-MASCAT is a research-based road traffic simulator. In this paper, we propose a plugin for MASCAT. The aim of this proposition is to provide a semantic data analysis for the simulators Multi-Agent System (MAS). The plugin introduce an interpretation phase during vehicle to vehicle communication (V2V). It will allow connected vehicles to make the most efficient behavioral decisions based on the state of surrounding environment. We designed a semantic web ontology to describe traffic data and ameliorate behavioral decisions. Our semantic plugin can link, structure, analyze MASCATs traffic data and can also optimize the Local Leader Election Protocol. Indeed, we demonstrate that a small percentage of connected cars can ensure traffic regulation specially in a shifting environment. Index Terms-Ontologies, Semantic Rules, Multi-agent System, Connected Vehicles, Local Leader Election Algorithm (LLEA) II. RELATED WORK
The progressive growth of the number of vehicles in our cities is considered as a cause of traffic congestion, but it is not the only reason for traffic jam. In order to study this problem thoroughly, it is necessary to simulate real traffic conditions by using an appropriate traffic simulator that would help us study the behavior of the vehicles.
Connected vehicles are able to sense and adapt their behavior according to their environment. How these vehicles will change the way we deal with traffic regulation? What percentage of connected vehicles is enough to ensure traffic regulation? In this paper, we will answer these questions, by describing the details of the framework we implemented as a semantic plugin for the MASCAT simulator. Semantic description provides the means to store information while giving it a logical meaning and a richer context.
This also means that we can construct advanced and intelligent queries over an ontology. The objective is to obtain the information inferred (by a reasoner) from the ontologys set of pre-defined relations and rules. Semantic rules are simply a set of IF-THEN statements for structuring complex axioms about a specific domain. We will start in section II with exploring the related work. In section IIIwe will describe our proposed solution. The solution design and implementation are detailed in section IV, while results are shown and analyzed in section V. Finally, conclusions and future work are exposed in section VI .
V2V (Vehicle-to-Vehicle) communications need to be tested through intensive experiments. Simulation models should include mobility models for providing accurate simulation of real time vehicular networking environment. The simulation tools should be selected based on their compatibility with application requirements and similarity to real-time traffic. We analyzed some existing traffic simulators. For the needs of our work, the chosen platform will not necessarily have to offer a very detailed definition of a mesh network.
Our affinities with free and community software push us to retain two main candidates: SUMO and MovSim, which offer in addition a lot of functionalities by their stage of maturity compared to the previous platforms. SUMO is widely used and well represented in research, and MovSim is a newer and less represented platform at the moment.
Simulation of Urban MObility (SUMO) is a road traffic
simulator [
Regarding V2V inter-vehicle communication, an extension of
SUMO is under development and aims to study the effects
of on-board applications on driver behavior [
We came up with a list of the most popular and free to
use GDBMS that support RDF/SPARQL as their data model:
Virtuoso, allegroGraph, StarDog, GraphDB, BlazeGraph.
And for rule-based inference [
In our case, we are interested in forward-chaining reasoning because query time is an essential factor that guarantees the required performance to keep up with the constantly ”shifting environment” of the connected vehicle in the traffic simulator MASCAT. Therefore, we need to dynamically create the Environment individuals and obtain the possible inferences before querying the semantic knowledge. Virtuoso and StarDog GDBMS only support Backward-Chaining Reasoning, which is not suitable for our case, so we are now only down to three Graph Database Management Systems: AllegroGraph, GraphDB and BlazeGraph. AllegroGraph [34] leads the largest deployment with loading and querying 1 Trillion triples. However AllegroGraph is available on a Windows platform through a Linux Virtual Machine, which could degrade the performance factor as directly declared on AllegroGraphs website. Therefore, we considered the use of Ontotexts GraphDB since it is more popular than BlazeGraph.
Graph database [
Based on our studies, we know that the vehicle driving
behavior is one of the main factors that can disturb traffic flow.
Each strange condition on the road can affect the surrounding
vehicles. In case of an accident for example because of snowy
weather, near surround- ing vehicles will be aware of the
situation and will try to control their behavior accordingly.
However, this reaction can have diverse effects on the flow,
because other vehicles on the road might not be aware of what
is happening. They only know that the road is blocked and they
cannot react appropriately. This is not the case of connected
vehicles which can react to regulate congested traffic in normal
traffic conditions as we demonstrated in our last work [
are respected by the highest number of surrounding CVs The CV must compute the optimal speed of the interval Ii which happens to be the median of the speed values of the surround- ing CVs in Ii The CV must choose the neighbor CV that has the closest speed to the computed median and send this decision to the surrounding CVs The Second Decision D2 is a collective decision that happens in a certain zone. It consists of three steps:
CVs The votes are counted for each CV in this zone The CV that received the most votes is elected to be the local leader for this zone and all CVs must adapt their speed to follow the speed of the local leader.
In our previous work, the implementation of this election protocol improved traffic flow in an optimal environment with no disturbances, by providing speed recommendations to the driver. This kind of recommendation will be offered to a connected vehicle after a negotiation with the surrounding vehicles based on a V2V communication. It allows a reduction of traffic congestion through an election of a local leader selected from the neighboring vehicles. These will adapt their behavior to follow the elected local leaders speed recommendation. Since this protocol is based on the concept of multi-agent system, each vehicle follows the cycle of Perception- Decision-Action However, many disturbances can affect the flow in real life like weather conditions, roadworks, unsafe and unexpected events, obstacles on the road, vehicle stopping, stuck vehicle, no protected accident area, emergency braking, unexpected queue end, ...
In order to be able to achieve traffic regulation in any context, it is strongly needed to provide a common framework that allows Traffic data to be shared and reused between entities. Therefore, it is necessary to Link and structure road traffic data exchanged between vehicles (agents) to simply access to the knowledge that it already contains and express accurately the real road traffic situation.
In other terms, we need to give our connected vehicles modeled as agents in the LLEA a particular knowledge of the environment to get them to: first understand the situation, then inform other connected vehicles on the road and finally emerge with a correct behavior whatever the conditions, without causing massive traffic congestion. As we showed in the state of art section, an ontology can provide a complete description of traffic data, and thus can allow an accurate interpretation of traffic context, which will give the connected vehicles adequate reactions. This ontology will represent the real road traffic road environment with the definition of all entities and properties that can affect the Connected Vehicles behavior. By augmenting the ontology with pre-defined traffic rules which respect to the European Traffic Norms and Regulations, the connected vehicles will adapt their behavior according to the knowledge inferred, at the interpretation level.
We choose to begin with inferring the adequate speed intervals needed for different situations on the roads, in order to provide CVs a with flexible intervals as we see in table (Fig.2). This will enhance the local leader election algorithm, since the local leader speed will be calculated based on dynamically inferred speed intervals. The framework we implemented is extensible, so in order to cover any environmental and contextual knowledge inference, we only need to add related concepts to the ontology and the corresponding rules. Fig.3 is a UML activity diagram that shows the control flow of the processing steps that enable a connected vehicle to understand its environment based on a realistic inferred knowledge. Individuals of our semantic data model, representing the environmental conditions, will be created dynamically on load time. Each connected vehicles can check its environment to know if there is a need to adapt its behavior. In our approach, the environment is controlled by three main criteria: the road type, weather, and visibility. As already explained, other criteria can be simply added further on. If the environment didn0t change, it would load the behavior of the connected vehicle in the LLEA directly without a level of interpretation. If it did change, a query is prepared with the Road Type and weather as well as the visibility detected by the corresponding sensor, as input parameters. The defined ontology is then queried with these aggregated values in order to infer new knowledge related to this correlation of values. If the query returns no result, this would mean that this is an occurrence of a new context, so new individuals will be created and added to the the ontology representing these detected conditions. If successful, the new inferred speed intervals will be returned by the query at the interpretation level in order to choose a Local Leader in the LLEA respecting the particular conditions at this particular time. The election protocol will be launched every 10 sec by every connected car, and based on the results of the election the Local Leader, vehicle adapt his speed in order to follow his Leader.
MASCAT is a research simulator based on a Multi-Agent System of vehicles, implemented in Java, over the MOVSIM simulator. In our previous research work on MASCAT we proposed and implemented a local leader election protocol. The purpose of this protocol was to ensure traffic fluidity and regulation on a highway. This solution has shown its effectiveness in tackling the congestion problem with a small percentage of connected vehicles. However, the MASCAT simulator does not include a very important feature which is the auto-adaption of connected vehicles under several disturbances in the environment. The connected vehicles in MASCAT do not take into account the environment in the ideal world under disturbances like the meteorological conditions such as the visibility, weather We will describe how we were able to alter the behavior of the MASCAT connected vehicles so that they would consider their surroundings, and this, without deteriorating the simulators performance while always ensuring traffic regulation.
The fact that the interactions between agents are highly information-dense raises many problems. Because of that our research was oriented to use a semantic method. This method can achieve a step of data interpretation by each agent on the multi-agent system. This goal can be achieved after implementing the Semantic Traffic Data Analysis plugin. Its main objectives are to Link, structure and analyze Traffic Data in order to optimize the Local Leader Election Protocol implemented in MASCAT. MASCAT Connected Vehicules would elect a local leader in a certain radius around them, based on the computed median speed value in the speed interval having the highest number of vehicles. The vehicle with the closest speed to this median is selected to become the local leader, so its speed will be adopted by surrounding connected vehicles. The Plugin will provide each connected vehicle with the adequate ”Speed Intervals” relatively to a certain context. The LLEA will be then executed, based on the inferred intervals, in order to compute an adequate recommended speed. Doing so will make the simulation more realistic by ensuring that the CVs behavior would adapt to a continuously shifting environment. In the following sub-sections we will start by describing the ontology based approach and the connection between the plugin and the simulator, then we will describe the graph database approach which we adopted for its high performance compared to the ontology based approach.
Firstly, as shown in the Fig. 4, we started to implement
our solution in Prote´ge´ with a preliminary prototype
including an initial ontology representing the context of
road traffic. This ontology includes main classes (Context,
Weather) and properties (hasWeather, hasSpeedLimit,
hasRecommendedSpeed)
Secondly, in order to obtain a recommended speed based
on knowledge about meteorological conditions, this
solution was augmented with semantic rules using semantic
web rule language (SWRL) to respond to each traffic
context (for each individual). This ontology uses the
rule engine Drools and the Pellet reasoner in order to
generate the axioms of the defined rules, inferring facts
and checking the consistency of the ontology.
Thirdly, based on this solution, we succeeded in returning
a value of ”the recommended speed” for a certain context
by using a SPARQL query. [
tor: In order to connect MASCAT to the semantic plugin, we implemented a Java module that uses the Jena API to query the OWL Ontology. The communication with the simulator was done through a dedicated Singleton Java class that we named the SemanticDecision class. A simple variable was used to toggle the semantic behavior On and Off, which will consequently toggle the interpretation phase On and Off, thus granting the vehicles with two possible behaviors: The First behavior where the CVs are following the steps local leader election algorithm (LLEA) without a semantic data analysis (without interpretation level) The second one where the CVs have an altered behavior: the Semantic Decision Behavior in which they adapt according to the inferred semantic knowledge.
This approach was tested in order to communicate an inferred recommended speed to the connected vehicles. The test showed that the ontology lacked the performance needed to keep up with the response delay of the simulator. Therefore, we opted to consider as an alternative approach the use of a graph database, which according to ref can perform much better than the ontology in terms of response time.
3) Graph Database approach: As we found when we explored the related work, and after severals tests and comparisons, we concluded that the use of the GraphDB Graph database would be the best choice to enhance the performance of our semantic plugin. We used Eclipse RDF4J (formerly known as Sesame) to connect GraphDB to the MASCAT code, and to process and handle RDF data. RDF4J also supports creating, parsing, scalable storage, reasoning and querying with RDF and Linked Data (Section II).
GraphDB can process an input OWL Ontology file and has
a built-in reasoner (TRREE) that automatically can make
inferences at load time (Forward-Chaining Reasoning), with
a very good performance level, enabling us to connect to
MASCAT and run our simulations appropriately. After
ensuring that the semantic plugin is highly functional at the
technical level, we went through enhancing it at the semantic
level, by extending the input ontology in order to be able
to infer speed intervals according to the vehicule’s context.
We explored many proposed existing ontologies [
Thus we added the following classes to the ontology Fig.5 : ”Environment, RoadType, Weather, Visibility” and the following object properties ”hasRoadType, hasWeather, hasVisibility” and the 4 inferred speed Intervals which are represented by 8 Data Properties which are the minimum and maximum bounds of each of these four intervals (hasInterval1Min, hasInterval1Max, ...)
We also augmented our Ontology with semantic rules to obtain the speed intervals when needed. In GraphDB these rules are called entailment rules and can be added in a custom ruleset .pie file. The main goal is to offer a connected vehicle with speed intervals (Fig.2) that correspond to its specific Environment individual, and thus respect road type, weather conditions, visibility and regulations.
As the simulations done in our latest work show, [
In this section, we describe the common experimentation
parameters to run our scenarios:
Since Highways suffer from an enormous daily amount of
vehicles, the road type studied in our tests is the Highway
type. The speed limitations are those defined by European
Laws (Fig.2). The baseline scenario consists of a dense traffic
state generated on a 3-lane straight highway. The input flow is
maintained to 1800 vehicles per hour during the 1200 seconds
simulation. Then, we gradually introduce CVs (from 0% to
30% ) which will execute the modified LLEA which takes
into consideration the inferred semantic knowledge. For now,
this knowledge consists of the speed intervals, needed for the
computation of the recommended speed, in each environment
and thus the election of the local leader. These scenarios
correspond to early CVs deployment phases that will occur
soon in real-life scenarios. Traffic flow is expected to be
gradually improved when we increase the number of CVs,
but we need to verify vehicles behaviors in critical weather
conditions based on the knowledge offered by our
implemented semantic plugin. Simulations results presented in this
paper were given by the mentioned Multi-Agent Simulator for
Connected and Automated Traffic (MASCAT) [
After the implementation of the GraphDB based Solution, it was time to test our overall solution to validate if the CVs in MASCAT were adapting their behaviors as a reaction to a specific disturbance (a specific weather condition) according to a deep interpretation realized by each agent in order to get correct intervals for this situation. In consequence, elect a local leader and respect the offered speed recommendation in the regulation strategy.
density in this scenario is set to 37 vehicles per kilometer. This setting tends to model a critical regime that can cause a network capacity drop due to the heterogeneities in the flow of numerous vehicles. We explained previously in Fig. ?? that connected vehicles should be able to adapt their speed after detecting the actual weather condition based on vehicle knowledge. For example, in case of snowy weather and low visibility, on a Highway Road, our semantic query will return the following inferred intervals: [0;30[ [15;50[ [17;30[ [26;40[
Figure 6 depicts the trajectories (and speeds) of all the vehicles in the right-most lane, for the baseline scenario (0% of CVs). Then, we gradually introduce CVs (from 0% to 30% ) executing the LLEA enabled with semantic knowledge. Traffic flow is expected to be gradually improved with the increasing number of CVs .
The baseline Scenario (0% CV) models a dense traffic state where congestion waves (in red color) appear spontaneously and grow, leading to the formation of massive traffic jams. The absence of connected vehicles in this baseline scenario will not allow the detection of environmental condition. The only possible way to control IDM vehicles speed in MASCAT simulator is to set their maximum speed to 40 km/h (equal to upper interval limit value). But, this maximum speed limit was not useful. Vehicles didnt respect this maximum value and no regulation strategy of the LLEA was used. Traffic congestion was mainly due to the speed heterogeneities between individuals, and it is accentuated by lane-changes.
The purpose of our approach is the self-adaptation in any weather condition using our knowledge-based electoral protocol thus maintaining traffic fluidity even in case of low speed (snowy weather). We can observe on the same figure, the results for the 10% of the connected vehicle which show how connected vehicles were able to adapt their speed based on our proposed semantic approach in LLEA. The local leaders tend to stabilize the flow in very difficult weather conditions. We notice after simulating our scenario with an increased percentage of connected vehicles (up to 50% of CV), connected vehicles were able to behave correctly as we see in the space-time diagrams. We also check our approach with all weather conditions mentioned in Fig. ?? for the road type Highway. We were satisfied by the results, our knowledgebased LLEA shows its effectiveness and reactivity in case of a weather disturbance.
Initial traffic density in this scenario is set to 15 vehicles per kilometer in order to detect the connected vehicles behaviors changes from one weather condition to another on a Highway Road. The aim of this scenario is to validate the concept of auto-adaptation of our approach and analyze what happens in the step of shifting between two extreme weather conditions. We tested this scenario in semantically enabled MASCAT version, with a Road length of 9000 m. This Road is divided into 9 road segments 1000 m each. From position 0 to 4500, we choose to have the first Environment (HighWay - Sunny - High) while the second Environment (HighWay - SnowyLow) will cover the position 4501 to 9000.In 7, we show how the vehicles started behaving at the beginning of the simulation.
As expected, we can see in Figure 7 that the highway is divided into two parts:
The first part (between the positions 0 and 4500) in which the CVs are moving noticeably fast in a Sunny Environment (Notice the green and blue colors of the CVs).
The second part (between the positions 4501 and 9000) in which the CVs are moving noticeably slower in a Snowy
Fig. 8. Space time Diagram for 100 percent of CVs During Shifting Scenario The space-time diagram shown in Fig.8 was obtained at the end of this scenario. Notice the middle road segment that is between the positions 4000 and 5000. As we can see, the cars are immediately slowing down when crossing from the Sunny to the Snowy Environment (Notice the red color of the CVs on the position 4500). This means that the CVs are actually abiding to the semantic LLEP result that was executed using the queried Speed Intervals. Each time a CV enters a new Environment, it queries the GraphDB for these Speed Intervals and will keep using them until this environment changes, therefore optimizing MASCATs performance. These results validates the overall behavior expected from our projects outcomes (the plugin and the ontology) that we detailed in the previous sections. Our solution was tested and validated after several simulations in MASCAT simulator. We can safely say that the Graph Database approach seems to be way more convenient than its semantic web ontology counterpart by effectively querying the Graph Database structure without having to rely on a less - performant exclusive ontology approach. We switched to the GraphDB GDBMS because we found out that the standard ontology could be easily inserted into a GraphDB repository. Doing so, we accomplished our projectsg´oals by implementing a plugin that would make the simulations more realistic. Being aware of the requirements and constraints, we ultimately did not deteriorate MASCATs pace while performing the necessary computations and tasks.
As we have seen in this paper, we have used many tools in order to develop this project and make it work. We design and build a working Semantic Web Ontology on Prote´ge´ and GraphDB, and successfully linked it to the MASCAT simulator by implementing a rigorous plugin. We also extended and improved our solution by dynamically creating the Ontologys individuals and activating the Visibility Sensor class in MASCAT. Our plugin helped achieve traffic regulation, by enabling the connected vehicles to sense and adapt to changing environmental conditions without deteriorating the simulator’s performance.
To summarize, results indicate that our approach can have a positive impact on the peoples lives where the driver of a CV could obtain a message in which the recommended speed of his vehicle is specified. He would then adapt to this speed knowing that its the most optimal, safe and law-abiding way to go. This way, he would make sure that the decision he takes would contribute in boosting the overall traffic conditions on the road and help in maintaining an efficient traffic situation that is conform to traffic regulations and laws. Additionally, our plugin made MASCATs simulations more realistic, and therefore more reliable by helping road traffic experts expand their studies and analyze the simulators results after introducing new factors that can now be taken into consideration thanks to the ever-expanding Semantic Web knowledge. The designed Ontology can be tailored to the experts needs in order to try to replicate any simulation environment thus contributing in finding solutions to the daily traffic congestions in shifting scenarios.
We propose now some of our recommendations for future works and implementations. We understand that this work is not fully complete in what concerns Semantic Traffic Data Analysis. The first noticeable thing that could be done is expanding the Ontologys designs in order to include more parameters and entities that can influence the behavior of the CVs in MASCAT such as the Drivers skills & abilities, traffic lights, road intersections, obstacles & road works, pedestrians, etc. Ultimately, the CVs sensors could be linked to a stream processing software platform like Apaches Kafka in order to simulate real-time environment detection in MASCAT.