Trac-related Knowledge Acquired by Interaction with Experts ?

Abstract

We present our research on acquiring domain knowledge related to urban vehicular trac by means of interaction with experts. Such knowledge is needed in knowledge discovery and data mining for approximation of complex vague concepts from the road trac. According to perception based computing paradigm, this can be done by construction of hierarchical classiers supported with expert knowledge. We treat trac, especially urban trac, as a complex process having hierarchical structure. Complexity of this process makes trac data massive and complex, what makes domain oriented hierarchical classiers indispensable here. We propose a method of trac domain knowledge acquisition by interaction with experts aimed at construction of such classiers.

Keywords

• vehicular trac
• interaction with experts
• vague concepts
• knowledge discovery
• perception based computing
• hierarchical classiers

-

Vehicular trac is a vital phenomenon for the contemporary city. It has a signicant impact on environment and life of many people. Understanding the phenomenon and learning how to manage it are crucial tasks for functioning and development of the contemporary city. One of the main issues here is to learn from data knowledge about urban trac as a complex process. It involves learning detection of trac jams and recognition of trac congestion levels. In order to learn such knowledge we have to learn basic trac concepts such as e.g. trac jam , trac congestion , trac jam formation . But here we face two challenges. These concepts are complex vague concepts, thus they are hardly mathematically dened. Instead of that we can learn them from experts, i.e. acquiring from experts the relevant concepts and approximating them by urban trac data, i.e. lower level data which describe urban trac. Thus some form ? The research has been supported by grant 2011/01/D/ST6/06981 from the Polish National Science Centre and by grant 2011-3/12 from Homing Plus programme of Foundation for Polish Science, co-nanced from EU, Regional Development Fund. Authors would like to express their gratitude to Andrzej Skowron for his inspiration and ceaseless support during conducting research presented in the paper. of interaction with experts is needed. Elaborating an interaction method is the rst challenge faced in this paper. Urban trac data are good example of Big Data, they are massive and qualitatively complex what makes their processing computationally expensive. They become even more computationally expensive in the task of adaptive, autonomous, on-line control which is one of the main tasks in trac research. This trac control task gives the second challenge. We do not face this challenge directly in our paper. We discuss it in the light of perception based computing paradigm [33, 34] pointing out a possible solution: construction of hierarchical classiers supported by domain knowledge [ 3 ].

In this research, we focus on a single basic trac concept - trac congestion on a single crossroad - and we elaborate methods for approximating this concept from sensory data. Our sensory data come from simulating trac using the Trac Simulation Framework (TSF) software [1012]. Data from the software may slightly dier from real-world trac data (which are very dicult to obtain), but are conrmed to be quite realistic [ 13 ], enough to conduct our research.

The paper has the following organization. Section 2 describes past approaches to trac modeling, recent approaches based on probabilistic cellular automata and the model developed by P. Gora [ 10, 12 ] which was implemented in TSF software and is used in our research. Section 3 outlines our approach for acquisition of trac knowledge, motivates why we chose such approach and explains how such knowledge may be applied to approximating complex, high-level trafc concepts from sensory data, according to the perception based computing paradigm. Section 4 presents the design of our experiment, the procedure of dialogizing with domain experts and values of simulation parameters used in our experiment. Section 5 describes the evaluation of data obtained in the experiment and conclusions that we drawn based on analyzing acquired data and feedback from experts. Section 6 concludes the paper. 2

Modeling urban vehicular trac

Despite many years of extensive research, it is still dicult and challenging task to model the urban trac with satisfactory accuracy, either using standard mathematical tools or computer simulations. The proper model should take into account many factors, such as: modeling drivers behavior, Origin-Destination matrix, location and conguration of trac signals, the weather, road works etc. The situation is much simpler in case of a trac on highways, where the trac is only in one or two directions, there are no trac lights, the number of possible interactions between cars is relatively small.

There are three major classes of trac models: macroscopic, microscopic and mesoscopic. Macroscopic trac models treat all cars aggregately and describe relationships between trac congestion, trac density and average speed. Such models are usually based on analogy to other, well-known physical phenomena, such as uid dynamics [ 19, 18 ] or kinetic gas theory [27]. Some models were able to reproduce properties of trac on highways, e.g. the pioneering macroscopic model, Lighthill-Whitham model [ 19 ], is able to reproduce shockwaves. Some macroscopic models have been already applied in commercial products, such as PTV VISUM [40], and are applied to planning public transport, construction and development of roads, analyzing economic eciencies of transport solutions, modeling travel demand etc.

The next class of models are microscopic models which model drive of every single car. For instance, the Nagel-Schreckenberg model (Na-Sch model) is based on a probabilistic cellular automaton and is able to explain and reproduce spontaneous trac jam formation on highways [21, 32]. Space, time and speeds are discrete in this model, the road is divided into cells, which may be empty or occupied by at least one car. Transition rules, determining driver’s behavior, are dened by properly selected rules. The most fascinating thing about the model is that it is based on a very simple, natural transition rules, and is able to reproduce trac on highways with very good accuracy [21]. The model was broadly investigated and generalized, e.g. to simulate 2-lane trac [29] or simple crossroads [ 7 ]. Also, generalizations of the model were applied in real-world trac simulations systems, e.g. in the system Autobahn [ 2 ], which simulates and predicts the trac in Germany, and in the software Trac Simulation Framework (TSF), [ 10, 12 ], which we use in our research. Recently, there are developed much more advanced microscopic models which take into account drivers behavior (e.g. Intelligent Driver Model [36]).

There are also mesoscopic models which propagates packets of cars. Such models also give good results in case of modeling large-scale urban trac [ 5 ]. 2.1

Trac Simulation Framework model

The Na-Sch model was used by P. Gora to develop a new trac simulation model, the TSF model [ 10, 12 ]. The model extends the standard Na-Sch model and allows conducting simulations on a realistic road network, represented as a directed graph. Cars drive through the road network on edges, which consist of tapes (representing road lanes), divided into cells, as in the original model. Some vertices contain trac lights, which are objects characterized by location, duration of a red phase, duration of a green phase and oset.

The model takes into account many factors, e.g. driver’s prole, road’s prole, location and conguration of trac signals, distributions of start and destination points. Currently driver’s prole species aggressiveness of driver, which determines the maximal car’s speed on a given road, but the prole could be easily extended in the future (in fact, there already exist microscopic models which include much more details with respect to driver’s behavior, e.g. Intelligent Driver Model [36]). Road’s prole determines number of lanes and normal distribution of maximal speed of drivers, which, together with a driver’s prole, determines the maximal speed of a car on a given road [ 10, 12 ]. Distributions of start and destination points are dened on graph vertices and specify how to choose these points for driver’s route (distributions may be replaced by the Origin-Destination matrix, if this is available). After specifying distributions of start and destination points routes are calculated using the A* algorithm [ 6 ]. The move of every car is specied by transition rules of a cellular automaton being an extension of the standard Nagel-Schreckenberg model [ 21, 32, 10, 12 ]. 2.2

Trac Simulator Framework implementation

The TSF model was implemented in Trac Simulation Framework, advanced software for simulating and investigating vehicular trac in cities. TSF is being developed in C# and runs using .NET Framework platform, it employs maps from the OpenStreetMap project [22] and real trac data for Warsaw from Municipal Administration of Urban Roads in Warsaw [42]. Currently TSF is able to simulate realistic trac in Warsaw with more than 105 cars faster than realtime1 and it was conrmed by Warsaw citizens that the software can reproduce trac jams in the same places as they occur in reality. TSF possesses a multifunctional Graphical User Interface (GUI) and allows modifying map parameters and simulation parameters from the GUI level. Currently it is possible to edit locations and congurations of trac signals, distributions of start and destination points, parameters of dierent types of road network segments (e.g. parameters specifying normal distributions of the maximal speed on a given road segment). Also, it is possible to generate large number of routes for cars, according to given distributions of starting points and destination points.

TSF is still being developed, its functionality was described in details in papers [ 10, 12 ]. The software has been already used for generating data for the IEEE ICDM 2010 contest on trac prediction [ 13, 37 ]. The contest was sponsored by TomTom, held under the patronage of IEEE, ICDM conference and the President of Warsaw, Mrs. Hanna Gronkiewicz-Waltz. It was an important data mining event, which attracted 575 participating teams, which submitted in total almost 5000 solutions. Many of those solutions are interesting data mining algorithms for predicting trac congestion, average speeds and trac jams occurrences. The best solutions were published in proceeding from the ICDM 2010 conference [ 9 ], some of them are also elaborated in the TunedIT blog [38]. Data for the contest was released for the public use to enable post-challenge research, resulting in few more interesting algorithms [30, 43].

Currently TSF is also used for designing evolutionary algorithms (e.g. genetic algorithms [ 11 ]) for optimizing trac by conguring trac lights. It is also used by scientists from many countries in their research on trac modeling, analysis and prediction, e.g. [28, 30, 43]. The recent application of the software is acquisition of trac-related domain knowledge by interaction with experts, which is a topic of the paper and is described in details in next sections. 3

Trac knowledge acquisition - Perception Based Computing approach

This research is aimed at acquisition of trac domain knowledge by interaction with experts. Trac is a very complex phenomenon and many high level concepts related to that phenomenon are complex and vague (e.g. large trac congestion or formation of a trac jam ). These concepts are hardly mathematically dened and may also depend on many factors such as city, type of a crossroad etc. In some cases there are engineering approaches which try to dene such concepts precisely (e.g. levels of service aim to approximate trafc congestion levels [ 15 ]), but in fact these are just approximations and many spatio-temporal concepts related to urban trac dynamics are beyond the scope of precise, mathematical denitions, because of their vagueness. However, human brain can recognize such concepts much better than machines. Drivers and pedestrians participating in urban trac are able to recognize the trac situation quickly and eortlessly and can easily decide whether in a given trac situation there is a trac jam or not. It should be highlighted that in most cases, drivers and pedestrians are not transportation science experts or trac engineers. Contrary, they can be viewed as experts - practitioners . They are practitioners because they are trac agents, taking part and interacting each other in urban trac. The way in which they make their decisions or results of those decisions inuence urban trac as a complex hierarchically structured process. One of the main principles in perception based computing states that perception is action oriented. In the case of algorithm evaluation it means that algorithms should be tested on the basis of eciency of actions that are managed or controlled by those algorithms.

The long-term goal of our research is an optimization of trac control. Knowledge collected from experts during this research will be used for construction of classiers evaluating trac congestion and, among others, recognizing appearance of trac jams. And nally these classiers can be used for evaluation of trac control optimization algorithms as one of the possible ways, in addition to delay measuring. Therefore, the choice of drivers and pedestrians as experts transferring knowledge about trac is not arbitrary since they are agents interacting in the urban trac and both, they and their knowledge, are elements of the urban trac complex system.

According to PBC approach, for concept learning hierarchical classiers will be used. Hierarchical classiers, as other classiers, are decision algorithms that map objects to decisions [ 1, 4 ], but they are doing that in a hierarchical way. Objects could be described by low-level numerical or symbolical attributes. Decisions, in many cases, are vague, complex concepts, which are semantically distant from original low-level data. Hierarchical classiers could be viewed as tools that may be used to cover that distance by approximating complex, vague concepts, using low-level data. In such classiers the classication process goes from input data to decisions through at least few hierarchy levels, from lower data levels to higher, more abstract, complex concepts levels. Objects and/or attributes on higher levels are constructed based on objects and/or attributes from lower levels [33, 34]. This process may be supported by domain knowledge, given e.g. in the form of ontologies. To cover the semantic distance, training sets can be constructed with experts support. Decisions could be also complex, temporal or spatio-temporal objects as automated planning of complex objects behavior, e.g. safe driving through a crossroad or medical diagnosis, see [ 3 ].

As we mention above, in this research, expert knowledge is collected for construction of classiers approximating the trac congestion concepts by means of low-level trac data. Low-level data, such as number of cars, car’s position, current car’s speed, are taken from TSF trac simulator created by P. Gora [ 10, 12 ]. Data collected from simulations generally can be treated as results of measurements returned by logical sensors (see [41]). In the case of hierarchical classiers approximating trac concepts, such data create the rst, sensory level of hierarchical approximation. Objects and attributes from consecutive hierarchy levels can be constructed on the basis of objects and attributes from lower levels of hierarchy by means of information systems, decision tables and decision rules taken from the rough set theory [2325] as it was done in [ 3 ]. 4

Experiment

In perception based computing one of the main factors in hierarchical information processing is a way of granule setting or construction at every level of the hierarchy, starting from basic granules. In our research, as a basic granule in hierarchical organization of urban vehicular trac we picked up a single crossroad. Thus the main aim of this research is to learn conceptual levels of trac congestion and a concept of a trac jam on a single crossroad by means of a dialog with experts.

First we prepared 51 trac simulations corresponding to dierent trac situations close to the crossroad of streets Banacha, Grjecka, Bitwy Warszawskiej 1920 r. using the Trac Simulation Framework. The area under investigation is presented in the Figure 1, it is a place where large trac congestion occurs very often. Then we selected values of all important simulation parameters based on our past research and experiments, see Table 1. Name of the Description Value parameter NrOfCars Initial number of cars for a single trac situation 100, 1000 Step Duration of a single simulation step 1000 ms TimeGap Time after which new cars start their ride 1 step NewCars Number of cars starting ride after every TimeGap steps 5; 3; 1 Steps Duration of a single simulation 600 steps Acceleration Acceleration of cars per simulation step 10 km/h CrossroadPenalty Percentage of speed reduction before the crossroad 25% TurningPenalty Percentage of speed reduction during turning 50%

Every simulation lasted 10 minutes, values of some parameters were common for all situations, but situations diered in the following parameters:

1. initial number of cars, 2. number of new cars that start drive in each simulation step, 3. start and destination points distributions.

We prepared 5 dierent distributions of starting points and 5 dierent distributions of destination points. Distributions of starting points were named From East, From West, From North, From South, Uniform, distributions of destination points were named To East, To West, To North, To South, Uniform. It gives us 25 congurations of pairs: (start points distribution, destination points distribution). Names of distributions indicates where is the major concentration of start or destination points, respectively. The detailed description of these distributions and procedures for editing start points and destination points is described in the paper [ 10 ].

For every combination of pairs (start points distribution, destination points distribution) we still have few degrees of freedom that can be manipulated in order to produce dierent simulation scenarios. Some of these degrees of freedom correspond to parameters named in the rst column of the table 1: N rOf Cars, N ewCars, Acceleration, CrossroadP enalty, T urningP enalty. Other parameters may be related to the initial conguration of trac signals at the crossroad or maximal speed permissible on a given street. For our current research we needed only 51 simulation scenarios, so we decided to manipulate parameters N rOf Cars and N ewCars. 5 dierent start points distributions, 5 dierent destination points distributions and 3 dierent values of the N ewCars parameter gives us 5 5 3 = 125 possible simulation scenarios, from which we chose 48, assuming that N rOf Cars = 100. In addition, for N rOf Cars = 1000 we chose U nif orm distribution of start points and destination points and generated 3 more situations with 3 dierent values of N ewCars. It gives in total 51 trac situations that were later simulated using the TSF software.

Every simulation was recorded - Trac Simulation Framework logged information about positions and speeds of cars during the simulation and presented the same trac situation to experts using Graphical User Interface of our software. The following information was logged out to the output le:

Timestamp (simulation step), Car positions (link in the road network, position within the link, geographical longitude and latitude), Current car’s speed (in km/h).

Such information enabled reconstruction of the situation. We assumed that duration of a single cycle of trac lights is constant and lasts 2 minutes for every trac signal, so every 10-minutes long situation consisted of 5 parts, each of which lasted 2 minutes and corresponded to one cycle of trac lights. Thus we divided logs from our 10-minutes long simulations into 5 such parts (each lasted 2 minutes), to which we refer simply as trac cases . Totally, it gave us 255 trac cases.

Each of 51 situations was evaluated by domain experts and their task was to provide information about a trac state in the area close to the crossroad. In our case (vehicular trac in cities) a domain expert may be any person who has experience with the city trac, the most preferable should be drivers, which use road networks in Warsaw often and have to cope with trac jams. 1 of 51 situations was analyzed by all experts, while every situation from the rest 50 was analyzed by 3 experts, which gives 50 3 situation evaluations. Every expert analyzed 3 situations: 1 common to all experts and 2 taken from the rest 50. Therefore, we constructed 150 / 2 = 75 dierent tests, one for each expert, so we needed 75 experts. In every test each 10-minutes long situation was divided into 5 trac cases, so it was possible to show to domain experts 2-minutes long trac cases separately. Thus, every test consisted of 15 trac cases. Additionally, for each expert 2 trac cases from every situation were randomly selected to be presented and labeled by an expert twice, to check consistency of experts (they were not informed that some trac cases are repeated in a test). Therefore, every test consisted of 21 trac cases, which were presented to the expert in a random order as a short movie in the TSF’s GUI.

After presentation of a particular movie, TSF displayed the question: What was the trac congestion? . Experts answered the question with one of ve possible answers: Small, Medium, Large, Trac jam , I don’t know. The answer was given by experts using the window presented in the Figure 2. If the experts selected I don’t know response in the rst window, the system asked for selecting the closest options by displaying the window presented in the Figure 3. In the next step, the system asked experts for the response justication, which they provided in natural language using the text window with no limited number of signs. After selecting the proper answer and submitting justication, the next movie was presented to the expert.

Evaluation of decisions and properties extraction

Evaluation of expert decisions can be either expert-oriented or case-oriented. In the expert-oriented evaluation we check a consistency of decisions made by a given expert. In this case, the evaluated situation should be labeled by an expert (before evaluation) at least twice for checking stability of the expert’s decision making. In order to do that, from every situation two phases were selected to be labeled by an expert twice. In the case-oriented evaluation we will analyze how a given case (phase or situation) is labeled by dierent experts. For this purpose, every phase was labeled by three dierent experts. Their decisions will be used either to determine the nal aggregated decision, e.g. by voting, or to nd a uniformity of decisions about a given phase. It should be noted that our approach is only one of possible approaches and that decision evaluation itself is a novel and interesting issue and a topic for further research.

After providing trac congestion answer experts were asked to justify their choice in natural language, e.g. one of experts answered that the congestion was Small-Medium and gave explanation: Small trac, but later density on the main crossroad increased.

We analyzed all such answers from experts in order to acquire important properties that were used by experts to justify their answers. Those information are our domain knowledge, which may be used to construct the ontology of trac concepts and hierarchy of classiers approximating such concepts, according to the perception based computing paradigm. Table 2 presents all acquired properties and their values. It is worth to emphasize that among properties we also consider street (as a spatial property) and time (as a temporal property).

We analyzed collected data and obtained decisions regarding trac congestion obtained from the experiment for each trac situation. We also extracted properties used by experts to explain their choice and analyzed feedback from experts. From experiment we got the following conclusions:

There was too many test cases for one expert and the time of the experiment was too long for experts. More experts should be assigned to each situation (for now we had 3 experts for 250 situations, only 5 cases were evaluated by all 75 experts). It would be better if experts were informed about the progress of the experiment by displaying information Situation nr k (from n) before or after every test case. 6

Conclusions and future work

In the paper we propose an interactive method for acquisition of vehicular trac domain knowledge by dialog with experts. The direct aim of our research is to prepare training urban data set for classiers. Training urban trac data set is needed for construction of hierarchical classiers based on rough set methods [ 3, 23, 31 ] for approximating the concept of a trac jam on a single crossroad. Hierarchical classiers approximating trac concepts can be used to construct methods for intelligent, adaptive urban trac control as well as to evaluate them. Therefore construction of trac hierarchical classiers and designing intelligent, adaptive control algorithms for urban trac are long-term goals of our research. Analyzing results of the conducted experiment, we decided to design the second experiment in order to acquire better structured data asking experts to describe every simulated situation according to properties constructed in this research and presented in Table 2. In the further steps these properties can be used as intermediating level in the construction of hierarchical classiers. Our method will be evaluated rstly by evaluation of hierarchical classiers induced on the basis of training sets constructed using our method. According to perception based computing paradigm our method will be also evaluated in the process of trac optimization: optimization algorithms constructed using our classiers will be compared to other optimization algorithms which are not supported by expert knowledge. 21. K. Nagel and M. Schreckenberg, A cellular automaton model for freeway trac,

Journal de Physique I 2 (12), pp. 22212229, 1992. 22. OpenStreetMap, http://www.openstreetmap.org 23. Z. Pawlak, Rough sets: Theoretical aspects of reasoning about data , Dordrecht:

Kluwer, 1991. 24. Z. Pawlak, Rough sets, International Journal of Computing and Information Sciences, vol. 18, pp.341356, 1982. 25. Z. Pawlak, A. Skowron, Rudiments of rough sets, Information Sciences, vol. 177, pp. 327, 2007. 26. T. Poggio and S. Smale, The mathematics of learning: Dealing with data, Notices of the AMS, vol. 50, no. 5, pp. 537544, 2003. 27. I. Prigogine and R. Herman, Kinetic Theory of Vehicular Trac , Elsevier, Amsterdam, 1971. 28. T.M. Quang, M.A. Baharudin, E. Kamioka, Synergistic approaches to mobile intelligent transportation systems considering low penetration rate, Journal Pervasive and Mobile Computing, Vol. 10, February, 2014, pp. 187-202. 29. M. Rickert, K. Nagel, M. Schreckenberg, and A. Latour, Two Lane Trac Simulations using Cellular Automata, Physica A: Statistical Mechanics and its Applications, vol. 231, no. 4, pp. 534550, 1996. 30. J. Rzesztko, S.H. Nguyen, Machine Learning for Trac Prediction, in Fundamenta Informaticae, Vol. 119, No 3-4 / 2012, pp. 407-420. 31. J. Stepaniuk, Rough-Granular Computing in Knowledge Discovery and Data Mining, Springer Verlag, 2008. 32. A. Schadschneider, The Nagel-Schreckenberg model revisited, The European

Physical Journal B, vol. 10, no. 3, pp. 573582, 1999. 33. A. Skowron and P. Wasilewski, Information Systems in Modeling Interactive Computations on Granules, Theoretical Computer Science , vol. 412, pp. 5939-5959, 2011. 34. A. Skowron and P. Wasilewski, Interactive information systems: Toward perception based computing, Theoretical Computer Science , vol. 454, pp. 240-260, 2012. 35. A. Skowron, P. Wasilewski, Introduction to Perception Based Computing, in: S. Ramanna, L. C. Jain i R. J. Howlett (Eds.) Emerging Paradigms in Machine Learning, 2013, pp. 249-275, Springer Verlag. 36. M. Treiber, A. Hennecke, and D. Helbing, Congested trac states in empirical observations and microscopic simulations, 2000, in Physical Rev. E 62, pp. 18051824. 37. IEEE ICDM Contest: TomTom Trac Prediction for Intelligent GPS Navigation http://tunedit.org/challenge/IEEE-ICDM-2010. 38. TunedIT blog, http://blog.tunedit.org/2010/10/26/ieee-icdm-contest-topsolutions-1. 39. V. Vapnik. (2012) Learning Has Just Started, an interview with Prof. Vladimir

Vapnik by R. Gilad-Bachrach. [Online]. Available: http://learningtheory.org 40. PTV VISUM, http://vision-trac.ptvgroup.com/en-us/products/ptv-visum/. 41. P. Wegner, Why interactions is more powerful than algorithms, in Communications of ACM, vol. 40, pp. 81-91, 1997. 42. Municipal Administration of Urban Roads in Warsaw, http://www.zdm.waw.pl/zarzad-drog-miejskich/zadania-zdm/badania-ianalizy/informacje-o-natezeniu-ruchu-na-drogach-i-mostach. 43. N. Zarei, M.A. Ghayour, S. Hashemi, Road Trac Prediction Using ContextAware Random Forest Based on Volatility Nature of Trac Flows, in Intelligent Information and Database Systems LNCS, Vol. 7802, 2013, pp 196-205.

References

1. E. Alpaydin , Introduction to machine learning , MIT Press, 2004 .

2. Autobahn system in Germany, http://www.autobahn.nrw.de/index_e.html.

3. J. Bazan , Hierarchical classiers for complex spatio-temporal concepts . Transactions on Rough Sets IX, Lecture Notes in Computer Science , 5390 , 2008 , pp. 474 - 750 .

4. C. Bishop , Pattern Recognition and Machine Learning , Springer, 2006 .

5. M. Bliemer , M. Raadsen , E. de Romph, E.S. Smits , Requirements for Trac Assignment Models for Transport Planning: A Critical Assessment . Australasian Transport Research Forum 2013 Proceedings , 2013 .

6. T.H. Cormen , C.E. Leiserson , R.L. Rivest , C. Stein , An introduction to algorithms , (3rd edition) , MIT Press and McGraw-Hill , 2009 .

7. D. Chowdhury and A. Schadschneider , Self-organization of trac jams in cities: eects of stochastic dynamics and signal periods , in Physical Review E, vol. 59 , no. 2 , pp. 1311 - 1314 , 1999 .

8. W. Dong and A. Pentland , A network analysis of road trac with vehicle tracking data , in AAAI Spring Symposium: Human Behavior Modeling , 2009 , pp. 7 - 12 .

9. W. Fan , W. Hsu , G.I. Webb , B. Liu , Ch. Zhang, D. Gunopulos , X. Wu (Eds.): Proceedings of the 10th IEEE International Conference on Data Mining Workshops, IEEE ICDMW 2010 , IEEE Computer Society 2010 .

10. P. Gora, Trac Simulation Framework - a cellular automaton-based tool for simulating and investigating real city trac, in M. A. K“opotek , A. Przepirkowski , S. T. Wierzcho«, K. Trojanowski (Eds.), in: Recent Advances in Intelligent Information Systems , 2009 , pp. 641 - 653 .

11. P. Gora , A genetic algorithm approach to optimization of vehicular trac in cities by means of conguring trac lights , in: Dominik Ry»ko, Henryk Rybi«ski, Piotr Gawrysiak, Marzena Kryszkiewicz, in: Emergent Intelligent Technologies in Industry. Studies in Computational Intelligence , Vol. 369 , 2011 , pp. 1 - 10 .

12. P. Gora, Trac Simulation Framework, in: 14th International Conference on Modelling and Simulation , 2012 , pp. 345349 .

13. P. Gora , S. H. Nguyen , M. Szczuka , J. wietlicka, M. Wojnarski, and D. Zeinalipour , IEEE ICDM 2010 Contest TomTom Trac Prediction for Intelligent GPS Navigation , in [9] , 2010 , pp. 13721376 .

14. P. Gora, P. Wasilewski , Inducing Models of Vehicular Trac Complex Vague Concepts by Interaction with Domain Experts, 2013 , in COGNITIVE 2013, The Fifth International Conference on Advanced Cognitive Technologies and Applications , pp. 120 - 125 .

15. P. Ryus, Highway Capacity Manual 2010 , Transportation Research Board, 2012 .

16. D. Helbing , M. Treiber , A. Kesting and M. Schonhof: Theoretical vs. Empirical Classication and Prediction of Congested Trac States ., The European Physical Journal B , Vol. 69 , Issue 4, pp. 583 - 598 , 2009 .

17. A. Jankowski and A. Skowron , Wisdom technology: A rough-granular approach , in M. Marciniak, A . Mykowiecka (eds.) Bolc Festschrift, Lectures Notes in Computer Science, 5070 , Springer Verlag, 2009 , pp. 341 .

18. Ch . Lustri, Continuum Modelling of Trac Flow , 2010 .

19. M.J. Lighthill and G.B. Whitham , On kinematic waves. II. A theory of trac ow on long crowded roads , in: Proc. of the Royal Society A , 229 , pp. 317 - 245 , 1955 .

20. J. Leskovec , A. Rajaraman and J.D. Ullman , Mining of Massive Datasets , Cambridge University Press, 2013 .