quick, and easy to navigate with some preference be-

ing given to routes which are more aesthetically pleas-

ing. In emergency situations, one would expect the

characteristics of directness and quickness to remain

prominent while the importance of scenery is irrele-

vant. Golledge also found that the route selection cri-

teria used differed for various route layouts and that

a combination of multiple criteria may give better re-

sults.

3 HEURISTICS

Our research evaluates a variety of heuristics. We de-

scribe several possible heuristics in the section.

Distance is a common heuristic used by pedestri-

ans when making a route selection and is the primary

heuristic in many search based strategies (Hoogen-

doorn and Bovy, 2004). Unless motivated by other

factors, pedestrians will choose the route which has

the shortest distance (Golledge, 1995). The dis-

tance heuristic is inapplicable in the absence of com-

plete route distance knowledge because missing data

causes the distance relation to be a partial order.

The shortest-leg ﬁrst heuristic is a greedy dis-

tance heuristic. Rather than utilizing the total route

distance, shortest-leg ﬁrst selects the route with the

shortest distance to the next decision point. Pedes-

trians only require distances to route goals which are

within their ﬁeld of view.

Another common heuristic used by pedestrians

when selecting a route is the least angle heuristic

(Dalton, 2003; Hochmair and Frank, 2000). This

heuristic selects the path which is closest in terms of

angularity to a direct line between their current po-

sition and the ﬁnal goal (B

uchner et al., 2007). For

example, in Fig. 1 the agent would select Route B

when using the least angle heuristic because it is the

route whose angle differs the least from the angle to

the goal. Unfortunately, the end goal location(s) is

not always known. In such situations, the least-angle

heuristic cannot be applied.

(a) Example using the least-

angle heuristic

(b) Depicts the values used to

calculate the angle heuristic

Figure 1: Least-Angle heuristic examples.

We can modify the least angle heuristic to accom-

modate unknown goals by selecting the closest path in

terms of angularity to the current walking orientation.

This new heuristic is actually a greedy version of the

fewest turns heuristic as it tries to minimize the num-

ber of direction changes an agent makes while exiting

from the building.

Most buildings include navigational signs to assist

pedestrians in locating their desired location (Kray

et al., 2005). These signs might include exit signs,

emergency exits signs, room signs, or navigational

maps. Pedestrians then use these signs to navigate

effectively through the building. One might think that

during an emergency evacuation pedestrians would

look for exit signs or emergency exit signs, but Ozel

found that emergency exit signs often have negative

connotation and are avoided even in emergency situ-

ations (Ozel, 2001). Exit signs, on the other hand, are

commonly used by pedestrians in navigational plan-

ning when in an unfamiliar building. Room signs can

also be used in egress route planning as they indicate

that an egress route through a particular door is un-

likely because individual rooms rarely contain egress

routes.

Main corridors tend to be preferred by pedestri-

ans especially when navigating an unfamiliar building

(Hillier, 1996). Like the relative width of various cat-

egories of roads, main corridors are wider than aux-

iliary hallways so the width of the corridor or door-

way can frequently be an indication of a main route

of travel leading to an exit.

The number of pedestrians currently gathered

around a location may be either a positive or a nega-

SIMULATING PEDESTRIAN ROUTE SELECTION WITH IMPERFECT KNOWLEDGE

147

tive indicator of a desirable exit route (Gwynne et al.,

2006; Pan, 2006; Koh et al., 2008). It may indicate

the route is a good choice because others are using

it. However, it may also indicate that another route

should be considered to avoid the congestion. Intu-

ition suggests that if pedestrians are unsure about the

situation, they will follow others. If they are more

conﬁdent and know multiple routes, they will seek al-

ternatives to avoid congestion.

Unless the pedestrians are traveling in a group,

they will not know the exact route of nearby pedes-

trians. However, by observing the velocity and past

movement, an agent may predict the immediate des-

tination of a neighboring pedestrian. Seeing agents

who are traveling in the same direction may bolster

the conﬁdence level of an agent in choosing a partic-

ular route (Ozel, 2001). Traveling in the same direc-

tion as others is also easier as one does not have to

ﬁght against the general direction ﬂow, improving the

overall pedestrian ﬂow.

Pedestrians will typically choose routes with

which they are familiar especially when under time

constraints (Ozel, 2001). Familiarity allows the

pedestrian to feel more comfortable and conﬁdent in

the route selected and allows the pedestrian to con-

centrate on other cognitive tasks. However, select-

ing a new route may or may not lead to a better solu-

tion. When the current known route cost is within an

acceptable limit, pedestrians are unlikely to change.

The cost of a route may be based upon any number

of factors such as distance, congestion, time, and so

on. Readers may refer to (Feuz, 2011) for details of

including distance and congestion costs as learned by

past experience.

4 SIMULATION ENVIRONMENT

This study is performed using the Pedestrian Leader-

ship and Egress Assistance Simulation Environment

(PLEASE). PLEASE is built upon the multi-agent

modeling paradigm where each pedestrian is repre-

sented as an individually rational agent capable of

perceiving the environment and reacting to it. In

PLEASE, pedestrian agents can perceive obstacles,

hazards, routes, and other agents. The agents use a

two tier navigational module to control their move-

ment within the simulation environment. The high-

level tier evaluates available routes and selects a des-

tination goal. The low-level tier, based on the social

force model (Helbing and Johansson, 2009), performs

basic navigation and collision avoidance.

PLEASE implements several of the heuristics out-

lined in Section 3. Here we describe the implementa-

tion details used for the heuristics of interest. Many

simulation models assume that total route distance is

known and/or the end goal location is known for at

least some subset of the available routes. However,

in this paper, we are interested in the case when no

additional information, beyond what is immediately

visible to the pedestrian, is provided. For this reason,

the distance, angle and learned-costs heuristics are not

considered.

Each heuristic can potentially be combined with

any other heuristic. To facilitate the integration of

multiple heuristics, all heuristic values are normalized

so that the unweighted cost falls between 0 and 1.

The leg cost of route x, L(x), is given by Formula

1 where w

is the weight applied to the shortest-leg

heuristic, and d

i,x

is the distance from agent i to route

goal x. The distance is normalized using the maxi-

mum distance between two points on the simulation

map. Agents in the simulation are able to accurately

estimate the distance to visible points within the sim-

ulation model. The distance to locations which are

occluded by walls or other obstacles cannot be esti-

mated without prior knowledge.

L(x) =

∗ d

i,x

(maxDistance)

(1)

The turn cost of route x, T(x), is given by Formula

2, where w

is the weight applied to the fewest turns

heuristic, and a

i,x

is the angle in radians between the

orientation of agent i and the direction to route goal x

from agent i. π acts as the normalization factor since

no angle will be greater than π. See Figure 1 for an

example.

T (x) = w

∗ a

i,x

/π (2)

The signage cost of route x, S(x), is given by For-

mula 3, where w

is the weight applied to the signage

heuristic, and getExitWeight(x) is the cost associated

with the given signage value. getExitWeight takes the

signage of a route and looks up the user-speciﬁed cost

of that sign. A user may specify arbitrary sign costs,

but by default, PLEASE uses the costs shown in ta-

ble 1. To be consistent with the other heuristics, cost

should be speciﬁed as a value between 0 and 1.

S(x) = w

∗ getExitWeight(x) (3)

Table 1: Default parameters used in PLEASE.

Sign Type Cost Width Cost

Emergency Exit 0.25 Small 1.0

Exit 0.0 Medium 0.5

Room 1.0 Large 0.0

None 0.5

The simple signage cost of route x, SS(x), is given

by Formula 4, where w

is the weight applied to the

ICAART 2012 - International Conference on Agents and Artificial Intelligence

148

simple signage heuristic, and getSimpleSignage(x) is

the cost associated with the simple signage of route x.

A route has a simple signage cost of 0 if the doorway

of the route is a direct exit and is marked with an exit

sign. All other routes have a simple signage cost of 1.

This causes a pedestrian to ignore all building signage

except for exit signs over direct exits. The simple sig-

nage heuristic allows us to compare the effect that dif-

ferent amounts of building signage and different lev-

els of attention to the building signage have upon the

egress times of pedestrians.

SS(x) = w

∗ getSimpleSignage(x) (4)

The width cost of route x, W (x), is given by For-

mula 5, where w

is the weight applied to the corri-

dor width heuristic, and getCorridorWeight(x) is the

cost assigned to the given corridor width. The getCor-

ridorWeight is a lookup table which takes the width

of a route and returns the user-speciﬁed cost for that

width. A user may enter any arbitrary width cost, but

by default, PLEASE uses the costs shown in table 1.

To be consistent with the other heuristics, cost should

be speciﬁed as a value between 0 and 1. Although

corridor width is a real number and can have an in-

ﬁnite number of values, the getCorridorWeight dis-

cretizes width into three categories, small, medium

and large. This is done to eliminate meaningless dif-

ferences between corridors. Two corridors which dif-

fer only slightly in width are equally likely to indicate

a main route and should be treated equally. The cut-

off values for these categories can be set by the user so

that the values are appropriate for the current building

layout. By default, PLEASE uses 1.5 m and 2.5 m as

the cutoff values. Agents in the simulation measure

width at the entry point of the corridor.

W (x) = w

∗ getCorridorWeight(x) (5)

The congestion cost of route x, Cong(x), is given

by Formula 6, where w

is the weight applied to

the congestion heuristic, sp

is the desired speed (a

normally distributed parameter value unique to each

pedestrian) of pedestrian i, sp

is the speed of pedes-

trian j, s

is 1 if sp

< sp

or 0 otherwise, n

p,x

is the

number of pedestrians along route x, and n

is the to-

tal number of pedestrians. This formula assigns cost

based upon the desired speed of the pedestrian and the

current speed of pedestrians along the selected route.

For each pedestrian along the selected route, if their

speed is slower than the desired speed, then a cost is

incurred relative to the speed difference. The cost is

raised to the square so that smaller speed differences

count less than larger differences. Finally, the result is

normalized by the worst case cost (i.e. if every pedes-

trian in the simulation was along the selected route

and was not moving).

Cong(x) = w

∗

p,x

∑

j=0

((sp

− sp

) ∗ s

)

∗ n

(6)

The consensus cost of route x, C(x), is given by

Formula 7, where w

is the weight applied to the con-

sensus heuristic, a

j,x

is the angle between pedestrian

j’s orientation and the direction to route goal x from

agent j (see Figure1), and n

p,i

is the number of pedes-

trians surrounding pedestrian i.

C(x) = w

∗

p,i

∑

j=0

j,x

/π)

p,i

(7)

The previously visited cost of x, V (x), is given

by Formula 8, where w

is the weight applied to the

visited heuristic, and visitedCount(x) is the number

of remembered times route x has been visited by the

pedestrian in this simulation run. Each pedestrian is

capable of remembering a speciﬁed limit of number

of routes for a speciﬁed amount of time to reﬂect the

ﬁnite memory of pedestrians. These limits may be

set to any arbitrary value by the user, but by default,

PLEASE has a limit of 10 routes for 1000 seconds.

V (x) = w

∗ visitedCount(x)/maxMemory (8)

5 EXPERIMENTAL DESIGN

To test the effectiveness of the various heuristics de-

scribed above, we use a combination of actual build-

ing layouts and building layouts constructed for the

purpose of these experiments. In the buildings shown

in Figure 2, blue lines represent walls of the build-

ings, dashed green lines represent exits signs, solid

red lines represent emergency exit signs, and wavy

gold lines represent door signs. The USU Business

building (see Figure 2(a)) is an approximation of the

ground ﬂoor of the actual building found on the Utah

State University Campus. Likewise, the CSULB FM

building (see Figure 2(c)) and the CSULB UP build-

ing (see Figure 2(b)) are approximations of the ac-

tual buildings found on the California State Univer-

sity, Long Beach campus.

The USU Business building and the CSULB UP

building are used because the ﬂoor plans follow ex-

pected conventions: rooms do not function as hall-

ways, main areas have wide corridors, and other ac-

cepted conventions are followed. The CSULB FM

building is used because of the lack of convention

with room placement. Rooms are found within rooms

and several rooms can even function as hallways. The

custom building is designed to represent a standard

SIMULATING PEDESTRIAN ROUTE SELECTION WITH IMPERFECT KNOWLEDGE

149

(a) USU Business Building Floor 1 (b) CSULB UP

Building

Figure 2: Building layouts used in the heuristic evaluation experiments.

building. The corridor widths are not an exact indi-

cation of main routes and areas, but are still closely

correlated to main routes. Rooms do not function as

hallways. The number of rooms and routes are not so

many as to be completely unmanageable by pedestri-

ans either.

For each building, we conduct a variety of tests.

We measure the total egress time of 100 pedestrians,

randomly distributed throughout the rooms, averaged

over 20 simulations using a single active heuristic. In

all the tests, the previously visited heuristic is always

active and is given a weight of ﬁve. This represents

the agents’ unwillingness to backtrack along a route

previously traveled.

The viability of the signage heuristic is completely

dependent upon the type and amount of signage found

with the buildings. For this reason, a comparison be-

tween the signage and simple signage heuristics is

helpful. The simple signage heuristic represents the

same building stripped of all signs except for exit

signs at actual exit doorways.

Based upon the performance of the individ-

ual heuristics, we then combine multiple heuristics,

weighting each heuristic by its relative performance

in relation to the total egress time achieved by the

heuristics. These weights are then adjusted to further

improve the performance of the combined heuristics.

Determining the exact weight speciﬁcation to opti-

mize performance when multiple heuristics are used

is outside of the scope of this paper.

6 RESULTS AND ANALYSIS

In this section, we discuss the results of the exper-

iment in four parts. First, we look at the effect of

building layouts on egress times. Second, we consider

in detail the result of applying only one heuristic at a

time. Third, we look at the results of applying mul-

tiple heuristics simultaneously. Fourth, we rank each

heuristic using a relative performance ratio.

6.1 Building Layouts

The USU Business building and the CSULB-UP

building have similar results for several of the heuris-

tics (see Figure 3). This makes sense because both

buildings have similar design characteristics. The

main differences in heuristic performance occurred

with the shortest-leg heuristic. The USU Business

building layout happens to be conducive to using the

shortest-leg heuristic. The outside rooms can actually

be used as hallways which lead directly to the exits,

and this is exactly what happens when the shortest-

leg heuristic is used. The ground ﬂoor of the USU

Business building layout appears similar to the cus-

tom building layout, yet the slight differences in spac-

ing and room layout create large differences in egress

times for the two buildings.

shortestLeg

signage

simpleSignage

width

congestion

random

Relative Performance Ratio

Heuristic Function

Effect of Building on Heuristic Performance

Custom

USU Business

CSULB FM

CSULB UP

Figure 3: Effect of building layout on heuristic effective-

ness. Performance ratio is calculated as the average time

taken for 90% of the pedestrians to evacuate using the given

heuristic divided by the average time taken for 90% of the

pedestrians to evacuate using perfect knowledge.

The CSULB FM building and the custom build-

ing both have similar results in egress times for the

shortest-leg, signage, congestion and random heuris-

tics. Although the actual layouts of the building are

quite different, the underlying patterns are similar:

ICAART 2012 - International Conference on Agents and Artificial Intelligence

150

width correlates with egress routes, long hallways

have many adjoining rooms, and exits are distributed

in a uniform manner. The main difference between

the results for each building is the performance of the

width heuristic. In the CSULB FM building, corridor

width corresponds closely with egress routes, while

in the custom building the correlation is weaker. The

simple signage heuristic also yields different results

in these two buildings. In the custom building, the

top left and top right exits are not visible throughout

most of the building and are thus highly underutilized.

Meanwhile, the exits in the CSULB FM building have

a much higher level of visibility throughout the build-

ing and are used more effectively.

6.2 One Heuristic

When only a single heuristic is used, the results vary

greatly between heuristics (see Figure 4). The signage

heuristic performs well in all types of building lay-

outs tested. This suggests that if pedestrians choose

an egress route based upon well-designed signage,

pedestrians can efﬁciently egress from a building even

when completely unfamiliar with the layout of the

building. Using simple signage, the egress times are

still as good as or better than any other heuristic in

the building layouts considered. This highlights the

importance of even minimal building signage in as-

sisting in pedestrian egress.

The shortest-leg heuristic leads to slow egress

times in almost every building layout considered. In

many instances, the shortest-leg heuristic does not

even outperform a random choice policy. As dis-

cussed in Section 3, when the end goal is not known,

choosing the route which is closest to the pedestrian

becomes the shortest-leg ﬁrst heuristic. This greedy

route selection heuristic provides no guarantee that

the route chosen will even lead to a direct exit. Ad-

ditionally, when distance is the sole means of eval-

uating a route, congestion is a common occurrence.

Pedestrians who are closest to a given doorway will

select that doorway regardless of what side they are

on or which direction other pedestrians are moving.

Thus, the pedestrians on opposite sides of the door-

way will converge at the doorway causing a bottle-

neck, and pedestrian ﬂow rates through the doorway

will be greatly inhibited. Interestingly, the shortest-

leg heuristic performs remarkably well in the USU

Business Building (see Figure 4). This building is

conﬁgured ideally so that greedily selecting the clos-

est visible route actually leads pedestrians to an exit

in a fairly efﬁcient manner. One reason that this is the

case is the double doors on most of the rooms. This

allows the pedestrian to explore routes without having

to backtrack, which is discouraged by the algorithm.

Additionally, the end rooms have doorways adjacent

to exits, which facilitates egress in this situation.

Similar to the shortest-leg heuristic, the fewest

turns heuristic also leads to poor performance. The

results are not shown here. The intuition behind the

fewest turns heuristic is to select a route that is as di-

rect as possible. However, considering only the next

route goal is too short-sighted and leads to routes

which are drastically less direct than they could be.

Without prior knowledge about the building layout,

though, this short-sightedness cannot be overcome.

For the building layouts considered in this paper,

the width heuristic leads to average egress times when

compared to the other heuristics. For the CSULB-FM

building, choosing the widest route leads to ﬁnding an

exit sooner than selecting a route by any other heuris-

tic except for signage. In the remaining buildings, the

width heuristic performs worse than the congestion

heuristic, but still signiﬁcantly outperforms a random

policy.

The congestion heuristic does not necessarily pro-

vide an indication of which route leads to an exit,

especially when none of the pedestrians have any

knowledge regarding the building layout. However,

avoiding congestion still improves the overall egress

time by helping prevent bottlenecks and increasing

the overall smoothness of pedestrian ﬂow. This al-

lows more routes to be explored in less time, which

leads to better egress times. Although (due to space

limitations) the results are not shown here, the con-

sensus heuristic also relieves congestions at bottle-

necks and improves pedestrian ﬂow so that routes

can be explored in a more efﬁcient manner. To be

most effective, the congestion and consensus heuris-

tics should be combined with another heuristic such

as width or signage which provide an indication of an

egress route.

6.3 Multiple Heuristics

After considering each heuristic individually, we then

combine several heuristics to further improve perfor-

mance (see Figure 5). Although many different com-

binations could be tried, in this paper, we consider

combining the simple signage and width heuristics

(SS-W) and the shortest-leg, simple signage, width,

and congestion (SL-SS-W-C) heuristics. When width

is the sole heuristic applied, the egress times are too

slow to be reliable in an emergency. The simple sig-

nage heuristic is also slower than desired but is still

the best alternative to the signage heuristic (which

may not be realistic for many buildings) when only

a single heuristic is applied. The goal of combining

SIMULATING PEDESTRIAN ROUTE SELECTION WITH IMPERFECT KNOWLEDGE

151

0.2

0.4

0.6

0.8

0 50 100 150 200 250 300

Percent Evacuated

Time (seconds)

Business Building

100 200 300 400 500

Time (seconds)

CSULB-FM Building

50 100 150 200 250 300

Time (seconds)

CSULB-UP Building

50 100 150 200 250 300

Time (seconds)

Custom Building

shortestLeg

signage

simpleSignage

width

congestion

random

shortestLeg

signage

simpleSignage

width

congestion

random

shortestLeg

signage

simpleSignage

width

congestion

random

shortestLeg

signage

simpleSignage

width

congestion

random

Figure 4: Agent egress times using multiple heuristics in the Simple building.

width and simple signage is to take advantage of the

signage when available and to fall back on the width

heuristic when the signage is not available . We then

include the other heuristics, namely shortest-leg and

congestion, to further improve the egress times. For

comparison purposes, the egress times of the signage,

simple signage and width heuristics are included in

the charts.

As can be seen in Figure 5, combining the sim-

ple signage and width heuristic (SS-W) did indeed

improve performance in most buildings in compari-

sion to either heuristic alone. The custom building

layout is the one exception. In this building, falling

back on width proved to be detrimental to the over-

all egress time as the widest areas did not have di-

rect exits. However, when several heuristics (shortest-

leg, simple signage, width and congestion, denoted

as SL-SS-W-C) are combined, performance is im-

proved in every single building layout when com-

pared to the performance of the heuristics separately.

In most cases, the egress times matched or beat select-

ing routes based upon perfect signage. This indicates,

that for a wide variety of buildings, the same heuristic

functions can be applied to successfully egress from

the building in a reasonably efﬁcient manner using

only the information which is directly perceivable by

the pedestrian.

Although the shortest-leg heuristic did not per-

form well by itself, when combined with other heuris-

tics, it leads to improved performance (results omit-

ted due to space limitations). This is indicative of the

value distance can play in route selection and justiﬁes

its use by actual pedestrians. However, it is important

to note the disastrous impact which relying only upon

distance can have upon the total egress time of indi-

vidual pedestrians when no additional information is

utilized.

6.4 Heuristic Rankings

The heuristic functions are ranked according to the

relative performance ratio (RPR) of each heuristic in

the above mentioned building layouts. For each build-

ing, the average time (t) it takes for 90% of the pedes-

trians to evacuate when each pedestrian has perfect

knowledge of route distances and congestion levels is

recorded. The average time (h) it takes for 90% of the

pedestrians to evacuate when each pedestrian is us-

ing the heuristic function of interest is also recorded.

The RPR of each heuristic function is then computed

as h/t. Table 2 displays the average relative perfor-

mance ratio for each heuristic.

Table 2: Ranking of heuristic functions by relative perfor-

mance ratio (RPR). A lower RPR signiﬁes better egress

times.

Heuristic RPR Heuristic RPR

signage 1.44 congestion 4.30

S-SS-W-C 1.59 consensus 4.65

SS-W 2.71 random 7.60

simpleSignage 2.89 shortestLeg 8.19

width 3.62 angle 9.41

7 CONCLUSIONS

Using heuristic estimations in selecting an egress

route is a natural and common process performed by

pedestrians on a daily basis, yet most simulation mod-

els do not adequately address this fact. This paper

highlights the importance of including heuristic costs

in pedestrian simulators, especially those designed to

model egress in an emergency situation. As Ozel indi-

cates in (Ozel, 2001), when in stressful situations and

under time constraints, pedestrians will react by ﬁlter-

ing and bolstering information (i.e. relying more upon

heuristic estimations). This leads to decisions which

are sub-optimal and, as shown, can have a signiﬁcant

impact on the total egress time of the simulation. In

emergency simulation models, it is not sufﬁcient to

assume pedestrians will make the best or even good

choices, the simulation model needs to consider the

ICAART 2012 - International Conference on Agents and Artificial Intelligence

152

0.2

0.4

0.6

0.8

0 50 100 150 200

Percent Evacuated

Time (seconds)

Business Building

50 100 150 200

Time (seconds)

CSULB-FM Building

50 100 150 200

Time (seconds)

CSULB-UP Building

50 100 150 200

Time (seconds)

Custom Building

SL-SS-W-C

SS-W

signage

simpleSignage

width

SL-SS-W-C

SS-W

signage

simpleSignage

width

SL-SS-W-C

SS-W

signage

simpleSignage

width

SL-SS-W-C

SS-W

signage

simpleSignage

width

Figure 5: Agent egress times using multiple heuristics in the Simple building.

possibility that pedestrians are forced to make split-

second decisions with little information.

When the total distance is unknown, using a

greedy strategy of selecting the closest route available

is seen to produce poor results in many circumstances.

If additional factors are included in the decision, then

the distance heuristic can help improve egress times,

even when the total distance is not known. Using four

main heuristics (the shortest-leg heuristic, the simple

signage heuristic, the width heuristic and the conges-

tion heuristic, each appropriately weighted) is shown

to produce good egress times even when no informa-

tion about the building layout is known before the

simulation begins. If only a single heuristic is used,

the signage heuristic gives the best results even when

the amount of building signage is minimal.

REFERENCES

uchner, S., H

olscher, C., and Strube, G. (2007). Path

choice heuristics for navigation related to mental rep-

resentations of a building. In Proceedings of the Euro-

pean Cognitive Science Conference, pages 504–509.

Taylor & Francis.

Dalton, R. (2003). The secret is to follow your nose. Envi-

ronment and Behavior, 35(1):107.

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

simulation environment. Master’s thesis, Utah State

University.

Fridman, N. and Kaminka, G. (2007). Towards a cogni-

tive model of crowd behavior based on social compar-

ison theory. In Proceedings of the National Confer-

ence on Artiﬁcial Intelligence, volume 22, page 731.

Menlo Park, CA; Cambridge, MA; London; AAAI

Press; MIT Press; 1999.

Golledge, R. (1995). Path selection and route preference

in human navigation: A progress report. Spatial In-

formation Theory A Theoretical Basis for GIS, pages

207–222.

Gwynne, S., Galea, E., and Lawrence, P. (2006). The intro-

duction of social adaptation within evacuation mod-

elling. Fire and materials, 30(4):285–309.

Gwynne, S., Galea, E. R., Lawrence, P. J., and Filippidis, L.

(2001). Modelling occupant interaction with ﬁre con-

ditions using the buildingexodus evacuation model.

Fire Safety Journal, 36(4):327–357.

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

crowd and evacuation dynamics. In Encyclopedia of

Complexity and Systems Science, pages 6476–6495.

Springer.

Hillier, B. (1996). Space is the Machine. Cambridge Uni-

versity Press, Cambridge.

Hochmair, H. and Frank, A. (2000). Inﬂuence of estima-

tion errors on wayﬁnding-decisions in unknown street

networks–analyzing the least-angle strategy. Spatial

Cognition and Computation, 2(4):283–313.

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

choice and activity scheduling theory and models.

Transportation Research Part B: Methodological,

38(2):169–190.

Koh, W. L., Lin, L., and Zhou, S. (2008). Modelling and

simulation of pedestrian behaviours. In PADS ’08:

Proceedings of the 22nd Workshop on Principles of

Advanced and Distributed Simulation, pages 43–50,

Washington, DC, USA. IEEE Computer Society.

Kray, C., Kortuem, G., and Kr

uger, A. (2005). Adaptive

navigation support with public displays. In Proceed-

ings of the 10th international conference on Intelligent

user interfaces, IUI ’05, pages 326–328, New York,

NY, USA. ACM.

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.

Thompson, P. and Marchant, E. (1995). Testing and ap-

plication of the computer model simulex. Fire Safety

Journal, 24(2):149–166.

SIMULATING PEDESTRIAN ROUTE SELECTION WITH IMPERFECT KNOWLEDGE

153