Intelligent Systems Lab Amsterdam


Please see the new ISLA website at: http://isla.science.uva.nl

Current Location

  1. 20102011.Afstudeerproject BSc Kunstmatige Intelligentie
  2. Project proposals

Project proposals

Content

  • Usage explanation for this folder (please read and act acccordingly)

    Status of the projects
    • The projects below shown in RED, have not yet been updated by their owners for this year.
    • The projects below shown in GREEN, are available for this year!
    What to do (Teachers):
    • Please add you projects below, using the 'item' option via the top-left pulldown menu. Make you sure to color the project title green.
    • Please modify 'old' projects for re-use in this year (via edit next to the project title).  Make you sure to color the project title green.
    What to do (Students):
    • Negotiate with teachers to obtain a project, and inform me (B.Bredeweg@uva.nl) when arrangements have been made
    • Define your own project, but this requires that you find a teachers who is willing to supervise the project (preferred option)
    • Ask me (B.Bredeweg@uva.nl) to assign you to a project (you have to take the project that will be assigned to you)

  • TweetereTweetereTweetereTweet

    supervisor Maarten Marx
    abstract No entity without identity is een beroemd dictum van Quine. Hij bedoelt hiermee dat je geen objecten moet introduceren als je niet weet hoe je kunt bepalen of twee instanties van dat objecten wel hetzelfde zijn.
    Dit lijkt misschien filosofisch geneuzel en helemaal niet practisch, maar er zijn genoeg voorbeelden in de wetenschap waarbij er veel gesproken wordt over objecten, maar men niet goed weet hoe identiteit te bepalen.
    De tweet is een goed voorbeeld. De retweet is een erg belangrijk onderdeel van Twitter en voor een groot deel verantwoordelijk voor explosieve verspreiding van informatie. Ook komt er veel informatie via retweets terecht bij mensen die nooit naar de originele bron hebben gekeken/geluisterd.
    Maar wanneer is een tweet eigenlijk een retweet, en zo ja, wat is zijn "origineel"?
    Deze vraag beantwoorden we in dit project. Wij hebben een collectie van 20M tweets van politici of over politiek, opgehaald sinds zomer vorig jaar.

    Nodig:     machine leren, omgaan met heel veel data. Statistiek. Om kunnen gaan met ruizige data.
    Literatuur:     stukken over Twitter op SIGIR, WSDM, ICWSM conferenties.

  • Media landschap van Twitter

    Supervisors: Maarten Marx
    Student
    Abstract:     We willen de volgende studie die gedaan is voor de VS repliceren in Nederland:
    We present a preliminary but groundbreaking study of the media landscape of Twitter. We use public data on whom follows who to uncover common behaviour in media consumption, the relationship between various classes of media, and the diversity of media content which social links may bring. Our analysis shows that there is a non-negligible amount of indirect media exposure, either through friends who follow particular media sources, or via retweeted messages. We show that the indirect media exposure expands the political diversity of news to which users are exposed to a surprising extent, increasing the range by between 60-98%. These results are valuable because they have not been readily available to traditional media, and they can help predict how we will read news, and how publishers will interact with us in the future. Zie http://talks.cam.ac.uk/talk/index/27945

    Nodig: Statistiek, Sociale netwerk theorie, Om kunnen gaan met heel veel data. Alle OW Linux machines tegelijk aanzetten om Twitter te crawlen. Structuur herkennen in noisy data.
    Uitkomst: Bij mooie resultaten is ons een artikel in NRC Next toegezegd.
    Literatuur: Het artikel behorend bij bovenstaande abstract is onder review, maar in bezit van de supervisor. Artikelen over Twitter netwerken in de ICWSM conferenties.

  • Elsevier in de Kamer, de Kamer in Elsevier

    Supervisor: Maarten Marx
    Student(s):
    Abstract:

    Elsevier bestaat in haar huidige vorm 65 jaar. Zowel Elsevier als de Handelingen der Staten Generaal zijn digitaal beschikbaar. Er wordt in beide bronnen veel over de ander "gesproken". Dit maakt deze twee archieven uitermate geschikt voor een Linked Data project. In dit project gaan we deze twee archieven aan elkaar koppelen. We willen door middel van machine leer technieken computers leren om een verwijzing naar een Elsevier artikel in de Handelingen om te zetten in een hyperlink naar het bewuste artikel in het Elsevier archief. Andersom willen we verwijzingen naar gebeurtenissen in het parlement die voorkomen in Elsevier voorzien van een hyperlink naar het verslag van die gebeurtenis in de Handelingen. Als we eenmaal de koppeling hebben kunnen we allerlei leuke vragen automatisch vanuit de databank laten beantwoorden. Wat voorbeelden:
    • Welke politici en partijen verwijzen veel naar Elsevier?
    • Op welke manier doen ze dat? In een betoog, in een interruptie, als bron voor een Kamervraag, etc? Hoe zijn de verwijzingen verdeelt over de beide Kamers?
    • Wat voor soort onderwerpen behandelen de stukken waarin Elsevier wordt aangehaald?
    • Over welke partijen en politici en over welke onderwerpen bericht Elsevier het meest?
    • ...

    Elsevier wordt volgens https://zoek.officielebekendmakingen.nl in 1152 parlementaire documenten genoemd. Dit bestand bevat de documenten sinds 1995. Op statengeneraaldigitaal.nl vinden we 483 hits voor de periode 1814-1995. Dit is hoogstwaarschijnlijk een grote onderschatting van het aantal voorkomens door 1) OCR fouten, en 2) het feit dat hier documenten veel groter/langer zijn dan bij https://zoek.officielebekendmakingen.nl.

    Nodig:  Data Mining, Machine leren, Text extractie.  Wens om iets te maken wat ook echt gebruikt zal worden.

    Resultaat: Uitkomsten worden gepubliceerd in Elsevier. De koppeling wordt toegevoegd aan Elsevier archief en aan de Handelingen bij de KoninKlijke Bibliotheek.

    Literatuur:   

  • Conceptions of causality

    Supervisor: Wouter Beek and Bert Bredeweg
    Student(s):
    Abstract: This research focuses on the various conceptions of causality that have been construed in the past. In classical and Medieval, i.e. Aristotelian, physics causality is a (quadripartite) ontological primitive. [1, e.g. 1013a] Causality is the central concept of the scientific description of the world. But this changes during the modern period. Under the influence of the work of Francis Bacon and, most influentially, David Hume [2] the concept of causality got dispersed from scientific discourse. Gradually other ontological primitives such as forces (e.g. gravity) have taken causality's place within scientific theorizing.

    We perform research in Qualitative Reasoning (QR), which is formalism that allows a description of the physical world in which causality takes, again, center stage. In QR causality is explicitly represented. [3] What is more, in QR causality is taken to be an ontological principle (though not with the same properties as in Aristotelian physics). We want to investigate the fundamental principles of causality as it is treated in QR.
    This investigation is not merely theoretic, but also computational, since we have a modeling environment in which models of causal systems can be simulated. [4] This modeling environment can be used to support or repudiate philosophical theories on causality, since the simulation results must be in line with how reality behaves. We are interested in the following research questions:
    • What has been, is, or could be causality's role in describing man's everyday interaction with the physical world?
    • What has been, is, or could be causality's role in representing the physical world in scientific theorizing?
    References:

  • It's the journey, not the destination: Mobile Recommender Systems

    Attached Files:
    Supervisor: Abdo El Ali

    Abstract
    In this project, you will build on current research aimed at developing a mobile recommender system that can find, personalize, and present routes, places, and possibly itineraries to pedestrians exploring a city (like Amsterdam or Barcelona). This has brought efforts to develop the PRIMER (Places, Routes  & Itineraries Mobile Exploratory Recommender) system    . The vision behind this is that people, when visiting a new city or exploring a familiar one, would sometimes like to get away from the tyranny of tour guides (e.g., Lonely Planet) and instead get a social and more local interpretation of the places and paths in the city. With the flood of social media content (e.g., Facebook, Twitter, Youtube, Picasa, etc.) nowadays, we can peek into the wisdom of the crowd and make use of that information to make real-time social recommendations to users of where to go, which paths to take, and inform who said what about a place. The need for this kind of work has even been recently highlighted in the Volkskrant [1]. To realize this vision, currently we make use of social media APIs (at present FlickR, but extensions on the way) to compute and present these social recommendations. Research in this direction includes [2][3].

    Taking first steps towards realizing the vision behind the PRIMER system, a basic system [4] that makes use of FlickR data (users, photos, tags, timestamps) has been developed that creates and presents popular routes on a given day using sequence alignment methods [5]. Currently, the system is being extended to provide personalized recommendations of places, using content-based collaborative approaches. There are a number of AI research areas that can significantly expand on this work:
     
    1) Prediction & Forecast
    2) Social Network Analysis
    3) Group Modelling

     If interested, you can read more about the specific projects in the attached pdf.

    References
    [1]http://www.volkskrant.nl/vk/nl/2702/Reizen/article/detail/1832650/2011/02/08/Je-smartphone-als-reisgids.dhtml
    [2] De Choudhury, Munmun and Feldman, Moran and Amer-Yahia, Sihem and Golbandi, Nadav and Lempel, Ronny and Yu, Cong (2010). Constructing Travel Itineraries from Tagged Geo-Temporal Breadcrumbs. In HT '10: Proceedings of the 21st ACM conference on Hypertext and hypermedia, p.35-44, ACM.
    [3] Maarten Clements, Pavel Serdyukov, Arjen P. de Vries, Marcel J. T. Reinders. Finding Wormholes with Flickr Geotags. In Proceedings of ECIR'2010, pp.658~661.
    [4] Sicco van Sas (2010). Generating route recommendations using geotagged Flickr photos and sequence alignment. BSc Thesis Artificial Intelligence. [sent upon request]
    [5] Shoval, N. and Isaacson, M (2007) Sequence Alignment as a Method for Human Activity Analysis, Annals of the Association of American Geographers vol. 97, 2: 282-297.

    Contact information
    Abdo El Ali
    e: a.elali@uva.nl
    t: 020 525 8661
    w:     http://staff.science.uva.nl/~elali/  

  • Automatic modelling of personality types

    Supervisor: Reinhard Blutner
    Student(s):
    Abstract:     Personality diagnostics has important application in the working world and practical live (personal & career development, leadership, teamwork, team building, workplace diversity and business management, marriage counseling, etc). The idea of psychological types originated in the theoretical work of C.G. Jung. Presently, there are two main approaches: (a) Myers-Briggs Type Indicator (MBTI): The corresponding assessment method is a psychometric questionnaire designed to measure psychological preferences in how people perceive the world and make decisions; (b) Singer-Loomis type development instrument (SLTDI): The central idea is that information is divided into eight different information elements, in compliance with the eight Jungian functions.
    Recently, Jungian personality theories were formalized by introducing spin operators into cognitive science. Further, the role of ordering effects (the frequencies of the answers are dependent on the ordering of he questions) was highlighted in a paper by Wang & Busemeyer (under review). Whereas classical model have problems to predict question order effects, the new models based on linear algebra and projection operators can predict such effects. 
    For realizing the project no background in personality diagnostics is required. However, basic know-ledge of linear algebra is required. Among the possible issues that can be investigated are the following:
    • Based on a given questionnaire both methods of personality classification (MBTI, SLTDI) should be implemented in a Java environment.
    • Building a small data base of anonymised individuals augmented by a list of famous personalities capturing all personality types.
    • Predicting question order effects. Discussing the role of order effects for finding a reliable answer pattern. 
    • Comparison with alternative models such as the “ big five”. 
    • Model fitting by using the 2-qubit model of personality.  Evaluating the quality of the fit. Note: the 2- qubit model allows for two different solutions given an MBTI questionnaire.
    This could be a concrete working plan for 2011:
    1. Model fitting by using the 2 qubit model and data from the MBTI.
    2. Handling order effects. Taking order effects into account when fitting the 2 qubit model.
    3. Do the data conform to the inner restrictions of the model? Evaluating the quality of the fit.
    4. Discussing the results. Comparison with alternative models (“big five”)
    References
    • Stevens, A. (1994). Jung - A very short Introduction. Oxford: Oxford University Press. 
    • Blutner, R. & E. Hochnadel (2008). Two Qubits for C.G. Jung's Theory of Personality. Available from http://www.quantum-cognition.de/Jung.pdf. 
    • Wang, Z. & Busemeyer, J. R. (2011). Explaining and Predicting Question Order Effects Using a Quantum Probability Model. Under review.
     

  • Qualitative Reasoning as a Constraint Programming Puzzle

    Supervisor: Jochem Liem and Bert Bredeweg
    Student(s):
    Abstract: In science, models are used to capture the important aspects of complex situations. By making use of Qualitative Reasoning (QR), simulations can be run on those models in order to generate prospective future behaviour. For example, a model of climate change may be used to plot the effects that rising temperatures will have on the melting of the polar caps.

    Constraint Satisfaction Problems, or CSPs, are mathematical problems that are defined as sets of objects whose state must satisfy a number of constraints or limitations. In this project, you will treat of simulations that are performed on models as if they were Constraint Satisfaction Problems.

    More specifically, Qualitative Models consist of quantities (e.g. amount_of_rain, or population_size) and relations (e.g. X_makes_Y_increase). You will translate:
    1. the quantities to variables,
    2. the values of the quantities to the domains of those variables, and
    3. the relationships to constraints. As a result of this, the Qualitative Reasoning paradigm will be reformulated in terms of CSP.
    This reformulation of the simulation problem in terms of CSP will give us more insight in theoretical underpinnings of model-based simulation. Also, the results of the new CSP approach allow a comparison with the reasoning results of the existing QR method. 

  • Automated modelling of conceptual knowledge

    Supervisor: Jochem Liem and Bert Bredeweg and Wouter Beek
    Student(s):
    Abstract: Ecology is an important but complicated field of research. Domain experts try to grasp the complexities of ecological behaviour by creating models. Such models play an important role in the acquirement of scientific knowledge, since they allow simulations to be run, so that predictions can be made (e.g. those regarding the future effects of global warming).

    One of the aims of Qualitative Reasoning (QR) is to capture the conceptual knowledge domain experts have of systems, particularly concerning causality, in a way that explanations about the behaviour of these systems are represented [1]. Garp3 is a workbench that is used by domain experts to develop such models and capture their insights on environmental science (e.g. [2,3,4]).

    In two earlier bachelor projects Hylke Buisman and Carsten van Weelden made important contributions to automating conceptual modelling. We would like to take this work a step further and focus on learning 'condition model ingredients', that is model statements that hold only under specific conditions (usually in/equality statements A=B, C>D+E, etc.). The software and ML techniques are implemented in Prolog. Extensions should also be implemented in Prolog.

    Via http://www.garp3.org you can run a movie to learn about how domain experts use the software. Since creating models is a time-consuming task and domain experts are relatively scarce, the result of this project will be both scientifically significant and practically important.

    References

    1. Cioaca, E., Linnebank, F.E., Bredeweg, B. and Salles, P., 2009. A qualitative reasoning model of algal bloom in the Danube Delta Biosphere Reserve (DDBR). Ecological Informatics, 4(5-6), 282-298.
    2. Symone C.S. Araújo, Paulo Salles, Carlos H. Saito. 2008. A case study on qualitative model evaluation using data about river water quality. Ecological Informatics, (3)1: 1-12.
    3. Salles, P., Bredeweg, B. and Bensusan, N. 2006. The ants' garden: Qualitative models of complex interactions between populations. Ecological Modelling, 194(1-3): 90-101.
    4. Tullos, D.D. and Michael Neumann, M. 2006. A qualitative model for analyzing the effects of anthropogenic activities in the watershed on benthic macroinvertebrate communities. Ecological Modelling, 196(1-2): 209-220.

  • Zoeksysteem voor Parlementaire documenten

    Supervisor: Maarten Marx, Anne Schuth
    Student(s):
    Abstract: Bekijk sectie 3 van dit artikel. Ga ook eens zoeken op de SGD site. Ben je het eens met de analyse die we daar geven? Wij hebben alle data die de SGD heeft ook op onze eigen machines in een XML zoeksysteem staan. Maar we hebben nog veel meer informatie dan zij: bijvoorbeeld van elk woord weten we wie het gezegd heeft.

    Maak daarmee een verbeterde versie van het SGD zoeksysteem. Een voor de hand liggende uitbreiding is die met facetten, zoals je ook hebt bij Marktplaats en LinkedIn.

    Mits goed werkend bestaat de mogelijkheid om het systeem in Den Haag aan de mensen van de Koninklijke Bibliotheek  te laten zien. Zij zijn hun website aan het vernieuwen en kunnen dit dan als voorbeeld gebruiken.

    Nodig: Affiniteit met Databases, information retrieval, XML en XPath en met het maken van gebruikers-interfaces. Wens om iets te maken wat ook echt gebruikt zal worden.

    Literatuur: http://searchuserinterfaces.com by Matti Hearst

  • Mobtwit - Mobile social storytelling with Twitter tweets

    Supervisor: Frank Nack
    Student(s):
    Abstract: Until recently we could access a location’s memory mainly through media surrogates, such as books, drawings, film or audio files, or through face-to-face encounters with people who were able to knit us into the rich but hidden experience fabric of a place. We are now able to leave traces of our emotional or intellectual experience as virtual attachments to any location. As a result we expect that any place, indoors or outdoors, reveals itself to us by confronting us with connection, context, and uncommon perspectives.

    Goal
    • Utilize natural language processing and event modeling techniques to combine geo-tagged twitter feeds into a story. This can be a general story or based on social collaboration in a group of friends. The story needs to be presented on a mobile phone. The development environment is Android mobile platform.
    Research question:
    1. What are general event characteristics of a story?
    2. How can a story automatically be generated in real time based on existing and non-annotated textual data?

  • Query completion voor televisiemakers

    Supervisor: Maarten de Rijke en Edgar Meij
    Student(s):
    Abstract: Vergelijk, test en verbeter algoritmes om automatisch zogenaamde voltooiingen (completions) van zoekvragen te genereren. Je kunt hiervoor gebruik maken van achtergrondkennis zoals vastgelegd in een thesaurus en van de informatie die is vastgelegd in query log files. In log files van zoekmachines wordt het zoekgedrag van gebruikers van de zoekmachine vastgelegd: waar ze naar zoeken, welke resultaten bekeken zijn, hoe de zoekvraag aangepast is, etc. Uit dergelijke digitale sporen kan nuttige kennis worden verkregen, bijvoorbeeld voor het voltooiien van zoekvragen.

    Dit is een mengeling van theoretisch en experimenteel werk. De data (thesaurus, collectie en log files) voor dit project is (onder licentie) beschikbaar gemaakt door het Nederlands Instituut voor Beeld en Geluid.

    Vereist: machine leren, data mining, omgaan met ruizige data.

    Wat achtergrond: http://en.wikipedia.org/wiki/Autocomplete
     

  • Ontology elements visualization in Collaborative Networks

    Supervisor: Hamideh Afsarmanesh and Ekaterina ErmilovaStudent(s):
    Abstract: One purpose for development and management of ontologies is sharing of common understanding about the information among people and software agents within a specific environment [1]. In Collaborative Networks environments, ontologies play an important role [2]. A Collaborative Network (CN) represents an alliance constituted by a variety of entities (e.g., organizations and people) that are largely autonomous, geographically distributed, and heterogeneous in terms of their operating environment, culture, social capital, and goals, but that collaborate to better achieve common or compatible goals, and whose interactions are supported by computer network [3]. Establishment of common understanding on the information accumulated in CNs among the wide variety of CN’s stakeholders is both crucial and challenging task, which needs to be supported by ontologies.
     
    This student project aims primarily at development and implementation of a web-application addressing functionalities and a graphical interface for viewing of the already developed (in the COLNET group) CN’s ontology by its stakeholders. The interface for the ontology needs to represent a force directed and hyperbolic graph [4]. The project requires research on both Collaborative Networks and ontological knowledge representation. The main skills required from the students include knowing Java programming, OWL, and CSS.

    References:
    1. Noy, N.F., McGuinness, D.L. (2001). Ontology Development 101: A Guide to Creating Your First Ontology, Stanford University, Stanford, CA, 94305.
    2. [2] Camarinha-Matos, L.M., Afsarmanesh, H. (2008). Concept of Collaboration. In the Encyclopedia of Networked and Virtual Organizations (G. Putnik, M. M. Cunha, Eds.), Idea Group
    3. Ermilova, E., Afsarmanesh, H. (2008): A unified ontology for VO Breeding Environments. In Proc. of DHMS'08 conference, Athens, Greece. March 9-12, 2008.
    4. Visual WordNet

  • Development of mobile, adaptive interfaces

    Supervisor: Vanessa Evers and Andi Winterboer
    Student(s):Abstract: The EC FP7 project Diadem deals with human interaction with distributed intelligent networks through mobile phones. The domain is environmental monitoring and the early detection of chemical incidents in urban industrial areas. At the University of Amsterdam we are concerned with developing and evaluating interfaces for communication and collaboration in emergency response settings.

    We are looking for an AI research student to develop and evaluate a prototype application of a mobile interface that adapts to users' state. Some direction in this project is still free, but at the moment we are considering adapting to (background) noise levels or user activity (walking/ manipulating/location).     We will be implementing the prototype application on a Google G1 Developer Edition mobile phone running the Android operating system.

    Alternatively, the student would improve an existing Android based mobile phone interface. This project is most suitable for a student with an interest in user experience and interface design. The idea is to increase the motivation to provide feedback for users of a mobile phone app for smell classification by integrating a location- and performance-based reward system, for example through gamification/playfulness (e.g., high scores, graphical feedback). Similar to the social network platform foursquare, users could receive virtual rewards for repeatedly providing smell descriptions when in a certain region (e.g., badges). This then could be used to generate a smell map of the Netherlands.

    In a second step, the student would perform a user study to evaluate the effectiveness of the new interface in terms of response rate, usability, and user preferences.

  • Automated extraction of narrative building blocks

    Supervisor: Benedikt Löwe, Sanchit Saraf
    Student(s): 
    Abstract: Pacuit and L    öwe defined a framework for analysing narratives that is called the "doxastic
    preference framework". In this framework, narratives are analysed as game trees, and an algorithm
    allows to calculate expectations of characters in the story and the audience. In a follow-up paper,
    Löwe, Pacuit and Saraf used this framework to formalize crime stories and identified a number of
    building blocks that can be used to generate narratives. In this project, the notion of "building
    block" should be made precise and an implementation should be created that allows the extraction of
    building blocks from a general doxastic game tree. The project requires some basic background in
    algorithms.

  • Mobflick - Mobile social storytelling with Flickr images

    Supervisor: Frank Nack
    Student(s):
    Abstract: Until recently we could access a location’s memory mainly through media surrogates, such as books, drawings, film or audio files, or through face-to-face encounters with people who were able to knit us into the rich but hidden experience fabric of a place. We are now able to leave traces of our emotional or intellectual experience as virtual attachments to any location. As a result we expect that any place, indoors or outdoors, reveals itself to us by confronting us with connection, context, and uncommon perspectives.

    Goal
    • Utilize image processing/annotation techniques and event modelling to combine geo-tagged Flickr images into a story. This can be a general story or based on social collaboration in a group of friends. The story needs to be presented on a mobile phone. The development environment is Android mobile platform.
    Research question:
    1. What are general event characteristics of a story and how can they be mapped to visual media?
    2. How can a story automatically be generated in real time based on existing visual data ?

  • Emotion-based Mobile Recommendation

    Supervisor: Frank Nack
    Student(s):
    Abstract: One of the major use cases of mobile applications is making or asking for recommendations. In most cases those recommendations are imprecise. For example, a recommendation might state that a restaurant is good but it it unclear what aspect(s) triggered that verdict.     This project looks at ways to contribute to the development of more comprehensive rating models in mobile recommendation systems.

    Goal: In this project, the influence of emotions associated with mood, temperament, personality, and disposition, will be investigated. The aim is to improve the     capture, querying and visualization of recommendations within mobile environments. An implementation of the models needs to be done in an Android mobile environment.Research questions
    1. How should emotions be represented so that they can contribute to recommendations?
    2. What type of interfaces are required to allow the capture as well as the presentation of recommendations within mobile environments?

  • Present and share - application in an experience-based context outdoor environment

    Supervisor: Frank Nack
    Student(s):
    Abstract: The central topic of the MOCATOUR project is to establish novel computational methods to facilitate tourists with personalised and contextualised access to cultural and historic information while they freely explore a city. One aspect of the project is to enable tourists to communicate through their own mobile devices with information servers of cultural institutions. In addition tourists are enabled to submit personal views in form of different media (photo, video, audio or text) to location servers as well as to the servers of collaborating cultural institutions. Both directions of interaction aim to support new forms of contextualised city experiences.

    Goal
    • The aim of this part is to develop new interface metaphors that allow to share registered experiences. The direction in this part is really to investigate the organisation of large amounts of output information on small mobile screens and on how AI techniques can help.
    Research questions
    1. What is required to present personalised experiences to a different user?
    2. What is necessary to cluster information adequately on a mobile screen (given its small size)?

  • Multimedia Knowledge Space

    Supervisor: Frank Nack
    Student(s):
    Abstract:     The goal is to support knowledge workers in their daily work within their domain through an evolving media-based knowledge system. Capturing, storing, finding, and using knowledge represented in various digital media need to become part of the standard infrastructure of a personal computing environment. Engelbart’s framework for augmenting human intellect approach  forms the basis of the proposed research, except that it is not applied to the syntax of processes, as Engelbart did, but to the understanding and handling of knowledge expressed explicitly and implicitly in digital media resources.

    This project addresses the underlying knowledge structures of such a system. In particular the representation of a scientific knowledge node will be in focus.

  • Een radar sensor voor USARsim

    Supervisor: Arnoud Visser
    Student:

    Abstract: USARSim (Unified System for Automation and Robot Simulation) is een gedetaileerde simulatie van robots en hun omgeving gebaseerd op de Unreal Tournament omgeving. Onderdeel van deze simulatie zijn sensoren die gebruikt worden in de robotica. Een sensor die nog niet onderdeel van USARsim is radar, omdat radar pas sinds kort lichtgewicht en betaalbaar  is geworden. Radar heeft voordelen ten opzichte van andere sensoren, o.a. omdat het de reflectie van meerdere obstakels tegenlijk kan teruggeven.

    Het is de taak van de student om de reactie van de sensor op obstakels te implementeren. Er is een uitgebreide dataset beschikbaar om het gedrag te valideren (ook ondere extreme omstandigheden). De voordelen van de nieuwe sensor kunnen aangetoond worden door het maken van een kaart in voor andere sensoren moeilijke omstandigheden.

    Literatuur:

    T. Peynot and S. Scheding, "Datasets for the evaluation of multi-sensor perception in natural environments with challinging conditions", in Euron GEM Sig Workshop on Good Experimental Methodology in Robotics at RSS, 2009.

  • Realistisch slachtoffer gedrag

    Supervisor: Arnoud Visser
    Student:

    Abstract: USARSim (Unified System for Automation and Robot Simulation) is een gedetaileerde simulatie gebaseerd op het spel Unreal Tournament. Deze simulatie wordt o.a. gebruikt bij de RoboCup Rescue, waar de situatie na een ramp zo goed mogelijk wordt nagebootst. Onderdeel van een ramp zijn slachtoffers, die door een robotteam gevonden dienen te worden. Een belangrijke eigenschap van nog levende slachtoffers is dat ze nog kunnen bewegen en geluid maken.

    Het is de taak van de student om de simulatie omgeving uit te breiden met slachtoffers die dynamisch gedrag vertonen. .

    Literatuur:

    Steven Nunnally and Stephen Balakirsky, "Acoustic Sensing in UT3 based USARSim", in Robots, Games, and Research: Success stories in USARSim", Workshop Proceedings of the International Conference on Intelligent Robots and Systems (IROS 2009), St. Louis, Missouri, USA, October 2009, p. 68-73.

  • Rescue robot behavior gebaseerd op XABSL

    Supervisor: Arnoud Visser
    Student:

    Abstract:  De Universiteit van Amsterdam is sinds 2006 actief in de virtual robot competitie van de RoboCup Rescue. Op het hoogste niveau worden de robots aangestuurd door behaviors, wat in de praktijk bestaat uit een paar bewegingen die aan elkaar zijn gekoppeld door een aantal transitities van de ene beweging naar de andere afhankelijk van de omstandigheden. De huidige implementatie lijkt op een finite state machine, maar is dat in de praktijk net niet.

    De opdracht voor de student is om het huidige gedrag te formaliseren, door die te vertalen in de Extensible Agent Behavior Specification Language XABSL. Onderdeel de taak van de student is om het voordeel van deze flexibele en formele aanpak aan te tonen.

    Literatuur:

    M. Risler. Behavior Control for Single and Multiple Autonomous Agents Based on Hierarchical Finite State Machines. Fortschritt-Berichte VDI Reihe 10: Informatik/Kommunikation, No. 801, Darmstadt, VDI-Verlag, 2009.  

    M. Risler and O. von Stryk. Formal Behavior Specification of Multi-Robot Systems Using Hierarchical State Machines in XABSL. In AAMAS08-Workshop on Formal Models and Methods for Multi-Robot Systems, Estoril, Portugal, 2008.    

    M. Lötzsch, M. Risler, and M. Jüngel. XABSL - A Pragmatic Approach to Behavior Engineering. In Proceedings of IEEE/RSJ International Conference of Intelligent Robots and Systems (IROS), pages 5124-5129, Beijing, China, 2006.    

    M. Lötzsch. XABSL - A Behavior Engineering System for Autonomous Agents. Diploma thesis. Humboldt-Universität zu Berlin, 2004.   M. Lötzsch, J. Bach, H.-D. Burkhard, and M. Jüngel. Designing Agent Behavior with the Extensible Agent Behavior Specification Language XABSL. In D. Polani, B. Browning, and A. Bonarini, editors, RoboCup 2003: Robot Soccer World Cup VII, volume 3020 of Lecture Notes in Artificial Intelligence, pages 114-124, Padova, Italy, 2004. Springer.   

  • 3-D user interface for a 3-D robot

    Supervisor: Arnoud Visser
    Student: 

    Abstract: The Universiteit van Amsterdam is actively participating in the Standard RobotLleague of RoboCup since 2004. In this competition soccer is played with a team of legged robots. The robots are autonomously controlled. To demonstrate the difficulty of this operation, we like to make an interface for an operator.

    It is the task of the bachelor student to analyze this situation, to invent multiple 3-dimensional control interfaces and to demonstrate the relative strength and weaknesses of those interfaces. Possible means for the control interface of a walking robot are for instance the keyboard, joystick, gamepad, speedpad, Wii-controller, Wii-balance-board, i-Phone, or voice-control. The control interface will also have a strong influence on the information feedback to the operator.

    Literatuur:

    [1] http://www.balanceboardblog.com/2008/06/wii-balance-board-used-to-control- robot.html

  • Belang van uitspraken voorspellen aan de hand van referentienetwerken

    Supervisor: Radboud Winkels
    Student(s):
    Abstract: Een belangrijke stap in het werk van juristen is het vinden van de juiste en liefst meest gezaghebbende bron(nen). Een indicatie van het belang van een bron kan zijn hoe vaak en door wie/wat er naar verwezen wordt. Wetgeving verwijst naar wetgeving, uitspraken van rechtbanken en andere colleges verwijzen naar wetgeving en naar andere uitspraken, rechtsgeleerde commentaren verwijzen naar wetgeving, uitspraken en naar andere commentaren.
    Bij het Leibniz Center for Law hebben we parsers gebouwd die zeer nauwkeurig (98% accuratesse) automatisch verwijzingen in wetgeving vinden. Deze parsers vinden ook (zij het nog iets minder nauwkeurig) verwijzingen naar wetgeving in uitspraken. Met een kleine aanpassing zouden deze ook verwijzingen naar uitspraken moeten kunnen vinden. Vervolgens willen we met netwerkanalyse onderzoeken of we maten kunnen vinden om goede voorspellingen te doen over welke uitspraken gezaghebbend zijn op basis van het aantal verwijzingen van en naar de uitspraak en bijvoorbeeld de aard van de instantie die verwijst (Hoge Raad of rechtbank bijvoorbeeld). Deze techniek is al toegepast op b.v. wetenschappelijke publikaties en patentaanvragen en Amerikanen hebben het uitgeprobeerd op uitspraken van hun Supreme Court.     

    Literatuur:

    • “ Law as a seamless web?: comparison of various network representations of the United States Supreme Court corpus (1791-2005)” in “ International Conference on Artificial Intelligence and Law – Proceedings of the 12th international conference on Artificial intelligence and law” p. 234-235, M.J. Bommarito II, D. Katz, J. Zelner, 2009, ACM, New York(NY), USA
    •  Automated detection of reference structures in law” in “Legal knowledge and information systems, volume 152 of Frontiers in Artificial Intelligence and Applications”� , pp. 41-50, E. de Maat, R. Winkels, T. van Engers, IOS, Amsterdam.

  • Fighting child abuse: detection of children in video

    Supervisor: Marcel Worring
    Student(s):
    Abstract:     Unfortunately child porn videos are captured by the police on a very regular basis and in increasing amounts. In such investigations the main goal of the police is to find any victims which might be somewhere in the video collections. This enormous task could be reduced considerably if methods would be available that can determine all children in the video automatically, so they could immediately start investigating the most promising leads.

    One method for detecting children in images has been developed by H. Weda and M. Barbieri (Automatic Children Detection in Digital Images, Multimedia and Expo, 2007 IEEE International Conference, 2-5 July 2007 Page(s):1687 – 1690)
    Paper summary: Distinguishing children from adults in digital images is beneficial for enabling ambient intelligent applications. For example, when an ambient intelligent shop window detects a child, it could adapt its content to the particular limitations and needs of children. This article presents a method to automatically detect children in digital images. The method is based on the fact that the size of a person's iris is practically constant after birth while the head grows as the person grows adult. Therefore adults can be distinguished from children based on the face/iris size ratio. The faces are detected using a standard Viola-Jones face detection technique. The irises are found and measured by using iterative Canny edge detection and a modified circular Hough transform. Our results show an accuracy of over 80% when tested on a set of 289 real life photographs of frontal faces.

    Project plan: This student project would take the above method as baseline, improve it and extend it to home video rather than photographs. Homevideos are choses as they are representative of a lot of child abuse material found and they typically contain many children. A major challenge will be the fact that in home video you will not only encounter frontal faces, but you will see faces from varying viewpoints. Resolution is also lower. The work will be part of the European project: The Investigator's Dashboard in collaboration with several police

  • Categorizatie leren

    Supervisor: Ingmar Visser
    Student(s):
    Abstract: In categorizatie experimenten leren proefpersonen stimuli indelen in verschillende categorien. Ze gebruiken hiervoor verschillende strategien, gebaseerd op een of meerdere dimensies van de stimuli. Deze verschillende strategien zijn ook uitgewerkt in computationele modellen. Het doel van dit project is om een (of meerdere) van deze modellen te implementeren en te toetsen aan experimentele data (de data van een aantal experimenten is reeds voorhanden).

    Denton, Kruschke, & Erickson (2008). Rule-based extrapolation: A continuing challenge for exemplar models. Psychonomic Bulletin & Review, 15 (4), 780-786. doi: 10.3758/PBR.15.4.780

  • Self-Training with Decision Trees

    Self-Training with Decision Trees

    Supervisors: Jafar Tanha and Maarten van Someren

     

    Semi-Supervised Learning means learning from both labelled and unlabelled examples. The self-training approach is that a  normal supervised learning algorithm first uses the labelled data tto construct a prediction model. This is used to make predictions for the unlabelled data and at the same time to give the confidence for the estimate. A small set of newly labelled data with high confidence are added to the labelled set. A new prediction model is learned from this. This is repeated until some  stop criterion holds. The self-training method can easily be combined with many supervised learning algorithms. The self-training procedure ``wraps'' around the ``base'' learner without changing its inner workings.

    Here, we consider decision tree classifiers as the ``base'' learner in self-training. Decision tree classifiers are used successfully in many diverse domains such as medical diagnosis, speech recognition. One of the most important features that contribute to the popularity of decision tree learners is the comprehensibility of the resulting trees. Our goal is to find out how a decision tree classifier can be used best in self-training.

     The main question in the selection step is how the selection function can find confidence in its predictions. Unfortunately in practice, decision trees have been found to provide poor probability estimates. Decision tree as the underlying classifier in self-training faces two obstacles to producing a good ranking of instances: the first is that the sample size on the leaves is almost always small, and the second is that all instances at a leaf get the same probability.

    Therefore, the research question in  Self-Training is to find a a good ranking method for the leaves of decision trees, which is crucial for self-training accuracy. The area under the receiver operating characteristic curve (ROC), simply AUC, has been recently used as a measure for ranking performance of learning algorithms. Thus, one solution is to apply AUC ranking in the leaves of the tree for better ranking. A second possibility is  to use decision tree learning with fuzzy labels in the leaves and combining multiple leaves to obtain better predictions and better confidence estimates.

  • Herkennen van ketosis

    Titel: Data Mining voor het herkennen van ketosis bij koeien

    Begeleiders: Maarten van Someren en Rik van der Tol (Lely)


    In het projekt Leren & Beslissen is gewerkt aan het het leren herkennen van koeieziektes. In een eerder projekt is met succes een
    systeem gemaakt voor het herkennen van mastitis. Het herkennen van "ziekte" in het algemeen bleek moeilijker. Dat kan komen door de variatie
    tussen de ziektes. Een duidelijke ziekte is "ketosis". Deze kan herkend worden aan de typische geur van de koe. Het gaat om een ander type ziekte dan
    mastitis, maar is specifieker dan algemene "ziekte".  Doel van het projekt is om een optimale ketosis herkenner te maken, zonder geur-informatie.

  • News is my story

    Supervisor: Frank Nack
    Student: Frank Smit

    Abstract:
    Daily news are distributed today through various channels, e.g. via mobile phones or particular applications such as news readers for the ipad.  This news is normally written for a broad audience but different in style depending n the newspaper  in the Netherlands there are several newspapers (like; de Telegraaf, NRC, AD, etc). Despite the need of being updated with current incoming news items (e.g. based on a user profile), readers are interested to gain larger analysis articles the summarize events (e.g. for a longer read over the weekend). The challenge here is that these longer pieces should be personalized too.
    As news report about events a potential solution to that problem might be to combine storytelling with event representation. This then can be used  to map multiple related news-articles to a new news-article, resulting in a news-story. This news-story can be told in different way according to the user interest and presentation preferences. The reader should be able to choose between a couple of different telling styles, so event structures need to be developed that facilitate various presentation styles.

    Goal:
    Creating event-structures to tell textual stories in different styles to facilitate personalized news presentation explaining the development of weekly events..

    Research question:
    - Is it possible to use multiple related news-articles, for creating a personal based news-story?

  • A story generator based on building blocks

    Instructor: Benedikt Löwe, Sanchit Saraf

    This project is related to the project "Automated extraction of narrative building blocks", and these two projects could be done by a group of two students.

    Based on the building blocks in the "doxastic preference framework" of Löwe and Pacuit, an
    automated story generator should be built that transforms a sequence of building blocks into a
    natural language text. Part of the task is to develop testing benchmarks for the quality of the
    implementation and to compare the software with other story generators such as Mexica by Perez y
    Perez.

  • Human understanding of Russian folktales

    Instructors: Rens Bod, Benedikt Löwe, Sanchit Saraf

    In the 1920s, Vladimir Propp presented one of the first formal systems for narratives and applied it to a collection of 100 Russian folktales. This system forms the historical basis for all formal models of narratives and is widely studied.

    However, no systematic study on the possibility of objectively representing natural language stories via Propp's system has been performed. This project is part of such a study. In the study, test subjects should learn to do formal representations in the Proppian system, in order to study the inter-coder agreement for such a formalization. The project consists of the development of an annotation software for the Proppian system together with a statistical evaluation tool. This software would then be used in the inter-coder study.

  • Conformal Geometric Algebra

    Supervisors: Carsten Cibura, Leo Dorst

    Conformal transformations are geometric transformations that preserve angles locally. A simple example would be a reflection in the surface of a sphere, but Euclidean transformations and rigid body motions are also contained as in this class. Geometric algebra offers a way to represent conformal transformations simply as orthogonal linear transformations, implemented as versors (spinors).

    Given exact minimal correspondence data, the conformal transformation relating the two data sets can be recovered. If the correspondence data is noisy, however, the true transformation between the two sets cannot be recovered even though some solution can be obtained. The question arises how well the obtained solution matches the true transformation depending on the magnitude and nature of the noise.

    Working on this bachelor project requires prior knowledge of basic geometric algebra. It will deepen you knowledge of the geometric algebra description of motions in space, and improve your skills in handling noisy geometrical data. Both are valuable tools in the future development of autonomous systems.

    In the course of this project you will implement the method of determining a conformal transformation between exact minimal correspondence data [1] in MATLAB and extend it so that it can handle noisy minimal correspondence data using an approach based on the singular value decomposition (SVD) [2]. You will create and use simulated data sets to evaluate the quality of the solution under different kinds of noise.

    [1] Cibura C., Dorst L.: Determining Conformal Transformations in R^n from Minimal Correspondence Data
    [2] Richard A., Fuchs L., Charneau S.: An Algorithm to Decompose n-dimensional Rotations into Planar Rotations, Computations Modeling of Objects Represented in Images, 2010; 60-71

  • Projective Geometric Algebra

    Supervisor: Leo Dorst

    Recently it has been indicated that projective transformations can be represented as orthogonal transformations in "line space'', a real space with metric (3,3). This is so new that many elementary and useful results can be obtained by the right person. We are going to investigate this, by means of handwork and implementation.

    Prior knowledge of geometric algebra is required. This project will improve your understanding of classical homogeneous coordinates (as used in robotics and computer vision), and of geometric algebra (which may replace it based on your work).
Maintained by Bas Terwijn. Last edited on Mon, 25 Jan 2010 17:10:55 +0100