real-time traffic management through knowledge - CiteSeerX

REAL-TIME TRAFFIC MANAGEMENT THROUGH KNOWLEDGEBASED MODELS: THE TRYS APPROACH Josefa Hernández, José Cuena, Martín Molina Department of Artificial Intelligence Technical University of Madrid Campus de Montegancedo s/n Boadilla del Monte, 28660-Madrid (Spain) Phone: +34 91 336 73 90, Fax: +34 91 352 48 19 Email: {phernan, mmolina, jcuena}@dia.fi.upm.es

ABSTRACT: This paper presents a general approach for real time traffic management using knowledge-based models named TRYS. Recognising that human intervention is still required to apply the current automatic traffic control systems, it is argued that there is a need for an additional intelligent layer to help operators to understand traffic problems and to make the best choice of strategic control actions. The need for an open architecture is stated, in order to allow users to modify decision criteria according to their experience. The result of applying this approach is the TRYS model, a knowledge-based traffic management model, which is in detailed described in this paper, that demonstrates the feasibility of the approach and the compatibility of this type of systems with state-of-the-art traffic control systems. The TRYS model was applied to develop several urban motorway control systems installed in the traffic control centres of different Spanish cities. At the end of the paper, the result of applying the TRYS approach to develop a traffic control management system for Barcelona is presented.

KEYWORDS: Traffic Management, Intelligent Systems, Knowledge Based Architectures, Inference Techniques, Real Time Information Systems.

1.- INTRODUCTION In the last years, the significant evolution of the telematics technology made possible the development of systems integrating the management of communications networks and real time data bases in order to provide information about the performance of traffic networks. This working scenario is based on the reception of a large amount of real time data from detectors that need to be interpreted and analysed by the operators to support the decision making process. Here, a careful evaluation of the adequate decisions in many cases requires the use of sophisticated models of behaviour. This type of operation can be supported by the use of automatic tools developed using knowledge-based technology that provides solutions to emulate the intuitive problem-solving steps made by the decision makers. However, the role of these systems in the traffic management process should be clear. This type of tools cannot be conceived as substitutes of the traffic operators in the decision making process but, on the contrary, these tools should be viewed as assistants that offer to the operators a wide range of exploratory actions useful to evaluate the different potential decisions. In this way, operators can get from the system justified suggestions that guide the decision making process but they still maintain the final responsibility on the management. This paper presents a general structure for modelling traffic control knowledge at the strategic level. In this model, criteria for problem understanding and problem solving in areas controlled by traffic lights and/or variable message sign (VMS) panels are described. This model was implemented in the TRYS project funded by the Spanish Directorate for Traffic (DGT, 1994) and gave rise to three different traffic control applications for the cities of Sevilla, Madrid and Barcelona. The contents of the paper are organised as follows. First, the main characteristics of the traffic control problem domain and therefore of the type of problems to be solved in this domain are described. Next, an overview of the state of the art in traffic management systems is presented to provide a rationale for the introduction of knowledgebased models into the current urban traffic control (UTC) technology and traffic management practice. Then, a general conceptual model for traffic analysis and management is described, followed by a formulation of the components of the

model at the symbol level. Finally, the application of this model to develop the TRYS system for Barcelona is described.

2.- THE DOMAIN OF TRAFFIC MANAGEMENT The domain of traffic management has recently experienced a significant demand of advanced information technology. Control centres for traffic management are on-line connected to different devices (such as detectors on roads, cameras, traffic lights, etc.) making possible that operators supervise the state of the road by consulting data bases with recent information from detectors and modify the state of control devices. The use of these traffic monitoring and management facilities requires sophisticated support tools for on-line operators, to help them in dealing with the information complexity and diversity of sensors and control devices. Figure 1 shows the typical information infrastructure for real-time traffic control that can be found in different cities. There are sensors (loop detectors) on major roads recording several traffic magnitudes such as speed (km/h), flow (vehicle/h) and occupancy (percentage of time the sensor is occupied by a vehicle). The distance between successive sensors on a freeway is usually about 500 meters. This information arrives periodically to the control centre (e.g., every minute). The control centre receives also information about the current state of control devices. Control devices include traffic signals at intersections, traffic signals at sideways entry-ramps, variable message signs (VMS) that can display different messages to drivers (e.g., warning about existing congestions or alternative path recommendation), radio advisory systems to broadcast messages to drivers, and reversible lanes (i.e., freeway lanes whose direction can be selected according to the current and expected traffic demand). In the control centre, operators interpret sensor data and detect the presence of problems and their possible causes. Problems are congested areas at certain locations caused by lack of capacity due to accidents, excess of demand (like rush hours), etc. In addition, operators determine control actions to solve or reduce the severity of existing problems. For instance, they can recommend increasing the duration of a phase (e.g., green time) at a traffic signal, or they may suggest displaying certain messages on some VMSs to divert traffic.

Traffic Control Center

TV Camera

Variable Message Panel

80

slow traffic ahead

Transmission Data Network

Vehicle Detectors

Figure 1: Typical information infrastructure for real-time traffic control.

According to this scenario, a real-time decision support tool that helps operators in detecting traffic problems and choosing appropriate control actions should generate information to satisfy three types of queries, Cuena, Hernández (1997):

• • •

What is happening?, i.e., what problem of a predefined set of problem classes is currently occurring in the traffic network. What may happen if?, i.e., what relevant events may happen in the near future if some external conditions change according to certain hypotheses. What should be done?, i.e., what control actions should be taken in order to improve the current traffic state considering hypotheses about external conditions.

3.- THE ROLE OF KNOWLEDGE-BASED SYSTEMS IN TRAFFIC MANAGEMENT Traffic management systems must be reactive to the different states of traffic flow in the controlled network. In the early systems, the approach was based on a library of signal plans applied on-line in different predefined situations according to some time-based criteria (fixed-time systems) or to the traffic data collected by roadside sensors (traffic actuated systems). However, this precalculated-plan approach usually lacked the conceptual granularity required by the system to be adaptive enough to the variety of situations, in time and space, which may occur in the network. As a result, this type of system was usually operated in a complementary way: a plan was selected from the library but, in certain situations, the operator had to introduce modifications according to her/his perception of reality together with her/his knowledge of traffic behaviour in the network. In the 1980’s more adaptive systems were introduced starting from SCOOT Hunt et al. (1981) and Bretherton, Bowen (1990) where, for the first time, an intelligence for understanding traffic situations in real time was designed and integrated with a model for decision making. Several other responsive systems were developed in this decade, including SCATS, Lowrie (1982), OPAC, Gartner (1983), PRODYN, Henry et al. (1983) and UTOPIA, Mauro, Di Taranto (1989). Dynamic systems like SCOOT introduced a very relevant change in traffic control systems, since such systems implemented for the first time an on-line intelligence for traffic control. This intelligence was organised as an evaluation step of local analysis at the level of single intersections together with a more general view evaluating a sequence of intersections. The analysis was performed along time to act on signal phase modifications during time slices of few seconds. However, although some parameters remained externally modifiable the reasoning model applied was hidden to the user. The experience in using SCOOT-like systems showed that a good performance of a traffic network might be attained when traffic situations are not critical, or congested conditions occur only for limited time intervals. In persistently congested situations, as in the precalculated plans approach, operator intervention is usually required because: • The evaluation of the traffic situation by the system may be biased due to the fact that sensors are sometimes insufficient for understanding the congestion process (e.g. the length of the queues at the stop line may overflow past the sensors placed for queue detection). • For the sake of stability in the network, the decisions taken by the system are small adjustments to the current signal plan, while more drastic changes might often be required. Studies performed in a number of sites have shown that operator intervention is almost customary in most UTC installations. The knowledge used by the human operator to decide which changes are required is mostly based on her/his understanding of the reasons causing the problem (i.e. the structure of the traffic demand, the existence of alternative paths to some conflicting nodes, etc.). The same need for taking into account specific aspects in problematic areas appears in motorway control through VMS panels. In this case, the advice displayed to drivers should be sufficient to: i) allow the users to take some of the available options and ii) not produce excessive flow outside the motorway that may congest alternative options. Every problem area therefore requires analysis of the situation using knowledge about traffic behaviour and control criteria specific for that area. The above considerations suggest a need to complement existing systems for traffic control (including pre-calculated plans systems, dynamic systems and VMS systems) with an additional layer where strategic knowledge is applied to understand the specific processes of congestion development, and corresponding actions for alleviating the problem may be modelled. As the knowledge to be introduced depends on the characteristics of the area, general models such as SCOOT may be not open enough since the form of the computer application supporting this type of model should allow the user to introduce, modify and maintain over time some conceptual formulations of his/her knowledge related to

every area of interest in the network. This may require more than a set of parameter values to be tuned to the specific application site, as used by existing traffic control systems. The technology of knowledge-based systems may help in designing and implementing suitable knowledge structures to formulate conceptual models for traffic analysis and management and to use such models for on-line strategic traffic management operations, Cuena et al. (1995), (1996), Molina et al. (1998). Work in knowledge-based systems applied to traffic management started at the end of the 1980’s (Scemama (1989), Cuena (1989)). At this time, the OECD created an expert group about expert systems in transport that produced two symposia, OECD (1990), (1992). Also, the European ATT/DRIVE and Telematics Applications Programmes promoted different experimental projects such as CLAIRE, Scemama (1992), KITS, Boero et al. (1994), Cuena et al. (1992) and FLUIDS, Hernández, Molina (1998).

4.- KNOWLEDGE-BASED ARCHITECTURE FOR TRAFFIC MANAGEMENT: THE TRYS MODEL The ideas presented above were put into practice in the TRYS system. TRYS is a knowledge representation environment supporting models to perform traffic management at a strategic level in urban, interurban or mixed areas. The city or traffic network where traffic has to be supervised is divided in several sections called problem areas. The decomposition of the city into problem areas allows a better analysis and understanding of the causes and evolution of traffic problems than if performed from a global perspective. This split does not define a set of disjointed areas whose sum is the whole city, but every area represents a part of the city where a determined traffic behaviour is usually present and where a set of signal elements can be managed to influence this behaviour. Then, a problem area may overlap with surrounding areas sharing, for instance, some signals but using them from different points of view. So, a problem area is a part of a city where traffic behaviour is locally studied and suitable control actions may be defined to improve the traffic state. An agent who understands the traffic conflicts that may appear, the usual behaviour of vehicles in the area and the signal and/or VMS actions that may improve the traffic state supervises every problem area. The control proposals generated by every agent are received by a higher level agent, called the co-ordinator, whose aim is to produce global proposals for the whole city by putting together the local proposals provided by the agents and removing the inconsistencies among them. Figure 2 shows the organisation of a set of agents.

Coordinator

Agent

Agent

Agent

Agent

Problem Area

Problem Area

Problem Area Problem Area

Figure 2: Basic structure of the control model

Based on this concept, TRYS organises the control knowledge in a set of local control entities named agents and a coordinator to synthesise the proposals. In the next sections, the knowledge used by each one of these units is detailed. 4.1.- MODEL OF AN AGENT The goal of an agent is twofold: (i) to diagnose the traffic problems present in a local problem area together with an explanation justifying such a diagnosis, and (ii) to propose control actions for the available signal devices to improve traffic conditions using the diagnosis information. To support this functionality, every agent needs three types of knowledge about the supervised problem area: • Physical structure: Knowledge about the traffic behaviour in the problem area. It includes procedures for interpreting sensor data (data abstraction) and understanding the network physical structure at different levels. • Traffic problems: Knowledge about how to detect and diagnose the presence of incidents or congestions. • Traffic control: Knowledge about the definition of control strategies adequate to solve the different problems. These knowledge areas are presented in detail in the next three sections.

4.1.1.- Physical structure The aim of the physical structure knowledge is to represent both static information about the topological structure of the problem area (components and their relationships) and the dynamic aspects of this structure. The physical structure is formulated using a declarative description of the network as a graph with nodes and links together with abstraction functions associated to components. For instance, figure 4 shows a graphical representation of a problem area where ellipses represent nodes, dots represent the location of detectors and sections, segments represent VMS panels and squares represent control zones. The procedures applied with this kind of knowledge are data abstraction methods that compute qualitative values from basic traffic parameters provided by sensors, i.e. speed, occupancy, saturation. Figure 3 shows the functions used to abstract the speed, saturation and occupancy numerical values. These functions are possibility functions (typically used in fuzzy logic) and although they are generally defined, it is possible to add particular versions of these functions for specific parts of the roads. POS low

10

20

high

medium

1.0

30

40

50

60

70

80

90

100

80

90

100

80

90

100

110

120

130

SPEED (Km/h)

POS low

1.0

high

medium

SATURATION (Flow/Capacity)*100 10

20

30

40

50

60

70

POS 1.0

low

high

medium

OCCUPANCY (Percentage) 10

20

30

40

50

60

70

Figure 3: Possibility functions to abstract traffic values from numerical data.

MATARO A19/13-PIV-1 13SMD2

A19 Barcelona PK_13100

12SMD2


A19 Barcelona salida a Pata Norte

NII A19 Barcelona incorporación desde la NII

12SMS2 A19 Barcelona PK_10000

12SMD4 11SMD2

A19 Barcelona PK_9500 A19/11-PIV-1

A19 Barcelona salida a Badalona Norte

BADALONA NORTE

11SMD4


10SMD2


9SMS2 9SMD2

A19 Barcelona incorporación desde Badalona Norte


8SMS2 A19/8-PIV-1

BADALONA CENTRO_SUR A19 Barcelona incorporación desde Badalona Centro

8SMD2


7SMD4


7SMD2


6SMS2 6SMD2

A19 Barcelona PK_6300 A19/5-PIV-1

A19 Barcelona salida a Badalona Sur A19 Barcelona incorporación desde Badalona Sur

5SMD4

A19 Barcelona PK_5600 A19 Barcelona PK_5300

5SMD2 5SMS1


4SMD2 3SMS2

A19 Barcelona desvío PK_3000

ZONA URBANA A19 Barcelona desvío PK_2000

3SMD2


3SMD4


2SMD4


2SMS3 A19/2-PIV-1 2SMD2


1SMD6


1SMS2 1SMD2


1SMD4


GLORIES

Figure 4: Graphical representation of a traffic network.

4.1.2.- Traffic problems The second knowledge area of an agent is specialised in the detection and explanation of traffic problems that may appear in the problem area, like congestion or incidents. The goal is to analyse the data recorded by sensors to • point out the current or foreseeable presence of problems taking into account the recent trend of traffic conditions and, • explain the severity and possible causes of these problems to be used later in the selection of the appropriate control actions. In general, an agent understands traffic problems as an imbalance between capacity and demand which generates an increase in density that impedes the fluidity of traffic. In a city network, the problems appear when a queue of vehicles propagates to the surrounding streets, blocking intersections and generating a so-called congestion tree. In a motorway, the problems are a consequence of a loss of capacity due to an unexpected incident or to the characteristics of the network. The problems are analysed not only by observing the instant where the problem has been detected but with certain

deepening in time. In this way, the presence of short-term congestion can be foreseen. This capability of short term prediction is very important for a system willing to propose control actions that contribute to the prevention of problems, and not only the solution of currently detected problems. This aspect is important because in many cases the effort needed to avoid a predictable congestion is significantly less than that necessary to eliminate an existing congestion. The approach made to describe a problem and to explain its causes is based on associating a traffic problem with an excess traffic volume, in such a way that a problem appears because the traffic demand received by a section of the network is bigger than the capacity offered by that section. According to this, the characterisation of a problem may distinguish three different issues: • Where is the problem? A problem is located in a so-called critical section, i.e. the point where the congestion starts due to a lack of capacity. Examples of critical sections are: works on the road disabling a lane, an entry ramp to a highway with a very high traffic demand, a turn movement controlled by a signal with a short green time, an accident that blocks a lane, etc. • How severe is it? The size of the imbalance between the network capacity and the traffic demand is a measure of problem severity. This imbalance is characterised by the excess of arrivals (measured in veh/hour), which is understood as the minimum demand decrease (or capacity increase) needed to solve the actual problem. It may be computed as the difference between the flow demand arriving at the section where the problem is located, obtained upstream of the congested area, and the capacity in that section. An accurate estimation of problem severity is a key aspect to pose an effective control plan. • What is the cause of the problem? Finally, the concept of participation is also managed to distribute the causes of the problem among the paths that cross the critical section carrying a significant traffic flow. Those paths are called involved paths. This information is important to understand the origin of the problem and to pose adequate actions on the points of the network that generate the traffic. The goal of problem identification reasoning is to analyse data from sensors to detect which are the active critical sections, indicating the excess of arrivals and the paths involved. In order to carry out this task, the necessary knowledge is represented using frames. The frame language provides a traffic language that can be used by an expert to express his experience in detecting and explaining problems. A frame is a collection of state variables for traffic and signals, characterising the situation in an area, whose values are included in intervals representing prototypical situations (each one of these variables is called a slot). Two types of frames are used to express the problem identification and diagnosis knowledge. Figure 5 shows an example of a frame used to detect traffic problems. In particular, that frame describes a traffic problem in a motorway corresponding to a traffic jam caused by an entry-ramp. The frame includes at the beginning a section (called variables definition) to define the set of variables that will be used within the frame. These variables are dynamically associated to particular values during the inference process by using the conceptual description of particular roads. In this example, several variables are used to define a generic part of the road where there is an entry-ramp (the three variables called previous, next and ramp identify the relevant points of the road area). Then, the next section of the frame (called description) is used to formulate the traffic state by associating values to attributes (slots). For instance, the speed of the previous section must be low, the occupancy of the previous section must be high, and so on. The value of the attributes can be a single qualitative value, a list (this means that the actual value must belong to that list), a number, a variable or a numerical expression (this is a kind of if-needed procedure that dynamically computes the value of the attribute). Each attribute has a label (e.g., labels a, b, c, d, etc. in this example). Labels are used in the last section of the frame (called relevance of characteristics) to indicate when (and to what extend) the whole frame can be deduced if some of the slots match the actual situation. Here, a rule representation (very simple) is used. Each rule includes on the left-hand side a logical expression of labels of slots (and is represented as a colon and or is represented as a semicolon). On the right-hand side, a percentage indicates the matching degree of the frame when the slots (whose labels are on the left-hand side of the rule) match the actual situation. For instance, in the example the third rule means that if slot a (or slot b) matches, and slot h matches, then the whole frame matches with a degree of 80%. Since slots may partially match the current state of the traffic, and there are several rules, an uncertainty model is required to

compute the matching degree for the frame. For this purpose, a particular version of the Mycin uncertainty model is used.

FRAME traffic jam caused by entry-ramp VARIABLES DEFINITION () type = on-ramp link, previous section = , next section = , ramp section = , DESCRIPTION () speed = low occupancy = high saturation = low () occupancy = [medium, high] saturation = low () speed = medium occupancy = [low, medium] saturation = high (critical section) location = state = overflow category = problem excess = EXP(dem() - capac()) RELEVANCE OF CHARACTERISTICS (a; b), d, h -> 100%. i, j -> 100%. (a; b), h -> 80%. d, h -> 70%. (a; b) -> 50%.

[a], [b], [c], [d], [e], [f], [g], [h], [i], [j], [k], [l],

Figure 5: Example of frame representing a type of traffic problem.

The second type of frame is used to formulate the demand model in the knowledge area and to support the problem diagnosis task. This knowledge is represented by using hierarchies of patterns of demand associated to temporal intervals, in such a way that it is possible to determine demand scenarios for a given present or future state. Figure 6 shows a partial example of such a frame. The first set of slots is used to define the temporal range where this demand is active. In this example, a labour day from 6:30 am to 9:30 am. Then, there are slots that represent origin-destination pairs with an approximate estimation of the traffic flow. For instance, from the origin Mataro to the destination Badalona Norte there will be an approximate flow of 1000 veh/h. This type of frames are not general, i.e. they must be given for each particular traffic area.

FRAME morning peak TYPE week day pattern type of day: labour day, temporal interval: 6:30 - 9:30, .... Mataro -> Badalona Norte: 1000 veh/h, Mataro -> Badalona Centro Sur: 1000 veh/h, Mataro -> Zona Urbana: 2000 veh/h, Mataro -> Glories: 2000 veh/h, NII -> Badalona Norte: 500 veh/h, NII -> Badalona Centro Sur: 500 veh/h, .... Badalona Centro Sur -> Glories: 1000 veh/h.

Figure 6: Partial example of frame representing a pattern of traffic demand.

4.1.3.- Control actions The goal of the traffic control process is to choose proposals of messages to be displayed in some VMS panels that induce drivers to take paths that do not pass through congested areas. The starting point is the set of paths with the

greatest influence in the state of the critical sections, called problem paths. The objective is to find recommendations (warnings of slow traffic, warnings of congestion, recommended paths, etc) that decrease traffic on the set of problem paths. The input of this process is the state of every problem focus indicating if it is free, overloaded or an incident. In this sense, there is another kind of frame used to describe control recommendations. These frames are called path-use frames. Every path-use frame includes three types of knowledge: • The specification of the control action proposed in terms of specific messages for VMS panels. • The characterisation of the traffic conditions that may be present in the problem area to apply the proposed control plan. • The estimation of the expected impact of the control proposal in the traffic behaviour. In this way, the control actions knowledge is formulated as a collection of triplets . Figure 7 shows an example of such a representation for a control action plan that warns about the presence of an incident (e.g., a traffic accident) at a particular location. The first section of this example shows a set of messages referred to the existence of an incident to be presented to the drivers using certain VMS panels. The second section (state conditions) includes a logical expression about the state of the road that has to be true to apply this plan. Finally, the third section shows the expected effect of the control action on the traffic behaviour. This is represented as the expected distribution of the traffic through different paths. For instance, the example says that to go from Meridiana to A18, between the 80% and 90% of traffic will take the path by A17 road, and between 10% and 20% will take the path by secondary road. This representation can be used to provide two different types of inference methods. The first one determines the expected effect of current control actions and it is useful to interpret the current traffic situation. The second one works on the opposite direction. It proposes control actions that are consistent with the current traffic state and achieve a set of required effects. CONTROL ACTION PLAN warning about incident at A17 CONTROL ACTIONS (B20/3-PIV-1) message = INCIDENT AT A17 (B20/6-PIV-1) message = INCIDENT AT A17 (A17/8-PIV-1) message = INCIDENT AHEAD STATE CONDITIONS (A17 Girona PK5300) state = free AND (A17 Girona in Montcada) state = incident AND (A17 Girona at Montmelo exit) state = free EXPECTED EFFECT (Meridiana -> A18) [80, 90] by A17 road [10, 20] by secondary road (Trinidad -> A18) [60, 80] by A17 road [20, 40] by secondary road

Figure 7: Example of control action plan. The reasoning process goes over the path-use frame knowledge base analysing which would be the effect of the proposed signal, and the corresponding traffic distribution, on the observed traffic excess, i.e. if it can remove or reduce the traffic arriving to the congestion focus. Finally, the most promising frames are selected and presented to the operator. 4.2.- MODEL OF THE CO-ORDINATOR The result of the individual reasoning performed by an agent is a set of local control proposals for its problem area. All the local control proposals defined by the agents are analysed by the so-called co-ordinator to build global proposals by coherently synthesising the local proposals. The co-ordinator specialises in integrating the different local proposals provided by a set of agents to build signal recommendations for the whole network according to a specific traffic state. With this aim, two types of knowledge support its reasoning: • Knowledge to determine the compatibility of different local proposals. • Knowledge to define a model of agent priority that is used to take decisions when conflicts appear.

4.2.1.- Compatibility of control actions This is the knowledge about coherence between control actions which detects when two local proposals are

incompatible during the process of building a global proposal. The incoherent situations between signal proposals include two possibilities: • Physical conflicts caused by different actions on the same signal device. For instance, two agents propose displaying different messages on the same VMS panel. • Semantic incoherence between proposals for different signal devices that may be incoherent from the point of view of a user traversing a given path. For instance, a panel recommending a certain speed and another one, in the same or a neighbouring area, suggesting a significantly different speed. The representation used to establish incompatible situations is rule based, expressing the inconsistency between signal actions. For example, a message suggesting a certain route is incompatible with a message of road works in the same path. Rules are managed by a procedure whose inputs are pairs of control actions and whose output is their qualification as compatible or incompatible.

4.2.2.- Priorities between agents Finally, another rule-based representation is also used to formulate knowledge to solve conflicts about incompatible control actions. The knowledge base includes facts and rules about levels of priority and rules that determine which problem area must change its proposal. A set of facts represent couples of traffic areas describing which is more priority. There can be also rules that deduce this priority from the severity of existing problems. In addition to that, there are generic rules that define transitivity of priority and rules to select the problem area that must change the proposal according to their priority. 4.3.- SUMMARY OF THE TRYS MODEL Then, the traffic management model previously described, which is shown in Figure 8, may be summarised as follows: • A set of data bases, each one describing the physical structure of a problematic area. • A collection of frames representing prototypical problematic situations, using the general format described in Figure 5, which the expert user will be able to formulate, consult and evaluate according to his/her experience through typical computer operations for text editing. • A collection of frames representing demand models of different problem areas, using the general format described in Figure 6, which can be also modified by the expert users. • A knowledge base to support control decisions organised as a collection of control frames where, according to the format described in Figure 7, different patterns of traffic behaviour are described in terms of the impact provoked on path uses and critical sections. This set of frames is also easy to consult and modify by expert users. It is interesting to remark on the way of using this frame base: it may be applied in two reasoning modes: (i) in the prediction mode to evaluate the possible impact of a set of messages on the traffic state and, (ii) in the control mode to identify which frames produce a significant change in the state of the problem focus with respect to the current situation. In the prediction mode, messages are premises and focus states are conclusions. In the control mode, the change in the state of the problem focus is the input and the messages to be displayed are the output. So, the same declarative knowledge is used to achieve two different goals.

Traffic Network

...

Agent 1

Problems Knowledge

Problem Scenarios

Control Actions Knowledge

Behaviour Knowledge

Demand Model

Control Actions

Agent N

Problems Knowledge

Problem Scenarios

Physical Structure

Co-ordinator


Behaviour Knowledge

Demand Model

Compatibility Knowledge

Priority Knowledge

Control Actions

Physical Structure

Figure 8: Structure of a TRYS model • The co-ordination knowledge is based on two rule bases: - One that models the different unacceptable displays of signals or VMS panels, that are incompatible both from the point of view of control and from the point of view of consistency for the users crossing the paths affected by the congestions (i.e. the user must find messages along a path that are consistent between each other and consistent with the traffic situation). - Another one with rules for solving the different detected conflicts that propose alternative messages for the conflicting VMS panels starting with general messages that maintain consistency. These rule bases may be inspected by traffic experts in such a way that the model is viewed by the user as a collection of text that is quite understandable and thus, easy to maintain and improve. The possibility of improvement derives both from the personal analysis of the expert and by inspection of the explanations provided by the system for specific decisions during on-line operation.

5.- THE TRYS MODEL FOR BARCELONA The proposed generic model allows an expert user on traffic management to define a specific domain model by fulfilling the declarative concepts used to formulate the knowledge about problem scenarios, network structure, coordination, etc. These knowledge bases can be inspected and improved by the traffic experts according to the experience gained with the daily operation. These criteria have been applied on an extensive network in Barcelona that includes the ring road of the city and all the main accesses (an urban ring that includes Ronda de Dalt and Ronda Litoral and five accesses: the roads A2, A17, A18, A19 and C246). The problem in Barcelona was to develop a traffic management system that receives real time data and generates control plan proposals for congestion situations that might simultaneously appear in different parts of the network. These control proposals need to be supported with explanations to give the traffic operator confidence on taking the final decision. The control requirements in the city determined the structure of the TRYS model with one coordinator and 22 local control agents and therefore 22 problem areas. Every problem area considers only one sense of traffic flow. The network areas assigned to the agents were control dependent and therefore some agents had to share control devices. The traffic control infrastructure available in Barcelona included 52 VMS panels, ramp metering at 7 entrances to the ring road, and more than 300 single and double loop detectors (this means more than 5150 variables are received and

processed by the knowledge-based system at each reasoning cycle). These devices were connected with the Traffic Control Center through fiber optics communications and delivered the recorded data every minute.

Barcelona Traffic Network

C246-outbound

...... A17-outbound A-17-inbound A17-Inbound Traffic Management Agent

Problems Knowledge

Problem Scenarios


Behaviour Knowledge

Demand Model

Control Actions

Physical Structure

Co-ordinator

Compatibility Knowledge

Priority Knowledge

Summary numbers: 22 problem areas 52 VMS panels 300 loop detectors 89 knowledge bases 126 problem frames 188 path-use frames

Figure 9: Summary of the TRYS model for Barcelona The domain model contained a total of 22 conceptual vocabularies that contain a total of 3680 traffic concepts, and 90 knowledge bases (22 problem scenarios, 22 demand models, 22 physical structures, 22 sets of control actions plans, plus one for the compatibility model and another one for priority). For instance, the model includes a total of 126 problem frames (with 3780 logical expressions) and 188 control action plans (with about 4700 logical expressions). Figures 9 and 10 show a summary of this model and a display of some of the windows in the user interface. The model was installed in the Traffic Control Centre of Barcelona (Jefatura Provincial de Tráfico) in July 1996 fully integrated with the rest of the information system already operating in this control centre. It was distributed in a set of Unix processes and integrated in the global architecture established for the information system of the control centre. This produced a better efficiency and a good integration with the global user interface. The application for Barcelona was developed during one year (four persons) which meant a significant decrease in the required effort compared to the development of previous TRYS-based applications, and in particular to the TRYS model for Madrid. This was possible thanks to the possibility of reusing the generic model applied to Madrid and the operational support based on software components. The main activities to develop the application of Barcelona were focused on knowledge elicitation following the knowledge areas established for the generic model, together with activities related to software development in order to integrate the knowledge-model in the control centre. As a result, this experience showed that the use of knowledge models at generic level based on high level structuring concepts (tasks, problem-solving methods, knowledge areas, etc.) was extremely useful in order to increase the productivity and the quality of the final system in the development of complex and large knowledge systems as it is the case of decisionsupport applications for traffic management.

Figure 10: Screen example of the user interface of the application for Barcelona.

6.- CONCLUSIONS The experience in the application of an open architecture for a real time decision support problem such as the one described in this paper has shown the feasibility and advantages of this approach both for development and maintenance. During the process of development both the generic conceptual model and its contents have been modified several times without relevant losses of time and resources. Also, a good level of user system interaction is available through explanations so it is easy for the user to take decisions based on the system advice. In the last years, there has been an important increase in the use of advanced information technology applied to adaptive traffic management in urban areas. In addition to the usual signals at intersections whose state may be changed from a remote control centre, there are other control resources such us ramp metering, variable message signs (VMS) or advisory radio systems, which provide a high capability for traffic management from control centres. In big cities, the number of these control devices is so great that automatic tools are necessary to help control centre personnel consistently determine the best control strategy for each traffic situation, for example, by ensuring that traffic signal timing is consistent with messages on VMS panels. The central difficulty of having automatic tools to manage together all those devices in a co-ordinated and adaptive way is that their management is mainly supported by strategic human intervention based on heuristic and imprecise knowledge, which is difficult or impossible to be captured in algorithms. Accordingly, one solution is to use artificial intelligence techniques that are capable of modelling reasoning procedures similar to the way persons solve problems. This line of work has created a new generation of systems and tools in traffic control using these techniques. Knowledge-based systems have several characteristics that make them suitable to develop and maintain applications for complex process management environments, as in traffic applications. Users can profit from these advantages in the following sense: • they can build up and make operative models of knowledge and expertise in the traffic domain, • the adaptation and maintenance of the models over time with new knowledge or results from experience is feasible because knowledge is expressed in terms familiar to the user,

• the interaction between the user and system is improved due to the transparency and accessibility of the application models (knowledge representation) and of system behaviour (exploration of alternative options of reasoning, explanations, etc.). The TRYS system, presented in this paper, proves that knowledge based models designed for on-line use can be developed to be integrated as components of advanced traffic control centres (TCCs) in order to improve traffic monitoring and management capabilities of current traffic control architectures. Such models identify an additional strategic level in current TCCs, a traffic knowledge processing layer developed on top of the available traffic control facilities. This layer allows the development and use of models of knowledge and expertise of traffic operators in the controlled area, providing additional traffic analysis and management criteria that can complement currently applied, optimisation-based automatic traffic control. Furthermore, they also allow modification and improvement over time of traffic control strategies as new knowledge is gained about long-term modifications of traffic behaviour through day by day operation of the traffic network. TRYS has been developed as a modelling environment to support development and application of knowledge based models for network management. A TRYS model provides traffic monitoring functions and control actions. The interface between TRYS and the control system allows the TRYS model to accept input data (i.e. speed and occupancy measurements) from the real-time data collection facilities (via the traffic control computer) and to send back control actions to the traffic control computer. Depending on the traffic control system available at the application site, control actions can range from a set of constraints limiting the selection to a library of predefined signal plans (or a library of predefined messages in VMS applications) to a set of constraints on signal setting parameters (i.e. cycle time, phase split and offsets) in a fully adaptive system. The design of the knowledge representation solutions provided by TRYS has been guided by the double objective of: • providing a model for an accurate representation of both the relevant aspects of traffic behaviour and the generation of the associated control decisions, and • keeping the model within an acceptable level of understanding by the users. Obviously, alternative approaches could be used but any design must try to keep an adequate balance between the models complexity and the user understanding required for reliable maintenance and on-line operation. The TRYS system and other initiatives carried out in European and US R+D emphasise the increasing interest in the application of artificial intelligence techniques in the transport domain, and in particular, in decision support systems for real time traffic control.

7.- REFERENCES Boero Marco, Cuena José, Kirschfink Heribert, Traetteberg Hallvard, Wild Dieter, 1994, “The Role of KnowledgeBased Models in Traffic Management and their Design” in “Towards an Intelligent Transport System” Vol. 2, edited by ERTICO. París. Bretherton R.D., Bowen G.T., 1990, “Recent Enhancements to SCOOT – SCOOT version 2.4”, Proceedings of the 3rd International Conference on Road Traffic Control, IEE, London. Cuena José, 1989, “AURA: A Second Generation Expert System for Traffic Control in Urban Motorways”, IX International Workshop on Expert Systems, EC2, Avignon. Cuena José, Ambrosino Giorgio, Boero Marco, 1992, “A General Knowledge-Based Architecture for Traffic Control: The KITS Approach”, Proceedings of the International Conference on Artificial Intelligence Applications in Transportation Engineering, San Buenaventura, California. Cuena José, Hernández Josefa, Molina Martín, 1995, “Knowledge-Based Models for Adaptive Traffic Management Systems”, Transportation Research, vol.3, nº 5, pp. 311-337, Pergamon Press. Cuena José, Hernández Josefa, Molina Martín, 1996, “An Intelligent Model of Road Traffic Management in the Motorway Network Around Barcelona”, 14th International Federation on Information Processing (IFIP) World Computer Congress, IFIP’96, pp. 173-180, Chapman & Hall. Cuena José, Hernández Josefa, Molina Martín, 1996, “Knowledge Oriented Design of an Application for Real Time Traffic Management: The TRYS System”, 12th European Conference on Artificial Intelligence, ECAI’96, pp. 308-312, Wiley. Gartner N.H., 1983, “OPAC: A Demand Responsive Strategy for Traffic Signal Control”, Transportation Research Record 906.

Henry J.J., Farges J.L., Tuffal J., 1983, “The PRODYN Real-Time Traffic Algorithm”, Proceedings of the 4th IFAC/IFORS Conference on Control in Transportation Systems, Baden-Baden. Hernández Josefa, Molina Martín, 1998, “Advanced Human-Computer Interaction for Decision Support Systems Using Knowledge Modelling Techniques”, 15th International Federation on Information Processing (IFIP) World Computer Congress, IFIP’98. IT&KNOWS Conference, Information Technology and Knowledge Systems, pp. 400-414, Austrian Computer Society. Hunt P.B., Robertson D.I., Bretherton R.D., Winton R.I., 1981, “SCOOT: A Traffic Responsive Method of Coordinating Signals”, Department of the Environment, Department of Transport, TRRL Report Nº LR1014, Transportation and Road Research Laboratory, UK. Lowrie P.R., 1982, “The Sydney Coordinated Adaptive Traffic System. Principles, Methodology, Algorithms”, Proceedings of the IEE International Conference on Road Signalling, London. Molina Martín, Hernández Josefa, Cuena José, 1998, “A Structure of Problem-Solving Methods for Real-Time Decision Support in Traffic Control”. International Journal on Human-Computer Studies, Special Issue on Problem Solving Methods, Nº 49, pp. 577-600, Academic Press. Mauro Vito, Di Taranto Carlo, 1989, “UTOPIA”, Proceedings of the 6th IFAC/IFORS Conference on Control, Computers and Communications in Transport, Paris. OECD, 1990, Workshop on Knowledge-Based Expert Systems in Transportation, Espoo, Finland. OECD, 1992, 2nd Workshop on Knowledge-Based Expert Systems in Transportation: Operational Experience and Perspectives, Montreal, Canada. Scemama G., 1989, “Creation of a Knowledge-Based System for Urban Traffic Control: Methodology”, Recherche Transport Sécurité, English issue 5, June 1990 (translation from R.T.S., num. 22, 1989). Scemama G., 1992, “Developpement d’un Systeme a Base de Connaissance Historique pour la Gestion du Trafic”, Recherche Transport Sécurité, English issue 5, June 90 (translation from R.T.S., num. 22, 1989).