User Modelling Languages for AmI: A Case Study on Road Traffic Alberto Fernández-Isabel and Rubén Fuentes Fernández1 Abstract. Traffic is an important phenomenon in modern societies. The work presented in this paper pursues reducing this effort by Its complexity and the difficulties to control the actual settings where providing base models for people acting in traffic phenomena. These it happens have made of simulation a key tool for its study. This models are part of a wider effort to build a general framework for approach requires suitable models to capture all its relevant aspects traffic simulation based on ABM, so they are designed looking for and their mutual influences. Among these aspects, people are the key reusability with different theories and contexts. For this purpose, the one. However, there is a limited understanding of people attitudes basis of the models is a classification of people with three dimen- and behaviours and their effect on traffic. Thus, simulation has here sions: their role in traffic, traits, and current state. an important component of exploration of hypotheses. Our research People role in traffic depends on their mean of transport and their contributes to this line of work with a set of general and extensible relation with it People are classified in drivers, passengers of vehicles agent-based models about people in traffic. These models integrate and pedestrians. All of them can be modelled at the individual level existing research from Social Sciences and simulation. The agent or as groups moving together. paradigm supports the explicit specification of processes of infor- People traits represent features of people that are permanent for the mation management, decision-making, action execution, and interac- travels considered in the simulation. They include physical attributes tion both with people and the environment. Such approach facilitates of the body, such as age, gender or disabilities. There are also atti- model reuse, and linkage between different elements relevant for the tudes to capture personality and mental features. For instance, people studies. The paper illustrates the application of these models with a can be more or less aggressive and have personal problems that in- case study that shows how to integrate in them a well-known model crease their stress. Moreover, people get more traffic experience over for drivers attitudes. time. This empowers them with additional knowledge and skills to face traffic situations through their learning. The current state captures dynamic features that depend on the 1 INTRODUCTION specific context and moment. For instance, they indicate if drivers attention is low because there are distracting passengers, or if traffic Life in modern societies is highly mediated by traffic. Every day, conditions bother the driver. millions of persons move on foot and by private vehicles or public The suitability of these models is illustrated with a case study that transport. These flows are organized according to certain social rules, specifies existent work on traffic simulation using the proposed prim- but also depend on individual attitudes and behaviours, unexpected itives. It considers the simulation in [13] about drivers attitudes and events happening in the environment, and their mutual influences. their influence on group behaviour. Given the difficulties to carry out these studies with real settings, The rest of the paper is organized as follows. Section 2 intro- models have emerged as a key tool to study traffic. duces the agent modelling language used in our approach. Section There are several approaches to model traffic [16]. Analytical 3 presents the models for people in traffic with that language, and models rely on a strong abstraction of the individual components de- grounds them in available research about traffic. Section 4 compares scribed mainly with mathematical formulas [8]. They are useful to these models with those in [13] regarding the phenomenon called consider phenomena with large populations, but have limitations re- dominance at junctions. These results are further compared with re- garding the specification of procedural and non-linear behaviors, and lated work in Section 5. Finally, Section 6 discusses some conclu- heterogeneous populations. As an alternative, simulation facilitates sions and future work. the specification of these kinds of behavior and population, but it is not usually intuitive the correspondence between the actual sys- tem and its computational representation. Agent-Based Modelling 2 AGENT-BASED MODELLING LANGUAGE (ABM) [8] addresses this problem using agents as its core modelling primitive. Agents are intentional abstractions conceptualized in terms This work specifies their models using the modelling language of the of elements such as knowledge, goals or capabilities. They are able INGENIAS methodology [14]. The key concept of INGENIAS is the to interact with other agents and their environment. These features agent. facilitate describing people behaviour with agents. However, mak- An agent is an intentional entity that follows its own agenda char- ing realistic models still demands a high effort to integrate different acterized by goals. In order to achieve these goals, the agent is able theories and give the needed information. to carry out certain tasks. An agent can trigger a task when it pursues an unsatisfied goal that the task is potentially able to fulfil, and all 1 University Complutense of Madrid, Spain, email: afernandezisabel, the elements required by the task are available. These elements are ruben@fdi.ucm.es usually pieces of information known by the agent or events coming from the environment. As the result of the execution of the task, the The group of physical attributes currently comprehends gender agent acts on the environment and produces or modifies information. and age. The gender attribute classifies people into male or female. Agents act on and perceive the environment through external ap- The age attribute uses four levels, young, adult-young, mature and el- plications. These are the sources of the events and tasks use their der. The age levels are different for drivers and pedestrians, as people methods to affect the environment. can walk before they can drive and the required capabilities for both A final element relevant for the models presented in this paper is activities are different. These attributes mainly affect perception and the AInherits relationships. This is an inheritance relationship that al- reaction parameters, such as sight distance and time to maneuver. lows defining a type of agent as an extension or constraint of another Other group of traits is the attitudes. Models consider in it a traf- type of agent. Thus, it highlights the common features of different fic profile and relationship problems. The traffic profile is based on types of agent and saves modelling effort. an extension of the selfish principle in [13], classifying drivers as aggressive, normal or moderate, and pedestrians as reckless, normal 3 TAXONOMY OF PEOPLE IN TRAFFIC or prudent. This classification differentiates, for instance, between drivers who always drive below or at speed limits, or on the contrary Traffic is the organized way of moving people using different means usually break them. The relationship problems acknowledge this as a of transport. This people have as their main goal to arrive fast and classical source of anxiety and distractions in traffic situations, mak- safely to their destinations [7, 9]. They can achieve this goal through ing more likely suffering an accident or taking greater risks [18]. alternative sequences of actions as long as they meet some con- The last group of traits is the traffic experience. It classifies indi- straints. First, people use different means of transport, and can con- viduals regarding their traffic learning with values between 0 and 5, trol them or be a passenger. This makes suitable only some routes being 5 the maximum experience. and implies certain rules. Our work only considers those means shar- ing spaces in our cities and roads, e.g. on foot or by car. Second, the sequence also depends on the physical and mental characteristics of 3.3 CURRENT STATE OF A PERSON people and their current state. However, models cannot consider all The traffic and personal conditions change during travels, and this the known people features and processes. This would be unsuitable affects people behaviour. For instance, drivers caught in a traffic jam regarding efficiency and abstraction, and even incorrect given our can start relaxed, but their frustration and impatience will rise as limited understanding of the phenomena. The taxonomy presented they waste more time stuck, which can cause risky situations in their proposes a number of features based on literature, mainly coming nearby environment. The models consider these dynamic features of from Social Sciences, widely accepted as relevant for traffic studies. behaviour with the attributes belonging to the current state of the Next subsections present in details all these aspects. person agent. Figure 2 shows these attributes classified in physical state and mood, depending on when they affect physical action and 3.1 ROLE AND MEAN OF TRANSPORT perception or thinking and attitudes respectively. The behaviour of a person in traffic is first limited by his/her mean The physical state influences the perception of the environment. of transport. Although passengers influence traffic, e.g. distracting Individuals do not receive objective information from the environ- drivers, these models focus on people controlling their mean of trans- ment, as this is really mediated by their own senses and depends port. on external conditions [15]. The personal conditions are represented The mean of transport requires certain processes to manage it, and with the values for this attribute, which are focused, drowsy, dis- also makes possible some processes. For instance, a person can know tracted and drugged. The influence of the external conditions is rep- how to brake a car, but needs driving one to perform the action. At the resented using the environment entity. same time, different means of transport obey different rules. These The environment entity has attributes for the weather conditions can be both explicit, e.g. traffic regulations, and implicit, e.g. drivers and type of environment [18]. The first one takes values between facilitating other drivers maneuvers. sunny, cloudy, rainy, heavy rain, windy, snowy, ice and foggy, while The models represent this information with a hierarchy of agents the second one is classified as familiar, unknown, difficult, affordable (see Figure 1). The basic agent is person, which incorporates the goal and straightforward. These attributes are linked to the physical state, of arriving to a destination following a certain route and perception pointing out that they affect its value. of obstacles and signals. According to the mean of transport, a person The mood considers that external factors influence people mental is extended as a driver or a pedestrian. These agents have additional state [7]. This state affects aspects such as decision making or level of goals, information and tasks to move by car or on foot respectively. attention to the environment. The specification of this attribute is fur- In the case of the driver, there is a related vehicle. The vehicle is ther decomposed into the attributes impatience and self-confidence. represented as an external application with, for instance, methods to The impatience represents the frustration of the person, perhaps be- brake or to manage the steering wheel and events from the speed cause she/he is in a hurry or the traffic conditions are adverse. The indicator. self-confidence represents the assurance of individuals on their own knowledge, capabilities and skills. Both attributes take values be- tween 0 and 5. Depending on the value, they can have a positive or 3.2 TRAITS OF PEOPLE INVOLVED IN negative effect on the person processes. For instance, a person with TRAFFIC self-confidence 5 can make risky decisions that are inadequate for The way of behaving in traffic also depends on the personal traits the perceived situation. In the case of pedestrians, the self-confidence of each person. A well-known example is the differences in acci- also indicates how crowds influence individual trajectories [9]. dents regarding age and gender [12]. These traits are static for each As it happened with the physical state, the value of the self- person, as they do not change during the travel, and thus in the sim- confidence is affected by other attributes. A familiar type of envi- ulation. There are three groups of traits: physical attributes, attitudes ronment and good weather conditions increase the self-confidence. and traffic experience. Moreover, people frequently move in groups, and this companion al- Figure 1. Inheritance of driver and pedestrian agents. Figure 2. Relationships between the elements of the taxonomy. Dotted lines represent that an attribute affects the calculation of other. ters the self-confidence with comments or actions. The companions nication between driver agents is not suitable, as drivers are in their attribute gathers this information. It considers the attitude of the com- own vehicles. The perceived agents could depend on new physical panions with values in silence, little chatty, chatty and fun-loving, attributes related to sight. and the number of individuals. This approach enables that the effects of the driver attitude in the Note that this presentation has pointed out several mutual relation- simulation can dynamically change. Such effects are modulated by ships between attributes. For instance, a bad physical state worsens attributes of the current state, which are influenced by traffic con- the perception of the road, reducing the self-confidence. In the mod- ditions, e.g. surrounding vehicles, speed or time waiting. Therefore, els, tasks managing the internal state of agents implement these mu- studies using these models provide behavior that is more realistic. tual influences. 4 CASE STUDY 5 RELATED WORK The attribute traffic profile presented in this paper is based on the classification of drivers in [13]. This classification uses the selfish principle, which assumes that any driver has a certain level of self- The current research must be framed within two main lines of work: ishness when pursuing his or her goals. This level classifies drivers in studies on people and their behaviour regarding traffic, and traffic moderate, normal and aggressive, determining their speed or prone- simulations. Both of them are sources of information to develop our ness to make risky decisions. The main limitation of that work is that models and validate them. two drivers of the same group do not differ in their behaviour, which Studies on actual people provide information on the relevant at- is not a realistic approach. This case study considers how the models tributes regarding traffic and the actual processes involved in it. proposed in our work cover the previous classification and facilitate These studies typically focus on obtaining data, statistics and rela- its extension. tionships among some factors under scrutiny. For instance, they try As previously mentioned, the traffic profile trivially supports the to identify aggressive behaviour and the reasons of their appearance classification in [13]. Its effect over driving depends on the imple- [17]. Some commonly considered attributes in these studies are gen- mentation of the tasks of the different agents. Note that since this der and age, as in [4, 12], presence of passengers in a vehicle, the is an attribute of the person agent, which is the base type of all our weekday and the most troubled hours [4], the driving experience agents, all the agents in our models include that attribute. of drivers [10], the physical state [18], and the mood [10]. There The heterogeneity of behaviour for agents with the same traffic are also behavioral studies more focused on the driving processes. profile is achieved with several attributes. The impatience is particu- These studies monitor, for instance, physiological signals, gestures larly relevant in this context, as it captures the anxiety produced by or speech to identify and/or predict decision-making, low-level ma- the current traffic situation. neuvers or drivers mood [19]. These studies propose attributes and The original work in [13] also discusses the phenomenon appear- processes that could be considered for simulation, but have some ing at junctions known as dominance. It happens when a driver or a limitations for this purpose. They are difficult to use to validate sim- group of them who are in a lane of a junction push their way, fol- ulation models or check the influence of new elements, as this would lowed by other cars, and get to block the other lanes. This lane of imply researchers to carry out new studies. cars will be the only one able to move forward as long as they do not The previous knowledge has been used in a variety of simulations free the junction. If drivers of two or more lanes of a junction exhibit with different goals. Regarding the level of abstraction at which traf- this behaviour at the same time, they can produce a deadlock where fic phenomena are considered, these simulations can be classified nobody will go forward. in: macroscopic, mesoscopic and microscopic simulations. The first With the presented models, this kind of behavior can be the con- ones attempt to capture the general principles governing the system sequence of the traffic profile and certain attributes present in the instead of individuals, in a way similarly to analytical models. They current state of the person. An aggressive driver is more dominant at are typically used to represent large areas of terrain with large quan- a junction than a moderate or normal one, and therefore the former tities of vehicles and traffic infrastructure conflicts [3, 16]. On the tries to cross the junction with greater determination. When drivers contrary, microscopic simulations present individual elements with have the same traffic profile, their current state is also crucial for the higher complexity. Most of ABM in traffic belongs to this category dominance. The attributes of current state more directly involved in [5]. The related computational costs make them suitable only to rep- this behavior are impatience and self-confidence. A high impatience resent small areas of terrain with few individuals. Moreover, it is dif- makes the driver prone to make quick decisions, not always enough ficult to embed general rules of behaviour in them, as rules usually meditated. With a high self-confidence, the driver dares to perform appear on each agent. Mesoscopic simulations are hybrid between maneuvers that in other circumstance she/he would not carry out. On the previous types. They try to solve their limitations locating each the contrary, a low self-confidence leads the driver to doubt about information or behavior at the most suitable level, either individuals maneuvers in complex settings (e.g. many cars around), causing that or groups [2]. Our work belongs to this category. As it is based on other more determined drivers cross before her/him. Furthermore, INGENIAS [14], it supports modelling both individual agents and drivers that are more impatient push others, which increase the frus- groups (not shown in this paper), as well as inheritance hierarchies tration of the later. This causes a widespread anxious mood in the involving those abstractions. junction [11], which makes it more hazardous. As a distinctive feature of our research, it works with simulations The update of the impatience level requires that agents know their at the level of models. As shown in [6], this facilitates the automated position and that of their neighbors. The own position can be in- generation of simulations for different target platforms from the spec- cluded as a new attribute position of the driver agent. The positions ifications using model-driven techniques. Other works have this in- of the other drivers are known through interactions of the driver with formation embedded in their programming tools [1], reducing the an external application that mediates its perception. Direct commu- possibilities of studying and reusing that knowledge. 6 CONCLUSION REFERENCES [1] Moshe Ben-Akiva, Michel Bierlaire, Haris Koutsopoulos, and Rabi Mishalani, ‘Dynamit: a simulation-based system for traffic prediction’, This work has presented the models of a taxonomy of people re- in DACCORS short term forecasting workshop, The Netherlands. Cite- garding their participation in traffic phenomena. These models are seer, (1998). intended to provide the basis for an extensible specification able to [2] Wilco Burghout, Haris Koutsopoulos, and Ingmar Andreasson, ‘Hybrid integrate available research and applicable to develop simulations. mesoscopic-microscopic traffic simulation’, Transportation Research Record: Journal of the Transportation Research Board, (1934), 218– The taxonomy is organized around three main dimensions: the 255, (2005). mean of transport used by people, their traits and current state. The [3] Nurhan Cetin, Adrian Burri, and Kai Nagel, ‘A large-scale agent- mean of transport currently only distinguishes between pedestrians based traffic microsimulation based on queue model’, in IN PROCEED- and people travelling using some motor vehicle, and among the later INGS OF SWISS TRANSPORT RESEARCH CONFERENCE (STRC), MONTE VERITA, CH. Citeseer, (2003). between drivers and passengers. Traits provide information about [4] Sean T Doherty, Jean C Andrey, and Carolyn MacGregor, ‘The situa- static features of people, both physical (e.g. gender and age), atti- tional risks of young drivers: The influence of passengers, time of day tudinal (i.e. traffic profile) and based on experience (i.e. traffic expe- and day of week on accident rates’, Accident Analysis & Prevention, rience). The current state considers several dynamic attributes that 30(1), 45–52, (1998). depend on the current environment and traffic state, such as self- [5] Patrick AM Ehlert and Leon JM Rothkrantz, ‘Microscopic traffic sim- ulation with reactive driving agents’, in Intelligent Transportation Sys- confidence and impatience. Among all these attributes, the discus- tems, 2001. Proceedings. 2001 IEEE, pp. 860–865. IEEE, (2001). sions have been focused on the traffic profile, which distinguishes [6] Rubén Fuentes-Fernández, José M Galán, Samer Hassan, Adolfo moderate, normal and aggressive people, and self-confidence and im- López-Paredes, and Juan Pavón, ‘Application of model driven tech- patience as dynamic modifiers of that profile. The relationships be- niques for agent-based simulation’, in Advances in Practical Applica- tions of Agents and Multiagent Systems, 81–90, Springer, (2010). tween these attributes illustrate how people behavior can be linked to [7] Ray Fuller, ‘Towards a general theory of driver behaviour’, Accident specific situations and combinations of attributes. Analysis & Prevention, 37(3), 461–472, (2005). The case study has shown how to use the proposed models to spec- [8] Nigel Gilbert and Klaus Troitzsch, Simulation for the social scientist, ify the simulation in [13]. The drivers attitude from [13] is the traffic McGraw-Hill Education (UK), 2005. profile of our models, but our modes additionally fix some of the [9] Dirk Helbing and Peter Molnar, ‘Social force model for pedestrian dy- namics’, Physical review E, 51(5), 4282, (1995). limitations pointed out in that work. In particular, they offer a sim- [10] Timo Lajunen, Dianne Parker, and Stephen G Stradling, ‘Dimensions ple way to introduce heterogeneity in the behavior of drivers in each of driver anger, aggressive and highway code violations and their medi- attitudinal group using the current state. At the same time, the spec- ation by safety orientation in uk drivers’, Transportation Research Part ifications in the case study were able to replicate other phenomena F: Traffic Psychology and Behaviour, 1(2), 107–121, (1998). [11] Timo Lajunen, Dianne Parker, and Heikki Summala, ‘Does traffic con- appearing in [13] such as dominance. gestion increase driver aggression?’, Transportation Research Part F: Working at the level of models facilitates comparing approaches Traffic Psychology and Behaviour, 2(4), 225–236, (1999). and reusing information between different studies. It also reduces [12] Dawn L Massie, Kenneth L Campbell, and Allan F Williams, ‘Traffic the costs of migration between different simulation platforms, as the accident involvement rates by driver age and gender’, Accident Analysis relevant information is available at a higher level of abstraction than & Prevention, 27(1), 73–87, (1995). [13] Praveen Paruchuri, Alok Reddy Pullalarevu, and Kamalakar Karla- that of code. Moreover, it promotes the design of domain specific palem, ‘Multi agent simulation of unorganized traffic’, in Proceedings modeling languages for different needs in traffic studies. of the first international joint conference on Autonomous agents and The current models are part of an ongoing effort to build a general multiagent systems: part 1, pp. 176–183. ACM, (2002). simulation platform for traffic. The current prototype integrates the [14] Juan Pavón, Jorge J Gómez-Sanz, and Rubén Fuentes, ‘The ingenias methodology and tools’, Agent-oriented methodologies, 9, 236–276, previous information in multi-agent systems based on the A-Globe (2005). platform and using geographical information from Google . The de- [15] Dario Salvucci, Erwin Boer, and Andrew Liu, ‘Toward an integrated velopment process is evolving to a fully model-driven approach in or- model of driver behavior in cognitive architecture’, Transportation Re- der to explore the actual benefits of these approaches for simulation. search Record: Journal of the Transportation Research Board, (1779), Regarding the models, those presented here still does not consider 9–16, (2001). [16] Jason Sewall, David Wilkie, Paul Merrell, and Ming C Lin, ‘Continuum several relevant aspects of traffic. Some of those to be included are traffic simulation’, in Computer Graphics Forum, volume 29, pp. 439– vehicles, companion agents or public transport. There is also need to 448. Wiley Online Library, (2010). consider extensions of the modelling language to facilitate the spec- [17] David Shinar, ‘Aggressive driving: the contribution of the drivers and ification of, for instance, relationships between attributes and the in- the situation’, Transportation Research Part F: Traffic Psychology and Behaviour, 1(2), 137–160, (1998). fluence of these on the execution of certain actions. [18] Michael Sivak, ‘Human factors and highway-accident causation: Some theoretical considerations’, Accident Analysis & Prevention, 13(2), 61– 64, (1981). [19] Shigeki Tezuka, Hitoshi Soma, and Katsuya Tanifuji, ‘A study of driver 7 ACKNOWLEDGEMENTS behavior inference model at time of lane change using bayesian net- works’, in Industrial Technology, 2006. ICIT 2006. IEEE International Conference on, pp. 2308–2313. IEEE, (2006). This work has been done in the context of the project “Social Ambient Assisting Living - Methods (SociAAL)” (grant TIN2011- 28335-C02-01) supported by the Spanish Ministry for Economy and Competitiveness, the research programme MOSI-AGIL-CM (grant S2013/ICE-3019) supported by the Autonomous Region of Madrid and co-funded by EU Structural Funds FSE and FEDER, and the “Programa de Creación y Consolidación de Grupos de Investigación” (UCM-BSCH GR35/10-A).