Artificial Intelligence/Cognition

Research Group


Compendium of
Research & Teaching Activities

March 1986 - May 1991




Institut für Informatik

Technische Universität München





Department of Computer Science (Informatics)
Technische Universität München

Research Unit on Theoretical Informatics
and Foundations of Artificial Intelligence

Director
Prof. Dr. Wilfried Brauer



Artificial Intelligence/Cognition Research Group




0 Table of Contents

  1. Members
    1. Researchers
    2. Student Employees
    3. Research Associates
    4. Guests
  2. Work in University
    1. Advanced Practical Work Done by Students
    2. Diploma Theses
    3. Dissertation
    4. Courses and Seminars
  3. Talks
  4. Publications
    1. Technical Reports (Forschungsberichte Künstliche Intelligenz)
    2. Scientific Publications



1 Members

1.1 Researchers

José Manuel Benedetti Nov. 89 - July 90
Petra Bräunling Sept. 89 - May 90
Marco Dorigo July 89 - Feb. 90
Christian Freksa, Ph.D. (Group leader) March 86 - May 91
Daniel Hernández since Jan. 88
Erwin Klöck Aug. 86 - Oct. 90
Matthias Marcinowski Dec. 87 - Dec. 89
Erwin Praßler Feb. 86 - March 89
Bernhard Schätz since Nov. 89
Jürgen Schmidhuber Nov. 88 - June 91
Gerhard Weiß since Jan. 90
Benjamin Werner April 89 - June 89
Dekai Wu Nov. 86 - June 87



1.2 Student Employees

Anton Beschta Aug. 87 - Sept. 90
Petra Bräunling June 86 - May 90
Anton Hartl Sept. 87 - Dec. 88
Paul Jetschny since Feb. 90
Sabine Karl Oct. 89 - March 90
Angela Marquardt Nov. 88 - April 91
Gerd Pauli since May 90
Thomas Röll since Nov. 89
Kai Zimmermann since Aug. 86



1.3 Research Associates

Gerhard Dirlich (Max-Planck-Institut für Psychiatrie, München)
Ulrich Furbach (Technische Universität München, Artificial Intelligence/Intellectics Research Group)
Manfred Gehrke (Siemens AG, München)
Thomas Laußermair (Siemens AG, München)
Kerstin Schill (Institut für Medizinische Psychologie, Ludwig-Maximilians-Universität, München)
Thomas Waschulzik (Kratzer Automatisierung GmbH, München)



1.4 Guests

Josep Aguilar-Martin LAAS du CNRS, Toulouse
Doug Appelt SRI International, Menlo Park
Barbara Becker GMD, St. Augustin
Harold Boley DFKI, Kaiserslautern
Nguyen Cat Ho National Center for Scientific Research of Vietnam, Hanoi
Lin Chen Beijing Laboratory of Cognitive Science
Thomas Christaller GMD, St. Augustin
Ulises Cortés Universitat Politècnica de Catalunya, Barcelona
Hartmut Dörner Universität Halle-Wittenberg
Scott Fahlman Carnegie-Mellon-University, Pittsburgh
Hans Werner Güsgen GMD, St. Augustin
Joachim Hertzberg GMD, St. Augustin
Michael Herweg Universität Hamburg
Markus Höhfeld Universität Karlsruhe
Petra Lichtenberg University of California, Irvine
R. López de Mántaras Centre d`Estudis Avançats de Blanes
Clayton Mc Millan University of Colorado, Boulder
Dieter Metzing Universität Bielefeld
Jürgen Müller Universität Kaiserslautern
Boris Petkoff Bulgarian Academy of Science
Enrique Ruspini SRI International, Menlo Park
Christoph Schlieder Universität Hamburg
Paul Smolensky University of Colorado, Boulder
Stefan Wermter Universität Dortmund






2 Work in University

2.1 Advanced Practical Work Done by Students

Reimplementation of the Left-Associative Grammar by Dr. Hausser, WS 88/89
Advisers: C. Freksa, E. Klöck
Student: Gerald Klix

Implementation of an User-Friendly Interactive Environment for Connectionist Network Modells in LOOPS and INTERLISP-D of an Xerox-Workstation, SS 88
Advisers: C. Freksa, E. Klöck
Student: Kai Zimmermann

Porting LISP-Dialects: General System and Concrete Example, SS 88
Adviser: D. Hernández
Students: Bernhard Brandmair, Rainer Lichtenfeld

Performance Comparison of LISP-Environments, SS 88
Adviser: D. Hernández
Student: Mario Myrenne

Maisy-Mailing System for Hypercard/Macintosh and Ultrix/Vax., SS 88
Adviser: D. Hernández
Students: Bernhard Schäfer, Alfred Jägel

Further Development of a Compiler for Neural Networks, WS 89/90
Adviser: T. Waschulzik
Student: Martin Eldracher

2- and 3-Dimensional Function Plotting in Common Lisp, SS 89
Adviser: E. Praßler
Student: Anton Beschta

Implementation of Different NEOCOGNITRON-Models for Position- Independent Recognition of Images (in Pascal for an Apple MAC II), SS 89
Advisers: C. Freksa, B. Kämmerer (Siemens)
Student: Stefan Lanser

Handling Diagrammatical Representations Using Funt's Parallel Retina, SS 90
Adviser: D. Hernández
Student: Thomas Hahn

An Implementation of Allen's Time Logic in Scheme, SS 90
Adviser: D. Hernández
Student: Roland Bergler

Expansion of a Test System for Neural Networks, WS 90/91
Adviser: T. Waschulzik
Student: Brigitte Menacher

Construction of Hebb-Networks with Reinforcement Input, WS 90/91
Adviser: G. Weiß
Student: Marcus Hutter

Implementation and Application of Neural Real-Time Algorithm for Reactive Environments, WS 90/91
Adviser: J.H. Schmidhuber
Student: Josef Hochreiter

Performance Comparison of Scheme-Environments and Acquirement of General Increasements, WS 90/91
Adviser: D. Hernández
Student: Christian Hohendorf



2.2 Diploma Theses

Ralf Kahler
Untersuchungen zur Anwendung selbstorganisierender assoziativer Netzwerke für komplexe Zugriffe auf Datenbanken, WS 86/87

Thomas Krempl
Untersuchungen zur semantischen Analyse natürlichsprachlicher Datenbankabfragen mit Hilfe assoziativer Netzwerke, WS 86/87

Thomas Hofbauer
Entwicklung eines Prototypen für ein wissenbasiertes Projekt-Verwaltungs- und Projekt-Management-Support-System, WS 87/88

Thomas Waschulzik
Optische Mustererkennung in neuronalen Architekturen, WS 87/88

Andreas Stolcke
Generierung natürlichsprachlicher Sätze in unifikationsbasierten Grammatiken, SS 88

Dagmar Böller
Realisierung eines hierarchischen neuronalen Netzwerks zur assoziativen Erkennung von Objekten in natürlicher Umgebung, WS 88/89

Karl-Heinz Krachenfels
Lernen von Regeln in neuronalen Netzen, SS 89

Bernhard Schätz
Ein konnektionistisches Modell zur Schemaadaption, SS 89

Christian Schmitt
Implementierung eines Expertensystems mit Informationszuwachsstrategie auf der Basis der Evidenztheorie von Dempster und Shafer, SS 89

Udo Wings
Objektrepräsentation und -benennung mit konnektionistischen Netzen, SS 89

Norbert Zeßel
Sprachverarbeitung mit konnektionistischen Netzen, SS 89

Martin Arnoldi
Mustererkennung mit Hilfe selbstorganisierender, adaptiver Mechanismen, WS 89/90

Meinrad Walzer
Entwicklung eines Software-Systems für den computerunterstützten Entwurf und die Ablaufsteuerung von Planungsaufgaben in der neuropsychologischen Rehabilitation, WS 89/90

Sylvester Claudio Diehl
Untersuchungen zur Informationscodierung und Verarbeitung in konnektionistischen Systemen, SS 90

Hans-Peter Dommel
Konzeptklassifikation und Generalisierung in konnektionistischen Netzen - Modelle humanen und maschinellen Lernens im Vergleich, SS 90

Jürgen Hollatz
Lernen und Erkennen in dynamisch stabilen Netzen, SS 90

Renate Schuster
Konzeption und Implementation einer Wissensbasis unter Einbezug der strukturellen Abhängigkeiten von Expertenwissen, SS 90

Klaus Eder
Satzerkennung mit konnektionistischen Methoden, WS 90/91

Martin Eldracher
Klassifikation großer Datenmengen anhand dynamisch extrahierter relevanter Merkmale mit Hilfe neuronaler Netze, WS 90/91

Rudolf Huber
Selektive visuelle Aufmerksamkeit: Untersuchungen zum Erlernen von Fokustrajektorien durch neuronale Netze, WS 90/91

Bernd Janosch
Dynamik und Plastizität in Netzen oszillierender Neuronen, WS 90/91

Margit Kinder
Repräsentation mehrdimensionaler funktionaler Abhängigkeiten in neuronalen Netzen, WS 90/91

Peter Menke-Glückert
Verallgemeinerung einer Netzwerkbeschreibungssprache für die Implementierung biologischer Eigenschaften, WS 90/91

Helga Menz
Kontextsensitives zeitliches Schlußfolgern, WS 90/91

Patrick Thomas
Populationsaspekte bei der Informationsverarbeitung selbstorganisierender neuronaler Netze, WS 90/91

Klaus Bergner
Untersuchungen zu neuronalen adaptiven Kritikern, SS 91

Bernhard Brandmair
Animation analogischer Wissensstrukturen, SS 91

Josef Hochreiter
Untersuchungen zu dynamischen neuronalen Netzen, SS 91

Marcus Hutter
Lernen in Klassifizierungssystemen, SS 91

Alfred Jägel
Erweiterung eines Neuronenmodells um einen strukturierten Dendritenbaum, SS 91

Evelyn Kliment
Kosten-Nutzen-Raum bei medizinischen Expertensystemen, SS 91

Arthur Steiner
Prototypische Signalverarbeitung zur sprecherabhängigen Einzelworterkennung mit neuronalen Netzen, SS 91

Kai Zimmermann
SEqO - Ein System zur Erforschung qualitativer Objektrepräsentationen, SS 91



2.3 Dissertation

J. H. Schmidhuber
Dynamische neuronale Netze und das fundamentale raumzeitliche Lernproblem, 1990



2.4 Courses and Seminars

Brauer, Freksa
Seminar: Representation of Space and Time in the field of Artificial Intelligence, WS 86/87

Freksa, Klöck, Praßler
Practical course: Development of AI-Programs using LISP (with Schneeberger), WS 86/87

Brauer, Freksa, Schneeberger
Seminar: Representation of Incomplete and Fuzzy Knowledge in the field of Artificial Intelligence, SS 87

Freksa, Klöck
Practical course: Introduction to Lexical Functional Grammar, SS 87

Brauer, Freksa, Praßler
Seminar: Cognition and Planning, WS 87/88
Practical course: Planning Techniques in Artificial Intelligence, WS 87/88

Brauer, Freksa, Hernández
Practical course: Methods of AI, Connectionism, SS 88

Lange, Freksa, Hernández, Marcinowski, Schmidhuber
Seminar: Connectionism, SS 88

Brauer, Freksa, Marcinowski
Practical course: Methods of AI, Connectionism, WS 88/89

Brauer, Freksa, Marcinowski, Schmidhuber
Seminar: Learning Neural Networks and Associative Memories, WS 88/89

Brauer, Freksa, Hernández
Seminar: Foundation of Knowledge Representations in the Cognitive Science, SS 89

Brauer, Klöck
Practical course: Methods of AI: Connectionism, SS 89 and WS 89/90

Brauer, Hernández, Zimmermann
Practical course: Methods of AI - Introduction to the LISP Programming Language, WS 89/90

Brauer, Freksa, Hernández, Schmidhuber, Weiß
Seminar: Highly Parallel Algorithms for Selforganization and Learning, WS 89/90

Brauer, Freksa, Hernández
Seminar: Representation of Space, SS 90

Brauer, Freksa, Schmidhuber, Weiß
Seminar: Learning and Selforganization in Highly Parallel Systems, SS 90

Brauer, Freksa, Schätz
Practical course: Methods of AI: Connectionism, SS 90 and WS 90/91

Brauer, Freksa, Jost, Schmidhuber
Seminar: Learning Neural Robots, WS 90/91

Brauer, Freksa, Laußermair, Weiß
Seminar: Artificial Life, WS 90/91






3 Talks

Brauer Brauer, Freksa, Klöck, Hernández, Marcinowski, Praßler Freksa Freksa, Bräunling, Zimmermann Hernández Hernández, Freksa Klöck Marcinowski Praßler Schätz Schmidhuber




4 Publications

4.1 Technical Reports (Forschungsberichte Künstliche Intelligenz)

FKI-84-88: Freksa, C.: Cognitive Science - eine Standortbestimmung. - Published in: Wissensarten und ihre Darstellung. Hrsg.: Heyer, G.; Krems, J.; Görz, G., Heidelberg: Springer, 1988, S. 1-12 (Informatik-Fachberichte). Also published in: Universitas 525, März 1990, S. 232-240.

The paper sketches the formation of the field Cognitive Science in the 1970s as an umbrella discipline for aspects of psychology, linguistics, artificial intelligence, anthropology, philosophy, and the neurosciences. The example of color naming is used to demonstrate the interdisciplinary approach to cognition. The AI phenomenon is considered from a cognitive point of view. Important topics and first results of Cognitive Science and the contributions of the subdisciplines are summarized.


FKI-85-88: Freksa, C. : Intrinsische vs. extrinsische Repräsentation zum Aufgabenlösen oder die Verwandlung von Wasser in Wein. - Published in: Wissensarten und ihre Darstellung. Hrsg.: Heyer, G.; Krems, J.; Görz, G., Heidelberg: Springer, 1988 (Informatik Fachberichte).

Real world problem solving is presented as a task that can be performed either by an action in the real world or by formalization, inference, interpretation in an artificial world. The paper focuses on the difficulty of constructing adequate formalizations of real world problems. Properties of relations and their intrinsic and extrinsic representation are discussed. The distinction between simualtion and explanation of real word behavior is drawn. The formalization problem is exemplified by a human approach to the wine mixing problem.


FKI-88-88: Hernández, D.: Module Fault Localization in a Software Toolbus Based System.

We incorporate a fault localization capability into POLYLITH ([Purtilo 86]), a system that supports the interconnection of heterogenous software modules. To this end we apply techniques developed in the context of diagnosis of technical systems. These techniques are based on explicit descriptions of the structure and behavior of the system to be diagnosed. The POLYLITH Module Interconnection Language (MIL) originally provides a description of software interconnectivity (structure), which is enhanced here by attributes specifying the high level behavior of the modules. Furthermore, the run-time support by the POLYLITH software bus gives us access to the actual behavior of the system under consideration. Given enough information we are then able to determine a module or set of modules that must be faulty in order to explain the given observations. The MOFALO system (implemented in Franz Lisp on a Sun Workstation) consists of a modified syntax for the POLYLITH MIL together with routines translating it into the object-oriented internal representation, some 'algebras' or behavior description languages for the example domains, and the core fault localization algorithm, which is based on [deKleer, Williams 87].


FKI-92-88: Klöck, E.: Utterance Generation Without Choice. - Published in: Künstliche Intelligenz. GWAI-88, 12. Jahrestagung. Hrsg.: Hoeppner, W., Berlin: Springer, 1988, S. 151-161 (Informatik-Fachberichte 181).

In this paper we discuss a parallel processing model for the generation of linguistic surface structures form a conceptual representation of the utterance content. We focus in particular on the verb selection task and its integration into a system for sentence production and introduce the notion of uttering pressure to control the moment of verbalization. The resulting model allows for different surface realizations of a single proposition without requiring an explicit choice among the alternatives. The system architecture presented consists of several independent spreading activation networks that communicate via a global blackboard. This setup combines the advantages of a classical modular system with the processing characteristics of the connectionist paradigm.


FKI-94-88: Zimmermann, K.: Der Netzeditor. Eine komfortable Umgebung zum Erstellen und Testen von konnektionistischen Netzen.

As connectionist networks become more and more important in artificial intelligence research we developed a system to support the design and testing of new networks. The system enables the user to construct networks from several unit and link types which are either provided by the system or defined by the user. He can do this interactively through a convenient menu-driven interface or by program, using a network description language. The user can simulate the behavior of the networks in a synchronous or asynchronous mode. He is provided with a graphical display of the network's state and can use it to change the activation of units during the simulation. The system is implemented in an object-oriented way using LOOPS and Interlisp-D on a Siemens-AI-Workstation. This paper describes the built-in high-level user interface, containing command windows, menus, editors and the network description language, and gives examples how to use the system efficiently. A special chapter explains the internals of the low-level parts which are useful for researchers who want to build their own applications.


FKI-95-88: Stolcke, A.: Generierung natürlichsprachlicher Sätze in unifikationsbasierten Grammatiken.

Die Arbeit gibt einen Überblick über vielschichtige Probleme der Sprachgenerierung. Dies geschieht im ersten Teil des Forschungsberichts in Abschnitt 2. Es folgt eine (notwendigerweise unvollständige) Übersicht über wichtige Arbeiten auf dem Gebiet der Sprachgenerierung, wobei sowohl die kognitionswissenschaftliche als auch die anwendungsorientierte Forschungsrichtung berücksichtigt wird. Es zeigt sich, daß sehr unterschiedliche kognitive Modelle und Formalismen in den verschiedenen Ansätzen Verwendung finden. Eines der kognitiven Paradigmen aus jüngerer Zeit, der sog. Konnektionismus, ist dabei vor allem bei der Modellierung psycholinguistischer Phänomene von Bedeutung. Aus der theoretischen Linguistik dagegen kommt ein Formalismus zur abstrakten Sprachbeschreibung, der in verschiedenen Varianten eine zentrale Rolle in vielen Generierungssystemen spielt: die unifikationsbasierten Grammatiken. Dieser Formalismus wird in Abschnitt 3 vorgestellt. Konnektionismus und unifikationsbasierte Grammatiken sind in vielerlei Hinsicht prototypisch für zwei grundverschiedene Verarbeitungsparadigmen innerhalb der KI und der Informatik, die beide sehr spezifische Vorzüge aufzuweisen haben. In Abschnitt 4 wird deshalb untersucht, inwieweit unifikationsbasierte Grammatiken mit dem konnektionistischen Paradigma vereinbar sind.


FKI-98-89: Brauer, W.; Freksa, C. and the AI/Cognition Group: Connectionist Approach to the Description of Spatial Knowledge.

This report collects five papers related to the ongoing project 'Connectionist Approaches to the Description of Spatial Knowledge'. The primary goal of this project is to study the applicability of connectionist aproaches to the task of describing spatial knowledge in a cognitively plausible way. The particular application is a system to give directions for going from a current location to a goal location in an urban setting. Given the current state of the art in connectionist systems, we are not seeking a closed 'connectionist solution' but rather exploring the usefullness for the various subtasks composing our system, like: representing spatial knowledge, finding adequate paths between a start and a goal location, and generating natural language descriptions of those paths. While the first paper gives an overview of the project, the others describe detailed work on particular aspects such as a path-finding algorithm, a knowledge theoretic frame work, the credit assignment problem in recurrent networks and ways to improve back-propagation.


FKI-124-90: Schmidhuber, J.H.: A Local Learning Algorithm for Dynamic Feedforward and Recurrent Networks. - Published in : Connection Science 1, 4/1990, pp. 403-412.

Most known learning algorithms for dynamic neural networks in non-stationary environments need global computations to perform credit assignment. These algorithms either are not local in time or not local in space. Those algorithms which are local in both time and space usually can not deal sensibly with 'hidden units'. In contrast, as far as we can judge by now, learning rules in biological systems with many 'hidden units' are local in both space and time. In this paper we propose a parallel on-line learning algorithm which performs local computations only, yet still is designed to deal with hidden units and with units whose past activations are 'hidden in time'. The approach is inspired by Holland's idea of the bucket brigade for classifier systems, which is transformed to run on a neural network with fixed topology. The result is a feedforward or recurrent 'neural' dissipative system which is consuming 'weight-substance' and permanently trying to distribute this substance onto its connections in an appropriate way. Simple experiments demonstrating the feasability of the algorithm are reported balancin.


FKI-125-90: Schmidhuber, J.H.: Networks Adjusting Networks.
A revised and extended version of 'Networks Adjusting Networks' in Kindermann, J. and Linden, A. (Eds.) Proceedings of 'Distributed Adaptive Neural Information Processing', St. Augustin, 24. - 25. 5. 1990, Oldenbourg-Verlag, 1990, pp. 197-208.

This paper describes extensions of previous 'adaptive critics' which have been onedimensional, acyclic, and suited only for feed-forward controllers. The extensions address the following issues:1. Feed-forward adaptive critics for fully recurrent probabilistic control nets. 2. Recurrent adaptive critics. 3. Vector-valued adaptive critics based on a system identification component. Furthermore an idea is described for approximating recurrent back propagation with a 3-network method which is local in time. In one experiment a linear adaptive critic adjusts a recurrent network such that it solves a non-linear task (a 'delayed XOR'-problem). In another experiment a four-dimensional adaptive critic quickly learns to solve a complicated pole balancing task.


FKI-126-90 : Schmidhuber, J.H.: Making the World Differentiable: On Using Self-Supervised Fully Recurrent Neural Networks for Dynamic Reinforcement Learning and Planning in Non-Stationary Environments.

First a brief introduction to reinforcement learning and to supervised learning with recurrent networks in non-stationary environments is given. The introduction also covers the basic principle of 'gradient descent through frozen model networks' as employed by Werbos, Jordan, Munro, Robinson and Fallside, and Nguyen and Widrow. This principle allows supervised learning techniques to be employed for reinforcement learning. Then a general algorithm for a reinforcement learning neural network with internal and external feedback in a non-stationary reactive environment is described. Internal feedback is given by connections that allow cyclic activation flow through the network. External feedback is given by output actions that may change the state of the environment thus influencing subsequent input activations. The network's main goal is to receive as much reinforcement (or as little 'pain') as possible. In theory, arbitrary time lags between actions and ulterior consequences are possible. The 'visible' environment may be 'non-Markovian'. Although the approach is based on 'supervised' learning algorithms for fully recurrent dynamic networks, no teacher is required. An adaptive model of the environmental dynamics is constructed which includes a model of future reinforcement to be received. This model is used for learning goal directed behavior. The algorithm may concurrently learn the model and learn to pursue the main goal. To attack certain problems with the parallel version of the algorithm, 'adaptive randomness' is introduced. The algorithm is applied to a reinforcement learning task in a non-Markovian environment. A connection to 'meta-learning' (learning how to learn) is noted. An extension of the algorithm is described which includes a vector-valued adaptive critic element (based on Sutton's TD-methods). The possibility of using the model for learning by planning (by 'mental simulation' of the environmental dynamics) is investigated. Finally it is described how the algorithm can be augmented by dynamic curiosity and boredom. This can be done by introducing (delayed) reinforcement for controller actions that increase the model network's knowledge about the world. This in turn requires the model network to model its own ignorance, thus showing a rudimentary form of introspective behavior.


FKI-127-90: Weiß, G.: Artificial Neural Learning.

Learning is a central area of research on artificial neural nets; major goal is the development of learning procedures that work efficiently even for complex real-world tasks. Starting from a general description of artificial neural nets, this report provides an introducing overview of the foundations and principles of artificial neural learning; thereby the main emphasis is on learning procedures. Basic characteristics of artificial neural learning are summarized. Relevant aspects of weight adaptation, of learning by the methods of gradient-following, and of the usual classification of learning procedures are reviewed. Several procedures for supervised learning, associative reinforcement learning and unsupervised learning are surveyed.


FKI-128-90: Schmidhuber, J.H.; Huber, R.: Learning to Generate Focus Trajectories for Attentive Vision.

One motivation of this paper is to provide an alternative for inefficient purely static 'neural' approaches to visual target detection. This is done by introducing a more efficient sequential approach. The latter is inspired by the observation that biological systems employ sequential eye-movements for pattern recognition. The other motivation is to demonstrate that there is at least one principle which can lead to the learning of dynamic selective spatial attention. A system consisting of an adaptive 'model network' interacting with a dynamic adaptive 'control network' is described. The system learn to generate focus trajectories such that the final position of a moving focus corresponds to a target to be detected in a visual scene. The difficulty is that no teacher provides the desired activations of 'eye-muscles' at various times. The only goal information is the desired final input corresponding to the target. Thus the task involves a complex temporal credit assignment problem, as well as an attention shifting problem. It is demonstrated experimentally that the system is able to learn correct sequences of focus movements involving translations and rotations. The system also learns to track moving targets. Some implications for attentive systems in general are discussed. For instance, one can build a 'mental focus' which operates on the set of internal representations of a neural system. It is suggested that self-referential systems which model the consequences of their own 'mental focus shifts' open the door for introspective learning in neural networks.


FKI-129-90: Schmidhuber, J.H.: Towards Compositional Learning with Dynamic Neural Networks.

None of the existing learning algorithms for neural networks with internal and/or external feedback addresses the problem of learning by composing subprograms, of learning 'to divide and conquer'. In this work it is argued that algorithms based on pure gradient descent or on temporal difference methods are not suitable for large scale dynamic control problems, and that there is a need for algorithms that perform 'compositional learning'. Some problems associated with compositional learning are identified, and a system is described which attacks at least one of them. The system learns to generate sub-goals that help to achieve its main goals. This is done with the help of 'time-bridging' adaptive models that predict the effects of the system's sub-programs. An experiment is reported which demonstrates the feasibility of the method.


FKI-132-90: Weiß, G.: Combining Neural and Evolutionary Learning: Aspects and Approaches.

This report focusses on the intersection of neural and evolutionary learning and shows basic aspects of and current approaches to the combination of these two learning paradigms. Advantages and difficulties of such a comblnation are described. Approaches from both the field of artificial intelligence and the neurosciences are surveyed. A number of related works as well as extensive references to further literature are presented.


FKI-134-90: Zimmermann, K.: Entwicklung einer bildorientierten Benutzungsoberfläche für wissensbasierte Systeme. - Published in:Graphik und KI. Hrsg.: Kansy, K. und Wißkirchen, P., Berlin: Springer, April 1990., S. 10-18 (Informatik-Fachberichte 239).

As graphical user interfaces become more and more important and sufficient hardware is easily available, we developed a picture oriented user interface for knowledge based systems. The paper discusses the architecture of the system and the advantages of using a picture oriented approach over the usual object oriented more pictogram like technique. It further discribes the different modes of user system interaction and how they may be implemented.


FKI-135-90: Hernández, D.: Relative Representation of Spatial Knowledge: The 2-D Case.
A revised and extended version of 'Relative Representation of Spatial Knowledge: The 2-D Case' in 'Cognitive and Linguistic Aspects of Geographic Space'.
Hrsg.: Mark, D.M. and Frank, A.U., Dordrecht: Kluwer, 1991.

There have been some straightforward efforts to extend Allen's interval-based temporal logic to spatial dimensions by using Cartesian tuples of relations (Güsgen 1989). We take a different approach based on a study of the kind of information that best relates two entities in 2-dimensional space qualitatively. The relevant spatial categories turn out to be 'projection' and 'orientation'. We define a small set of spatial relations and stress the importance of making their reference frames explicit. Furthermore, we introduce 'abstract maps', an analogical representation that inherently reflects the structure of the represented domain, and demonstrate their use in spatial reasoning. This scheme also facilitates 'coarse'' reasoning and the hierarchical organization of knowledge. These representational issues form the basis for an experimental system to develop 'cognitive maps'' from 2-D scanned layout plans of buildings.


FKI-136-90: Freksa, C.: Qualitative Spatial Reasoning. - Published in: Workshop Räumliche Alltagsumgebung des Menschenn. Hrsg.: Hoeppner, W. Universität Koblenz-Landau, Oktober 1990. To appear in: Cognitive and Linquistic Aspects of Geographic Space. Eds.: Mark, D.M. and Frank, A.U., Dordrecht: Kluwer.

Physical space has unique properties which form the basis of fundamental capabilities of cognitive systems. This paper explores some cognitive aspects of perception and knowldge representation and explains why spatial knowledge is of particular interest for cognitive science. It is suggested that 'spatial inference engines' provide the basis for rather general cognitive capabilities inside and outside of the spatial domain. The role of abstraction in spatioal reasoning and the advantages of qualitative spatial knowledge over quantitative knowledge are discussed. The usefulness of spatioal representations with a low degree of abstraction is shown. An example from vision ( the aquarium domain) is used to illustrate in which ways knowledge about space may be uncertain or incomplete. Parallels are drawn between the spatial and the temporal domains. A concrete approach for the representation of qualitative spatial knowldge on the bassis of 'conceptual neighborhood' is suggested and some potential application areas are mentioned.


FKI-137-90: Dorigo, M.; Schätz, B.: Mapping a Generator for Neural Network Simulators to a Transputer System.

Transputers are a kind of coarse grain parallel hardware produced by Inmos. Their main characteristic is the high degree of modularity and scalability they allow. Using trans- puters it is possible to build application programs composed of communicating tasks resident on different processors. These tasks communicate through the use of hard- ware links (channels). Every processor may have 1-4 MBytes of RAM. We have used a transputer system to implement a generator of neural network simulators. Starting from the system described in [l] we have tried to map the system to the transputer parallel ar- chitecture. In the report we shortly overview the computational model given by trans- puters and some of the characteristics of the Parallel-C language that we have been using to implement the system.Then a general overview of how the sequential generator of simulators works is given. We also explain the architectural decisions we have made to change the sequential system to a parallel one. At the end a description of the algorithms used to control a routing system is given and some results about the speed-up obtained are reported.


FKI-138-90: Bräunling, P.; Freksa, C.; Zimmermann, K.: The SpaceGarden Bibliography. Published in: Repräsentation und Verarbeitung räumlichen Wissens. Hrsg.: Freksa, C.; Habel, C., Berlin: Springer, 1990, S. 267-353 (Informatik-Fachberichte 245).


FKI-140-90: Thomas, P.: Beyond Hebb Synapses: Biological Building Blocks for Unsupervised Learning in Artificial Neural Networks.

This paper briefly reviews the neurobiology of synaptic plasticity as it is related to the formulation of learning rules for unsuperviscd learning in artificial neural networks Presynaptic, postsynaptic and heterocellular mechanisms are discussed and their relevance to neural modelling is assessed These include a variety of phenomena of potentiation as well as depression with time courses of action ranging from milliseconds to weeks The original notion put forward by Donald Hebb stating that synaptic plasticity depends on correlated pre- and postsynaptic firing is reportedly inadequate. Although postsynaptic depolarization is necessary for associative changes in synaptic strength to take place (which conforms to the spirit of the hebbian law) the association is understood as being formed between pathways converging on the same postsynaptic neuron The latter only serves as a supporLing device carrying signals between activated dendritic regions and maintaining long-term changes through molecular mechanisms It is further proposed to restrict the interactions of synaptic inputs to distinct compartments The hebbian idea that the state of the postsynaptic neuron as a whole governs the sign and magnitude of changes at individual synapses is dropped in favor of local mechanisms which guide the depolarization-dependent associative learning process within dendritic compartments. Finally, a framework for the modelling of associative and non-associative mechanisms of synaptic plasticity at an intermediate level of abstraction, the Patchy Model Neuron, is sketched.


FKI-146-91: Freksa, C.: Conceptual Neighborhood and its Role in Temporal and Spatial Reasoning. - Published in: Proc. of the IMACS Workshop on Desicion Support Systems and Qualitative Reasoning. Eds.: Singh, M.; Travé-Massuyès, L., Elsevier Science Publishers, Amsterdam, 1991.

An extension of Allen's approach to interval-based temporal reasoning is presented. The new method allows for temporal and spatial reasoning on the basis of incomplete or imprecise knowledge of the kind that is available from inference and perception processes. The central idea of the representation method is the structuring of knowledge according to the conceptual neighborhood of temporal and spatial relations. This representation allows for integration of coarse and fine knowledge. Logical reasoning on the basis of such knowledge therefore takes place within a unified scheme. The method presented not only is more efficient than Allen's method, it also is more `cognitively adequate' in comparison with previous approaches.


FKI-147-91: Schmidhuber, J.H.: Learning to Control Fast-Weight Memories: An Alternative to Dynamic Recurrent Networks.

Previous algorithms for supervised sequence learning are based on dynamic recurrent networks. This paper describes alternative gradient-based systems consisting of two feed-forward nets which learn to deal with temporal sequences by using fast weights: The first net learns to produce context dependent weight changes for the second net whose weights may vary very quickly. The method offers a potential for STM storage efficiency: A simple weight (instead of a full-fledged unit) may be sufficient for storing temporal information. Various learning methods are derived. Two experiments with unknown time delays illustrate the approach. One experiment shows how the system can be used for adaptive temporary variable binding.


FKI-148-91: Schmidhuber, J.H.: A Neural Sequence Chunker.

This paper addresses the problem of learning to 'divide and conquer' by meaningful hierarchical adaptive decomposition of temporal sequences. This problem is relevant for time-series analysis as well as for goal-directed learning, particularily if event sequences tend to have hierarchical temporal structure. The first neural systems for recursively chunking sequences are described. These systems are based on a principle called the 'principle of history compression'. This principle essentially says: As long as a predictor is able to predict future environmental inputs from previous ones, no additional knowledge can be obtained by observing these inputs in reality. Only unexpected inputs deserve attention. A focus is on a class of 2-network systems which try to collapse a self-organizing (possibly multi-level) hierarchy of temporal predictors into a single recurrent network. Only those input events that were not expected by the first recurrent net are transferred to the second recurrent net. Therefore the second net receives a reduced discription of the input history. It tries to develop internal representations for 'higher-level' temporal structure. These internal representations in turn serve to create additional training signals for the first net, thus helping the first net to create longer and longer 'chunks' for the second net. Experiments show that chunking systems can be superior to the conventional training algorithms for recurrent nets.


FKI-149-91: Schmidhuber, J.H.: Adaptive Confidence and Adaptive Curiosity.

Much of the recent research on adaptive neuro-control and reinforcement learning focusses on systems with adaptive 'world models'. Previous approaches, however, do not address the problem of modelling the reliability of the world model's predictions in uncertain environments. Furthermore, with previous approaches usually some ad-hoc method (like random search) is used to train the world model to predict future environmental inputs from previous inputs and control outputs of the system. This paper introduces ways for modelling the reliability of the outputs of adaptive predictors, and it describes more sophisticated and sometimes more efficient methods for their adaptive construction by on-line state space exploration: For instance, a 4-network reinforcement learning system is described which tries to maximize the expectation of the temporal derivative of the adaptive assumed reliability of future predictions. The system is 'curious' in the sense that it actively tries to provoke situations for which it learned to expect to learn something about the environment. An experiment with an artificial non-deterministic environment demonstrates that the method can be faster than the conventional model-building strategy.


FKI-151-91: Schmidhuber, J.H.: An O(n3) Learning Algorithm for Fully Recurrent Networks.

The fixed-size storage learning algorithm for fully recurrent continually running networks [2][7] requires O(n4) computations per time step, where n is the number of non-input units. We describe a method which computes exactly the same gradient and requires fixed-size storage of the same order as the previous algorithm. But, the average time complexity per time step is O(n3).


FKI-152-91: Laußermair, T.; Weiß, G.: Artificial Life - Eine Einführung.

This paper provides a general introduction to the young field of Artificial Life (AL for short). The first part of this paper briefly outlines the biological background of AL, its synthetical character, and its relation to Artificial Intelligence. The central question underlying the AL research is that about the essential processes and structures of 'life' in contrast to 'non-life'. The central methodology of AL research is to synthesize these characteristics from structural and functional primitives. And the central working hypothesis underlying the AL research is that the interactions of the primitive components and not the components themselves make up the characteristics of living systems. Thereby AL is not restricted to the study of natural living systems ('life as it is') but also investigates man made, synthesized living systems ('Life as it could be'). The second part of this paper focuses on the central concepts ('keynotes') and issues of AL. Locality and parallelism in the behavior of the primitive components of a system lead to a new process model that fundamentally differs from the traditional sequential von Neumann model. A consequence of this locality and parallelism at the component level, as well as of the nonlinearity in the behavior of the components, is the emergence of new properties at the system level, that means, the arising of new system properties that are not based on the components tllemselves but on their interactions. As a special type of emergence, AL studies the self-organization of structural and behavioral patterns in multi-component systems. (These concepts-- locality, parallelism, emergence, self organization--are primarily used for exploring the transition from non-living to living systems.) Closely related to the concept of self-organization is tllat of the adaptability of a system, that means, the ability of a system to functionally and/or structurally adapt to environmental changes. Finally, based on the biologically motiviated genotype/phenotype distinction, in the field of AL the concept of the evolution of systems plays a very important role. (These concepts--adaptability, genotypes versus phenotypes, evolution--are primarily used for exploring the transition from simple to complex systems.) In this paper tllree central issues of AL are identified and outlined: the development of structure and form, the development of behavior and function, and collective bellavior and communication. The paper concludes with a citation of C. G. Langton on the essence of AL.


FKI-153-91: Freksa, C.: Temporal Reasoning Based on Semi-Intervals.
A revised and extended version of TR-90-016, Intemational Computer Science Institut, Berkeley 1990. To appear in: Artificial lntelligence, Winter 1991/92.

A generalization of Allen's interval-based approach to temporal reasoning is presented. The notion of 'conceptual neighborhood' of qualitative relations between events is central to the presented approach. Relations between semi-intervals rather than intervals are used as the basic units of knowledge. Semi-intervals correspond to temporal beginnings or endings of events. We demonstrate the advantages of reasoning on the basis of semiintervals: l) semi-intervals are rather natural entities both from a cognitive and from a computational point of view; 2) coarse knowledge can be processed directly; computational effort is saved; 3) incomplete knowledge about events can be fully exploited; 4) incomplete inferences made on the basis of complete knowledge can be used directly for further inference steps; 5) there is no trade-off in computational strength for the added f,exibility and efficiency; 6) for a natural subset of Allen's algebra, global consistency can be guaranteed in polynomial time; 7) knowledge about relations between events can be represented much more compactly.


FKI-154-91: Zimmermann, K. : SEqO - Ein System zur Erforschung qualitativer Objektrepräsentationen.

In the field of Artificial Intelligence different mathematical formalisms have been adopted sofar to allow for the modelling of natural inferencing mechanisms. Standard logic has been extended into the fuzzy logic, relation theory is used in constraint systems. All these systems share the property of being directly based on the set of real numbers. For example fuzzy logic allows the representation of fuzzy values but only by describing them through a function f:|R->|R that returns for every value x the fuzzy grade f(x). Thus, paradoxically, you end up representing fuzzy knowledge using real numbers with potentially infinite precision and resolution. In most symbolic constraint propagation systems the underlying variables are taken to be quantitative, i.e. their value set is the set of real numbers, even if one does not care for the exact value. The usage of qualitative values for the variables is typically not supported. Some widely used systems even translate qualitative constraints into numeric ones. This report describes a method to represent such values qualitatively. It is based on cognitive requirements and constraints that lead into a system using positive absolute values, together with sign attributes, and an extended half order representation. This system can be used in a stand alone fashion or combined with existing ones.



4.2 Publications

Beckstein, C.; Görz, G.; Hernández, D.; Tielemann, M.: An Integration of Object-Oriented Knowledge Representation and Rule-Oriented Programming as a Basis for Design and Diagnosis of Technical Systems. - In: Annals of Operations Research 16, 1988,
pp. 13 - 32.

Bräunling, P.: Umfrage zum Thema Valenzwörterbücher. - In: Lexicographica, Band 5.
Eds: Kucera , A. et al., Tübingen: Niemeyer, 1989.

Bräunling, P.: Zur kontextabhängigen Objektbenennung - Ist ein großes gelbes Haus immer groß und gelb? - In: Workshop Räumliche Alltagsumgebungen des Menschen. Hrsg.: Hoeppner, W., Universität Koblenz-Landau, 8.-10. Okt. 1990, S. 1-8.

Bräunling, P.; Freksa, C.; Zimmermann, K. (eds.): The SpaceGarden Bibliography. - In: Repräsentation und Verarbeitung räumlichen Wissens. Hrsg.: Freksa, C.; Habel, C., Berlin: Springer, 1990, S. 267-353 (Informatik-Fachberichte 245).

Brauer, W. and the AI/Cognition Group: Approaches to the Representation of Knowledge.
In: Advanced Information Processing. Proceedings of a Joint Symposium Information Processing and Software Systems Design Automation. Eds.: Schwärtzel, H. and Mizin, I., Academy of Science of the USSR and Siemens AG, Berlin: Springer, 6/1990,
pp. 29-38.

Brauer, W.; Freksa, C. (Hrsg.): Wissensbasierte Systeme. Proc. 3. Internationaler GI-Kongreß, München, 16.-17. Okt. 1989, Berlin: Springer, 1989 (Informatik-Fachberichte 227).

Dirlich, G.; Benda, H.v.; Freksa, C.; Furbach, U., Müller, A.; Wimmer, K.: Computerunterstützte Planung von Ferienreisen. Ein fiktives Beispiel. - In: Kognitive Aspekte der Mensch-Computer-lnteraktion. Hrsg.: Dirlich, G. u.a., Berlin: Springer, 1986, S. 22-36 (Informatik-Fachberichte 120).

Dirlich, G.; Freksa, C.; Schwatlo U.; Wimmer, K. (Hrsg.): Kognitive Aspekte der Mensch-Computer-lnteraktion., Berlin: Springer, 1986 (Informatik-Fachberichte 120).

Erol, N.; Freksa, C.: An Approach to Structuring and Formalizing Knowledge for a Design Support System. - In: Artificial Intelligence II - Methodology, Systems, Applications. Eds.: Jorrand, Ph.; Sgurev, V., Amsterdam: North-Holland 1987, pp. 271 - 279.

Freksa, C.: Knowledge Representation for Interactive Aircraft Design. - In: Expert Systems and Knowledge Engineering. Ed.: Bernold, T., Amsterdam: North-Holland, 1986,
pp. 221-230.

Freksa, C. (Hrsg.): Kognition - Wissensstrukturen beim Aufgabenlösen.- In: GWAI-87, 11th German Workshop on Artificial Intelligence. Hrsg.: Morik, K., Berlin: Springer, 1987, S. 277-305 (Informatik-Fachberichte 152).

Freksa, C.: Mit welchen Themen soll sich die KI auseinandersetzen? - In: GWAI-88, Künstliche Intelligenz. Hrsg.: Hoeppner, W., Berlin: Springer, 1988, S. 327-333.

Freksa, C. : Intrinsische vs. extrinsische Repräsentation zum Aufgabenlösen oder die Verwandlung von Wasser in Wein. - In: Wissensarten und ihre Darstellung.
Hrsg.: Heyer, G. u.a., Berlin: Springer, 1988 (Informatik Fachberichte).

Freksa, C.: Cognitive Science - eine Standortbestimmung.- In: Wissensarten und ihre Darstellung. Hrsg.: Heyer, G. u.a., Berlin: Springer, 1988, S. 1-12 (Informatik-Fachberichte). Also published in: Universitas 525, März 1990, S. 232-240.

Freksa, C.: Wissensdarstellung und Kognitionsforschung. - In: Informationstechnik it 31 (1989) 2, S. 134-140. Auch erschienen in: Wissensrepräsentation. Hrsg: Struß, P., Oldenbourg-Verlag, München 1991, S. 61-68.

Freksa, C.: Qualitative Spatial Reasoning. - In: Workshop Räumliche Alltagsumgebung des Menschenn. Hrsg.: Hoeppner, W., Universität Koblenz-Landau, Oktober 1990. To appear in: Cognitive and Linguistic Aspects of Geographic Space. Eds.: Mark, D.M. and Frank, A.U., Dordrecht: Kluwer, 1991.

Freksa, C.: Conceptual Neighborhood and its Role in Temporal and Spatial Reasoning.
In: Proc. of the IMACS Workshop on Desicion Support Systems and Qualitative Reasoning. Eds.: Singh, M.; Travé-Massuyès, L., Elsevier Science Publishers, Amsterdam, 1991.

Freksa, C.; Bräunling, P.; Zimmermann, K.: The SpaceGarden Interface for an Interdisciplinary Bibliography System. Extended Abstract of the talk given at the workshop 'ISI - Intelligente Schnittstellen zu Informationssystemen / Intelligent Access to Information Systems', Darmstadt, 1990.

Freksa, C.; Habel, C. (Hrsg.).: Repräsentation und Verarbeitung räumlichen Wissens, Berlin: Springer, 1990 (Informatik-Fachberichte 245).

Freksa, C.; Habel, C.: Warum interessiert sich die Kognitionsforschung für die Verarbeitung räumlichen Wissens?. - In: Repräsentation und Verarbeitung räumlichen Wissens, Berlin: Springer, 1990, S. 1-15.

Freksa, C.; Hernández, D.; Marcinowski, M.; Schmidhuber, J.H.: Sprachartige Beschreibung von Objekten und räumlichen Gegebenheiten in einem konnektionistischen System. Skizze eines Forschungsprojektes. - In: Arbeitspapiere der GMD 329, Proc. 1. Konnektionismus-Workshop 1988, St. Augustin, 3.-4. Feb.. Hrsg.: Lischka, C.; Kindermann, J., St. Augustin: GMD, 1988.

Freksa, C.; Kobsa, A.: Erhebung zur Kognitionsforschung im deutschsprachigen Raum. 1985/86.

Furbach, U.; Freksa, C.; Dirlich, G.: Wissensrepräsentation in künstlichen symbolver-arbeitenden Systemen. - In: Wissenspsychologie. Hrsg.: Mandl, H. u.a., München: Psychologie Verlags Union, 1988, S. 505-528.

Hartl, A.: Kognitive Karten und kognitives Kartieren. - In: Repräsentation und Verarbeitung räumlichen Wissens, Berlin: Springer, 1990 (Informatik-Fachberichte 245).

Hernández, D.: Bericht über den Workshop Konnektionismus an der GMD. - In: KI 3/88, Oldenbourg-Verlag, Juni 1988.

Hernández, D.: Zur Implementierbarkeit analogischer Repräsentationen. - In: GWAI-89, Proceedings. Hrsg.: Metzing, D., Eringerfeld, 18-22 Sept.1989, Berlin: Springer, 1989, S. 479-481 (Informatik-Fachberichte 216).

Hernández, D.: A Principled Approach to Knowledge Representation in Connectionist Systems. - In: Connectionist Approaches to the Description of Spatial Knowledge and Related Papers. Eds.: Brauer, W. et al., Technical Report FKI-98-89, Institut für Informatik, Technische Universität München, 1989.

Hernández, D.: Relative Representation of Spatial Knowledge: The 2-D Case. - In: Cognitive and Linguistic Aspects of Geographic Space.Eds.: Mark, D.M. and Frank, A.U., Dordrecht: Kluwer, 1991.

Hernández, D.; Kanal, L.; Purtilo, J.M.: A Debugging Assistant for Distributed Systems. - In: Proc. Northwest Software Quality Conference, Portland, Oregon, 19.-20. Sept. 1988, pp. 301-316.

Klöck, E.: Spreading Activation Networks for Natural Language Generation: the Verb Selection Problem. - ln: Arbeitspapiere der GMD 329, Proc. 1. Konnektionismus-Workshop 1988, St. Augustin, 3.-4. Feb.. Hrsg.: Lischka, C.; Kindermann, J., St. Augustin: GMD, 1988, S.185-195.

Klöck, E.: Utterance Generation Without Choice. - In: Künstliche Intelligenz. GWAI-88, 12. Jahrestagung. Hrsg.: Hoeppner, W., Berlin: Springer, 1988, S. 151-161 (Informatik-Fachberichte 181).

Klöck, E.: Erst denken, dann Sprechen? - Die kognitive Basis der Sprachproduktion. -In: Workshop Räumliche Alltagsumgebung des Menschen. Hrsg.: Hoeppner, W., Universität Koblenz-Landau, 1990, S. 92-101.

Praßler, E.: Electrical Networks and a Connectionist Approach to Pathfinding. - In: Proc. of the International Conference Connectionism in Perspective, Amsterdam: Elsevier, 1989.

Schätz, B.: Portierung eines neuronalen Netzwerksimulators auf ein Transputersystem. - In : Abstractband des 2. bundesweiten Transputer-Anwender-Treffens TAT'90, RWTH Aachen,17.-18. Sept.1990, S. 48-49.

Schätz, B.: Ein konnektionistischer Ansatz zur verteilten Darstellung von Objektpositionen. -In: Workshop Räumliche Alltagsumgebungen des Menschen. Hrsg.: Hoeppner, W., Universität Koblenz-Landau, 8.-10. Oktober. 1990, S. 147-158.

Schmidhuber, J. H.: The Neural Bucket Brigade. - In: Connectionism in Perspective. Eds.: Pfeifer, R. et al., Zürich, Switzerland, 11.-13.0kt. 1988, Amsterdam: Elsevier, North-Holland, 1989, pp. 439-446.

Schmidhuber, J. H.: Accelerated Learning in Backpropagation Nets. - In: Connectionism in Perspective. Eds.: Pfeifer, R. et al., Zürich, Switzerland, 11.-13. Okt. 1988, Amsterdam: Elsevier,North-Holland, 1989, pp. 429-438.

Schmidhuber, J. H.: A Local Learning Algorithm for Dynamic Feedforward and Recurrent Networks. - In: Connection Science 1, 4/1990, pp. 403-412.

Schmidhuber, J. H.: Networks Adjusting Networks. - In: Proceedings of 'Distributed Adaptive Neural Information Processing', St. Augustin, 24. - 25. 5. 1990. Eds.: Kindermann, J.; Linden, A., Oldenbourg-Verlag, 1990, pp. 197-208.

Schmidhuber, J. H.: Recurrent Networks Adjusted by Adaptive Critics. - In : Proc. IEEE/INNS International Joint Conference on Neural Networks, Washington, D. C., 1990, pp. 719-722.

Schmidhuber, J.H.: Temporal-Difference-Driven Learning in Recurrent Networks. - In: Parallel Processing in Neural Systems and Computers. Eds.: Eckmiller, R.; Hartmann, G. and Hauske, G., North-Holland, 1990, pp. 209-212.

Schmidhuber, J.H.: Response to G. Lukes review of 'Recurrent Networks Adjusted by Adaptive Critics'. Neural Network Review, in press (1990).

Schmidhuber, J.H.: Reinforcement-Lernen und adaptive Steuerung.
Nachrichten Neuronale Netze 2/1990, S.1-3.

Schmidhuber, J.H.: An On-Line Algorithm for Dynamic Reinforcement Learning and Planning in Reactive Environments. - In: Proc. IEEE/INNS International Joint Conference on Neural Networks, San Diego, 1990, pp. 253-258.

Schmidhuber, J.H.: Reinforcement Learning with Interacting Continually Running Fully Recurrent Networks. - In: Proc. INNC International Neural Network Conference, Paris, 1990, pp. 817-820.

Schmidhuber, J.H.: Learning Algorithms for Networks with Internal and External Feedback. In: Proc. of the 1990 Connectionist Models Summer School. Eds.: Touretzky, D.S. et al., San Mateo, CA: Morgan Kaufmann, 1990, pp. 52-61.

Schmidhuber, J.H.: A Possibility for Implementing Curiosity and Boredom in Model-Building Neural Controllers. - In: Proc. of the International Conference on Simulation of Adaptive Behavior: From Animals to Animats. Eds.:Meyer, J.A.; Wilson, S.W., MIT Press/Bradford Books, 1991, pp. 222-227.

Waschulzik, T.; Geiger, H.; Arnoldi, M.; Böller, D.; Nischwitz, A.; Brauer, W.: New Concepts for Information Processing in Connectionistis Systems. - In: Parallel Processing in Neural Systems and Computers. Eds.: Eckmiller, R.; Hartmann, G. and Hauske, G., Amsterdam, 1990, pp. 323-328.

Wu, Dekai: Concretion Inferences in Natural Language Understanding. -
In: GWAI-87, 11th German Workshop on Artificial Intelligence. Ed.: Morik, K., Berlin: Springer, 1987.

Zimmermann, K.: Entwicklung einer bildorientierten Benutzungsoberfläche für wissensbasierte Systeme. - In:Graphik und KI. Hrsg.: Kansy, K. and Wißkirchen, P., Berlin: Springer, April 1990., S. 10-18 (Informatik-Fachberichte 239).