Schriftenreihe des Instituts für Angewandte Informatik / Automatisierungstechnik Karlsruher Institut für Technologie Band 40 F. Hoffmann, E. Hüllermeier (Hrsg.) Proceedings 21. Workshop Computational Intelligence Dortmund, 1. - 2. Dezember 2011 # ! !#! %$()! !#%!!#$ Graph !"" $((% (!$$''&#! ""( # $*!#"$# Lernalgorithmus $" !(!$ #"!# !$#!$''&&"# $"#! $! #'# $ #!$#$! &" F. Hoffmann, E. Hüllermeier (Hrsg.) Proceedings 21. Workshop Computational Intelligence Dortmund, 1. - 2. Dezember 2011 Schriftenreihe des Instituts für Angewandte Informatik / Automatisierungstechnik am Karlsruher Institut für Technologie Band 40 Eine Übersicht über alle bisher in dieser Schriftenreihe erschienenen Bände finden Sie am Ende des Buchs. Proceedings 21. Workshop Computational Intelligence Dortmund, 1. - 2. Dezember 2011 F. Hoffmann E. Hüllermeier (Hrsg.) Impressum Karlsruher Institut für Technologie (KIT) KIT Scientific Publishing Straße am Forum 2 D-76131 Karlsruhe www.ksp.kit.edu KIT – Universität des Landes Baden-Württemberg und nationales Forschungszentrum in der Helmholtz-Gemeinschaft Diese Veröffentlichung ist im Internet unter folgender Creative Commons-Lizenz publiziert: http://creativecommons.org/licenses/by-nc-nd/3.0/de/ KIT Scientific Publishing 2011 Print on Demand ISSN: 1614-5267 ISBN: 978-3-86644-743-1 Inhaltsverzeichnis P. Held, R. Kruse 1 (Universität Magdeburg) Estimation of hidden driver properties based on the driving behavior Ch. Kühnert, L. Gröll, M. Heizmann, R. Mikut 15 (Fraunhofer IOSB, Karlsruher Institut für Technologie) Ansätze zur datengetriebenen Formulierung von Strukturhypothesen für dynamische Systeme F. Langer, K. Bertulies 31 (Fraunhofer ESK) Self Learning Anomaly Detection for Embedded Safety Critical Software A. Schrodt, A. Kroll 47 (Universität Kassel) Zur Fuzzy-Clusterungs-basierten Identifikation eines Verbrennungsmotor- kennfelds: Methodenvergleich und Parametrierungsstrategie J. Walkenhorst, T. Bertram 61 (Technische Universität Dortmund) Multikriterielle Optimierungsverfahren für Pickup-and-Delivery-Probleme St. Moses, C. Gühmann, J. Jäkel 77 (Volkswagen AG, Technische Universität Berlin, HTWK Leipzig) Optimierung von Fahrzeugkonzepten in der frühen Entwicklungsphase mit Hilfe Genetischer Algorithmen und Künstlicher Neuronaler Netze K. Gerstl, G. Rudolph, O. Schütze, H. Trautmann 93 (Technische Universität Dortmund) Gleichmäßige Paretofront-Approximationen für mehrkriterielle Kontrollprobleme unter Verwendung des gemittelten Hausdorff-Maßes Th. Fober, W. Cheng, E. Hüllermeier 107 (Universität Marburg) Focusing Search in Multiobjective Evolutionary Optimization through Preference Learning from User Feedback M. Friese, M. Zaefferer, Th. Bartz-Beielstein, O. Flasch, P. Koch, 119 W. Konen, B. Naujoks (Fachhochschule Köln) Ensemble Based Optimization and Tuning Algorithms S. Schäfer, U. Lehmann, M. Schneider, M. Stieglitz, C. Radisch, 135 S. Turck, J. Krone, J. Brenig (Fachhochschule Südwestfalen) Optimizing the training of artificial neural networks using evolutionary algorithms P. Koch, W. Konen, O. Flasch, M. Friese, B. Naujoks, M. Zaefferer, 147 Th. Bartz-Beielstein (Fachhochschule Köln) Tuned Data Mining in R M. Stieglitz, U. Lehmann, M. Schneider, F. Calcagno, S. Schäfer, 161 S. Buhl, J. Brenig, J. Wiggenbrock, J. Willms (Fachhochschule Südwestfalen) Beschleunigung des Backpropagation Algorithmus mit CUDA T. Cui, A. Grumpe, M. Hillebrand, U. Kreßel, C.Wöhler 171 (Technische Universität Dortmund) Analytically tractable sample-specific confidence measures for semi-supervised learning A. S. Phung, J. Malzahn, F. Hoffmann, T. Bertram 187 (Technische Universität Dortmund) Datenbasierte Modellierung der Kinematik eines dreigliedrigen elastischen Roboterarms H. Schulte, M. Zajac 203 (Hochschule für Technik und Wirtschaft Berlin) Modellgestützte Diagnose von Sensorfehlern in Rotorblatt-Verstellsystemen von Windenergieanlagen M. Bauerdick, S. Hafner 217 (Fachhochschule Südwestfalen) Adaptive Neural Temperature Control for Coffee Machines D. Fiß, M. Wagenknecht, R. Hampel 233 (Hochschule Zittau/Görlitz) Dynamische Simulation von Siedeprozessen S. Flachs, Ch. Arnold, S. Lambeck 245 (Hochschule Fulda) Webbasierte Diagnose und Entscheidungshilfe zum Klimamanagement in der präventiven Konservierung S. Kittan, W. Kästner 257 (Hochschule Zittau/Görlitz) Analyse des Potentials von Multi-Agenten-Simulation und Zellulären Automaten zur Nachbildung der Anlagerungsprozesse von Isolationsmaterial an Rückhaltevorrichtungen K. Voth, A. Dicks, M. Sasse, K. Becker, V. Lohweg 269 (Hochschule Ostwestfalen-Lippe) Konfliktlösende Informationsfusion zur Maschinendiagnose am Beispiel von Extrusionsanlagen Th. Runkler 283 (Siemens AG) Learning Takagi Sugeno Fuzzy Systems Using Robust Statistics H. Schulte, P. Schubert 295 (Hochschule für Technik und Wirtschaft Berlin) Nichtlineare Folgeregelung eines pneumatischen Schwingungsprüfstands mit Takagi-Sugeno Fuzzy Systemen W. Jakob, S. Strack, G. Bengel, A. Quinte, K. Stucky, W. Süß 301 (Karlsruher Institut für Technologie, Institut für Angewandte Informatik, Karlsruhe) Vergleich Memetischer Algorithmen für das schnelle Rescheduling von Gridjobs Estimation of hidden driver properties based on the driving behavior Pascal Held Rudolf Kruse Otto-von-Guericke University of Magdeburg Faculty of Computer Science Department of Knowledge Processing and Language Engineering Universitätsplatz 2, D-39106 Magdeburg Tel.: +49 391 67 18718 Fax: +49 391 67 12018 E-Mail: {pheld,kruse}@iws.cs.uni-magdeburg.de Abstract Given a better knowledge about the driver the car’s driver assistance systems could be specifically adapted, potentially increasing the systems’ acceptance. Our purpose is to estimate hidden driver properties underlying the driving behavior. We used a linear combination of different quantities as basis of the similarity calculation. With the aid of an evolutionary algorithm we determined these quantities as well as their weights. In a study with 26 subjects three hidden target properties were examined and estimated. We reached a correlation of approximately 0.7 between the predicted values and given measurements. In order to exclude overfitting of the model, an additional optimization was carried out using random values. In this case only a considerably smaller correlation of 0.36 could be achieved. As a consequence we observed that the basic assumption applies at least to the three examined target properties. We conclude that an estimation based on driving behavior is possible. 1 Introduction In modern vehicles the incorporation of assistance systems is constantly increasing. Their goals are, among other things, to reduce accident consequences or completely avoid them altogether. To do so, the system activates the brake for example. The greatest challenge these systems face is the decision when to intervene. An early intervention can increase the efficiency of the system. This results, however, in drivers more frequently perceiving the activation as necessary. Given a better knowledge about the driver, the systems could be specifically adapted, potentially increasing the systems’ acceptance. Such an estima- tion of the driver is the goal our work. The main idea how to estimate from given data is explained in Section 2. Our purpose is to estimate hidden driver properties underlying the driving behavior. This estimate is based on the assumption that drivers with similar properties behave similarly concerning certain quantities. To select which properties to use, we developed an evolutionary algorithm, which is descriped in Section 3. We tested our algorithm with different data sets in Section 4. The results obtained could be used to for example parameterize assistance systems. A similar approach was presented from Bauer et al. in [1]. Q prrqvt! X xu8hvhyDryyvtrpr9 q ! !! 2 Methods 2.1 Used Data We used data from two different sources for our analysis. The first dataset we examined was recorded from a driving simulator, where 21 test persons passed both a city and a country road scenario. Additionally one target property was determined for every person. As a second source we used a field study with 26 test persons. Every journey was split into three parts, i.e. freeway, country road, city scenario. For every test person three target properties were determined [2]. While driving we collected different driving parameters every 100 ms. These parameters were distance to the vehicle driving ahead, acceleration, gas pedal value, velocity, v and cross acceleration. The distance was detected with the help of a radar system. The other parameters were directly derived from the car’s CAN bus. We transformed this distance into three mea- sures, i.e. the absolute distance dabs in meter, the time gap Δt in seconds, and the time to collision TTC in seconds. [3] d Δt = abs (1) vown d TTC = abs (2) vown − vahead During the whole driving session we can distinguish the following situations: free driving (no car ahead), driving behind a car, free stopping (no car ahead), stopping behind a car, free start (no car ahead), starting behind a car and global behavior. Stopping processes were detected by velocity analysis. 2.2 Generating Parameters The question is how to generate features from collected data. One requirement is that there should be a cumulative evaluation of the previous driving session. We used statistical measures such as mean value, variance, and percentage above a given threshold. These are measures which are easy to update online. In our work we used about 40 different characteristic values from all