THE EMPEROR’S NEW THEORY Editor-in-Chief Consciousness is regaining interest in the fields of natural sciences. For a long time, thinking (or even talking) about consciousness was regarded as unworthy of natural scientists. It was better to leave the subject of consciousness to philosophers and psychologists who are used to think about things that are hard to grasp. But now, all this has changed due to pioneers such as Francis Crick and Christof Koch who took the bold step to put the study of consciousness on the agenda of neurocomputational research. Nowadays, theorizing about consciousness is very common in the natural sciences and has even entered the bastion of physics. Sir Roger Penrose, renowned amongst others for his work in collaboration with Stephan Hawking on the origins and evolution of the universe, has given consciousness a central role in his theorizing about physics and the observing mind. In a recent lecture in Maastricht he talked about time in physics and psychology. His exposé on the fundamentals of physics revealed his great knowledge and erudition. With a nonchalance characteristic of a great scientist he presented a very clear story hardly affected by his chaotic handling of hand-drawn overhead sheets. His lecture focused on the gap between the two most powerful theories in physics: quantum theory and classical theory (including relativity theory). Although both theories have great explanatory power in their own domains (i.e., the very small versus the very large, respectively) present-day physics has not succeeded in merging them into a single theory. Penrose noted that in both theories time is symmetrical whereas in our experience time is perceived as asymmetrical. The theory linking quantum and classical theory should therefore account for the “arrow of time” as perceived by conscious agents. After discussing his hypothesis on how the asymmetry of time arises, Penrose outlined his hypothesis on the origin of consciousness. According to this hypothesis, consciousness resides at the level of synapses where quantum effects are present and are affecting the mental processes at the larger scale of neural networks. The nonlocality of quantum mechanics, now an active subject of study in the context of quantum computing (see also Edwin de Jong’s report on the CASSYS’99 conference on page 136 of this newsletter) could arise from these synaptic quantum effects. In Penrose’s view, consciousness emerges from the quantum level and affects the neural processes underlying mental processes. The heroic efforts by Penrose and others have great appeal as they address one of the oldest riddles of science. Consciousness is central to ourselves, our culture, and our language; it deserves to be taken seriously as a subject of study. However, the route followed by Penrose seems to be a futile one. Explaining consciousness in terms of the gap between quantum and classical theory is in some sense like reducing visual perception in terms of the gap between fotons and nerve impulses. The point is that consciousness is a not a physical concept. Conscious experience probably arose as an evolutionary adaptation to the Paleolithic, some 200,000 years ago. Since then, it probably underwent many changes due to changing cultural (and environmental) demands. In this view the basis of conscious experience should primarily be sought in the complexity of the environment (i.e., other conscious agents), rather than in the complexity of the brain alone. How do these reflections on consciousness relate to artificial intelligence? Currently, studies of (artificial) consciousness are very rare if present at all, but I expect them to (re-)emerge soon. The emphasis on autonomous agents and robots is bound to lead to a reconsideration of consciousness in artificial-intelligent systems. Although I consider it much too early to put forward any sensible question related to consciousness, I nevertheless encourage our members to start thinking along these lines. After all, that is what science is all about.
Cover illustration: the BNAIC’99 logo (designed by Hans Hoornstra) The agenda and financial overview for the BNVKI Board meeting (to be held at the BNAIC on Thursday November 4 from 12.35 to 13.45 hours) can be found via the BNVKI website http://www.cs.unimaas.nl/~bnvki/
BNVKI newsletter
124
October 1999
TABLE OF CONTENTS The Emperor’s New Theory (Editor in Chief) .....................................................................................................124 Table of Contents .................................................................................................................................................125 BNVKI-Board News (Joost Kok) .......................................................................................................................126 BNAIC’99 Update (Floris Wiesman and Eric Postma) .......................................................................................126 BNAIC’99 Programme ........................................................................................................................................127 Logic and Gambling (Jaap van den Herik) ..........................................................................................................130 Book Review: Function, Selection, and Innateness (Bart de Boer) .....................................................................130 The Sixteenth International Joint Conference on Artificial Intelligence (Edwin de Jong).................................132 Casys ’99: Computing Anticipatory Systems (Edwin de Jong) ..........................................................................136 Nouvelle AI (Rens Kortmann) .............................................................................................................................138 Gecco – 99 Student Workshop (William Langdon).............................................................................................143 Verslag van de Landelijke KION-dag (Ida Sprinkhuizen – Kuyper)..................................................................144 Section Knowledge Systems in Law and Computer Science (Radboud Winkels)...............................................145 Boekbesprekingen De Bouw van Juridische Kennissystemen (Bart Verheij)............................................................................145 Hercules of Karneades Hard Cases in Recht en Rechtsinformatica (Kees de Vey Mestdagh)....................149 Section Computational Linguistics (Antal van den Bosch) .................................................................................151 BNVKI Tutorials on AI and Language Processing......................................................................................151 Section SIKS (Richard Starmans)........................................................................................................................152 Call for Papers......................................................................................................................................................154 LOFT 4 (Wiebe van der Hoek)...................................................................................................................154 Call for Participation............................................................................................................................................155 The Jurix ’99 Conference............................................................................................................................155 Conferences, Symposia, Workshops....................................................................................................................156 Email-adresses, Board Members/ Editors BNVKI newsletter/ How to become a member?/ Submission ...........157 Back Issues/ Change of Address .........................................................................................................................157
The BNVKI is sponsored by AHOLD and by BOLESIAN The photographs on pages 143 and 144 are provided by William Langdon BNVKI newsletter
125
October 1999
In 1999, the publication of the BNVKI newsletter is also supported by the Division of Computer Science Research in the Netherlands (previously called SION, now ACI)
BNVKI newsletter
126
October 1999
hungry people in third-world countries. Fortunately, we do not have to await these future developments to offer help. If you want to do something nice today, please click on the button on http://www.thehungersite.com/ and donate some rice to a hungry person.
BNVKI-BOARD NEWS Joost Kok
Chairman BNVKI One of the nice things of AI research is that it generates many new technological ideas. Some of these ideas are surprising. The dial-a-coke concept from Finland is a good example. The concept is explained in the following press release by NOKIA.
After you have done this, you can dial 043 – 3883477, the BNAIC registration desk and register for the conference in case you have not done so already. As any reader can imagine I look forward to the demonstrations of our BNAIC in Vaeshartelt, Maastricht. See you at the Conference.
Singapore (June 5, 1998) – Nokia today announced the “Dial A Coke” concept – a new application for the Nokia Card Phone designed to extend greater convenience to the consumer. The innovative application displayed at CommunicAsia ’98 integrates the wireless data capabilities of a Nokia Card Phone with a Coca-Cola vending machine, and allows consumers to purchase a drink simply by using their mobile phone. To purchase a beverage through “Dial A Coke”, customers simply use their mobile phone to dial a phone number indicated on the vending machine. The drink pops out automatically and the purchase is confirmed through a short message to the customer’s mobile phone bill. The “Dial A Coke” concept works with all mobile phones.
BNAIC UPDATE Floris Wiesman and Eric Postma www.cs.unimaas.nl/~bnvki/bnaic99 The BNAIC’99 is rapidly approaching and we are busy preparing the conference. In the previous newsletter we reported on our main sponsor Lucent Technologies and that there will be three invited lectures by Tom Mitchell, Edward Rietman, and Jonathan Schaeffer. Now we are happy to announce a fourth, albeit short, invited lecture. Professor Frans Groen will present an introduction to the Robot Soccer demonstration entitled Robot Soccer: Game or Science? The RoboCup soccer demonstration is sponsored by Tryllian, a young start-up company specializing in mobile software agents. The demonstration takes place in one of the smaller conference rooms. To allow all participants to have a close look at the robots and their actions, the demonstration will be repeated several times. The demonstration and the happy hour (sponsored by Bolesian) take place at the same time so that participants can visit the demonstration in small groups. Below, we give a global overview of the BNAIC programme. Full details can be found on the following pages. We look forward to seeing you in Maastricht in the first week of November!
It does not take too much imagination to elaborate on this concept. Whatever the end result of this elaboration may be, it is clear that it involves a convergence of mobile-phone, television, credit-card and computer technologies. Intelligent technology is certainly needed for integrating these different technologies. So, interesting challenges lie ahead for Artificial Intelligence.
A GLOBAL OVERVIEW
NOKIA’s innovation could serve a much better cause than satisfying the thirst of rich people. Imagine that we install “Dial a cup of Rice” machines in poor countries which can be called from any country in the world. Then, we can offer direct help to BNVKI newsletter
TUESDAY NOVEMBER 2 17.00 – 18.00 hours Early bird reception at the Maastricht town hall WEDNESDAY NOVEMBER 3 09.45 – 10.20 hours opening 10.20 – 11.20 hours Tom Mitchell
127
October 1999
11.20 – 11.35 hours 11.35 – 12.25 hours 12.35 – 13.30 hours 13.30 – 14.45 hours 14.45 – 15.00 hours 15.00 – 16.15 hours 16.15 – 16.30 hours 16.30 – 17.30 hours 17.30 – 18.00 hours 18.00 – 20.00 hours 20.00 – 23.00 hours
coffee break parallel sessions lunch parallel sessions coffee break parallel sessions coffee break Edward Rietman (Lucent) Frans Groen RoboCup demo (Tryllian) Happy Hour (Bolesian) Conference Dinner
THURSDAY NOVEMBER 4 09.00 – 09.50 hours 09.50 – 10.05 hours 11.05 – 11.20 hours 11.20 – 12.35 hours 12.35 – 13.45 hours 13.45 – 15.00 hours 15.00 – 15.15 hours 15.15 – 16.30 hours 16.30 – 17.00 hours
parallel sessions Jonathan Schaeffer coffee break parallel sessions lunch and BNVKI Board meeting parallel sessions coffee break parallel sessions awards and closing
BNVKI newsletter
128
October 1999
13.30 – 14.45 Three Parallel Sessions (13.30 – 13.55; 13.55 – 14.20; 14.20 – 14.45)
PROGRAMME BNAIC’99 Tuesday November 2 17.00 – 18.00 Early-bird reception (at the Maastricht town hall)
BELIEF NETWORKS N. Peek, V. Coupé‚ and J. Ottenkamp Focused quantification of a belief network using sensitivity analysis
Wednesday November 3 09.00 – 09.45 Reception and coffee 09.45 – 10.20 Opening 10.20 – 11.20 Invited lecture by Tom Mitchell: Extracting Information from the World Wide Web 11.20 – 11.35 Coffee break 11.35 – 12.25 Three Parallel Sessions (11.35 – 12.00; 12.00 – 12.25)
S. Renooij and L.C. van der Gaag Exploiting Non-monotonic Influences in Qualitative Belief Networks J. Donkers, R. Ferreira, J. Uiterwijk, and H.J. van den Herik VAS: Quantifying a Qualitative Network SEARCH
LOGIC AND REASONING 1
E. Marchiori and A. Steenbeek A Genetic Local Search Algorithm for Random Binary Constraint Satisfaction Problems
Shan-Hwei Nienhuys-Cheng The Complexities of a Refinement Operator for Prenex Conjunctive Normal Forms
D.M. Breuker, J.W.H.M. Uiterwijk, and H.J. van den Herik Investigating pn2 Search
A.Bos, N. Roos and C. Witteveen Computing with Computational Histories EVOLUTIONARY COMPUTATION 1
P. van Bael, D. Devogelaere, and M. Rijckaert HESSA solves the Job Shop Scheduling Problem
D.D.B. van Bragt, C.H.M. van Kemenade and J.A. La Poutré The Influence of Evolutionary Selection Schemes on the Iterated Prisoner's Dilemma
AGENT TECHNOLOGY 1 B.G.W. Craenen, A.E. Eiben and E. Marchiori Solving Constraint Satisfaction Problems with Heuristic-based Evolutionary Algorithms
K.V. Hindriks, F.S. de Boer, W. van der Hoek, and J.-J. Meyer A Formal Semantics of the Core of AGENT-0
MACHINE LEARNING AND NEURAL NETWORKS 1
K. Jonkheer Intelligent agents, markets and competition - consumers' interests and functionality of destination sites
W. Pijls and J.C. Bioch Mining frequent itemsets in memoryresident databases
C. Castelfranchi, F. Dignum, C.M. Jonker, and J. Treur Deliberate Normative Agents: Principles and Architecture
L.F.A. Wessels, M.J.T. Reinders, R. Baldocchi and J. Gray Statistical analysis of gene expression data 12.25 – 13.30 Lunch BNVKI newsletter
129
October 1999
Manufacturing of Integrated Circuits 17.30 – 18.00 Introduction to the RoboCup soccer demonstration by Frans Groen: Robot Soccer: Game or Science? 18.00 – 19.30 Happy hour sponsored by Bolesian and RoboCup soccer demonstration sponsored by Tuyllian 20.00 – 23.00 Conference Dinner
14.45 – 15.00 Coffee Break 15.00 – 16.15 Three Parallel Sessions (15.00 – 15.25; 15.25 – 15.50; 15.50 – 16.15) KNOWLEDGE REPRESENTATION AND SYSTEMS A. Mittal and K.K. Biswas A Knowledge-Based Framework For Satellite Video Indexing P. Beys and M. Jansen Automatic Reuse of Knowledge: A Theory
THURSDAY NOVEMBER 4
09.00 – 09.50 Three Parallel Sessions (09.00 – 09.25; 09.25 – 09.50)
A. ten Teije and F. van Harmelen Describing Problem Solving Methods using Anytime Performance Profiles
LOGIC AND REASONING 2
AI AND LAW (JURIX SESSION)
H.van Ditmarsch The Logic of Knowledge Games: showing a card
L. Mommers Transfer of knowledge in the legal domain
J. Kamps On Criteria for Formal Theory Building
R. Winkels, D.J.B. Bosscher, A.W.F. Boer and J.A. Breuker Generating Exception Structures for Legal Information Serving B.Verheij Automated Argument Assistance for Lawyers
EVOLUTIONARY COMPUTATION 2
A.J.M.M. Weijters and J. Paredis Discovering Rules with a Genetic Sequential Covering Algorithm (GeSeCo) I.G. Sprinkhuizen-Kuyper, C.A. Schippers, and A.E. Eiben On the real arity of multiparent recombination
ROBOTICS AND VISION 1
J. Van Looveren Multiple Word Naming Games S.H.G. ten Hagen, D. l' Ecluse and B. Kröse Q-Learning for Mobile Robot Control
MACHINE LEARNING AND NEURAL NETWORKS 2
E.D. de Jong Autonomous Concept Formation
M. Schuemie and J. van den Berg Information Retrieval Systems using an Associative Conceptual Space and SelfOrganising Maps
16.15 – 16.30 Coffee Break 16.30 – 17.30 Invited lecture by Edward Rietman (Lucent Technologies): AI Techniques in
W. Peng, J. Nijhuis, and L. Spaanenburg Notes on Embedding a Trained Neural Network 09.50 – 10.05 Coffee Break
BNVKI newsletter
130
October 1999
10.05 – 11.05 Invited lecture by Jonathan Schaeffer: The Games Computers (and People) Play 11.05 – 11.20 Coffee break 11.20 – 12.35 Four Parallel Sessions (11.20 – 11.45; 11.45 – 12.10; 12.10 – 12.35)
R.M. van Eijk, F.S. de Boer, W. van der Hoek, and J.-J.Ch. Meyer Open Multi-Agent Systems: Agent Communication and Integration A.G. Pérez and V.R. Benjamins Overview of Knowledge Sharing and Reuse Components: Ontologies and Problem-Solving Methods
DEMONSTRATIONS 1
R. Ekkelenkamp IGUANA: A Web Crawler Aimed at Creating a Domain Specific Web Search Engine
C.M. Jonker, R.A. Lam, and J. Treur A Multi-Agent Architecture for an Intelligent Website in Insurance ROBOTICS AND VISION 2
M.C. van Wezel, J. Sprenger, R. van Stee, J.A. La Poutré, and J.B.M. van Wieringen Neural Vision 2.0-Exploratory Data Analysis with Neural Networks
P. Vogt Grounding a Lexicon in a Coordination Task On Mobile Robots
J.N.H. Heemskerk, R. Klopman, M.R. Vonder and R. de Wit Agent Based Customer Service
R. Kortmann, E. Postma, and H.J. van den Herik The trade-off between spatial and temporal resolution in visual systems
AI IN MEDICINE (SPECIAL SESSION)
12.35 – 13.45 Lunch and BNVKI meeting 13.45 – 15.00 Four Parallel Sessions (13.45 - 14.10; 14.10 14.35; 14.35 - 15.00)
N.L.W. Keijsers, M.W.I.M. Horstink, and C.C.A.M. Gielen Detection and Assessment of the Severity of Levodopa Induced Dyskenesia in Patients with Parkinson's Disease by Neural Networks
DEMONSTRATIONS 2
S. Spreeuwenberg and R. Gerrits, A Knowledge Based Tool to Validate and Verify an Aion Knowledge Base
D.M.H. Van Hyfte, P.A. de Clercq, T.B. Tjandra-Maga, F.G. Zitman, and P.F. de Vries Robbé Modelling the psychoactive drug selection application domain at the knowledge level
J.H. van Lieshout and E.C. van de Stadt KMD-MATE-An analysis and design environment for Knowledge Management Support
P.A. de Clercq, J.A. Blom, A. Hasman, and H.H.M. Korsten GuiDE: an architecture for the acquisition and execution of clinical guidelineapplication
J.I. van Hemert and A.E. Eiben Mondriaan Art by Evolution EVOLUTIONARY COMPUTATION 3
J. Eggermont, A.E. Eiben and J.I. van Hemert Comparing genetic programming variants for data classification
AGENT TECHNOLOGY 2
BNVKI newsletter
131
October 1999
DEMONSTRATIONS 3
W.B. Langdon Size Fair Tree Genetic Programming Crossover
P.A.N. Bosman and D. Thierens Interactive and continuous visualizations of EAs: The EA Visualizer
P.A.N. Bosman and D. Thierens On the Modelling of Evolutionary Algorithms
W. Teepe "Wij kiezen partij voor u" Online Voting Advise
AGENT TECHNOLOGY 3
L. Hulzebos (Special talk of Bolesian) Promising Practical Fruits of AI
M. Albers, C.M. Jonker, M. Karami and J. Treur An Electronic Market Place: Generic Agent Models, Ontologies and Knowledge
LOGIC AND LEARNING THEORY G. Bontempi and M. Birattari A bound on the cross-validation estimate for algorithm assessment
D.E. Herlea, C.M. Jonker, J. Treur and N.J.E. Wijngaards Specification of Behavioural Requirements within Compositional Multi-Agent System Design
H. Jurjus and H. de Swart Implication-with-possible-exceptions M. Bertolino, S. Etalle, and C. Palamidessi The Replacement Operation for CCP Programs
L. van der Torre and Yao-Hua Tan Rights, Duties and Commitments between Agents
MACHINE LEARNING AND NEURAL NETWORKS 3 M. Bot and W.B. Langdon Application of Genetic Programming to Induction of Linear Classification Trees
ROBOTICS AND VISION 3 T. Belpaeme Evolving Visual Feature Detectors
S.M. Bohté‚ H. La Poutré and J.N. Kok Unsupervised Classification in a Layered RBF Network of Spiking Neurons
A. Kröse, R. Bunschoten, N. Vlassis, and Y. Motomura Appearance based robot localization
F. Verdenius and M.W. van Someren Top-down Design and Construction of Knowledge-Based Systems with Manual and Inductive Techniques
P. Andras, E. Postma, and H.J. van den Herik Dealing with Environmental Dynamics
16.30 - 17.00 Closing and Awards. LOGIC AND GAMBLING
15.00 - 15.15 Coffee Break 15.15 - 16.30 Three Parallel Sessions (15.15 –15.40; 15.40 – 16.05; 16.05 – 16.30)
BNVKI newsletter
Jaap van den Herik IKAT, Universiteit Maastricht
132
October 1999
clearly present in this world: Internet, ICT, and AI are combining their forces.
The world is full of structure. For any Ph.D. student it is a real challenge to reveal part of the structure to her/his supervisor. Those who succeed in doing so are bestowed with the honour of being listed in the pages of the BNVKI newsletter and as a consequence they receive the degree of Doctor after a successful thesis defence. This is logical and therefore full of logic. But how about finding structures in Chaos and Gambling? Is there any structure in chaos? Can gambling help us to find structures in chaos? Or should we use Informed Gambling?
H. Vandecasteele (May 26, 1999) Constraint Logic Programming: Application and Implementation. Katholieke Universiteit Leuven. Promotor: Prof.dr. Danny De Schreye. Additional promotor: Prof.dr.ir. Maurice Bruynooghe G. Zwaneveld (June 4, 1999) Kennisgrafen in het Wiskundeonderwijs. Open Universiteit Nederland Promotor: Prof.dr. J. van Craats. Additional promotor: dr. A. van Streun A. de Moor (October 1, 1999) Empowering Communities: a Method for the Legitimate User-Driven Specification of Network Information Systems. Katholieke Universiteit Brabant. Promotor: Prof. dr. R.A. Meersman; co-promotor: Dr. H. Weigand.
Of the six thesis defences announced below, four are a repetition of the announcement in the August issue. Two are new. These are the ones of which the dates have already passed. The Editorial Board congratulates H. Vandecasteele and G. Zwaneveld with the receipt of their Doctor titles. The same holds for A. de Moor (October 1). We wish Fernandez de Montessinos, Lenting, and Willemse much energy with the preperation of their defence. Moreover, we are happy to announce that we publish three reviews. The review below by Bart de Boer is not specifically on a Ph.D. thesis, but on a book written by Simon Kirby which resulted from a Ph.D. Thesis. A propos, this thesis has never been announced in the BNVKI Newsletter. Two Ph.D. theses which have been announced in our pages, are now reviewed. First, Marnix Weusten’s thesis is reviewed by Bart Verhey, and second, Ronald Leenes’ thesis by Kees de Vey Mestdagh. Both reviews are placed in the Section Knowledge Systems in Law and Computer Science. Readers interested in applications of AI in the real world, especially in the world of Law are encouraged to read the reviews. They constitute one more sign of progress in a world which consisted only a decade ago mostly out of disbelievers of Artificial Intelligence. The change is
M. Aznar Fernandez de Montessinos (October 25, 1999) Aeroplane Design: AI applications. Technische Universiteit Delft. Promotor: Prof. dr. H. Koppelaar. J.H.J. Lenting (December 3, 1999) Informed Gambling: Conception and Analysis of a multi-agent mechanism for discrete reallocation. Universiteit Maastricht. Promotor: Prof. dr. H.J. van den Herik; co-promotor: dr.P.J. Braspenning. W.J. Willemse (December 3, 1999) Computational Intelligence: Life without Tables for the Actuary. Technische Universiteit Delft. Promotor: Prof.dr. H. Koppelaar.
Book Review
FUNCTION, SELECTION, AND INNATENESS
BNVKI newsletter
133
October 1999
Kirby begins his book by posing the central question: how is it possible that languages tend to be close to optimal from a functional point of view? He calls this “The puzzle of fit” (borrowing a phrase from Gary Cziko) or the “Problem of linkage”. If we look at a large sample of languages, we find they have universal properties that make them easier to process. How can it be explained that constraints of processing influence the form of language?
SIMON KIRBY
A review by Bart de Boer AI Lab, VUB
Sometimes I encounter a book that is a real intellectual adventure to read. Such a book must be related to my own research interests, there must be lots of fascinating material in it, but there must also be enough in it that I do not agree with. That way I am constantly faced with the question why I do not agree with it, and what arguments I can find to defend my own position. Simon Kirby’s book “Function, Selection and Innateness, The Emergence of Language Universals” from Oxford University Press is just such a book. The book is based on Kirby’s Ph.D. thesis and has been published in a series of books on the evolution of language, in which Andrew Carstairs-McCarthy’s book “The Origins of Complex Language” has also been published.
KIRBY’S HYPOTHESIS
In the second chapter, Kirby presents the hypothesis that constructions that are harder to process are harder to learn. The basis of this hypothesis is that the difficulty of parsing linguistic input determines how often it will be presented to the language acquisition device. The fact that a new generation has to learn the language from the previous one under these pressures then might cause some constructions to disappear from the language and some to be favoured. Kirby tests this with a computer simulation. In his model, certain constructions are assigned an a-priori parsing difficulty (based on the complexity of the parsing tree). He then generates a population of language users that all have a random language. The probability of learning a construction is based on the difficulty of parsing. He shows that under these circumstances the most functional type of language will be preferred, and that the composition of the language changes over time follows the same sigmoid curve as language change observed in human languages. Kirby recognises that his simulation has a problem: it will generally result in only one language type emerging, while in reality, different types co-exist. He therefore introduces what he calls “competing motivations” in his third chapter. Competing motivations are in fact pressures that work in opposite directions.
UNIVERSAL PROPERTIES
In his book, Kirby addresses the problem of how universal properties of human languages can emerge. At first sight, this question might seem to be remotely related to research in the field of artificial intelligence, but Kirby’s solution to this problem as well as his methodology are of interest to the AI-community. Kirby investigates what aspects of language can be explained by functional pressures (and are therefore learnt) and what aspects must be innate, as well as what mechanism can explain how functional constraints can make language to be maximally functional. The debate of what is learnt and what is innate is also central in artificialintelligence research. Also, Kirby uses computer simulations of populations of language learners in order to test his hypotheses.
BNVKI newsletter
134
October 1999
preferred over others. The language acquisition device on the innatist side, causes some languages to be learnable and others not. Furthermore, evolutionary pressure causes the language acquisition device to become more and more adapted to functional languages. This way, Kirby argues, the puzzle of fit is (at least partly) solved.
Kirby uses p-complexity (parsing complexity) and m-complexity (morphological complexity). If a sentence is hard to parse, this can be compensated for by introducing more morphology (inflections, cases etc.) so that the listener is aided in understanding the sentence. By using m-complexity and p-complexity of which the relative influence change randomly over time and by using a spatial distribution of agents, so that only agents that are neighbours in a grid communicate, it is shown that co-existing different language types emerge.
Kirby provides a laudable synthesis between functionalist and innatist approaches to explaining language universals. He shows that these two schools of thought are not at all incompatible with each other, but that they are, on the contrary, both needed in order to explain language universals. He even goes beyond creation of mere theories, and puts his to the test with computer simulations. I think in combining innate and functional criteria Kirby is right on the mark, and that his ideas are applicable to other domains of cognition, so that his book is not only interesting to people interested in language universals or the origins of language.
LIMITS
Then Kirby argues that there are limits to the explanatory power of functional constraints. Using examples from different languages, he shows that sometimes functionally advantageous constructions do not occur in human language, and that there might even be phenomena that are maladaptive. These, according to Kirby, are the result of the innate properties of the “Universal Grammar” or “Language Acquisition Device” shared by all humans. Universals of human language can therefore be explained as an interaction between functional constraints that tend to push languages in a certain direction, and the universal grammar, that causes some languages to be learnable and others not.
CRITIQUE
Still, as I mentioned in my introduction, I did not agree with everything Kirby has done. First of all, the computer simulations Kirby uses are rather simple and ad hoc from the perspective of an AI-researcher such as myself. In principle there is nothing wrong with simple computer simulations, in fact they are to be preferred over overly complex ones, as it is clearer to investigate their behaviour. However, Kirby has simplified his model almost too far, in my opinion. What is left over of a very interesting and plausible theory of the influence of parsing complexity on language is basically a difference equation in which the emerging behaviour (the sigmoid transition towards preferred language types) can be predicted from the equations by direct mathematical analysis.
He then goes on in the last chapter before his conclusions to show how it is possible for Darwinian evolution in combination with functional constraints on language use to influence the human Language Acquisition Device through the Baldwin effect. In this way it becomes possible that the human capacity for language becomes adapted for acquiring functional languages only. Kirby finally combines the different influences on language universals in one diagram, with a functional side and an innatist side. Functional pressure on the one side causes some languages to be BNVKI newsletter
135
October 1999
The behaviour is therefore no longer really emergent. This is probably also a cause why Kirby has to introduce what are in my opinion rather ad hoc solutions, such as a randomly changing influence of mcomplexity and p-complexity and a spatial distribution in order to get different language types in one population.
A report by Edwin de Jong
VUB AI Lab http://arti.vub.ac.be/~edwin From July 31 to August 6, I visited the Sixteenth International Conference on Artificial Intelligence (IJCAI) in Stockholm. IJCAI is one of the oldest international AI events. It is also one of the largest conferences on the subject; this year’s edition featured 29 workshops, 20 tutorials, and 194 out of 750 submitted conference papers. Clearly, this report will not cover more than a fraction of the papers that were presented; it describes the presentations that I found most interesting of the ones I have visited.
STRUGGLING WITH ARTE FACTS
Also, I think that some problems Kirby needs to explain are caused by the adherence to certain aspects of generative grammar, such as for example the autonomy of syntax. Of course, as generative grammar is very influential in linguistics, it is necessary to argue that a theory is not in conflict with it. However, to a relative outsider like myself, it seems as if some of the problems Kirby struggles with are artefacts of the generative theory, rather than real problems.
THE STATISTICAL MACHINE LEARNING FOR LARGE SCALE OPTIMIZATION WORKSHOP
The workshop Statistical Machine Learning for Large Scale Optimization was based on the observation that techniques from machine learning are becoming sufficiently powerful to be utilized for real world optimization problems. One of the application domains listed by the organizers is Computer Aided Design (CAD). A frequently recurring task within this domain is that of graph partitioning, where nodes of different classes have to be separated from each other, cuts are associated with costs, and the total costs of the cuts have to be minimized. Recent methods addressing this problem start by finding a clustering of the nodes of the graph. The initial problem is then transcribed into that of partitioning the graph of clusters, which decreases the problem size by an order of magnitude while maintaining an adequate performance. Graph partitioning methods were presented by Wray Buntine and by Ted Carson.
CONCLUSION
Concluding, I think Kirby’s book is a valuable contribution to the question how linguistic universals can be explained as well as to the wider debate of the role of innateness versus functionalism in explaining cognitive phenomena. His use of computer simulations in investigating his theories provides good additional support to his arguments. Furthermore, it is compact and very well written. I would therefore recommend it to linguists, as well as cognitive scientists and artificial intelligence researchers.
THE SIXTEENTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE
BNVKI newsletter
COUNTING LOCAL MINIMA
136
October 1999
Rich Caruana of JustResearch presented a method for estimating the number of local minima in large complex search spaces. His method was inspired by a problem familiar in statistics, called the birthday problem: how many people must be in a room for the probability of two or more people sharing the same birthday to exceed ½ 1. A variation of this is the Martian Year problem: how long is a Martian year if on average we have to coax k Martians into a room before two of them have the same birthday? Analogously, one may randomly sample a search space and count the number of local minima before a duplicate is found to obtain an estimate of the total number of local minima. The lower bound provided by this creative technique may for example be used for selecting an appropriate solution method; the required information becomes available during optimization anyway, and does not incur much extra computational costs.
proverbial pushing of a button, given selected basis functions. As a step towards this ambitious goal, he presented a TD approach to optimal stopping, along with theoretical results concerning its convergence and approximation error. INSTRUCTION SCHEDULING
Andrew Barto, a main contributor to the increased formal understanding of reinforcement learning that has spurred developments in this branch of research, presented an application of reinforcement learning to instruction scheduling. Instruction scheduling is the problem of how the instructions in a basic block of machine code may be reordered in order to minimize their execution time. TD learning was shown to yield good schedulers without using any heuristics. The benefit of the method does not lie in the overall performance of the learned schedulers, which is a bit lower than that of commercial schedulers, but rather in the fact that it provides a means of quickly generating good schedulers. In addition, it outperformed a commercial scheduler when trained and tested on specific problems, suggesting a potential for generating high performance special purpose schedulers. The approach may save development time in settings where schedulers for new architectures are frequently required.
OPTIMAL STOPPING
Benjamin Van Roy presented a method developed by John Tsitsiklis and himself that approximates solutions of optimal stopping problems. The problems consist of deciding when to stop a Markov chain that returns values at each time step and a terminal reward at the end such that the expected discounted sum of rewards is maximized. The talk started with a clear and well-presented overview of Temporal Difference learning. Temporal Difference (TD) methods, named such by Richard Sutton, learn to predict a temporally distant value by making adjustments based on differences between temporally successive intermediate estimates, rather then the final outcome, and are the basis for most modern reinforcement learning methods. General convergence results for reinforcement learning methods require the use of lookup tables, which limits the scope to small problems. Autonomous systems need to use approximations of the value function, which may cause divergence. Van Roy envisioned a time when spreadsheet programs will have built-in support for reinforcement learning problems which can be activated by the 1
REGRESSION TREES
Tom Dietterich presented a regression tree approach to value function approximation. A novel aspect of this work is the use of three error terms: the supervised training error, a Bellman error term, and an advantage error term, all of which were found to be important for obtaining good performance. These terms are combined into a single composite error by weighting. Experimental results with a job-shop scheduling task indicated that performance of the regression tree method was comparable to an earlier neural network approach to the same problem, suggesting a potential for regression trees as approximators of the value function. But several problems need to be overcome first; the current method requires estimates of state values in order to compute the supervised error term, regression trees are not well suited for incremental learning, and all features are assumed to be equally relevant and uncorrelated. PANEL SESSION
Organizer Wray Buntine opened the panel session by stressing the importance of turnaround time as a
The answer is surprisingly low: at least 23 people
BNVKI newsletter
137
October 1999
hierarchy of categories is the result. Although current results are interesting on themselves, I think a promising extension of the work would be to take relevance of concepts for the robot’s behavior into account in the concept formation process.
factor governing the applicability of machine learning methods to optimization. If learning methods can be rigged up quickly and produce answers within specified time limits, they may obtain a place within optimization practice, especially where traditional methods are troubled by large search spaces. Dietterich stressed the importance of using appropriate representations. Barto stated that learning value functions is hard, that no turn-key methods are available, and that new algorithms and theory are necessary. Boyan was more optimistic, and suggested that learning methods employing rollouts might become important. A rollout algorithm uses a given reasonable policy to estimate the values of different actions, and changes its policy according to these estimates. Although computationally intensive, rollouts can often give good performance.
EMBEDDED LANGUAGES
THE NEURAL, SYMBOLIC, AND REINFORCEMENT METHODS FOR SEQUENCE LEARNING WORKSHOP
Mikael Bodén has investigated a recurrent neural network that learns a deeply embedded language. Whereas regular languages have been learned with success by several researchers (Pollack, Cleeremans, Elman and others) these more difficult languages may cause instability in networks. Bodén trained a set of networks on the task of learning the language anbn. The hidden state that is necessary to recognize this language is encoded in oscillations of two hidden neurons, which requires considerable precision, especially for large n. An analysis of the error gradients revealed a spiky landscape, especially near the region of interest. Given the difficulties any gradient method will have on such an error surface, this explained the problematic behavior. Current work therefore considers whether choosing a different search space may improve learning performance.
CLUSTERING TIME SERIES
THE TECHNICAL PROGRAMME
Tim Oates presented an interesting method for clustering times series with Hidden Markov Models (HMMs) and Dynamic Time Warping (DTW). The method automatically determines the number of HMMs that generated the data, along with their parameters. This determination was done with DTW. DTW computes a warping of two time series that minimizes the difference between the series. The remaining differences between the time series, computed as the area between them, is used as a distance measure in an agglomerative hierarchical clustering method. The rough clustering that results from this is then refined by a clustering method that iteratively trains an HMM on the set of sequences and throws out the sequences that do not fit the resulting HMM. This hybrid clustering method yielded promising experimental results, and is the subject of continued research.
BOOSTING
Robert Schapire delivered an invited lecture on the theory and practice of boosting. Boosting is a recent class of methods that builds strong classifiers by computing weighted combinations of simple decision rules. The simple decision rules are easily found using a weak classifier, and weighting is such that emphasis is put on the difficult cases that are classified wrongly by the simple rules. These cases often turn out to be outliers. Several problems that affected early boosting methods were fixed in later methods, such as AdaBoost [Freund and Schapire, 1995]. A counterintuitive result is that for certain problems, boosting does not only decrease the training set error (exponentially in the number of passes), but also the test set error, even when the training set error has already dropped to zero. This is a result of boosting’s property of increasing the margins of the training examples, which links boosting to Support Vector Machines. However since there is no such thing as a free lunch, there must also be problems for which boosting does not work so well. As might be expected, this is the case for problems with substantial noise, as shown by Tom Dietterich [Dietterich, to appear].
CONCEPT FORMATION
Paul Cohen from the University of Massachusetts presented work with Michael Rosenstein on autonomous concept formation with a mobile robot. Events were detected based on correlations of the sensor time series with several templates. A time window around the event was then fed into a clustering procedure, and after computing signatures and performing a second clustering, a BNVKI newsletter
BAYESIAN NETWORKS
138
October 1999
how a robot may receive information about causal relationships from its designer. Although this was not a question for Pearl himself anymore, he was was willing to attempt to convey his tranquility to the audience, as he put it himself. The central topic was a procedure for evaluating counterfactual questions, such as “would A have been different if it were not for B?” Based on the principle of ‘mutilate and simulate’, which involves forcing variables in the network to take particular values, these questions can in certain cases be answered. He gave several examples of this procedure, and mentioned actual applications to policy evaluation and troubleshooting. Unfortunately, when he came to the point of a mid-talk summary, the time allotted to the lecture was almost over. Judea Pearl’s lecture was rewarded with a standing ovation.
David Heckerman of Microsoft Research gave an overview of work on Bayesian networks. In the early days of Bayesian networks, they were constructed by hand, based on interviews with domain experts. Nowadays, it has become common practice to construct automatically Bayesian networks by learning them from data. An example is collaborative filtering, where the preferences of a large group of consumers are searched to find relationships between their interests. Bayesian networks are enjoying increased attention, as can be seen from the overwhelming presence of work in this area at the recent Uncertainty in AI conference (UAI) and the increasing number of commercial Bayesian network packages. Technically, the core idea of a Bayesian network is that a probability distribution over a set of random variables is dissected into a set of local distributions. The dependencies between the different variables are graphically represented in a directed graph. If these dependencies are sparse, it becomes feasible to compute the complete jointprobability distribution from the local distributions. The presentation described the construction of Bayesian networks by combining both expert knowledge and data. The focus was on how these construction techniques can be used to determine causal relationships from observational data. Although a long held view of statisticians is that causal relationships cannot be learned from observational data, Heckerman holds that this is possible under certain assumptions. The argument is based on the causal Markov assumption [Spirtes, 1993], which states that if a direct acyclic graph C is a causal graph for a set of variables, i.e. nodes correspond to variables and arcs to causal relationships, then C is also the structure of a Bayesian network for the joint physical probability distribution of those variables. Furthermore, a limited set of possible causal networks is assumed. If a particular Bayesian network can then be learned from the data with high probability, this rules out those causal networks that are inconsistent with the Bayesian network that was found. If the assumptions are correct, this elimination process may lead to a single causal model.
MEASURING MOTION
Radu Horaud was originally interested in visual servoing and the coordination of vision and action. When he realized the importance of how motion is measured he began to use a vision system with two cameras instead of one so that a motion representation consistent with stereo vision could be investigated. Under the condition that the position of the cameras relative to each other is fixed, this approach yields accurate and reasonably fast estimations of motion, and practical applications of the method are expected. ROBOCUP
Whereas in earlier robotic soccer events games were sometimes rather static and the purpose behind moves was not always clear, this year’s Robocup event provided some quite spectacular shows. Several problems that had been encountered on earlier occasions had now been addressed. For example, some of the small size league teams had incorporated a solution for getting the ball out of the corner by spinning their small robot. The speed of the game in this league was sometimes truly amazing. The Robots from Korea for example were able to catch up with the ball, get behind it, and drive it into the goal of their opponents. The competition in this league was won by Cornell University. The team from Iran made up for last year’s disappointment when they could not enter France in time to play in the Paris RoboCup by winning the mid size league competition. As Minoru Asada remarked during his presentation, engineering issues are still a concern in RoboCup. It is to be hoped that the technological progress that has been made will allow the teams to focus on strategy in the future. Asada has a clear ultimate goal in mind for robotic soccer; to beat a human
CAUSAL RELATIONSHIPS
Causal relationships were also the subject of Judea Pearl’s lecture. Pearl received the IJCAI-99 award for research excellence, of which past recipients include John McCarthy, Alan Newell, Marvin Minsky and Herbert Simon. For the question of how a robot can acquire causal information from its environment, he referred to David Heckerman’s work. The focus of Pearl’s own presentation was on BNVKI newsletter
139
October 1999
team with a team of humanoid robots. As a video of the Honda humanoid robot showed, this will still take a while, but progress is being made.
Papers of widely varying quality and from different perspectives have been accepted for presentation. The relations between these different perspectives on anticipatory systems often remained unclear. In the following, a few presentations will be described that may serve as examples of different approaches to the subject.
NEXT EDITION
The Seventeenth IJCAI will be held August 5-10 2001 in Seattle, Washington. More information will become available through www.ijcai.org. REFERENCES
In an introductory paper on computing anticipatory systems Daniel Dubois, organizer of the CASYS conferences, describes two types of anticipatory systems. One class of anticipatory systems is that of systems with multiple potential future states for which the actual states the system visits are determined by the events “at each current time". This describes the class of dynamical systems. The other class mentioned by Dubois is that of systems that use an expectation about the future in determining their current action. In computer science, such a system would be called an autonomous agent. Both classes are clearly of interest to AI.
[Dietterich, to appear] T.G. Dietterich. An experimental comparison of three methods for constructing ensembles of decision trees: Bagging, boosting, and randomization. Machine Learning. (2000) [Freund and Schapire, 1995] Y. Freund and R.E. Schapire. A decision-theoretic generalization of on-line leaning and an application to boosting. Journal of Computer and System Sciences, 55(1):119-139, August 1997. [Spirtes et al., 1993] P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction, and Search. Springer-Verlag, New York, 1993.
CASYS’99: COMPUTING ANTICIPATORY SYSTEMS
A report by Edwin de Jong VUB AI lab
On August 9 and 10, I visited the first two days of CASYS'99, the third conference on Computing Anticipatory SYStems. The CASYS conference took place from August 9 to 14 in Liège, Belgium. There were 165 papers in four parallel tracks, so I could only attend a fraction of the presentations. Although artificial intelligence is in principle concerned with the construction of systems that anticipate, I had not come across the term anticipatory systems before in AI literature. Therefore, my main purpose in visiting the conference was to find out what computing anticipatory systems are and whether the research that is being done on them is relevant to AI. The conference programme listed many diverse subjects, and I was not sure what to expect; this might indicate that the field is well developed and has many different branches. However, my observations during the conference did not confirm this. BNVKI newsletter
Tom Quick gave an interesting presentation on embodiment. The presentation is a good example of research at the conference that is aimed towards an increased understanding of anticipatory systems, since it discusses what properties of such systems are fundamental, so that research may focus on systems possessing those properties. His main point was that what matters is not whether autonomous systems are physical systems or software systems, but rather what the nature is of the interaction between the system and its environment. George Mobus presented work on foraging agents. The agents in his experiments had to learn causal relations to anticipate resources or threats, and were carried out on a robot and with software agents. These presentations are examples of research that can also be found at AI conferences such as ECAL or SAB. The 140
October 1999
important to notice that not just any incursive relation represents a consistent system. This is because the variables x(t+k) are not arbitrary variables, but successive values of a single variable, and the relation thus specifies a system of equations when different values are substituted for t. In my view, systems that can be described by incursive relations are special in that successive states bear a particular relationship to each other, rather than demonstrations or explanations of how some future cause would affect the present. Furthermore, the definition of incursion does not take into account the idea that anticipation may involve a prediction about future states (be it of the system itself or its environment), since it only involves actual states. Its relevance to anticipation may therefore be questioned.
following sections describe presentations that were more specifically aimed at CASYS. INCURSION
The source of the term anticipatory systems is a book by Robert Rosen (1985) with the same title. In his own presentation, Dubois gave a review on computing incursive, hyperincursive and anticipative systems. He presented incursion and hyperincursion as two principles to be used for modeling, simulating and controlling anticipatory systems. At time t, the state of a system with incursion at the next timestep x(t+1) can be described as a function of among others x(t+1) itself. Hyperincursion is similar, but with the additional property that the equation has multiple solutions. In [1], Dubois defines an incursion as a relation that can be written as:
QUANTUM MECHANICS
Many papers at the conference referred to quantum mechanics. In the field of quantum computation, it is investigated whether quantum mechanical principles can be used to achieve a qualitative increase of the computational power of computers. In contrast, these were not the reasons for the interest in quantum mechanics at CASYS. Rather, the implicit assumption is that quantum mechanical effects may explain properties of anticipatory systems.
x(t+1) = F[..., x(t-1), x(t), x(t+1),...] "where the value of a variable x(t+1) at time t+1 is a function of this variable at past, present and future times." Thus, x(t+k) denotes the state of the system at time t+k, as usual. The next state of the system may therefore depend on future states of the system. In some cases, e.g. the example given in Dubois’s paper [1], such relations can be rewritten such that no knowledge about the future is necessary. In other cases it is not possible to determine these values at the current point in time. Dubois repeatedly mentioned the “final causation” principle of Aristotle in connection with anticipatory systems. In [2], this principle is described as follows: “A future cause could produce an effect at the present time. Then the causality principle seems reversed.”
In his abstract for the conference titled "Anticipation – A spooky computation", Professor Mihai Nadin expresses his view that anticipatory processes are related to quantum non-locality. He states that "anticipation is, of course, different from expectation or from forecasting". Furthermore, "in the realm of the living, correlations among separated but entangled parts of a system defy the accepted notions of causality (at least in its classical deterministic sense) and of unidirectional time progression".
The definition of incursive relations may appear to be a formalization of the final causation principle. However, it is BNVKI newsletter
141
October 1999
difficult to discern any relations between them. Ideally, these relations would be clear from links with the underlying paradigm of anticipatory systems. However, the implications of the presented research on knowledge about anticipatory systems more often than not remained undiscussed. This can be a consequence of obscurities in or unfamiliarity with the subject of the conference, such as the unclear role of incursive relations, and of a lenient review process. Therefore, my recommendation for possible future editions of the conference would be to focus on presentations that clearly aim at an increased understanding of or knowledge about anticipatory systems. This could be implemented in the review process, and would have the side-effect that the duration of the conference (6 days) will decrease somewhat, which would be welcome considering the chronic lack of time of the modern researcher.
Nadin does not explain how anticipation in living systems might relate to quantum nonlocality, other than in broad and general terms. This is a pity. Although there is little doubt today that quantum effects play a role at the smallest scales in physical systems, I know of no experimental evidence that quantum effects (such as entanglement) in living systems are related to thinking or anticipation. If quantum mechanics is to be a basis for research into anticipation, it is important to clarify how its properties may be related to anticipation. FURTHERMORE…
[1] D.M. Dubois (1998). Computing Anticipatory Systems with Incursion and Hyperincursion. AIP Conference Proceedings 437, Computing Anticipatory Systems: CASYS-First International Conference 1997. D.M. Dubois (ed.). American Institute of Physics.
Finally, there were papers that cannot easily be classified under an existing scientific dicispline. Maybe the best way to explain this is to give an example. To quote from one of the abstracts: “The II-III transition to humanness, exhibiting language, holistic thinking, consciousness and sentiency, arises by ‘en passant’ growth of Pavlovian nonholistic circuits into Steinbuch-Taylor (S-T) holistic matrix circuitry plus elaborations found in the cortex. In the context of the hamiltonian organization of man and animal, this transition from Pavlov and the von Neumann externally programmed computer to the self-programming S-T matrix computer and mankind requires only a few mutations.” [emphasis in the original]
[2] D.M. Dubois. Introduction to Computing Anticipatory Systems.
DISCUSSION
An extremely wide variety of presentations has been accepted for presentation at the conference, as can be judged from the titles of the presentations: these contain such diverse topics as "general axiomatic theory of everything", space flight, spooky computation, holism, diffusion in chemical systems, and the size of the Internet. In general, assembling researchers with different experience and viewpoints can lead to very fruitful events. At CASYS however, the differences between the presentations were often so large that it was BNVKI newsletter
142
October 1999
The programme of the second European workshop on neuromorphic systems (EWNS-2) covered five different topics: general papers, auditory systems, visual systems, robotics, and hardware for neural network models. The talks were spread out over three days and were followed by a panel discussion on Sunday afternoon.
NOUVELLE AI: A REPORT ON FOUR SEPTEMBER EVENTS
EWNS-2, PhD course on behavior based robotics, ECAL-99, and EWLR-8 Rens Kortmann
IKAT Universiteit Maastricht
[email protected]
On the first day, we were welcomed on the scenic campus of the University of Stirling by Leslie Smith, the main organiser of the workshop. In his Scottish accent he announced Pedro Marijuan, professor at the University of Zaragoza, who gave an invited lecture on the quest for a neurodynamic optimization principle. The quest took us from Darwin to Ramon y Cajal, a famous Spanish physiologist and explained how animals developed a neural signaling system. Then, Barbara Webb (well-known for her robot model of cricket phonotaxis) proposed a framework for talking about bio-mimetic models. Very often, people discuss models of biological behaviour without adopting a common terminology. The careless use of terms, though, often causes misunderstandings that can be avoided when adopting a shared framework. The framework she proposed is not in its final stage yet, but very much worth further investigations. The last general talk was given by Catherine Breslin who is interested in the morphology of neurons and proposed a technique for investigating how the form of a cell affects its function.
The first three weeks of September featured four scientific events on what is sometimes referred to as Nouvelle AI, behaviour-based AI, or Artificial Life. There exist no exact definitions of these terms, so, hopefully, this report will give an impression of what is meant here. Also, hopefully, readers will feel inspired by the topics covered in the events, so that also here, in Belgium and the Netherlands, it will be possible to establish a Nouvelle AI tradition. First, the computer science department of the University of Stirling hosted the second European workshop on neuromorphic systems (ewns-2) from 3 to 5 September. As the call for papers says: `Neuromorphic systems are implementations in silicon of systems whose architecture and design are based on neurobiology.' Second to fourth, the Swiss federal institute of technology (EPFL) offered (2) a PhD course on behaviourbased robotics, (3) the fifth European conference on artificial life, (ECAL-99) and (4) the eighth European workshop on learning robots (EWLR-8). The common theme of all events was the situated approach to artificial intelligence. Rather than studying, e.g., disembodied reasoning systems, the situated approach investigates relatively simple, though robust, interaction patterns of artificial systems with the real world. Starting from this position, the situated approach aims to scale up to higher levels of intelligence.
AUDITORY SYSTEMS
The sessions on auditory systems mainly featured talks on techniques, inspired from biology, for detecting directionality of sound sources with a silicon pair of ‘ears’. The hardware solutions are particularly interesting when one needs cheap and robust devices that use little power. A drawback is that bio-inspired systems are usually very task-specific.
EWNS-2, SEPTEMBER 3 - 5 SILICON BRAINS AND ROBOT ANIMALS
BNVKI newsletter
143
October 1999
HARDWARE IMPEMENTATIONS VISUAL SYSTEMS
The last session concentrated on hardware implementations of neural systems. The first speaker (Shih-Chii Liu) implemented a silicon model of the fly visual sensors and generated optomotor response and fixation behaviour after mounting the chip on a Koala robot. Another interesting device, amongst others, was a programmable neural network chip (integrate-and-fire type), presented by Leslie Smith. The chip consisted of four neurons and 128 programmable synapses. Performance was still somewhat noisy, but had already considerably improved with respect to the prototype.
In the sessions on vision, again, the emphasis was on cheap, bio-inspired systems that can be applied more robustly to visual tasks than traditional computer vision systems. For example, Reid Harrison, who later received the best-paper award, implemented gain control in an existing model of fly visual motion detection. The addition, based on the function of particular interneurons in the fly brain, improved the performance of the model significantly. Usually, neuromorphic chips are based on the neural systems of small animals on which the most accurate data exist. Still, two speakers (Peter McOwan and F. Wörgötter) based their designs on human cortical processes and wondered ‘why just to model flies’.
In the final discussion round, I brought forward the issue of noise in a neural system. A motivation for implementing neural networks or neuromorphic sensors in hardware is the fact that in a softwarebased simulation, one is not always sure to have represented the environment correctly (e.g., noise factors). When using a hardware device, one gets the environment ‘for free’, i.e., noise in signal transfer through the medium (e.g., the propagation of sound waves through the air) is the same for natural systems as for bio-mimetic hardware models. Still, noise within the system is not necessarily the same in silicon as in neural substrate. Therefore, although hardware models are more accurate representations of natural sensory systems than software simulations - they deal with the complexity of the real world still, they do not incorporate possibly relevant internal noise factors. Finally, after a very enthusiastic British Telecom representative encouraged us to apply for BT funding, the meeting was closed.
ROBOTICS
The session on robotics was introduced by Avis Cohen, professor at the University of Maryland. She talked about lessons to be learnt from biology for application in robotics. As an example, she discussed central pattern generators (CPGs), neural circuits that transform tonic input into periodic muscle activation. CPGs are responsible for motion patterns found in many (vertebrate) animals and recover easily from perturbations employing sensory feedback. Other talks in this session concentrated either on high-level controllers involving neural maps for navigation (R. Mudra), or on low-level robot control modelling insect behaviour. Examples of the latter approach are the robot models of wind-mediated escape behaviour in crickets (Tim Chapman), of collision avoidance response in locusts (Mark Blanchard), and of visual homing behaviour found in various insects (Ralf Möller).
BNVKI newsletter
ROBOTICS COURSE, SEPTEMBER 8 LITTLE BRAIN, FEW MUSCLES, IMPRESSIVE BEHAVIOUR
144
10
October 1999
could evolve relatively simple control mechanism for complex tasks through selforganisation in the controller. There were examples of evolving the synaptic weights of a neural network, the learning rules governing the life-time learning of a robot, and of co-evolution between either different controllers (e.g., predator-prey) or between the controller and the robot's task/environment. Third, Takashi Gomi spoke about applications in ‘dark’ and ‘not-so-dark’ domains. He and his company decline all project-proposals related to military, `resource guzzling', or nature destructive purposes (dark domains). Rather, they aim at `not-so-dark' applications, such as service robots (household or office environments), autonomous wheelchairs, or robots similar to the Sojourner that investigated Mars in 1997. Although his talk was rather much focused on ethics, it gave a clear idea of possible applications of BBR in domains other than the military.
After a rapid train journey through Britain and France, I arrived in Lausanne, Switzerland on September 7. The day after, a 3-day PhD course on behaviour-based robotics started in the premises of EPFL, the Swiss federal institute of technology. Behaviour-based robotics (to which I will refer hereafter as BBR) is an alternative to the traditional approach to building robot controllers. The emphasis is not on building environmental maps and deliberative planning, but on simple, tight couplings of sensors and actuators instead. On the one hand, it appears that BBR is much better at handling a complex, changing world - the robots navigate more robustly through cluttered, dynamic environments than traditional, deliberative robots: the latter need to rebuild their environmental maps and re-plan their actions every time there is a small change in the environment which makes real-time performance infeasible. On the other hand, purely behaviour-based robots are just reactive agents and therefore difficult to design for more complex tasks than, e.g., light-seeking or obstacle avoidance. As a result, many industrial projects now adopt a hybrid deliberative/reactive approach in the design of their robots. The course was taught by three well-known researchers from the field. First, Ronald Arkin, who wrote one of the first comprehensive textbooks on BBR, gave a two-day lecture on: the objectives of BBR; background (animal behaviour); techniques and architectures for BBR; representational and perception-related issues; and finally, social and adaptive behaviour in robots. The emphasis was on engineering robust and reliable controllers for real-world tasks, mainly in the military domain (e.g., autonomous mine-sweepers). The other two lecturers showed different interests. Second, Dario Floreano spoke about evolutionary robotics applying evolutionary algorithms to robot controllers - from a scientific point of view. He showed how a genetic algorithm BNVKI newsletter
ECAL-99, SEPTEMBER 13 - 17 THE MEANING OF (A) LIFE
The fifth European conference on artificial life (ECAL) took place from 13 – 17 September also held at EPFL was chaired by Dario Floreano. The conference consisted of a day of tutorials, four consecutive days of single-track talks, demonstrations and poster presentations, and a guided tour to see artificial life in Switzerland from the 18th century (mechanical automata). Since it is too much to summarize all talks, I highlight only a few of all remarkable presentations. On Monday, six tutorials were given in parallel sessions. Amongst others, Marco Dorigo spoke about ‘ant algorithms’. Ant algorithms exploit artificial stigmergy, an indirect communication scheme adopted by many social insects. Applications of ant 145
October 1999
species involved. Symbiosis allows a group of species to evolve in ways that are often impossible for a single species: by ‘borrowing’ each other's traits, the fitness landscape changes for each species. For instance, a rhinoceros is not able to get rid of its parasites alone, whereas symbiosis with a bird species that feeds on the parasites does enable adaptation to them. The work presented here was much more abstract, i.e., it did not model real animals. Still, the evolutionary mechanisms themselves became clearly visible. Later, Tim Taylor presented his analysis of selfreproducing programs in artificial life platforms such as Tierra (Thomas Ray). In contrast to Von Neumann's genetic architecture for self-reproducing machines, he proposed to encode the reproduction process implicitly in the environment (in Von Neumann's architecture, the machines carry the reproduction process explicitly in their own description). He suggested the focus of artificial life models of openended evolution should be moved away from self-reproduction per se, towards issues on environmental interaction. This shift of attention will increase the potential for evolving a more varied ecology. Work on gene expression was presented, e.g., by Torsten Reil who had designed an artificial genome with biologically plausible properties (i.e., it was based on template matching in a nucleotide-like sequence). Especially now that more and more genome decoding projects come to an end - e.g., the complete genome of the nematode C. elegans has been sequenced the demand for techniques to study gene expression (i.e., how the genotype is translated in the phenotype) becomes more pressing. Notice that even a complete description of an animal's genome does not explain how the genotype is mapped to the phenotype!
algorithms are, for instance, internet routing, task allocation, and sorting. Victor Jongeneel spoke on the difficulties in genomics concerning the analysis of the results of the many genome sequencing projects that finish in the near future. After the quest for the genome, it will be time for the second question: how is a genome translated into a living system? See also below (session gene expression). In the proceedings (published by Springer as vol. 1674 of the series ‘lecture notes in computer science’) the contributions were classified in the following categories: epistemology; evolutionary dynamics; evolutionary cybernetics; bio-inspired robotics & autonomous agents; selfreplication, self-maintenance & gene expression; societies & collective behaviour; and communication & language. From the wide variety of subjects, one notices the strong multidisciplinary disposition of the young field (artificial life, since 1987). Participants in the conference came from various disciplines, such as theoretical biology, ethology, social sciences, linguistics, robotics, cognitive science, philosophy and computer science. Some people investigate ‘life as it is’, using computer simulations or robots to model existing biological systems; others are interested in `life as it could be' and study, e.g., the mechanisms of open-ended evolution that effect the complexity and variety of life-forms in our current world. RHINOCEROS AND PARASITES
Much of the work using the latter approach (i.e., studying `life as it could be') was presented in the sessions on evolutionary dynamics and the session on selfreplication, self-maintenance, and gene expression. For example, Richard Watson studied how lifetime interaction of animals parasites with other species (in particular symbiosis) affects the evolution of the BNVKI newsletter
BIO – INSPIRED AGENTS
146
October 1999
to another island, the agents drowned. Still, the bodies of drowned agents acted as a means for the other agents to cross the canals. Therefore, the willingness to walk into a canal indicated the amount of altruism present. It appeared that the placement of offspring on the lattice (i.e., close to their parents or not) crucially determined the emergence of altruism. Another poster was presented by Tom Quick who studied the topic ‘embodiment’. Rather than the traditional use of the term by roboticists (embodied = ‘possessing a physical body’) he widened the scope by focussing on the relationship between an agent and its environment in general. Now, he argued, "it is possible to analyse the significance of physical qualities without grounding the analysis itself in a material ontology."
The sessions on evolutionary cybernetics and bio-inspired robotics & autonomous agents contained work studying (the evolution of) sensori-motor control in biological and artificial systems. Just to mention a few examples: Joseba Urzelai presented work on evolutionary robotics with the emphasis on compact encoding of the genotype; Auke Jan Ijspeert showed how he had evolved swimming and walking controllers (central pattern generators) for simulated lampreys and salamanders; myself, I gave a talk on the evolution of cricket calling song divergence in real and robot crickets; Yasuhisa Hasegawa presented adaptation in a brachiating robot (that swings from branch to branch, similar to monkeys); Tony Belpaeme's poster was on the evolution of visual feature detectors, partly based on natural visual systems; and Andrew Adamatzky proposed the use of chemical controllers (e.g. BelousovZhabotinsky chemical processors) for various robotic tasks.
CHRIS LANGTON
Later that evening, during the conference dinner, Chris Langton, the founding father of the field, gave a talk. He had coined the field in 1987 based on the idea that simulating open-ended evolution of artificial organisms would lead to an artificial ecology that was as diverse and complex as the world we live in currently. An evolved artificial ecology (‘life as it could be’) is not necessarily carbon-based. Still, Langton claims, we can deduce general principles on complex systems by comparing ‘life as it could be’ to ‘life as it is’.
DEMOS
Thursday afternoon featured a series of demonstrations and poster presentations. One of the demos - that was awarded the young researchers award later that evening - was from Maciej Komosinski et al. The project, called Framsticks, was similar to Karl Sims' platform for artificially evolved creatures (well-known from the VPRO video on artificial life). The robot doll Robota by Aude Billard also attracted a lot of attention. It was built to study manmachine communication and learning by imitation by letting children play with it. Poster presentations included the work of Paul den Dulk and Martijn Brinkers who studied the evolution of non-reciprocal altruism in a population of simulated agents. The agents inhabited a lattice of islands separated by canals. Each island contained a limited amount of food but when crossing a canal, in order to migrate BNVKI newsletter
SOCIETIES AND COMMUNICATION
So far, one branch of Alife research has been left undiscussed in this report: the work on societies & collective behaviour and communication & language is not only concerned with behaviour of single systems, as seen in previous sessions, but focusses especially on distributed problem solving, cultural transmission, and the emergence and role of explicit 147
October 1999
to investigate how self-organisation developed sets of vowels in changing populations, as to draw conclusions on how similarities in unrelated human languages (called universal tendencies) came to exist.
communication between members of both natural and artificial societies. Examples include Charlotte Hemelrijk's studies of dominance in primate (ape) species. She proposed an explanation on why sometimes - in tight groups - female primates dominate males whereas usually the situation is the other way round. The explanation was more parsimonious than previous explanations and involved computer simulations of ape societies of varying density. In the simulations, two very simple interaction rules (grouping and acting dominantly) explained why, in cohesive groups, male and female dominance showed more overlap than in looser groups. Jason Noble investigated the handicap principle in relation to sexual signalling. The handicap principles ensures that male sexual signals of partner quality, such as an excessively long tail, are honest: a long tail, e.g., impairs movement and therefore really only `high-quality males' are able to carry one. He investigated three variants of the handicap principle using simulated evolution and showed that all variants enforce honesty, contradicting earlier studies. Edwin de Jong presented a poster on the evolution of communication, analysed from a dynamical system's perspective. In his multi-agent system, individual agents develop concepts describing possible situations in the shared environment. Then, by interacting, they couple concepts to ‘words’, which can be viewed as the workings of a dynamical system. Analysis of the phase space of the system revealed, e.g., how the association strengths between concepts and words developed in the population. The dynamical systems approach appeared to be very fruitful. As a last example I mention the presentation by Paul Vogt who also spoke about the evolution of language. But instead of the coupling of sound and meaning, the paper by Bart de Boer and him focussed on the emergence of speech sounds in a population of agents that imitate each other. Their key interest was BNVKI newsletter
The conference closed with a trip to Neuchatel to visit a few museums of early `artificial life' projects: mechanical automata from the 18th century. Especially `the scribe', a mechanical doll that could be programmed to write 6 line notes, was an amazing piece of machinery that was appreciated by everyone. EWLR-8, SEPTEMBER 18 HOW TO TEACH YOUR MOBILE ROBOT
The Saturday following ECAL was dedicated to the 8th European workshop on learning robots (EWLR-8), chaired by Jeremy Wyatt and John Demiris. The 11 talks of this one-day seminar could be roughly clustered into two groups: those that followed the reactive, behaviour-based approach and those that were concerned with deliberative path-planning and environment representation. The day started with an invited talk by Henrik Lund, whose robotics lab is partly funded by LEGO, the toy company. He is currently working on a very high-level, graphical programming language for children to programme LEGO mindstorms robots to play soccer. Following the behaviour-based approach, Lund and his team implemented a fixed set of behaviours that the children could use as building blocks for their controllers. Consecutively, the primary school pupils were allowed to evolve their controllers. For this they adopted hand-selection, as the use of a mathematical fitness criterium is obviously beyond the scope of young children. A remarkable demonstration was presented by Sevan Ficici and Richard Watson. On a table of approximately 1.5 by 1.5 meters, a group of small mobile 148
October 1999
workshop closed almost 1.5 hours late, but I had the impression nobody really minded staying somewhat longer.
robots was learning a light-searching task. They started with random controllers and received an amount of energy when they came in the vicinity of the light source. Meanwhile, they emitted parts of the genetic code representing their neural network controller through a short-range infra-red module. The genetic code was subject to random mutations and the robots emitted more frequently as they received more energy from the light-source. Samewise, after receiving a piece of code from another robot, robots were more prone to introduce the piece into their own controller with decreasing energy levels. This method for embodied (on-line) evolution, i.e., using real robots, reduces the time needed to teach the task when compared to other on-line methods, due to the distributed, a-synchronous evolutionary process.
NOUVELLE AI IN THE LOW COUNTRIES
Looking back, nouvelle AI, as mentioned in the beginning of this report, is mainly a matter of `foreign affairs'. That is, both at the PhD course and the workshop on learning robots, I was the only participant from the Benelux. Still, especially at ECAL-99, I was surprised by the number of Belgian and Dutch researchers either working in their own countries or abroad. I even spoke to a PhD student from London who referred to us as `the Dutch speaking clan'. Perhaps the largest centres of nouvelle AI can be found in the AI lab of the VUB and at the French speaking ULB. From the Netherlands, there were delegates from Amsterdam (UvA), Utrecht and Maastricht.
The second invited talk was given by Jun Tani who gave an overview of his work since 1991. His approach to robot control and learning is to view the robot and its environment as one single dynamical system. As an example of the bottom-up approach to AI, the recurrent neural network controller in his mobile robot created a rational attractor map of its environment through explorative behaviour. An example from the second category, mentioned above, is the work presented by Nobuhiro Inuzuka et al. They applied inductive logic programming to develop action rules for a mobile robot. The robot was taught by an erroneous teacher (as is the case in real life very often too) which made it necessary for the robot to find many examples (preconditions of action rules) in order to induce the correct action rules from the teacher's directions based on the examples.
THE FUTURE
The situated approach, that emphasises the importance of an intelligent system's interaction with its environment, is becoming a more and more established research theme. Soon, we shall start a robotics lab here in Maastricht to investigate adaptive behaviour in embodied systems and to study theories from psychology and biology. Also BNAIC-99, in Maastricht on the 3rd and 4th of November, will feature work on situated systems. The next large international conferences will be ALIFE VII in Portland, USA, 1-6 August and the international conference on the simulation of adaptive behaviour, SAB (better known by the name of its proceedings: From animals to animats), in Paris, summer 2000. For further reading I refer to the proceedings mentioned below in the selected reference list. On-line information can be found at www.cs.stir.ac .uk/EWNS2/ (for the euromorphic systems
Thanks to the limited number of participants (approximately 40), the interaction between speakers and audience was lively. Due to time delays, the BNVKI newsletter
149
October 1999
panel (and fellow students) asked, sometimes probing, questions, made suggestions for their work and gave feedback on their presentations. Getting through 24 presentations in one day is no mean feat but Conor managed it with grace and kindness but still kept the workshop to time. Conor also organised (and heavily participated in) informal discussions late into the night. In addition to the workshop day the students were also given the opportunity to present their work in special student sessions running in parallel with the main conference sessions. The formal aspect of the workshop concluded with the students and the panel voting for three prize winners.
workshop) or at diwww.epfl. ch/lami/ecal99/ (for the events in Lausanne). SELECTED REFERENCES D. Floreano, J-D Nicoud, and F. Mondada (eds.) Advances in Artificial Life, Lecture notes in computer science 1674, Springer, Berlin, 1999 R. Pfeifer, B. Blumberg, J-A. Meyer, and S.W. Wilson (eds.) From Animals to Animats 5 Proceedings of the Fifth International Conference on Simulation of Adaptive Behavior, MIT press, Cambridge, 1998 C. Adami, R.K. Belew, H. Kitano, and C.E. Taylor (eds.) Artificial Life VI, Proceedings of the Sixth International Conference on Artificial Life, MIT press, Cambridge, 1998 A.F. Murray and L.K. Hansen (eds.) International Journal of Neural Systems, special issue on neuromorphic systems, World Scientific Publishing Co, Singapore, in press.
GECCO-99 Student Workshop W.B. Langdon CWI, Amsterdam The Genetic and Evolutionary Computation Conference (GECCO-99) in Orlando, Florida, USA- this year was the biggest ever conference on evolutionary computation. It was a joint meeting of the long standing biannual International Conference on Genetic Algorithms (ICGA) which has been running since 1985 and the more recent annual Genetic Programming (GP) conference (which started in 1996). Before the start of each GP conference there has been a workshop for PhD students research in GP. This year Dr. O'Reilly (MIT AI lab) organized a joint graduate student workshop for research students principally concerned with any aspect of evolutionary computation, Evolvable Hardware, DNA computation, Artificial Life or Agents. As in previous years, the main part of the workshop took place on the day before the start of the main conference. In contrast to GP conferences, at GECCO'99 there were many other workshops on the same day. The workshop itself was chaired by Conor Ryan (University of Limerick, Ireland). He ran a tight, if informal, ship. This was a tough task with 48 students from all around the world and a similarly international panel of experts in Genetic Programming, Genetic Algorithms, DNA Computing, Artificial Life, Evolvable Hardware and from Industry. Approximately half the students gave brief presentations of their work to the panel and their fellow students. This was a golden opportunity for the students to put their work in front of some of the best names in their field. The BNVKI newsletter
[Best Research] Maarten Keijzer, Danish Hydraulic Institute, Denmark Title: “Scientific Discovery using Genetic Programming” Maarten received a copy of Advances in Genetic Programming 3 signed by the editors and joined the panel for dinner http://www.cwi.nl/ bill/gecco99/graduate.htmldinner. Maarten recently moved from the Netherlands where he worked for CAP to the Danish Hydraulic Institute, Horsholm.
150
October 1999
[Best Presentation] Anna Marino, University of Southampton, UK Title: “Sexual vs. Asexual Recombination for the Graph Colouring Problem with Hybrid Genetic Algorithms”
Verslag van de LANDELIJKE KION-DAG
Anna received a copy of Genetic Programming and Data Structures signed by the author, W.B. Langdon.
vrijdag 10 september 1999, Utrecht Ida Sprinkhuizen-Kuyper (UM) KION staat voor Kunstmatige Intelligentie Onderwijs in Nederland. Op de eerste landelijke KION dag waren bijna alle universitaire KIopleidingen vertegenwoordigd. De verschillende opleidingen verzorgden een korte presentatie over hun opleiding. Tijdens de presentaties en vooral in de wandelgangen (lunch, koffie- en theepauzes) was er gelegenheid om van gedachten te wisselen over KI-onderwijs en -onderzoek en hoe beide goed gecombineerd konden worden. De organisator / initiatiefnemer van deze dag was Albert Visser. Een dag als deze is zeker voor herhaling vatbaar. Tijdens de dag werd duidelijk dat de verschillende KI opleidingen dezelfde problemen tegenkomen. Het bespreken van deze problemen in landelijk verband kan sneller leiden tot het vinden van geschikte oplossingen. Aangezien naar de verhalen van de deelnemers te oordelen de onderwijsbelasting bij de KI opleidingen zeer hoog is, is een dergelijke gezamenlijke aanpak zeer welkom. Zo'n landelijke aanpak sluit goed aan bij de plannen voor een vervolg op het BOK project (zie de bijdrage van Jaap van den Herik en Evert van de Vrie over BOK) in de BNVKI newsletter Vol. 15, No.5, pp. 140-143), de Landelijke Onderwijsinnovatie Kennistechnologie (LOK).
[Best Questioner] Mark Voss, Markette University, USA Mark received a copy of the Proceedings of EuroGP, 1999 donated by EvoNET. Details of the workshop and links to many of the students can be found on the Internet at http://www.ai.mit.edu/people/unamay/geccophd/gecco-student-index.html Details of next year's conference can be found at http://www.genetic-algorithm.org/ Many of the students were generously supported by the American Association for Artificial Intelligence (AAAI) travel bursarys.
BNVKI newsletter
Uit de presentaties bleek dat op veel universiteiten gewerkt wordt aan robosoccer, zowel in het onderwijs als in het onderzoek. Robosoccer is een applicatie waarin veel soorten KIonderzoek getest kunnen worden. Bovendien spreekt het soccer-spel de studenten zeer aan, vooral als er met echte robots gewerkt kan worden. Een laatste punt dat ter sprake kwam betreft de bekendheid van de verschillende opleidingen. Aangezien er onder de deelnemers nogal wat onduidelijkheid was over het bestaan en de aard van de 151
October 1999
Nederlandse KI opleidingen is er besloten om een overzichtslijst te maken van alle KI-opleidingen in Nederland. Binnenkort zal deze lijst beschikbaar zijn via de WWW-pagina van de BNVKI (http://www.cs.unimaas.nl/~bnvki/).
BNVKI newsletter
152
October 1999
Ten eerste vroeg hij zich af of het mogelijk is een methodologische beschrijving te geven van het proces van het ontwikkelen van een juridisch kennissysteem. Ten tweede onderzocht hij mogelijkheden om programmatuur en instructies voor ontwikkelaars te maken die de bouw van een juridisch kennissysteem aan de hand van de methodologie konden ondersteunen. Ten derde wilde hij weten of een volgens de methodologie geproduceerd juridisch kennissysteem in de praktijk gebruikt kon worden. Het aan de Universiteit Utrecht uitgevoerde onderzoek is gestart in 1986. Het resulterende proefschrift werd op 10 maart 1999 succesvol verdedigd. Promotores waren prof. dr. A.W. Koers en prof.dr. H.J. van den Herik.
Section Knowledge Systems in Law and Computer Science Section-Editor Radboud Winkels DE BOUW VAN JURIDISCHE KENNISSYSTEMEN
M.C.M Weusten Proefschrift Universiteit Utrecht Deventer: Kluwer (1999) Bart Verheij
Dr. B. Verheij is als onderzoeker verbonden aan de capaciteitsgroep Metajuridica van de Universiteit Maastricht
[email protected], http://www.metajur.unimaas.nl/~bart/
OVERZICHT BOEK
Het proefschrift begint met een explicitering van uitgangspunten en doelstelling. De doelstelling die Weusten zich stelt wordt kracht bij gezet door het motto van het boek: the proof of the pudding is in the eating. Het gaat Weusten niet alleen om het bouwen van een systeem, hij wil bovendien dat het succesvol in de praktijk kan worden ingezet. In het tweede hoofdstuk worden methoden besproken om juridische kennis zo weer te geven dat die automatisch verwerkt kan worden. Het derde en vierde hoofdstuk betreft de bespreking van KRT, een methodologie voor het bouwen van juridische kennissystemen. In hoofdstuk vijf en zes worden twee juridische kennissystemen besproken, namelijk WVP en OVB. Het boek eindigt met een evaluatie en conclusies. Hieronder bespreek ik de inhoud van het proefschrift uitgebreider.
INLEIDING
Marnix Weusten heeft onderzoek gedaan naar de bouw van juridische kennissystemen. Een juridisch kennissysteem is een computertoepassing die - gebruik makend van in het systeem opgeslagen juridische informatie ondersteuning biedt bij juridische taken. Bij de bouw van juridische kennissystemen komt heel wat kijken: aan de ene kant is rechtswetenschappelijke kennis nodig, aan de andere kant informatietechnologische. Het samenvloeien van die twee zeer verschillende disciplines, de rechtswetenschap en de informatietechnologie, heeft geleid tot een interdisciplinair onderzoeksgebied dat wel als rechtsinformatica wordt aangeduid. Nederlandse onderzoekers zijn - verenigd in de stichting JURIX1 - internationaal actief op dit gebied.Weusten heeft zich in zijn onderzoek gericht op drie problemen. 1
UITGANGSPUNTEN
Zoals gezegd expliciteert Weusten in het eerste hoofdstuk zijn uitgangspunten ten
Zie de webpagina jurix.nl.
BNVKI newsletter
153
October 1999
oplossing van een casus geleid via het beantwoorden van een aantal vragen. Een andere aanpak is een 'goal driven' aanpak, waarbij geprobeerd wordt een beoogde conclusie te onderbouwen. Als laatste onderzoeksuitgangspunt noem ik nog de leus dat validatie tot acceptatie leidt. Hij spreekt van de professionele ongelovigheid van juristen, die leidt tot extra hoge kwaliteitseisen aan juridische kennissystemen. Weusten stelt zich ten doel instrumenten te ontwikkelen die de kwaliteit van rechtspleging kunnen verhogen. Hij denkt aan de mogelijkheid de beschikbaarheid van specialistische kennis te verhogen, aan de ondersteuning van personeel bij taken van hoger niveau dan hun kwalificaties, en aan publieksvoorlichting. Weusten wil een bijdrage leveren aan de rechtswetenschap en de rechtspraktijk, en niet aan de kunstmatige intelligentie en de cognitiewetenschap.
aanzien van het recht en van juristen. Ik noem er een aantal. Ten aanzien van het recht gaat Weusten ervan uit dat de kennisacquisitie ten behoeve van de bouw van juridische kennissystemen beperkt kan worden tot schriftelijke bronnen, namelijk wet- en regelgeving en jurisprudentie. Expertkennis zoals alleen domeindeskundigen die hebben en die niet op papier staat, kan naar zijn mening buiten beschouwing worden gelaten. Volgens Weusten kan een juridisch kennissysteem in de praktijk goed bruikbaar zijn, ook als het systeem geen domeinexpert is. Daarbij tekent hij aan dat juridische bronnen regelmatig veranderen, en dat het onderhouden van kennissystemen dus speciale aandacht nodig heeft. Vervolgens stelt hij dat rechtsregels niet universeel gelden of van toepassing zijn. Hij ziet dit vooral als een reden om bij de kennisacquisitie ten behoeve van de bouw van het systeem behalve op de explicitering van de regel zelf, ook op de toepassingsomstandigheden van de regel te letten. Weusten pleit ervoor dat de structuur van de juridische informatie de structuur van de representatie bepaalt. Dit heeft voordelen voor het onderhoud en voor de verificatie van een kennissysteem. Tot slot noem ik nog dat Weusten stelt dat het recht een zekerheidsfictie hanteert. Hij doelt hier op de omstandigheid dat een met rechtspraak belaste persoon een concrete beslissing moet nemen, ook als er sprake is van onzekerheid. Op grond van deze zekerheidsfictie kan de juridische informatie volgens Weusten in een kennissysteem gestructureerd worden als een beslissingsboom.Ten aanzien van juristen noemt Weusten onder andere de volgende uitgangspunten. Ten eerste geeft hij aan juridisch redeneren niet te willen nabootsen; het gaat hem om de juistheid van de conclusies die het systeem trekt. Vervolgens noemt hij de keuze een 'data driven' kennissysteem te bouwen. In zo'n systeem wordt de gebruiker naar de BNVKI newsletter
JURIDISCHE KENNISREPRESENTATIE
In het jargon wordt het weergeven van juridische informatie op zo'n manier dat die informatie automatisch verwerkt kan worden, juridische kennisrepresentatie genoemd. In hoofdstuk twee van zijn boek behandelt Weusten een aantal vormen van kennisrepresentatie, en bekijkt in hoeverre hij ze bruikbaar acht voor de bouw van juridische kennissystemen. Tot de vormen van kennisrepresentatie die door Weusten bruikbaar worden bevonden, behoren productieregels, beslissingsbomen en frames. Ik behandel ze hieronder kort. Productieregels hebben een 'ALS ... DAN ...'-structuur. Weusten geeft de volgende weergave van een van de regels over meeren minderjarigheid (art. 233 Boek 1 Burgerlijk Wetboek, oud): ALS de persoon in kwestie de leeftijd van achttien jaren heeft bereikt, OF ALS de persoon in kwestie is gehuwd, OF 154
October 1999
weer worden ingetrokken. In het recht komt dit veel voor, bijvoorbeeld door het bestaan van uitzonderingen op regels. Weusten kiest Reason-Based Logic, geïnitieerd door Hage en mede verder ontwikkeld door mijzelf, er als voorbeeld uit. Al acht Weusten Reason-Based Logic aantrekkelijk doordat specifieke vormen van juridische argumentatie, zoals het afwegen van redenen, er expliciet in kunnen worden weergegeven, vooralsnog vindt hij het niet geschikt voor juridische kennisrepresentatie. De reden die hij geeft is dat er volgens hem problemen kleven aan implementatie.
ALS de persoon in kwestie gehuwd is geweest, DAN is hij meerderjarig. Weusten noemt als voordelen van productieregels onder andere de uniformiteit van de resulterende weergave van juridische informatie, en de begrijpelijkheid van de resulterende informatie. Als nadelen noemt Weusten dat het gebruik van productieregels niet altijd efficiënt is en dat productieregels een weinig genuanceerde weergave van juridische informatie toestaan. Beslissingsbomen kunnen onder andere bestaan uit een als een boom gestructureerde reeks vragen en antwoorden. In een beslissingsboom kan het van het antwoord op een vraag afhangen wat de volgende vraag wordt. Zo kan bijvoorbeeld voorkomen worden dat een op grond van eerdere informatie overbodig geworden vraag opnieuw wordt gesteld. Zo zou bij de eerdergenoemde kwestie van meerder- of minderjarigheid de op zich relevante vraag of een persoon gehuwd is niet gesteld hoeven te worden als al vaststaat dat de persoon boven de achttien is. Frames kunnen als een soort invulschema's worden opgevat: ze bestaan uit een aantal velden die een waarde van een bepaald type kunnen aannemen. In een frame worden concepten, objecten of processen op een hiërarchische manier geordend naar abstractieniveau. Een frame dat een juridische beslissing weergeeft zou onder andere de beslissende instantie, de datum van de beslissing en de beslissing als velden kunnen hebben. Weusten acht frames weliswaar geschikt voor juridische kennisrepresentatie, maar geeft geen eigen analyse van de voors en tegens. Tot de vormen van kennisrepresentatie die door Weusten niet geschikt worden bevonden, hoort niet-monotoon redeneren, een onderwerp dat in het onderzoeksveld veel aandacht heeft gekregen. Bij nietmonotoon redeneren staat centraal dat aanvullende of gewijzigde informatie er toe kan leiden dat aanvankelijke conclusies BNVKI newsletter
KRT ALS METHODOLOGIE EN ALS GEREEDSCHAP
In de hoofdstukken drie en vier beschrijft Weusten KRT, een methodologie voor het ontwerpen van juridische kennissystemen. Na het kort noemen van andere ontwikkelmethodologieën behandelt Weusten de methodologie KRT in hoofdstuk drie op hoofdlijnen. In KRT worden drie hoofdfasen in de ontwikkeling van juridische kennissystemen onderscheiden. De eerste fase is die van definitie en analyse. In deze fase worden drie onderdelen onderscheiden, namelijk kennisacquisitie, kennisrepresentatie en validatie/verificatie. Deze drie onderdelen kunnen cyclisch doorlopen worden. Bij de kennisacquisitie worden bijvoorbeeld de grenzen van het domein en de gewenste mate van detaillering bepaald. Bij de kennisrepresentatie wordt een beslissingsboom opgesteld van het gekozen domein. Het domein wordt opgesplitst in 'beslispunten'. Afhankelijk van het antwoord op de bij een beslispunt horende vraag wordt naar de volgende vraag doorverwezen. Bij het onderdeel validatie/verificatie wordt gekeken of het systeem biedt waarvoor het bedoeld is en of de juridische informatie juist is weergegeven. Er vindt in dit onderdeel onder andere een automatische syntaxiscontrole plaats op de gerepresenteerde juridische informatie. Het resultaat van deze fase is onder andere een beslissingsboom voor het gekozen domein en een verzameling productieregels die deze boom weergeven. De tweede fase bestaat uit het programmeren. De in de eerste fase geproduceerde domeinweergave wordt automatisch omgezet in de 155
October 1999
programmeertaal Prolog. De resulterende code wordt gekoppeld aan een 'shell', dat wil zeggen een kennissysteem dat nog niet met domeininformatie is gevuld. Het resultaat is een voltooid kennissysteem. De derde fase in de KRTmethodologie is het onderhoud van het systeem. Omdat het recht voortdurend wijzigt is aandacht voor het onderhoud van een juridisch kennissysteem in de praktijk onontkoombaar. In hoofdstuk vier komt KRT als gereedschap aan de orde. KRT richt zich op twee soorten auteurs, namelijk aan de ene kant juridische experts voor het juridische vakgebied dat wordt gemodelleerd en aan de andere kant kennisingenieurs, die de juridische experts begeleiden. Weusten beschrijft onder andere de auteurshandleiding, de syntaxismodule en de beslissingsboommodule. De juridische expert maakt eerst een grafische weergave van de domeinanalyse in de vorm van een beslissingsboom en voert die vervolgens in het KRT-programma in als tabellen. Na onder andere syntaxiscontroles en enkele tussenrepresentaties genereert de beslissingsboommodule opnieuw een beslissingsboom. Bij wijze van verificatie kan deze automatisch gegenereerde boom door de juridisch expert met de handmatig gevonden grafische weergave van de boom worden vergeleken. Aan het eind van het hoofdstuk wordt een voorbeeld uitgewerkt.
een DOS-versie. Bij de ontwikkeling van WVP is een juridische expert van het ABP betrokken. Bij het doorlopen van de vragenreeks betreffende de verevening van pensioenrechten geeft WVP ook steeds relevante informatie, zoals relevante wetteksten en jurisprudentie. Het systeem OVB betreft de overdrachtsbelasting, dat wil zeggen de belasting op de juridische verkrijging van onder andere onroerende goederen, zoals die is vastgelegd in de Wet op belastingen van rechtsverkeer. Met behulp van het systeem kan worden bepaald of overdrachtsbelasting verschuldigd is. Als dit zo is, kan ook worden bepaald hoeveel. Bij de ontwikkeling van het systeem zijn een fiscaal jurist en een wetgevingsjurist van het Ministerie van Financiën betrokken. Weusten noemt de ontwikkeling van OVB ambitieuzer dan die van WVP doordat er twee experts bij betrokken waren, het domein groter is en het systeem berekeningen maakt. Op een laat tijdstip werd besloten aan te sluiten bij de huidige internettechnologie. OVB werd dan ook in HTML en Javascript ontwikkeld. Dit had tot gevolg dat diverse van de KRT-gereedschappen niet gebruikt konden worden. Wel konden de voordelen van hypertext worden uitgebuit. Een gebruiker van OVB kan naar behoefte door de juridische informatie heen 'klikken'2
DE KENNISSYSTEMEN WVP EN OVB
In de hoofdstukken vijf en zes bespreekt Weusten twee kennissystemen. Het eerste heet WVP en betreft de Wet verevening pensioenrechten bij scheiding. Het tweede systeem, OVB, betreft de overdrachtsbelasting. Beide systemen zijn op de markt verkrijgbaar. Bij het proefschrift is een cd-rom met een demonstratieversie van beide programma's gevoegd. Het systeem WVP geeft informatie over de vraag in hoeverre pensioenrechten overgaan naar een nietpensioengerechtigde echtgenoot in geval van scheiding. Het systeem kwam in mei 1995 op de markt, tegelijk met de inwerkingtreding van de Wet verevening pensioenrechten bij scheiding. Er zijn regelmatig nieuwe versies verschenen, soms vanwege juridische veranderingen, soms voor technische verbeteringen. Aanvankelijk was bijvoorbeeld geen Windows-versie beschikbaar maar alleen BNVKI newsletter
CONCLUSIES VAN HET PROEFSCHRIFT
In het slothoofdstuk vat Weusten zijn bevindingen samen en geeft hij een evaluatie van het beschreven onderzoek. Weusten concludeert dat het mogelijk is om een methodologie voor het ontwikkelen van juridische kennissystemen te geven, dat zo'n methodologie ondersteund kan worden 2
Voor meer informatie over de systemen verwijs ik naar het proefschrift en mijn productbesprekingen in het tijdschrift R&EM. WVP is besproken in R&EM, nr.1 (1996), OVB in R&EM, nr.4 (1998) 156
October 1999
Weusten niet diep in op de verhouding van zijn juridische kennissystemen tot andere, die ook in de praktijk werden of worden gebruikt. Over de technische kant van de systemen en hun ontwikkeling heb ik enkele vragen. Weusten gebruikt diverse 'tussenrepresentaties' van de relevante juridische informatie die uiteindelijk leidt tot de programmacode. De ontwikkeling begint met een grafische weergave van een beslissingsboom door de juridische expert. Die voert die in de KRT programmatuur in in de vorm van tabellen. Via onder andere een automatisch gegenereerde representatie als grammatica en als productieregels en na diverse controles op syntaxis en consistentie wordt weer een beslissingsboom gegenereerd. Die kan vervolgens gebruikt worden voor verificatie en validatie. Als ik het goed begrijp zijn de diverse weergaven op hoofdlijnen equivalent. Het is mij niet voldoende duidelijk geworden waarom al die tussenrepresentaties nodig zijn. Kan niet volstaan worden met de grafische weergave van de beslissingsboom door de juridisch expert en één interne representatie waarop alle controles worden uitgevoerd? Dit lijkt een grote besparing op benodigde gereedschappen te kunnen opleveren. Gezien de mogelijkheden van moderne bureaucomputers kan in de toekomst wellicht gedacht worden aan een ontwikkelsysteem waarmee de juridische expert rechtstreeks op zijn scherm een beslissingsboom kan bouwen en testen. Dit zou het benodigde werk voor zowel de juridische expert als de kennisingenieurs wel eens flink kunnen vereenvoudigen en verminderen. De basisrepresentatie van de juridische informatie in de kennissystemen is de beslissingsboom. Dat is de weergave waar de juridische expert mee begint en waarvan na semi-automatische verwerking uiteindelijk de juridische houdbaarheid wordt onderzocht. De beslissingsboom is een bekende representatievorm, die al vaak
door gereedschappen zoals programmatuur en handleidingen, en dat zo ontwikkelde kennissystemen in de praktijk bruikbaar kunnen zijn. De methode van KRT en het op de markt zijn van de twee ermee ontwikkelde systemen tonen dit volgens hem aan. Ten aanzien van zijn uitgangspunten nuanceert Weusten enigszins: de keuze voor een 'data driven' aanpak (in plaats van daarnaast de mogelijkheid van een 'goal driven' aanpak) acht hij niet voor alle domeinen geschikt. EVALUATIE VAN HET PROEFSCHRIFT
Na de bespreking van de inhoud van het proefschrift volgt nu mijn evaluatie van het proefschrift en het erin beschreven onderzoek. Het proefschrift is helder en toegankelijk geschreven. De ervaring van de auteur bij het uitleggen van technische materie aan juristen heeft zijn weerslag gevonden in de stijl van het boek. Weusten presenteert zijn hoofdboodschap, dat in de praktijk bruikbare juridische kennissystemen gebouwd kunnen worden, dan ook succesvol en met verve. De belangrijkste resultaten van het onderzoek, namelijk de twee kennissystemen, zijn van goede kwaliteit. De demonstratieversies van de twee kennissystemen (beide gebruik makend van een webbrowser) ogen overzichtelijk en zijn eenvoudig in het gebruik. De gebruikersvriendelijkheid is groot en de juridische inhoud ziet er (voor een leek als ikzelf) degelijk uit. Ik had wel graag iets meer informatie 'achter de schermen' willen krijgen. Na de proefsessie bleef ik met de vraag zitten hoe het kennissysteem precies tot zijn conclusie was gekomen. Waarom er geen recht op uitbetaling van een vereveningsdeel bestond en hoe het aan overdrachtsbelasting verschuldigde bedrag was berekend werd niet uitgelegd. Ook had ik graag een overzicht gekregen van alle gegevens die ik had ingevoerd, al was het maar om te kunnen controleren of ik me had vergist. Jammer genoeg gaat BNVKI newsletter
157
October 1999
velen in het onderzoeksveld - in dit opzicht veelbelovend. De nadruk die Weusten legt op de verificatie en validatie van juridische kennissystemen is terecht. Juist het feit dat Weusten zich nadrukkelijk richt op praktisch bruikbare kennissystemen die aan professionele maatstaven voldoen onderscheidt zijn werk van dat van veel anderen. Hij heeft het hierbij vooral over de verificatie en validatie van de beslissingsboom zoals die ten grondslag ligt aan de kennissystemen. Uiteraard is dit zeer belangrijk; de beslissingsboom vormt als het ware het hart van de naar KRT gemodelleerde kennissystemen. Aan de recente versies van de door Weusten ontwikkelde kennissystemen zitten echter ook veel interessante interface- en hypertext-aspecten. Gekozen is voor kennissystemen die de tegenwoordig breed beschikbare webbrowsers als 'omhulsel' gebruiken. Het kennissysteem biedt door middel van hyperlinks toegang tot relevante informatie waar doorheen 'geklikt' kan worden. Ook het hyperlinkmodel (dat aangeeft welke informatie via hyperlinks wordt verbonden met andere informatie), en meer in het algemeen de vormgeving van de gebruikersinterface van de kennissystemen, lenen zich voor toetsing. Er zijn legio interessante vragen. Voldoet de manier waarop de informatie aangeboden wordt aan de wensen van de gebruiker? Werkt de interface prettig en gemakkelijk? Met recht wijst Weusten er op dat uit de beschikbaarheid van de kennissystemen op de markt al impliciet volgt dat ze nuttig en bruikbaar zijn. Het zou mij echter niet verbazen als een systematisch gebruikersonderzoek tot nuttige inzichten leidt. Het zou bijvoorbeeld voor de praktijk interessant kunnen zijn als een gebruiker niet alleen de huidige stand van de wetgeving kan zien, maar ook kan instellen dat hij de stand van de wetgeving op een eerdere datum te zien krijgt.
is toegepast, ook buiten het recht. Eén van de eigenschappen van het gebruik van beslissingsbomen in KRT is dat elke casus tot een antwoord leidt. Dit past bij éen van de onderzoeksuitgangspunten, namelijk dat in het recht een zekerheidsfictie wordt gehanteerd. Weusten bedoelt hiermee onder andere dat een rechter nu eenmaal in alle gevallen een beslissing moet nemen. Indien nodig moet hij de knoop doorhakken en een keuze maken. Ik betwijfel of dit uitgangspunt ook zonder meer moet gelden voor juridische kennissystemen. Het is heel wel mogelijk om een systeem te bouwen dat niet in alle gevallen een beslissing neemt, maar naar 'zijn beste weten' aangeeft waarom er geen of geen ondubbelzinnig antwoord is. In zekere zin is dit zelfs juist voor juridische kennissystemen te verwachten. In een dynamische wereld wordt het recht voortdurend geconfronteerd met nieuwe situaties en omstandigheden. Telkens moet - gebruik makend van het staande recht een afweging worden gemaakt tussen de voors en tegens van een juridisch oordeel. Nieuwe concrete situaties leiden zo telkens tot juridische noviteiten. Dit is precies de reden dat rechters altijd beslissingen moeten nemen. Zij vormen zo mede het recht en zorgen er voor dat het op de nieuwe omstandigheden wordt toegespitst. De rol van een juridisch kennissysteem is vooralsnog niet om beslissingen te nemen, maar om ze te ondersteunen. Het kan juist een uiterst nuttige functie van juridische kennissystemen zij om verschillende overwegingen en opvattingen over het geldend recht en de huidige casus zo goed mogelijk weer te geven. Mijn vermoeden is echter dat hiervoor beslissingsbomen een minder geschikte representatie vormen, juist door de zekerheidsfictie die er welhaast in verankerd ligt. De door Weusten kort besproken representatiemethodes die gebaseerd zijn op typische aspecten van juridische argumentatie (zoals het aanvoeren van tegenargumenten) lijken mij - en met mij BNVKI newsletter
158
October 1999
onwrikbare standpunten onmogelijk zijn. Het beeld is mooi. De rechtsinformaticus Hercules (die het overigens niet van zijn intelligentie moet hebben) neemt de taak op zich om de Augiasstal van de hard cases te reinigen. Het is communis opinio onder de rechtsgeleerden dat het modelleren van hard cases onmogelijk is. In hard cases is niets volkomen zeker. De strijd tussen Hercules en Karneades is begonnen.
SLOT
Weusten heeft degelijke en gebruikersvriendelijke juridische kennissystemen ontwikkeld. Hij heeft zich hierbij niet gericht op een theoretische bijdrage aan de kunstmatige intelligentie of de cognitiewetenschap. Zijn onderzoek onderscheidt zich door de praktische haalbaarheid en de professionele bruikbaarheid van de juridische kennissystemen op de eerste plaats te zetten. De juridische kennissystemen die hij, zijn motto 'the proof of the pudding is in the eating' indachtig, op de markt heeft gebracht, smaken naar meer. Professionals op het gebied van de Wet verevening pensioenrechten bij scheiding en de overdrachtsbelasting kunnen inmiddels nuttig gebruik maken van de kennissystemen WVP en OVB. En dan te bedenken dat nog veel meer juridische domeinen op vergelijkbare wijze toegankelijk kunnen worden gemaakt ...
PROBLEEMSTELLING
Waarom zijn sommige juridische geschillen moeilijk oplosbaar? Het antwoord op deze vraag is van belang voor rechtsinformatici omdat zij zich de taak hebben gesteld om het oplossen van juridische geschillen te modelleren. Ronald Leenes pakt de beantwoording van deze vraag aan door het stellen en voor zover mogelijk beantwoorden van een aantal deelvragen: 1. Wat zijn moeilijke gevallen (hard cases) in het recht? 2. Welke problemen levert dit op voor het modelleren van hard cases? 3. Welke rechtsinformatica-toepassingen zijn er in het licht van de antwoorden op de vragen 1. en 2 mogelijk?
HERCULES OF KARNEADES HARD CASES IN RECHT EN RECHTSINFORMATICA
Proefschrift van R.E. Leenes Twente University Press, Enschede, 1998 ISBN 903651245X
WAT ZIJN HARD CASES?
De vraag naar de eigenschappen van hard cases wordt door Ronald Leenes beantwoordt aan de hand van de opvattingen van een aantal rechtstheoretici (Hart, Dworkin, Raz en MacCormick). Een hard case is een geval waarin het onduidelijk is of een rechtsregel van toepassing is of waarin meerdere toepasselijke regels tot onverenigbare resultaten leiden en er geen conflictenregel is die een keuze mogelijk maakt. In het eerste geval kunnen rechtsbeginselen mogelijk een oplossing bieden.
Bespreking door Kees de Vey Mestdagh, Rijksuniversiteit Groningen INLEIDING
Het proefschrift van Ronald Leenes heeft niet alleen een intrigerende titel maar voor de rechtsinformaticus ook een prikkelende inhoud. Hercules is bekend als zinnebeeld van de overwinning van moed, kracht en doorzettingsvermogen op een schijnbaar onmogelijke tegenstander in de vorm van onoplosbare opgaven (hoewel het uiteindelijk slecht met hem afloopt). Karneades is onbekend en terecht, omdat hij het onwrikbare standpunt innam dat BNVKI newsletter
ZIJN HARD CASES MODELLEERBAAR?
159
October 1999
Ronald Leenes aangeeft is die van dialogische systemen. Het onderscheid tussen easy en hard cases is dan niet gebaseerd op het al dan niet ontbreken van kennis in een gesloten systeem, maar in het al dan niet beschikbaar zijn van tegenargumenten in een open systeem. Het enige bezwaar tegen deze benadering is gelegen in de schijnbare tegenstelling tussen een gesloten systeem dat alle mogelijke argumenten al bevat en een open systeem dat alle mogelijke argumenten kan opnemen. De sting zit hem in het ‘alle mogelijke argumenten’. In beide gevallen is het de vraag waar deze vandaan moeten komen en wie of wat bepaalt of ze geleverd zijn of niet. De oplossing die Ronald Leenes kiest is pragmatisch. Het gaat niet om alle mogelijke argumenten, maar om de daadwerkelijk geleverde argumenten. Het criterium dat het systeem derhalve moet hanteren is: ‘er worden geen nieuwe argumenten meer geleverd’.
Aan de hand van twee rechtsinformaticastudies (van Anne Gardner en Tom Gordon) laat Ronald Leenes zien wat pogingen om het onderscheid tussen easy en hard cases te modelleren opleveren. In beide onderzoeken worden modellen ontwikkeld die hard cases modelleren als gevallen waarin de kennis te kort schiet om een definitieve oplossing te kiezen. Ofwel omdat er uit het model geen oplossing voortvloeit, ofwel omdat er meerdere oplossingen uit voortvloeien waaruit geen keuze kan worden gemaakt. De conclusie moet zijn dat deze modellen hiermee een zinvol onderscheid tussen easy en hard cases mogelijk maken, maar niet de oplossing van hard cases beschrijven. De functie van het vinden van de ontbrekende (nieuwe) kennis die hiervoor nodig is geen onderdeel van deze modellen. Ronald Leenes voegt hier het volgende aan toe: ‘Een RI-systeem opereert binnen de grenzen van zijn kennis en is niet in staat op te merken dat een probleem buiten deze grenzen valt. Het systeem kan daardoor oplossingen genereren die onjuist zijn zonder dat de gebruiker daar zicht op heeft.’ Deze conclusie is discutabel. Het is zeker dat de meerderheid van de operationele juridische kennissystemen technisch gesloten is. In eenvoudige gevallen doordat er maar één oplossing wordt aangedragen en, wat ernstiger is, in moeilijke gevallen door het gebruik van bijvoorbeeld negation by failure. Het is echter eveneens zeker dat er technisch open juridische kennissystemen operationeel zijn, die de mogelijkheid van alternatieven (ook in eenvoudige gevallen), het ontbreken van gegevens, het tekortschieten van de regels en het niet beschikken over conflictenregels aangeven en evenals (ik zou zelfs durven beweren in tegenstelling tot) hun menselijke collegae zeggen ‘ik weet het niet (zeker)’. Dit neemt niet weg dat het punt van het niet kunnen vinden van de benodigde nieuwe kennis voor het oplossen van hard cases blijft staan. De oplossingsrichting die BNVKI newsletter
WELKE ZINVOLLE RECHTSINFORMATICATOEPASSINGEN ZIJN ER NOCHTANS? Aan de gekozen oplossingsrichting, die van de dialogische systemen geeft Ronald Leenes zowel een theoretische als praktische uitwerking. Het dialogisch karakter van de rechtsvinding wordt uitgebreid besproken en uitgewerkt aan de hand van de dagvaardingsprocedure en het bewijsrecht. Vervolgens wordt er een concrete casus, die van een civiel geschil, waarin de bewijslast voor de eigendom van een tenthuisje centraal staat, aan een dialectische analyse onderworpen. Hiermee wordt overduidelijk aangetoond dat de rechtsvinding op dialogische wijze kan worden beschreven en daarmee wordt de houdbaarheid van de stelling dat de rechtsvinding een dialogisch karakter heeft onderschreven. Het recht laat zich alleen vaststellen in concrete procedures met een dialogisch karakter. De vraag was echter of er in het licht van het veronderstelde onderscheid tussen easy en hard cases 160
October 1999
bereikt. Ten slotte wekt de negatieve conclusie ten aanzien van de mogelijkheid voor RI-toepassingen om zelfstandig nieuwe argumenten en gezichtspunten te genereren ook genoeg irritatie over het tekortschieten van de huidige modellen op om hard verder te zoeken naar toepassingen die wel in staat zijn om het menselijk vermogen tot het vormen en toepassen van argumenten en gezichtpunten te evenaren en zelfs te overtreffen. Het feit dat argumenten een bepaalde, welomschreven vorm hebben, dat de spelregels voor hun toepassing blijkbaar expliciteerbaar zijn en dat hun elementen verwijzen naar op zichzelf beschrijfbare entiteiten als feiten, regels en andere argumenten leidt onvermijdelijk tot de conclusie: Geef mij een goed gegevensmodel en ik geef jou alle mogelijke argumenten voor en tegen alle mogelijke conclusies. Als ik maar over voldoende geheugen en tijd beschik. En wie beschikt er nu eigenlijk over voldoende geheugen en rekentijd?
sprake kan zijn van zinvolle rechtsinformatica-toepassingen. Gezien de hiervoor getrokken conclusie is de vraag nu of het modelleren van concrete procedures met een dialogisch karakter mogelijk is. Het antwoord op het tweede deel van de vraag kan kort zijn: ‘ja, het is technisch mogelijk om procedures met een dialogisch karakter te modelleren’. Ronald Leenes laat dit ook uitgebreid en op heldere wijze zien aan de hand van de eerder genoemde theoretische en praktische uitwerkingen. Het probleem is echter gelegen in het eerste deel van de vraag: ‘is het mogelijk om concrete procedures te modelleren’. Het antwoord hierop is afhankelijk van de realiseerbaarheid van een model van alle mogelijke concrete argumenten. Is een dergelijk model niet realiseerbaar dan zijn hiermee de grenzen van de mogelijke rechtsinformatica-toepassingen duidelijk. Dialogische systemen die in staat zijn alle mogelijke argumenten te ontvangen en mee te wegen zonder deze argumenten zelf te kunnen genereren. Laten we eens zien wat Ronald Leenes hierover te zeggen heeft. In zijn eindconclusie blijft staan dat de huidige RI-toepassingen geen ruimte bieden voor nieuwe argumenten en nieuwe gezichtspunten. Deze argumenten en gezichtpunten worden vastgesteld in dialogen. Deze dialogen worden beheerst door spelregels. RI-toepassingen zijn weliswaar niet in staat om zelf nieuwe argumenten en gezichtspunten te genereren maar wel om toe te zien op de handhaving van de spelregels.
END OF SECTION KNOWLEDGE SYSTEMS IN LAW AND COMPUTER SCIENCE SECTION COMPUTATIONAL LINGUISTICS Antal van den Bosch
FIRST ANNOUNCEMENT AND CALL FOR PRE-REGISTRATION BNVKI TUTORIALS ON ARTIFICIAL INTELLIGENCE AND LANGUAGE PROCESSING
Tilburg, January 10, 2000 Organized by BNVKI, with support from SIKS and CLIN, and sponsored by CLIF
CONCLUSIES
Organization: Walter Daelemans & Guy De Pauw (CNTS, University of Antwerp) http://pcger33.uia.ac.be/bnvki
Het proefschrift van Ronald Leenes levert helder en interessant materiaal voor wie geïnteresseerd is in het snijvlak van rechtstheorie en rechtsinformatica en met name in (de modelleerbaarheid van) het onderscheid tussen easy en hard cases. Het levert ook veel en belangwekkend theoretisch materiaal voor het verder ontwikkelen van dialogische systemen, die de juridische dialoog begeleiden en waarborgen dat een rationele uitkomst wordt BNVKI newsletter
Artificial Intelligence and Natural Language Processing (Computational Linguistics) have the same roots, but have largely developed separately the last two decades, especially in Europe. The Belgian-Dutch Association for Artificial 161
October 1999
there is also a growing application of “symbolic” machine learning methods to problems in naturallanguage processing. This tutorial will review approaches in which learned knowledge is represented in the form of decision-trees or logical rules. We will review basic algorithms and some advanced techniques for tree and rule induction and review applications of these methods to various problems in natural-language processing such as text categorization, syntactic parsing, word-sense disambiguation, semantic analysis, information extraction, and anaphora resolution.
Intelligence (BNVKI) invites you to two tutorials which focus on the connections between both areas, and address two active areas of research in their intersection: machine learning of natural language and communication in multi-agent systems. These tutorials are open to researchers and PhD students in Artificial Intelligence and Computational Linguistics. VENUE
Tilburg University, Room YZ-1 Travel Instructions: http://cwis.kub.nl/~buro/english/about /how.htm
COMMUNICATION IN MULTI-AGENT SYSTEMS
REGISTRATION
Registration includes tutorial materials, coffee, lunch, and drinks. BNVKI and CLIF members: 50 fl Others (includes BNVKI membership for the year 2000): 100 fl Pre-registration is required before December 15, 1999. Payment at registration on January 10 in cash, Dutch currency. Please Register on-line via http://pcger33.uia.ac.be/bnvki or by sending email to Guy De Pauw (
[email protected]) stating affiliation and membership to BNVKI or CLIF if applicable.
Dr. Frank Dignum (University of Eindhoven) In recent years the interest in multi-agent systems (MAS) has grown tremendously. The applications of MAS ranges from digital libraries through cooperative engineering to electronic commerce. All these applications have one thing in common. The agents operating within these systems have to communicate. In this seminar we will discuss both theoretical backgrounds of agent communication as well as some practical applications. The theory will cover the semantics of messages, the use of ontologies and conversation policies. The practical aspects cover the implementation of agent communication in KQML and possible other implementations.
PROGRAMME
09.00 - 10.00: Registration & coffee 10.00 - 13.00: Symbolic Machine Learning for Natural Language Processing (Raymond Mooney) 13.00 - 14.00: Lunch 14.00 - 17.00: Communication in MultiAgent Systems (Frank Dignum)
Section -Editor Richard Starmans
17.00 - 17.30: drinks Coffee/tea breaks at 11.15-11.30 and 15.15-15.30
CALL FOR PARTICIPATION AND REGISTRATION
ABSTRACTS SYMBOLIC MACHINE LEARNING FOR NATURAL LANGUAGE PROCESSING
BASIC COURSE “System modeling and Knowledge modeling”
Prof. Raymond Mooney (University of Texas) Most current empirical (corpus-based) research in computional linguistics employs “statistical” methods which represent and reason with knowledge using explicit probabilities. However, BNVKI newsletter
29 November - 3 December 1999 Hotel Apeldoorn, Apeldoorn 162
October 1999
For SIKS-Ph.D. Students and others interested in both topics.
Prof. dr. Arie de Bruin (EUR), dr. Guust Schreiber (UVA) and Prof. Dr. Roel Wieringa (UT). Several lecturers and guest-speakers (both academic and nonacademic) will contribute.
INTRODUCTION
From November 29 till December 3, the School for Information and Knowledge Systems (SIKS) organizes two basic courses: "System modeling (B1) and Knowledge modeling (B2)". Both courses will be given in English and are part of the obligatory Basic Course Program for SIKS-Ph.D. Students. Although these courses are primarily intended for SIKSPh.D. Students, other participants are not excluded. However, their number of passes will be restricted and depends on the number of regular students taking the course.
COSTS
SIKS-Ph.D. students will not be billed. Other groups will be charged as follows: - Students of ASCI, GST, OzsL and IPA: f. 1050,- Other academic participants: f. 1500,- Non-academic participants: f.2000,A reduction is possible for those participants that follow only part of the program. REGISTRATION
Anyone who is interested in the courses, is kindly requested to contact Margreet van Soest for registration. You can reach her preferably by e-mail:
[email protected] Please inform her whether you are SIKS Ph.D. Student or not. After your registration you will receive the final program and other details aboutthe courses as soon as possible. For questions about the content of the course, please contact the coordinator of SIKS.
CONTENT
The final program is not completed yet. The following topics give a first impression of the content. SYSTEM MODELLING
Petri nets Data modeling with a Z-like language Process algebra Algebraic specifications (Formal) specification of systems Verification and validation Heterogeneity Interoperability Software/process metrics
SIKS BASIC COURSE PROGRAM
On behalf of some new students we make a few general remarks about the place of the courses in SIKS’s educational program. SIKS offers a four-year Ph.D.-program which intends to provide students with broad basic knowledge as well as specialized advanced training. In order to reach these aims, the school has developed a pyramid model , that comprises five stages: homogenization, the basic course program, advanced components, general research skills and independent research. Although it is assumed that students already have completed an undergraduate degree in computer science, information
KNOWLEDGE MODELLING
Knowledge level modeling Knowledge acquisition Architectures for KBS Methods for KBS development Formal specifications of knowledge models, Ontologies ORGANIZATION
BNVKI newsletter
163
October 1999
the role in the Basic Course Program of SIKS or the individual education and supervision plan, please contact the coordinator of SIKS.
science or comparable discipline, background knowledge sometimes needs slight improvement c.q. adjustment. The homogenization stage is especially intended to raise the knowledge of beginning SIKS-students to a common level and to improve the background knowledge. This can be achieved by independent study or by taking undergraduate courses at some department or faculty in the Netherlands. It should not exceed a period of 4 weeks in the first year of the appointment. The goal of the Basic Course Program, the second stage of the educational program, is to bring the student’s general knowledge of the field of Information and Knowledge Systems to an international level. The core of the program is a two-year cycle of basic courses, that covers important paradigms, topics and trends in the beforementioned field. Unlike the courses that can be taken in the homogenization stage, the basic courses are entirely developed by SIKSstaff members and they primarily focus on the Ph.D-students. Currently, the following components give shape to the program
Voor alle vragen en opmerkingen over SIKS kunt u (op maandag, woensdagmiddag en donderdag) terecht bij: Richard Starmans, Coördinator SIKS Postbus 80.089, 3508 TB UTRECHT tel. 030- 253 4083 / 1454 fax. 030- 251 3791
B1 System modeling
B2 Knowledge modeling B3 Databases B4 Combinatory methods B5 Intelligent systems B6 Interactive systems B7 Varia B8 Varia B7 and B8 do give staff and students some additional degrees of freedom. Usually SIKS organizes one week basic courses, comprising two components. The basic courses are an obligatory part of the education and supervision plan of each SIKS - Ph.D.student. They are supposed to take at least 6 out of 8 courses. If a student chooses this option, the “remaining” part (2 courses) can be compensated for by taking courses organized by another research school. For more questions about BNVKI newsletter
164
October 1999
FORMAT OF THE CONFERENCE AND SUBMISSION
CALL FOR PAPERS
The four-day conference will include 11 invited lectures and 11 contributed papers. Submission for contributed papers are encouraged from all those who are working in one or more of the broad areas above, whatever their professional fields. Potential contributors should send one copy of an extended abstract (not more than 3 pages) to: The Organizing Committee, LOFT4 International Centre for Economic Research Villa Gualino Viale Settimio Severo, 63 10133 Torino, ITALY. (Fax: +39.011.6600082, E-mail:
[email protected], URL: http://pages.inrete.it/icer)
LOFT 4 ICER, Torino (Italy), June 29 – July 2, 2000
Logic and the Foundations of the Theory of Games and Decisions AIMS OF THE CONFERENCE This is the fourth in a series of conferences on the applications of logical methods to foundational issues in the theory of (individual and interactive) decision-making. It will be held in Torino (Italy) from June 29 to July 2, 2000. The previous three conferences took place at CIRM (Marseille, France) in January 1994 and at ICER (Torino, Italy) in December 1996 and December 1998. LOFT4 has been planned in cooperation with the TARK community and it is hoped that it will attract researchers who have participated in past TARK conferences. The aim of the LOFT conferences is to promote exchange across different disciplines. The organizers express their preference for papers which bring together the work and problems of several fields, such as game and decision theory, logic, computer science and artificial intelligence, philosophy, cognitive psychology, mathematics and mind sciences. The complete programs of the last two LOFT conferences (LOFT 2 and LOFT 3) are available at: http://www.econ.uc davis.edu/~bonanno/loft3a.html :http://www.econ.ucdavis. edu/~bonanno/loft2.htm. Among the topics of particular relevance are: 1) Modal logic: multi-agent logic, temporal logic, dynamic logic, probabilistic and multi-valued logic, logic of belief revision, the logicalomniscience problem. 2) Game and decision theoretic applications of modal logic: epistemic foundations of decision theory, epistemic foundations of equilibrium concepts in games, reasoning and belief revision in extensive-form games, applications of complexity theory. 3) Learning and information-processing models: economic aspects of information processing, learning in game-theoretic contexts, inductive learning and inductive decision making.
BNVKI newsletter
Complete papers are also welcome, but they will not be considered if they are not accompanied by an extended abstract. If not obvious from the abstract, the authors should mention whether and how their work fits the list of topics suggested above. The deadline for submission is February 15, 2000, and authors will be notified of acceptance decisions by March 31, 2000. People presenting contributed papers will have their local expenses (accommodation and meals) covered by ICER, while they are expected to rely on other sources for travel expenses. Further details will be communicated following acceptance of the papers. Those who wish to participate in the conference without submitting a paper should express their interest to the Organizing Committee at ICER. PROCEEDINGS
Selected papers from the previous LOFT conferences were published (or are forthcoming) in special issues of Theory and Decision, Mathematical Social Sciences, in a volume by Kluwer Academic Press and in Games and Economic Behavior. It is the intention of the organizers to publish a selection of the papers presented at LOFT4 in one or more special issue of a suitable journal. Details will be given at the conference. Wiebe van der Hoek
[email protected] Computer Science Department Fax 31 30 251 3791 Utrecht University Phone 31 30 253 3599 The Netherlands http://www.cs.uu.nl/people/wiebe/
165
October 1999
Marie-Francine Moens, Katholieke Universiteit Leuven, Belgium.
CALL FOR PARTICIPATION
PRESENTATIONS OF DEMOS
14. 00 - 14.15 Welcome by Dr. Jaap Hage, chairman of JURIX 14. 15 - 17.00 Automation of the legal practice: Demonstrations A knowledge added Internet database for texts of environmental legislation D. Vancolen Cosmos for Enterprises P. Lardinois
THE JURIX CONFERENCE Leuven, december 9 – 10, 1999 The Dutch Foundation for Legal Knowledge Systems, JURIX, is a forum for Research on Artificial Intelligence in the Legal domain. Since 1988, JURIX has organised annual international conferences on current research in this field. On December 9th and 10th, 1999, the twelfth conference will be hosted, for the first time outside the Netherlands, by the research units InfoSoc and ICRI of the K.U. Leuven, Belgium. Research on legal knowledge systems addresses both the theory of legal reasoning and the development of practical decision-support systems. Some traditional JURIX themes are: Investigation in legal reasoning Representation of legal knowledge in intelligent systems (e.g., logics, ontologies) Legal decision support systems, legal drafting systems, legal knowledge discovery systems, legal tutorial systems Storage and retrieval of legal information Applications of machine learning to law.
Electronic Annual Belgian Case Law Report (R.A.J.B.i) J. Lepaffe, Ch.Lepaffe, C. Lepaffe EXIT L. Windmolders, G. Francois, T. Van Buggenhout Expert system for environmental permit law K. de Vey Mestdagh JURIDISK INTERNET: the most complete social-juridical databank on the WWW K. Moeremans, G. Bal. Intelligent Quantum (IQ) - Tort Valuation by Precedent Instead of 'Rules' R. Douglas, D. Toulson SOLON - A legislative drafting system for the Flemish government S. Debaene, R. Van Kuyck, B. Van Buggenhout
Conference Website: http://www.law.kuleuven.ac.be/icri/educati on/ JURIX99
Tools to effectively access the content of legal texts M.-F. Moens, C. Uyttendaele, M. Logghe K. Van De Kerckhove, J. Dumortier
December 9, 1999 December 10, 1999 TUTORIAL
10.00 -13.00
CONFERENCE PROGRAMME
Jon Bing, University of Norway, Oslo BNVKI newsletter
166
October 1999
09.30 – 09.40 Welcome by Dr. Jaap Hage, Chairman of JURIX
Mapping inference trees to document structure for text generation in refuge determinations
09.40 – 10.05 Opening of the conference by Mr. Frank Robben, Chairman
14.50 – 15.15 On the formal analysis of normative conflicts. A. Elbag, J. Breuker and P.W. Brouwer
A.Stranieri, J. Yearwood and C. Anjara
10.05 – 10.30 Invited lecture by Prof. Dr. Jon Bing, University of Oslo, Norway
15.15 – 15.40
Tea break 15.40 - 16.05 Managing legal precedents with Case Retrieval Nets M. Costa, O. Sousa, J. Neves
10.30 – 10.55 Logic, context and valid inference. Or: can there be a logic of law. B. Verhey 10.55 – 11.15
16.05 - 16.30 Rule consistency J. Hage
Coffee Break 11.15 – 11.40 On formalising burden of proof in legal argument H. Prakken
16.30 – 16.55 Closing of the conference: Invided lecture by Prof.Dr. J. Zeleznikow, La Trobe University, Australia
11.40 – 12.05 Computer-mediated collaborative learning of legal argumentation T.J.M. Bench-Capon and P.H. Leng
17.00 Reception 19.00 Conference Dinner
12.05 – 12.30 Knowledge criteria for the evaluation of legal beliefs L. Mommers, H.J. van den Herik
For participation, please send an email to:
[email protected]
12.30 – 12.55
DiaLaw: levels, dialog trees, convincing arguments A. Lodder 13.00 – 14.00
Lunch 14.00 – 14.25 Legislative technique as basis of a legislative drafting system S. Debaene, R. Van Kuyck and B. Van Buggenhout
CONFERENTIES SYMPOSIA WORKSHOPS
14.25 – 14.50 BNVKI newsletter
167
October 1999
Dr. Y.H. Tan EURIDIS, Erasmus Universiteit Rotterdam Postbus 1738, 3000 DR Rotterdam Tel.: (010) 4082255. E-mail:
[email protected]
Below, the reader finds a list of conferences and web sites or email addresses for further information. A more extensive list of conferences can be found in the Calender 1999, as published in AI Communication and in the SIGART Newsletter.
Dr. E.O. Postma Universiteit Maastricht, IKAT Postbus 616, 6200 MD Maastricht Tel.: (043) 3883493. E-mail:
[email protected]
November 1-2, 1999 SIKS course, Maastricht, The Netherlands Information: Richard Starmans, email:
[email protected]
Dr. R. Verbrugge Rijksuniversiteit Groningen, Cognitive Science and Engineering Grote Kruisstraat 2/1, 9712 TS Groningen. Tel.: (050) 3636334. E-mail:
[email protected]
November 3-4, 1999 BNAIC’99, The 11th Belgian Netherlands Artificial Intelligence Conference, Maastricht The Netherlands. Information:http://www.cs.unimaas.nl/~bnvki/bnaic99.htm
Dr. W. van der Hoek Universiteit Utrecht, Department of Computer Science, P.O. Box 80089, 3508 TB Utrecht Tel.: (030) 2533599. E-mail:
[email protected]
November 5, 1999 BENELEARN’99, Maastricht, The Netherlands. Information: Luc Dehaspe, email: Luc.Dehaspe@cs. kuleuven.ac.be
Dr. L. de Raedt Institut für Informatik, Albert-Ludwigs-Universität Am flughafen 17, D-79110 Freiburg, Germany Tel: (+49)761203 8005. Email:
[email protected]
November 5, 1999 BENELOG, The 11th Belgian Netherlands Logics Conference, Universiteit Maastricht, The Netherlands. Information: Sandro Etalle, email:
[email protected]
G. Beijer BOLESIAN BV, Steenovenweg 19, 5708 HN Helmond Tel.: (0492) 502525. E-mail:
[email protected]
November 5, 1999 Database Conference, Universiteit Maastricht, The Netherlands. Information: Arno Siebes, CWI, Amsterdam. Email:
[email protected].
Dr. W. Daelemans Katholieke Universiteit Brabant, Vakgroep TaalLiteratuurwetenschap, Postbus 90153, 5000 LE Tilburg. Tel.: (013) 4663070. E-mail:
[email protected]
November 15-16-17, 1999 Ismick ’99. Six International Symposium on the Management of Industrial and Corporate Knowledge, Erasmus University of Rotterdam Information: http://www.fbk.eur.nl/FBK/AI/call99.html
EDITORS BNVKI newsletter Dr. E.O. Postma (editor in chief) (See addresses Board Members)
November 22-23, 1999 GEOMED’99, Second International Conference on Geography and Medicine, Paris, France. Information: http://www.b3e.jussieu.fr/geomed99
Prof. dr. H.J. van den Herik Universiteit Maastricht, IKAT Postbus 616, 6200 MD Maastricht Tel.: (043) 3883485. E-mail:
[email protected]
December 9-10, 1999 JURIX’99, Twelft International Conference on Legal Knowledge-Based Systems, Leuven, Belgium. Information: http://www.law.kuleuven.ac.be/icri/education/ JURIX99/
Dr. C. Witteveen Technische Universiteit Delft, Department Informatica, Julianalaan 132, 2628 BL Delft Tel.: (015) 2782521. E-mail:
[email protected]
December 10-15, 1999 Workshop on Artificial Intelligence and Judicial Proof at the Second World Conference on New Trends in Criminal Investigation, Amsterdam. Information: http://www.eurocongres.com/criminallaw
Dr. R.G.F. Winkels Universiteit van Amsterdam, Rechtsinformatica Postbus 1030, 1000 BA Amsterdam Tel.: (020) 5253485. E-mail:
[email protected]
January 4-7, 2000 HICSS, Thirty-third Annual Hawai’i International Conference on Systems Sciences, Maui, HI. Information: http://www.hicss.hawaii.edu
Dr. S.-H. Nienhuys-Cheng Erasmus Universiteit Rotterdam, Informatica Postbus 1738, 3000 DR Rotterdam Tel.: (010) 4081345. E-mail:
[email protected]
Technische
January 9-12, 2000 IUI2000, International conference on Intelligent User Interfaces, New Orleans, Louisiana, USAInformation: http://www.iuiconf.org Ir. E.D. de Jong Vrije Universiteit Brussel, AI Lab Pleinlaan 2, B-1050 Brussel, Belgium Tel.: +32 (0)2 6293713. E-mail:
[email protected]
MAIL ADRESSES BOARD MEMBERS BNVKI Prof.dr. J. N. Kok Wiskunde en Natuurwetenschappen, Dept. of Computer Science Universiteit Leiden, Niels Bohrweg 1, 2333 CA Leiden Tel: (071) 5277057. E-mail:
[email protected]
BNVKI newsletter
en
Dr. A. van den Bosch
168
October 1999
Katholieke Universiteit Brabant, Taal- en Literatuurwetenschap, Postbus 90153, 5000 LE Tilburg Tel.: (013) 4360911. E-mail:
[email protected] Dr. R.J.C.M. Starmans Coordinator Research school SIKS, P.O. Box 80089, 3508 TB, Utrecht Tel.: (030) 2534083/1454. E-mail:
[email protected] Dr. B. de Boer Vrije Universiteit Brussel, AI Lab Pleinlaan 2, B-1050 Brussel, Belgium Tel.: +32 (0)2 6293703. E-mail:
[email protected] HOW TO SUBSCRIBE The BNVKI/AIABN Newsletter is a direct benefit of membership in the BNVKI/AIABN. Membership dues are Fl. 75,- or BF 1.400 for regular members; NLG 50,- or BF 900 for doctoral students (AIO's); and NLG 40,- or BF 700 for students. In addition, members will receive two issues of the European journal AI Communications. The newsletter appears bimonthly and contains information about conferences, research projects, job opportunities, funding opportunities, etc., provided enough information is supplied. Therefore, all members are encouraged to send news and items they consider worthwhile to the editorial office of the BNVKI/AIABN newsletter. Subscription is done by payment of the membership due to RABO-Bank no. 11.66.34.200 or Postbank no. 3102697 for the Netherlands, or Argenta Bank no. 979-9307518-82 for Belgium. In both cases, specify BNVKI/AIABN in Maastricht as the recipient, and please do not forget to mention your name and address. Sending of the BNVKI/AIABN newsletter will only commence after your payment has been received. If you wish to conclude your membership, please send a written notification to the editorial office before December 1 1999. COPY The editorial board welcomes product announcements, book reviews, product reviews, overviews of AI research in business, and interviews. Contributions stating controversial opinions or otherwise stimulating discussions, are higly encouraged. ADVERTISING It is possible to have your advertisement included in the BNVKI/AIABN Newsletter. For further information about pricing etc., please contact the editorial office. CHANGE OF ADDRESS The BNVKI/AIABN newsletter is sent from Maastricht. The BNVKI/AIABN board has decided that the BNVKI/AIABN membership administration takes place at the editorial office of the Newsletter. Therefore, please send address changes to: Editorial Office BNVKI/AIABN Newsletter Universiteit Maastricht, FdAW, Vakgroep Informatica, Postbus 616, 6200 MD Maastricht, Nederland, Tel: ++31-(0)43-3883477 E-mail:
[email protected]
BNVKI newsletter
169
October 1999
