ERAM

Evacuation Routing using Ant Colony Optimization over Mobile Ad Hoc Networks

Alejandro Avilés del Moral

, Munehiro Takimoto

and Yasushi Kambayashi

Department of Information Sciences, Tokyo University of Science, Tokyo, Japan

Department of Computer & Information Engineering, Nippon Institute of Technology, Saitama, Japan

Keywords: Ant Colony Optimization, Global Positioning System, Mobile Ad Hoc Networks, Mobile Agents,

Multiagent Systems.

Abstract: This paper proposes a distributed multi-agent framework for discovering and optimizing evacuation routes

on demand. We name it Evacuation Routing using Ant Colony Optimization over Mobile Ad hoc Networks

(ERAM). Taking advantage of ant colony optimization (ACO) on mobile ad hoc networks (MANETs)

composed of smartphones with geo-location capabilities, ERAM aims for adaptability and layout

independence, relying exclusively on crowd’s knowledge during mass evacuations. Such knowledge is

inserted into the system: actively, by users’ indication of having reached safe areas on their smartphones;

and passively, by smartphones tracking their own movement. In the framework, agents migrate through

nodes of a MANET towards safe areas based on an indirect communication mechanism called stigmergy,

which is a behaviour that social insects show. Once an agent finds such an area, it traces its path backwardly

collecting geographical information of intermediate nodes for composing an evacuation route. During the

backward travel, agents lay pheromone down while they migrate back based on the ACO algorithm,

strengthening quasi-optimal physical routes, and hence guiding succeeding agents. This scenario is

analogous to data-packet routing on Internet or resource discovery on P2P networks, except it routes people

through physical environments towards safe areas instead.

1 INTRODUCTION

In cases of emergency, people in a building, facility

or even whole cities might need to be evacuated.

Known routes, however, found to be congested or

compromised can lead crowds to panic due to

occasional lack of indications to reach a safe area.

Therefore, it is desirable to have an infrastructure-

less adaptable framework to assist finding routes on

demand without any previous specific preparation of

roads or buildings in order to increase the survival

rate.

We apply swarm intelligence based routing

mechanisms that are well known to be useful in

virtual environment as shown in (Ducatelle et al.,

2010); (Ducatelle, 2007) and (De Rango and

Socievole, 2011) for crowd flow guidance in the real

world through handheld devices such as

smartphones.

Nowadays, smartphones are increasingly

relevant in daily life among all kinds of users. This

incessant growth is derived from some features such

as portability, which traditional computers lack. This

feature as well as relatively small size can be applied

to solve old problems in new ways. In particular,

there are two key features for the implementation of

our framework. Whereas the smartphone’s wireless

connectivity allows the deployment of a swarm

intelligent system, the built-in Global Positioning

System (GPS) reception can provide information

about the physical environment.

The combination of these technologies on a

single portable device along with swarm

mechanisms assists to find not only main evacuation

routes but also alternative ones, while taking in

account of the congestion in the routes. The found

routes can reflect unpredictable human behavior

through dynamically supporting safety policies

(Proulx, 2001).

We propose a distributed multi-agent distributed

framework for discovering optimized evacuation

routes on demand. We henceforth call this

framework Evacuation Routing Using Ant Colony

Optimization over Mobile Ad hoc Networks

118

Avilés del Moral A., Takimoto M. and Kambayashi Y..

ERAM - Evacuation Routing using Ant Colony Optimization over Mobile Ad Hoc Networks.

DOI: 10.5220/0004261301180127

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 118-127

ISBN: 978-989-8565-38-9

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

(ERAM). This framework came from AntHocNet

(Di Caro et al., 2004), (Di Caro, Ducatelle, &

Gambardella, 2005) and (Di Caro et al., 2005), a

framework for routing in Mobile Ad hoc Networks

(MANETs) and the algorithm for resource discovery

in P2P networks presented in (Kambayashi and

Harada, 2007) and (Kambayashi and Harada, 2009).

Both approaches use Ant Colony Optimization

(ACO) providing indirect communication between

agents in order to achieve self-organized

optimization.

ACO is a biologically inspired framework based

on ant foraging behaviors, as proposed in (Colorni,

Dorigo, & Maniezzo, 1992) and (Dorigo and

Gambardella, 1997). Ant agents laying pheromone

down reinforces paths that are more likely to lead

towards fruitful nodes. This way, pheromone level

indicates the goodness of a node regarding the goal,

and by constantly being modified by ant agents, the

system get quasi-optimized in run-time. This indirect

method is called stigmergy.

The structure of the balance of this paper is as

follows. In the second section, we describe an

evacuation scenario in which ERAM performs the

guidance for the users. The third section describes

ERAM framework. ERAM consists of several static

and mobile agents. The static agents interact with

users and store routing information, and the mobile

agents effectively work together with stigmergy to

find optimal evacuation routes. The fourth section

demonstrates the usefulness of the framework with

the results of numerical experiments on a simulator

we built. Finally, the fifth section discusses

conclusion and future works.

2 SCENARIO

ERAM presents a novel hybrid solution for network-

physical environment routing, a field that has not

investigated so far. In contrast with other swarm

intelligence based routing frameworks such as the

ones proposed in (Kambayashi and Harada, 2007)

and (Ziane and Melouk, 2005), which basically

focus on data packet routing over networks, ERAM

aims to discover physical resources over physical

layouts instead of virtual ones. ERAM, however,

uses networks as a means to locate those resources,

and thus previous researches on ACO routing

provide a robust background.

2.1 Scenario Description

As derived from the previous section, in an

evacuation scenario resources correspond to physical

safe areas that users proactively mark on their

smartphones as soon as they are reached, each

referred to as node. Paths discovered with ERAM

take into account physical layouts. The GPS is used

to provide geographic information that will be

collected on ERAM as described later. This

information is also used to evaluate the congestion

of crowd flows and to optimize them.

In order to perform resource discovery and

routing, active and passive sources of information

are required. Whereas users provide active input

when they reach safe areas, smartphones provide

passive input by constantly storing GPS tracks.

Finally, optimal routing discovery both in

network and in physical environments is achieved by

indirect communication through migrations of

mobile agents over nodes of a MANET applying the

ACO algorithm described in the next section.

In order to make the scenario mentioned above

technically feasible, we assume that each node has

the following functionalities:

 Built-in Wi-Fi connectivity to construct the

MANET with other smartphones.

 Making use of decent GPS signal.

2.2 Scenario Difficulties

A real implementation of ERAM framework,

however, implies some difficulties mainly related

with the reliability of the information.

One problem, which is easier to handle, is that of

relying on human knowledge in a stressful situation

to obtain the routes. To deal with this, the

framework assumes that, at least, a few people are

able to reach a safe area, i.e. shelter, out of the

building, etc. and mark it on their smartphone. Then,

by the unsupervised usage of the pheromone, those

nodes closer to these areas will be more likely to

lead people to a secure one.

Another problem, not yet fully addressed in this

paper, is the impossibility to avoid fake marking of

safe areas. A proposal of solution is weighting safe

flags in relation of its geographic density based on

the fact that the majority of people will not lie. In

this manner, the more safe flags of which nodes are

close to, the more likely they are to attract mobile

agents, and thus, to produce correct evacuation

routes avoiding those leading to fake safe areas.

Another way to enforce reliability is a distributed

trust protocol, as proposed in (Nakamoto, 2008),

ERAM-EvacuationRoutingusingAntColonyOptimizationoverMobileAdHocNetworks

119

which makes almost impossible to fake the system.

Additionally, GPS signal is maybe the most

troublesome aspect of the framework. It is hardly

available indoors and obstacles like trees or

buildings outdoors can easily hinder precision. In

order to obtain precise GPS samples, there are some

solutions widespread at the moment. One of them is

the usage of differential GPS, which relies on static

devices with known position, if available. Another

one is using external sensors or measurement of Wi-

Fi signal to estimate absolute position like the one

proposed in (Woodman and Harle, 2008), but this

relies on infrastructure or preloaded information.

There are also more innovative approaches such as

(Bejuri et al., 2011), which propose a taxonomic

GPS that would provide precise and ubiquitous

positioning.

3 ERAM FRAMEWORK

In this section, ERAM framework is described in

detail. It first covers the agents that are involved in

the consecution of the goal, conditions of the

algorithm to guarantee the negotiability of the

evacuation route, and the formal definition of the

pheromone calculation.

3.1 Agents

Due to the fact that similarities found in

formulations of ERAM’s scenario with the one

proposed in (Kambayashi and Harada, 2007), the

solutions are also similar. Especially, its concept of a

set of collaborative static and mobile agents for

discovering resources in P2P networks has also been

adapted to ERAM. This concept is required to deal

with the complexity and uncertainty of what hybrid

network‐physical scenario presents.

Fig. 1 depicts the different kinds of agents

participating in the framework. The following

descriptions are brief explanation of their tasks.

1. Information Agent (IA): it is a static agent

resides on each node. It is in charge of

perceiving physical context as well as storing

knowledge about its hosting node. It also acts as

an interface with the user, receiving input and

providing directions to follow.

2. Node management Agent (NA): it is a static

agent also residing on each node. It is in charge

of storing knowledge of the network context,

namely storing information about other nodes.

3. Goal Agent (GA): it is a mobile agent that

floods the network with information about the

Figure 1: ERAM participators. Nodes host mobile agents

(RAs and GAs), and static agents (IA and NA). Users

communicate with IA. IA creates RAs and GAs. Mobile

agents migrate towards other nodes.

safe area recently reached. An IA on a safe node

creates GA.

4. Routing Agent (RA): it is a mobile agent that

tries to find safe nodes in order to obtain

evacuation routes. IA on a node in search of

evacuation routes, which is called swarm node,

creates RA.

The interactions between them can be illustrated as

shown in Figure 2. We describe the details of each

agent below.

Figure 2: Interactions between agents. Here, IA provides

the node GPS log and receives GPS log collections for

calculations; NA provides IP address to mobile agents,

and both RAs and GAs updates pheromone value on the

NA.

3.1.1 Information Agent (IA)

Each node hosts an IA and it keeps track of the GPS

log by taking position samples periodically.

Whenever another one node comes within Wi-Fi

range of the node on which the IA resides, it starts

the handshaking protocol in order to initiate session.

This can be achieved as shown in SLS (Gajurel and

Heiferling, 2009), but storing the whole GPS log

instead of a single position. Adjacent connected

nodes constantly exchange GPS logs between IAs

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

120

and all received information is stored in the NA. We

describe NA in the next section.

User can interact with the IA by activating the

safe flag or by activating the evacuation search.

These two modes are mutually exclusive.

 Safe flagged: When a user reaches a safe area,

he or she hoists the safe flag in his or her

smartphone. Once the safe flag is activated,

the IA starts creating GAs periodically in

order to flood the network with information

about the safe area position.

 Searching evacuation: When a user is in need

of evacuation route, he or she sets his or her

smartphone in search mode. In this process,

the IA produces RAs constantly, making them

migrate to search negotiable paths towards

safe areas. Once one RA returns from a safe

flagged node storing GPS logs of intermediate

nodes, the IA composes an evacuation route

from them. If the estimated time of the new

evacuation route is better, the current one is

replaced by it. Finally, the route is provided to

the user in a form of directions to be followed.

As pointed in (Rodriguez and Amato, 2010) and

(Rodriguez and Amato, 2011), routes composed of

pairs of positions should be smoothened to reduce

unnecessary loops.

3.1.2 Node Management Agent (NA)

In addition to an IA, every node hosts an NA. The

NA keeps information about neighbour nodes such

as IP address, GPS log and other data used for

pheromone calculation. Each entry is created when

the IA successfully initiates session with neighbour

nodes, which are updated as mobile agents arrive

and give new information.

Table 1: Knowledge encoded within an NA.

IP:port GPS log Jumps SAV GA id

9.163.251.45:50000 {(3,4), ...} 4 0.8 2435

92.71.112.26:50001 {(0,1), ...} 1 0.1 8431

… … … … …

As shown by Table 1, an NA encodes the basic

knowledge in its matrix to cover the following

fields.

 Jumps: the number of hops from a safe

flagged node. It is updated by GAs.

 Speed Ahead Value (SAV): this encodes

congestion level and it is modified by RAs. It

is covered later in detail.

 GA id: the unique identification of the GA

that last updated the entry. This is used for

stopping GAs from migrating endlessly.

When mobile agents need to migrate, they request

the NA to provide them the IP addresses of the

nodes to which they migrate next.

3.1.3 Goal Agent (GA)

A safe flagged node’s IA creates GAs in order to

disseminate the information in the network about the

safe area. The GA, therefore, requests the NA

information of all the connected nodes, and clones

itself on each of them. As the GA migrates through

the network, it counts the number of hops it has

performed. When it steps into a new node, this

information is provided to the NA of that node in

order to update the Jumps entry of the neighbouring

nodes. This information is used for subsequent

pheromone calculation.

Figure 3: Migration activity of a GA through a simple ad

hoc network. Numbers represent Jumps value. Arrows

with an X at the end mean the GA is killed.

Figure 4: Condition for RA to perform a migration.

The terminating condition of GA’s migration is

as follows. When it is created, every GA has a

unique identification. Thus, if it steps into a node

with the same GA id as it has, it means that has been

previously visited by one of the other clones of the

GA, and therefore it is killed. GA’s life span is

graphically explained in Figure 3.

3.1.4 Routing Agent (RA)

A routing agent is an agent created by an IA on a

node that searches for safe areas. The RA migrates

while it satisfies the condition where a GPS log on a

ERAM-EvacuationRoutingusingAntColonyOptimizationoverMobileAdHocNetworks

121

new node to which it is migrating has some

intersections with the GPS log of the current node as

shown in Figure 4. It means that there are some

intersections between GPS logs such as a chain, and

this is possible because the NA of the hosting node

stores the GPS log of nodes surrounding that node.

This condition guarantees that nodes can

physically reach the safe areas following paths

derived from GPS logs without meeting any

obstacles.

In order to optimize this evacuation route, the

RA will choose the node with the highest

pheromone value at present. As ACO states, this is

merely probabilistic as explained in (Di Caro et al.,

2006); therefore, there are situations in which

random paths are chosen in order to explore other

possible evacuation routes.

If a situation in which a sequence of connecting

nodes meets the mentioned condition and it reaches

a safe flagged node, the RA becomes a backward

RA immediately. The backward RA follows the

same path as the forward RA but in the opposite

direction.

There may be some situations in which the

forward RA hits node having an evacuation route

already created. In this case the RA will copy the

evacuation route and become a backward RA.

Figure 5: When GPS logs are not exactly intersecting but

proximity can be used for calculating the evacuation route.

On its way back, the RA collects the GPS logs of

the nodes while laying pheromone down in the NAs

on the nodes. This pheromone, represented by Speed

Ahead Value (SAV), reflects the congestion present

along the path. When the RA hits the original node

where it was created, the GPS logs are provided to

the IA. Once this happens the evacuation route is

established.

3.2 Route Construction

The route composed from GPS logs is expected to

be an evacuation route because a chain of

intersecting GPS logs ends in a safe flagged node,

which is in a safe area.

This is the expected result of an algorithm used

to compose the evacuation route, when intersecting

GPS logs have been received. The composed route

avoids loops and unnecessary paths by deleting

those parts that do not lead towards a safe flagged

node.

As depicted in Figure 5, sometimes the GPS logs

may be not exactly intersecting but should be

considered for having physical reachability. In such

a case, fuzzy approaches provide flexibility for

dealing with uncertainty as well as with constraints

that accepts some degree of error.

In our system, instead of binary functions

returning whether some intersections exist or not, a

radial basis function is utilized to weaken

intersection constraints. Notice here that such a

radial basis function needs a predefined threshold of

distance to be regarded as an intersection between

GPS tracks.

3.3 ACO Algorithm

The following equation determines the pheromone

level of a node.

pheromone







speed



jumps



∗SAV



(1)

speed

is the average speed of node n, using the

GPS track stored in its IA. The length of the time

window of the GPS track affects the sensibility of

this parameter. This implies that the shorter the

length of the window becomes, the fewer samples

for calculating the parameter we get. That results in

a highly responsive value, causing the last few

seconds of the movement to be reflected more

precisely.

jumps

reflects how many migrations are needed

to reach the closest safe flagged node. Due to an

agents wireless capabilities this is a heuristically

determined parameter as it positively estimates how

far the node is from a safe area. Even if agents can

migrate through obstacles such as walls, human

beings will not be able to pass through.

SAV

stands for Speed Ahead Value. It serves the

purpose of estimating congestion ahead. For SAV,

having a lower value means that nodes ahead are

moving more slowly, therefore the probability of

congestion is higher and RAs will be less likely to

follow these nodes while seeking a safe flagged

node.

As described in (Kambayashi and Harada, 2007),

our framework does not implement a pheromone

decay system through evaporation. Instead, thanks to

RA constantly migrating through the network, this

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

122

value is frequently updated and serves the purpose

of pheromone decay mechanism.

SAV of any safe node is 1. SAV of a non-safe

given node n starts at 1 and it is recalculated based

on: its SAV previous value and both the SAV and

average speed from the node straight ahead. The RA

carries that information. In this way, knowledge

about congestion is propagated backwardly through

the network while updating each node through

iterations.

The formal calculation is as follows.



















1





1



(2)

Where 01, and













1

, w

and w

are predefined weights that

determine how SAV will be modified through

iterations. Fine-tuning these three constants is a key

to obtain a self-optimizing routing system.

 w1: represents how slowly SAV is modified

through iterations. The higher the weight is,

the slower SAV will be affected. Having

values near to 1 means that it will hardly

change, and near to 0 means that will be

overridden every iteration.

 w2: represents how much overall congestion

ahead from node n influences its SAV

calculation. Having values near to 1 means

that initial SAV will be carried all the way

back.

 w3: represents how much the speed of the

node straight ahead influences on the node

SAV. Having values near to 1 means that

SAV only represents congestion of the node

straight ahead.

In addition, the variable speedNorm is a

normalizer of speed that keeps the value between 0

and 1.

In summary, the pheromone level of a node is

directly proportional to the average speed of

movement. The pheromone level is inversely

proportional to the number of migrations to the

closest safe area. Thus, the pheromone level is

penalized by distance from a safe area and

congestion such as a pheromone decay mechanism.

4 RESULTS

In order to demonstrate the effectiveness of our

framework, we have built a simulator and conducted

Figure 6: Simulation screenshot with 300 nodes (10%

swarm). Blue nodes are informed nodes, pink nodes

swarm nodes, and red nodes swarm nodes with evacuation

route. Green grounds represents safe area, white ground

floor within the building, and black impassable walls. It

also shows Wi-Fi connections between nodes, GAs and

RAs, and GPS logs in red trajectories.

numerical experiments. The results addressed in this

section were acquired using a simulation

environment that provides graphical representation

of the framework and statistical results for further

analysis. We call this simulation environment

ERAMsim.

First, we give a brief explanation of the

simulator; second, we describe the simulation

settings; and last, we discuss the simulation results.

4.1 Simulator Overview

ERAMsim is a software program developed for

running basic simulations of ERAM framework.

Providing simplified both physical and network

environment simulations, it implements the ACO

algorithm as well as the behaviour of individuals

trying to reach a safe area in a floor plan, as depicted

in Figure 6.

It can be perceived that white ground represents

the interior of a building whereas green grounds are

safe areas outside the building and black lines are

walls. Consequently, the basic unit for space is the

pixel, and the time is discrete.

The simulator takes account of physical

constraints such as maximum density of people or

the impossibility of them to pass through others,

producing bottlenecks in doors. It also assumes that

there are some nodes that already know where to go,

referred as to informed nodes, and others using

ERAM exclusively, swarm nodes. As soon as any of

them reaches a safe area it becomes a safe node,

ERAM-EvacuationRoutingusingAntColonyOptimizationoverMobileAdHocNetworks

123

starting to produce GAs, and then walking away

from the building.

4.2 Simulation Settings

Table 2 shows the default values of the main

parameters used for running the simulator.

These parameters are clarified as follows:

 ‘Total nodes’ determines how many

individuals-smartphone are participating in the

evacuation. 200 nodes provide a density of

4.71%.

 ‘Swarm nodes’ configures the percentage of

the total nodes using ERAM to find their way.

 ‘GPS log length’ refers to the maximum

amount of GPS samples that are stored within

a NA with a frequency of 0.25 samples/round.

 ‘Wi-Fi range’ determines the range, in pixels,

of the Wi-Fi signal of the smartphones.

 ‘ACO random’ determines the chance of a RA

to migrate randomly instead of following the

best pheromone value.

 w

1~3

are the weights for SAV calculation.

 ‘GA/RA cooldowns are the number of rounds

that takes for a node to produce the next

mobile agent.

Table 2: Parameters default values.

Parameter Value Parameter Value

Total nodes 200 w

0.1

Swarm nodes 10% nodes w

0.5

GPS log length 100 samples w

0.4

Wi-Fi range 70 pixels GA cooldown 50 rounds

ACO random 10% RA cooldown 5 rounds

We use discrete time, a floor plan with

800px*600px, and nodes represented by

circumferences of 6px radius moving at a maximum

speed of 2px/round. Approximating these values to

meters, nodes are 70cm diameter circumferences,

from which it can be derived that a pixel is 5.83cms

and they can move up to 1.16 m/s in a 46.7m x 35m

physical environment.

Finally, default values have been chosen for the

purpose of keeping the performance metric low,

which enables the behavior to be easily analyzed

through modifying them.

4.3 Simulation Results

The results focus on the success rate of finding the

evacuation route rather than on the overall route

optimization. This is due to the fact that for

measuring the improvement, comparison between

virtual evacuation time using ERAM and real

evacuation time is required, and in order to obtain

the real one, sophisticated human behaviour

simulation is needed.

In these results, the performance is measured as

the success rate of finding evacuation route in terms

of the previously defined variables previously

defined. Assuming that all swarm nodes that

successfully reach a safe area are caused by the

usage of ERAM, success rate of the route discovery

algorithm can be formally defined as follows.

success rate 

swarmnodessaved

swarmnodes

(3)

Because of the probabilistic nature of the simulator,

in order to get reasonable results, each sample

reflects the average of five simulations using the

same parameter values.

4.3.1 GPS Log Length

When configuring a real implementation of ERAM,

one of the main parameters is the GPS log length.

Every node needs to exchange this information

constantly with connected nodes in order to reflect

lag-free geographic information. This data exchange

is critical for the framework but it is also the most

bandwidth consuming, and thus, it is crucial to size

it optimally.

Figure 7: Success rate depending on GPS log length.

As can be seen in Figure 7, the longer the length

of the GPS log is, the higher the success rate of route

discovery becomes. This is caused by the fact that, it

is more probable to find intersections between GPS

logs, and consequently for the system to allow more

agent migrations. By easing agent migration, route

discovery of swam nodes increases and so does the

success rate.

Nevertheless, this metric is mainly bounded by

two facts, both responsible of the plateau at around

70% success rate in Figure 7. The first one is related

20%

40%

60%

80%

100%

10 50 100 150 200 250 300 350 400

GPSloglength(samples)

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

124

to a physical constraint and is impossible to avoid. It

is that regardless the GPS log length there might be

nodes located where others have never come,

making any GPS intersection nonexistent. The

second one has to do with technical limitations and

is interesting as can be predicted more easily. It is

that although there are GPS intersections, if the Wi-

Fi signal is much shorter than the GPS log, nodes

cannot be aware of it, and therefore migrations will

not take place.

4.3.2 Wi-Fi Signal

Studying how Wi-Fi signal influences the success

rate is interesting, because it not only determines the

bounding the GPS log length, but also it eventually

provides minimum technical requirements for the

ERAM framework to work properly.

Figure 8: Success rate depending on Wi-Fi range.

Figure 8 shows an improvement in success rate

similar to the one exposed for the GPS log length.

Again, by increasing the Wi-Fi signal more nodes

will be connected in the MANET, and this way,

more GPS log intersections will be found by agents.

On the other hand, the GPS log length also

bounds the improvement of Wi-Fi signal. Regardless

the range of the signal connecting nodes, technically

allowing agents to migrate, when GPS logs are not

long enough, intersections will hardly occur and

then migration constraint will not be satisfied.

With this simulation configuration, having a Wi-

Fi signal 1.5 times wider than the GPS log length

produces a success rate of 93%.

4.3.3 Density and Swarm Nodes

Finally, we discuss the interest of using ERAM

depending on the density of people in the evacuation

and how the framework performs when only a few

of them know the way to a safe zone.

Note that swarm nodes in this simulation walk

towards safe zones only when they have obtained an

evacuation route through ERAM. Hence

measurements do not concern about 0% or 100%

swarm nodes scenarios. Having none swarm node

would mean there is nothing to measure, and having

all swarm nodes would mean no human knowledge

about safe areas and therefore success rate equal to

zero.

Figure 9: Success rate depending on density and

percentage of swarm nodes.

As shown in Figure 9, as density grows the

overall success rate increase, highlighting that

ERAM presents superior performance for mass

evacuations. Given the map the simulation uses, 100

nodes mean a density of 2.36%, and 500 nodes

11.78%.

This improvement is derived from two main

facts: first of all, the topology of the MANET mesh

is more interconnected; and secondly, more GPS

logs stored in the system leads to more GPS

intersections.

Another dimension that Figure 9 explores is the

impact of swarm nodes percentage. As can be

observed, as the number of swarm nodes increases

the success rate decreases in low-density settings.

This can be explained because the more swarm

nodes participating in the evacuation, the less

informed nodes, and thus, only a few safe nodes will

be produced. In addition, having low densities

makes the MANET connectivity too reduced,

producing some isolated sub-MANETs with fewer

chances to transmit agents towards one of those

scarce safe nodes.

When density increases, however, and

consequently so does MANET connectivity, RAs

laying pheromone down makes, even for highly

isolated MANETs, the success rate of this

migrations increase up to almost 100% of swarm

nodes. Eventually this makes almost all the swarm

nodes reach safe areas even when only a few nodes

know the way to the safe area.

20%

40%

60%

80%

100%

50 75 100 125 150 175 200

Wi-Fi range (pixels)

20%

40%

60%

80%

100%

10% 40% 70% 90%

Swarm nodes (%)

500 nodes

400 nodes

300 nodes

200 nodes

100 nodes

ERAM-EvacuationRoutingusingAntColonyOptimizationoverMobileAdHocNetworks

125

These results indicate that encouraging the active

usage of ERAM (swarm nodes) for evacuations in

scenarios such as public buildings, sport events, or

cities will provide good results.

It should be clear that this metric is extremely

pessimistic due to the simplistic human behavior

simulation. As in a real situation a swarm node can

become an informed one asking other people,

following signs or by visual contact with the safe

area.

5 CONCLUSIONS AND FURTHER

WORKS

ERAM represents a novel usage of ACO for hybrid

environments. Routing people by the passive and

active usage of the smartphone opens a new horizon

of study. This project provides not only the first

formal definition of the framework but also a

software tool to test it.

A great amount of uncertainty and imprecision

will be present in a real scenario because of the

usage of commercial GPS, and unsuitable conditions

for geo-location. The simulation results, however,

indicate that the direction of our investigation is

right. It is now clear that smartphone and ACO are a

suitable combination to develop the new safe

evacuations systems.

Our experiments on the simulator so far show

promising results regarding the evacuation route

discovery algorithm, especially for high-density

evacuation. This first iteration of ERAM is,

however, the initial stage, and still far from

achieving the optimal evacuation routing. Further

works can be summarized in three main topics:

framework refinement, simulator realism, and

extensive experimentation.

In regard of framework refinement, first,

pheromone calculation of the ACO algorithm needs

to find the proper values for weights. One approach

to solve this while at the same time making it even

more flexible should be applying some sort of neural

network to dynamically adapt the weights to the

different type of context, i.e. pedestrian or car

evacuations. It is also interesting incorporate into the

framework solutions for scenario difficulties as

described in section 2.

ERAMsim needs further works in order to

demonstrate the feasibility of the framework in

different scenarios under much more complex

environments. For doing so, it will be necessary to

adopt more realistic simulation approaches. On the

one hand, a realistic network simulation would allow

measuring the congestion of agents’ interactions

produce as well as obtaining the actual time it takes

to find an evacuation route. On the other hand,

utilizing physical environment simulations as shown

in (Rodriguez and Amato, 2010) and (Rodriguez and

Amato, 2011) would allow to measure how route

optimizing routing improves survival rate.

The simulator, although must be improved, is

good enough to show a current state and to explore

the limits of this framework by retrieving results

beyond the ones shown in this paper.

Concluding, and regarding the feasibility of a

future real implementation, we believe that we have

enough technological components to implement

ERAM. The usage of smartphones is being

widespread. We are trying to implement an ad hoc

network using solely smartphones. Upon completion

of such network, it is possible to have a mobile agent

system without explicit network infrastructures. It is

also possible to have alternative location system

other than GPS. For example, it should be relatively

easy to construct infrastructure using RFID rather

than GPS that provides the precision needed to

provide dynamic routing in controlled environments

such as museums for providing new features besides

evacuation routing.

REFERENCES

Bejuri, W. M., Mohamad, M. M., & Sapri, M. (2011).

Ubiquitous Positioning: A Taxonomy for Location

Determination on Mobile Navigation System.

International Journal of Signal & Image Processing ,

2 (1), 24-34.

Colorni, A., Dorigo, M., & Maniezzo, V. (1992).

"Distributed Optimization by Ant Colonies". Proc. 1st

European Conference on Artificial Life, (pp. 132-142).

Cambridge.

De Rango, F., & Socievole, A. (2011). "Meta-Heuristics

Techniques and Swarm Intelligence in Mobile Ad Hoc

Networks". In Mobile Ad-hoc Networks: Applications.

Di Caro, G., Ducatelle, F., & Gambardella, L. (2006). "An

analysis of the different components of the AntHocNet

routing algorithm". Proc. 5th International Workshop

on Ant Colony Optimization and Swarm Intelligence

(ANTS2006). Brussels.

Di Caro, G., Ducatelle, F., & Gambardella, L. (2005).

"AntHocNet: An adaptive nature-inspired algorithm

for routing in mobile ad hoc networks". EUROPEAN

TRANSACTIONS ON TELECOMMUNICATIONS, 16,

pp. 443-455.

Di Caro, G., Ducatelle, F., & Gambardella, L. (2004).

"AntHocNet: an Ant-Based Hybrid Routing Algorithm

for Mobile Ad Hoc Networks". Proc. 8th International

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

126

Conference on Parallel Problem Solving from Nature

(PPSN VIII). Birmingham,.

Di Caro, G., Ducatelle, F., & Gambardella, L. (2005).

"Using Ant Agents to Combine Reactive and Proactive

Strategies for Routing in Mobile Ad Hoc Networks".

International Journal of Computational Intelligence

and Applications (IJCIA) , 5 (2), 169-184.

Dorigo, M., & Gambardella, L. (1997). "Ant Colony

System: A Cooperative Learning Approach to the

Traveling Salesman Problem". IEEE Transactions on

Evolutionary Computation, 1, pp. 53-66. Brussels.

Ducatelle, F. (2007). "Adaptive Routing in Ad Hoc

Wireless Multi-hop Networks". PhD Thesis, Università

della Svizzera Italiana, Istituto Dalle Molle di Studi

sull'Intelligenza Artificiale, Lugano.

Ducatelle, F., Di Caro, G., & Gambardella, L. (2010).

"Principles and applications of swarm intelligence for

adaptive routing in telecommunications networks".

Swarm Intelligence , 4 (3).

Gajurel, S., & Heiferling, M. (2009). "A Distributed

Location Service for MANET Using Swarm

Intelligence". Proc. IEEE Mobile WiMAX Symposium,

(pp. 220-225).

Kambayashi, Y., & Harada, Y. (2007). "A Resource

Discovery Method Based on Multi-agents in P2P

Systems". Proc. 1st KES International Symposium on

Agent and Multi-Agent Systems: Technologies and

Applications, (pp. 364-374). Verlag.

Kambayashi, Y., & Harada, Y. (2009). "Integrating Ant

Colony Optimization in a mobile-agent based resource

discovery algorithm". Proceedings of the IADIS

International Conference on Intelligent Systems and

Agents (pp. 149-158). Antonio Palma dos Reis.

Nakamoto, S. (2008). "Bitcoin: A Peer-to-Peer Electronic

Cash System". From Bitcoin:

http://bitcoin.org/bitcoin.pdf

Proulx, G. (2001). "Occupant Behaviour and Evacuation".

Proc. 9th International Fire Protection Symposium,

(pp. 25-26).

Rodriguez, S., & Amato, N. (2010). "Behavior-based

evacuation planning". Proc. IEEE International

Conference on Robotics and Automation (ICRA), (pp.

350-355). Anchorage.

Rodriguez, S., & Amato, N. (2011). "Utilizing Roadmaps

in Evacuation Planning". Proc. Computer Animation

and Social Agents, (pp. 67-73).

Woodman, O., & Harle, R. (2008). Pedestrian localisation

for indoor environments. Proc. of the 10th

international conference on Ubiquitous Computing

(pp. 114-123). New York: ACM.

Ziane, S., & Melouk, A. (2005). "A swarm intelligent

multi-path routing for multimedia traffic over mobile

ad hoc networks". Proc. Q2SWinet, (pp. 55-62). New

York.

ERAM-EvacuationRoutingusingAntColonyOptimizationoverMobileAdHocNetworks

127