SPECIAL SECTION ON HETEROGENEOUS CROWDSOURCED DATA ANALYTICS

Received November 16, 2016, accepted January 6, 2017, date of publication July 11, 2017, date of current version July 31, 2017.
Digital Object Identifier 10.1109/ACCESS.2017.2725984

An Agent Based Traffic Regulation System for

the Roadside Air Quality Control
ABDELAZIZ EL FAZZIKI1 , DJAMAL BENSLIMANE2 , ABDERRAHMANE SADIQ1 ,
JAMAL OUARZAZI3 , AND MOHAMED SADGAL1
1 Computer Systems Engineering Laboratory, Cadi Ayyad University, 40000, Marrakesh, Morocco
2 LIRIS Laboratory, Université Claude Bernard Lyon 1, 69622, Villeurbanne, France
3 Laboratoire Physico-Chimie des Matériaux et Environnement (URAC 20), Cadi Ayyad University, 40000, Marrakesh, Morocco
Corresponding author: Djamal Benslimane ([email protected])

ABSTRACT This paper describes an on-road air quality monitoring and control approach by proposing
an agent-based system for modeling the urban road network infrastructure, establishing the real-time and
predicted air pollution indexes in different road segments and generating recommendations and regulation
proposals for road users. This can help by reducing vehicle emissions in the most polluted road sections,
optimizing the pollution levels while maximizing the vehicle flow. For this, we use data sets gathered from a
set of air quality monitoring stations, embedded low-cost e-participatory pollution sensors, contextual data,
and the road network available data. These data are used in the air quality indexes calculation and then
the generation of a dynamic traffic network. This network is represented by a weighted graph in which the
edges weights evolve according to the pollution indexes. In this paper, we propose to combine the benefits of
agent technology with both machine learning and big data tools. An artificial neural networks model and the
Dijkstra algorithm are used for air quality prediction and the least polluted path finding in the road network.
All data processing tasks are performed over a Hadoop-based framework: HBase and MapReduce.

INDEX TERMS Air quality management, crowd-sourcing, Dijkstra algorithm, mobile sensors, pollution
prediction, traffic regulation.

I. INTRODUCTION the spatial and temporal resolution of pollution sampling is

Several studies have shown that the risks associated with car- very low. For example, there exist about 3 active monitoring
diovascular and respiratory morbidity increase with chronic stations in Marrakech, separated from each other by tens of
exposure to air pollution. Also, acute inhalation of pollutants kilometers. This requires the use of mathematical models to
even for a short period of time can lead to a dysfunction in calculate and estimate pollutant’s levels in a large areas of
the cardiovascular system and lung function [1], [2]. As in the city, which can be both complex (requiring parameters
most cities, the basic pollution problem arises because the such as the topography data, weather variables and chemical
residential areas are far from workplaces, resulting in daily compositions) and inaccurate and therefore, can lead to incor-
large population movements [3]. Thus, transport, especially rect conclusions [2], [4]. The estimation of personal exposure
road traffic, is a major source of air pollution in most of is essential for individuals to manage the risks associated
the cases. It contributes to 15% of the CO2 emissions in with this problem, both through a retrospective understanding
Europe . In Moroccan urban environment, there is a range of the levels of pollutants that affect their health and the
of pollutants in the atmosphere with the capacity to cause prospectively decision making to reduce the risks of being
harm to both humans and the ecosystem, including Carbon exposed to the pollution pikes [5].
monoxide, Nitrogen oxides, Sulphur dioxide, Particular mat- Local authorities face a difficult situation with a constant
ter, Volatile organic compounds, Ozone, and Hydrocarbons; increase in road traffic, leading to congestion and travel times
thus, our particular interest in addressing this issue. that are constantly growing. This leads to an increase in fuel
Therefore, many government agencies are increasingly consumption and pollutant emissions, especially during peak
interested in more monitoring and publication of data on hours. Emissions in cities often exceed the tolerated level;
air quality. At this level, two fundamental limitations on the studies have shown that the effects of air pollution reduce life
approaches to the monitoring of air quality are imposed: First, expectancy and induce enormous cost to society. However,

2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
13192 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 5, 2017
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

the existing techniques for air quality and traffic congestion 5) The use of Dijkstra algorithm in the shortest path (least
management have not been very effective because of the polluted path) finding [9], [10]. This algorithm has
lack of autonomy, adaptation, and collaboration between the been widely used in route management systems, with
relevant entities. This issue needs the use of an integrated variations to improve its performance [9].
system for traffic management and air quality control, which 6) We tested and deployed the system by simulating the
aims to jointly optimize traffic and air quality. Therefore, data collecting by a small number of users to validate
an effective intelligent transport system taking into account our proposal.
air quality is of great importance for society. The rest of this paper is illustrated over a few sections
In this work, we are particularly interested in traffic regu- starting with a brief literature review followed by an overview
lation and generating recommendations to take into account of the proposed system and the traffic regulation process
in order to improve air quality. The proposed system must be details in section 3. Section 4 is devoted to the use of ANN
able to establish air quality indices in urban areas, to generate for air quality prediction. The MapReduce based data analysis
relevant information to users and propose recommendations process is described in Section 5. In section 6 we present the
about the best paths to take. This is done based on pollutant used method for path finding in a road network. Section 7 is
data collected from users embedded devices that provide dedicated to the multi-agent system description. In section 8,
properties related to road infrastructure and air quality to the a case study and the experimental results are presented, fol-
system. It must take into account the constraints of geograph- lowed by a conclusion and perspectives in section 9.
ical distribution and dynamics of the problem seen the fact
that air quality levels change over time and the exact number II. RELATED WORKS
of users which demand recommendations and then the num- In the past few years, many researches were centered around
ber of vehicles to redirect to alternative paths is unknown. the use of static air quality monitoring stations along with
Therefore, the system must be adapted to user needs, taking crowdsourcing and participatory sensing. Such solutions lend
into account all these factors while ensuring a great per- themselves not only to monitoring the state of the physical
formance in handling the large amounts of data (pollutants, world but can also help raising peoples awareness of issues
weather, etc.) gathered from different data sources (users related to air quality and pollution [2]. One of the main
devices, monitoring stations, weather data records, etc.) that projects which proposed such a solution is the MESSAGE
will continually increase over time. For this, we propose (Mobile Environmental Sensing System Across Grid Envi-
the use of an approach based on agent technology [6], [7], ronments) project [11]. It aims to develop fixed and portable
crowd-sourcing, and Big Data analysis tools in order to mea- devices for high-density measurement of concentrations of
sure, monitor and control the quality of air with a higher carbon monoxide and nitrogen oxides in urban areas. They
spatiotemporal resolution and to involve users in track- have very recently reported their development and deploy-
ing their exposure to pollution through custom embedded ment experience [12] in the Cambridge area, and demon-
tools. strated that the use of low-cost fixed and portable devices
Moreover, due to large amounts of data, the processing deployed in high densities can give a much more accurate
task has become a great challenge. To address this problem picture of the spatial and temporal structure of air quality
we suggest the use of Hadoop framework to ensure a great in the urban environment. In [13] authors introduced a low-
flexibility and speed in performing the needed prediction and cost mobile sensors based participatory air quality monitoring
analysis algorithms applicable to large scale data. Our main system. this work presents the design and implementation of
contributions in this work are: GasMobile a small and portable measurement tool suited to
1) Offer a low-cost component used by users who want to be used by a large number of people.
contribute to the monitoring of air quality by partici- In recent years, many works have been interested in using
pating in the data collection task. This presupposes the artificial intelligence and machine learning methodologies
development of a monitoring sensor unit to use onboard to address problems related to quantitative and qualitative
vehicles or by users, an application for data loading and measures more effectively. They can allow solving complex
transfer, and a server for data hosting. issues related to air quality management, traffic regulation
2) A component for extracting and loading meteoro- and transportation systems by integrating intelligent compo-
logical and pollution data from fixed monitoring nents [14] Multi-agent based systems have been considered
stations. as an efficient tool for large-scale system such as intelligent
3) Predicting the level of pollution near roads using the traffic and air quality management [15]. They were known
available data and ANN that have a strong ability to by the ability to model complex systems where a set of
predict fuzzy data and efficiency in the modeling of autonomous entities interact to produce global solutions [16].
dynamic systems. Many works propose the use of the MAS technology in traffic
4) Calculate the real-time and predicted air quality control and management systems, such as [9] which proposes
indexes, based on Murena method [8], and generating an integrated approach for modeling transport infrastructure
a spatio-temporal map for the distribution of pollution and optimizing traffic in urban areas in order to reduce
level in the study area. carbon-dioxide emissions. Authors in [14] have proposed

VOLUME 5, 2017 13193

 A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

an adaptive multi-agent system based on the ant colony

behavior and the hierarchical fuzzy model. The solution is
based on an adaptive vehicle route guidance system which
allows adjusting efficiently the road traffic according to the
real time changes in road networks. In most works based
on MAS technology, such as [7], authors have integrated
advanced analysis components and tools. Thus, the software
agents use data mining methods for knowledge discovery,
FIGURE 1. Overview of the system structure.
which will be used as a foundation for recommendation
generation. Namoun et al. [9] have used an improved version
of the Dijkstra algorithm as a graph search algorithm. This
algorithm has a powerful potential to solve the problems of handful of fixed points across the study area (As illustrated
traffic congestion in a road network, which can be consid- in Figure 1).
ered as a collection of graphs. The suggested solution in [9] One novel aspect of the system is that it makes use of
combines the benefits of MAS and real time traffic data immediate contextual parameters which can be retrieved from
forecasting and analysis in the recommendation generation the road users’ devices, such as time of day, location, speed
process. Machine learning tools such as ANN has been fre- and direction. These services help estimating traffic flow and
quently used to predict air quality and pollution levels using congestion. As a result, users receive more relevant informa-
a set of inputs, like pollutant concentrations, meteorological tion and recommendations that are based on a combination of
data and available traffic information [17], [18]. Moreover, their historical preferences and contextual parameters.
many research works have been dedicated to implement- All these types of inputs feed the system, which uses an
ing such computationally expensive algorithms and analysis advanced computational module for analyzing and extracting
tasks on parallel or distributed computing systems such as knowledge based on required algorithms and OLAP anal-
Hadoop [19]. Thus, information system enhancement using ysis. This module uses the pollutants data, contextual data
Hadoop as a data hub to optimize the information systems and meteorological information for forecasting the pollutants
performance is considered as a new emerging strategy [20]. concentrations in a specific zone by using an ANN based
Many works such as [21] has proposed a study on big prediction model. The real-time and predicted results are
data integration with data warehouse built using relational then used in order to calculate an Atmospheric pollution
technology mainly for operational sources. Regarding the index (API) based on Murena method (REF). The API is
presented related works, the development of management used to determine the quality of air in a simple manner and
and control systems must be based on emerging technolo- provides information that can be understood by the general
gies such as intelligence systems, Big Data and advanced public. The API is based on predicted rates of each pollutant
analytics tools. It has been proved that MAS is a powerful comprising of sulfur dioxide (SO2), ozone (O3), nitrogen
computing paradigm to manage the complexity in dynamic dioxide (NO2), Atmospheric particulate matter (PM10) and
distributed environment and the combination of MAS and carbon monoxide (CO).
Big data analytics tools can considerably improve the system The computational module uses the resulting information
performance. in addition to the available road infrastructure data to build
a weighted road network in which the edges weights are
III. AIR QUALITY MANAGEMENT: TRAFFIC REGULATION calculated based on the obtained APIs and the road segment
The aim of this work is to propose a hybrid solution to lengths. The generated network is then used to find the most
the issue related to the on-road air quality optimizing. This environment-friendly routes, provide forecasts about on-road
solution relies on both the data gathered from a set of static pollution level and formulate recommendations to road users.
air quality monitoring stations and datasets gathered from The objective is to avoid the most polluted sections while
several low-cost mobile sensors units. These units are embed- averting the generation of traffic congestion. Figure 1 illus-
ded in users vehicles and connected to these users mobile trates an overview of the system structuring.
devices (i.e. mobile phones) in order to upload the obtained As result, the main features of the system are:
data along with other additional data which we can get from • An effective communication with the users. It provides
these devices such as the GPS (Global Positioning System) route solutions and recommendations that satisfy the
location and temperature. The use of mobile sensors has the preferences of users (e.g. Time of travel).
advantage of providing real time measurements from differ- • Integrate and manage a wide range of data sources.
ent locations (all the places visited by users) and also the con- • Use of real-time data: receive real-time information
textual information that we can collect from the used mobile about pollution levels, perform needed analysis and gen-
phones. Furthermore, the static air quality monitoring stations erate the air quality indexes for each urban path segment.
provide measurements datasets for various pollutants as well • Predicted air qualitys data: use historical data about pol-
as meteorological parameters (Temperature, wind speed and lutants and meteorological data to perform predictions
direction, precipitation, pressure, etc) very accurately at a and calculate the air quality indexes for path segments.

13194 VOLUME 5, 2017

A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

A. POLLUTION MEASUREMENT SENSORS C. ROAD NETWORK MODELING

One of the most critical and challenging component of the A road network of the studied area is used in order to repre-
architecture is the device that will be used to measure air sents the transport supply. It is described by the different road
pollution levels. There are several options and we considered segments and their intersections. In this work, we simulate
few criteria to choose the most suitable one: the road segments for which properties are gathered from
• Portability: In order to face the issue related to the spatial different data sources and provided to the system [22]. These
coverage, the device must be made portable enough for collected data are used in the weighted graph generation
a user to carry himself (e.g. on vehicles). (see Figure 4). A planar graph is used to formalize the
• Complexity: Having decided to make our devices suit-
able for vehicle-mounting, the next major decision we
were confronted to was regarding target cost and com-
plexity of the device.
• Sensor Type: depends on the pollutants to measure.
In this work we analyze the case where several users are
equipped with ozone sensors along with a GPS module.
These sensors are connected (via Bluetooth) with mobile
devices which are equipped with a 4G communications unit.
Mobile devices upload all gathered data to a central server
which performs all required analysis and generates the suit-
able recommendations. Figure 2, illustrates the proposed
process. FIGURE 4. Road segment details and an example of road paths with
different data (by Google Maps).

road network. In this graph,the roads are represented by a

set of arcs and the junction points by nodes. Arcs are char-
acterized by features such as APIs, path lengths and traffic
capacity. Nodes are associated with intersections character-
istics. The cartography is represented by a weighted graph
G = (V, E) where V is a set of vertices representing and
E = V ×V is a set of edges e = (vi, vj). Each segment joining
adjacent vertices is represented by either one or two directed
edges. The edge weight wij, vj between two vertices vi and vj
FIGURE 2. Data gathering and processing tasks. is a dynamic factor, which is used in order to reflect prop-
erties related to the edge (vi, vj) which represents. Figure 5
illustrates an example of a road graph.
B. THE TRAFFIC REGULATION PROCESS
The traffic regulation and recommendation generation pro- D. ANN FOR AIR QUALITY PREDICTION
cess is divided into the following steps (See Figure 3): The main stage of air quality control is the process of predict-
ing the on-road pollution level. The present study investigates
the advantage of using neural networks for forecasting the on
road air pollution. Artificial neural network (ANN) models
are mathematical models inspired by the functioning of ner-
vous systems, which are composed by a set of interconnected
artificial neurons. These neurons can be associated in many
different ways, depending on the characteristics of the issue to
FIGURE 3. The traffic regulation and recommendation generation address. To construct an ANN model the air pollution system
process.
is considered as a system that receives information from
distinct sets of inputs and produces a specific output [23].
• Pollution, meteorological and road network data The proposed network has been used for predicting pol-
gathering. lutants concentration using three years data which contains
• Users’ contextual information integration. hourly records of the different pollutants concentration and
• Predict the air quality indexes. meteorological records. This prediction process is based on
• Road network generation. three stages. The first one is the data extraction stage, in which
• Data processing and the road network weights the objective is to choose the more significant data for the
calculation. learning phase. The learning stage is the second phase. It aims
• Apply the shortest path algorithm. to find the optimal configuration of hidden layers, the transfer

VOLUME 5, 2017 13195

 A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

sub-indexes by applying the Murena [8] method. At the same

time we use meteorological, geographical and road data files
to generate the final API for each road segment [21], [24].
The intermediate results are stored into the output databases.
In the Second Map phase (Figure 7) the API values are
combined with users data and road network data in order
to generate the weighted network. The intermediate results
use the pair of segment identifier, and timestamp as the
key map and the API as well as edges length as the value.
During this Map stage the Dijkstra algorithm is applied in
order to generate three shortest paths in response to each user
query. In the Reduce phase, the paths with the same key are
cumulated together.

V. PATH FINDING IN A ROAD NETWORK

A. WEIGHTED NETWORK GENERATION
The weights of the network segments are calculated by com-
bining the predicted API records and the segment lengths.
Murena method [8] was adopted to calculate the API. The
evaluation of the API at stations for a pollutant p (PIs, p)
FIGURE 5. Example of a two directed edge based road graph. is carried out by a linear interpolation between the reference
scale values reported in Table 1 and is given by (Equation(1)):
PIhi − PIlo
function, and the performance index. The objective during PIs,p = [ (Cp − BPlo ) + PIlo ]s,p (1)
BPhi − BPlo
this stage is to minimize the prediction error. The third stage
is the prediction part in which we predict the pollutants Where:
concentration for a given time and location from a previously • PIs,p :The value of the pollution index for a pollutant p
calibrated neural network during the learning phase. at site s.
In this work, we used static monitoring stations data which • BPhi : The lowest break-point of a pollutant p that is
have hourly frequency in the prediction model. The avail- greater than or equal to Cp
able data records for each road segment is variable (from • BPlo : The highest break-point of a pollutant
4625 to 10124 records) due to the sensors deficient time. Each • p that is lower than or equal to Cp
record contains 13 attribute: The five pollutants concentration • PIhi : The PI value corresponding to BPhi
(SO2, O3, NO2, PM10 and CO), solar radiation, Wind speed, • PIlo : The PI value corresponding to BPlo
Temperature (Celsius), Humidity ratio, Hour, Day, Month • Cp : The pollutant p daily concentration
and year. The attributes and the hidden layers to use in the An individual score is assigned to the level of each pollu-
ANN model depend on the pollutant to predict. The network tant and the final API index is equal to the highest sub-index
requires a data set to be calibrated. This set is called training determined for each of the considered pollutant. The other
set. The evaluation of the network is carried out using dif- required road segments properties are prepared and stored in
ferent data from those used during training, which is called internal databases. The weights of the network are updated
the test set. The training set corresponds to the first 80% of when new forecasts are available. In this case, the system
records, and the tests set corresponds to the last 20%. We recalculates the API and dynamically updates the network
choose backpropagation as a learning rule since it is most with the new values. Otherwise the current weights are used
suited and generally used technique for prediction problems. to generate paths and recommendations.
Along with, we choose to take advantage of the momentum
to improve the efficiency of the algorithm. The evaluation of B. THE LEAST POLLUTED PATH SEARCHING:
the neural network performance is carried through the Mean Dijkstra ALGORITHM
Square Error (MSE). Dijkstra’s algorithm is a graph search algorithm that solves
the single source shortest path problem for a graph with
IV. MAPREDUCE BASED DATA ANALYSIS PROCESS nonnegative edge path costs, producing a shortest path tree.
The weighted graph generation task uses a multi-phase This algorithm is often used in routing and as a subroutine
MapReduce process to get pollution level on different road in other graph algorithms [25] In this work, we apply the
segments [21]. The predicted and monitoring air pollution Dijkstra shortest path finding algorithm over the generated
data (pollutants concentration) are loaded from the HBase in traffic weighted networks in order to find the least polluted
which they are stored. In the First Map stage (see Figure 6), possible paths in the network. The weight of each segment
we use the pollutants level datasets to calculate the pollution is a combination of the different cost nature of using this

13196 VOLUME 5, 2017

A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

FIGURE 6. The first stage of the MapReduce process.

FIGURE 7. The second stage of the Map and Reduce process.

TABLE 1. Breakpoints for the proposed API (µg/m3 for all pollutants and mg/m3 for CO)(Murena 2004 [8]).

particular segment. As we are interested in the air quality opti- order to optimize the weight calculation function. In this work
mizing by traffic regulation, we use the on-road calculated we modify the Dijkstra’s shortest path algorithm to obtain the
API values and the segment lengths [9]. three selected shortest paths in order to provide the decision
By using this weighted network the system can pro- making and risk evaluation to the users.
cess user queries, generates recommendations and calcu-
VI. THE MULTI-AGENT SYSTEM ARCHITECTURES
late potential roads according to the Dijkstra shortest path In order to provide an adequate solution in terms of robustness
algorithm. For a particular source node, the algorithm finds and agility, we use a multi-agent framework to represent the
the path with the lowest cost between that node and any on-road air quality management system. The objective is to
other node. The used function for calculating the shortest path propose an architecture that consists of a set of autonomous
will depend on the road segment length and the calculated agents able to set their own goals and actions and interact and
API for each path segment. Thus, the value of the path collaborate with each other through a communication pro-
can be provided by the following formula (Equation(2) and tocol [26]. The agents can be categorized into three groups:
Equation(3)): 1). A data integration agent group which continuously mon-
n
X itorsing and collectsing air pollution data, meteorological
Pi = Xi × Wi (2) data, etc, 2). The agents in the data processing group analyze
i=0 the pollution data provided to generate recommendations for
Wi = (C1 × Ai + C2 × Li )/(C1 + C2 ) (3) road users, and 3). While the interface group agents which are
responsible for contextual data gathering.
Where, Xi is a configurable weight and Wi is the weight of A two-level hierarchical MAS can also be adopted. The
the edge number i and n is the potential number of edges to data processing group is cognitive agents that give instruc-
browse. Ai and Li are successively the current API and length tions while the worker agents in UI and data integration
of the edge i. C1 and C2 are two coefficients related to each groups are reactive agents that carry out the tasks assigned.
cost parameter. They are controlled by decision makers in Figure 8 illustrates the multi-agent system structure. In the

VOLUME 5, 2017 13197

 A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

FIGURE 8. The Multi-agent air quality system for traffic regulation.

development process we chose to use the Prometheus • Network agent: This agent is responsible of the road
methodology to design and implement the multi-agent sys- paths network generation and path finding in this
tem. Our choice is motivated by the efficiency of message weighted network by searching for the shortest possible
communication and lightweight nature of the framework. route between two road nodes. The path is represented
Prometheus methodology [27] has been developed to support by a set of segments to browse from a source node to the
the complete software development life-cycle from problem destination node.
description to implementation. It offers an environment for • Managing agent: Responsible for the reliability of the
analyzing, designing, and developing heterogeneous multi- whole system, and manages the operation of the indi-
agent systems. This methodology consists of three phases: vidual agents, especially the station agents, model, data
System Specification, architectural design and the detailed provider and network generation.
design. The Prometheus Development tool is extended with • User’s Data agent: This agent uses the mobile app to
the ability to generate skeleton code in the JACK agent- gather data concerning traffic, meteorological and
oriented programming language [28] using a PDT code road conditions using users device sensors readings
generator extension which maintains also synchronization (e.g. Accelerometer, GPS).
between the generated code and the design when either of • Analytics agent: The purpose of the OLAP agent is to
them changes. convert the amount of monitoring data into valuable
The derived multi-agent system is composed of seven main information by applying quick and effective analysis and
agents: create recommendations for users using various views
• Data integration agents: These agents represent all mon- and representations of data. It provides all the basic
itoring stations and mobile sensors distributed in the functionality of an OLAP system and also the missing
study area and provide the required functionality during intelligence in traditional OLAP systems. The aim is
the data extraction, transformation and loading process. performing OLAP analysis on behalf of an agent or a
A data integration agent is also responsible for data user and reporting its result back to the requesting entity
validation, accuracy, the type conversion, etc. It prepares and all other entities that should be informed [29].
all input data required for the good functioning of the
model agent. It sets up a register of emissions for the VII. CASE STUDY
region in order to make regional modeling and prepare A. THE STUDY AREA
the needed domain data and meteorological parameters. The study area used for the testing scenario in this work
Collaboration between these agents will allow a better is located in Marrakech-City. This study area suffers from
understanding of the spatiotemporal evolution of surface the effects of pollutants produced by vehicle exhaust sys-
air quality. tems. This study is based on a set of sensors that provide
• Model agent: The Model agent performs the determinis- information and measures of the air pollutant concentration.
tic modeling. It brings together a set of equations repre- It focused on air quality indexes related to the following
senting the transport and chemistry of gaseous species, pollutants: Sulfur Dioxide (SO2), Nitrogen dioxides (NO2),
allowing the quantification of the evolution of a set of Carbon Monoxide (CO), Particulate Matter, and Ozone (O2).
pollutants, according to time on road segments, taking The map used for the testing scenario in this paper was cut in
into account all parameters. This agent is responsible order to reduce the simulation time. The area considered in
for the generation of the air qualitys prediction using the simulations is represented in Figure 9.
ANN, which is capable of modeling highly nonlinear
relationships while taking into account the data distribu- B. AIR QUALITY PREDICTION
tion factors. This agent uses the predicted and real time We use a three-layer perceptron ANN model and data con-
data in order to calculate API values which are then used cerning the study area described above to predict pollutants
to generate the weighted network. All resulting data are level. In this case study we generate prediction of the
stored into an HBase. Ozone concentration; A six neurons in the input layer are

13198 VOLUME 5, 2017

A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

optimal number of hidden layers and neurons, or the value of

momentum and the learning rate. We experimented multiple
configurations until convergence of the MSE, in a way to
find the optimal one. We concluded that more neurons do not
help getting lower mean square error. In addition, the momen-
tum and the learning rates must be low to get the best
results.
The following graph (Figure 11) shows the performance of
the network above. It represents a comparison of the observed
and the predicted Ozone concentrations evolution based on
the MSE according to time (during two days).

FIGURE 9. Inset map of an area of Marrakech used for simulations

(by Google Maps).
FIGURE 11. Comparison of one day observed and predicted Ozone
concentration values from the ANN for one road segment .

FIGURE 10. The three-layer ANN perception configuration.

used including temperature, solar radiation, the NO2 con-

centration, CO concentration and wind velocity. The number
of hidden layers and neurons in each hidden layer are the
FIGURE 12. A screenshot from the used mobile user interface .
parameters to choose in the model construction. In our case,
we tested different configuration and we found that a model
based on one hidden layer and four neurons gives the best
ANN performance. The last layer is the output, which con- C. API VALUES GENERATION
sists of the target of the prediction model. Here, O3 was used The system generates the analysis queries based on selected
as the output variable and a hyperbolic tangent sigmoid func- options from an easy to use user interface to get the result-
tion was used as the transfer function. A two year data set was ing API. Figure 12 presents a screenshot of the systems
divided into two parts: 80% used for training the networks and final user interface which allows users to select the desired
the remaining 20% employed in testing the networks. The options from a web form and Figure 13 is an example of
mean square error was chosen as the statistical criteria for a query result which allows selecting the API during one
measuring the network performance. Figure 10 illustrates this week (On May 2015) in a selected location from the study
ANN configuration. There is no exact procedure to meet the area.

VOLUME 5, 2017 13199

 A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

loading, fast query processing and efficient storage. The sys-

tem is currently being tested using real-time Marrakech city
data. The case study addresses the generation of pollutants
API in the road segments. The pollution level is then used in
order to calculate the best route a user can use.
Our experimental results show that the data processing
operations and the algorithms deployed in the large-scale data
processing system are feasible and efficient. The use of a
system based on Hadoop has improved the performance and
decreased significantly the processing time. The perspectives
FIGURE 13. Three segments pollution index in the simulation area during
one week . of this work are the integration of multi-criteria decision
support tools and the use of historical traffic data in the deci-
sion making process in order to generate more appropriate
D. THE LEAST POLLUTED PATH FINDING recommendations for final users.
In the implemented system prototype, the road graph is gen-
erated in a specific location in the study area according to ACKNOWLEDGMENTS
users’ needs. For each segment in this location the system The authors wish to acknowledge the contributions of other
calculates or updates the costs which are calculated based on members of computer systems engineering laboratory and
the available data and the Dijkstra algorithm is then applied. physical-chemistry of materials and environment laboratory
Figure 14 describes a road network example for a user who (URAC 20) for their helpful discussions and the availability
wants to navigate from the node N1 to N14 and the generated of all data and resources that have helped make this work in
shortest path. the best conditions.

REFERENCES
[1] N. L. Mills et al., ‘‘Adverse cardiovascular effects of air pollution,’’ Nat.
Clin. Pract. Cardiovascular Med., vol. 6, no. 1, pp. 36–44, 2009.
[2] V. Sivaraman, J. Carrapetta, K. Hu, and B. G. Luxan, ‘‘HazeWatch:
A participatory sensor system for monitoring air pollution in Syd-
ney,’’ in Proc. 38th Annu. IEEE Conf. Local Comput. Netw. Work-
shops, Oct. 2013, pp. 56–64. [Online]. Available: http://ieeexplore.ieee.
org/lpdocs/epic03/wrapper.htm?arnumber=6758498
[3] EUROPEAN COMMISSION and Climate Action.
Road Transport: Reducing CO2 Emissions From Vehi-
cles, accessed on Nov. 11, 2016. [Online]. Available:
http://ec.europa.eu/clima/policies/transport/vehicles/index en.htm
[4] L. J. Young, C. A. Gotway, J. Yang, G. Kearney, and C. DuClos, ‘‘Linking
health and environmental data in geographical analysis: It’s so much
more than centroids,’’ Spatial Spatio-Temporal Epidemiol., vol. 1, no. 1,
pp. 73–84, 2009.
[5] D. Olaru and J. Powell, ‘‘What activity-based analysis and personal sam-
pling can do for assessments of exposure to air pollutants?’’ in Air Pollution
Modeling and Its Application XIX. Dordrecht, The Netherlands: Springer,
2008, pp. 717–718.
[6] I. T. Ćirić, Ž. M. Ćojbašić, V. D. Nikolić, P. M. Živković, and M. A. Tomić,
‘‘Air quality estimation by computational intelligence methodolo-
gies,’’ Thermal Sci., vol. 16, no. 2, pp. S493–S504, 2013. [Online].
Available: http://www.scopus.com/inward/record.url?eid=2-s2.0-8487480
8919&partnerID=40&md5=9af13791ba66850493af418a57387bff
[7] M. V. Sokolova and A. Fernández-Caballero, ‘‘Modeling and imple-
menting an agent-based environmental health impact decision sup-
FIGURE 14. Example of a weighted road graph . port system,’’ Expert Syst. Appl., vol. 36, no. 2, pp. 2603–2614,
2009.
[8] F. Murena, ‘‘Measuring air quality over large urban areas: Development
and application of an air pollution index at the urban area of Naples,’’
VIII. CONCLUSION Atmos. Environm., vol. 38, no. 36, pp. 6195–6202, 2004.
[9] A. Namoun, C. A. Marín, B. S. Germain, N. Mehandjiev, and
The main contribution of this work is the definition of a J. Philips, ‘‘A multi-agent system for modelling urban transport
development process based on big data and intelligent sys- infrastructure using intelligent traffic forecasts,’’ in Proc. 6th Int.
tems concepts for a traffic regulation system according to Conf. (HoloMAS), vol. 8062, 2013, pp. 175–186. [Online]. Avail-
able: http://www.scopus.com/inward/record.url?eid=2-s2.0-84884969849
air quality data. We have, through this paper presented the &partnerID=40&md5=0a85fc95883c3c7911c22956260c410c
implementation of an air quality system for recommendation [10] H. Zahmatkesh, M. Saber, and M. Malekpour, ‘‘A new method for
and traffic regulation over distributed data gathered from urban travel rout planning based on air pollution sensor data,’’ Current
World Environ., vol. 10, pp. 699–704, May 2015. [Online]. Available:
different air quality sensors, users’ devices and other external http://www.cwejournal.org/vol10nospl-issue-may-2015/a-new-method-
databases, that are managed using Hadoop to ensure fast data for-urban-travel-rout-planning-based-on-air-pollution-sensor-data/

13200 VOLUME 5, 2017

A. El Fazziki et al.: Agent Based Traffic Regulation System for the Roadside Air Quality Control

[11] R. Anderson and J. Giddings, ‘‘Message: A mobile envi- ABDELAZIZ EL FAZZIKI received the M.S.
ronmental sensing system across grid environments,’’ Dept. degree from the University of Nancy, France,
Comput., Imperial College London, London, U.K.: Tech. in 1985, and the Ph.D. degree in computer science
Rep., 2009. from Cadi Ayyad University in 2002.
[12] M. I. Mead et al., ‘‘The use of electrochemical sensors for monitoring urban He has been with Cadi Ayyad University since
air quality in low-cost, high-density networks,’’ Atmos. Environ., vol. 70, 1985, where he is currently a Professor of com-
pp. 186–203, May 2013. puter science. He has been responsible for the
[13] D. Hasenfratz, O. Saukh, S. Sturzenegger, and L. Thiele, ‘‘Partic-
master’s degree program in information system
ipatory air pollution monitoring using smartphones,’’ in Proc. 2nd
engineering since 2006.
Int. Workshop Mobile Sens., 2012, pp. 1–5. [Online]. Available:
ftp://ftp.tik.ee.ethz.ch/pub/people/hdavid/HSST2012.pdf He was the Director of the Engineering Com-
[14] H. M. Kammoun, I. Kallel, J. Casillas, A. Abraham, and A. M. Alimi, puter Systems Laboratory between 2011 and 2015. He has co-authored
‘‘Adapt-Traf: An adaptive multiagent road traffic management sys- several papers on agent-based image processing, and is the main Author of
tem based on hybrid ant-hierarchical fuzzy model,’’ Transp. Res. C, over 20 papers in software engineering and data analytics field. His research
Emerg. Technol., vol. 42, pp. 147–167, May 2014. [Online]. Available: interests are related to software engineering, decision support, big data, data
http://www.sciencedirect.com/science/article/pii/S0968090X14000692 analytics, crowdsourcing, and e-government. In the MDA field, he has been
[15] X. Jin and L. Jie, ‘‘A study of multi-agent based model for urban intelligent involved in agent-based systems, service-oriented systems, and decisional
transport systems,’’ Int. J. Adv. Comput. Technol., vol. 4, pp. 126–134, systems.
Apr. 2012.
[16] G. Phillips-Wren and L. Jain, ‘‘Recent advances in intelligent decision
technologies,’’ in Knowledge-Based Intelligent Information and Engineer- DJAMAL BENSLIMANE is currently a Full
ing Systems (Lecture Notes in Computer Science), vol. 4692. Berlin, Professor of Computer Sciences with Université
Germany: Springer, 2007, pp. 567–571. Claude Bernard Lyon 1 and a member of the
[17] T. Fontes, L. M. Silva, S. R. Pereira, and M. C. Coelho, ‘‘Applica- Lyon Research Center for Images and Information
tion of artificial neural networks to predict the impact of traffic emis- Systems, Service Oriented Computing Research
sions on human health,’’ in Progress in Artificial Intelligence (Lecture Team, Lyon, France.
Notes in Computer Science), vol. 8154. Berlin, Germany: Springer, 2006, He has authored papers in well-known jour-
pp. 21–29. nals, including the IEEE TKDE, ACM TOIT,
[18] I. García, J. G. Rodriguez, and Y. M. Tenorio, ‘‘Artificial neural network SIGMOD Record, the IEEE TSC, the IEEE
models for prediction of ozone concentrations in Guadalajara, Mexico,’’ in
INTERNET COMPUTING, the WWW Journal, and the
Air Quality-Models and Applications. Rijeka, Croatia: InTech, Jun. 2011.
DPDB Journal. His research interests include distributed information sys-
[Online]. Available: http://www.intechopen.com/books/air-quality-
models-and-applications/artificial-neural-network-models-for-prediction- tems, Web services, ontologies, and databases.
of-ozone-concentrations-in-guadalajara-mexico
[19] Apache.Org. Apache Hadoop Documentation. [Online]. Available: ABDERRAHMANE SADIQ received the M.S.
http://hadoop.apache.org
degree in 2012 and the Ph.D. degree in computer
[20] T. Das and A. Mohapatro, ‘‘A study on big data integration with
science from the University of Marrakesh in 2017.
data warehouse,’’ Int. J. Comput. Trends Technol., vol. 9, no. 4,
pp. 188–192, Mar. 2014. [Online]. Available: https://doaj.org/
He has authored or co-authored several papers
article/3d9781bcab794e26b11f44ccefbc97fc on software engineering, data analytics, and air
[21] J. Zhao, J. Zhang, S. Jia, Q. Li, and Y. Zhu, ‘‘A MapReduce framework quality management. His research interests are
for on-road mobile fossil fuel combustion CO2 emission estimation,’’ in related to software engineering, decision sup-
Proc. 19th Int. Conf. Geoinformat., Jun. 2011, pp. 1–4. [Online]. Available: port, data analytics, crowdsourcing, and air quality
http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=5980759 management.
[22] G. Mitran, ‘‘The estimation of air pollution from road traffic by transport
modeling,’’ Int. J. Sci., Techn. Innov. Ind. Mach., Technol., vol. 5, no. 4,
pp. 60–63, 2007.
[23] A. Russo, F. Raischel, and P. G. Lind, ‘‘Air quality prediction using optimal JAMAL OUARZAZI is currently a Full Professor
neural networks with stochastic variables,’’ Atmos. Environm., vol. 79, of Chemistry with Cadi Ayyad University and a
pp. 822–830, Nov. 2013. member of the Laboratory of Physical chemistry of
[24] W. Fang, V. S. Sheng, X. Wen, and W. Pan, ‘‘Meteorological data Materials and Environment, Marrakesh, Morocco.
analysis using MapReduce,’’ Sci. World J., vol. 2014, pp. 1–10, He has authored papers in well-known jour-
Feb. 2014. nals, including Chemical Engineering, Inorganic
[25] D. Fan and P. Shi, ‘‘Improvement of Dijkstra’s algorithm Chemistry, Thermochimica Acta, Magnetic Res-
and its application in route planning,’’ in Proc. 7th Int. Conf. onance in Chemistry, Inorganica Chimica Acta,
Fuzzy Syst. Knowl. Discovery (FSKD), Yantai, China, 2010, and SpringerPlus. His research interests include
pp. 1901–1904. statistical and deterministic modeling of the air
[26] D. Lavbič and R. Rupnik, ‘‘Multi-agent system for decision support in
quality.
enterprises,’’ J. Inf. Org. Sci., vol. 33, no. 2, pp. 269–284, 2009.
[27] P. Lin, J. Thangarajah, and M. Winikoff, ‘‘AUML protocols and code
generation in the Prometheus design tool,’’ in Proc. 6th Int. Joint Conf. MOHAMED SADGAL received the Ph.D. degree
Auto. Agents Multiagent Syst. (AAMAS), Honolulu, HI, USA, 2007, in computer science from the University of Lyon
p. 270. in 1989, and the Ph.D. degree in computer science
[28] P. Busetta, R. Rönnquist, A. Hodgson, and A. Lucas, ‘‘JACK Intelligent from Cadi Ayyad University in 2005. From 1985
Agents-Components for Intelligent Agents in Java,’’ AgentLink News,
to 1987, he was an Associate Researcher with
vol. 1, no. 2, pp. 2–5, 1999.
Lyon I, France. He is currently a Professor with
[29] S. Muhammad, ‘‘Development and implementation of air quality
data mart for Ontario, Canada: A case study of air quality in Cadi Ayyad University, Marrakesh, Morocco. His
Ontario using OLAP tool,’’ Ph.D. dissertation, Lund Univ., Lund, research interests include computer vision, artifi-
Sweden, 2010. [Online]. Available: http://lup.lub.lu.se/student-papers/ cial intelligence, and multi-agent systems.
record/3559141/file/3559170.pdf

VOLUME 5, 2017 13201