An Overview of Agent-based Traffic Simulators
Johannes Nguyen, Simon T. Powers, Neil Urquhart, Thomas Farrenkopf, Michael Guckert
AA N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS P REPRINT
Johannes Nguyen, Thomas Farrenkopf,Michael Guckert
Kompetenzzentrum für InformationstechnologieTechnische Hochschule Mittelhessen61169 Friedberg, Germany {johannes.nguyen,thomas.farrenkopf,michael.guckert}@mnd.thm.de
Simon T. Powers, Neil Urquhart
School of ComputingEdinburgh Napier UniversityEH10 5DT, Edinburgh, United Kingdom {s.powers,n.urquhart}@napier.ac.uk
February 16, 2021 A BSTRACT
In most countries population in urban areas is growing, while available travel infrastructure andresources are limited. At the same time desires to minimise environmental impact and energy usehave led to new requirements in the field of inner-city transportation. As a result, the portfolioof mobility services provided is developing in order to improve the use of the available resources.Computer-based simulation is an accepted means for investigating the effects of new transportationpolicies and services. Many researchers are faced with the question of choosing a suitable simulatorfor their specific research question. In this paper, we review a broad spectrum of recent and historicallyimportant applications, in order to provide an overview of available work and to help researchersmake a more informed decision on the selection of a suitable simulator. We discuss strengths andweaknesses of the applications and identify gaps for which we argue that more detailed work isrequired. K eywords Traffic Simulation · Multi-agent Systems · Simulation Software
Over the last decades, structures of transportation and personal mobility have repeatedly faced radical changes. Today,terms such as smart cities and
Intelligent Transportation Systems (ITS) dominate the discussion. ITS use advancedinformation technology to improve traffic management. For example, central traffic control systems are deployed toprovide real-time information on road closures, parking space availability, etc. in order to minimise avoidable trafficproblems. In many countries, ITS are already in use with the main objectives to increase general traffic safety andto make more efficient use of the existing infrastructure. Driven by technological innovation and changing societaldemands, current transportation structures have evolved into complex systems. Based on this, [1] have emphasised theimportance of a more user-centric approach in modern ITS, focusing on social and environmental aspects of the system.Computer-based simulations can be used to plan and assess the effects of new policies in advance, and provide decisionsupport for transport planners and authorities. As ideas on traffic simulation date back to the 1970s [2, 3] a variety ofcomputer-based simulators has been developed. The implementation of these applications has been primarily driven bytwo types of simulation models: cellular automata and multi-agent based . The traditional simulation model is based oncellular automata. A cellular automaton is a computer model consisting of regular sets of identical cells. Each cell canadopt certain states and interacts with a defined number of neighbouring cells. Cellular automata are used to modelspatially discrete dynamic systems, in which the development of individual cells at time t + 1 depends primarily onthe cell states in a given neighbourhood and on its own state at time t . According to [4], the main difference betweencellular automata and multi-agent models can be described as follows. In models based on cellular automata, entitiesare represented as static cells while spatial and other processes move across or through the cells. Models based oncellular automata have been criticised for being an oversimplification of reality. Therefore, state of the art research a r X i v : . [ c s . M A ] F e b N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021has shown a growing interest in the application of multi-agent models in the field of traffic simulation. In contrast tothe traditional simulation model, the modelled entities in multi-agent models are represented as autonomous softwareagents that actively move in space, perceiving and interacting with other agents and their surrounding environmentthrough sensors [4].Due to the broad spectrum of traffic simulators that have emerged over the years, end-users working on specific researchquestions, such as testing of traffic light algorithms, are faced with the issue of finding an appropriate simulationenvironment. [5] describe the phenomenon that in many cases instead of exploiting the potential of already existingsimulators, researchers have implemented their own research specific applications. This may be the result of nothaving sufficient overview on the set of available simulators and their features which requires a considerable amount ofin-depth research. For this reason, this paper provides a review of available and established traffic simulators. Giventhe large number of available applications it is not possible to review all of them in this paper. We have selected 10applications that have attracted attention and have had impact on the research community. We aim at covering a broaddiversity of applications that range from historically important foundation work such as TRANSIMS and SUMO tomore recent and advanced multi-agent approaches such as MATSim, AgentPolis, ATSim etc. In particular, applicationfocus and implemented features as well as possible limitations and gaps of the simulators will be discussed in order tohelp end-users make an informed assessment of the suitability of the applications for their specific research questions.This paper is organised as follows: The next section provides a short overview on the diversity of traffic simulators andpresents the scope and perspectives of comparison of this paper. Following this, a review of established simulation toolsis given. The review is primarily based on publications and publicly available discussions of expert communities. Afterthat we categorise the simulators presented by area of application and discuss relevant gaps where we identified missingfunctionality that we consider relevant for effective traffic simulations. The paper ends with a discussion of future stepsthat can help close these gaps.
In order to be able to discuss the suitability of different traffic simulators, it is important to obtain an overview of state ofthe art research and available systems. Based on available publications (e.g. [6, 7, 1]), discussions as well as extensiveinternet research using academic search engines such as Google Scholar, we filter out current research subjects andcategorise the simulators based on their underlying models. We included a wide spectrum of applications that rangefrom historically important foundation work to more recent and advanced multi-agent approaches. In the literature,traffic simulators are often divided into four different groups [1, 8]:1. Macroscopic simulations focus on traffic flow modelling based on high-level mathematical models. This typeof simulation can be used for the analysis of wide-area systems in which no detailed modelling is required, e.g.the simulation of motorway traffic. Given the low level of detail, macroscopic simulations are relatively fastand require less computing power.2.
Microscopic simulations focus on modelling individual entities based on a high level of detail. Possibleentities include travellers, vehicles, traffic lights, etc. This type of simulation is often used for the analysis ofurban traffic. It is possible to analyse both macroscopic and microscopic aspects (e.g. traffic lights algorithm,multimodal traffic) of the system. Consequently, microscopic simulations may result in longer computingtimes.3.
Mesoscopic simulations are a mixture of macroscopic and microscopic simulation models. Traffic entities aremodelled at a higher level of detail than macroscopic approaches, however, interaction and behaviour of theindividuals appear to be less detailed.4.
Nanoscopic simulations are even more detailed that microscopic approaches. This type of simulation is appliedin the field of autonomous driving, in which internal functions of the vehicles such as gear shifting or vehiclevision have to be examined.Depending on the group of simulators, different aspects of the transport system are covered with a different level ofdetail. In general, it should be outlined that traffic simulation is a combination of algorithms and data. The use ofreal-world data makes simulation models realistic and has the potential to increase the accuracy and relevance of theresults. Relevant data which is frequently used for the modelling of traffic scenarios include geographical data andpublic transport timetables as well as census information. Such data may be classed as either scenario data or instancedata. Scenario data defines the underlying transport networks, whilst instance data defines the particular transportation N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021demand – it is possible to have many instances within one scenario. Geographical data can supply the underlyinggraph structures of road networks including capacities, distances and travel times. Depending on the granularity ofthe simulation, such data could also include details of junctions and lane layouts which can have an impact on times,distances and available routes. The underlying street graph can also be enhanced with rail links and tram/metro linkswhere these are relevant to the problem being investigated. A common source for such data is OpenStreetMap (OSM)[9], which can provide transport network graph data for roads, metro routes, railway lines etc. for most areas of Europeand North America and other areas. OSM data is released under an Open Source license allowing it to be widely used byresearchers. Scenario data is often derived from sources such as census information and results of travel surveys. UnlikeOSM there is no defined format for such data, which may differ in terms of availability, format and license dependingon the geographical area that it refers to. In addition to this, the selection of algorithms significantly influences theoptions and limitations of the underlying simulation models. In this work, a distinction is made between the followingcategories. An extract of reviewed example applications for each of the categories is given.•
Fully Agent-Based : MATSim [10], ITSUMO [11], MovSim [12], MASCAT [13], MATISSE [14], POLARIS[15], AgentPolis [5], OPUS [16], MOSAIIC [17], MARS [18], SimMobility [19]•
Featuring Agent-Technology : ATSim [20], FastTrans [21]•
Not Agent-Based : TRANSIMS [22], SUMO [23], OpenTraffic [24, 25], CONTRAM [26], PTV VIS-SIM/VISUM [27], GETRAM/AIMSUN [28], PARAMICS [29], MITSIM [30], FreeSim [31], TSIS/CORSIM[32], VATSIM [33], DRACULA [34], RENAISSANCE [35], SimTraffic [36]Primary focus is put on the approaches of the first and the second category as state of the art research has shown in agrowing interest in the application of agent-based simulation models. However, important applications that have laidthe groundwork for the development of agent-based systems are worth mentioning and have also been included. Theapplications are studied with regard to three key aspects. The first aspect refers to the background of the application. Inthis section licensing, focus and area of application are presented. Based on this, architecture and internal processeswithin the application are described. This provides insight into the functions of the application as well as openness forcustom extensions. Finally, the modelling capabilities of the application are reviewed. In particular, methods for themodelling of demand, decision-making and learning capabilities are discussed.
Approaches to the simulation of traffic have been published since the 1970s. In 1990 these ideas were combined withfurther concepts developed within the TRANSIMS project [37]. The TRANSIMS project is an integrated collection ofapplications for regional transport system analysis based on a cellular automaton [38]. This approach requires trafficmovements to be modelled at a high level of detail. Although the TRANSIMS approach is not agent-based it is worthmentioning as it lays the foundation for later works in the field of computer-based traffic simulations. The TRANSIMSproject was sponsored by the US government with the intention to support metropolitan planning organisations [22].[39] have demonstrated use of application for
Chittenden County, Vermont, a medium-sized urban area with a populationof approximately 145.000 citizens. Given that TRANSIMS is an application with general focus, it can be applied tostudy a wide range of research questions. [40] have collected and discussed a selection of case studies conducted withTRANSIMS. For example, evacuation scenarios [41] and industrial land use in the city of Moreno Valley [42] havebeen simulated with TRANSIMS. Apart from this, it can be noted that the source code is written in C++ and publishedas an open-source project Architecture
The TRANSIMS architecture is structured in five functional components [43, 44]. Each component has been designedfor a specific purpose (see Figure 1). The first component is referred to as the population synthesiser . This componentdeals with the initialisation of agents for the simulation. For this, demographic census information is used in order tocreate a synthetic population in which travellers are assigned to specific households [45]. Each traveller possesses a listof activities that defines the number of trips. A
Household Activity Generator is responsible for producing this dailylist of activities. The generator is based on data provided by household activity surveys, workplace surveys, as wellas land use information. Activities include information about priority, start time, duration, mode preference, location,etc. For each activity, a
Route Planner component calculates various aspects of the travel journey such as preferredroute, transportation modes, parking locations, travelling companions, etc. This process requires data on the geographic see https://sourceforge.net/projects/transims/ - (access on 05/05/2020) N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021network in addition to output information provided by the first and second component. The required geographic networkdata includes information on locations, links, mode accessibility as well as travel times. Based on this information,the fourth component finally, executes the micro-simulation . In comparison to other microscopic approaches, theTRANSIMS model appears to be less detailed [43]. The final component becomes active once the simulation has ended.In this component, the output data of the simulation is assessed and traffic emissions are calculated.
Population Synthesiser Household Activity Generator Route PlannerMicro-Simulation
Emissions
Calculator
Figure 1: TRANSIMS - Architecture.
Modelling Capabilities
Although complex computation is required, it is possible to model subtle details such as carpooling, lane switching andmultimodal transportation [38, 46]. The microscopic approach allows for detailed modelling. However, the overallmodelling approach is based on static mathematical distributions using real-world data. Concerning traveller demand,TRANSIMS implements an activity-based approach in which surveyed information (census data, origin-destinationmatrices, etc.) is used to produce pseudo activities for trip generation [47]. Survey information is used in order to modelevery traveller and vehicle in TRANSIMS explicitly. For this, TRANSIMS provides a wide number of parametersthat can be configured. Regarding the aspects of decision-making and learning, the TRANSIMS approach appearsto be static based on the nature of the underlying model. However, as TRANSIMS is considered an early approach,dynamic behaviour has not been a central focus. This becomes evident especially as travellers statically choose theirtransportation mode based on survey data instead of dynamically trying to optimise individual preferences [38].
Basic concepts and ideas from the TRANSIMS project were further refined in 1998 and introduced in a new agent-basedsimulation project, MATSim, by Kai Nagel from ETH Zurich. In 2004, Kay W. Axhausen joined the project. Withinthe course of this project, knowledge from various fields such as physics, engineering, traffic flow, computing, choicemodelling and complex adaptive systems has been incorporated [10]. Further development is currently being conductedas part of a collaboration between the TU Berlin, ETH Zurich and CNRS Lyon. MATSim is an agent-based softwareframework implemented in Java. The framework has a general focus and is particularly designed for the simulation oflarge-scale transportation scenarios. Hence a particular effort was made for efficient computational processing andparallelisation [48, 49]. MATSim has been used for simulating various scenarios such as transport energy demandplanning [50] or autonomous taxi services in multimodal traffic [51].
Architecture [10] provide a detailed documentation of the MATSim system. The framework consists of five modules for
InitialDemand, Execution, Scoring, Replanning and
Analysis (see figure 2). Based on the modular approach, custom modulescan be implemented and integrated into MATsim in order to replace or to upgrade existing modules. The first moduledeals with modelling and generation of agents. This module makes use of real-world census data in order to create aninitial population of agents. Following the TRANSIMS approach, agents in MATSim also possess a list of activities(activity chains). Generation of activities is also based on survey data. Furthermore, every agent possesses a list ofplans in which different combinations of actions and choices are defined. This includes choices not only about classicaltraffic properties such as routes and transportation mode but also time scheduling. The execution module is responsiblefor the selection of plans and their execution within the simulated environment. The scoring module assigns a score4 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021to every plan based on a utility function. MATSim uses a co-evolutionary algorithm for optimising individual utility(selfish routing). An ordinary evolutionary algorithm would have searched for a global optimum. The co-evolutionaryalgorithm is applied to evolve the set of agent plans of the travellers at the same time regarding both travel choices aswell as activity scheduling. This is performed in the replanning module . The simulation cycle (execution - scoring -replanning) repeats until a MATSim equilibrium is reached. This is the case when agent scores stabilise. Finally, theoutput data of the simulation is aggregated in the analysis module . Initial Demand Execution Scoring Analysis
Replanning
Figure 2: MATSim - Architecture.
Modelling Capabilities
As [10] describe, MATSim simulations are commonly modelled for single days. However, experiments have demon-strated the simulation of multi-day scenarios [52]. Similar to the TRANSIMS approach, for demand modelling,travellers are modelled explicitly. MATSim provides options for generating an initial population based on user input andcensus data. MATSim also uses an activity-based approach. Using additional survey data, agents are assigned a list ofgenerated activities. It should be noted, that travel demand changes with every iteration of the simulation as replanningmechanisms include rescheduling of activities. Regarding the aspects of agent decision-making, MATSim uses adiscrete-choice model [10]. This implies that quantitative methods are used to determine probabilistic distributions foraction alternatives. Agents select plans based on the assigned score. A higher score increases the probability of a planto be chosen. The selection process of agents is based on a multi-nomial logit (MNL) choice model (see [53]). Anotheraspect reviewed is agent learning. In MATSim agents are described to not display conventional learning capabilities inthe sense of traditional human learning behaviour. Instead, the learning capabilities refer more to moving the modelcloser to a stable equilibrium state. For this, MATSim implemented the co-evolutionary algorithm in combination witha stochastic
Monte-Carlo approach [54].
MovSim (Multi-model open-source vehicular-traffic Simulator) is an agent-based microscopic traffic simulator writtenin Java. The project was started in the late 1990s for educational purposes by Martin Treiber and Arne Kesting both,being researchers in the
Department of Transport and Traffic Sciences at the Technische Universität Dresden [12]. Incontrast to most commercially available traffic simulation tools that model specific road networks (e.g. cities), MovSimfocuses on the simulation of fundamental issues in the field of traffic dynamics. For example, MovSim has been used tosimulate the effects of the drivers longitudinal movement (acceleration and braking) on traffic jams, such as stop-and-gowaves . Specific attention has been given to modelling lane changes and roadside control such as ITS [55]. Because ofthis particular focus on traffic dynamics, Movsim has also been applied for the simulation of rather unconventionalscenarios such as ski marathons [56]. There are also a series of MovSim extensions that can be used for specificresearch questions such as car-to-car (C2C)- or car-to-infrastructure (C2I) communication, driving assistance (adaptivecruise control), fuel-consumption, etc [57, 58, 59, 60]. Furthermore, MovSim also includes a number of referenceimplementations for established mathematical car-following models as described in [61]. Because of its simplicity,MovSim has become known to the public as the application has been presented at renowned exhibitions such as CeBIT(2009 & 2010) and the Volkswagen Autostadt, as well as in popular media, e.g. the Wallstreet journal (July 1, 2005).
Architecture
As described in [55], the MovSim structure is organised in three layers (see figure 3). In the input layer , simulationsettings and parameters are defined. For this, MovSim provides three options. The user can input information using agraphical user interface (GUI), command line or as XML files. This information is forwarded to the main loop layer .In this layer, agent control and movement are implemented. The simulation controller continuously calculates thesimulation in a loop as MovSim is based on a time-continuous model. It can be noted that the simulation controller5 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021primarily focuses on quantitative models. Different submodules implement logic for aspect-specific agent behavioursuch as acceleration, braking, lane-changing, etc. Parameters and user input can be changed at any time during thesimulation loop. The effects and consequences of these changes are instantly applied to the simulation and can thereforebe observed in real time without having to pause the application. At the same time, two additional modules act asobservers to the simulation loop in order to extract information for the output layer . The SimViewer module deals withinformation relevant for the visualisation of the simulated scenarios. MovSim includes implementation for both, 2Dand 3D visualisation. Users can choose between a microscopic (cockpit perspective) or macroscopic (bird’s eye) viewof the simulation. Apart from the visualisation, simulated data can also be output as files.
Input Main Loop
Command LineXML FilesGUI Simulation
Controller
SimViewerMeasurements
Output
VideoScreenFileRoad SectionVehicle AgentAspect-specific Models
Figure 3: MovSim - Architecture.
Modelling Capabilities
In MovSim agents represent vehicles or drivers of vehicles in a general road network in order to investigate thefundamental issues in the field of traffic dynamics. However, it can be noted that MovSim does not focus on modellingaspects such as multimodal transportation that are irrelevant to its specific research objective. The MovSim approachprimarily focuses on quantitative modelling. Concerning the aspect of demand modelling, MovSim implements aseries of car following models in which vehicles are represented as moving particles in the network with no specificroute being assigned. Hence, traffic volume can be defined using numerical input parameters. In the movsim demoapplication , demand is parameterised using individual sliders on the GUI. With regard to agent-decision-making,MovSim focuses on movement-related driver decisions such as lane-changing, acceleration, braking, etc. For this,MovSim considers discrete-choice modelling. Finally, the last aspect to be reviewed refers to agent learning which inMovSim is not a central focus. Simulation of Urban MObility (SUMO) is a software package for microscopic traffic simulation. The project was createdby the German Aerospace Center (DLR) out of the necessity for an appropriate open-source solution. Other projectssuch as TRANSIMS, which are now open-source, were difficult to obtain at that time [10]. Available traffic applicationswere mainly used as black-boxes with no options to examine the underlying simulation model [23]. Researcherswere restricted by the given parameterisation and modelling. Existing solutions essentially included no options toimplement custom ideas. Addressing these problems, the first version of SUMO was published in 2001 [23]. Sincethen, SUMO has been accepted by a wide community. In particular, SUMO should rather be considered as a collectionof software applications to prepare and perform traffic simulation. SUMO is implemented using the programminglanguage C++. It includes several sub-applications for handling network data (netgen, netconvert), demand modelling(od2trips, activitygen), routing calculation (jtrrouter, dfrouter, duarouter), and other aspect-specific modules such asvehicle communication (V2X) [62]. Concerning the road network, it is possible in SUMO to either manually generatedata or import data from external sources (inter alia OSM, VISSIM, VISUM, XML, CSV). SUMO has been appliedparticularly for the analysis of traffic-light algorithms. For this purpose, the
Traffic Control Interface (TraCI)
API wasdeveloped in order to interact with an external application via a socket connection [8]. Originally, SUMO is based oncellular automata. The SUMO approach is therefore not agent-based. Nevertheless, SUMO is worth mentioning as it isconsidered one of the most important approaches in the field of applied traffic simulation together with the TRANSIMSproject. In 2014, SUMO was integrated with the
Java Agent Development Framework (JADE) (see [63]) in order tocreate SUMO-based simulations that are compatible with recent agent technologies [64, 65]. JADE is an open-source N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021software framework that is used for the implementation of agent-based applications. This combination of SUMO andJADE has been used for simulating and assessing the effects of traffic management systems [65, 66]. The followingsection on the SUMO architecture focuses on the combination of SUMO and JADE.
Architecture [64] have implemented a software connector that enables communication between the two software environments.This connector is referred to as
TraSMAPI (Traffic Simulation Manager Application Programming Interface) . Fromthe SUMO perspective, the TraCI API is the central component for the integration of SUMO and JADE. TraSMAPIcommunicates with the TraCI API and acts as an intermediary. Although the project focuses on the integration ofSUMO and JADE, TraSMAPI is abstracted to be able to handle various simulators besides SUMO (see Figure 4). Thismakes it possible to compare the results of different simulators. The combination of SUMO, JADE and TraSMAPI cantherefore be categorised as an
Artificial Transportation System (ATS) which is an extension of traditional modelling andsimulation approaches with the ability to integrate different simulation models in a virtual environment [67].
JADE FrameworkTraSMAPI
SUMO other simulator
TraCI other sim. API
Figure 4: TraSMAPI - Architecture.In this approach, JADE agents represent drivers that are linked to vehicles in SUMO. A separation of strategic andtactic-reactive agent behaviour has been implemented with two layers. This is also referred to as the delegate-agentconcept and allows for more parallel computation [68]. Basically in this scenario the delegate-agent concept canbe understood as a separation of cognitive and reactive actions from the executing driving tasks [64]. In particular,the strategic layer deals with collection and processing of information from the surrounding environment. Based onthis information the agent chooses its travel route, also in the strategic layer. In addition, in the tactic-reactive layerdriving related behaviour such as acceleration, braking, lane changing, etc. is implemented. Based on the functionalrequirements of the two layers, the strategic layer was kept in JADE whereas the tactic-reactive layer was realised inSUMO (see Figure 5).
Driver Agent
Strategic
Tactic-Reactive
JADE
StrategicSUMOTactic-Reactive
Figure 5: Delegate-Agent Concept.
Modelling Capabilities
The original SUMO package provides two options for demand modelling [8]. Travel demand can be defined manuallyusing an origin-destination matrix (od2trips). In this case vehicles have to be defined explicitly, specifying the id, route,travel times, etc. Alternatively, a synthetic population of travellers can be generated using an activity-based approach(activitygen). As SUMO is described as a microsimulator, it is possible to model subtle details such as lane changes ormultimodal transportation. With the exception of walking pedestrians to not being simulated but instead travel times7 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021are estimated [23]. Referring to the modelling of decision-making processes as well as learning, it can be noted thatthis has not been a central focus of the SUMO project. In this context, the only related aspect is the selection of travelroutes. These are usually calculated based on established routing algorithms that make use of different cost functions[62]. However, for these situations, SUMO has prepared the TraCI API for extension by external modules. For example,with the integration of agent technology it is possible to further extend these capabilities. [64] have demonstrated thepotential of their approach to represent human behaviour by applying reinforcement learning techniques.
ITSUMO (Intelligent Transportation System for Urban Mobility) is an open-source agent-based microscopic simulatorfor traffic scenarios. The simulator is written in C++ and Java, and was first presented in 2006. Since then ITSUMO hasbeen continuously refined and advanced [69, 11]. Apart from the similarity in name, there is no direct link betweenITSUMO and the previously described SUMO project. ITSUMO has also been applied for the simulation of routechoice scenarios. However, the primary focus of the application is on traffic control [70, 11]. For example, ITSUMOcan be used for testing traffic light algorithms as it is possible in ITSUMO to model control measures and driverreactions. In ITSUMO, traffic actors such as drivers and traffic lights are modelled as agents. As the creators describe,ITSUMO was developed out of the lack of customising options in available simulation tools, as most of the existingsolutions were developed for specific purposes. Other drawbacks described are for related simulation tools to not beingfully agent-based, for them to be relying on strong simplifying assumptions, or deficiencies in the demand planningoptions [11]. Therefore, it can be noted that the ITSUMO approach is fully agent-based which aims at addressing thedeficiencies mentioned above.
Architecture
The system architecture is structured in five components [71, 70] (see Figure 6). The first component is a database .This database contains information about the geographic traffic network as well as data used in the simulation such asinsertion rate of vehicles or origin and destination of the drivers. Network data can be imported in the XML format.ITSUMO uses map data from OpenStreetMap. The second component is described as the simulation kernel . Thiscomponent accesses data stored in the database, executes the simulation and manages agent interaction. Apart from this,there is a control component in which traffic light and other control entities are implemented. Communication betweenthe simulation kernel and the control component is established using a socket connection. The control componentpasses information to the simulation kernel for example, to provide instructions for traffic light agents or other simulatedcontrol entities. Finally, results of the simulation are output in a separate component. For this, sensors and detectors areused during the simulation in order to collect relevant data such as travel times, average speed, etc. The output moduleprovides two visualisation options for both, a microscopic and macroscopic view of the simulation. In addition, resultscan also be exported as data files.
Database SimulationKernel VisualisationControl
Figure 6: ITSUMO - Architecture.
Modelling Capabilities
In ITSUMO demand can be modelled using a defined origin-destination (OD) matrix or by generating synthetic demandusing uniform probabilities for a set of locations. For each combination of origin and destination, vehicles are generatedand a route is determined. Regarding the aspects of agent decision-making, the ITSUMO approach is fairly detailed.ITSUMO distinguishes between prejourney planning and the en route (re)planning. En route replanning refers tochanges in decision-making during the journey. ITSUMO primarily focuses on route planning. For this, an initial routechoice is made prejourney using established routing algorithms such as Dijkstra, A*, ARA*. Regardless of the selectedrouting algorithm, routing can be either centralised or decentralised. Centralised routing assumes complete knowledgeof the network and in reality is therefore only applicable if all drivers use a navigation system. In contrast, decentralisedrouting is based merely on local information. More specifically, knowledge that is based on how an agent perceives8 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021its local environment. ITSUMO implements two strategies for en route replanning. Agents can either replan at everyintersection ( intersection replanning ) or in case of a delay during the journey ( delay replanning ). For delay replanning,each time an agent arrives at a new link, the expected arrival time is calculated based on local information of the agent.The agent then starts replanning (searches for new routes), if the calculated arrival time exceeds his former assumption.Finally, concerning the aspects of agent learning, methods of reinforcement learning have been applied particularly forthe implementation of traffic control agents [72].
MATISSE is a large-scale multi-agent based simulation platform for Intelligent Transportation Systems (ITS) writtenin Java [14, 73]. The simulator was created at a time when only a few fully agent-based approaches existed, andhas been released for non-commercial use under the name
DIVAs 4 . Within this set of fully agent-based simulationmodels, the creators of MATISSE criticised the lack of core agent mechanisms. Based on this, MATISSE focuseson the representation of core agent simulation mechanisms such as sensing, diverse communication types, and thecapability to model collisions in traffic networks [14]. Agents are used for the representation of both vehicles as wellas intersection controllers. The simulator considers both autonomous and non-autonomous vehicles. In MATISSE,inter-vehicle communication is possible as well as communication with intersection controllers. The simulator usesOSM data for the representation of the network. In particular, MATISSE specialises in the simulation of scenariosrelated to traffic safety. MATISSE has been used to investigate strategies to manage evacuation operations in urbanareas [74]. It is possible to change properties of the simulation at runtime without having to interrupt calculations.
Architecture
The MATISSE architecture is structured in three layers (see Figure 7) [14]. The first layer is described as
MATISSEControl and Visualisation Module . It includes a control GUI for parameterisation and configuration of the simulationmodel. Furthermore, 2D/3D visualisation is implemented in this layer. Apart from this, there is a communication layer.This layer includes a
Message Transport Service that acts as a controller in order to enable communication betweenthe user interface and the simulation system. The third layer
MATISSE Simulation System is the core element of theapplication. In this layer, calculations are performed in order to run the simulation. The layer is divided into threesubsystems. The first subsystem is called
Agent System . This subsystem is responsible for the creation and controlof various agents types (vehicles, traffic lights, etc.). During the simulation, agents are able to communicate using adedicated
Agent-to-Agent Message Transport Service . The second subsystem is described as the
Environment System .This subsystem creates and controls additional simulation elements related to the traffic environment. For example, thisincludes elements such as the traffic network. A dedicated
Agent-Environment Message Transport Service connects theenvironment system with the agent system. Finally, a third subsystem is the
Simulation Microkernel . This subsystemhandles all tasks related to the simulation workflow.
Message Transport ServiceMATISSE Control and Visualisation Module
Simulation Control GUI 2D/3D VisualisationAgent System Environment System Simulation
Microkernel
Agent Environment Message Transport Service
MATISSE Simulation System
Figure 7: MATISSE - Architecture.
Modelling Capabilities
Simulation modelling and execution of MATISSE are described in detail on the official MATISSE website [73]. Fordemand modelling, a total number of simulated agents is specified by the user. Based on this, MATISSE either uses a9 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021normal distribution or a user specified distribution in order to initialise agents for defined user entry and exit points. InMATISSE, agent behaviour modelling considers several aspects. First, agent movement is based on car-following andlane-changing models. During the simulation, agents are given a destination. A route is determined using a non-trivialdynamic routing algorithm based on the presented traffic situation. Agents are able to perform en route replanning, ifthe original goal has to change. For example, this is the case if an emergency notification is issued. Furthermore inMATISSE, agents act on decentralised knowledge that is theoretically based on information perceived within their sensorrange, also described as circle-of-influence . However, for software design purposes the implementation differs from thetheoretical idea. MATISSE introduces two design related concepts [75]. A cell which contains all information related toa specific traffic area as well as a cell controller . The cell controller is a special purpose agent that continuously providesother agents located within its cell with a correct perception of their surrounding. Driver agents can be assigned a virtuallevel of distraction, causing unpredicted traffic behaviour. Although not specifically mentioned, the internal agentstructure resembles a mental-level model (see [76]). There is an interaction module that deals with agent perception andcommunication as well as a knowledge module to store agent knowledge. Furthermore, there is a separate planning andcontrol module that determines agent actions as wells as a tasks module for managing the agents tasks. Agent learningis not a particular focus of this project.
AgentPolis is a fully agent-based software framework for the implementation and simulation of transport systemswritten in Java [77, 5]. The framework provides a set of abstractions, code libraries and software tools for buildingsimulation models. The AgentPolis framework focuses particularly on the simulation of interaction-rich transportsystems. In 2013, the creators noted that existing simulation approaches fail to implement the ability to model ad hocinteractions among the entities of the transport system as well as the just-in-time decision-making that is required forthis form of interaction. However, current mobility services such as real-time on-demand transport (e.g ridesharing) relyon frequent, ad hoc interactions between various entities of the transport system. The creators observed that generaltoolkits, explicitly mentioning MATSim and SUMO, are only rarely used for the modelling of interaction-rich transportsystems, even though there are a lot of modelling concepts (e.g. road network, vehicles, traffic flow) that can be sharedwhen modelling different mobility services such as collective taxis, car-sharing etc. Instead, reference is given to workin which model-specific simulation tools have been developed from scratch (see [78, 79, 80]). Based on this, the creatorsconclude that similarities between such simulation models have not been exploited sufficiently due to existing toolsnot taking into account the multi-agent nature of interaction-rich transport systems. For example, MATSim treats theindividuals as passive data structures whose state can only be updated synchronously by central modules at infrequentpredefined points in time, whereas in reality agents in transport systems make just-in-time decisions asynchronously[5]. Leveraging the features implemented in AgentPolis, the simulator has been used for example as a testbed forbenchmarking on-demand mobility services [81].
Architecture
The AgentPolis framework was created in order to address the limitations mentioned above [5]. The framework isstructured in four main components. The first component is described as the modelling abstraction ontology . Thetheoretical concept of this component is to separate defined modelling abstractions from implementations of specificmodelling elements. It uses an ontology in order to define more general concepts of multi-agent systems that resultin a tailored structure for object-oriented programming when extending the simulation models for research-specificscenarios. This allows for enforcement of implementations that consider interoperability of existing and additionalresearch-specific modelling elements in their design. The second component is a library of implemented modellingelements based on the given abstractions specified in the ontology. The library contains a set of modelling elements thatrepresent common entities in transport systems. Apart from this, the third component can be described as the simulationengine. This component performs all calculations for running the simulation based on a discrete event model. Finally,the last component is a set of tools for user interaction, particularly for configuration and creation of the simulationmodel, data import, visualisation, etc.
Modelling Capabilities
In AgentPolis travellers are represented as agents. For demand modelling, AgentPolis includes a tool that generatesa large number of agents using statistical distributions based on census data [77]. The structure of the simulationmodel is given by the abstraction ontology (see Figure 8). This ontology defines concepts for
Activities, Actions,Sensors, Environment Objects, Queries and
Reasoning Modules . As described in [5], activities specify agent behaviourfor initiating agent actions. Agent actions model the effects of the agent on its environment. An example is giventhat a
DriveVehicle activity may result in a
MoveVehicle action. Moreover, physical entities placed in the agentssurrounding such as traffic lights are described as environment objects. Each agent has a set of sensors in order toperceive information on the current status of the other agents and its environment. For this purpose, the agent has to10 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021
Agent
Reasoning ModuleActivity ActivitySensor Sensor
Agent
Reasoning ModuleActivity ActivitySensor Sensor
Environment
Action
Environment
Object Environment Object Environment ObjectQuery
Figure 8: AgentPolis - Simplified Architecture.actively perform queries to obtain information. Agent behaviour is primarily determined by the reasoning module.In this module, complex algorithms can be implemented. For this, the AgentPolis framework only comes with theimplementation of
JourneyPlanner at the simulation core. This reasoning module models multimodal route planningbased on a time-dependent graph. Detailed description of the applied method is given in [82]. Apart from the reasoningmodule provided, different routing algorithms such as A* have been implemented and used for reasoning in otherexperiments [77]. The reasoning modules are reused for execution of different activities which leads to a connectedmodelling of driver decision-making and vehicle control. [64] have argued that at this stage a clear separation of theseaspects is possible. Referring to the aspect of agent learning it can be noted that this has not been of central focus inAgentPolis. [15] have argued that transportation research has focused on different aspects of the transportation system only in anisolated manner. Such aspects include travel demand, traffic flows, emissions, etc. However, simulation of complexsystems requires a combined method. Early attempts to integrate the isolated models into a unified system have shownthat resulting solutions are either inflexible, non-modular or inefficient. Based on this, the need for a unified solutionthat enables inter-operability between the isolated models has been established. As a response to this, the POLARISframework has been proposed [15]. POLARIS is an open-source agent-based modelling software development kit thatintegrates dynamic simulation of travel demand, network supply and network operations. The framework is written inC++ and combines different traffic-related modelling aspects that otherwise require a number of separate standalonesoftware applications. The approach differs from early integration attempts by implementing the various aspectsas dedicated agent capabilities within a general agent-based framework. POLARIS has been used for example, toanalyse vehicle energy consumption in the city of Detroit comparing scenarios that include current and future vehicletechnologies [83].
Architecture
The framework consists of two major conceptual components. The first component deals with agent-based modellingspecifically for the traffic domain. The second component provides a software development kit (SDK) which simplifiesthe development, execution and review of a developed simulation model. The internal architecture was designed using alayered model. Aspect-specific subcomponents are assigned to a layer depending on the level of modelling detail. Thisensures rather abstract concepts which are commonly used across possible variations of traffic simulation models to beless likely to change. Instead, users are supposed to make research-specific customisations on a more detailed level.This creates reusability of frequently used modelling aspects. Based on this assumption, layer 0 is the most abstractlayer of the POLARIS framework which contains a set of core libraries such as the discrete event engine (see Figure 9).The discrete event engine provides an API for handling the agent life-cycle. Simulations are performed by executing a11 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021list of events. In layer 1, POLARIS contains a set of fundamental extensions. This includes components for 2D/3Dvisualisation (Antares) or data import/export services. Layer 2 is described as an open-source versioned repository . Inthis repository, there is a set of model fragments that can be used for the implementation of custom models. The modelfragments included are tested and chosen by universal applicability. Typical model fragments for example are referenceimplementations of well-established routing algorithms. Finally, layer 3 is described as the user playground . In thislayer, custom components can be included in order to extend the POLARIS framework with research-specific modellingaspects. Based on the provided elements from all layers, the user can build his custom application for agent-basedtraffic simulation.
Core Libraries
POLARIS Meta Structures Memory
Allocator
Discrete Event
Engine
InterprocessEngine Custom Data Containers
Fundamental Extensions
Antares Scenario Manager Data Interchange Tools
Open-source Versioned Repository
Resusable Tranportation Prototypes Modular Transportation Algorithms Fundamental Transportation Data Layouts
Open-source Playground
Experimental Transportation
Algorithms
Specialised Transporation
Data Layouts
Final User Application Layer 0Layer 1
Layer 2
Layer 3
Figure 9: POLARIS - Architecture.
Modelling Capabilities
The POLARIS model fully incorporates the agent paradigm and uses agents to represent various modelling aspects. Atthe center of the model is the person agent that represents an individual traveller. These person agents operate withina defined environment whose entities are modelled as network agents. Furthermore, the POLARIS model includes a traffic management center (TMC) agent for monitoring the network agents and controlling the ITS system. Demandmodelling in POLARIS is based on an adjusted version of the ADAPTS (Agent-based Dynamic Activity Planning andTravel Scheduling) model, which is an activity-based approach for activity planning and scheduling [84]. Originally, theADAPTS model has been designed as a standalone application for simulating the occurrence of travel demand patternsthat result from travel planning and scheduling processes. For integration into the POLARIS framework, the ADAPTSmodel has been reorganised in order to match the agent paradigm. This resulted in a separate activity planning agentwhich in addition to the person agent models the traveller’s cognition of the activity planning process. For other aspectsof agent behaviour, the person agent is composed of a set of subagents. These include agents for the perception of thetraveller, movement handling and routing. The combination enables the modelled travellers to be able to generate, planand reschedule activities as well as choosing a route using established routing algorithms. The POLARIS model alsoconsiders both, prejourney and en-route information and (re-)planning. For this purpose, a bounded rationality en-routeswitching model is used [85]. The POLARIS framework as such does not come with dedicated components for agentlearning but can be extended by custom implementations.
Agent-based Traffic Simulation System (ATSim) is a simulator developed at TU Clausthal [20]. The creators arguethat for the modelling of the latest advances in the transport domain such as autonomous vehicles, an agent-basedapproach can be crucial to represent aspects such as communication, goals and plans. However, existing agent-based12 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021simulators have not focused on an intuitive graphical user interface and exhibit a lack of tools for data collection anddata analysis. This is why in the ATSim approach, the commercial simulator AIMSUN [28] has been integrated withthe JADE platform [63] in order to extend the simulator by agent modelling capabilities. This allows the reusabilityof all features already implemented in AIMSUN. In this approach, AIMSUN is used for modelling and simulation oftraffic scenarios while implementation of multi-agent behaviour is realised in JADE. ATSim has been used for exampleto simulate group-oriented traffic coordination in which groups of agents coordinate their speed and lane choices [86].
Architecture
The ATSim architecture is structured in four components (see Figure 10). The first component is the commercialAIMSUM simulator with all its features for modelling and simulating traffic scenarios. The second component is themulti-agent system based on JADE. This component is responsible for managing and controlling the agent life-cycle. InATSim, agents represent various kinds of traffic objects in AIMSUN in order to extend AIMSun objects with agentcapabilities. Communication between agents and traffic objects is possible based on the AIMSUM API. AIMSUMprovides an application programming interface (API) for the integration of external services in Python and C++.However, JADE is based on Java and it is therefore necessary for ATSim to make use of a middleware in order to allowcommunication between AIMSUM and JADE.
AIMSUN AIMSUN API Middleware MAS JADE
ATSim
Figure 10: ATSim - Architecture.
Modelling Capabilities
Basically, AIMSUN is a microscopic simulator based on established car following and lane change models. Thesemodels have been extended by agent capabilities such as perception and interaction for the individual traveller. Theoverall traffic model distinguishes three types of objects: static objects, objects with dynamic states and mobile objects .For example, the road network is represented as a static object whereas traffic lights are modelled as objects withdynamic states. Moreover, vehicles are presented as mobile objects. Traffic objects can be assigned to an agent in JADE.Each agent can only control a single object in AIMSUM. The link between the agents and traffic objects is based ontwo assumptions. First, the agent life-cycle is synchronised with the life-cycle of the associated traffic object. Second,agents constantly receive updated information from the assigned traffic object after each simulation step. In AIMSUMtraffic demand is modelled using classical origin-destination matrices. Regarding the aspects of agent decision-making,ATSim implements decentralised agent behaviour. ATSim also implements en route replanning. It is possible to modelmultimodal transportation in ATSim. Agents in ATSim are based on the INTERRAP Model [87] with abilities forperception, action, interaction and existing as sole processes in the system. However, agent learning has not been acentral focus of this project.
SimMobility is an agent-based multi-scale simulation platform that integrates a set of aspect-specific models relevant tothe transportation domain [19]. Such aspect-specific models include land-use, demographic movement, transportationand communication interactions, etc. The platform focuses on modelling the effects on the traffic infrastructure,intelligent transportation services and environment. This allows for the simulation of alternative planning optionsspecifically with regard to technology, policies and investment. SimMobility has been used for example to simulate theeffects of different fleet sizes for on-demand autonomous mobility [88] or to assess disruption management strategies inrail transit [89]. Furthermore, it can be noted that SimMobility is completely written in C++ and emphasises a modular,parallel and distributed system architecture.
Architecture
The system architecture of SimMobility is described as a multi-level approach. SimMobility consists of three submodulesthat each simulate a different perspective of the traffic system (Figure 11). The first submodule is the
Long-term (LT) simulator. This component deals with generation and updating the population of agents. The LT simulator particularlymodels long term aspects such as house location, job location, land development and car ownership. Job changes as wellas household relocations are simulated over longer time periods (months/years) in order to achieve realistic demographic13 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021movements. Apart from this, other long-term effects such as environmental consequences can be examined in thismodule. The second submodule is described as the
Mid-term (MT) simulator [90]. This module is primarily designedfor the simulation of agent behaviour in time scales of minutes and hours. In general, SimMobility distinguishestwo types of agent behaviour. The first type refers to high level travel decisions such as route choice, transportationmode, departure times, etc. Only this type of decisions is simulated in the MT simulator. In contrast, low level traveldecisions consider movement aspects such as car following and lane changing. This type of decisions is simulated inthe
Short-term simulator which is the last component to complete SimMobility’s multi-level architecture [91]. The STsimulator is a microsimulator based on MITSIM which has been extended by agent capabilities. A special characteristicof this architecture is that each module can be used as a standalone application. Furthermore, all simulators share thesame database so that simulated individuals exist across all simulation levels simultaneously.
Long-Term Simulator household relocation, job market, land development, car ownership, etc.
Mid-Term Simulator depature times, route choice, transportation mode, etc.
Short-Term Simulator lane changing, braking, accelerating, gap acceptance, etc. DB Figure 11: SimMobility - Architecture.
Modelling Capabilities
In SimMobility, agents are used to represent all sorts of entities and communication in the system such as travellers,vehicles, phones, traffic lights, etc. However, modelling aspects are distributed across the three submodules. Referringto the aspect of demand modelling, SimMobility implements an activity-based approach in the MT simulator [19].For each simulated day, the MT simulator generates an activity schedule that includes destination, departure time,route and mode choice. The resulting output is an activity schedule that contains a set of trip chains. This approachhas been integrated with methods of dynamic traffic assignment as demand calibration is usually performed usingorigin-destination matrices. For this purpose, generated trip chain information is aggregated to create origin-destinationmatrices. This demand information is then passed on to the ST simulator. In the ST simulator agent behaviour ismodelled in more detail. It distinguishes pre-day and within within-day decisions. Agents are able to spontaneouslyperform rerouting as well as rescheduling based on the agents perception. It is described that route choices are based ona probabilistic model that captures the impact of travel times and biases toward routes that use freeways over urbanstreets [91]. The ST simulator also indicates a mechanism that enables day-to-day agent learning. Information on theperformance of the previous day is given in order to update the agent’s knowledge. For this purpose, the MT simulatorprocesses day activities of each agent and calculates a figure that represents maximum expected utility of activity-travelpatterns at given conditions [19]. In correspondence with the authors, it was described that learning applies specificallyto the calculation of travel times. The daily learning process is described as an aggregated learning filter, in whichtravel times are updated using a filter with exponential flattening (see [92]). The travel times of the current iteration aredetermined by a combination of knowledge of the previous day and the present situation in the road network.
Apart from the above mentioned systems, specifically designed for the simulation of traffic scenarios, there is a varietyof examples of general purpose agent platforms that have been applied to traffic simulation such as NetLogo [93],Madkit [94], JANUS [95], GAMA [96], Repast Simphony [97], SWARM [98], MASON [99], etc. In these works, thegeneral purpose agent platforms were customised for the analysis of specific subjects [100, 101, 102, 103, 104]. Forexample, in [103] a traffic simulation plugin for the GAMA platform has been developed. This plugin aims at providinga less complex solution for non-computer scientists in comparison to established tools such as MATSim and ITSUMO.According to the author, the established tools often require complex customisation for the analysis of more specificapplication contexts such as car accidents. In another example, [102] investigated different aspects of carpooling on thebasis of an agent-based model. In this project the JANUS Platform was used to implement the simulation.14 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021
In this section, we categorise the simulation tools presented by area of application and discuss benefits and drawbacks aswell as relevant gaps where we identified missing functionality that we consider relevant for effective traffic simulations.Assignment of the simulators to a specific category has been determined based on their primary focus. Thus, simulatorsassigned to a category do not have to be used exclusively for this area of application, but are particularly suitable. Basedon the previous literature research conducted for the detailed analysis of the simulators, five major areas of applicationhave been identified.1.
General Focus Toolsets : TRANSIMS, MATSim, POLARIS2.
Traffic Control and ITS : SimMobility, ITSUMO, SUMO + JADE3.
Interaction Rich Services : ATSim, AgentPolis4.
Traffic Safety : MATISSE5.
Theoretic Flow Dynamics : MovSimThe first category of tools can be referred to as the
General Focus Toolsets . As the name suggests, this type of simulatorsincludes all-rounder applications with a general focus. Applications covered by this category include
TRANSIMS, MAT-Sim and
POLARIS . The idea is to provide an environment for other researchers to investigate aspect-specific researchquestions about traffic without having to bother with modelling of basic and recurring traffic structures. For example,important models for route choice and resource consumption have been implemented. While this type of application isvery powerful and universally applicable, it is argued that customisation can only be achieved with considerable effortand a profound understanding of the application architecture. In particular, when crucial aspects for specific researchquestions are missing, researchers have decided to develop their own simulation software, as familiarisation andmodifications to the available platforms are overly complex and time-consuming. As a result, [5] describe that often alot of shared potential remains unused and discusses examples in which individual solutions have been developed instead.The second category includes simulators with a primary focus on the simulation of
Traffic Control and ITS . This coversapplications such as
SimMobility, ITSUMO and
SUMO + JADE . These applications provide dedicated design conceptsfor the implementation of traffic control. In addition to moving travellers or vehicles, traffic objects such as traffic signsor traffic lights are also implemented as intelligent agents that can be controlled interactively. In particular, specificinterfaces are provided in order to modify the behaviour of these traffic objects more easily. Therefore the applicationsare well-suited for the integration and testing of custom ITS strategies such as new traffic light algorithms.Furthermore, another major category of simulators can be described with a focus on
Interaction Rich Services .Applications such as
ATSim or AgentPolis can be assigned to this type of simulators. Particular attention has been givento exploit the communication capabilities of the agent paradigm in order to represent the dynamics and interactionbetween traffic participants. This is particularly relevant for the simulation of modern mobility services that emergefrom car-sharing and ride-sharing, or autonomous vehicles. An important issue in traffic which is frequently addressedinvolves
Traffic Safety . There are numerous simulations on this topic, engaging with different aspects of traffic safety,such as the risk potential at different types of intersections or unexpected evacuation scenarios. While generalist toolssuch as MATSim are often used for route-choice simulations of evacuation scenarios, there is a specific tool that shouldbe outlined for the simulation of safety scenarios. The
MATISSE simulator specifically addresses this topic in depth andsupports the modelling of more detailed aspects such as driver distraction or intersection traffic.Finally, the last type of application with the most significant distinction to the other categories applies to the MovSimsimulator. Instead of simulating traffic scenarios on specific road networks (e.g. cities), MovSim focuses on thesimulation of theoretic flow dynamics . In practical application, MovSim is best suited to simulate the emergence of flowphenomena such as phantom traffic jams. This is particularly helpful in the field of academic teaching in order tovisually demonstrate these theoretical effects.Throughout all categories of simulators, we have identified two modelling aspects which so far have only been dealtwith marginally. We argue that for the simulation of modern traffic questions further research in these areas is requiredto achieve even more realistic and effective simulations. The first aspect addresses what can be referred to as dynamicdecision-making . In scenarios, travellers are continuously provided with the latest information for example, due toavailable mobile devices. As a result, travellers often act within moments to determine whether they deviate from theiroriginal travel planning. For example, in the event of unexpected delays in public transport, travellers may change15 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021connections to still be able to arrive on time. We have observed that only a few simulators have so far addressed theissue. ITSUMO for example implemented en route replanning for both centralised and decentralised routing [11]. Thiscan be considered a first step towards modelling dynamic decision-making processes. However, traffic participants areinfluenced by a large number of factors at any moment during the journey that lead to choices not only concerning thetravel route but also transportation mode and many others. For example, when weather conditions change and it startsraining, a pedestrian might search for the nearest bus stop and continue their remaining travel route using the bus.These dynamic decision-making processes become particularly relevant for research on the deployment of new mobilityservices such as autonomous vehicles or e-scooters. Hence, further work should be done in this area.We also observed that only a few simulators addressed the issue of agent learning in traffic simulation. However, this isan important process that has a significant impact on real traveller behaviour. In particular, learning behaviour becomesessential for simulations that are modelled over longer periods of time and that deal with the consequences of measuresin traffic planning. In this case, SimMobility was one of the leading applications, which includes a custom day-to-daylearning module to update the agents’ knowledge. ITSUMO had a slightly different focus and applied reinforcementlearning to traffic control agents for improving traffic light control of the overall system. Nevertheless, we believe that alot of potential is left unused and that further work is required on this matter. While research questions often focus onfinding better global equilibria (social learning), in real-world problems individuals can be self-interested which maylead to a suboptimal global outcome. This phenomenon has been demonstrated in game theory (e.g. see the Prisoner’sDilemma [105]). In particular, in repeated games, decisions from previous rounds have significant influence on newdecisions. Commuter scenarios or other recurring forms of daily mobility are essentially repeated games of this form.In order to appropriately capture these effects for the creation of realistic simulation models, capabilities for individuallearning by self-interested agents cannot be neglected.
In order to cope with the changing requirements in inner-city traffic, computer-based simulations can be used to plan andassess the effects of new policies. There is a variety of agent-based simulation systems that range from general purposeagent platforms (e.g. NetLogo, GAMA, Repast Simphony) to systems specifically designed for traffic simulation(MatSim, ITSUMO, AgentPolis etc.). As early ideas on traffic simulation have already been published in the 1970s[2, 3], the number of existing simulation tools is significant with each of the tools focusing on different aspects of thetransport system and differing in the underlying methods. Models based on cellular automata and multi-agent systemshave dominated the field of available traffic simulators. As state of the art research shows a growing interest in theapplication of multi-agent models, we have primarily focused our survey on simulators that adopt this type model. Withthe aim to cover a broad spectrum of applications, we also included simulators that have been of historical significanceas they provide important foundation work. Based on this, a selection of 10 representative applications has beenexamined in detail with regard to their background, architecture and modelling capabilities. It can be noted that with theexception of ITSUMO, all simulators are still under active contribution. Following this, we have identified 5 categoriesthat express the diversity in the focus of the applications. Table 1 provides a summary on the properties of the reviewedapplications to help end-users make an informed assessment of the suitability of the applications for their specificresearch questions. More particularly, the table shows level of detail of the underlying simulation model (microscopic,macroscopic, mesoscopic or nanoscopic). The table also gives information about licensing of the applications which isessential for end-user selection. Regarding the aspect demand modelling, it has been shown that three methods havebeen applied: origin-destination matrices (OD), activity-based and location-specific probabilities (LSP). Simulators thatmake use of OD matrices include both approaches based on static values as well as statistical distributions. In the LSPapproach travellers usually move based on established car-following models with no route specification. Instead, eachlocation is assigned a pair of probabilities for the number of travellers starting as well as stopping at the location. Forexample, location at A 20% of the trips originate and 10% of them are collected, while at the same time at location B5% of the trips are starting and 20% are collected. Apart from this, information on the methods applied to the areaof agent decision-making is given. For this, one aspect included refers to route choice (RC) which can be central ordecentral. Another aspect refers to the extent to which en route replanning has been implemented. In addition to this,special features that also fall under the same category are outlined in Table 1.Finally, in this paper we have discussed possible limitations and gaps of the applications reviewed in this paper. We haveidentified two modelling aspects for which we argue that further research in these areas is required. While agent modelsincreasingly consider traffic as the collective result of a large number of individuals, available models fail to fullycapture the essence of individual behaviour. Although the role of the individual has been acknowledged, aspects such asagent learning or dynamic decision processes have so far been only been dealt with marginally. As a consequence a lotof potential is left unused in order to achieve even more realistic simulation models.16 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021Table 1: A summary of reviewed simulation applications
Application Model Licensing Demand Modelling Decision-Making Agent Learning
TRANSIMS Microscopic Open Source Activity based RC based on statistical distributions (central) Out of focusMATSim Microscopic Open Source Activity based RC based on discrete-choice-model Game theoretic Learning (systemmoving to stable equilibrium)SUMO + JADE Microscopic Open Source Activity basedor OD matrices RC based on routing algorithms (central) Out of focusITSUMO Microscopic Open Source OD matricesor LSP RC based on routing algorithms (central & decentral)en route replanning (route choice) Reinforcement learning fortraffic control agentsMovSim Microscopic Open Source LSP RC based on discrete-choice-model Out of focusMATISSE Microscopic Open Source LSP RC based on routing algorithms (decentral)en route replanning (route choice)Special Feature: Parameter for driver distraction Out of focusAgentPolis Mesoscopic Open Source OD matrices RC based on routing algorithms (central) Out of focusPOLARIS Mesoscopic Open Source Activity based RC based on routing algorithms (central & decentral)en route replanning (route choice) Out of focusATSim Microscopic Commercial OD matrices RC based on routing algorithms (decentral)en route replanning (route choice) Out of focusSimMobility Mesoscopic Open Source Activity based RC based on statistical distributions (central)en route replanning (route choice & activities) Day-to-day traveller learning
This research has been supported by a grant from the Karl-Vossloh-Stiftung (Project Number S0047/10053/2019).
References [1] L.S. Passos, R.J.F. Rossetti, and Z. Kokkinogenis. Towards the next-generation traffic simulation tools: a firstappraisal. In , pages 1–6. IEEE,2011.[2] M. Poeck and D. Zumkeller. Die anwendung einer massnahmenempfindlichen prognosemethode am beispiel desgrossraums nürnberg. In
DVWG-Workshop Policy Sensitive Models, Giessen , 1976.[3] K.W. Axhausen and R. Herz. Simulating activity chains: German approach.
Journal of TransportationEngineering , 115(3):316–325, may 1989.[4] K.C. Clarke. Cellular automata and agent-based models. In
Handbook of Regional Science , pages 1–16. SpringerBerlin Heidelberg, 2018.[5] M. Jakob and Z. Moler. Modular framework for simulation modelling of interaction-rich transport systems. In , pages 2152–2159. IEEE,IEEE, oct 2013.[6] S. Algers, E. Bernauer, M. Boero, L. Breheret, C. Di Taranto, M. Dougherty, K. Fox, and J.F. Gabard. Review ofmicro-simulation models.
Review Report of the SMARTEST project , 1997.[7] M. Pursula. Simulation of traffic systems-an overview.
Journal of geographic information and decision analysis ,3(1):1–8, 1999.[8] P.A. Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y. Flötteröd, R. Hilbrich, L. Lücken, J. Rummel,P. Wagner, and E. Wießner. Microscopic traffic simulation using SUMO. In , pages 2575–2582. IEEE, IEEE, nov 2018.[9] M. Haklay and P. Weber. OpenStreetMap: User-generated street maps.
IEEE Pervasive Computing , 7(4):12–18,oct 2008.[10] A. Horni, K. Nagel, and K.W. Axhausen.
The Multi-Agent Transport Simulation MATSim . Ubiquity Press, aug2016.[11] A.L.C. Bazzan, M.d.B. do Amarante, T. Sommer, and A.J. Benavides. Itsumo: an agent-based simulator for itsapplications. In
Proc. of the 4th Workshop on Artificial Transportation Systems and Simulation. IEEE , page 8,2010.[12] M. Treiber and A. Kesting. An open-source microscopic traffic simulator.
IEEE Intelligent TransportationSystems Magazine , 2(3):6–13, 2010.[13] M. Guériau, R. Billot, N.E. El Faouzi, J. Monteil, F. Armetta, and S. Hassas. How to assess the benefitsof connected vehicles? a simulation framework for the design of cooperative traffic management strategies.
Transportation Research Part C: Emerging Technologies , 67:266–279, jun 2016.17 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021[14] B. Torabi, M. Al-Zinati, and R.Z. Wenkstern. MATISSE 3.0: A large-scale multi-agent simulation system forintelligent transportation systems. In
Advances in Practical Applications of Agents, Multi-Agent Systems, andComplexity: The PAAMS Collection , pages 357–360. Springer International Publishing, 2018.[15] J. Auld, M. Hope, H. Ley, V. Sokolov, B. Xu, and K. Zhang. POLARIS: Agent-based modeling framework devel-opment and implementation for integrated travel demand and network and operations simulations.
TransportationResearch Part C: Emerging Technologies , 64:101–116, mar 2016.[16] P. Waddell, H. Ševcıková, D. Socha, E. Miller, and K. Nagel. Opus: An open platform for urban simulation. In
Computers in urban planning and urban management conference, London , 2005.[17] G. Czura, P. Taillandier, P. Tranouez, and É. Daudé. Mosaiic: City-level agent-based traffic simulation adaptedto emergency situations. In
Proceedings of the International Conference on Social Modeling and Simulation,plus Econophysics Colloquium 2014 , pages 265–274. Springer International Publishing, 2015.[18] J. Weyl, D. Glake, and T. Clemen. Agent-based traffic simulation at city scale with mars. In
Proceedings of theAgent-Directed Simulation Symposium , ADS ’18, San Diego, CA, USA, 2018. Society for Computer SimulationInternational.[19] M. Adnan, F.C. Pereira, C.M.L. Azevedo, K. Basak, M. Lovric, S. Raveau, Y. Zhu, J. Ferreira, C. Zegras, andM. Ben-Akiva. Simmobility: A multi-scale integrated agent-based simulation platform. In , 2016.[20] V.H. Chu, J. Görmer, and J.P. Müller. Atsim: Combining aimsum and jade for agent-based traffic simulation. In
Proceedings of the 14th Conference of the Spanish Association for Artificial Intelligence (CAEPIA) , volume 1,2011.[21] S. Thulasidasan, S. Kasiviswanathan, S. Eidenbenz, E. Galli, S. Mniszewski, and P. Romero. Designing systemsfor large-scale, discrete-event simulations: Experiences with the fasttrans parallel microsimulator. In , pages 428–437. IEEE, IEEE, dec 2009.[22] Texas Transportation Institute. Early deployment of transims: Issue paper, 1999.[23] D. Krajzewicz, G. Hertkorn, C. Rössel, and P. Wagner. Sumo (simulation of urban mobility)-an open-sourcetraffic simulation. In
Proceedings of the 4th middle East Symposium on Simulation and Modelling (MESM20002) ,pages 183–187, 2002.[24] M. Miska, E. Santos, E. Chung, and H. Prendinger. Opentraffic-an open source platform for traffic simulation. In
Australasian Transport Research Forum , 2011.[25] G. Tamminga, Peter K., and J.W.C. Van Lint. Open traffic: A toolbox for traffic research.
Procedia ComputerScience , 32:788–795, 2014.[26] N.B. Taylor. The contram dynamic traffic assignment model.
Networks and spatial economics , 3(3):297–322,2003.[27] M. Fellendorf. Vissim: A microscopic simulation tool to evaluate actuated signal control including bus priority.In , volume 32, pages 1–9. Springer, 1994.[28] J. Barceló and J. Casas. Dynamic network simulation with aimsun. In
Simulation approaches in transportationanalysis , pages 57–98. Springer, 2005.[29] G.D.B. Cameron and G.I.D. Duncan. PARAMICS?parallel microscopic simulation of road traffic.
The Journalof Supercomputing , 10(1):25–53, 1996.[30] Q. Yang, H.N. Koutsopoulos, and M.E. Ben-Akiva. Simulation laboratory for evaluating dynamic trafficmanagement systems.
Transportation Research Record , 1710(1):122–130, jan 2000.[31] J. Miller and E. Horowitz. FreeSim - a free real-time freeway traffic simulator. In , pages 18–23. IEEE, IEEE, sep 2007.[32] L.E. Owen, Y. Zhang, L. Rao, and G. McHale. Traffic flow simulation using corsim. In , volume 2, pages 1143–1147. IEEE, 2000.[33] J. Lei, K. Redmill, and U. Ozguner. Vatsim: A simulator for vehicles and traffic. In
ITSC 2001. 2001 IEEEIntelligent Transportation Systems. Proceedings (Cat. No. 01TH8585) , pages 686–691. IEEE, 2001.[34] R. Liu. Traffic simulation with DRACULA. In
Fundamentals of Traffic Simulation , pages 295–322. SpringerNew York, 2010.[35] Y. Wang, M. Papageorgiou, and A. Messmer. RENAISSANCE – a unified macroscopic model-based approachto real-time freeway network traffic surveillance.
Transportation Research Part C: Emerging Technologies ,14(3):190–212, jun 2006. 18 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021[36] D.K. Sorenson and J. Collins. Practical applications of traffic simulation using simtraffic. In
Compendiumof Papers. Institute of Transportation Engineers 2000, District 6 Annual Meeting, Institute of TransportationEngineers (ITE) , 2000.[37] M. Balmer, K. Meister, and K. Nagel.
Agent-based simulation of travel demand: Structure and computational per-formance of MATSim-T . ETH, Eidgenössische Technische Hochschule Zürich, IVT Institut für Verkehrsplanungund Transportsysteme, 2008.[38] L. Smith, R. Beckman, and K. Baggerly. TRANSIMS: Transportation analysis and simulation system. Technicalreport, Los Alamos National Lab., NM (United States), jul 1995.[39] S. Lawe, J. Lobb, A.W. Sadek, S. Huang, and C. Xie. TRANSIMS implementation in chittenden county, vermont.
Transportation Research Record: Journal of the Transportation Research Board , 2132(1):113–121, jan 2009.[40] K.S. Lee, J.K. Eom, and D.S. Moon. Applications of TRANSIMS in transportation: A literature review.
ProcediaComputer Science , 32:769–773, 2014.[41] L. Pasupuleti, O. Ghrayeb, C. Mirman, H. Ley, and Y. Park. Disaster planning for a large metropolitan city usingtransims software. age , 14:1, 2009.[42] J.A. Kerenyi et al. Moreno valley’s transims traffic model: Microsimulating the mpo’s model. In
ManagingOperational Performance... Exceeding Expectations. 2012 ITE Technical Conference and ExhibitInstitute ofTransportation Engineers (ITE) Parsons BrinckerhoffTraficonClary, The Continuous Power CompanyFLIRE-conolite, USASiemensSam Schwartz Engineering , 2012.[43] L.R. Rilett and J. Zietsman. An overview of the transims micro-simulation model: application possiblities forsouth africa.
SATC 2001 , 2001.[44] M.D. Williams, G. Thayer, and L. Smith. Development of the transims environmental model, 1997.[45] R.J. Beckman and M.D. Baggerly, K.A.and McKay. Creating synthetic baseline populations.
TransportationResearch Part A: Policy and Practice , 30(6):415–429, nov 1996.[46] K. Nagel, P. Stretz, M. Pieck, R. Donnelly, and C.L. Barrett. Transims traffic flow characteristics. arXiv preprintadap-org/9710003 , 1997.[47] S. Washington.
Incorporating safety into long-range transportation planning , volume 546. TransportationResearch Board, 2006.[48] C. Dobler and K.W. Axhausen. Design and implementation of a parallel queue-based traffic flow simulation.
Arbeitsberichte Verkehrs-und Raumplanung , 732, 2011.[49] D. Charypar.
Efficient algorithms for the microsimulation of travel behavior in very large scenarios . PhD thesis,ETH, Eidgenössische Technische Hochschule Zürich, 2008.[50] T. Novosel, L. Perkovi´c, M. Ban, H. Keko, T. Pukšec, G. Krajaˇci´c, and N. Dui´c. Agent based modelling andenergy planning – utilization of MATSim for transport energy demand modelling.
Energy , 92:466–475, dec2015.[51] S. Hörl. Implementation of an autonomous taxi service in a multi-modal traffic simulation using matsim. Master’sthesis, CHALMERS UNIVERSITY OF TECHNOLOGY, 2016.[52] A. Horni and K.W. Axhausen. Matsim agent heterogeneity and a one-week scenario.
Arbeitsberichte Verkehrs-und Raumplanung , 836, 2012.[53] G. Flötteröd and B. Kickhöfer. Choice models in MATSim. In
The Multi-Agent Transport Simulation MATSim ,pages 337–346. Ubiquity Press, aug 2016.[54] G. Flötteröd. MATSim as a monte-carlo engine. In
The Multi-Agent Transport Simulation MATSim , pages327–336. Ubiquity Press, aug 2016.[55] A. Kesting, M. Treiber, and D. Helbing. Agents for traffic simulation. In
Multi-Agent Systems , volume 11, pages325–356. CRC Press, jun 2009.[56] M. Treiber, R. Germ, and A. Kesting. From drivers to athletes: Modeling and simulating cross-country skiingmarathons.
Traffic and Granular Flow ’13 , pages 243–249, 2015.[57] M. Schönhof, M. Treiber, A. Kesting, and D. Helbing. Autonomous detection and anticipation of jam frontsfrom messages propagated by intervehicle communication.
Transportation Research Record: Journal of theTransportation Research Board , 1999(1):3–12, jan 2007.[58] A. Kesting, M. Treiber, M. Schönhof, and D. Helbing. Adaptive cruise control design for active congestionavoidance.
Transportation Research Part C: Emerging Technologies , 16(6):668–683, dec 2008.19 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021[59] A. Kesting, M. Treiber, and D. Helbing. Connectivity statistics of store-and-forward intervehicle communication.
IEEE Transactions on Intelligent Transportation Systems , 11(1):172–181, mar 2010.[60] M. Guériau, R. Billot, S. Hassas, F. Armetta, and N.E. El Faouzi. An extension of MovSim for multi-agentcooperative vehicles modeling. In ,pages 859–860. IEEE, IEEE, nov 2014.[61] M. Treiber and A. Kesting.
Traffic Flow Dynamics . Springer Berlin Heidelberg, 2013.[62] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz. Sumo–simulation of urban mobility: an overview. In
Proceedings of SIMUL 2011, The Third International Conference on Advances in System Simulation . ThinkMind,2011.[63] F. Bellifemine, F. Bergenti, G. Caire, and A. Poggi. Jade — a java agent development framework. In
Multi-AgentProgramming , pages 125–147. Springer US, 2005.[64] G. Soares, Z. Kokkinogenis, J.L. Macedo, and R.J.F. Rossetti. Agent-based traffic simulation using SUMO andJADE: An integrated platform for artificial transportation systems. In
Simulation of Urban Mobility , pages 44–61.Springer Berlin Heidelberg, 2014.[65] T. Azevedo, P.J.M. De Araújo, R.J.F. Rossetti, and A.P.C. Rocha. Jade, trasmapi and sumo: A tool-chain forsimulating traffic light control.
CoRR , abs/1601.08154, 2016.[66] I.J.P.M Timóteo, M.R. Araújo, R.J.F. Rossetti, and E.C. Oliveira. Using TraSMAPI for the assessment ofmulti-agent traffic management solutions.
Progress in Artificial Intelligence , 1(2):157–164, may 2012.[67] F.Y. Wang and S. Tang. A framework for artificial transportation systems: From computer simulations tocomputational experiments. In
Proceedings. 2005 IEEE Intelligent Transportation Systems, 2005. , pages1130–1134. IEEE, 2005.[68] J. Wahle, A.L.C. Bazzan, F. Klügl, and M. Schreckenberg. The impact of real-time information in a two-routescenario using agent-based simulation.
Transportation Research Part C: Emerging Technologies , 10(5-6):399–417, oct 2002.[69] B.C. Silva, A.L.C. Bazzan, G.K. Andriotti, F. Lopes, and D. Oliveira. ITSUMO: An intelligent transportationsystem for urban mobility. In
Innovative Internet Community Systems , pages 224–235. Springer Berlin Heidelberg,2006.[70] R.J.F. Rossetti and R. Liu.
Advances in artificial transportation systems and simulation . Academic Press, 2014.[71] A.L.C. Bazzan, M.d.B. do Amarante, G.G. Azzi, A.J. Benavides, L.S. Buriol, L.F.S. Moura, M. Ritt, andT. Sommer. Extending traffic simulation based on cellular automata: From ParticlesTo autonomous agents. In
ECMS 2011 Proceedings edited by: T. Burczynski, J. Kolodziej, A. Byrski, M. Carvalho , pages 91–97. ECMS,jun 2011.[72] A.L.C. Bazzan, K. Nagel, and F. Klügl. Integrating matsim and itsumo for daily replanning under congestion. In
Proceedings of the 35th Latin-American Informatics Conference, CLEI
Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems , pages 1454–1456.International Foundation for Autonomous Agents and Multiagent Systems, 2017.[75] F. Araujo, J. Valente, and R.Z. Wenkstern. Modeling agent-based traffic simulation properties in alloy. In
Proceedings of the 2012 Symposium on Agent Directed Simulation , pages 1–8, 2012.[76] J. McCarthy. Ascribing mental qualities to machines. Technical report, STANFORD UNIV CA DEPT OFCOMPUTER SCIENCE, 1979.[77] M. Jakob, Z. Moler, A. Komenda, Z. Yin, A.X. Jiang, M.P. Johnson, M. Pˇechouˇcek, and M. Tambe. Agentpolis:towards a platform for fully agent-based modeling of multi-modal transportation. In
Proceedings of the11th International Conference on Autonomous Agents and Multiagent Systems-Volume 3 , pages 1501–1502.International Foundation for Autonomous Agents and Multiagent Systems, 2012.[78] M.E.T. Horn. Multi-modal and demand-responsive passenger transport systems: a modelling framework withembedded control systems.
Transportation Research Part A: Policy and Practice , 36(2):167–188, feb 2002.[79] L. Quadrifoglio, M.M. Dessouky, and F. Ordóñez. A simulation study of demand responsive transit systemdesign.
Transportation Research Part A: Policy and Practice , 42(4):718–737, may 2008.20 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021[80] S.F. Cheng and T.D. Nguyen. TaxiSim: A multiagent simulation platform for evaluating taxi fleet operations. In
Advanced Agent Technology , volume 2, pages 359–360. Springer Berlin Heidelberg, 2012.[81] M. ˇCertick`y, M. Jakob, R. Píbil, and Z. Moler. Agent-based simulation testbed for on-demand mobility services.
Procedia Computer Science , 32:808–815, 2014.[82] J. Hrnˇcíˇr and M. Jakob. Generalised time-dependent graphs for fully multimodal journey planning. In , pages 2138–2145. IEEE,IEEE, oct 2013.[83] E. Islam, A. Moawad, J. Auld, D.A. Karbowski, and A. Rousseau. Impact of advanced vehicle technologies onenergy consumption for the city of detroit using transportation system simulations. In , pages 1–6. IEEE, IEEE, dec 2017.[84] J. Auld and A. Mohammadian. Framework for the development of the agent-based dynamic activity planningand travel scheduling (ADAPTS) model.
Transportation Letters , 1(3):245–255, jul 2009.[85] R. Jayakrishnan, H.S. Mahmassani, and T.Y. Hu. An evaluation tool for advanced traffic information andmanagement systems in urban networks.
Transportation Research Part C: Emerging Technologies , 2(3):129–147,sep 1994.[86] J. Görmer and J.P. Müller. Multiagent system architecture and method for group-oriented traffic coordination. In , pages 1–6. IEEE,IEEE, jun 2012.[87] J.P. Müller.
The design of intelligent agents: a layered approach , volume 1177. Springer Science & BusinessMedia, 1996.[88] K.A. Marczuk, H.S.S. Hong, C.M.L. Azevedo, M. Adnan, S.D. Pendleton, E. Frazzoli, et al. Autonomousmobility on demand in SimMobility: Case study of the central business district in singapore. In , pages 167–172. IEEE, IEEE, jul 2015.[89] M. Adnan, F.C. Pereira, C.L. Azevedo, K. Basak, K. Koh, H. Loganathan, Z.H. Peng, and M. Ben-Akiva.Evaluating disruption management strategies in rail transit using simmobility mid-term simulator: a study ofsingapore mrt north-east line. In ,2017.[90] Y. Lu, M. Adnan, K. Basak, F.C. Pereira, C. Carrion, Vahid H. S., H. Loganathan, and M.E. Ben-Akiva.Simmobility mid-term simulator: A state of the art integrated agent based demand and supply model. In , 2015.[91] C.L. Azevedo, N.M. Deshmukh, B. Marimuthu, S. Oh, K. Marczuk, H. Soh, K. Basak, T. Toledo, L.S. Peh,and M.E. Ben-Akiva. SimMobility short-term: An integrated microscopic mobility simulator.
TransportationResearch Record: Journal of the Transportation Research Board , 2622(1):13–23, jan 2017.[92] G.E. Cantarella and E. Cascetta. Dynamic processes and equilibrium in transportation networks: Towards aunifying theory.
Transportation Science , 29(4):305–329, nov 1995.[93] U. Wilensky et al. Center for connected learning and computer-based modeling. In
NetLogo . NorthwesternUniversity, 1999.[94] O. Gutknecht and J. Ferber. MadKit. In
Proceedings of the fourth international conference on Autonomousagents - AGENTS ‘00 , pages 78–79. ACM Press, 2000.[95] S. Galland, N. Gaud, S. Rodriguez, and V. Hilaire. Janus: Another yet general-purpose multiagent platform. In
Seventh AOSE Technical Forum, Paris , 2010.[96] E. Amouroux, T.Q. Chu, A. Boucher, and A. Drogoul. GAMA: An environment for implementing and runningspatially explicit multi-agent simulations. In
Agent Computing and Multi-Agent Systems , pages 359–371. SpringerBerlin Heidelberg, 2009.[97] M.J. North, N.T. Collier, J. Ozik, E.R. Tatara, C.M. Macal, M. Bragen, and P. Sydelko. Complex adaptivesystems modeling with repast simphony.
Complex Adaptive Systems Modeling , 1(1):3, mar 2013.[98] N. Minar, R. Burkhart, C. Langton, M. Askenazi, et al. The swarm simulation system: A toolkit for buildingmulti-agent simulations. Technical report, SANTA FE INSTITUTE, 1996.[99] S. Luke, C. Cioffi-Revilla, L. Panait, K. Sullivan, and G. Balan. MASON: A multiagent simulation environment.
SIMULATION , 81(7):517–527, jul 2005. 21 N O VERVIEW OF A GENT - BASED T RAFFIC S IMULATORS - F
EBRUARY
16, 2021[100] J. Ge and G.J. Polhill. Exploring the combined effect of factors influencing commuting patterns and CO2emissions in aberdeen using an agent-based model.
Journal of Artificial Societies and Social Simulation , 19(3),2016.[101] H.M. Kammoun, I. Kallel, J. Casillas, and A. Alimi. A road traffic multiagent simulation using turtlekit undermadkit. In
Proc. of the 9th International Conference on Artificial Intelligence and Soft Computing ICAISC,Academic Publishing House Exit , pages 503–514, 2008.[102] S. Galland, J. Buisson, N. Gaud, M. Gonçalves, A. Koukam, F. Guiot, and L. Henry. Agent-based simulation ofdrivers with the janus platform.
Procedia Computer Science , 32:738–743, 2014.[103] P. Taillandier. Traffic simulation with the gama platform. In
Eighth International Workshop on Agents in Trafficand Transportation , pages 8–p, 2014.[104] M. Zargayouna, B. Zeddini, G. Scemama, and A. Othman. Agent-based simulator for travelers multimodalmobility. In
KES-AMSTA , pages 81–90. Citeseer, 2013.[105] N. Lacey.