The Autumn School will take place from September 12th to September
16th, 1997 at Freiburg. It will take place just after the
21st German
Conference on Artificial Intelligence (KI'97), which will
also be held at Freiburg from the 9th of September to the 12th of
September,
1997.
Pat Langley: Machine Learning for Adaptive Advisory Systems
Unlike traditional expert systems, advisory systems attempt to aid
experts rather than replace them. Both approaches rely on substantial
domain knowledge, but advisory systems differ in another important
respect - their interaction with the user provides them with data
about the user's preferences. Methods for machine learning, embedded
within the advisory system, can use these data to acquire a user model
and thus improve the system's ability to meet the user's needs. In
this talk, I review a number of such adaptive advisory systems and
propose some tentative principles about their study and development. I
also consider briefly the relation between such user models and
cognitive simulations of human behavior.
Hans Spada: Abschlußveranstaltung
- Teil 1: Ein erstes Resümee zur 7. Herbstschule
Kognitionswissenschaft
-
Einleitend wird ein erster kritischer Rückblick auf die 7.
Herbstschule Kognitionswissenschaft gegeben. Anhand einer Analyse
der Programme aller bisherigen Herbstschulen wird die Entwicklung
dieses Treffens beschrieben. Ihre Bedeutung für die
Kognitionswissenschaft in Deutschland wird skizziert. Konsequenzen
sind aus dem absehbaren Auslaufen der die Herbstschulen tragenden
Graduiertenkollegs zu ziehen.
- Teil 2: Kognititve und emotionale Verarbeitung von
Informationen über Umweltrisiken
- Im zweiten Teil berichte ich
über gemeinsam mit Josef Nerb und Stefan Wichmann gewonnenen
Arbeitsergebnissen zur Informationsrezeption bei Umweltrisiken. Wir
formulierten ein Modell, das einerseits die Entstehung von Emotionen
gegenüber Umweltrisiken beschreibt und andererseits die Mechanismen
des Erwerbs und der Verwendung des oft stereotypen und schemaartigen
Wissens von Laien aufzeigt und mit der vorwiegend
ereignisorientierten Medienberichterstattung in Verbindung bringt.
Mit dem Ziel der Integration und Präzisierung relevanter
psychologischer Theorien wurden Teile des Modells unter Nutzung der
Ansätze der parallel constraint satisfaction von Holyoak und
Thagard (1995) als Computersimulation realisiert. Das Modell ist in
der Lage, eine Reihe von empirischen Befunden zu erklären und
konnte zum Teil bereits validiert werden (Nerb & Spada, 1997).
Literatur
- 1
- Holyoak, K. J., & Thagard, P. (1995).
Mental leaps: Analogy in creative thought.
Cambridge, MA: MIT Press.
- 2
- Nerb, J., & Spada, H. (1997) The role of
controllability of the cause in cognitive and emotional evaluation
of an environmental risk. Proceedings of ECO-INFORMA`97.
How can biological systems be used as models for artificial systems?
What can robots and artificial creatures tell us about human and
animal behavior? The goal of the course is to show how insights from
biology can be used to guide the construction of artificial systems
and how such systems in turn can act as models for their biological
counterparts.
The starting point will be to ask what we know about
intelligence in animals? It was once taken for granted that learning
in animals and man could be explained with a simple set of general
learning rules, but over the last hundred years, a substantial amount
of evidence has been accumulated that points in a quite different
direction. In animal learning theory, the laws of learning are no
longer considered general. Instead, it has been necessary to explain
behavior in terms of a large set of interacting learning mechanisms
and innate behaviors. Artificial intelligence is now on the edge of
making the transition from general theories to a view of intelligence
that is based on an amalgamate of interacting systems. In the light of
the evidence from animal learning theory, such a transition is to be
highly desired.
The developments from the early conceptions of animal
behavior to contemporary neurophysiological models will be traced
through such diverse areas as ethology, behaviorism, cognitivism,
control theory, cybernetics, and brain theory. All these areas have
something to offer to the explanation of intelligent behavior. Taken
together they suggest that different behaviors can be of very
different complexity. Some behaviors can adequately be described as
simple reflex-loops. Many other require very intricate machinery.
Is
behavior reactive or deliberate? How are behaviors controlled and
combined? The view from ethology in the tradition of Lorenz and
Tinbergen will be connected to the recent development in reactive
robotics. Brooks' subsumption architecture can be seen as a robot
ethology that can be used to synthesize reactive behaviors. However,
many animals are clearly capable of behavior that cannot be purely
reactive. This is especially vivid in shortcut and detour problems
where animals seem to reason about different alternatives.
A distinction among four types of behaviors will be suggested: (1)
appetitive behavior that is directed toward an object or a situation,
(2) aversive behavior that is directed away from an object or a
situation, (3) exploratory behavior that relates to objects or
situations that are unknown, and finally, (4) neutral behavior that is
guided by objects that are known to be neutral, that is, neither
appetitive nor aversive.
Two types of behaviors will be worked out in
more detail. The first is the orienting behavior that directs the
sensory apparatus of an animal toward the source of unexpected
stimuli. It combines a number of interesting properties such as multi
modal integration, sensory-motor mapping, interaction with
exploration, learning and attention. The neurological mechanisms of
the behavior and number of robot implementations will also be
presented.
The second example is the approach behavior, that is,
behavior that moves an organism, or parts of it, closer to an external
object. A number of variants of the basic goal-directed behavior will
be given. Starting with Braitenberg's vehicles, the path will be
worked out to visually guided servoing during locomotion and reaching
behavior. Examples from the animal kingdom as well as from robot labs
will be shown.
Apart from being caused by the complexity of the
innate mechanisms, intelligent behavior is the result of learning. The
presentation will center around two types of learning that have
traditionally been considered very simple: conditioning and
habituation. It turns out however, that a complete explanation of
these very simple learning situations require many complex interaction
among a number of learning systems. Starting with Pavlov's early
model, the analysis will touch on a number of computational models
such as Hull's classical work, the Rescorla-Wagner model and more
modern real-time models such as the Klopf model and the Sutton-Barto
model. Connections to recent development in reinforcement learning
will also be made. Their merits and limitation, both as explanations
of animal behavior and for robot control, will be discussed.
When the
demands of conditioning and habituation are taken seriously, no simple
associative explanation is sufficient on its own. Mechanisms for
category formation, generalization and the formation of expectations
must be added before any model comes close to explaining the available
data. These more advanced systems leads the way to the processing of
expectancies, a type of structure that shows up for example in spatial
navigation and top-down influences on perception.
The final topic of the course will be the role of motivation and
emotion in autonomous robots. Based on Rolls' neurophysiological
theory of emotion the basis for artificial motivation and emotions
will be presented. With the risk of taking the mystery out of
emotions, motivation will be seen as a type of arbitration, while
emotions are seen as states caused by reinforcing events.
References
- 1
-
Balkenius, C. (1994). Biological learning and
artificial intelligence. / Lund
University Cognitive Studies, 30.
- 2
- Balkenius, C. (1995). Natural
intelligence in artificial creatures. / Lund University Cognitive
Studies 37. (available as on-line book at http://lucs.fil.lu.se/Staff/Christian.Balkenius/Thesis/)
- 3
- Balkenius, C. (1995). Multi-modal
sensing for robot control. In Niklasson, L. F.,
Bodén, M. B. (Eds.) Current trends in connectionism
(pp. 203-216). / Hillsdale, NJ: Lawrence Erlbaum.
- 4
- Balkenius, C. (1996). Generalization
in instrumental learning. In Maes, P., Mataric, M.,
Meyer, J.-A., Pollack, J., Wilson, S. W. (Eds.) From
Animals to Animats 4: Proceedings of the Fourth International
Conference on Simulation of Adaptive Behavior. /
Cambridge, MA: The MIT Press/Bradford Books.
- 5
- Balkenius, C., Kopp, L. (1996).
Visual tracking and target selection for mobile robots. In
Jörg, K.-W. (Ed.) Proceedings of EUROBOT '96. /
IEEE Press.
All of the above are avilable on-line at
http://lucs.fil.lu.se/Staff/Christian.Balkenius/.
Karl Gegenfurtner: Neurowissenschaftliche
Wahrnehmungsforschung
Neurowissenschaftliche Forschung zeichnet sich durch eine
interdisziplinäre Vorgehensweise aus. In der jüngsten
Vergangenheit hat sich dieser Ansatz vor allem auch für die
Erforschung des visuellen Systems ausgezeichnet. Die bedeutendsten
neueren Ergebnisse der Wahrnehmungsforschung haben sich aus der
Verknüpfung von psychophysischen Verhaltensuntersuchungen,
elektrophysiologischen Ableitungen von Neuronen und
Neuronenverbänden, und theoretischen Überlegungen
ergeben. In diesem Kurs wird ein Überblick über die
Informationsverarbeitung im visuellen System gegeben, anhand einiger
ausgewählter Beispiele, die diesen interdisziplinären Ansatz
benutzen.
Nach einer einleitenden Übersicht über die
Struktur und Funktionsweise des Sehsystems soll zunächst auf die
erste Stufe der Informationsverarbeitung in der Retina eingegangen
werden. Dies geschieht in zweierlei Hinsicht. Zum einen wird gezeigt,
wie sich unser Sehsystem im Laufe der Entwicklung optimal an unsere
Umwelt angepaßt hat. Zum anderen soll dargestellt werden,
inwieweit elementare periphere Faktoren visuelle
Wahrnehmungsleistungen ermöglichen bzw. einschränken
können.
Der zweite Teil beschäftigt sich mit der zentralen
Verarbeitung visueller Information. Am Beispiel der
Bewegungswahrnehmung wird aufgezeigt, wie die Verquickung von
neurophysiologischen und psychophysischen Daten zu direkten
Aufschlüssen über die neuronalen Verarbeitungsmechanismen
führen kann. Anschließend soll am Beispiel der Verarbeitung
von Bewegung und Farbe untersucht werden, zu welchem Grad das visuelle
System Strategien der parallelen Informationsverarbeitung
benutzt. Dies ist insbesondere wichtig, da das Prinzip der
gleichzeitigen Verarbeitung visueller Information in getrennten
Kanälen in den letzten 20 Jahren als eine Richtlinie für das
Verständnis des Gehirns diente.
Im dritten und letzten Teil soll
schließlich auf die Verarbeitung komplexer Reize eingegangen
werden, wie sie bei der Betrachtung natürlicher Szenen oder dem
Erkennen von Objekten auftauchen.
References
- 1
-
Spillmann, L. (1996). Wie das Gehirn sieht. Freiburger
Universitätsbltter, 134, 9-47.
- 2
- Gegenfurtner, K.R. (1997).
Farbwahrnehmung beim Menschen.
http://www.mpik-tueb.mpg.de/people/personal/karl/html/heidelberg/heidelberg.html
- 3
-
Newsome, W.T., Britten, K.H., Salzman, C.D. & Movshon, J.A. (1990).
Neuronal mechanisms of motion perception. Cold Spring Harbor Symposia
on Quantitative Biology, 55, 697-705.
- 4
-
Gegenfurtner, K.R. & Hawken, M.J. (1996). Interactions of color and
motion in the visual pathways.Trends in Neurosciences, 19, 394-401.
- 5
-
Logothetis, N.K. & Sheinberg, D.L. (1996). Visual object
recognition. Annual Review of Neuroscience, 19, 577-621.
- 6
-
Barlow, H. (1996). The neuron doctrine in perception. In
M. S. Gazzaniga (Ed.) The
cognitive neurosciences /, pp. 415-435. Boston:MIT Press.
Udo Hahn: Performanzmodellierung in der Sprachverarbeitung
Der Kurs gibt einen Überblick über verschiedene Ansätze,
die sprachliche Performanzaspekte in Modellen der Sprachverarbeitung
explizit berücksichtigen.
Den Ausgangspunkt der Betrachtungen bildet die von Chomsky in die
linguistische Theoriedebatte eingeführte Unterscheidung zwischen
``Kompetenz'' und ``Performanz'', zwischen dem allgemeinen Wissen
über Sprache und dem spezifischen Sprachverhalten. Unter dem
deklarativen Blickwinkel der Kompetenz wird vollständig von
prozeduralen Bedingungen der Analyse (Parsing) und Produktion
(Generierung), d.h. von konkreten Sprech- und Verstehensvorgängen
abstrahiert. Solche Beschreibungsaspekte der Sprachverwendung werden
dem Bereich der Performanz zugeordnet und umfassen z.B. die
Untersuchung von allgemein gültigen kognitiven Rahmenbedingungen der
Sprachanalyse (Begrenzung des Kurzzeitspeichers, Präferenzen bei der
Verarbeitung und andere heuristische Ökonomieprinzipien), von
natürlichen Diskursszenarien (Dialoge und Texte statt isolierten
Sätzen) und von an Vorwissen, Interessen u.ä. adaptierten
Sprachverstehensstrategien, aber auch von Fehlern bei der
Sprachproduktion und -rezeption (aufgrund von motorischen oder
artikulatorischen Defiziten, Verrauschungen des Sprachsignals,
Aufmerksamkeitsschwankungen usw.). Wie aber verknüpft man diese
Performanzphänomene in methodisch angemessener Form mit
kompetenzgrammatischen Beschreibungen? Sind Performanzbeschreibungen
nur außergrammatische prozedurale Aufsätze auf deklarative
Kompetenzgrammatikkerne, oder sind sie integraler Teil eines um
performatives Wissen substantiell erweiterten Grammatikkonzepts?
Nicht nur aus theorielinguistischer Perspektive sind diese
Fragestellungen bedeutsam. In dem Maße, in dem der kompetenzfundierte
Grammatikansatz zu erheblichen theoretischen (ungünstige
Komplexitätseigenschaften der jeweils verwendeten
Grammatikformalismen bzw. Parsing-Algorithmen) und praktischen
Problemen (etwa kombinatorische Effekte bei unbeschränkt ambigen
sprachlichen Strukturen) im Rahmen der automatischen Verarbeitung
natürlicher Sprache führt, stellt sich auch innerhalb der
Computerlinguistik zusehends die Frage nach dem theoretischen
Stellenwert von Performanzbetrachtungen. Sie wird übrigens aus zwei
voneinander eher unabhängigen Richtungen aufgeworfen -- einerseits
von seiten der Sprachtechnologie, die die Perspektive der Effizienz
funktional komplexer Sprachverstehenssysteme betont, andererseits von
seiten der (computergestützten) Sprachverstehenssimulation, die
kognitive Plausibilitätserwägungen in den Vordergrund ihres
Erklärungsinteresses rückt.
Im Kurs werden die derzeit verfügbaren Beschreibungskonzeptionen
behandelt. Es können vier Hauptkategorien von Lösungsvorschlägen
für die Einbeziehung von Performanzeffekten bei der
Sprachbeschreibung bzw. -analyse unterschieden werden:
- Statistische Kriterien
-
Im Kontext statistischer Performanzmodelle werden
Auftrittshäufigkeiten lexikalischer Elemente, syntaktischer
Strukturen und semantischer Interpretationen auf der Grundlage
umfassender Korpusanalysen erhoben. Häufigkeitsdaten beruhen auf
direkt beobachtbaren Ausprägungen sprachlicher Performanz und
bilden die Grundlage für die Abschätzung von
Auftrittswahrscheinlichkeiten entsprechender sprachlicher
Phänomene, die sich in der aktuellen Sprachverwendung als
``Erwartbarkeit'' von bestimmten Verwendungsmustern oder Lesarten
äußern. Diese Überlegungen führen direkt zu probabilistischen
Grammatiken. Hierbei wird Performanz als statistische
Widerspiegelung von Kompetenz gedeutet und ein um statistische
Verwendungsinformationen erweitertes Konzept von
Kompetenzgrammatiken vorgeschlagen. In der Computerlinguistik
finden diese Vorschläge entweder a priori in Form von numerischen
Gewichtungsfaktoren zur Steuerung der Regelauswahl oder a posteriori
zur Plausibilitäts- bzw. Präferenzbewertung von alternativen
Parsing-Ergebnissen ihren Niederschlag.
- Strukturelle Konfigurationen
-
Die Reduktion von Ambiguitäten kann auch durch die Annahme
bevorzugter struktureller Anbindungen als Grundlage für
heuristische Parsing-Strategien gesteuert werden (wie minimal
attachment oder right association). Hintergrund dieser stark von
psycholinguistischen Evidenzen geprägten Überlegungen sind meist
informelle Kostenabschätzungen und darauf aufbauende
Optimierungsversuche für den Aufbau von Strukturbeschreibungen
(beispielsweise werden weniger komplexe Phrasenstrukturbäume
bevorzugt) oder Hypothesen über die Auswirkungen von
Beschränkungen der Ladekapazität temporärer Speicher (etwa des
Kurzzeitgedächtnisses) auf die Parsebarkeit von Sätzen.
- Formale Restriktionen
-
Aus der direkten Umsetzung geläufiger Kompetenzgrammatiken als
Modelle für die menschliche oder automatische Sprachverarbeitung
resultiert in den meisten Fällen die Turing-Äquivalenz
entsprechender Automatenmodelle für die Sprachanalyse. Dies hat
fatale Auswirkungen auf die prinzipielle Berechenbarkeit von
linguistischen Beschreibungen und mögliche Anforderungen bzw. den
Verbrauch von Ressourcen (Zeit, Speicher) beim Berechnungsprozeß.
Die aus theorielinguistischer Sicht gerechtfertigten, aber für die
praktische Verarbeitung nachteiligen, weil übergenerellen Annahmen
haben deshalb auch in der Computerlinguistik Überlegungen Auftrieb
gegeben, die empirisch unstrittigen Rahmenbedingungen bei der
Analyse natürlicher Sprache (endliche Sätze, endliche
Rekursionstiefen, endliche Speicher usw.) in ein diese
Restriktionen direkt widerspiegelndes formales
Performanzgrammatikmodell einfließen zu lassen. Entsprechende
Konzeptionen orientieren sich vor allem am Konzept deterministischer
Grammatiken bzw. Analysen durch LR(k)-Grammatiken oder an noch mehr
restringierten endlichen Automaten mit im wesentlichen linearem
Laufzeitverhalten. Dieser Ansatz axiomatisiert somit
Performanzbeschränkungen und fixiert Performanzgrammatiken
orthogonal zu Kompetenzgrammatiken.
- Differentielle Ressourcenallokation
-
Eine Alternative zur starren formalen Beschränkung der
Analyseapparaturen ist die flexible Vergabe von Ressourcen für
Berechnungsprozesse über Steuerungsparameter eines
Sprachanalysesystems. Einige der Vorschläge ähneln statischen
compile-time-Verfahren, während andere dynamischeren
run-time-Verfahren zugeordnet werden können.
Zur ersten Gruppe zählt das Skimming. Ausgangspunkt ist die
gezielte Beschränkung bzw. Ausdünnung der Grammatik- und
Konzeptsystemspezifikationen auf bestimmte, als inhaltlich signifikant
erachtete lexikalisch/konzeptuelle Beschreibungselemente. Treten diese
beim Parsing in einem Text auf, wird das einschlägige Textsegment einer
tieferen syntaktisch-semantischen Feinanalyse unterzogen, während
unsignifikante Textpassagen ohne Analyse überlesen werden. Eine andere
Konzeption, das Skimming zu realisieren, besteht in der Vorgabe von
an Leserpräferenzen orientierten Interessemodellen, die (obwohl
möglicherweise vollständige Domänen- und Grammatikspezifikationen
vorliegen) die Tiefe des Textverstehens aufgrund der Übereinstimmung des
Textinhalts mit dem, was Leser ``interessant'' finden, steuern.
Der Wechsel von ignorierten und tiefer
analysierten Textpassagen kann durch sog. partielle Parser
weiter beschränkt werden, die durchgängig auf die
strukturell ``flache'' Beschreibung oder gar nur auf die Erkennung
bestimmter Strukturmuster (etwa Nominalphrasen) in sprachlichen
Äußerungen ausgelegt sind.
Zur zweiten Gruppe der laufzeitorientierten Verfahren zählen beispielsweise
dynamisch an der Systemlast oder der Wichtigkeit von Berechnungen orientierte,
von Scheduling-Techniken inspirierte Anytime-Algorithmen.
Ihr wesentliches Kennzeichen ist, daß sich zumindest eine Subklasse dieser
Algorithmen zu jedem Zeitpunkt der Ausführung unterbrechen läßt und die
bis dahin erzielten Berechnungsergebnisse unmittelbar zugreifbar sind.
Bei Bedarf können präzisere (qualitativ ``bessere'', aber
auch nur durch höheren Ressourceneinsatz zu erzielende) Ergebnisse erzeugt
werden, wobei Optimalitätskriterien ungerechtfertigten Ressourcenverbrauch
unterbinden. All diesen Verfahren gemeinsam ist die
Komplexitätsreduktion der Sprachverarbeitung, die mit Einbußen bei der
Qualität (Präzision) der Analyseergebnisse erkauft wird.
Wie sollte ein wirklich / intelligenter Roboter (also so einer
wie in science-fiction-Filmen) bestimmen, was er als nächstes tut?
Wie kann der Tages-Arbeitsplan - sagen wir: einer großen
Autofabrik automatisch generiert und im Tagesverlauf angepaßt
werden an unvorhergesehene Ereignisse wie Ausfall von Maschinen oder
Eintreffen von Eilaufträgen? Wie kann ein Hilfe-System Nutzern mit
unterschiedlichen Graden an Erfahrung im Umgang mit einer gegebenen
Anwendungssoftware Tips geben, wie sie ihre Anwendungsprobleme mit der
Software am besten lösen können?
In allen diesen Problemen stecken als Teile Handlungsplanungsprobleme:
Gegeben Wissen über die Welt (das mehr oder weniger vollständig
und zutreffend sein kann), über Aktionsmöglichkeiten des
betrachteten Systems (die ebenfalls mehr oder weniger vage bekannt
sein können) und über (möglicherweise unpräzise) subjektive
Ziele des Systems in der Welt, ermittle einen Aktionsplan, dessen
Ausführung die Ziele in der Welt erfüllen wird oder wenigstens den
Lauf der Dinge positiv beeinflußt.
Handlungsplanung, oder kurz: Planen, ist eines der klassischen
Teilgebiete der KI. Der Kurs führt in das Gebiet aus dem Blickwinkel
der Informatik ein: Im Vordergrund stehen Algorithmen,
Repräsentationen und die Charakterisierung der jeweils bearbeitbaren
Planungsprobleme; Fragen der Übertragbarkeit in die
Kognitionsforschung und der Überführbarkeit in
Anwendungs-Technologie stehen für diesen Kurs im Hintergrund.
Der Kurs besteht aus fünf Teilen:
- ,,Klassisches`` Planen.
- Zum Einstieg gibt es einen
Abriß der ältesten, bestuntersuchten und in gewisser Weise
einfachsten Klasse von KI-Planungstechniken.
- Varianten des Planens.
- Dann entwickeln wir aus einer genaueren
Charakterisierung des klassischen Planens eine weiter gefaßte
,,Landschaft`` von Planung.
- Operatorkalküle.
- Theoretisches und praktisches Herzstück
vieler Planungsalgorithmen ist ein Kalkül, der ermittelt, wie
die Welt voraussichtlich aussieht nach Anwendung eines gegebenen
Operators in einer gegebenen Weltsituation. Dieser Kursteil beschreibt
einige einfache Varianten solcher Kalküle.
- Planen mit Zeit.
- Eine Klasse von Planungsvarianten verwendet
reichere Repräsentationen von Zeit - beispielsweise Zeitintervalle,
quantitative Zeitdauer oder globale Uhrzeit. Hier geben wir einen
Abriß zweier einfacher Techniken dazu.
- Planen unter Unsicherheit.
- Viele, besonders neuere Arbeiten
modellieren und verwenden explizit bekannte Unsicherheit des
Vorhandenen Welt- oder Operatorwissens zur Planungszeit. Zum
Abschluß des Kurses skizzieren wir einige
wahrscheinlichkeitsbasierte Ansätze.
Teilnahmevoraussetzungen
Es wird Grundlagenwissen (maximal Vordiplomsniveau) über KI und
Informatik vorausgesetzt.
References
- 1
-
J. Hertzberg.
Planen. Einführung in die Planerstellungsmethoden
der
Künstlichen Intelligenz.
BI Wissenschaftsverlag, Mannheim u.a., 1989.
- 2
-
S. Russell and P. Norvig.
Artificial Intelligence: A Modern Approach.
Prentice Hall, Englewood Cliffs, NJ, 1995.
Part ,IV: Acting Logically und Part ,V: Uncertain
Knowledge and
Reasoning (teilweise).
- 3
-
D.S. Weld.
An introduction to least commitment planning.
AI Magazine, 15(4):27-61, 1994.
Keith Holyoak: Analogy: Psychological, Computational and Neural Approaches
This course will focus on thinking and the representation of
knowledge, with particular emphasis on the role of analogy in
inference and learning. Analogy provides a case study of research in
the interdisciplinary field of cognitive science. The importance of
analogical thinking has been widely recognized in such diverse fields
as cognitive, developmental, social, and comparative psychology, as
well as philosophy, artificial intelligence, education, law, political
science, literary criticism, and anthropology. Analogy provides
important insights into human cognition because it depends on the
ability to form and manipulate explicit representations of complex
relations.
We will consider analogical thinking from three general perspectives,
each the focus of one of the three lectures. From a psychological
perspective, we will examine experimental research on human use of
analogy in thinking. We will survey what is known about the major
stages of analogy use: retrieval, mapping, inference, and
generalization. From a computational perspective, we will examine
current simulation models of analogy use. We will consider the
assumptions they make about knowledge repreesentation and their
adequacy in accounting for the psychological evidence. In particular,
the relevance of analogy to evaluating the psycholgical adequacy of
connectionist models of thought will be discussed.
Finally, we will consider related work in neuropsychology that
suggests connections between analogical reasoning and the neural
substrates of working memory in the prefrontal cortex.
Some background knowledge of cognitive psychology or knowledge
representation will be helpful. However, I will endeavor to make the
course accessible to a broad audience.
Relevant Background Readings
- 1
- Beardsley, T. (1997). Machinery of thought.
Scientific American, August 1997, 78-83.
- 2
- Falkenhainer, B., Forbus, K. D., & Gentner, D. (1989).
The structure-mapping engine: Algorithm and examples.
Artificial Intelligence, 41, 1-63.
- 3
-
Holyoak, K. J., & Thagard, P. (1989). Analogical mapping by
constraint satisfaction. Cognitive Science, 13, 295-355.
- 4
- Holyoak, K. J., & Thagard, P. (1995).
Mental leaps: Analogy in creative thought.
Cambridge, MA: MIT Press.
- 5
-
Hummel, J. E., & Holyoak, K. J. (1997). Distributed representations
of structure: A theory of analogical access and mapping.
Psychological Review, 104, 427-466.
- 6
- Keane, M. T., Ledgeway, T., & Duff, S. (1994).
Constraints on analogical mapping: A comparison of three models.
Cognitive Science, 18, 387-438.
- Lecture 1. Hybrid architectures for autonomous mobile robots
-
This lecture covers fundamentals of mobile robot programming,
concentrating on the hybrid 3-level architecture that is currently
the best practical system for intelligent autonomous operation of
mobile robots. We briefly review behavior-based control, behavior
combination schemes, and reactive control paradigms. The last
part of the lecture gives more in-depth coverage of the Procedural
Reasoning System (PRS) and its implementation in the Colbert
language.
Practical: programming simple activity patterns on the
Pioneer robots using the Colbert language.
- Lecture 2. An Open Agent Architecture for agent coordination
-
This lecture discusses the theory and implementation of SRI's Open
Agent Architecture, a system for the coordination of heterogenous
agents on the World Wide Web. We cover the use of a facilitator
agent as the organizing principle in the OAA, registering agent
capabilities and distributing tasks. We present particular input
and output agents (speech, gestures, text), along with examples of
their coordination.
Practical: programming multimodal agents to perform
simple alerting and message delivery tasks.
- Lecture 3. Mobile robots as physical agents
- Here we combine
the ideas of the previous two lectures to formulate a multiagent
system that incorporates physical agents. We discuss issues of
human communication, tasking, and robot coordination. Examples of
the problems and opportunities presented by the agent framework
are given.
Practical: a small cooperative task using the Pioneer
robots.
In this course, I cover the basic
principles and methods of machine learning, including their
relevance to computational models of human learning. The aim is to
provide participants with broad overview of the field, rather than
to focus on specific research results.
The first lecture defines machine learning and reviews its major
paradigms, then concentrates on techniques for supervised concept
induction, including decision-tree and rule induction, case-based
learning, connectionist learning, and probabilistic methods.
Research on these basic algorithms and their extensions constitute
the core activity in machine learning. They also play a central role
in work on knowledge discovery and data mining, as well as in
studies of human categorization.
The second lecture motivates the need to move beyond classication
tasks to more complex domains like problem solving and natural
language, then considers the role of supervised learning methods in
these more challenging areas. This includes a review of methods that
learn search-control heuristics and grammatical knowledge, including
recent techniques for reinforcement learning. Such algorithms are
important in building intelligent agents and also have implications
for complex human learning.
The final lecture deals with the issue of incremental learning,
which any computational model of human learning must address. After
considering some alternative definitions of the term ``incremental'',
we examine some effects of training order on incremental learners
and consider the factors that can mitigate these effects. In
closing, we revisit some of the basic techniques for supervised
learning and consider their psychological plausibility along these
dimensions.
References
- 1
-
Langley, P. (1995). Elements of machine learning. San Francisco:
Morgan Kaufmann.
- 2
- Langley, P. (1995). Order effects in incremental
learning. In P. Reimann & H. Spada (Eds.). Learning in
humans and machines: Towards and interdisciplinary learning
science. Oxford: Elsevier.
- 3
- Langley, P. (1997). Machine learning for
intelligent systems. Proceedings of the Fourteenth National
Conference on Artificial Intelligence. Providence, RI: AAAI Press.
(http://www.isle.org/~langley/papers/invite97.ps).
- 4
- Langley, P. (in press). Machine learning for
adaptive user interfaces. Proceedings of the 21st German
Annual Conference on Artificial Intelligence. Freiburg, Germany:
Springer. (http://www.isle.org/~langley/papers/adapt.ki97.ps).
- 5
- Schlimmer, J. C., & Langley, P. (1992). Machine
learning. In S. Shapiro (Ed.), Encyclopedia of artificial
intelligence (2nd ed.). New York: John Wiley & Sons.
- 6
- Langley, P., & Simon, H. A. (1995). Applications of
machine learning and rule induction. Communications of the ACM, 38,
November, 55-64. (http://www.isle.org/~langley/papers/app.cacm.ps).
Knowledge discovery in databases (KDD), often called data mining, aims at
the discovery of useful information from large collections of data. The
discovered knowledge can be rules describing properties of the data,
frequently occurring patterns, clusterings of the objects in the
database, etc. The lectures cover some of the research issues in
knowledge discovery.
We start by briefly discussing the KDD process, basic data mining
techniques, and listing some prominent applications. Then we discuss
the role of machine learning and statistics in KDD and data mining.
Next, we move to the role of databases in knowledge discovery by looking
at a simple example of data mining, namely the problem of discovering
association rules. We present a simple algorithm for this task, and
show how the same ideas can be used for other types of data, too. We
continue by looking at some additional examples: episodes in sequences
and keys in relational databases. After that, we present a generic data
mining algorithm, and discuss some of the architectural issues in data
mining systems. Finally, we present the notion of inductive databases
and show how this concept can be used to model data mining as querying.
References
- 1
-
Heikki Mannila:
A tutorial on data mining.
Proceedings of
International Conference on
Database Theory (ICDT'97),
Delphi, Greece, January 1997, F. Afrati and
P. Kolaitis (ed.), p. 41-55.
- 2
-
Rakesh Agrawal, Heikki Mannila, Ramakrishnan Srikant, Hannu Toivonen,
and A. Inkeri Verkamo:
Fast discovery of association rules.
Chapter 12 in Usama M. Fayyad, Gregory Piatetsky-Shapiro, Padhraic
Smyth, and
Ramasamy Uthurusamy, editors,
Advances in Knowledge Discovery and Data Mining,
1996. AAAI Press.
- 3
-
Heikki Mannila, Hannu Toivonen, and A. Inkeri Verkamo:
Discovery
of frequent episodes in event sequences.
Report C-1997-15, University of Helsinki, Department of Computer
Science,
February 1997.
A revised version is to appear in
Data
Mining and Knowledge Discovery, 1997.
- 4
-
Tomasz Imielinski and Heikki Mannila: A Database Perspective on
Knowledge Discovery.
Communications of the ACM, November 1996.
- 5
-
Heikki Mannila and Hannu Toivonen:
Levelwise search and borders of theories in knowledge discovery
Report C-1997-8, University of Helsinki, Department of Computer Science,
January 1997.
A revised version is to appear in
Data
Mining and Knowledge Discovery, 1997.
- 6
-
Heikki Mannila:
Inductive databases and condensed representations for data mining.
, International Logic Programming Symposium, 1997, to appear.
Ziel des Praktikums ist eine Einführung in Methoden zur Modellierung
ausgewählter Aspekte des kooperativen Problemlösens. Ausgehend von
grundlegenden Techniken der Metaprogrammierung werden schrittweise
komplexer werdende Systeme entwickelt, welche die Modellierung eines
wissensbasierten Informationsaustausches während des kooperativen
Problemlösens gestatten. Die Entwicklung der verschiedenen Modelle
erfolgt mit Hilfe der Programmiersprache Prolog. Vorkenntnisse in der
Programmierung mit Prolog sind wünschenswert, aber keine
Voraussetzung.
- Teil I: Einführung.
- Einführung in (1) ausgewählte Aspekte
des Informationsaustauschs beim kooperativen Problemlösen, (2) die
kognitive Modellierung aus psychologischer Sicht und (3) die
Entwicklung wissensbasierter Systeme mit Hilfe der
Programmiersprache Prolog.
- Teil II: Metaprogrammierung.
- Es werden grundlegende Techniken
der Metaprogrammierung in Prolog und deren Anwendungen vorgestellt.
Ausgehend von einfachen Metaprogrammen werden Erweiterungen zur
Erzeugung von (generalisierten) Beweisbäumen sowie zur
Durchführung kognitiver Selbstdiagnosen behandelt.
- Teil III: Modellierung des Informationsaustauschs beim
kooperativen Problemlösen.
- Vorgestellt wird ein Metaprogramm, das
einzelne Aspekte des Informationsaustauschs beim kooperativen
Problemlösen auf Grundlage kognitiver Selbstdiagnosen zu
modellieren gestattet. Im Vordergrund wird dabei die wissensbasierte
Erzeugung von Fragen an den Problemlösepartner stehen.
Abschließend soll auf dem Hintergrund der behandelten
Modellierungen die Frage diskutiert werden, wie Lernen, das während
des kooperativen Problemlösens stattfindet, in den Modellen
berücksichtigt werden kann.
References
- 1
-
Opwis, K. & Plötzner, R. (1996). Kognitive Psychologie mit dem
Computer - Ein Einführungskurs zur Simulation geistiger Leistungen
mit Prolog. Heidelberg: Spektrum Akademischer Verlag.
- 2
- Plötzner, R., Fehse, E., Hermann, F. & Kneser, C.
(in press). Modeling the knowledge-based exchange of
information during collaborative problem solving on the basis of
deductive self-diagnosis. Proceedings of the Eigth World
Conference on Artificial Intelligence in Education.
Steven Payne: Cognition in Human-Computer Interaction
This course reviews a programme of work on
the cognitive science of human-computer interaction (HCI). The main
reason for the existence of this research area is a widely-held
belief that the design of usable and learnable computer systems
might be informed by an applied science of user cognition. I will
try to support this argument, but I will also try to argue that the
study of human-computer interaction can and should play a key role
in the theoretical development in cognitive psychology.
The first lecture will argue that the design of technologies can
benefit from cognitive psychology, and that the science of
psychology can benefit from serious consideration of technologies.
It will discuss some of the high-level arguments that have been made
in favour of both propositions, and it will review some successful
cases from the HCI literature. The argument will incorporate some
empirical work that shows that peoples intuitions about the
usability of designs do not always accord with well-founded human
factors principles, especially stimulus-response compatibility.
References I
- 1
- Payne, S. J. (1996). Cognitive psychology
and cognitive technologies. The Psychologist, 9, 309-312.
- 2
- Payne,S. J. (1995). Naive judgments of
stimulus-response compatibility. Human Factors, 37,
495-506.
The second lecture will review attempts to provide cognitive
descriptionsöf user interfaces. Such descriptions can be thought of
as cognitive models of the knowledge needed to use (an aspect of) a
computer system. The idea is that such descriptions can be designed
and inspected for usability in advance of a system being implemented.
The most celebrated representation for cognitive description is the
GOMS model of Card, Moran and Newell. I will briefly review the GOMS
approach before describing Task-Action Grammars (TAGs), which attempt
to predict the relative learnability of different user interfaces. I
will review recent attempts to incorporate TAG into computational
learning models, as well as ongoing attempts to generate TAGs
automatically from user manuals.
References II
- 1
- Payne, S. J. & Green, T. R. G. (1986).
Task-action grammars: a model of the mental representation of task
languages. Human-Computer Interaction, 2, 93-133.
- 2
- Howes, A. and Young, R. M. (1996). Learning
consistent, interactive and meaningful task-action mappings: a
computational model. Cognitive Science, 20,
301-356.
Whereas the second lecture is concerned with the knowledge users need
to work a machine, the third and fourth are concerned with users'
knowledge of how machines work: their theme is mental models. I will
describe a theory of the role and content of mental models of computer
systems, and compare HCI work on mental models with work from the
mainstream cognitive literature. I will illustrate my arguments with
some descriptive work on the formation of mental models by naive
users, and some experimental work on the internalisation of models
that are provided by instructions. Finally in these lectures I will
depart from HCI to defend the proposal that memory for a mental model
is often dominated by memory for the processes that construct and
consult the model. This theme of process-based memory will then
re-emerge in the final lecture.
References III + IV
- 1
- Payne, S. J., Squibb,H. R. and Howes, A. (1990).
The nature of device models: The yoked state space
hypothesis and some experiments with text editors.
Human-Computer Interaction, 5, 415-444.
- 2
-
Payne, S. J. (1991a).
A descriptive study of mental models.
Behaviour & Information Technology, 10(1), 3-21.
- 3
-
Payne, S. J. (1993).
Memory for mental models of spatial descriptions: An
episodic-construction-trace hypothesis.
Memory & Cognition, 21(5), 591-603.
- 4
- Bibby, P. A. & Payne, S. J. (1996).
Instruction and practice in learning about a device.
Cognitive Science , 20, 539-578.
The fifth lecture will deal with the issue of display-based problem
solving. When computer users perform tasks with modern user
interfaces, they engage in a particular style of problem solving which
is characterized by the vital role of an external information display.
I will review empirical evidence that even expert users performing
routine tasks rely to a large extent on information flow from device
to user. I will describe several recent experiments that throw light
on the nature of such display-based problem solving and, at the same
time, address current controversies in recognition memory. One theme
of this work is that display-based contexts can lead to exploratory
strategies that are adaptive in the short-term, but which undermine
longer-term learning.
References V
- 1
-
Payne, S. J. (1991).
Display-based action at the user interface.
International Journal of Man-Machine Studies, 35(3),
275-289.
- 2
-
Trudel, C.-I. & Payne, S. J. (1995).
Reflection and goal management in exploratory learning.
International Journal of Human-Computer Studies, 42(3),
307-339.
Frank Ritter & Richard Young: Introduction to the Soar Cognitive Architecture and Example Models that Learn
This tutorial will cover the fundamentals of the Soar architecture and
present two advanced models that learn. Soar is a cognitive
architecture built around multiple problem spaces for representing
knowledge and implemented as a production system. It includes a
simple built-in learning mechanism called chunking that has been used
to realise several complex learning mechanisms. Soar has been
proposed by Allen Newell as a candidate ``unified theory of cognition''.
It has been used to model behavior in natural language processing,
planning, HCI, abductive reasoning, and various laboratory tasks. The
tutorial will be of interest to cognitive psychologists, cognitive
scientists, and HCI user modellers. The emphasis will be on the
cognitive implications of Soar rather than its AI and Knowledge
Engineering applications. There will be facilities to run and modify
Soar programs, and also opportunities to discuss the psychological
implications of Soar and other topics of interest to the audience.
The most relevant text for the tutorial is the book by Allen Newell,
``Unified Theories of Cognition'' (Harvard University Press, 1990).
Waldrop's articles in Science (p. 27-29 & 296-298, vol. 241, 1988) are
far shorter and provide a short but correspondingly shallow
introduction. The BBS target article (p. 425-492, vol. 15, 1992) is a
compromise between them.
Some Background Readings
- 1
-
Newell, A. (1990).
Unified Theories of Cognition.
Cambridge, Massachusetts: Harvard University Press.
- 2
-
Newell, A. (1992).
Precis of unified theory of cognition.
The Behavioral and Brain Sciences, 15(3), 425-492.
With Open Peer Commentary and Author's Response.
- 3
-
Waldrop, M. M. (1988a).
Soar: A unified theory of cognition?
Science, 241, 296-298.
- 4
-
Waldrop, M. M. (1988b).
Toward a unified theory of cognition.
Science, 241, 27-29.
We expect attendees to be familiar with the concept of modelling
cognition. Given sufficient notice of who will attend, we will post
them introductory articles.
- Lectures covering the basic concepts of Soar as a unified theory
of cognition, Soar as a problem space architecture (spaces, states,
operators, impasses, chunks, and so on), and Soar as a programmable
architecture (production rules, preferences, elaboration and
decision cycles, attributes, default rules, the chunking mechanism,
etc.).
- Practical experience running a number of supplied programs, and
carrying out a series of graded examples involving running the
programs in different ways, modifying the programs slightly,
extending them with new rules, and debugging further examples.
- Question answering and discussion about Soar, its relation to
cognition, technical aspects of Soar programming, and the
applicability of Soar to problems of interest to the audience.
However, the tutorial will NOT be organised into three blocks
(lectures, practical, discussion). Rather, after an initial
orientation and introduction, we will intersperse hands-on practice,
discussion, and further lectures. We will also relate aspects of Soar
theory to the corresponding aspects of the programming, and present
two more advanced models that learn, a diagrammatic reasoning model
matches aggregate behavior (with an r2 > .9) and protocol data, and a
model of menu exploration.
There now also exists a hypertext version of the tutorial. It will be
on the diskette taken away from the tutorial by each student, and will
also be made use of during the tutorial itself.
References
- 1
- Ritter, F. E., & Larkin, J. H. (1994).
Using process models to summarize sequences of human actions.
Human Computer Interaction. 9 (3). 345-383.
- 2
- Ritter, F. E., & Bibby, P. (1997).
Modelling learning as it happens in a diagramatic reasoning task
(Tech. Report No. 45). ESRC CREDIT, Dept. of Psychology, U. of
Nottingham.
- 3
- Nerb, J., Krems, J., & Ritter, F. E. (1993).
Rule learning and the power law: A computational model and
empirical results. In Proceedings of the Fifteenth Annual Conference
of the Cognitive Science Society. 765-770. Hillsdale, New Jersey:
LEA.
- 4
- Ritter, F. E. & Young, R. M. (1994).
Practical introduction to the Soar cognitive architecture: Tutorial
report. AI and Simulation of Behavior Quarterly, 88, 62.
- 5
- Young, R. M. (1979). Production systems for
modelling human cognition. In D. Michie (Ed.) Expert Systems in the
Micro-Electronic Age. Edinburgh University Press, 35-45.
[Reprinted in E. Scanlon & T. O'Shea (Eds.) Educational Computing,
Wiley, 1987. 209-220.]
- 6
- Rieman, J., Young, R. M. & Howes, A. (1996).
A dual-space model of iteratively deepening exploratory
learning. International Journal of Human-Computer
Studies, 44, 743-775.
- 7
- Howes, A. and Young, R. M. (1996). Learning
consistent, interactive and meaningful task-action mappings: a
computational model. Cognitive Science, 20,
301-356.
This course will focus on the development of thinking, reasoning,
and problem solving. It will present both traditional and
contemporary approaches to issues in these areas. Students will be
encouraged to relate the specific research and ideas on children's
thinking that I present to issues of relevance in their own areas of
specialization. The course will include three sessions:
- Session 1: How Do Children Decide What to Do?
- The first
session will focus on general issues regarding strategy choice.
These issues have traditionally been approached in terms of
discontinuous changes between qualitatively different types of
thinking. This is true regardless of whether the particular
theorist proposed changes in children's general stages of
reasoning (e. g., Piaget, Vygotsky), their domain-specific theory
(e.g., Wellman, Carey) or their basic form of representation
(Bruner, Case). However, a great deal of recent data indicates
that these approaches underestimate the variability in children's
thinking, where more and less mature ways of thinking typically
coexist and compete for prolonged periods of time. This
underestimation of cognitive variability, in turn, leads to
underestimation of the importance of choice for understanding
children's thinking. Both of these problems interfere with our
efforts to understand the most important single issue about
development: How change occurs. After considering these issues,
we then will consider methods for assessing which strategy someone
is using to solve a particular problem. The session will conclude
with presentation of a computer simulation model, based on these
ideas, that accounts for a wide range of data regarding the
development of strategy choices.
- Session 2: How Do Children Discover New Strategies?
-
In the second session, we will examine the process of strategy
discovery. The methods for assessing strategy use on a trial by
trial basis, described in Session 1, allow identification of
exactly when a child discovers a new strategy. Once the trial of
discovery is identified, we can examine what led up to the
discovery and how it was generalized once it was made. Studies
examining strategy discovery on a trial by trial basis have led to
very consistent findings, many of which are quite surprising. For
example, even when children generate insights that lead to
bneeficial new strategies, they continue to use inferior old
strategies as well. (Adults are no different, as will also be
discussed.) Also, children often discover new approaches even
though they have been succeeding using previous approaches. Among
the other topics considered in this session will be the effects of
consciousness on discovery and generalization of new strategies.
As in Session 1, we will conclude by discussing a recent computer
simulation, this one modeling how children discover a new
strategy.
- Session 3: How Can We Promote Strategy Discoveries?
- The
final session will focus on ways of increasing the likelihood of
children discovering and appropriately generalizing new
strategies. An approach that will receive particular attention is
encouragment to explain the reasoning of other, more knowledgable
people. This method has been shown to be effective with people of
all ages from 5 years to adulthood. It combines the major
advantage of traditional, didactic instruction (efficiency) with
the major advantage of discovery-oriented learning (excitement and
deep understanding). Research on self-explanation has shown that
children benefit from explaining why correct answers are correct,
and that they benefit even more from explaining both why correct
answers are correct and why incorrect answers are wrong. The
session will conclude with an overview of the issues considered in
the brief course, and with consideration of issues likely to be of
great importance in the future.
References
- 1
- Siegler, R. S. (1996). Emerging
Minds. New York: Oxford University Press.
This recent book presents the main themes I will be discussing.
- 2
- Siegler, R. S. (1995). How does change
occur: A microgenetic study of number conservation. Cognitive
Psychology, 28, 225-273.
This article focuses on strategy discovery and on methods for studying
it. It also describes the self-explanation approach that will be the
focus of Session 3.
- 3
- Siegler, R. S., & Shipley, C. (1995).
Variation, selection, and cognitive change. In T. Simon
and G. Halford (Eds.) Developing cognitive competence: New
approaches to process modeling. Hillsdale, NJ: Erlbaum.
This chapter presents the current computer simulation model of
strategy choice, which will be discussed in Session 1.
Born 1966. B.A. at Lund University 1990. Ph.D. in Cognitive Science 1995.
Currently research assistant at Lund University Cognitive Science. His
primary interest is neural network modelling of various forms of cognitive
processes, including sequential processing, induction, categorisation,
motivation and action selection in autonomous agents, as well as spatial
learning, conditioning and habituation. Another of his research areas is
how representations can be described from different perspectives at both
the conceptual and the neural level and the emergent properties in networks
of interacting systems for motivation, perception, and action. He is
currently working with computer simulations of neural network controlled
artificial creatures and with visually controlled mobile robots.
Here is his homepage: Christian Balkenius
Karl R. Gegenfurtner, geboren am
7. Juli 1961 in Straubing (Bayern). 1981-1986 Studium der Psychologie
in München und Regensburg. 1986 Diplom in Psychologie an der
Universität Regensburg. 1986-1990 Ph.D. Programm für Experimentelle
Psychologie an der New York University. 1990 Ph.D. 1990-1993 PostDoc
am Center for Neural Science an der New York University. seit 1993 am
MPI fuer biologische Kybernetik (Abt. Bülthoff).
Forschungsschwerpunkte: Visuelles System des Menschen, insbesondere
Farb- und Bewegungswahrnehmung. Methoden: Psychophysik, Physiologie.
Here is his homepage: Karl Gegenfurtner
Udo Hahn ist Professor für Linguistische
Informatik und Computerlinguistik an der Universität Freiburg.
Er arbeitet an der Entwicklung von
Textverstehenssystemen, deren Eingabe authentische Texte
sind. Ein Schwerpunkt seiner Arbeiten liegt auf der
Entwicklung von computerlinguistischen Modellen und
Systemarchitekturen, mit denen die dabei auftretenden
natürlichsprachlichen Performanzphänomene angemessen erfaßt
werden können.
Here is his homepage: Udo Hahn
Joachim Hertzberg, Jahrgang 1958, hat in
Braunschweig und Bonn Informatik studiert, in Bonn 1986
promoviert und sich in Hamburg 1995 habilitiert. Er ist -
durch Forschungs- und Lehraufenthalte unterbrochen - seit
1986 Mitarbeiter der GMD. Sein Forschungsgebiet ist KI,
insbesondere Handlungsplanung, Schließen über Zeit,
logisches Schließen über Wandel und Constraint-basiertes
Schließen. Derzeit arbeitet er zur Anwendung von KI-Methoden
in der Service-Robotik.
Here is his homepage: Joachim Hertzberg
Keith J. Holyoak received his PhD in
psychology from Stanford University in 1976. Currently he is
professor for cognitive psychology at University of California,
Los Angeles (UCLA). His research interests are in the
general area of reasoning and problem solving. Much of his
work is concerned with the role of analogy in thinking. One of
the major themes of this work is the way in which analogy
serves a psychological mechanism for learning and transfer of
knowledge.
Here is his homepage: Keith Holyoak
Kurt Konolige is a Senior (born 17/11/53)
Computer Scientist at the Artificial Intelligence Center of SRI
International, and a Consulting Professor in Computer Science at
Stanford University. His research interests are broadly based on
issues of commonsense reasoning, fuzzy control for reactive
systems, constraint-based planning and inference systems, reasoning
about perceptual information, and realtime robotics and vision
systems.
Here is his homepage: Kurt Konolige
Dr. Pat Langley was born in 1953 and
received his PhD in cognitive psychology from Carnegie Mellon
University in 1979, where he studied with Herbert Simon. Since then,
he has worked in academia (at Carnegie Mellon and the University of
California, Irvine), in government (NASA Ames Research Center), and in
industry (Siemens Corporate Research). Dr. Langley has published
widely on the topics of machine learning and scientific discovery,
including the recent text Elements of Machine Learning, and he is a
well-known advocate of the application and experimental study of
learning algorithms. He is an editor of the journal Machine Learning
and edits the Morgan Kaufmann series on that topic. He currently heads
the Intelligent Systems Laboratory at Daimler-Benz Research and
Technology Center, he serves as Director of the Institute for the
Study of Learning and Expertise, and he is a Consulting Professor at
Stanford University.
Here is his homepage: Pat Langley
Heikki Mannila was born in 1960. He is a professor of
computer science at the University of Helsinki. He has been an
associate professor at the Universities of Tampere and Helsinki, a
visiting professor at the Technical University of Vienna, and a guest
researcher at the Max Planck Institut fuer Informatik in Saarbruecken.
He has done research on algorithms, logic programming, and database
design. Recently, his research has concentrated on of data mining,
especially on rule discovery from large data sets, with applications in
telecommunications, medical genetics, etc. His research interest
include also the theory of data mining, computational data analysis, and
text databases. He is an editor-in-chief of the Data Mining and
Knowledge Discovery journal.
Here is his homepage: Heikki Mannila
Klaus Opwis studierte an der Universität Heidelberg Psychologie
und Mathematik. 1982 wechselte er nach Freiburg, wo er 1985 promoviert und 1991 habilitiert wurde. Seit vier Semestern vertritt er
die Professur fuer Allgemeine Psychologie und Methodologie an der
Universität Basel. Seine Arbeitsschwerpunkte liegen in den Bereichen
Kognitive Psychologie (Denken/Problemlösen und Gedächtnis) sowie
kognitionswissenschaftliche Methodologie.
Here is his homepage: Klaus Opwis
Stephen James Payne.
Born 24 December 1956.
Education:
1980 Loughborough University of Technology:
BSc Ergonomics
1985 Sheffield University: PhD Psychology.
Career:
1984-1989 Lancaster University, Departments of
Psychology and Computing. Lecturer.
1989-1991 IBM Thomas J. Watson Research Center, Research Staff Member
Yorktown Heights, New York, Computer Science Department.
1991- University of Wales, College of Cardiff,
School of Psychology. Currently Professor.
Research Interests: The cognitive psychology of human-computer interaction. Human problem solving and
learning.
Here is his homepage: Steven Payne
Rolf Plötzner studierte an der Universität
Freiburg Psychologie und Mathematik. Nach einem einjährigen
Forschungsaufenthalt an der University of Pittsburgh (U.S.A.)
schloss er seine Promotion 1993 als Dr. phil. an der Universität
Freiburg ab. Zur Zeit ist er als Wissenschaftlicher Assistent am
Psychologischen Institut der Universität Freiburg tätig. Seine
Arbeitsschwerpunkte liegen in den Bereichen (kooperatives)
Problemlösen und Lernen sowie Anwendung wissensbasierter Systeme.
Here is his homepage: Rolf Plötzner
Frank Ritter was awarded a doctorate by
Carnegie-Mellon University in AI and psychology. As his thesis he
developed a methodology and software for testing the sequential
predictions of Soar models (and cognitive models in general). As part
of this work, he built an interface to Soar. He has worked with
the Soar architecture since 1987, and has recently thought about the
implementation of emotions in cognitive architectures. He is a
lecturer in psychology, an associate lecturer in computer science at
the University of Nottingham, and co-chairman of the 2nd European
cognitive modelling workshop to be held in Nottingham in April 1998.
Here is his homepage: Frank Ritter
Year of birth: 1949 PhD: SUNY at Stony
Brook, 1974. Academic Appointments: 1974-1997 Psychology Dept,
Carnegie Mellon University. Current Position: Heinz Professor of
Psychology, Carnegie Mellon University. Research Interests: Development
of scientific and mathematical thinking in children. Development of
strategy choice and strategy discovery. Computer simulation models of
cognitive change. Application of research findings to improving
education.
Here is his homepage: Robert Siegler
Geboren 1944 in Wien. Dort Studium der
Psychologie mit Doktorat 1969. Habilitation in Wien und Kiel 1976.
Mehrere Jahre am Institut für die Pädagogik der
Naturwissenschaften in Kiel. Seit 1980 Lehrstuhl für Psychologie an
der Universität Freiburg. Sprecher des Freiburger
Graduiertenkollegs und des ESF Schwerpunktprogramms "Learning in
Humans and Machines". Forschungsschwerpunkte: Modelle des Denkens
und Lernens, Umweltpsychologie.
Here is his homepage: Hans Spada
Richard M Young has recently moved to the
Psychology Department at the University of Hertfordshire, after
working for many years at the Medical Research Council's Applied
Psychology Unit in Cambridge. He gained his PhD in Psychology in 1973
from Carnegie Mellon University, where he was a student of Allen
Newell. His research lies in the general area of cognitive science
and cognitive modelling, and especially in the topic of cognitive
architectures. He is interested in the modelling of complex, everyday
tasks, including those involving human-computing interaction. He has
followed the development of Soar since its emergence in 1982, and has
been actively working with Soar since 1987.
Here is his homepage: Richard Young
Posterpresentations
The members of the ``Graduiertenkolleg''
(graduate college) Intelligence in Humans and Machines at the
University of Freiburg will present
posters of their ongoing work during extendet coffee breaks.
Of course, it is possible to catch people during the whole Autumnschool.
General Information
Timetable (preliminary)
Freiburg is a small university town with ``high touristic
value'' in the very South-West of Germany; the state is called
Baden-Württemberg. Freiburg has around 200.000 inhabitants. The city
is located right in front of the Black Forest, with France and
Switzerland very close by. The university is mid-sized (about 23.000
students) and has a long history, dating back more than 600 years.
