Multi-agent Systems for Estimating Missing Information in Smart Cities

Davide Andrea Guastella

1,2

, Val

erie Camps

and Marie-Pierre Gleizes

Institut de Recherche en Informatique de Toulouse, Universit

e de Toulouse III - Paul Sabatier, France

Universit

a degli Studi di Palermo, Italy

Keywords:

Smart City, Cooperative Multi-agent Systems, Missing Information Estimation.

Abstract:

Smart cities aim at improving the quality of life of citizens. To do this, numerous ad-hoc sensors need to be

deployed in a smart city to monitor the environmental state. Even if nowadays sensors are becoming more and

more cheap their installation and maintenance costs increase rapidly with their number. This paper makes an

inventory of the dimensions required for designing an intelligent system to support smart city initiatives. Then

we propose a multi-agent based solution that uses a limited number of sensors to estimate at runtime missing

information in smart cities using a limited number of sensors.

1 INTRODUCTION

The concept of Smart City refers to a territorial

context where the use of human and natural re-

sources, properly managed through different Informa-

tion and Communications Technologies (ICTs), al-

lows the creation of an ecosystem that provides in-

tegrated and more intelligent systems (Roscia et al.,

2013). The concept of smart city remains strongly

ambiguous; it has been growing from empirical ex-

periences and therefore a systemic theoretical study

about this phenomenon still lacks (Dameri, 2013). A

work commissioned by the European Union deﬁnes

smart city initiatives as multi-stakeholder municipally

based partnerships aimed at addressing problems of

common interest with the aid of ICTs, which sup-

ports the smart classiﬁcation (Manville et al., 2014).

Dameri (Dameri, 2013) analyzed ﬁve areas of study

concerning the smart city: (i) intelligent city, (ii) dig-

ital city, (iii) sustainable city, (iv) technocity and (v)

well-being city. However, even if such distinction al-

lows to deﬁne precisely what is a smart city, these

concepts share some common aspects. Thus, they do

not represent disjoint areas of analysis. In this paper

we focus on intelligent, sustainable and well-being

city. An intelligent city is able to produce knowl-

edge and to translate it into unique and distinctive

abilities; this city is smart because it is able to cre-

ate intellectual capital and to ground development and

well-being on this intellectual capital. A sustainable

city uses the technology to reduce CO2 emissions, to

produce efﬁcient energy, to improve the buildings ef-

ﬁciency; it aims at becoming a green city. A well-

being city aims at producing the best quality of life

for citizens, but also to create regional attractiveness

both for people and for business. The technology is

only a part of the means used to reach these goals;

also culture, climate, history and monuments are con-

sidered as important success factors (Dameri, 2013).

Well-being is commonly related to the user’s comfort.

Thermal, visual, luminosity and noise are some of the

main indicators used to deﬁne a comfortable environ-

ment (Frontczak and Wargocki, 2011). Thermal stan-

dards are required to help building designers to pro-

vide an indoor climate that occupants will perceive

as thermally comfortable (Wong et al., 2014; Ghahra-

mani et al., 2018; Herkel et al., 2008).

Because the smart city concept embraces multidis-

ciplinary ﬁelds, it is important to provide a short de-

scription of Internet Of Things (IoT) and Ambient In-

telligence concepts. IoT aims at providing a global in-

frastructure for the information society, enabling ad-

vanced services by interconnecting physical and vir-

tual things based on existing and evolving interop-

erable information and communication technologies.

The main challenge of the IoT is to achieve full in-

teroperability of interconnected devices while guar-

anteeing the trust, privacy and security of communi-

cations (Piette et al., 2016). These interconnected de-

vices become more unobtrusive and thanks to their

embedded sensors they can perceive the physical en-

vironment in which they are situated. Ambient Intel-

ligence provides to ambient systems the mechanisms

necessary to carry out environmental reasoning using

214

Guastella, D., Camps, V. and Gleizes, M.

Multi-agent Systems for Estimating Missing Information in Smart Cities.

DOI: 10.5220/0007381902140223

In Proceedings of the 11th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2019), pages 214-223

ISBN: 978-989-758-350-6

a representation of the environment perceived by IoT

devices. Ambient systems are designed to provide

adapted services that respond to an individual, collec-

tive, and social requirement. The term environment

refers to a physical space enriched with sensors and

computational entities that are seamlessly and invisi-

bly interwoven (Pirttikangas et al., 2010). To be con-

sidered as smart or intelligent, an environment needs

to be associated with a representative description that

can be constructed from the perceptions of the ambi-

ent components. These includes the IoT devices that

must be able to interact with other components that

are not known a priori by humans/users. The inter-

actions of ambient components enable a smart city to

enhance its services such as transports, health, cul-

tural events and so on. Nevertheless, avoiding the in-

stallation of new components in an ambient system to

provide precise everywhere and anytime information

on the environment is a difﬁcult task.

After a presentation of some different application

domains for the smart-city in section 2, we present

the main dimensions in designing a system to sup-

port smart cities. Section 3 introduces our research

context as well as our proposition to estimate miss-

ing environmental information in smart contexts. Our

contribution, based on a cooperative multi-agent sys-

tem, allows to avoid the installation of ad hoc sensors.

In section 4 we evaluate the proposed solution using

a real weather dataset. Section 5 brieﬂy describes our

research direction and perspectives.

2 SMART CITIES AND

MULTI-AGENT SYSTEMS

As stated in the previous section, the deﬁnition of

smart city embraces different area of analysis and

ﬁnds application in different domains. Reviewing all

the smart city applications goes beyond the purpose

of this paper. For this reason we explore different

applications in which the concept of smart city takes

place in order to emphasize the relevance of multi-

agent systems in this domain.

2.1 Multi-agent Systems Application

Domains in Smart Cities

The growing power of sensors and connected devices

makes the Smart Grid gain much attention. A smart

grid can be deﬁned as an autonomous electrical net-

work able to adapt itself to client’s needs in a se-

cured, ecological and economical way. It enables

bidirectional exchanges of electricity and information

through lines. Perles et al. propose an approach based

on multi-agent systems to estimate the voltage of each

bus in an electric network without having sensors in-

side each bus (Perles et al., 2017). As the system

is speciﬁcally devoted to the smart grid, information

employed are not heterogeneous.

Roscia et al. propose a model of smart city

that that employs multi-agent systems (Roscia et al.,

2013). The proposed model is based on a system of

systems that embraces different technologies to pro-

vide a basic infrastructure for the deﬁnition and the

creation of a smart city. The composition of these

systems will change as technology evolves, generat-

ing new businesses and new interactions. For each

model domain of smart city, each individual device

is associated to an software agent: its behavior is de-

cided speciﬁcally according to the domain in which

the agent takes place.

Smart Health provides artiﬁcial intelligence and

cognitive computing in order to assist the doctors

when they have to interpret medical data and to es-

tablish the right diagnosis for their patients. In the

context of the 3Pegase project, whose main goal is to

offer an efﬁcient solution to follow-up old people at

their home, an approach based on a multi-agent sys-

tem has been proposed to detect at runtime behav-

ioral anomalies by using feedbacks from the medical

staff (Verstaevel et al., 2018). The system employs

different sensors that track the activity of the users.

This analysis lets the system rise alerts when a deviant

behavior is detected. This solution is not suitable for

large scale applications. As the system uses a limited

number of sensors deployed in a home context, its de-

ployment at large scale remains difﬁcult. Moreover,

the amount of data processed is limited with respect

to the quantity of data produced by sensors at smart

city scale.

Cook et al. propose the MavHome project whose

goal is to create a smart home that acts as a ratio-

nal agent, perceiving the state of the home by means

of sensors (perception of light, humidity, tempera-

ture, smoke, gas, motion) and acting on the envi-

ronment through effectors (in this case, device con-

trollers) (Cook et al., 2006). An agent acts in order

to maximize its goal, which is a function that maxi-

mizes comfort of the inhabitants and minimizes op-

erational costs. This solution assumes that data from

sensors are always available; thus unpredictable situa-

tions where information from sensors are unavailable

are not taken into account.

Karnouskos et al. propose an agent-based solu-

tion for simulating the dynamic behavior of a smart

city (Karnouskos and Nass de Holanda, 2009). Their

proposition simulates discrete heterogeneous devices

Multi-agent Systems for Estimating Missing Information in Smart Cities

215

that consume and/or produce energy. The software

agents, associated to real-world devices, are able to

monitor efﬁciently the consumption of a high num-

ber of devices. agents are not cooperative and their

behavior is decided a priori. However cooperation

would enable agents to acquire more knowledge and

experience from other agents in order to improve their

own knowledge or act in a more appropriate way in

the environment in which they are situated.

2.2 Challenges in Smart Cities

When dealing with systems to support smart cities, we

have to consider ﬁve different challenges (Guastella

and Valenti, 2018). (i) Openness: the system must

be able to work with intermittent devices. For ex-

ample, devices such as city users’ smartphones are

not always available; (ii) Heterogeneity: the obser-

vations that come from heterogeneous devices pro-

duce large volumes of data that have to be pruned and

correlated in order to generate valuable knowledge;

(iii) Large-scale: due to the amount of entities (phys-

ical and virtual) involved in smart cities and the huge

amount of data to process, there is a need for efﬁ-

cient data storing and manipulation techniques. Also,

the data must be always available to the ﬁnal user;

(iv) Unpredictability: systems to support smart city

initiatives have to be able to continuously self-adapt

to changes that may occur in a high dynamic environ-

ment; (v) Privacy: non-intrusiveness is a key point

when collecting data from ad-hoc sensors in a smart

city context.

Multi-Agent Systems are a promising way to

address these challenges. Indeed, agents are au-

tonomous, they are capable to reasoning and are pro-

acting, thus enabling a system to be intelligent and

able to make anticipations (Olaru et al., 2013). More-

over, each agent has its own local perceptions, knowl-

edge and goals (Georg

e et al., 2011). The conception

of a system for addressing the described challenges

requires an ever-increasing reliability that centralized

systems could not provide due to their low perfor-

mance in precise tasks. On the other hand multi-agent

systems are able to get high performance thanks to

their local, distributed intelligence and self-adaptation

ability. The propositions presented in section 2.1

show how multi-agent ambient systems can be used

to assist users in a smart city context. However, there

is a need to design multi-agent systems that are able to

operate in a highly dynamic environment with hetero-

geneous and intermittent sensors by using entry-level

instrumentation. This means that the system must not

require any speciﬁc type of device to operate.

Despite the advantages of the reviewed state-of-art

propositions, to the best of our knowledge the prob-

lem of conceiving an efﬁcient technique to estimate

missing environmental information in smart contexts

using mobile and intermittent devices is a ﬁeld that

remains unexplored. Therefore, our primary inter-

est is to propose a solution to avoid the deployment

of a high number of sensors in smart environments.

In fact, even if nowadays the ambient sensors have

a more and more affordable cost, the deployment of

numerous high-quality sensors in a smart context can

still be an expensive operation due to their installation

and maintenance costs notably. Thus, the deﬁnition of

intelligent systems to reduce the costs of both deploy-

ment and maintenance of sensors enables these tech-

nologies to be more attractive for smart cities initia-

tives. Furthermore, our goal is to design an open sys-

tem being capable of performing environmental esti-

mation by means of mobile and intermittent devices

(Figure 1). The openness is crucial in developing a

system that can be deployed at large scale, in the case

of a university campus as well as a smart city.

3 MULTI-AGENT SYSTEM

PROPOSITION

After a brief description of the neOCampus project

devoted to the construction of the campus of the fu-

ture, we present our multi-agent solution to estimate

missing environmental information in smart contexts

by limiting the number of sensors to deploy.

3.1 Context and Objectives

The neOCampus project, supported by the University

of Toulouse III - Paul Sabatier, plays a major role in

terms of technologies that could be employed in smart

cities by doing experiments in a university campus

context (Gleizes et al., 2018). Due to their size, to

the number of users and their mixed activities, univer-

sity campuses can be considered as districts or small

cities. As a matter of fact, more and more researchers

consider university campuses as great places to ex-

periment innovative services and techniques for smart

cities, building what is called a smart campus. With

an area of more than 264 hectares, the campus of the

University of Toulouse III - Paul Sabatier can be con-

sidered as a small city where several thousand data

streams come from heterogeneous sensors placed in-

side and outside the buildings (CO2, energy and ﬂuid

consumption, humidity, luminosity, ...). In such a

context, it is important to collect and integrate in-

formation that come from a large number of ad-hoc

sensors. Moreover, installing and maintaining a large

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

216

Figure 1: We aims at performing environmental estimation by means of mobile and intermittent devices in smart cities

(obtained through OpenStreetMap (OpenStreetMap contributors, 2017)).

number of sensors to monitor environmental param-

eters of the campus can be expensive. The costs are

mainly related to the installation, the maintenance and

the infrastructures of sensors in existing buildings.

For this reason, we propose a system able to pro-

vide environmental information everywhere and ev-

ery time without having to install numerous sensors.

In our work, we consider that in the University of

Toulouse III - Paul Sabatier the number of people per

day on the campus is around 36 000 among which

80% own a smartphone, but only 30% are volunteer

to share information from their smartphone. We can

consider that these people stay on the campus on aver-

age 6 hours a day, their smartphones provide correct

data about 10% of the time from 5 onboard sensors,

about 40 times each hour. By multiplying these pa-

rameters, we acquire more than 500 000 data every

day only using people’ smartphones. Thus, the use

of such a device represents a good solution for avoid-

ing the installation of ad-hoc sensors, but this solution

opens new challenges:

• Intermittent information: smartphones move

along the campus together with their owners, so

the required environmental information are not al-

ways available. The system has to take into ac-

count the unavailability of smartphones as well as

their displacement within the campus. This en-

ables the system to deal with both openness and

unpredictability;

• Data privacy: personal users’ information have to

be secured in order to prevent malicious activity

or users tracking;

• Data heterogeneity: smartphones are often

equipped with different types of sensors. Because

the aim of our proposition is to design a system

able to provide the maximum number of environ-

mental information, we must deﬁne an efﬁcient

solution to correlate and estimate missing infor-

mation;

• Large scale: the system has to be able to col-

lect and process data through an effective and

distributed architecture in order to ensure a high

quality service to users.

These requirements enable to address the chal-

lenges discussed in section 2.2. In neOCampus, we

assume that users agree on installing third-party soft-

ware to support our project: it is easier to ﬁnd stu-

dents, teachers and researchers motivated to experi-

ment innovative services and techniques.

3.2 Multi-agent System Proposal

Our goal is to propose a multi-agent system to es-

timate missing information in smart environment by

using a limited number of ad hoc devices. To do this

we propose a system based on two different types of

agents: (i) a Real Sensor Agent (RSA) which is any

physical instrumentation that can provide accurate en-

vironmental information value (such as a tempera-

ture), and a (ii) Virtual Sensor Agent (VSA) which is

responsible for the estimation, at a given point of the

environment, an information that a real sensor would

perceive if it was situated at the VSA’s location. The

goal of a VSA is to provide an accurate estimation

and an approximated conﬁdence zone, that represents

an area of the physical environment where mobile de-

vices provide, through their sensors, reliable values

to be used by the VSA in order to estimate the en-

vironment state. The conﬁdence zone is deﬁned as

a polygon centered in the position where the corre-

sponding VSA is situated. The RSAs are autonomous

and aware of the state of their local physical environ-

ment; they send their perceptions to VSAs. In this

Multi-agent Systems for Estimating Missing Information in Smart Cities

217

way, RSAs support VSAs in pursuing their goal. This

is the basis of the cooperative process; it consists in

the exchange of information perceived by RSAs with

VSAs in order to allow the latter to estimate environ-

mental information whereas real sensors are actually

missing at the VSA location.

Algorithm 1 describes the behavior of a RSA. At

line 2, the RSA perceives the physical environment

through an ad hoc sensor. We assume that a VSA

is associated to a RSA and a RSA are associated if

the RSA is situated within the conﬁdence zone of the

VSA. At line 4 the RSA checks if a VSA sent an as-

sociation request. At line 5 the RSA gets the list V

of associated VSAs. At line 6 the RSA sends its last

perception to the VSAs in V .

Algorithm 1: RealSensorAgent.

1: {—perceive—}

2: p ← perceiveFromSensor()

3: {—decide and act—}

4: checkAssociations()

5: V ← getAssociatedVSAs()

6: send(p,V)

Algorithm 2 describes the behavior of a VSA. Ini-

tially a VSA is associated to a predeﬁned conﬁdence

zone that has an octagonal shape that further evolves.

The algorithm starts by getting the list R of associated

RSAs within the conﬁdence zone (line 2). At line 3

the VSA sends an association request to the RSAs

inside its conﬁdence zone. Once associated, a RSA

sends regularly its perception to a VSA. At line 4 the

VSA receives the perceptions of the RSAs in its con-

ﬁdence zone. At line 6 the VSA evaluates the pairs of

RSAs and returns the set D of data ﬁelds. The pairs

of RSAs are chosen to be the most aligned with re-

spect to the VSA. A data ﬁeld between two sensors

is a vector ﬁeld in the Euclidean space. Each point is

associated to a vector which is oriented towards the

sensor which provides a higher data value; the mag-

nitude is the value of the gradient between the data

perceived by the sensors. Figure 2 shows an example

of gradient between two sensors T

and T

Algorithm 2: VirtualSensorAgent.

1: {—perceive—}

2: R ← getRSAsInConﬁdenceZone()

3: associateTo(R)

4: P ← receivePerceptionsOfRSAs()

5: {—decide and act—}

6: D ← evaluateRSAsPairs(R, O)

7: e ← evaluateEstimation(D)

8: updateconﬁdenceZone(D)

=19°

=21°

Figure 2: Example of a data ﬁeld between T

and T

. The

data ﬁeld is oriented towards T

since t

> t

At line 7 the environment state is estimated using

the data ﬁelds in D. Finally, the conﬁdence zone is

enlarged or reduced according to the data ﬁelds in D

(line 8). Standard techniques are used to determine

the pairs of RSAs which provide data ﬁelds that are

considered as outliers, so to be excluded from the con-

ﬁdence zone. Then the VSA cooperates with RSAs

and modiﬁes its conﬁdence zone. However, the VSA

does not have any knowledge about how to shape the

conﬁdence zone in an optimal way in order to keep

only the best pairs of RSAs. Thus, the VSA reasons

on the perceptions and the location of the RSAs in or-

der to determine if they have to be excluded or not

from the conﬁdence zone of a VSA.

3.3 Case Study

In the scenario illustrated in Figure 3 there are two

rooms including four temperature sensors whereas the

corridor does not contain any real sensor available.

In our proposed multi-agent system solution, RSAs

correspond to ad hoc devices they are associated with

, T

), whereas T

, T

are VSAs.

Figure 3: Case study illustration. The dashed zones repre-

sent the conﬁdence zones of VSAs T

, T

and T

The solving process for estimating the tempera-

ture in T

, T

and T

is roughly the following:

1. Each RSA perceives the physical environment

through its sensor and spreads it into the virtual

environment.

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

218

2. The RSAs situated inside a VSA’s conﬁdence

zone (represented by a black dotted line in Fig-

ure 3) are associated to the corresponding VSA

and the perceptions of the RSAs are sent regularly

to this VSA. Thus T

, T

send their percep-

tions to T

, while T

, T

send their perceptions to

3. The VSA determines pairs of RSAs within its

conﬁdence zone. It sorts the RSAs in its conﬁ-

dence zone according to their distance. Each RSA

is coupled with the most aligned RSA with respect

to the VSA. The pairs (T

, T

) and (T

, T

) are

formed by T

, the pair (T

, T

) by T

4. Each VSA evaluates the data ﬁeld provided by the

pairs of RSAs.

5. The data ﬁelds are used to estimate the environ-

mental state.

6. The VSA T

cannot estimate its current value

because there is no real sensor in its conﬁdence

zone. Its conﬁdence zone cannot be updated.

This is only a didactic example with homogeneous

sensors but in a real world scenario, different situa-

tions have to be considered: (i) an information has

to be estimated in an environment where there is no

sensor capable of providing the same type of infor-

mation; (ii) there is no sensor available in a certain

room; (iii) sensors can be intermittent and imprecise:

there is no certainty of obtaining precise and reliable

informations at a given instant from sensors.

4 EXPERIMENTAL RESULTS

As the campus of the University of Toulouse III - Paul

Sabatier is currently being instrumented and not oper-

ational, we evaluated our solution on a freely weather

dataset using available temperatures in degree Celsius

recorded by Arpae-SIMC, a weather service of the

Emilia-Romagna region in Italy. This service pro-

vides weather warnings to the Italian Civil Protec-

tion Department (Bressan et al., 2017). We consider

the average daily air temperatures at 2 meters of al-

titude collected over a period of time that goes from

September 8 2017 to April 25 2018 (196 days) from

80 weather stations (Figure 4). We do not take into

account the days where stations were not operational.

The dataset consists in an array of 196 × 80 numer-

ical values. For each station the dataset is provided

with geographic coordinate (longitude and latitude)

that were projected to the Cartesian plane through the

Mercator projection (Monmonier, 2010).

4.1 Evaluation of the Multi-agent

System Proposal

To evaluate our proposition we applied a leave-one-

out cross validation: for each experiment a precise

station has been replaced by a VSA in order to eval-

uate the estimation from the remaining stations. This

evaluation has been done using the optimal conﬁ-

dence zone for each sensor. That is, for each sen-

sor the method has been executed once for evaluat-

ing the best conﬁdence zones and then for estimating

the temperature using the conﬁdence zone previously

found. Each VSA contains initially all the weather

stations within its conﬁdence zone. Moreover, the

formed pairs of RSAs have different weights accord-

ing to the position between the VSA and the RSAs:

this inﬂuences the estimation of the VSA. For this

reason, the initial error of each station varies. Each

station, replaced by a VSA to do the leave-one-out

validation, tries to reduce its conﬁdence zone to keep

only the stations that are near the VSA and situated in

a homogeneous environment. Figure 5 shows the av-

erage error and the standard deviation of each station

during the considered 196 days together with a line

plot that marks the error produced by the technique

based on cluster analysis and normalized convolution

presented in section 4.2. The average overall error

is just 0.036 degree. Figure 6 shows the cumulative

error: each bar indicates the amount of absolute av-

erage error for a certain percentage of samples. The

difference between the absolute error of the proposed

technique is comparable to a pipeline of state of art

methods we present in section 4.2. Moreover, there is

a small number of samples (10%) for which the pro-

posed solution behaves better by evaluating a more

reliable conﬁdence zone and thus obtaining a more

precise estimation.

Figures 7 shows the conﬁdence zones of weather

stations 3, 10, 80 and 29 respectively. Figures 7a–

7c show conﬁdence zones of VSAs that take into ac-

count weather stations placed in different, far envi-

ronments because their data ﬁelds are not identiﬁed

as outliers. As shown in the error bar plot in Figure 5

these stations have a relevant average error. Figure 7d

shows an example of conﬁdence zone where the in-

volved weather stations are in a similar environment.

The conﬁdence zones in Figure 7a, 7b and 7c assume

a star shape because in the corresponding directions

there is not enough information to decide if they have

to be modiﬁed. This is due to the fact that the sen-

sors used in the experimentation are ﬁxed. By using

mobile devices as sensors, the VSAs will be able to

estimate the conﬁdence zone in a more precise man-

ner.

Multi-agent Systems for Estimating Missing Information in Smart Cities

219

Figure 4: The map of Emilia-Romagna (obtained through OpenStreetMap (OpenStreetMap contributors, 2017)). The stations

are indicated by the red dots.

Figure 5: Error bar of estimated temperatures (degree Celsius) for each station.

Figure 8 shows the error plot for the weather sta-

tions 3, 10, 80 and 29. The error for station 3 in-

creases while reducing its conﬁdence zone: as shown

in Figure 7a the conﬁdence zone takes into account

weather stations that are not situated in a homoge-

neous environment. The stations 10 and 29 maintain a

low error rate. The error for station 80 is also reduced

as the conﬁdence zone of the weather station reduces.

The proposed solution was coded in Java lan-

guage. For each station the algorithm takes about a

second to execute the entire leave-one-out validation

process for all the samples. The experiments were

carried out on an entry-level machine equipped with

i7-7820HQ, 32GB RAM and Windows 10.

4.2 Comparison to Standard

Techniques

We compared the results of the proposed solution

to a pipeline of standard techniques using the same

dataset and leave-one-out validation (Guastella and

Valenti, 2018). This pipeline, implemented as a pre-

liminary study for the problem of estimating missing

environmental information, includes Voronoi tessel-

lation to determine the conﬁdence areas of the sta-

tions (Okabe et al., 2000), hierarchical clustering to

group together the stations that behave in a similar

manner (Rafsanjani et al., 2012) and normalized con-

volution to estimate missing information (Knutsson

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

220

Figure 6: Cumulative error (degree Celsius).

and Westin, 1993). The areas of two or more stations

are merged if their Voronoi areas are adjacent and

they are grouped together by the clustering process.

Finally, for a given point, a normalized convolution

is used to estimate a missing environmental informa-

tion using the stations in the corresponding relevance

area. Normalized convolution is a standard method

used to reconstruct incomplete or uncertain data from

a spatio-temporal signal (Pham et al., 2006), widely

used in geo-statistical applications (Higdon, 1998;

Lemos and Sans

o, 2006).

The pipeline based on Voronoi, clustering and nor-

malized convolution shows better results with respect

to the multi-agent proposal because the optimal sub-

set of sensors is used to estimate missing informa-

tion. This is due to the fact that the clustering and the

Voronoi tessellation establish a priori the best groups

of sensors to be used for the estimation by using all

the available data. In this case the conﬁdence zones

are evaluated after the clustering process; thus any

modiﬁcation at a single real sensor implies a recon-

ﬁguration of the entire system. So this pipeline can

be used in the case of weather stations, but cannot

be used when considering mobile and intermittent de-

vices.

With respect to the pipeline, the multi-agent pro-

posal has the advantage of being able to enlarge or

reduce the conﬁdence zone of a VSA in a dynamic

way according to the knowledge of the RSAs that are

within the conﬁdence zone at a given time. More-

over, we assume that a RSA is not ﬁxed because it is

related to a smartphone device: if one or more RSAs

are being moved, classical approaches such as cluster-

ing need an entire reconﬁguration of the system that

could require a signiﬁcant amount of time. Also, the

multi-agent system has the advantage of being able to

continuously learn using intermittent devices without

reconﬁguring the entire system at runtime.

5 CONCLUSION AND

PERSPECTIVES

After a discussion of the state-of-art contributions that

exploit multi-agent systems in the context of smart

city, we presented the neOCampus project devoted

to our work. Then, we introduced a multi-agent sys-

tem to estimate missing information in smart environ-

ments. The goal of the proposed solution is to provide

anytime and everywhere accurate information where

ad hoc sensors are missing. The solution does not

require any parameter and is able to provide an esti-

mation of the environmental state at runtime.

Our work is a preliminary study to address the

challenges discussed in section 2.2. The autonomous,

adaptable and pro-acting behavior of agents allows

the system to work with intermittent devices, thus ad-

dressing the openness problem. Moreover, distribut-

ing the intelligence among all the agents allows a de-

ployment of the system at large scale. Unpredictabil-

ity is addressed through the dynamics of the agents.

In fact, a VSA is currently able to provide an esti-

mation through the RSA that are situated within its

conﬁdence zone. Moreover, a VSA will be capable of

providing an estimation even if different RSAs are ex-

cluded from the conﬁdence zone or they move within

it. This avoid the system to do an entire reconﬁgu-

ration. Even if it has not been considered, privacy

can beneﬁt from multi-agent systems: agents have

a local behavior, so a security problem involving an

agent would be locally limited, so that the system is

being able to easily identify and restore the entity con-

cerned.

As soon as the campus of University of Toulouse

III - Paul Sabatier is sufﬁciently instrumented we aim

to conduct an in vivo experimentation using data com-

ing from sensors within the campus and to compare

the obtained results to state-of-art techniques to pro-

vide an innovative infrastructure for the smart cam-

pus. The proposed solution currently focuses on ho-

mogeneous information and provides promising re-

sults. Our future work will focus on processing het-

erogeneous and intermittent information.

ACKNOWLEDGEMENT

This work is supported by the neOCampus initiative

(neocampus.org) from the University of Toulouse III

- Paul Sabatier.

Multi-agent Systems for Estimating Missing Information in Smart Cities

221

(a) (b)

Figure 7: Resulting conﬁdence zones of sensors 3 (a), 10 (b), 80 (c), 29 (d).

Figure 8: Error plot for stations 3, 10, 80 and 29.

REFERENCES

Bressan, L., Valentini, A., Paccagnella, T., Montani, A.,

Marsigli, C., and Tesini, M. (2017). Sensitivity of

sea-level forecasting to the horizontal resolution and

sea surface forcing for different conﬁgurations of an

oceanographic model of the Adriatic Sea. Advances in

Science and Research (European Meteorological So-

ciety), 14:77–84.

Cook, D. J., Youngblood, M., and Das, S. K. (2006). A

Multi-agent Approach to Controlling a Smart Envi-

ronment. In Augusto, J. C. and Nugent, C. D., edi-

tors, Designing Smart Homes: The Role of Artiﬁcial

Intelligence, pages 165–182. Springer.

Dameri, R. P. (2013). Searching for smart city deﬁnition:

a comprehensive proposal. In International Journal

of Computers & Technology, volume 11, pages 2544–

2551. Council for Innovative Research.

Frontczak, M. and Wargocki, P. (2011). Literature survey

on how different factors inﬂuence human comfort in

indoor environments. Building and Environment (El-

sevier), 46(4):922 – 937.

Georg

e, J.-P., Gleizes, M.-P., and Camps, V. (2011). Co-

operation. In Di Marzo Serugendo, G., Gleizes, M.-

P., and Karageorgos, A., editors, Self-organising Soft-

ware: From Natural to Artiﬁcial Adaptation, pages

193–226. Springer.

Ghahramani, A., Castro, G., Karvigh, S. A., and Becerik-

Gerber, B. (2018). Towards unsupervised learning of

thermal comfort using infrared thermography. Applied

Energy (Elsevier), 211:41–49.

Gleizes, M.-P., Boes, J., Lartigue, B., and Thi

ebolt, F.

(2018). neocampus: A demonstrator of connected,

innovative, intelligent and sustainable campus. In

De Pietro, G., Gallo, L., Howlett, R. J., and Jain,

L. C., editors, Intelligent Interactive Multimedia Sys-

tems and Services, pages 482–491. Springer.

Guastella, D. A. and Valenti, C. (2018). Estimating miss-

ing information by cluster analysis and normalized

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

222

convolution. In 2018 IEEE 4th International Forum

on Research and Technology for Society and Industry

(RTSI), Palermo, Italy.

Herkel, S., Knapp, U., and Pfafferott, J. (2008). Towards a

model of user behaviour regarding the manual control

of windows in ofﬁce buildings. Building and Environ-

ment (Elsevier), 43(4):588–600.

Higdon, D. (1998). A process-convolution approach to

modelling temperatures in the North Atlantic Ocean.

Environmental and Ecological Statistics (Springer),

5(2):173–190.

Karnouskos, S. and Nass de Holanda, T. (2009). Simula-

tion of a smart grid city with software agents. In 2009

Third UKSim European Symposium on Computer

Modeling and Simulation, pages 424–429, Athens,

Greece.

Knutsson, H. and Westin, C.-F. (1993). Normalized and

differential convolution. In Proceedings of IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 515–523. IEEE.

Lemos, R. T. and Sans

o, B. (2006). Spatio-temporal vari-

ability of ocean temperature in the portugal current

system. J. Geophys. Res. (AGU), 111(C4):C04010.

Manville, C., Cochrane, G., Cave, J., Millard, J., Peder-

son, J. K., Thaarup, R. K., Liebe, A., Wissner, M.,

Massink, R., and Kotterink, B. (2014). Mapping

Smart Cities in the EU. Technical report.

Monmonier, M. (2010). Rhumb lines and map wars: a so-

cial history of the Mercator projection. University of

Chicago Press.

Okabe, A., Boots, B., Sugihara, K., and Chiu, S. N. (2000).

Spatial Tessellations: Concepts and Applications of

Voronoi Diagrams. Probability and Statistics. Wiley.

Olaru, A., Florea, A. M., and El Fallah Seghrouchni, A.

(2013). A context-aware multi-agent system as a mid-

dleware for ambient intelligence. Mobile Networks

and Applications (Springer), 18(3):429–443.

OpenStreetMap contributors (2017). Planet

dump retrieved from https://planet.osm.org.

https://www.openstreetmap.org.

Perles, A., Camilleri, G., and Gleizes, M.-P. (2017). Self-

adaptive distribution system state estimation. In Cri-

ado Pacheco, N., Carrascosa, C., Osman, N., and

Juli

an Inglada, V., editors, Multi-Agent Systems and

Agreement Technologies, pages 202–216. Springer.

Pham, T. Q., van Vliet, L. J., and Schutte, K. (2006). Robust

fusion of irregularly sampled data using adaptive nor-

malized convolution. EURASIP Journal on Advances

in Signal Processing (Springer), 2006(1):083268.

Piette, F., Caval, C., Dinont, C., Seghrouchni, A. E. F.,

and Tailliert, P. (2016). A multi-agent solution for

the deployment of distributed applications in ambi-

ent systems. In Baldoni, M., M

uller, J. P., Nunes, I.,

and Zalila-Wenkstern, R., editors, Engineering Multi-

Agent Systems, pages 156–175. Springer, Singapore,

Singapore.

Pirttikangas, S., Tobe, Y., and Thepvilojanapong, N. (2010).

Smart environments for occupancy sensing and ser-

vices. In Nakashima, H., Aghajan, H., and Augusto,

J. C., editors, Handbook of Ambient Intelligence and

Smart Environments, pages 825–849. Springer.

Rafsanjani, M. K., Varzaneh, Z. A., and Chukanlo, N. E.

(2012). A survey of hierarchical clustering algo-

rithms. Journal of Mathematics and Computer Sci-

ence (ISRP), 5(3):229–240.

Roscia, M., Longo, M., and Lazaroiu, G. C. (2013). Smart

city by multi-agent systems. In 2013 International

Conference on Renewable Energy Research and Ap-

plications (ICRERA), pages 371–376, Madrid, Spain.

IEEE.

Verstaevel, N., Georg

e, J.-P., Bernon, C., and Gleizes, M.-P.

(2018). A self-organized learning model for anoma-

lies detection: Application to elderly people. In IEEE

International Conference on Self-Adaptive and Self-

Organizing Systems (SASO), pages 70–79, Trento,

Italy. IEEE.

Wong, L. T., Mui, K. W., and Cheung, C. T. (2014).

Bayesian thermal comfort model. Building and En-

vironment (Elsevier), 82:171–179.

Multi-agent Systems for Estimating Missing Information in Smart Cities

223