A NEW TAKE-OFF FOR AI? Editor-in-Chief Most scientists would agree that AI has not achieved his goal of creating an intelligent machine. The original high hopes fueled by the extravagant claims of early AI researchers and by the impressive performances of the first computers, soon turned into a general feeling of disappointment. It became clear that there is more to intelligence than fast flawless computation. Nowadays, artificial intelligence still has many critics stating that intelligent machines cannot be constructed and that the search for artificial intelligence is deemed to fail. Nevertheless, artificial intelligence research is prospering. Apparently, the goal of artificial intelligence is still inspiring a large part of the scientific community. But what is this goal, anyway? In the winter 1998 special issue of Scientific American on Exploring Intelligence, Patrick Hayes and Kenneth Ford rethink the goal of artificial intelligence. According to Ford and Hayes the traditional goal of AI is “to create a machine that can successfully imitate human behavior”. This goal is directly reflected in the Turing test in which an intelligent machine can fool the judge by imitating a human being. Ford and Hayes argue that the traditional goal is too limited for present-day AI research and should be replaced by a more contemporary goal: “to provide a computational account of intelligence”. To support their argument, they draw an analogy between the efforts to build a flying machine with those of building an intelligent machine. Originally, the developers of artificial flying machines took birds as an example. The reader may recall the soundless blackand-white movies showing the desperate attempts of early artificial-flight researchers to fly in machines equipped with flapping wings. In their contribution, Ford and Hayes show that the quest for a bird-like machine hampered rather than helped the development of flying machines. In a similar vein, the quest for humanoid machines may hamper the development of intelligent machines. Therefore, the goal of AI has to be broadened to encompass mental abilities of both humans and non-humans. Ford and Hayes appear to say that by focussing on the intelligence (flying) rather than on the human behavior (flapping wings), artificial intelligence is ready to take off. The current goal of AI research is certainly much broader than creating a machine that can pass the Turing test. It might even be something like “providing a computational account of intelligence”. What bothers me is the parallel that Ford and Hayes draw between a flying machine and an intelligent machine. The point is that a bird knows what flying is and a human being knows what intelligence is. To a bird, an artificial flying machine is a kind of brute-force approach to flying. It can bring you anywhere in a short time, but it makes a lot of noise, it smells, and its flight pattern does not attract females. To a human being, an artificial intelligent machine, such as Deep Blue, is a brute-force approach to playing chess, it can bring you a victory, but it makes noise, and you can not intimidate it. Many AI researchers have reached the insight that intelligence is closely related to embodiment. The physical make-up of an intelligent being is closely related to its way of perceiving the world, interacting with the world and thinking about the world. As a result, human intelligence, bird intelligence, and cockroach intelligence are all very different. So, if your goal is to “provide a computational account of intelligence” in one of these (or any other) species, mimicking is allowed and even required. This newsletter contains (amongst others) reviews on the sessions held during the NAIC'98 at the CWI in Amsterdam. Bart de Boer was awarded with the best-paper award at the NAIC and as a result became a new editor for the newsletter. I would like to congratulate Bart with his prize and welcome him to our editorial board. In addition, I am happy to welcome Antal van den Bosch who is also strengthening our editorial board. Fotoos on page ?????are made by Eric Postma. De uitgave van de NVKI-Nieuwsbrief wordt in 1998 mede mogelijk gemaakt door de Stichting Informatica Onderzoek in Nederland (SION). Sponsoring AHOLD
NVKI-Nieuwsbrief
162
December 1998
TABLE OF CONTENTS (Editor-in-Chief).................................................................................................................................................162 Table of contents ....................................................................................................................................................1 NVKI-Board News (Joost Kok)........................................................................................................................... 1 The 10th NAIC in Amsterdam ..............................................................................................................................1 Session Invited Speaker (Joke Hellemons en Eric Postma)........................................................................ 1 Session Language and Linguistics 1(Ton Weijters) .......................................................................................1 Session Learning (Ida Sprinkhuizen-Kuyper).................................................................................................. Session Programming (Maurice Bruynooghe) ................................................................................................. Session Logics 1 (Luc de Raedt) ........................................................................................................................ Session Knowledge Engineering 1 (Cees Witteveen) ....................................................................................... Session Decision Networks 1 (Maarten van Someren)..................................................................................... Session Robotics (Frans Groen) ........................................................................................................................ Session Search (Cor Bioch) ................................................................................................................................ Session Language and Linguistics 2 (Edwin de Jong) ..................................................................................... Session Agent Technology 1 (Peter Braspenning)............................................................................................ Business Session on Electronic Commerce 1 (Han La Poutré) ....................................................................... Session Evolutionary Algorithms (Dirk Thierens)........................................................................................... Session Agent Technology 2 (J.-J. Meyer) ........................................................................................................ Business Session on Electronic Commerce 2 (Frank van Harmelen)............................................................. Session Decision Networks 2 (Joost Kok) ......................................................................................................... Session Multi-Agent Demonstrations 1 (Erica van de Stadt) .......................................................................... Session Electronic Commerce (Gert-Jan Beijer) ............................................................................................. Session Knowledge Engineering 2 (Pierre Yves Schobbens)........................................................................... Session Multi-Agent Demonstrations 2 (Robert van Liere) ............................................................................ Session Neural Networks (Eric Postma) ........................................................................................................... Session Logics 2 (Yao-Hua Tan)........................................................................................................................ Machine Learning of Phonotactics (Antal van den Bosch)............................................................................... 1 The Minimum Description Length Principle and Reasoning Under Uncertainty (Ronald de Wolf)................ Taalkundige Analyse van Zakelijke Conversaties (Hans Weigand) .................................................................... ?? (Jaap van den Herik).........................................................................................................................................1 BENELOG 1998 (Sandro Etalle) ............................................................................................................................ SIKS .......................................................................................................................................................................... ANTS’98: From Ant Colonies to Artificial Ants (Katja Verbeek) ....................................................................... Artificial Intelligence Research at CNTS (Walter Daelemans and Steven Gillis)............................................... Artificial Intelligence and Beyond (..)..................................................................................................................... AI Abroad: A Year at the University of Calgary (Niek Wijngaards) ................................................................1 Section Knowledge Systems in Law and Computer Science (Section-editor Radboud Winkels) ....................1 PROSA- een Computerprogramma als instructieomgeving (Raf van Kuyck and Stijn Debaene)................. 1 Rechtsinformatica en Hard Cases (Ronald van den Hoogen) .............................................................................. Power - Programma Ondersteuning Wet en Regelgeving (Arno Lodder) ........................................................1 Conferenties, Symposia, Workshops ................................................................................................................... 1 Email-adressen bestuursleden / Redactie NVKI-Nieuwsbrief / Hoe word ik lid? / Kopij / Oude nummers/Advertenties / Adreswijzigingen ............................................................................................... 1 Joost Kok Chairman
NVKI-BOARD NEWS
NVKI-Nieuwsbrief
163
December 1998
ASSEMBLY
The NAIC’98 was a succesful conference: 150 participants from Belgium and the Netherlands. One participant even had to take the nightbus from Oxford to be able to attend the event. It was a lively event, which was well organized by the Evolutionary Computing group of Han La Poutré at the CWI in Amsterdam. During the conference, many interesting papers and demo's were presented. Also, the conference dinner was a great succes. Especially, the Italian wine was very nice.
November18, 1998 Frank van Harmelen, Secretary 1. Agenda 2. Yearly Report 3. Financial Report 4. Selection of New Board 5. Merge with Belgian AI Society 6. NAIC Conferences The chairman opens the meeting at 13.30h. Yearly Report. The chairman mentions the following points as main activities of the board in the past year: - Acquisition of sponsorship by Bolesian (major sponsor) and AHOLD - Recruited Belgian Newsletter editor (Edwin de Jong) - Publicity activities aimed at new Belgian members (poster action, Web-site) - Invited Gert-Jan Beijer from Bolesian B.V. as informal board-member to replace Henk Venema who has left Bolesian. - Supervised organisation of NAIC'98 - Financial Report
At the general assembly meeting a new board has been elected. Besides myself, the members of the nes board are: Rineke Verbrugge, Wiebe van der Hoek, Gert-Jan Beijer, Walter Daelemans, Eric Postma and Yao-Hua Tan. Last year the board spent quite some time on the Belgian-Dutch integration. The result is reflected in the society’s new name: BelgischNederlandse Vereniging Voor Kunstmatige Intelligentie/ Association pour Intelligence Artificielle Belgique-Neerlandais (BNVKI/AIABN). Ida Sprinkhuizen-Kuyper, Henk Venema, Bernard Manderick and Frank van Harmelen have left the board. I would like to thank them for all the work they have been doing. Frank van Harmelen will assist the board (especially the new members) in the next year.
The expenses over this financial year were somewhat lower than expected, mainly due to underspending on subsidising workshops and tutorials. At the same time, the income was higher than expected, due to the acquisition of sponsorship. This has led to a much smaller negative balance in the budget than foreseen.
A discussion on a new format for the BNAIC was initiated in the last newsletter and also during the general assembly meeting at the NAIC in Amsterdam. Many useful remarks and suggestions were made by the participants. In the coming year the board of the BNVKI/AIABN will make an effort to further improve the format of the BNAIC and to recruit new members from AIrelated fields.
The accounts committee has inspected the books, has given some recommendations to the treasurer and was satisfied that the books were in order. By general acclamation the meeting accepts the recommendation of the accounts committee to discharge the treasurer of her duties. The treasurer presents the new budget to the meeting. A small negative balance is foreseen, but this is deemed acceptable in the light of the available reserves. The meeting accepts the proposed budget by general acclamation.
foto MINUTES OF THE NVKI GENERAL (Inserted after announcement of the chairman during the conference closing session). The chairman proposes that the new accounts committee consists of Catholijn Jonker, Walter Thoen en Edwin de Jong.
Berndard Manderick all step down from the board. Henk Venema has already resigned during the past year. The board proposes Gert-Jan Beijer (Bolesian BV), Rineke Verbrugge (Universiteit Groningen), Wiebe van der Hoek (Universiteit Utrecht) and Walter Daelemans (Universiteit Antwerp/KUB) as new members. Joost Kok is proposed as chair of the
Election of new board Frank van Harmelen, Ida Sprinkhuizen-Kuyper and
NVKI-Nieuwsbrief
164
December 1998
board. This proposal is accepted by the meeting by general acclamation.
buy the various publications on offer from the BNVKI/AIABN. The chairman closes the meeting at 14.45 h.
Merge with Belgian AI Society The board reports the new and succesful Dutch-Belgian collaboration in the past year, the increase in the number of Belgian members, and the significant Belgian participation in NAIC'98. The board proposes to go ahead with the proposed extension of the NVKI to include the Belgian as well as the Dutch AI community. Some discussion follows concerning the new name of the association. The board proposes BNVKI/AIABN, standing for: Belgisch-Nederlandse Vereniging voor Kunstmatige Intelligentie/Association pour Intelligence Artificielle Belgique-Neerlandais. In favour of this proposal is the continuity with the current name. Some discussion follows regarding the length of the acronym. After some discussion, the meeting accepts the proposal of the board with 12 in favour, 5 against and 4 abstentions.
SESSION INVITED SPEAKER
foto Luc Steels
NAIC'98 The NAIC'98 has succeeded in attracting 130 participants plus 20 student participants. There were 65 submissions of which 50 were accepted. Some 40 of the 65 were new publications not submitted elsewhere. The new business-track on E-commerce has attracted significant interest from commercial partners. The meeting is very grateful to Han la Poutré and Jaap van den Herik for their efforts in organising the NAIC'98. Future of the NAIC Some discussion follows concerning the future format of our annual conference. Some of the aims of the conference are mentioned: presentation of scientific results, a forum for young researchers, a meeting place for the Belgian-Dutch AI community, and a meeting place for academia and industry. Possible improvements to the current format are suggested: aiming for a special issue of a journal for a selected set of NAIC contributions, re-integration with the applied AI conference, the additional tutorials, and aiming for internationally published proceedings. No consensus was reached on the merits of the various proposals. The board will ensure that these and other proposals are taken into account when organising next year's conference. BNAIC'99 The 1999 conference will be called BNAIC'99, and will be organised in Maastricht. Any other business Jaap van den Herik encourages all those present to
NVKI-Nieuwsbrief
165
December 1998
SESSION LANGUAGE AND LINGUISTICS 1 Ton Weijters Eindhoven University of Technology
Emergence of sound systems through selforganization B. de Boer Vrije Universiteit Brussel
The first parallel session on Language and Linguistics was an elaboration of invited lecture by Luc Steels: Bootstrapping Cognition Through Language. This is not surprising because both Luc Steels and all speakers in this session are members of the Artificial Intelligence Laboratory of the Vrije Universiteit Brussel.
The third and last paper was nominated for two NAIC’98 awards: the best paper award and the best Ph.D. student-paper award (ultimately Bart appears the winner of the best-paper award). His paper describes a model for explaining the emergence and the universal structural tendencies of vowel systems. Both are considered as the result of self-organization in a population of language users. The language users try to imitate each other and to learn each other's vowel systems. The speaker showed through computer simulations that coherent and natural (similar to human) sound systems can indeed emerge in populations of artificial agents. The important parameters in these simulation were only human speech production characteristics, perception characteristics and noise. Really a very interesting a very readable paper that deserved to be awarded as the best paper award.
The evolution of a lexicon and meaning in robotic agents through self-organization P. Vogt Vrije Universiteit Brussel The first paper discusses interdisciplinary experiments, combining robotics and evolutionary computational linguistics. The goal of the experiments is to investigate if robotic agents can originate the lexicon of language (naming objects). The lexicon is propagated through social interactions of the individual agent with its environment including other agents, and a natural selection-like ontogenetic development between the agents. The Development of a Lexicon Based Behavior E. de Jong Vrije Universiteit Brussel
foto
The subject of the second presentation was strongly related to the subject of the preceding presentation. However, this paper investigates whether a group of agents may develop a common lexicon; relating words to situations (not objects). Each agent independently decides which situations are useful to distinguish, based on experience with the environment. The question is of whether a process of self-organization results in the development of a shared lexicon. The experimental results presented by the speaker showed that this question can be answered positively.
SESSION LEARNING Ida Sprinkhuizen-Kuyper Universiteit Leiden Relational Reinforcement Learning S. D_eroski, L. De Raedt, and H. Blokeel Katholieke Universiteit Leuven The goal of the first talk was to show that a combination of reinforcement learning and relational learning can handle new learning tasks by using a more expressive representation language. Planning in the blocks world served as an example.
It is not always clear for me whether the reported experiments are frustrated or stimulated by the use of real robots. Problems with sensing capabilities and radio link reliability seems to take a lot of research effort. On the other hand, the videos displayed during the presentations were entertaining and illustrative.
Goal-driven Learning for Knowledge Acquisition M. Van Someren Universiteit van Amsterdam
In the second talk Maarten van Someren gave an overview of an architecture to generate and select
NVKI-Nieuwsbrief
knowledge acquisition operators from existing knowledge systems and sources of knowledge (e.g.,
166
December 1998
human experts). It was quite a large system to describe in a twenty-minute talk, but Maarten was able to give an impression of a useful architecture for generating knowledge systems.
Katholieke Universiteit Leuven and D.A. de Waal, Potchefstroom University, South Africa Some programs explore an infinite search space (e.g. in planning) and run forever when the query is such that the problem has no solution. The absence of a solution can be proven by showing a model of the program in which the query is false. This work develops methods for searching such models and compares the methods with program-analysis methods which can, - as a byproduct - show failure of queries and with some model generation methods used in theorem proving.
Unsupervised Learning of Subcategorisation Information and its Application in a Parsing Subtask S. Buchholz Katholieke Universiteit Tilburg The last talk of this session was about subcategorization, i.e., a lexical property of (mainly) verbs. Sabine showed that unsupervised learning of subcategorization information can improve the complement-adjunct distinction task by 1%, which is 2/3 of the improvement obtained by using this information from tree-bank annotations, for which much more work is necessary. The accuracy improvement of 1% corresponds to an error reduction of 15%. The session on learning was an interesting opening session of the NAIC'98.
A Lazy Logic programming Language S. Etalle , University of Maastricht, and F. van Raamsdonk, CWI Lazyness - employing a call-by-need evaluation mechanism - is a major feature in many functional programming languages (e.g., Haskell). The notion does not fit standard logic programming where an answer is only returned when there are no more subgoals to resolve. This paper proposes a lazy logic language. It is obtained by adding two kind of annotations to standard logic programming: requests and strictness, and by replacing the notion of successful derivation by the notion of adequate derivation.
SESSION LOGIC PROGRAMMING Maurice Bruynooghe Katholieke Universiteit Leuven A Framework for Bottom-up Specialisation of Logic Programs W. Vanhoof, D. De Schreye, and B. Martens Katholieke Universiteit Leuven
SESSION LOGICS 1 Luc de Raedt Katholieke Universiteit Leuven SESSION KNOWLEDGE ENGINEERING 1 Cees Witteveeen Technische Universiteit Delft
Traditionally, partial deduction of logic programs --- a specialisation technique --- is performed top-down and specialises programs by exploiting part of the program input (query) which is already available at specialisation time. This work proposes a framework for bottom-up partial deduction. It is argued that such an approach is simpler and gives better results in situations where no partial input is available as a part of the query, but rather as a set of predicate definitions. For example, as the predicates defining an abstract data type, or as the object program to be manipulated by a meta-interpreter.
Three rather diverging contributions characterized the first session on knowledge engineering, thereby illustrating the potential broad scope of this field and showing the possible benefits particular approaches in AI can have for other areas both within and outside AI. Characterizing approximate problem-solving by partially fulfilled pre- and postconditions F. van Harmelen and A. Ten Tije Vrije Universiteit Amsterdam
Detecting Unsolvable Queries for Definite Logic Programs M. Bruynooghe, H. Vandecasteele, M. Denecker The first contribution was the presentation of Frank van Harmelen and Annette ten Tije. Their work is an intriguing attempt to apply approximation methods developed in AI to a rather well-established software engineering
NVKI-Nieuwsbrief
pre-and postconditions paradigm. Where traditionally the precondition has to be satisfied completely in order to evaluate the fulfillment of the postcondition, these authors proposed a more refined framework where also a relationship can
167
December 1998
be established between partially preconditions and postconditions.
fulfilled Donkers, Uiterwijk and Van den Herik presented a method for finding the best decisions in a Markov Decision Network. This is a representation for describing the relations between tests, actions and states. The idea is to represent a "groups" of states as one node with a description that abstracts from some of the state variables. The networks are "markovian". Donkers et al. adapt a method for finding the optimal action or test for a POMDP (Partially Observed Markov Decision Process) to linear Markov Decision Networks. In this context the term “Linear” means that the value of an episode is a linear function of costs and rewards of individual steps. This allows for an efficient method to find the optimal action or test. Their example was quite timely: optimising the control of dikes to prevent floods.
Version Space Retraction with Integrated Instance/Concept-Based Boundary Sets E. Smirnov and P. Braspenning Universiteit Maastricht The second talk was presented by the first author and was based on their award-winning paper at the ECAI'98. In the paper they show that revision methods can be incorporated in the classical version space concept-identification framework in both an effective and efficient way. The importance of such an approach cannot be overlooked easily: it opens a possible way to merge the hitherto largely isolated (AI) paradigms of theory revision and inductive learning.
Decision Trees: Equivalence and Propositional Operations H. Zantema Universiteit Utrecht
Specification of Dynamics for Knowledge-Based Systems P. van Eck, J. Engelfriet, D. Fensel, F. van Harmelen, Y. Venema and M. Willems Vrije Universiteit Amsterdam
Hans Zantema presented results on combining decision trees into new decision trees. These trees are a notation for boolean functions and this raises (by analogy) questions about detecting equivalence between trees and combining trees efficiently into new trees. Zantema gives an algorithm for the first point. This turns out to be NP-complete as for arbitrary boolean functions but the complexity is approximately in the order of the product of sizes of the trees. He then shows that a tree, that represents the conjunction or disjunction of two given trees, cannot be represented efficiently (much more compact) than by simply adding trees. These results are useful for the design of algorithms for refinement or revision of decision trees.
The last talk was a final illustration of the usefulness of, not only merging the ideas of several authors, but also thereby comparing different knowledge specification formalisms. In this case the problem was to specify the dynamic reasoning behaviour of a knowledge-based system. The method followed was a detailed comparison of the resulting formalizations of a specific example. The comparison was made on two dimensions: the first one dealing with the kind of concepts used to analyze the example, the second one dealing with the way in which these concepts were represented. SESSION DECISION NETWORKS 1 Maarten van Someren Universiteit van Amsterdam
An Algorithm for Generating Quasi-Monotone Decision Trees for Ordial Classification Problems R. Pothartst and J.C. Bioch Erasmus Universiteit Rotterdam
Solving Markov Decision Networks using Incremental Pruning H.H.L.M. Donkers, J.W.H.M. Uiterwijk, and H.J. van den Herik Universiteit Maastricht Potharst and Bioch presented a method for learning quasi-monotonic ordinal decision trees. In some problems variables have ordered values and one can assume that the values have a monotonic relation with the criterion: at some point on the ordered scale the criterion flips but then it will not flip back. The method that the decision tree learner C4.5 uses cannot be forced to construct decision trees that satisfy this
NVKI-Nieuwsbrief
constraint. Potharst and Bioch give a method that enforces this constraint during learning. For domains that actually have this "quasi-monotonic" ordered structure, this gives better results than C4.5.
SESSION ROBOTICS
168
December 1998
Frans Groen Universiteit van Amsterdam
The last paper discussed the tracking of objects using an active camera. The presentation gave a very nice overview of the different approaches to visually track moving objects. In active vision the direction in which the camera looks is dynamically adapted, so that the moving object stays in the focus of the camera. Active vision is based on reactive dynamic closed-loop control and only fast algorithms can be used for that. Therefore, the method applied in the described experiment is a very simple and fast one. It calculates the difference between two successive images to find the moving objects of which the centre of gravity is calculated. That is used to control the pan and tilt motor of the camera. Results showed that a coke bottle on a string could be followed with 20 frames a second.
Planning Strategies for Decision-Theoretic Robotic Surveillance N.A. Massios and F. Voorbraak Universiteit van Amsterdam The first paper in this session discussed a decisiontheoretic approach to planning strategies for robotic surveillance. When a robot is employed in a surveillance task, it has to travel the environment to detect relevant events, such as a fire, using its sensor that has a certain range. Based on a formal model of the environment, different surveillance strategies are presented and illustrated by examples. The strategies were also implemented in a simulator, of which results were shown. The conclusion is that the minimum expected cost policy behaves well in situations where the probabilities and costs matter and early detection is important.
SESSION SEARCH Cor Bioch Erasmus Universiteit Rotterdam Solving Job Shop problems with Critical Block Neighbourhood Search P. van Bael and M. Rijkaert Katholieke Universiteit Leuven
AIACS: A Robotic Soccer Team Using the Priority/Confidence Model J. Lubbers, R.R. Spaans, E.P.M. Corten, and F.C.A. Groen Universiteit van Amsterdam
In this lecture Patrick described a neighbourhood search scheduling algorithm to solve the Job-Shop Scheduling well-known in Operations Research. The general problem is NP-hard and therefore many different heuristics/algorithms has been developed in the literature each with its own advantages and disadvantages. The most popular algorithms are iterative and use local search. The most important part of these algorithms is the way neighbourhood solutions are created and how to move to one of these solutions. The main contribution of the authors is the use of a neighbourhood structure based on socalled critical blocks and simulated annealing combined with a new iterative improvement algorithm. This algorithm is tested on the famous 10 * 10 problem of Mute and Johnson. The experiments show that the iterative algorithm gives very good results and that the neighbourhood structure performs better than the older one. Patrick also discussed the importance of the algorithm for industrial applications.
The second paper discussed the architecture of the robot soccer teams of the UvA, which participated in the Robocup'98 world championships in Paris. It is a 3-layer structure consisting of the basic behavior layer, the skilled behavior layer and an action manager. The two UvA teams differ mainly in the action manager. The AIACS team has an action manager based on a priority/confidence model. Decisions are made according to a confidence measure, which is based on the importance of action and the satisfaction of preconditions. The AIACS team became ninth in the Robocup championship. The AIAACS team is compared with the other UvA team: the Windmill Wanderers, which has an action layer based on a decision tree, and reached the third place in the Robocup championship. Tracking Objects Using an Active Camera T. Belpaeme Vrije Universiteit Brussel Simulated Annealing with estimated temperature: a new efficient temperature schedule based on the notion of acceptance E. Poupaert and Y. Deville Université Catholique Louvain
NVKI-Nieuwsbrief
Simulated annealing is a well-known local search optimisation algorithm. This algorithm proposed by Metropolis simulates the behaviour of a system at a given temperature. The neighbour of the current
169
December 1998
solution is generated randomly at each iteration. If it is better than it is accepted as the new current solution, otherwise it is accepted with a probabiltiy depending on the energy difference between the two solutions and the temperature. The most important part of the simulated annealing algorithm is the temperature schedule used to decrease the temperature. The aim of this schedule is to reach thermal equilibrium. A good schedule must be efficient, general and robust. In his lecture Eric Poupaert discussed several annealing algoritms and proposed a new algorithm based on the idea on maintaining an evolving target acceptation probability throughout the optimisation process. To compare this algorithm with the classical ones the authors have implemented a platform for general local search optimisation problems. They compared their algorithm with three classical ones. The algorithms were tested on two classes of problems: the geographic travelling salesman problem and real function optimisation. The new algorithm outperforms the classical algorithms and compares well with the best algorithm known of date (Saef). So the idea of an evolving target acception during optimisation is very promising. In the discussion generalisations to other local search algorithms were mentioned. For example tuning of natural selection pressure in genetic algorithms.
Antal van de Bosch presented this paper which is to appear in the prestigious Machine Learning Journal. Several thorough experiments have been performed, all of which support the claim stated in the informative title. Instead of classical treebased methods, memory-based learning methods were applied using Tilburg's own publicly available TiMBL package. The experiments included grapheme-to-phoneme conversion (GS), part of speech tagging (POS), base noun phrase chunking (NP), and prepositional-phrase attachment (PP). As an example of POS, the word 'man' in 'the old man the boats' should be identified as a verb phrase in this context, whereas probably in most contexts it functions as a noun phrase. In NP, the difference in function of the latter parts of 'she ate pizza with a fork' and 'she ate pizza with anchovis' should be recognized. A first question that was investigated concerned 'editing', the practice of removing outliers to improve generalization, which is common in supervised learning. This involved a measure of typicality (from Zhang) and an indicator for class prediction strength (from Salzberg). The conclusion was that editing is always harmful. Moreover, and this is the main thesis of the paper, it turned out that removing exceptional instances (instances with low typicality or prediction strength) is much more harmful than removing typical instances.
SESSION LANGUAGE AND LINGUISTICS 2 Edwin de Jong Vrije Universiteit Brussel Lexical Cohesion and Authorship Attribution H. Paijmans Katholieke Universiteit Tilburg
Furthermore, the benefits of memory based learning in the domain of language learning were discussed. IB1-IG, a memory-based method based on the information gain criterion, was found to perform better than eager learning. Disjunctiveness, a characteristic property of language data, explains the importance of exceptional instances.
Unfortunately this paper was not presented. In the paper, text-cohesion methods are applied to authorship attribution, i.e. determining the author of a text. Even though this was not a primary goal in Paijmans' research, it yielded interesting results, and sheds new light on the question of who wrote the 'Federalist papers', a historical collection of writings.
Argue! - An Implemented System for Computer-Mediated Defeasible Argumentation B. Verheij Universiteit Maastricht
Forgetting Exceptions is Harmful in Language Learning W. Daelemans, A. Van den Bosch, and J. Zavrel Katholieke Universiteit Brabant This paper was presented by Bart Verheij. Defeasible argumentation consists of inference (drawing conclusions from premisses), justification (giving reasons for a premiss) and attack (giving counterarguments to an argument). Since no available formalism is generally agreed upon,
NVKI-Nieuwsbrief
Verheij uses his own formalism called cumulA, which specifies argumentation as a tree of arguments. The speaker showed exemplary prudence by presenting a weakness of his own formalism, in that counterarguments constitute a rather general class. The question with defeaters
170
December 1998
is: what is a counterargument to what. Experiments with a graphical notation, involving boxes drawn around the elements of an argument, demonstrated that this issue is more difficult than it may seem at first sight.
uitgevoerd. Een voorbeeld is een architect, waarvan de expertise zich betrekt op het ontwerpen van gebouwen. Algemeen gesproken genereert een design agent een design-object-beschrijving op basis van de informatie die van andere agenten wordt verkregen, en stelt de design-agent ook zelf wederom resultaten ter beschikking aan andere agenten.
Finally, the pros and cons of the Argue! system were discussed. The positive aspects are that it provides a graphical representation of attack, and that the argumentation is free and not bounded by the system. The less desirable features include the lack of rules and a not so intuitive user interface. The author concluded with the consideration that because the field of defeasible argumentation is still young, experiments such as these are useful since they provide a testbed or showcase, and can also be a practical aid. Future work may include constructing a template-based interface, as in Tom Gordon's Zeno system.
De bijdrage bediscussieerde hoe een generieke design-architectuur kan worden geintroduceerd door op geeigende wijze een generiek agent-model (gebaseerd op DESIRE) samen te brengen met een (verfijning van een) generiek design model. Als zodanig verenigt het resultaten uit het veld van Multi-Agent Systemen met resultaten uit het veld van AI & Design. De auteurs beogen onder andere deze design architectuur toe te passen binnen het domein van Electronic Commerce.
SESSION AGENT TECHNOLOGY 1 Peter J. Braspenning Universiteit Maastricht
Meer bepaald heeft men de ontwikkeling van een multi-agent 'broker'-architectuur op het oog, waarin zowel 'broker'-agenten als PersonalAssistant-agenten en andere taak-specifieke agenten huizen. Iedere 'broker'-agent kan dan dynamische reconfiguratie van agenten uitvoeren alsmede de realisatie/introductie van nieuwe agenten of de modificatie van reeds bestaande agenten. Het is duidelijk dat de daartoe benodigde conceptuele design zeer goed gebruik kan maken van de eerder aangeduide generieke design agent. Aanbevelingswaardig is om op zoek te gaan naar de originele paper wanneer men preciezer kennis wil nemen van de synergie tussen MAS en AI & Design.
Deze eerste sessie gewijd aan Agent Technologie (AT) omvatte bijdragen van de Vrije Universiteit, de Universiteit Utrecht en een gezamenlijke bijdrage van de Universiteiten van Warschau en Groningen. Hieronder bespreek en plaats ik elk van deze bijdragen in dit turbulente subdomein van de Artificiële Intelligentie. Compositional Design of a Generic Design Agent F.M.T. Brazier, C.M. Jonker, J. Treur, en N.J.E. Wijngaards Vrije Universiteit Amsterdam Deze presentatie door Catholijn Jonker was gebaseerd op een abstract welke refereert naar een uitgebreide bijdrage van dezelfde auteurs in de ‘Proceedings of the AAAI Workshop on Artificial Intelligence and Manufacturing’ van 1998. Het hoofdthema is "ontwerpen"; een taak die vaak door meerdere gespecialiseerde agenten wordt Communicatie binnen Multi-Agent Systemen neemt de vorm aan van uitwisseling van informatie tussen agenten. Echter, verschillende agenten kunnen verschillende conceptualisaties van de omgeving hanteren en dus ook een ander vocabulair gebruiken om hun informationele attitudes te representeren. Dat heeft tot gevolg dat de betekenis die verzendende agenten aan hun gegevens hechten kan verschillen van die welke ontvangende agenten daaraan hechten. Om
NVKI-Nieuwsbrief
Constructing Translations Between Individual Vocabularies in Multi-Agent Systems R.M. van Eijk, F.S. de Boer, W. van der Hoek, en J.-J. Ch. Meyer Universiteit Utrecht werkelijk begrepen te worden dienen beide typen van agenten dus te beschikken over de gebruikte semantiek van de ander, oftewel beide typen van agenten dienen te beschikken over de interpretaties van de gebruikte constanten en relatie-tekens. Er werd een logisch raamwerk gepresenteerd, dat gebaseerd is op de mogelijke-werelden semantiek, waarin het mogelijk is om de communicatie van
171
December 1998
agenten te modelleren die een verschillend vocabulair gebruiken om hun informatie te representeren. Agenten kunnen zo hun eigen vocabulair uitbreiden met termen uit een "vreemd" vocabulair als ook hun 'belief' expanderen met nieuw verworven informatie. Aldus kunnen bruggen worden geslagen tussen individuele lexicons.
De gebruikte architectuur kan helpen bij het verhelderen van de onderlinge afhankelijkheden van de betrokken agenten, die, bijvoorbeeld, bezig zijn met individuele probleemoplossing of meer verantwoordelijk zijn voor de geëigende organisatie van de cooperatieve probleemoplossing. De vier stadia, 1) potentiële herkenning [van een planningsprobleem], 2) teamvorming, 3) plan-vorming, en 4) team-actie, van de architectuur hebben allemaal een inherent dynamisch karakter en vereisen daarom ook daarop toegesneden methoden. De auteurs abstraheren van deze methoden, maar definiëren de stadia door middel van de te bereiken resultaten en associeren deze met de eerder aangeduide motivationele attitudes.
Het is ook mogelijk om de 'beliefs' van andere agenten te vergelijken met hun eigen 'belief' om zo uit te maken of de interpretaties van de symbolen in het vreemde vocabulair corresponderen met de interpretatie van een van de tekens in hun eigen vocabulair. De abstract in de NAIC Proceedings verwijst door naar de aanzienlijk meer uitgebreide bijdrage in de Proceedings van de 8e AIMSA conferentie (AIMSA'98) die dit jaar is uitgegeven door Springer Verlag.
Opgemerkt wordt, dat het thema collectieve team-actie relatief schaars wordt bediscussieerd in de Multi-Agent Systeem (en AI) literatuur in tegenstelling tot de drie daaraan voorafgaande stadia. Wanneer tijdens planexecutie een collectieve intentie wordt onderhouden, is het cruciaal dat agenten op gepaste en efficiente wijze her-plannen wanneer enkele leden (agenten) de hen toebedeelde subtaken niet vervullen. De hoofdbijdrage is dan ook het noodzakelijke reconfiguratie-algoritme dat wordt geformuleerd in termen van de abstracte stadia (van de Wooldridge/Jennings-architectuur) en hun complexe wisselwerking. Echter, ook om andere redenen die te maken hebben met een nadere kennismaking met het thema motivationele attitudes is de paper zeer de moeite waard.
A Methodology for Maintaining Collective Motivational Attitudes during Teamwork B. Dunin-Keplicz en R. Verbrugge Rijksuniversiteit Groningen Deze bijdrage addresseert het gepaste onderhoud van individuele, sociale en collectieve motivationele attitudes binnen een groep van heterogene agenten. Meer speciaal gaat het om een methodologie van teamwerk (van agenten) gericht op cooperatieve probleemoplossing in een dynamische, constant veranderende omgeving. In feite wensen de auteurs met name de notie van collectieve 'commitment' te onderzoeken, waaraan ze achtereenvolgens onderscheiden: de constructie, het onderhoud en de realisatie ervan. Deze hoofdfases van cooperatieve probleemoplossing worden toegewezen aan een abstracte architectuur bestaande uit vier stadia (afkomstig van Wooldridge & Jennings) voor collectieve probleemoplossing. Vervolgens worden deze vier generieke stadia en hun onderlingen afstemming/samenwerking als uitgangspunt gebruikt voor het opstellen van een flexibel re-configuratie algoritme teneinde een initiële planning voor het bereiken van een 'overall' groepsdoel te kunnen bijstellen.
BUSINESS SESSION ON ELECTRONIC COMMERCE 1 Gert-Jan Beijer Bolesian A formal specification of automated auditing of trustworthy trade procedures for open electronic commerce R.W.H. Bons, F. Dignum, R.M. Lee, and Y.H. Tan Erasmus Universiteit Rotterdam
The paper presented by Yao-Hua Tan consisted of two innovative ingredients: the application of existing trade procedures in an electronic commerce (EC) environment and the use of different kind of logics to prove the trustworthiness of the first. The first part of the presentation focussed on current trade procedures and the different ways companies handle risk while
NVKI-Nieuwsbrief
trading. In an EC environment, with trusted third parties and EDI as important cornerstones, companies will only trade goods and services if the same level of security can be offered. For example, current e-mail services do not offer any of the following basic required functionality for trushworthiness: guaranteed, on time, one-time sending (no more, no less) that is traceable and
172
December 1998
management. The solution had to take the following into consideration: 1. an employee is fully responsible for his or her own agenda, 2. an employee can change or refuse changes to his or her agenda, and 3. changes in the agenda should reflect different individual settings, like working hours.
archived. The application of existing procedures requires the following: 1. formal representations of trade principles and procedures, and 2. formal representations of (automated) audit principles and procedures
In the architecture different agents are discriminated: 1. client, 2. call centre agent, 3. work manager (for managing all agenda’s, one work manager for every local office), 4. personal assistent (one for managing the agenda of every employee), and 5. employee.
The techniques that were used to realise this consisted of different kind of logics: Epistemic, dynamic, deontic, and temporal logic in which primitives can be found to represent: a) directed obligations, b) general obligations, c) dynamic logic, d) action performance, and e) temporal to-do obligations.
The advantage of multi-agent technology lies in the fact that in this case, scheduling is a distributed effort. The prototype was developed based on principled design at the conceptual level, using a compositional development method called DESIRE.
The specification method has the appeal of a generic method for proving the trustworthiness of existing and new (still to be discovered) trade procedures. In addition, it visualised where the aspect of trust is allocated.
A logical model of directed obligations and permissions to support electronic contracting Y.H. Tan and W. Thoen Erasmus Universiteit Rotterdam
Distributed scheduling to support a call center: a co-operative multi-agent approach F.M.T. Brazier, C.M. Jonker, F.J. Jungen, and J. Treur Vrije Universiteit Amsterdam
In this lecture Y.H. Tan elaborated his ideas on representing directed obligations in a logical model. It can be seen as a theoretical basis for the first lecture summarised in this track. It was the most theoretical lecture from the three.
In the paper and in the presentation, a prototype system is described that was developed on behalf of, and in cooperation with, the Rabobank. During their search to add value to customers, the Rabobank decided to make their Call Center available 24 hours per day. Customers should always be able to make appointments, whether employees are there or not. Processing a request implies deciding on a procedure to follow and schedule this procedure and the necessary resources. Given the decentralised structure of the Rabobank, this implicated the support of some sort of distributed information technology. The type of support they decided to develop was based on intelligent (multi) agent technology that takes over business after regular office hours, especially agenda Logical preliminaries to represent such models come from standard deontic logic and action logic from Santos and Carmo. The best attempt from to represent this model comes from Herrestad and Krogh, but even their approach has some disadvantages. An alternative could be found in the so-called institutional power operator from Jones and Sergot. These observations have lead to the conclusion that for logically representing models for electronic contracts some new primitives should be used, for instance, on behalf of
NVKI-Nieuwsbrief
A formal model of a contract for electronic commerce and transactions is useful and necessary for automatic processing and auditing of such transactions. Just like in the real world, next to the prize, the payment and delivery conditions are important in electronic trade. Directed obligations play an important role herein. An simple example is: a seller’s obligation to deliver goods is directed from the seller to the buyer. the notion of directed permissions. Permissions, on the other hand, can have more than one meaning: 1. being allowed to without formal approval, 2. being allowed to after formal approval, or 3. involuntary consequence (if one does not bring you your stuff, you have to get it yourself). More research into this area is definitely needed. SESSION EVOLUTIONARY ALGORITHMS
173
December 1998
Dirk Thierens Universiteit Utrecht
Solving 3-SAT using Adaptive Sampling M.B. de Jong and W.A. Kosters Universiteit Leiden
Building Block Filtering and Mixing Cees van Kemenade CWI
In the last paper presented, the authors claim that it is beneficial to study Evolutionary Computing and Neurocomputing in a new unifying framework called Adaptive Sampling. Unfortunately they spend little time at discussing this framework, but instead immediately propose two algorithms - inspired by the Adaptive Sampling view - for solving 3-SAT problems. Experimental results indicate that one of the algorithms outperformes the SAW-ing evolutionary algorithm (also discussed in the previous talk) that was the best incomplete 3-SAT method the authors could find in the literature.
The first lecture was given by the organizational chair of this 10th NAIC conference, Cees van Kemenade (CWI). Building blocks are groups of non-linear interacting bits that have to be present all together in order to find optimal solutions. When these bits are not closely positioned on the representation string and no substrings give correct partial information about the value of the building block, then standard genetic algorithms will not be able to find the optimal solution. In this paper a non-standard genetic algorithm is outlined that deals with this problem. First, building blocks are explicitly identified, and next they are juxtaposed - or mixed - to create optimal or near-optimal solutions. Experiments are conducted on artificial problems specifically designed to test the limits of standard GAs, and results show the potential of the proposed algorithm.
BUSINESS SESSION ON ELECTRONIC COMMERCE 2 Frank van Harmelen Vrije Universiteit Amsterdam The second session on E-commerce consisted of two presentations.
Solving Binary Constraint Satisfaction Problems using Evolutionary Algorithms with an Adaptive Fitness Function A.E. Eiben, J.I. van Hemert, and E. Marchiori Universiteit Leiden and A.G. Steenbeek, CWI
The POWER project: ProgrammaOndersteuning in Wet en Regelgeving T.M. van Engers Belastingdienst
The second paper presented joint work of the Universiteit Leiden and CWI and was given by Jano van Hemert. The paper experimentally compares three evolutionary algorithms (COE, SAW, and MID) on a test suite of randomly generated binary Constraint Satisfaction Problems with finite domains. All three algorithms adapt the fitness (penalty) function during the search process. While the performance of the Co-Evolutionary (COE) approach is rather unsatisfactory, the other two algorithms seem to trade off success rate versus computational effort. The Microgenetic Iterative Descent (MID) performs best with respect to success rate, but the Stepwise Adaptation of Weights (SAW) is only slightly worse on harder problems, and achieves this with less fitness evaluations. - to help in checking the consistency of legislation, - to simulate the effects of introducing new legislation, and - to make explicit existing internal knowledge within the revenue service.
The first was by Mr. van Engers from the Dutch tax office. He presented the POWER project. This project tackles the following problems: - tax regulations are often difficult to understand for citizens , - expert knowledge on tax legislation is sometimes scarce, even within the revenue service, and - it is often hard to predict the results of introducing new legislation. To tackle these problems, the POWER project has the following aims: - to make legislation more transparent, - to enable a uniform interpretation of legislation, following block-diagram summarises the steps that must be taken in the POWER project:
The POWER project is a collaboration between the University of Tilburg, the University of Amsterdam, the Ministry of Justice and the Treasury. The
NVKI-Nieuwsbrief
174
December 1998
set of attribute-value pairs plus possibly relations constraining the possible values. Van Rijn presented an on-line demonstration of some of the software developed at Data Distilleries for analysing credit-risks in banking and in cross-selling products to customers (cross-selling is a nice word for selling customers a product they didn't ask for). After the presentation, a very interesting and lively debate followed on the legal and moral implications of these techniques. Is it allowed to use data for purposes for which the data was not originally volunteered? What are the moral and legal implications of judging individuals on the basis of statistical profiles? In this reporter's humble opinion these serious issues were not satisfactorily dealt with by the speaker.
Plaatje Van Engers finished his interesting presentation by discussing a number of the hurdles that must be taken by the POWER project: - the translation between the source legislation and a more formal representation must be both bi-directional and transparent for all parties involved, - the programme will have to operate in a culture that is not very IT oriented, - the programme will have to be integrated with other areas of legislation, even though these areas often use very different conceptualisations (e.g., civil law vs. criminal law), and - results of the programme will have to be integrated in a traditional IT environment .
SESSION DECISION NETWORKS 2 Joost Kok Universiteit Leiden
Customized E-Commerce by Data Mining F. van Rijn Data Distilleries
Variational Belief Networks for Approximate Inference W. Wiegerink and D. Barber Universiteit Nijmegen
The second presentation was by Mr. van Rijn from Data Distilleries, one of the many recent CWI spin-off companies. Data Distilleries was described as a hightech market leader in the data-mining area. Van Rijn described how many companies these days have large and fast growing data-bases at their disposal. The main task of data-miners is to exploit this information. Traditionally, statistical techniques were used for this task. The main drawback of these techniques is that they can only be used to confirm or deny an already formulated hypothesis. They do not help in formulating the hypothesis itself. Data-mining techniques, on the other hand do present new (and often unexpected) hypotheses to the user. These techniques often generate and test many thousands of potential hypotheses, of which only the most interesting ones are ultimately presented to the user. In this context, a hypothesis often has the form of a relationship between a particular customer profile and a customer's behaviour, where a customer profile is a
The first paper in the session on Decision networks was presented by Wim Wiegerinck of the Foundation for Neural Networks, University of Nijmegen. Approximate techniques are needed for belief networks with larger clique sizes. The paper proposes to use mean-field theory. At first sight, mean field theory is not suited because only completely factorized models can be used. The paper shows that it is also applicable to a larger class of models. One of the advantages is that a lower bound on the likelyhood is obtained. Variational Approximations in a Broad and Detailed Probabilistic Model for Medical Diagnosis W. Wiegerinck, E. Ter Braak, W. Ter Burg, M. Nijman, Y. O, J. Neijt, and H. Kappen Utrecht University Hospital networks. The second part of the presentation consisted of a demo of the system. It was interesting to see that such a large network could be manipulated on-line.
The second paper was co-presented by Wiegerinck and Ter Braak. Ter Braak is affiliated with the department of internal medicine of the Utrecht University Hospital. A larger belief network for medical diagnosis was presented. It consisted of two parts: a parent network and a child network conditioned on the state of the parent network. The nodes in the child network are independent. Similar techniques to the ones used in the first paper in the session were used to attack such
NVKI-Nieuwsbrief
Representation and Learning Capabilities of Additive Fuzzy Systems D. Ettes and J. van den Berg Erasmus Universiteit Rotterdam The third paper was presented by Ettis of the
175
December 1998
via the World Wide Web. People interested in performances and the theater's program, can have a preview of the Hall and information about the performances at there own homes.
Erasmus University. An interesting comparison between Feedforward Neural Networks (FFN) and Additive Fuzzy Systems (AFS) was given, and some of the theory about representation and generalization for FFN was lifted to AFS. The representation capabilities of several AFS were discussed and a fuzzy decision tree algorithm was proposed. A distinction in the fuzzy rule base is made by distinguishing the more important fuzzy rules. It still has to be tested in practice whether it can be used as a tool for data mining.
The demonstration showed how users can exploit the Music Hall and experience the view on the stage from particular seats in the hall. One can imagine that this kind of visual information indeed adds to the information that can be acquired by a telephone call to the Theater. On the other hand, it is questionable whether dragging the mouse to navigate though the entrance and up the stairs of the building is the obvious interaction model. From the discussion after the demonstration became clear that perfectionizing this particular interaction had not been the primary research goal of the project. Instead, the emphasis of the research had been on the interactions with Karin. The demonstrations showed some examples of more or less complex natural language interactions with Karin. Using natural language a user can ask information on different performances and make reservations.
SESSION MULTI-AGENT DEMONSTRATIONS 1 Erica C. van de Stadt WizWise Technology To promote the feeling with real applications, last NAIC's program included System Demonstrations in addition to paper contributions. This short discussion reports on two systems demonstrations: A Virtual Reality Environment for Information and Transaction authored by Hendir Handorp and Anton Nijholt, both from the University of Twente. And a second one entitled A Model for Distributed Multi-Agent Traffic Control from Christine Bel, Wim van Stokkum and Rob van der Ouderaa, all authors work at Kenniscentrum CIBIT.
The second demonstration showed a graphical representation of a railway traffic simulation. The underlying system uses distributed real time traffic control and is based on multi-agent concepts. The agents in the model are sticked to the trains and the elements of infrastructures (crossings). In case where trains arrive at the same crossing at the same time, trains and elements of infrastructure negotiate to form a plan for passing the railway crossing. The planning is based on quality criteria formulated in terms of the interest (satisfaction) of the passengers. The model of control is based on local optimization between arriving trains and an element of infrastructure. The demonstration showed a prototype implementation of the system for the railway section Arnhem-Utrecht.
The research underlying the first demonstration was rather on interaction modalities than on virtual reality modeling. Research on interaction modalities includes natural language interactions with computerized agents (by keyboard or speech). In the demonstrated system the agent was not merely a computer module but instead, this agent was personified into a woman called Karin. Karin (modeled by VRML specifications) is situated in a virtual version of the Muziek centrum of Enschede. (Which I happen to know and indeed could recognize in de moving pictures presented.) The application can be accessed Compared to a more traditional central control model, the results of the simulation show that on the passengers satisfaction criteria, the prototype-systems scores increase. Unfortunately, the demonstration did not address scalability issues. In effect, the conditions for or the situations in which the proposed local optimization strategy will be successful did not became clear.
editions of the NAIC will continue to reserve program space for demonstrations and applications. SESSION NEURAL NETWORKS Eric Postma Universiteit Maastricht Environment learning and localization in Sensor Space B. Kröse Universiteit van Amsterdam
Having attended both system demonstration sessions my personal opinion is that system demonstrations certainly fit in the overall program of NAIC. AI techniques applied in real situations make the theoretical concepts concrete and I hope that future
NVKI-Nieuwsbrief
The first contribution to the Neural Networks
176
December 1998
session focussed on the visual navigation of a robot vehicle through a virtual room. Traditional approaches to this problem represent the environment in some manner. Often, a geometric model of the world is constructed and the location of the vehicle is tracked by keeping a record of the number of revolutions of the wheels. An estimate of the position in the world is then obtained from the distance traversed and the number and directions of turns taken from some starting point. However, measuring the distance in this way is not very accurate. Therefore, an additional or alternative source of information is required for estimating the position. After some initial experiments with range sensors feeding their outputs into a Kohonen map, Kröse turned to appearance modelling of the environment. Appearance modelling is inspired by the work of Murase and Nayar, who projected 2D views of 3D objects onto a low-dimensional shape space. They showed that novel views of an object can be recognised by matching them to their nearest neighbours in shape space. Appearance modelling in navigation proceeds in a similar fashion: from a large number of positions, snapshots of the environment are taken and stored along with their location. Subsequently, the vehicle is placed in an arbitrary position. Again, snapshots are taken and compared to the stored snapshots. An estimation of the position is obtained, for instance, by interpolating between the positions of nearest-neighbouring stored snapshots.
a vehicle was placed at a few hundred different positions. From each position, snapshots were taken and stored as images. To reduce the dimensionality of the images, they were mapped onto their principal components. A subset of the components served as a (50-dimensional) subspace on which the images were projected. In the test phase, the vehicle was placed at 100 random positions. At each position, a snapshot was taken and the probability of being in a certain location was computed. If the probability was below some threshold, the camera was turned to a suitable position to take a new snapshot taken. The results proved the feasibility of the approach. When allowing multiple snapshots, in about 90% of the cases a correct localisation was obtained. The approach of Kröse appears to be a viable alternative to traditional model-based approaches. In the near future experiments will be performed using a real robot placed in a real-world environment. Given the presence of uncontrollable factors such as variations in lighting conditions and inaccuracies in the positions of the camera and vehicle, the move from the virtual to the real world will not be without difficulties. Maximum Likelihood Weights for a Linear Ensemble of Regression Neural Networks M. van Wezel, W. Kosters, and J. Kok Universiteit Leiden
Experimental studies were performed in a virtual replica of Kröse’s lab. Within this virtual environment By combining the outputs of multiple networks trained on the same task, the overall generalisation performance may be enhanced considerably. The generalisation performance of a collection of neural networks, a so-called ensemble, can be optimised by appropriately weighting their outputs. Michiel van Wezel presented work that attempted to find an appropriate set of weights for the linear combination of networks in an ensemble. He discussed three general ways of defining the weights, i.e., ‘bagging’, ‘bumping’, and ‘balancing’. In bagging, all weights have equal values whereas in bumping there is exactly one non-zero weight. Balancing is the approach in which an optimal set of weights is determined by a quadratic programming technique. Van Wezel presented an alternative approach based on the maximum likelihood principle. After some elaborate derivations, he arrived at what he claimed to be an effective formula for determining the ensemble weights. The determination required the use of a conjugate gradient technique, the Bates-Granger technique, or the simulated-annealing technique.
presentation corroborated the claim of effectiveness. Maximum likelihood weighting outperformed the other methods on sets of marketing data, stock exchange data, and a wave-height data. For large ensembles, the proposed method largely outperforms the bagging method. Bumping and balancing performed better than bagging, but worse than the maximum likelihood weighting. Interpreting Knowledge Representations in BP-SOM T. Weijters and A. van den Bosch Technische Universiteit Eindhoven, Katholieke Universiteit Tilburg One of the disadvantages of multilayer perceptrons (MPs) is that they are like black boxes. It is very hard to interpret the internal representations of a trained network. The reason is that, unlike rule-based systems, the representations are distributed patterns of weights, rather than sets of rules. In the final talk of the Neural Networks session, Ton Weijters presented his approach to opening the black box of his BP-SOM network.
The empirical evidence presented at the end of the
NVKI-Nieuwsbrief
177
December 1998
help of some simple examples, Weijters showed that the SOM representation can be translated into a set of understandable IF-THEN rules. As a consequence, the black box is opened and the hidden strategies used by the trained MP can be uncovered.
BP-SOM is a hybrid network composed of a multilayer perceptron and a Kohonen Self-Organizing feature Map (SOM). The network integrates the error-based learning of the MP with the similarity-based learning of the SOM. The integration is achieved by connecting a SOM to the hidden layer of the MP. During training, the SOM receives the activation patterns of the hidden layer as inputs. In addition, the SOM influences the formation of the hidden patterns through feedback connections to the hidden layer. The overall effect of the SOM is that it reduces the dimensionality of the hidden representations by using information from the hidden representations themselves. Previous studies have shown BP-SOM to outperform both standard MP and MP with weight decay on a variety of tasks.
Although the three contributions to the Neural Networks session were entirely different in scope and presentation, they provided an interesting overview of neural-network research. Now that the hype is over, theoretical and application-oriented neural network research form an indispensable part of artificial intelligence. The distribution of neural-network research over disparate domains such as statistics, robotics, and information retrieval provides a case in point. SESSION LOGICS 2 Yao-Hua Tan Erasmus Universiteit Rotterdam
In his presentation Weijters concentrated on interpreting the representations arising in BP-SOM during learning. Earlier observations suggested that the BP-SOM representations are highly structured. In particular, the SOM-part appears to organise the examples in the data set into clusters of similar representations belonging to the same class. With the The paper was presented by André Bos. Many problems in AI are intractable. One way to deal with this problem is so-called Knowledge Compilation (KC). The basic idea of knowledge compilation is that a distinction is made between the fixed part of a problem and a varying query part. The fixed part can then be computed off-line and reused for specific queries, hence reducing the actual run-time complexity. Bos gave a nice example of knowledge compilation. In expert systems for the process industry like oil refinery or a chemical plant the actual model of the process is usually quite complicated. Instead of reasoning with the full model each time a query is computed a table can be computed off-line with input-output behaviour of the process. When querying the process this table is used instead of recomputing every time the whole model. In strict knowledge compilation there are specific constraints such as (1) that the result of the off-line computed fixed part should be polynomial-space bounded, and (2) that the on-line query reasoning can be done in polynomial time. The problem with these constraints is that they are so strict that they exclude some of the most obvious applications of KC. For example, neither clausal inference nor logical abduction satisfy these constraints. In the presentation was explained how the first constraint can be relaxed to obtain a new type of knowledge compilation, the so-called nonstrict knowledge compilation (NKC) such that clausal inference as well as abduction can be compiled in NKC. However, relaxing the constraint does not solve
NVKI-Nieuwsbrief
Non-Strict Knowledge Compilation A. Bos and C. Witteveen Technische Universiteit Delft all practical problems. It was discussed that in NKC problem compilation critically depend on the construction of compilation algorithms that can handle the size of the set of prime implicates of a real-life problem, which is known to be hard to compute. This will be the topic of further research carried out by the authors. Generated Preferred Models and Extensions of Nonmonotonic Systems J. Engelfriet and H. Herre Vrije Universiteit Amsterdam The paper was presented by Joeri Engelfriet. Stable model semantics is a well-known semantics for normal logic programming and it can also be used to provide a semantics for normal default logic via a simple translation from default logic into logic programming. Normal logic programming and default logic cannot be used to represent disjunctive information. Special extensions of logic programming and default logic, the so-called disjunctive logic programming and disjunctive default logic, have been developed to represent disjunctive information. Herre and Wagner introduced a special type of stable models, the socalled stable generated models that can be used to provide a semantics for these disjunctive nonmonotonic formalisms. In the presentation it was explained that the stable generated models correspond with the temporal models, which were
178
December 1998
introduced by Engelfriet and Treur to provide a semantics for default logic. In particular, it was shown that the temporal models can be used to model disjunctive default logic in a similar way as was done with the stable generated models. Also the relation was explained between temporal models and semiconstructive extensions in default logic. In particular, it was shown that temporal models model very precisely the step-wise construction of semiconstructive extensions. Engelfriet presented the paper in an original way. First, he presented the conclusions, and then he reasoned backwards to the introduction of the paper, explaining for each step what was a necessary prerequisite to arrive at this The paper was presented by Shan-Hwei Nienhuys-Cheng. Inductive logic programming (ILP) is a type of machine learning that is based on so-called refinement operators. The basic idea of downward refinement operators is that you start with the most general theory, namely a tautology that implies every sentence, and you gradually refine the most general theory by specializing it until it fits with the examples that have to be learned. Technically, this specialization is the result of a substitution applied to a formula. An example of such a specialization is that you refine a universally quantified sentence like ∀xB(x), which means 'x is a block', to B(a) if it is known that the object a is a block. Since ILP has usually been implemented in Prolog it has been developed mainly for the underlying logic of Prolog, namely Horn clause logic. Horn clause logic is a fragment of first-order predicate logic. One of the restrictions is that Horn clauses are universally quantified formula that do not contain existential quantifiers. In this presentation it was explained how refinement operators can be generalized to Predicate Calculus in conjunctive normal form, which is equivalent to full first-order predicate logic. One of the typical problems that had to be solved is that substitutions on existentially quantified sentences sometimes lead to generalizations instead of specializations. For example, if we substitute the variable y for x in the matrix of the formula ∀x∃yB(x, y), then the result of the substitution ∀xB(x, x) is more general than the first formula. The authors adapted the definition of substitution in such a way that when applied to an arbitrary PCNF formula the result is always a specialization. Some of the adaptations are quite complicated, and the authors found some very ingenious solutions for this adaptation. By redefining downwards refinement to PCNF they showed the way how the rich tradition of ILP machine learning can be applied to full first-order predicate logic. The talk was presented by Shan-Hwei Nienhuys-Cheng in a very lively way. We are very
NVKI-Nieuwsbrief
step. Probably this is the best way to present technical papers, because you do not have to work your way through technical details before you can grasp the conceptual ideas.
Substitutions and Refinement operators for PCNF S.-H. Nienhuys-Cheng, W. van Laer, and L. de Raedt pleased to see such an active and creative AI researcher back at the NAIC after her illness of last year. SESSION AGENT TECHNOLOGY 2 J.-J. Ch. Meyer Universiteit Utrecht A Formal Embedding of AgentSpeak(L) in 3APL K. Hindriks, F. de Boer, W. van der Hoek, and J.-J. Meyer Universiteit Utrecht The varied session started with a presentation of Koen Hindriks on A Formal Embedding of AgentSpeak(L) in 3APL. After giving a short overview of the agent language AgentSpeak(L) developed by Anand Rao and the agent language 3APL (triple-A-P-L) that we have proposed ourselves, Koen showed (or rather made plausible in a rather audience-friendly way, leaving the technical details to the reader of the paper) that it is possible to embed the rather involved language AgentSpeak(L) into a subset of the 'much simpler' language 3APL. In order to make precise in what sense this embedding / simulation works, a notion of bisimulation (originally stemming from both concurrency theory and modal logic) was introduced, ensuring that the embedding is faithful with respect to observable steps of computation. (Without this important restriction the embedding would be rather trivial since both languages are Turingcomplete.) As a consequence it appeared that AgentSpeak(L) could be simplified without sacrificing expressive power. In reply to a question of Joeri Engelfriet about the involvement of the simulation, i.e., a terribly complicated code as a result of the translation, Koen answered that in fact the resulting 3APL
179
December 1998
code is generally much simpler than the original code in AgentSpeak(L), since in 3APL various implementation details (regarding e.g., stacks of plans) are suppressed.
was clear that the presenter did not want to adhere to any 'strong' mentalistic interpretation of the resulting agent, I myself thought that (a posteriori) it was perhaps still possible to describe the behaviour of the agents in a BDI-like way, the agent's agenda providing already for an obvious I(ntention) aspect. At least it would be interesting to look at the agent in this way, since it may yield a neat way to reason about its behaviour. During the discussion afterwards Frank Dignum raised the question whether the work reported was really that specific for "a corporate environment" as was suggested by the title of the talk. Unfortunately this interesting discussion had to be ended by the chairman for reasons of time.
Agent-Based Information Gathering in a Corporate Environment A. Wan, F. Wiesman and P. Braspenning Universiteit Maastricht The second paper was presented by Fred Wan. The corporate environment mentioned in the title of the talk pertained to KPN, and the paper described an agent system for data-mining and knowledge discovery for this company, where a 'weak agent' (i.e. non-mentalistic) position was taken in the design of the agent system. Fred described the architecture of the system in some detail. In this architecture the agent's autonomy is realised by manipulation of an agenda according to priority ratings on relevance and through a sparse need of user interaction. Pro-activeness is realized by means of various retrieval and filtering actions not explicitly instigated by the user. Although it The last paper of the Agent Technology 2 session was written by a "whole bunch" (as I somewhat disrespectfully announced, but no disrespect intended!) of authors, viz., a mix from the VUA and Karlskrona/Ronneby University. It was presented by Catholijn Jonker who had quite a busy day that Thursday (I witnessed at least 3 excellent presentations by her that day). Catholijn illustrated a method of compositional designing multi-agent systems in DESIRE developed at the VU by a very nice real-life example, viz. that of negotiation in the context of electricity companies dealing with the balancing of the load of the network. Here the company is willing to cut down electricity prices if customers are willing to avoid use of electricity during certain peak hours. The negotiation that is performed between company and customers can be done automatically by the multi-agent system designed, and I believe it is actually running in Sweden. An important feature of the design & verification method presented is that the system can be viewed at several levels of process abstraction, and that properties at different levels of process abstraction are involved in the verification process.
Compositional Design and Verification of a Multi-Agent System for One-to-Many Negotiation F. Brazier, F. Cornelissen, R. Gustavsson, C. Jonker, O. Lindeberg, B. Polak, and J. Treur Vrije Universiteit Amsterdam and Karlskrona/Ronneby University Promotor: Prof.dr.ir. J.A. Nerbonne review by Antal van den Bosch Tilburg University On October 19, 1998, Erik Tjong Kim Sang successfully defended his thesis "Machine Learning of Phonotactics" in Groningen. It concluded an eight-year period of research which Erik spent partly as a researcher in Groningen, and partly as a researcher/lecturer at Uppsala University in Sweden. Prof. John Nerbonne (Groningen University) was the promotor, and the thesis committee consisted of Anton Nijholt, Ger de Haan, and Nicolay Petkov. At the defence, the promotion committee was furthermore strengthened by Gertjan van Noord, Ronan Reilly, Walter Daelemans, and Ben Spaanenburg. Erik's thesis concerns the application of a variety of machine learning techniques to the problem of modelling the phonotactics of a language. The phonotactics of a language is the set of constraints on what constitutes a valid syllable (or a valid word, made up of one or more syllables) in that language. The phonotactics of English allow "pand" to be a possible English word, and would reject "padn". The phonotactics of languages differ; "mloda" would not be an acceptable Dutch or English word, but it is in Polish.
MACHINE LEARNING OF PHONOTACTICS Ph.D. dissertation of Erik F. Tjong Kim Sang
NVKI-Nieuwsbrief
180
December 1998
How does one model the phonotactics of a language? One possibility is to sit down and think up all the rules and constraints that could discriminate between allowed and disallowed syllables and words. This has been done for many languages by linguists, which has resulted in many good generic descriptions of language-dependent but also some language-universal phonotactics. The approach taken in this thesis differs from the traditional linguistic approach in that it describes methods to derive phonotactical models automatically from examples of existing words. The methods used in the thesis are categorised under the Because it is obvious that such choices may affect learning and the success of it, and because one would like to have a grip on their effect, the thesis covers a matrix of combinations of representation choices. It comes as no surprise that this matrix dictates the structure of the body of the thesis in which the methods are applied to the problem. The first dimension in the matrix is formed by the three learning methods (HMM, SRN, and ILP), described respectively in chapters 2, 3, and 4. The second dimension concerns the lowest-level representation of the data, which is alternated between spelling and phonology. While spelling is of course a common format for words, phonology (pronunciation) is commonly and generally considered to be the actual level at which phonotactics operate (its spelling counterpart being referred to sometimes as graphotactics). Spelling is not phonology; in Dutch and English it has grown to be a distorted reflection of it because it has developed at a different pace, and because it also represents phenomena that are unrelated to (and sometimes even defy) pronunciation: morphology (e.g., the "-dt" and the conjunction "-n-" in Dutch), etymology (the Dutch word "synthese" mirrors the non-Dutch spelling of old non-Dutch lemmas; it is not written as "sintese"), and historical word-image conservatism (the old "-isch" and "-lijk" at the end of Dutch words). Although I would take these aspects of spelling to be a reason to discard it from the study and only focus on phonology, Erik simply continues attempting both. In the concluding chapter, he is "surprised" to find that learning phonotactics from phonology was easier than learning it from spelling. He shouldn't have been. A general finding in the literature (cf. Selkirk, 1984; Blevins, 1995) is that almost all syllables in languages like English and Dutch adhere to the sonority principle. When people speak, their vocal tract opens and closes to the rhythm of syllables, which is pretty directly correlated to the sonority of the uttered phonemes. Sonority peaks at vowels, and is lowest around syllable boundaries (with only the /s/ as the occasional odd one out). This simple regularity already governs a major part of phonotactics; "padn" is not valid
NVKI-Nieuwsbrief
umbrella term "Machine Learning", and include Hidden Markov Modelling (HMM), Simple Recurrent Networks (SRN) and Inductive Logic Programming (ILP). Basically, all three methods are given examples of Dutch monosyllabic words to derive their model of Dutch phonotactics from. How these "examples" are represented depends not only on the method, but also on several choices the experimenter can make; for example, whether to include some linguistic abstraction, estimated by linguists to be useful in a phonotactical model. because the "d" has a lower sonority than "n" while in a syllable's coda sonority is only allowed to decrease gradually (which is why the two-syllable "padna" would be OK again - the "n" is the onset of the second syllable and the sonority rhythm is not violated). Of course, this "simple" general principle assumes a non-trivial abstraction - sonority. The third dimension is whether the low-level representation is augmented with linguistic expert knowledge. Learning may be facilitated and the result of learning may be more successful when at the onset of learning some general abstract knowledge about phonotactics is available to the learner, e.g., by representing the example data in a format that focuses on the aspects of the data linguists claim to be relevant. The knowledge that is thus included in half of the experiments is a general model of the phonological structure of the syllable. Of course, it would have been a give-away to explicitly tell the learner about real phonotactic constraints before it starts learning phonotactics; the chosen syllable model cleverly represents knowledge that is just below that level of specificity. It does not tell the learner what phonotactics is; however, it brings the learner quite close to the point at which it would be daft to be unable to see the solution. The thesis describes empirical work, and everything follows from the evaluation of the experimental results. Each experiment, in which one learner is trained on 5577 Dutch monosyllabic words, results in some way in an automatically learned model of Dutch phonotactics. To evaluate each experiment, Erik measures (1) the model's acceptance rate (in percentages) of 600 Dutch words that were not in the learning material, and (2) its rejection rate on 600 implausible words such as "ywua" and "odhnf", generated by randomly sticking letters to each other (with probabilities derived from the real-word list). Of course, any good model of phonotactics should be able to reject alien words or syllables. Reference is made to the very classic discussion on the
181
December 1998
possibility of learning language only on the basis of positive examples, without negative examples: Gold said in 1967 that it was impossible to build a perfect model for a general (i.e., not necessarily natural) language which can generate infinite numbers of strings (e.g., words) by only looking at positive examples of the data; it can be expected to perform badly on rejecting alien strings. Erik's learners are learning only from positive data, so the fate predicted by Gold is imminent. However, it is encouraging that in real life, children pretty much lack and ignore negative examples, while they all tend to succeed in learning quite a bit of language. Innate knowledge, Erik postulates, may be their key aid, and the inclusion of linguistic background knowledge, one of the experimental dimensions described above, may be seen as a model of hardwiring useful innate (initial) knowledge into the learner.
enforced by the specific biases of the learners themselves. First, the HMM method is found to work better with letter or phoneme bigrams than with unigrams; the latter is more content with "pajn" than with "pijn" because it prefers "a" over "i" at that specific position while it does not know that "aj" is rather unlikely in Dutch. It is not surprising that the minimal amount of context one could ask for, viz. one neighbouring letter as in bigrams, helps to learn phonotactics dramatically. In the end, bigram HMMs are found to be rather accurate (around 99%) in accepting positive data, trained on any representation; rejection of negative data is almost as accurate, at least on phonological data. Rejection of negative data in the spelling representation is worse: 91.0% without, and 94.5% with built-in linguistic knowledge. Erik concludes that the learners found it easier to discover regularities in the phonological data as compared to the spelling data. Again, this comes as no surprise: spelling is not phonology, and phonology at the syllable level is known to behave quite nicely.
Returning to the first dimension, the three learning methods, it becomes clear when reading the chapters 2, 3, and 4 that some serious choices of data representation that the experimenter has to make are Second, the SRNs turn out to perform quite disappointingly. They are trained indirectly on phontactics by having them learn to predict the next phoneme in a sequence of phonemes. SRNs learn their own memory, and the hope was that this memory, which is quite small and typically grows to represent general aspects of the learning data rather than example-specific information, will represent the underlying grammar of the sequences, viz. the phonotactics. Although this leads to networks accepting valid words at a high accuracy, deriving rejection of alien words from clues in the network's output turns out infeasible; most alien strings are simply accepted. The SRN's classification task, redicting a phoneme given its predecessing sequence, could have been one by any classification learning method, including decision trees and memory-based learners. Although Erik corrected this claim lateron, the thesis mentions that decision trees and memorybased learners cannot be trained on positive examples only - but they can be trained straightforwardly on the SRN's task. Moreover, Erik discards them from is study because they do not generate rules such as the ILP approach - but neither do the SRNs or the HMMs.
prefix and suffix hypotheses: when a specific sequence of letters is valid, and some specific letter gets attached to the left or right of this string, the new string is also valid. Fed with all these hypotheses and the data themselves, the ILP systems tested need a few training cycles to arrive at what they estimate to be a good set of hypotheses about the phonotactics of words and some selected prefix and suffix hypotheses. While the acceptance rate is of HMM-quality, ILP's rejection rate is considerably lower (63.2% for the spelling data, and 86.2% for the phonetic data). For instance, it accepts "kwrpn", due to blind prefixing of consonants left of the "n" which on its own is a valid word according to the data. The problem disappears when strange words such as "n", "t", "sst", and "pst" are removed from the data by hand. Apparently, the ILP approach is very sensitive to these noisy instances in the data, more so than the HMM approach of which the statistical smoothing behaviour is able, apparently, to ignore this level of noise. When augmented with the mentioned syllable structure model, the ILP learners are roughly as accurate in acceptance and rejection as the HMM learners. The general discussion focuses on the question whether the HMM or the ILP method yields the best models of Dutch phonotactics, SRNs having left the backdoor. Erik recommends ILP because (1) it can do a good job when it is equipped with innate linguistic knowledge, (2) it generates rules which can be interpreted by humans, and (3) it is faster than its HMM counterparts. The results that fill the
The third learning method, ILP, is a strong contestant. It induces rules from data, building on a background knowledge model. Of course, this background model correlates with the third dimension in the global matrix: no linguistic knowledge versus some relevant innate linguistic knowledge. The "no linguistic knowledge" variant tells the ILP explicitly what the HMM represents implicitly, namely a collection of
NVKI-Nieuwsbrief
182
December 1998
global matrix support claim (1), and claims (2) and (3) should be taken as important considerations for anyone interested in inducing an interpretable set of constraints (e.g., a grammar) from raw sequences of language data (or, for that matter, gene sequences or stock exchange time series). The main strength of the thesis is that it has done what was asked at the onset; it has come up with recommendations for successful automatic learning of phonotactics, and has done that in a process that very explicitly introduces a wide and interesting variety of techniques along the way. The thesis can be read as a primer in HMMs, SRNs, and ILP.
reduced" data. It seems that the learning data covered enough of Dutch to contain almost every possible "Dutch" bigram; words not in the training set appear to be composed of these bigrams, while alien words perhaps have at least one or two unknown bigrams too many. Having a good bigram memory is an adequate basis for doing phonotactics. More than just a bigram context may do just a bit better, but that is part of future research, and fellow researchers of Erik in Groningen have already taken up the challenge. Above all, I have found the thesis to be an enjoyable piece of work. It is a clearly written study in machine learning of natural language, a blooming field to which this thesis is definitely a valuable contribution. I should also note that the style of writing contributes to this positive impression. Erik's prose is direct, gets to the essentials without deviations, and represents a clear flow of argumentation.
The raw fact that phonotactics can be learned is less surprising, given the regularity of the domain, especially on the phonological level. Moreover, giving the learner clues about syllable structure is almost a give-away. On the other hand, this criticism is easily outweighed by the big and pleasant surprise of the reported rejection rates of the learners which had no negative examples and no linguistic background: the HMMs and the ILPs with "noiseFortunately, the field of machine learning of natural language has not lost Erik upon finishing his thesis. Recently, Erik started as a postdoc researcher at the Centrum voor Nederlandse Taal en Spraak (Centre for Dutch Language and Speech), at the Universitaire Instelling Antwerpen (Antwerp University). His research project is part of the European "Learning Computational Grammars" network project, financed by the EC as a part of the TMR programme (Training and Mobility of Researchers). Erik is now focussing on tasks concerning the recognition of the structure of noun phrases in texts.
AND REASONING UNDER UNCERTAINTY Dissertatie van Peter Grünwald, CWI Promotor: Prof.dr. P.M.B. Vitányi Bespreking door Ronald de Wolf UvA en CWI INLEIDING Op 8 oktober van dit jaar verdedigde Peter Grünwald, OIO aan het Centrum voor Wiskunde en Informatica (CWI), zijn dissertatie in de Oude Lutherse Kerk van de Universiteit van Amsterdam. De meeste dissertaties in de exacte wetenschappen zijn tegenwoordig niet veel meer dan bundelingen van gepubliceerde artikelen, vooraf gegaan door een snel geschreven inleiding. Zo niet hier. Grünwalds dissertatie is een flink boekwerk van zo'n 300 pagina's. Hoewel sommige delen snel en onder hoge druk zijn geschreven, maakt het toch de indruk van een doorwrocht geheel, dat meer is dan de som van de gepubliceerde artikelen waarop sommige hoofdstukken gebaseerd zijn.
References Blevins, J. (1995). The syllable in phonological theory. The handbook of phonological theory (ed. J. A. Goldsmith), pp. 206-244. Cambridge, MA. Blackwell. Selkirk, E. O. (1984). On the major class features and syllable theory. Language sound structure (eds. M. Aronoff and R. T. Oehrle), pp. 107-136. Cambridge, MA. The MIT Press. Tjong Kim Sang, E. F. (1998). Machine learning of phonotactics. Ph.D. thesis, Rijksuniversiteit Groningen. Groningen Dissertations in Linguistics, number 26.
In brede zin gaat het proefschrift over het modelleren van gegeven data via zogenaamde Minimum Description Length (MDL) principes, en over de inductie en generalisatie die daarvan het gevolg zijn. De gebruikte methoden zijn grotendeels statistisch, coderingstheoretisch en informatietheoretisch. Wanneer we de dissertatie binnen het kader van de kunstmatige intelligentie
THE MINIMUM DESCRIPTION LENGTH PRINCIPLE
NVKI-Nieuwsbrief
183
December 1998
willen plaatsen, is de kortste omschrijving van het onderwerp waarschijnlijk `statistische Machine Learning'. Machine Learning (ML) is het deelgebied van de kunstmatige intelligentie dat zich bezighoudt met automatisch leren, dat wil zeggen met het algoritmisch ontdekken van verbanden en regelmatigheden in gegeven data. Er zijn verschillende motivaties te geven voor Machine Learning. Ten eerste is menselijk leergedrag bij uitstek een gebied waar intelligentie ten toon gespreid wordt, en dus een voor de hand liggend object van studie voor AI-onderzoekers en cognitief psychologen. Een tweede, meer praktische reden is dat goede leeralgoritmes kennis genereren: de regelmatigheden die leeralgoritmes vinden kunnen gebruikt worden zowel voor voorspelling van toekomstige gebeurtenissen, als voor verklaring van vroegere gebeurtenissen. Bij de bouw van expertsystemen en andere AI-toepassingen blijkt keer op keer hoe moeilijk het vaak is voor mensen om hun kennis "op te schrijven". Wanneer we dus een kennissysteem willen bouwen (bijvoorbeeld een medisch expert-systeem dat verbanden legt tussen symptomen en ziektes), is machinaal-geleerde kennis
vaak het enige goede alternatief voor onvoldoende beschikbare menselijk-gegenereerde kennis. Binnen de ML bestaan verschillende "frameworks" of "paradigma's", die zich met name onderscheiden door de structuren die gebruikt worden om kennis te representeren. Zo heb je bijvoorbeeld leren in neurale netwerken, in beslissingsbomen, en in logica (inductive logic programming). Een aanpak die minder afhangt van de kennis-representatie en daardoor algemener is, is het Minimum Description Length principe. Dit principe ontstond in de jaren 60 in een geïdealiseerde vorm in het werk van Ray Solomonoff. De meer praktische uitwerking kwam in de jaren 70 en 80, met name in het werk van Jorma Rissanen. Het grootste deel van Grünwalds proefschrift beslaat theoretisch en experimenteel onderzoek naar dit MDL principe. De dissertatie valt uiteen in drie delen, die we hieronder zullen bespreken. DEEL I. THEORY OF MDL bits codeert. Data met een hoge kans krijgt dus een kort codewoord, data met een kleine kans een lang codewoord. Iedere H staat nu een beschrijving van D toe met lengte L(H)+L(D|H). Deze beschrijving codeert eerst het model H, en codeert vervolgens de data D met behulp van het model H. Vaak zal deze beschrijving korter zijn dan de oorspronkelijke lengte van D. MDL zegt nu dat we het model H moeten kiezen dat L(H)+L(D|H) minimaliseert. Er zijn hier twee uitersten. Als we voor H een "lege" hypothese kiezen zal L(H) laag zijn maar L(D|H) hoog, en zal er weinig geleerd zijn. Omgekeerd, als we voor H een hypothese kiezen die D heel precies beschrijft, dan zal L(H) hoog zijn en L(D|H) laag, en er zal nu "te veel" geleerd zijn (overfitting). Het kiezen van een H die de som van L(H) en L(D|H) minimaliseert, zoekt de gulden middenweg. De uiteindelijk gekozen H staat de meeste compressie van de data toe, en zal dus waarschijnlijk de meeste regelmatigheden in de data bevatten. Hierin herkennen we het bekende principe van de 14e eeuwse filosoof Occam: als er verschillende hypotheses of verklaringen consistent zijn met je data, kies dan de simpelste ("Occam's razor").
Zoals gezegd, Machine Learning gaat om het vinden van regelmatigheden in een gegeven set data. MDL baseert zich op de volgende cruciale observatie: Iedere regelmatigheid in de data kan gebruikt worden om de data te comprimeren (korter te representeren). Simpel voorbeeld: stel de data is de bitstring 011011011011011011011011011011. De regelmatigheid is duidelijk: het patroon 011 wordt steeds herhaald. Maar het herkennen van dit patroon staat ons ook toe om de data korter te beschrijven, namelijk als "10 keer 011". Gegeven een klasse van mogelijke modellen, stelt het MDL principe dat dat model moet worden gekozen dat de data het meeste kan comprimeren. Gegeven bovenstaande observatie, zal dit ook het model zijn dat de meeste regelmatigheden in de data gevonden heeft, en dat dus de beste generalisaties van de data geeft. Hoe kan een model H de data D comprimeren? Het eenvoudigst is dit uit te leggen aan de hand van Twopart MDL. Stel dat ieder model H ("hypothese") een kansverdeling is op de mogelijke data, dus P(D|H) is de kans op data D als H "waar" zou zijn. Stel dat L(H) de lengte van een beschrijving van H is in een of ander coderings-schema voor de hypotheses. De beroemde Shannon-Fano code uit de informatieheorie vertelt ons dat we kansen kunnen omzetten in codewoorden: gegeven de kansverdeling P(D|H), kunnen we een coderings-schema ontwerpen dat data D met een codewoord van lengte L(D|H)=-log P(D|H)
NVKI-Nieuwsbrief
Welke klasse van mogelijke modellen H moeten we nemen? In het ideale geval zouden we de klasse van alle "berekenbare" modellen nemen, en de lengte L(H) van een model uit die klasse zou gegeven worden door de zogenaamde Kolmogorov complexiteit K(H) van H (= de lengte van het kortste programma dat H "berekent"). K(H) is een
184
December 1998
objectieve maat voor de complexiteit van H. Onoverkomelijk nadeel hiervan: K(H) is niet berekenbaar, dus de ideale versie van MDL kan niet in een algoritme geïmplementeerd worden. Als we echter onze ambities wat laten zakken, dan kunnen we een beperktere modelklasse kiezen, toegespitst op het gebied wat we willen modeleren, en daarvoor een (efficiënt) berekenbaar coderings-schema maken. Er allerlei goede formele en informele redenen om de H uit je modelklasse te kiezen die L(H)+L(D|H) minimaliseert. Deze beperkte, praktische versie van MDL (relatief aan een keuze van modelklasse en coderings-schema) is Rissanens MDL en is het onderwerp van de dissertatie.
baseert op een prior distributie P(H) op de mogelijke hypotheses. Inderdaad functioneert P(H) bij de Bayesianen formeel gezien op ongeveer dezelfde wijze als de codelengte L(H) bij MDL. Interpretatief is er echter een wereld van verschil tussen beide. Terwijl prior probabilities een dubieuze status hebben (eigenlijk zijn het bijna metafysische aannames over wat waarschijnlijk is in de wereld, voorafgaand aan het zien van data uit die wereld), zijn MDL's codelengtes uitermate concreet en onafhankelijk van aannames over hoe de wereld zich gedraagt. MDL's interpretatie van waarschijnlijkheden als codelengtes (via ShannonFano's L(D|H)=-log P(D|H)) leidt tot een heel andere kijk op waarschijnlijkheden in het algemeen, en op statistische inferentie in het bijzonder.
MDL wordt vaak vergeleken met de zogenaamde Bayesiaanse inferentie, die zich bij model-selectie Een verwant leerprincipe is het Maximum Entropy (ME) principe. Dit zegt dat je uit verschillende consistente modellen het beste degene met maximale entropie (= grootste mate van uniformiteit) kunt kiezen. Dit principe is verbonden met de naam van de fysicus Jaynes en is altijd tamelijk omstreden geweest, hoewel het in de praktijk vaak succesvol is toegepast. Een oud resultaat zegt dat ME beschouwd kan worden als een speciaal geval van MDL, en Grünwald laat zien hoe. Aan de hand hiervan valt theoretisch beter in te zien wat de sterke en zwakke punten van ME zijn.
model te simpel en dus strikt genomen "onwaar" is. DEEL II. EXPERIMENTS WITH MDL Het tweede deel van de dissertatie beschrijft experimentele vergelijkingen van het MDL principe met ander leermethodes uit de klassieke en Bayesiaanse statistiek. De meeste gevonden experimentele verschillen kunnen uit de bestaande theorie verklaard worden. De hoofdconclusie van Deel II is dat geavanceerde vormen van MDL en Bayesiaanse methodes vaak verrassend goede resultaten leveren wanneer slechts kleine data-sets gegeven zijn. Bij grote data-sets werken alle methodes ongeveer even goed, maar dat is weinig verrassend wegens de wetten van de grote getallen (bij veel data gaan gemiddeldes van data bijna altijd naar hun verwachting toe). Opvallend is dat MDL bij kleine data-sets ook iets beter lijkt te werken dan het zeer-verwante-maar-subtiel-verschillende MML (Minimum Message Length) principe. DEEL III. REASONING UNDER UNCERTAINTY
De belangrijkste nieuwe bijdrage van de dissertatie aan de theorie van MDL en ME is het onderscheid tussen "safe" en "risky" toepassingen van statistische inferentie. Een van data geleerd model zal bijna altijd een (grove) simplificatie zijn van het gemodelleerde domein. Toch kunnen we met zo'n simpel model vaak goede voorspellingen doen, beter dan met een zeer complex model. Hoe kan dat? Grünwald geeft een aanzet tot een theoretische verklaring. Stel dat we een model van bepaalde data geleerd hebben, en daarmee sommige dingen willen voorspellen. Als de data waaruit we het model geleerd hebben op een bepaalde manier representatief was voor wat we willen voorspellen (dus voldoende relevante regelmatigheden bevat), dan zal het geleerde model meestal redelijk goede voorspellingen doen. Maar het is ook mogelijk dat het geleerde model gebruikt wordt om voorspellingen te doen over dingen die niet in de data terug te vinden waren. Het eerste geval is een "safe" gebruik van het geleerde model, het tweede is "risky". Grünwald analyseert voor een model dat op de MDL/ME manier geleerd is (en dus relatief simpel zal zijn), welke soorten voorspellingen "safe" zijn en welke "risky". Door je te beperken tot "safe" voorspellingen krijg je wiskundige garanties dat je voorspellingen betrouwbaar zijn en er waarschijnlijk niet te ver naast zullen zitten, ondanks het feit dat je
NVKI-Nieuwsbrief
Het derde deel van de dissertatie valt onder het logicistische paradigma van de AI en ontwikkelt een formele theorie van common-sense redeneringen over gebeurtenissen en veranderingen. Dit deel heeft minder direct met MDL en statistiek te maken. Desondanks is er een duidelijke link met delen I en II van de dissertatie, aangezien zowel leren-van-data als common-sense redeneren gaan over de vraag: wat is het beste om te doen in een situatie waarin we slechts onvolledige informatie hebben? Opnieuw is het antwoord op deze vraag statistisch geïnspireerd. Grünwald baseert zijn logisch systeem op een formalisering van de notie van "causaliteit" en op het Beginsel van de Voldoende Oorzaak (sufficient
185
December 1998
cause principle). Deze noties zijn afkomstig van statistisch werk van Judea Pearl. (De epiloog van Deel III geeft verdere connecties tussen systemen voor nietmonotone logica en statistische redeneersystemen.)
principe ons toestaat om een groot deel van zowel de successen als de mislukkingen van bestaande redeneersystemen te verklaren. CONCLUSIE
Het blijkt dat het voorgestelde systeem een generalisering is van verschillende bestaande systemen voor common-sense redeneren, zoals die van McCain & Turner, Lin, en Baral, Gelfond & Provetti. Deze bestaande systemen maken vaak impliciet gebruik van het sufficient cause principle, maar worden problematisch wanneer ze daarvan afwijken. Grünwald gebruikt zijn formalisering van common-sense reasoning om het aloude Yaleshooting probleem nog maar eens te lijf te gaan en een (hopelijk definitieve) oplossing daarvoor te geven. Hoofdconclusie van Deel III is dat het sufficient cause TAALKUNDIGE ANALYSE VAN ZAKELIJKE CONVERSATIES Dissertatie van Ans Steuten, TUD Promotores: J. Dietz en P. Hengeveld
Een indrukwekkend proefschrift: zo'n 300 pagina's, variërend tussen theorie en praktijk, tussen statistiek, informatietheorie, informatica, machine learning, logica en AI. Wegens haast nog wat slordig aan de randjes (zoals ook toegegeven), maar desondanks precies en tamelijk leesbaar. Grünwald is nu voor een jaar als post-doc verbonden aan Stanford University, waar hij met name de theoretische innovaties van Deel I van zijn proefschrift verder uit zal werken. bevestigen. Onderzoek vanuit LAP houdt zich bezig met de vraag welke taalhandelingen onderscheiden moeten worden, hoe deze samenhangen en waarom de ene interactie “beter” is gestructureerd dan een ander. Dit alles tegen de achtergrond van de mogelijke invoering van informatiesystemen. Vandaar dat men zich doorgaans beperkt tot zakelijke communicatie, in een bedrijf of in de handel tussen bedrijven.
Bespreking door Hans Weigand Katholieke Universiteit Brabant Wat heeft een informaticus die bezig is informatiesystemen in te voeren in bedrijven van doen met taalkunde? Meestal niets. Wanneer echter die informatiesystemen tot doel hebben de communicatie in de organisatie te ondersteunen, is een analyse van de bestaande communicatiepatronen nodig alvorens we iets kunnen zeggen over hoe de situatie, al of niet met behulp van informatietechnologie, kan worden verbeterd. Om een dergelijke analyse uit te voeren maakt de informaticus doorgaans gebruik van modellen. Een modelleermethode is de methode DEMO die is ontwikkeld door prof. Jan Dietz van de Technische Universiteit Delft.
PROEFSCHRIFT In juni is door Ans Steuten van de TUD een proefschrift verdedigd (Steuten, 1998) waarin geprobeerd wordt een brug te slaan tussen de taalkundige analyse van zakelijke conversaties en de modelleermethode DEMO. Centrale vraag is of een taalkundige analyse - eventueel uitgevoerd door een computerprogramma - een informaticus kan helpen bij het opsporen van de in het betreffende domein relevante taalhandelingen. Als sprekers zich zouden houden aan de performatieve zinnen zoals die door Searle worden gehanteerd - “ik beloof hierbij ...” - dan zou dit een triviale zaak zijn, maar in de praktijk is dit natuurlijk niet zo. De praktijk is dat taalhandelingen veelal indirect zijn en ook vaak impliciet, dus niet eens uitgesproken. Voor de taalkundige analyse heeft Steuten gebruik gemaakt van de Functionele Grammatica (Dik, 1989) en de Conversatie Analyse. Het empirisch materiaal is afkomstig uit telefoongesprekken in een hotel en bij een arbeidsbureau.
DEMO is gebaseerd op het zogenaamd Language/Action perspectief (LAP), een stroming in de informatica die als zodanig is geïnitieerd door Winograd en Flores (1986), en waarin de taalhandelingentheorie van Searle een belangrijke plaats inneemt. Volledigheidshalve is het goed te vermelden dat behalve in Delft ook LAPonderzoek wordt verricht in Tilburg en Eindhoven (zie bijvoorbeeld het proefschrift van Verharen, 1997).Waar de meeste benaderingen in de informatica geneigd zijn communicatie te beschouwen als de overdracht van gegevens, gaat het volgens het LAP om wat mensen in communicatie doen. Bijvoorbeeld, een kamer reserveren, een opdracht geven, of een opdracht
NVKI-Nieuwsbrief
BRUG TUSSEN DEMO EN FG/CA Met het door Steuten opgestelde geïntegreerde model is de brug tussen DEMO en FG/CA gelegd.
186
December 1998
Dit is het goede nieuws. Het slechte nieuws is echter dat het nog lang zal duren voordat taalkundige analyse met enige betrouwbaarheid uitspraken kan doen over illocuties, laat staan automatisch de taalhandelingen opleveren zoals DEMO die zou willen zien. Gegeven een DEMO model, is het wel mogelijk de verschillende stappen in de loop van de conversatie aan te wijzen. Maar om het model uit de conversatie te induceren is een brug te ver.
momenteel zijn binnen de FG gemeenschap over discourse representaties. Omgekeerd kan het LAP de inbreng van de taalkunde goed gebruiken. Men baseert zich daarin meestal op het klassieke model van de filosoof Searle, aangevuld of gecorrigeerd door filosofen als Grice en Habermas. Taalkundigen hebben in het verleden nogal wat kritiek geuit op Searle. Het is dan ook zeker nuttig de taalkunde te laten meepraten over de verdere ontwikkeling van het LAP.
Ondanks dit negatieve resultaat heeft het proefschrift zeker zijn waarde. Het onderzoek levert een bijdrage aan de discussies die er In het eerste hoofdstuk geeft Steuten aan wat het hoofddoel van het onderzoek is: een linguïstische onderbouwing van zakelijke conversaties vanuit het LAP perspectief en met het oog op automatische analyse van deze conversaties. Gegeven deze doelstelling is het eigenlijk jammer dat slechts twee cases zijn bestudeerd, waarvan vervolgens slechts één echt uitgewerkt is. Met behulp van een grotere verzameling (verschillende transacties, verschillende talen, ...) was het wellicht mogelijk geweest terugkerende patronen dan wel interessante onderscheidingen te vinden. Er wordt gesteld dat een gedegen begrip van zakelijke conversaties helpt bij het formuleren van richtlijnen voor een informatieanalIst die een communicatiemodel wil opstellen van een organisatie. Naar mijn mening wordt dit in het proefschrift onvoldoende onderbouwd. Uiteindelijk is een lexicon van werkwoorden nodig om de tekstanalyse tot resultaten te laten komen. Daarmee wordt het probleem dus eigenlijk verschoven. Hiermee wil ik niet zeggen dat een corpus van zakelijke conversaties niet een rol kan spelen in het analysetrajekt, maar het lijkt erop dat dit beter gebruikt kan worden ter validatie van een model dat op andere wijze verkregen is. Aangezien de kern van het communicatiemodel gevormd wordt door de zogenaamde essentiële acties, zou je kunnen beginnen met een actiemodel te maken voor de organisatie (of het domein in kwestie).
OVERZICHT Grammatica, waarmee individuele uitingen kunnen worden gerepresenteerd (naar syntactische, semantische en pragmatische inhoud). Aan de andere kant zijn dat de conversatie-analyse en discourse analyse waarmee de samenhang van uitingen in een discourse of dialoog kan worden weergegeven. Het gaat dan om zaken als turn-taking, feedback, opening en afsluiting, etc. HIERARCHISCH MODEL De verschillende theorieën worden geïntegreerd in een zogenaamd hiërarchisch model van zakelijke conversaties. Van boven naar beneden gelezen ziet dit model er als volgt uit. Het hoogste niveau is dat van de business transactie. Die bestaat uit een aantal fasen, en omvat ook het niet-linguïstische deel, de uitvoering van de acties waar het om gaat. Een business transactie bestaat uit een aantal exchanges (niveau 2). Een exchange is in het eenvoudige geval een combinatie van een actie van de spreker gevolgd door een reactie van de hoorder. De onderdelen van de exchange worden interactionele acties genoemd. Dat zijn per definitie de kleinste betekenisvolle eenheden van een conversatie. Voorbeelden zijn directieven (verzoek) en acceptaties. Interactionele daden worden gerealiseerd door illocutionaire daden. Daarbij gaat het dus om de talige uiting als zodanig. De relatie interactionele daad - illoctionaire daad is niet één-op-één, hoewel er prototypische verbanden zijn (bijvoorbeeld: directief - imperatieve zin). Het is ook geen deel-geheel relatie, maar zoals gezegd een doel-middel. In hoofdstuk zes van het proefschrift wordt aangegeven welke combinatiemogelijkheden er allemaal zijn. Zo kan een vraag als “Kunt u ..?” niet alleen maar een question, maar ook een directief realiseren.
In hoofdstuk twee wordt het LAP perspectief geschetst en in het bijzonder de DEMO benadering. DEMO beschouwt de speech acts niet in isolatie, maar als onderdelen van transacties c.q. conversaties, zoals de “conversation for action” (actagene conversatie). In hoofdstuk drie en vier worden dan de taalkundige theorieën ingevoerd die nodig zijn voor de taalkundige analyse. Aan de ene kant is dat de theorie van de Functionele
NVKI-Nieuwsbrief
187
December 1998
Maar aangezien er zoveel mogelijkheden zijn, is het heel moeilijk om hier richtlijnen uit af te leiden. Dit wordt nog eens verergerd doordat het model soms wat teveel detailleringen wenst (zoals het onderscheid tussen een informatieve en actagene exchange).
Wat verder nog lijkt te ontbreken als het gaat om de context is de juridische en organisatorische achtergrond. Zo maakt het om bij de case van de hotelkamerreservering te blijven wel uit of de reservering al tot stand komt in het telefoongesprek, of pas met de daaropvolgende fax. Wat ook een rol kan spelen zijn de instructies die telefonistes ontvangen ten aanzien van de vragen die ze moeten stellen. Het is duidelijk dat dergelijke context niet uit de conversatietekst zelf te halen is. Overigens zou het ook niet zo eenvoudig zijn om die context in DEMO te specificeren, het lijkt erop dat daarvoor de modellen (nog) ontbreken.
Aangezien de grammaticale structuur dus geen uitsluitsel biedt, wordt in het laatste hoofdstuk de hulp ingeroepen van, zoals gezegd, het lexicon, en de context. De laatste wordt in de vorm van scripts weergegeven. Nu zal ongetwijfeld de ervaringskennis van menselijke subjecten, al of niet in de vorm van scripts, het hen een stuk gemakkelijker maken om illocutionaire acties te duiden. Het is echter onduidelijk wat de rol van scripts is in de methode DEMO en in het beoogde automatische analyseproces. Al met al een aardig proefschrift. Het bevat een hoop bruikbare linguïstische hulpmiddelen. De waarde ervan ligt denk ik vooral in de link naar de empirie. Helaas is dat iets waar informatici zich in het algemeen niet zoveel aan gelegen laten liggen. Zij zijn meer geïnteresseerd in de gebruikswaarde. Die lijkt vooralsnog laag.
EVALUATIE Odijk (12-01), A. Bos (26-01), J. van den Akker (3003), F. Wiesman (7-05), E. Oskamp (13-05), A. Lodder (5-06), F. Ygge (6-06), A. Steuten (22-06), L. Combrink-Kuiters (10-09), A-.W. Dutler (22-09), G.-J. Zwenne (29-09), P. Grünwald (8-10), W. de Waard (9-10), D. Breuker (16-10), E.F. Tjong-KimSang (19-10), A. Hoekstra (20-10), H. Blockeel (1812), L. Dehaspe (21-12).
Bibliografie Dik, S.C. (1989). The Theory of Functional Grammar, Part 1: The Structure of the Clause. Foris, Dordrecht.
20 x waarvan 2 in 1997. Hendrik Blockeel (December 18, 1998). Top-down Induction of First Order Logical Decision Trees. Katholieke Universiteit Leuven. Promotor: Prof.dr. M. Bruynooghe, co-promotor: Dr. L. de Raedt.
Steuten, A. (1998). A contribution to the linguistic analysis of business conversations, within the Language/Action perspective. Proefschrift, Technische Universiteit Delft, Delft .
Luc Dehaspe (December 21, 1998). Frequent Pattern Discovery in First-order Logic. Katholieke Universiteit Leuven. Promotor: Prof.dr. M. Bruynooghe, co-promotor: Dr. L. de Raedt.
Verharen, E. (1997). A Language-Action perspective on the design of Cooperative Information Agents. Proefschrift, Katholieke Universiteit Brabant, Tilburg.
Ronald Leenes (January 7, 1999). Hercules of Karneades; Hard cases in Recht en Rechtsinformatica. Universiteit Twente. Promotor: Prof.mr. D.W.P. Ruiter, co-promotor: Dr. J.C. Hage.
Winograd, T, F. Flores (1986). Understanding Computers and Cognition - a New Foundation for Design. Ablex, Norwood NJ. ? ?
Ina Enting (January 14, 1999). Zovex, a Knowledge Based System to Analyse Factors Associated with Pig-health Disorders. Universiteit Utrecht. Promotor: Prof.dr.ir. M. Tielen.
Jaap van den Herik Universiteit Maastricht
Joeri Engelfriet (February 4, 1999). The Dynamics of Reasoning. Vrije Universiteit Amsterdam. Promotor: Prof.dr. J. Treur. H. Jurjus (3-12-97), K. van Belleghem (19-12-97), M.
NVKI-Nieuwsbrief
188
December 1998
Katholieke Universiteit Tilburg. Promotores: Prof.dr. J.E.J. Prins and Prof.dr. P.M.E. de Bra, co-promotor: dr. W.J.M. Voermans.
Marco Wiering (February 17, 1999). Explorations in Efficient Reinforcement Learning. Universiteit van Amsterdam. Promotor: Prof.dr.ir. F.C.A. Groen, co-promotor: Dr. H.J.H. Schmidhuber.
BENELOG 1998
Marnix Weusten (March 10, 1999). De Bouw van Juridische Kennissystemen. KRT: methodologie en gereedschap. Promotores: Prof.dr. A. Koers and Prof.dr. H.J. van den Herik.
Sandro Etalle Universiteit Maastricht Benelog 1998 was the Tenth Benelux Workshop on Logic Programming. It took place on Friday November 20, 1998 at the CWI - the Center for Mathematics and Computer Science - in Amsterdam, on the day after the closing of the NAIC'98. The editors and local organisers were Krzysztof Apt and Femke van Raamsdonk.
Pierre van de Laar (March 12, 1999). Selection in Neural Information Processing. Katholieke Universiteit Nijmegen. Promotor: Prof.dr. S. Gielen, co-promotor: Dr. T. Heskes. Luuk Matthijssen (April 9, 1999). Interfacing Between Lawyers and Computers: An Architecture for Knowledge-Based Interfaces to Legal Databases. First-time visitors to the Benelog may have had the impression that it was the 100th edition, rather than the 10th. People entered the conference as if they had been there the day before, much like students entering their lecture hall. Before the start, everyone took a coffee and chatted a bit with everyone else. The atmosphere was very informal and charming.
Curry and to answer to the questions of myself and others. Every group gave one or two presentations showing their very latest research results. Also, for each group, some time was reserved for presenting an overview of the research carried out in the past year. In this way it was possible to have - within one day - a good insight into the hot topics while still retaining a global overview of the research done in the various institutions. There were interesting lectures (the possible exception: my own presentation, where I misspelled "append" twice; a crime punishable with life-long exile from the net), always followed by interesting debates.
I should mention that Krzysztof (I do have a macro for writing his name) is known in the research circuit not only for his outstanding scientific results, but also for his concise and crystal-clear writing style. His welcome oration, precisely on time, reflected these virtues: "It is nine thirty, shall we start?". And the presentations - all presentations - magically started exactly on time, as if it was the most normal thing in the world.
The tenth Benelog was a very inspiring and interesting experience. What is the secret? Maybe it is the simplicity of the formula or maybe it is just the participation of good groups with good leaders. I don't really know, but I wished that more workshops and symposia were like it. I also wish that next year's edition, that will take place in Maastricht on November the fifth, will be able to replicate the success of this year. I am looking forward to meeting you at the 11th Benelog!
Benelog is actually more a discussion forum than anything else. This year it attracted the logic programming groups of six outstanding research centers. Besides the already mentioned CWI, the University of Leuven was represented by the groups of Maurice Bruynooghe, Bart Demoen and Danny de Schreye. There were the groups of Yves Deville, of the University of Louvain-La-Neuve and the ones of Jean-Marie Jacquet and Baudouin Lecharlier, of the University of Namur. Luxembourg was well represented as well by Raymond Bisdorff, of the Centre Universitaire of Luxembourg. The invited speaker was Michael Hanus, from the Technical University of Aachen. He gave a talk on Multi-Paradigm Declarative programming in Curry. Hanus remained for quite some time after the closing of the workshop to patiently show how to solve some problems in
o g o l E I T C E S S K I S
NVKI-Nieuwsbrief
Readers interested in the program or in the papers that have been presented at the 10th Benelog may consult http://www.cwi.nl/~femke/ benelog98/ benelog98. html. Preliminary information about the 11th Benelog can be found at http://www.cs.unimaas.nl/ ~etalle/benelog99/ benelog99.html.
189
December 1998
SYSTEMS Nico Roos Universiteit Maastricht From November 30 till December 4, 1998, the SIKS course on Interactive Systems and Multi Agent Systems was organized. This course consisted of two separate courses: two days on Interactive Systems and three days on Multi Agent Systems. Both courses will be discussed below.
SIKS COURSE ON INTERACTIVE SYSTEMS AND MULTI AGENT
INTERACTIVE SYSTEMS
The course on Interactive Systems was given by Gerrit van der Veer and Charles van der Mast. They presented a road from a current working situation, via a desired working situation, to a new situation in which an interactive system has been implemented and introduced. The first step in this process is analyzing the current situation. Gerrit van der Veer presented several analytical methods from psychology as well as from ethnography to analyze the current situation. The psychological methods aim at determining the explicit knowledge of human experts, while the ethnographical methods aim at determining the implicit knowledge of a group. Using these methods, we get a picture of what people are doing in a situation, how they work together, and what knowledge they are using.
interactions with the user. An other important consideration in the design of an interactive system is the choice of a consistent metaphor. A wrong or inconsistent metaphor can lead to confusion. Finally the choice of the interaction style, for example, whether or not to use direct manipulation, and the interactions themselves are important. To get the above mentioned issues right, evaluation is important. Design guidelines can help in this stage of the development process. Another aspect to be evaluated is the cognitive complexity of the interactive system. Several techniques that can be used, were briefly discussed. The importance of this stage was underlined with a video in which an airplane crash was reconstructed. The presentation of information in the newly designed cockpit, played an important role in the crash.
After analyzing the current situation it is time to look at the desired new situation. One has to envision the new desired situation. To determine whether one’s vision makes any sense, it is important to evaluate it. Playing the new situation in several scenario’s is a technique that can be used in an early stage is. At the SIKS course a video from 1989 was shown in which HP showed the office of 1995. An interesting thing in this video was that it presented personal agents (the subject of the second part of the course). Other forms of evaluation, such as prototyping, as well as the different ways of using a prototype were also addressed.
The last aspect presented in this course concerned the management of the development of an interactive system. Charles van der Mast emphasized that the traditional waterfall method is not suited for the design of an interactive system. Because of the importance of regular evaluations during the whole development of an interactive system, several stages have to be repeated. The revision of earlier stages becomes difficult if the people who carried out that stage are no longer available.
Charles van der Mast discussed the design of the interactive system that is to be used in the desired situation. An important issue in such a design is the fact that humans see a computer, the screen and keyboard, as a social actor. This implies that the interactive system must be polite in its
NVKI-Nieuwsbrief
MULTI AGENT SYSTEMS The course on Multi Agent Systems was given by Catholijn Jonker, Peter Braspenning, John-Jules Meyer, Han la Poutré, Hans Weigand, Floris Wiesman and Etiënne Mathijsen. They gave an
190
December 1998
overview of the new developments in the field of multi agent systems. Since Multi Agent Systems is a relatively new field, there is yet little consensus on what an agent is.
intentions, commitments, awareness, and so on. Peter Braspenning discussed these agents in more detail but did not formally define them. Instead he pointed out that there is no consensus and that the best thing that we can do is looking at examples.
Catholijn Jonker presented the notion of a weak agent. Such an agent should exhibit four types of behavior; autonomous behavior, social behavior, reactive behavior and pro-active behavior. All these behaviors must be present in the eyes of an external observer. What it means for an agent to be autonomous or pro-active is an issue that caused much debate. A strange thing was that these properties were no longer considered necessary properties for designing an agent.
John-Jules Meyer discussed the BDI (Beliefs, Desires and Intentions) model, which gives a formal characterization of a strong agent based on modal logic. The idea behind this characterization is that an agent has beliefs about the world, it has wishes it desires to fulfill, and it has the intention to actually fulfill some selected subset of the wishes. The wishes the agent desires to fulfill, need not form an consistent set. Based on the current beliefs some of these wishes are selected to become intention, the agent tries to realize.
Beside the notion of a weak agent, there is also the notion of a strong agent. According to Catholijn Jonker, a strong agent is a weak agent with additional properties such as beliefs, desires, When an agent tries to realize its intentions, it should not re-evaluate its intentions each time it tries to perform an action. Re-evaluating the actions requires computational resources. As a consequence the agent might spend more time on re-evaluating its intentions than on performing actions that realize these intentions. So in the context of limited computational resources, it is irrational if an agent re-evaluates its intentions too often. This requirement is denoted by the term commitment. Peter Braspenning discussed the importance of commitment and the various ways in which it is formalized in the literature.
agent that performs a task by manipulating the objects through their operations. He also sees it as a way to encapsulate existing Information Systems, and in this way enable them to collaborate. Besides this diversity of agent definitions, there was also little disagreement on the meaning of concepts such as autonomy, beliefs, desires, intentions, commitment, and so on. Especially, participants working on multi agents made objections. Some even pointed out that one should not use common sense psychological concepts to characterize properties of agents. This would not justify the proper meaning of these concepts. I personally see no objection against using terms in the context of agents. One must, however, keep in mind that in the context of agents these terms have a technical meaning, which has some correspondence with their meaning in the term in the real world.
Han la Poutré presented a different picture of an agent. In his view an agent is a module that models a part of the behavior of real agents such as humans or companies. He uses these agents to make predictions about the behavior of humans and of economic systems. An important aspect of these agents is that they evolve in time. Genetic algorithms are used for this purpose.
So what is an agent? In my opinion it is a bad thing not to have a proper definition of an agent. Since agent technology is a popular topic, without a proper definition one can easily get wrong expectations, which may result in disappointments. Therefore, I will try to formulate what I think an agent is, based on the views presented at the SIKS course. To those of you that think that I am partially or completely wrong, I propose that you send your view to the NVKI newsletter.
Yet another view was presented by Hans Weigand. In his view an agent presents a new level of modeling. In a relational database, relations between objects are described. Object oriented modeling goes one step further by describing an object by its relation with other objects and by adding operations that can be performed on the object. To perform a task, in objects oriented modeling one often has to introduce a problem object with a do-it operation that performs the task by activating operations of other objects. Hans Weigand argued that it is more natural to have an
NVKI-Nieuwsbrief
An agent is a modeling concept. It is used to
191
December 1998
describe an object that can make perceptions, that can perform actions, that has knowledge about the environment and that tries to fulfill goals in a rational way. Again these are properties seen by an outside observer. Knowledge and goals need not be represented in an explicit way in the agent. Notice that this definition is similar to Newell’s Knowledge Level. This definition of an agent seems consistent with a weak agent presented by Catholijn Jonker and with Hans Weigand’s description of an agent. Properties, such as autonomy, can be present in different grades. Autonomy, for example, can vary between a standalone system that does not require external interference to fulfill its goals if the environment is changing, and on the other hand agents that learn from experience. The BDI model is a refinement of an agent as a modeling concept. Beliefs are an explicit representation of knowledge, and desires and intentions are a refinement of goals. Multi Agent Systems are systems consisting of a set of agents that interact by observing the (results of) actions of other agents. A special form of In order to let agents communicate, a common language is required. John-Jules Meyer and Hans Weigand discussed several agent languages. Hans Weigand pointed out that communication should be put in a larger setting, called a scenario. A scenario is a sequence of transactions in which communi-cation is one of the components. Scenario’s are important in electronic commerce where an agreement to buy/sell something should be followed by the obligation to deliver and the obligation to pay. It seems that on this aspect of the interaction between agents, the field of Interactive Systems could have a valuable contribution.
interaction is communications. The most interesting agents are those that, as a group, perform a task or fulfill some goal. Due to this, the agents must interact, possibly by communications. Catholijn Jonker demonstrated several of these agent systems. One demonstration concerned the load balancing of electricity use. Here consumer agents negotiate with the supply agent over the reduction of electricity use. A consumer agent is a kind of personal assistance that may negotiate as a representative of the actual consumer according to the consumer’s preferences. Another interesting example was the call center in which the call center agent tries to make an appointment for a client. To do this it asks the agent of a particular office to schedule an appointment for this client. The office agent then has to negotiate with the personal agents of the employers in order to schedule the appointment.
ANTS' 98 FROM ANT COLONIES TO ARTIFICIAL ANTS: FIRST INTERNATIONAL WORKSHOP ON ANT COLONY OPTIMIZATION Brussels, Belgium, October 15-16 Katja Verbeeck Free University Brussels (VUB)
Peter Braspenning discussed the attemps of the FIPA (Foundation of Intelligent Physical Agents) to develop an agent architecture for the internet. The FIPA is not concerned with the architecture of an agent itself but with the platform on which an agent lives and the language the agent uses to communicate. For the latter purpose the FIPA developed an agent language called ACL.
Mid-October, the first international workshop on ant colony optimization took place at the Université Libre de Bruxelles (ULB), organized by Marco Dorigo (Research Associate at the FNRS, the Belgian National Fund for Scientific Research) and colleagues of ULB's IRIDIA lab.
Looking back on the Multi Agent Systems course, it seems that one aspect was missing; namely the social behavior of agents. Agents can collaborate or compete with each other in many different ways. Some forms of collaboration were shown in the course. It seems, however, that more can be said about this topic.
INTRODUCTION
S K I S o g o L
Ant colony optimization (ACO) studies artificial systems that are inspired by real ant colonies behavior. The resulting systems seem to be very well suited for discrete optimization. Problems like travelling salesman, sequential ordering, quadratic assignment, partitioning, graph coloring, routing in communication networks, and so on, are already
NVKI-Nieuwsbrief
192
December 1998
addressed successfully.
path. Eventually, this autocatalytic process causes all the ants to take the shortest path.
The main observation on which ACO is based is that real ants are capable of finding shortest paths from their nest to food sources and back. They are even capable of adaptation, that is, of finding new shortest paths when their environment changes. They can perform this behavior thanks to a simple pheromone laying mechanism. In fact, while walking ants deposit some amount of pheromone on the ground. When ants move from their nest to the food source they move mostly randomly, but their random movements are biased by pheromone trails left on the ground by preceding ants. Because the ants that initially choose the shortest path to the food arrive first, this path will be seen as more desirable by the same ants during their journey back to the nest (this is called "differential length effect"). This in turn will increase the amount of pheromone deposited on the shortest A quick overview of the state-of-the-art in the field, which included Ant System, the first ACO system introduced in 1991 by Marco Dorigo, as well as a number of more recent ACO algorithms, was given in a tutorial on Wednesday evening by Dr. Dorigo himself. The actual workshop started on Thursday morning. The first speaker, Dr. Owen Holland from the university of West of England, immediately grasped the attention of the audience by raising some interesting questions, like: What are the principles of ACO algorithms? How do they work? Why do they work? What does control them? What is the difference between ACO algorithms and other algorithms based on physical, chemical (diffusion), electrical (electrical flows), or even human behaviour? Answers to these questions were expected to come up during the workshop.
Artificial ants take advantage of the differential length as well as of the autocatalytic aspects of real ant behavior to solve discrete optimization problems. Artificial ants are software agents that move on a graph, and that modify some variables associated to graph elements so to favor the emergence of good solutions. In practice, to each graph's edge is associated a variable, called pheromone trail for analogy with the real ants. Ants add pheromone to those edges they use and by so doing they increase the probability with which future ants will generate good solutions. Artificial ants, as real ones, move according to a probabilistic decision policy biased by the amount of pheromone trail they "smell" on the graph edges. systems were tested against known routing algorithms, with very good results. The second part of the session was dedicated to applications to combinatorial optimization. The domains of application are very broad here: from bus driving scheduling to mobile telephony (the problem of covering regions), till water irrigation distribution. CONCLUSION Certainly not all the questions posed by O. Holland were answered during this two-day workshop. Nevertheless, some of the results presented are excellent, and give researchers a good motivation to pursue further investigations in this new exciting area. More information on ants and on ant colony optimization can be found at the ACO homepage: http://iridia.ulb.ac.be/dorigo/ACO/ACO.html. The workshop did not include proceedings, but a selection of the papers will be published in the "Future Generation Computer Systems Journal", special issue on Ant Colony Optimization, Elsevier North-Holland, next year. The Ants 2000 workshop will most likely be held again at ULB, Brussels, in September 2000.
REAL ANT MODELS The word was then given to biologists and people who were working on modeling real ants and on the study of simulations of these models. Some of the topics people are working on are: How does a colony of ants allocate tasks in a dynamical environment without a coordination center? Or, how do they move? The resulting models were simulated on computer, and their results were tested against results obtained by real ants observation.
AI IN VIDEO GAMES
APPLICATIONS
Stephane Assadourian AI Researcher, APPEAL / IMLABS
The first part of the next session was dedicated to applications to communication networks as routing and load balancing. Several routing algorithms were presented by different people (L.J.M. Rothkrantz, G. Di Caro & M. Dorigo, D. Snyers). Some of these
Appeal Software is a video game development company based in Belgium in Namur (Namen). It was founded in 1995 by Yann Robert, Yves Grolet and Franck Sauer, who have 12 years of
NVKI-Nieuwsbrief
193
December 1998
experience in the game industry. Currently, we are busy finishing a game called Outcast, which will be one of the most advanced real time 3D action/adventure games ever produced on PC. The aim of Appeal is no less than giving the gameplay a new dimension. This involves the development of new technologies, as done within the Himalaya project. This project breaks down into several sub-projects, each relating to a certain field of expertise including rendering, physics, and Artificial Intelligence, which will be the subject of this article.
monsters, but, depending on the design of a game, they may take other shapes. We're talking virtual world, so you can follow your imagination as far as it goes... A NEW AI ENGINE Himalaya's AI sub-project is called Lhotse, and it is focused on a new AI engine for future games. This engine is constrained by many parameters since it is embedded in a game application. Such parameters are expressed in terms of CPU time allowed. An important difference compared to academic research in AI is that an engine will not be built before its consumption of processor time is known to be adequate. Therefore the implementation phase is preceded by both a design phase and a test phase.
Concerning Artificial Intelligence, Appeal's R&D department is interested in the simulation of behavior of what we call NPCs. NPC stands for Non Playable Character, meaning everything the player does not control. NPCs should be able to reason, plan, take actions, react to events, and show emotions. Typically these are people or The aim of Lhotse is to give birth to a high level tool that will simulate the major aspects of realistic behavior, where realistic is taken to mean believable. During planning for example, we are not looking for the best solution, but rather for a solution which produces believable behavior. This is another difference compared to academic research in AI. What matters in this type of game, is the effect on the player.
same way and do things according to their mood, which is determined by emotions. COMPETING ACTIONS Finally, to allow for parallel processing of actions, an agent based architecture is used. Actions are designed as operators, which require resources. The TAKE operator e.g. requires (at least) a hand to pick up the object. Actions may be performed simultaneously, so one may see an NPC taking an object while looking in another direction, or performing more complex behaviors such as running to hide while targeting, shooting, and taking fire. Clearly, this may result in a competition for resources between different actions.
Within Lhotse, two objectives are important. First, NPCs should have an individual semantic interpretation of what is happening in the world and of the meaning of (re)actions of others. Furthermore, parallel processing of actions is required since a game is a real time application. Therefore we are designing an engine that will allow us to code behavior via planning. Thus, the behavior of an NPC is not hard-coded. Knowledge needed for planning is encapsulated in the objects of the world. These objects support constraints. To give an example, there will be rocks that some NPCs are able to lift and some aren't, according to their strength. As a direct consequence of the constraints specification feature, there is a need for cooperation if NPCs are to perform tasks they can not handle alone. A communication protocol will be developed that will allow NPCs to ask for help and accept or deny these requests according to parameters such as mood, or emotional state. Indeed, emotions will allow us to capture an important part of behavior. NPCs are characterized by many features, allowing to create characters that look different and act in different ways because they are not sensitive to events in the
NVKI-Nieuwsbrief
The resources feature only solves the problem of deciding which actions are allowed to be performed for a given NPC, but not which action to select from several competing ones. An Applicable Operator List is created for the Active Goal. Based on various heuristics and priorities attached to each branch, it can then be decided which operator should be developed. Thus, we are subgoaling from operators to operators, propagating potential constraints along the tree. One might wonder what should happen if an agent wants to pick up an object and open a door (also requiring the hand) at the same time. Let's assume that both the object and the door to be opened are close enough to perform the action. If both hands are free, there's no competition for resources between the operator
194
December 1998
TAKE and the operator OPEN_DOOR, since all required resources can be allocated perfectly. If at least one hand is not free, there's a competition for the resource left hand (or right hand) and the winner will be determined according to heuristics such as which operator is in the highest priority branch, or which one satisfies the largest number of goals.
with students for either a Ph.D. work or an internship. So do not hesitate to contact us at
[email protected], for a possible collaboration.
This was a general view of the project and if you want to know more about it, feel free to visit the following web page, http://www.appeal.be/research/ Lhotse/lhot_par.htm, dedicated to the Lhotse project.
Walter Daelemans & Steven Gillis University of Antwerp
ARTIFICIAL INTELLIGENCE RESEARCH AT CNTS
CNTS (Centrum voor Nederlandse Taal en Spraak) is the Centre for Dutch Language and Speech of the University of Antwerp (UIA). CNTS specializes in applying computational modeling and AI techniques in the fields of computational linguistics, language technology, artificial intelligence and computational psycholinguistics. In addition, there is also a fair amount of more general linguistics, phonetics, and psycholinguistics research. library and information science students looking for an introduction to the AI approach to computational linguistics.
Also, if you have some remarks feel free to contact me at
[email protected]. Finally, I've got to point out the fact that Appeal is clearly willing to support any kind of collaboration Computational Linguistics has a long tradition at UIA. Luc Steels (now VUB Brussels and Sony Computer Science Lab Paris) started his career in AI in Antwerpen in the seventies with a doctorate in computational linguistics. His student, Koenraad De Smedt (now in Bergen, Norway) started his work on object-oriented knowledge representation for language processing in Antwerpen before moving to Nijmegen in the early eighties. Willy Martin (now in Amsterdam, VU) and his co-workers developed tagging (word class disambiguation), lemmatization, and lexicographic technology, mainly for English, throughout the eighties, and started with a set of courses in computational linguistics. The nineties saw the start of the CNTS, and a shift of research activities to the application of machine learning techniques in language technology, linguistics, and psycholinguistics. We believe that the application of machine learning and statistical pattern recognition techniques will make a difference in these three areas of language research. Currently, the Artificial Intelligence research activities in CNTS are directed mainly by Steven Gillis and Walter Daelemans, and the group working on these issues has recently grown to about 8 researchers, thanks to financial support from FWO Vlaanderen, IWT, and European funding. There is now also a fairly complete curriculum in Natural Language Processing allowing linguistics students to specialize in language technology both at an undergraduate and postgraduate level. The courses are also used by computer science and
NVKI-Nieuwsbrief
In the remainder of this text, we will describe our main research topics and results. Within Machine Learning we have focused our attention (in close cooperation with the ILK research group in Tilburg, http://ilk.kub.nl/) on memory-based learning. This approach is based on the general idea that cognitive tasks (e.g. language processing) originate from the direct reuse of experiences of the task rather than from the application of rules or other abstractions extracted from the experience. It is interesting to see that the approach has been advocated under different guises in Artificial Intelligence (e.g. work on instance-based learning, memory-based reasoning, case-based reasoning, etc.), but also in the "linguistic underground" (as an alternative to Chomskyan linguistics, e.g. Skousen, Bybee, Derwing, Ohala and others). The last decade, we have investigated this hypothesis in the context of the three research fields mentioned earlier, viz. language technology, computational psycholinguistics and linguistics. In language technology, where knowledge acquisition bottlenecks have always hindered practical application of rule-based computational linguistics, we have shown that for problems such as speech synthesis and text analysis, the memory-based approach is often
195
December 1998
superior in accuracy to alternative rule-based and statistical approaches. It also allows very fast development of language processing modules to be used in language technology applications.
research has only recently started systematically, results indicate that memory-based computational models can indeed mimic and motivate the `rule-based' behaviour (including regularization and irregularization) that is observed so often in human language processing. As such, memory-based computational models may become an interesting alternative to both dual route and connectionist single route models of human language processing.
Y T I S R E V I N U E H T T A R A E Y A
In computational psycholinguistics, our group has begun investigating the psychological relevance of memory-based language processing by comparing the output and errors made by our algorithms to that of children in first language acquisition, and adults performing linguistic tasks. Although this In linguistics, finally, we have shown that the memory-based model matches well with current thinking in cognitive linguistics. As memory-based learning crucially relies on the availability of huge amounts of training materials, i.c. linguistic lexica and corpora, CNTS is currently involved in several projects that aim at collecting huge databases of spoken language and is involved in research efforts to enrich these databases with linguistically relevant annotations. For instance, CNTS is participates in the collection and annotation of the 'Corpus Gesproken Nederlands', an important initiative of the Dutch and the Flemish governments to collect 10 million words of spoken Dutch that will eventually be available to the research community. CNTS is also involved in a Dutch-Flemish project that aims at collecting a phonetically rich and balanced database of Standard Dutch. In addition, CNTS houses the European headquarters of CHILDES, an organization that archives and linguistically annotates corpora of children's (and adults') spontaneous conversational speech. In CHILDES corpora from more than 20 different languages are available.
Y R A G L A C F O y r sg a d r l a a a C g f n jo i y Wi t s kr e e i v N i n U : m a d r e t s m A r e d i s n o C n o e l p o e p y n a m , r e t n e c y t i c d l o n a h t i w y t i c t c a p m o c a d i m u h a n i g n i v i l s t n a t i b a h n i 0 0 0 , 0 0 7 t u o b a , s t e e r t s e h t . l e v e l a e s e v o b a y l e r a b r o r e d n u e t a m i l c e t a r e d o m : y r a g l a C r e d i s n o C n r e d o m a h t i w y t i c g n i d n a p x e r e v e , t u o d a e r p s y r e v a g n i v i l s t n a t i b a h n i 0 0 0 , 0 0 7 t u o b a , s t e e r t s e h t n o e l p o e p t o h r o ) C o 5 2 w o l e b ( g n i z e e r f r e h t i e , r i a y r d y r e v n i . l e v e l a e s e v o b a s e r t e m 0 0 0 1 t u o b a t a , ) C o 5 2 + r e v o ( d l u o w y h w o S . s e t i s o p p o r i e h t t s o m l a e r a s e i t i c e s e h T n i g n i v i l r e d i s n o c m a d r e t s m? Ar a e y l l u f a r o f y r a g l a C
n i g n i v i l y d o b y n a
T R A T S E H T e j i r V
e h t t a h c r a e s e r D h P y m h t i w d e t r a t s l l a t I
e c n e g i l l e t n I l a i c i f i t r A e h t t a , m a d r e t s m A t i e t i s r e v i n U e h d n a p u o r g t a h t f o r o s s e f o r p e h t s i r u e r T n a J . p u o r G I . h c r a e s e r y m d e s i v r e p u s r e i z a r B s e c n a r F d n a l a n o i t i s o p m o c f o n g i s e d e r n o h c r a e s e r n i d e t a p i c i t r a p c i t s o n g a i d a f o n g i s e d e r e h t o t d e i l p p a , s e r u t c e t i h c r a a f o n o i t a c i f i d o m f l e s
e h t d n a m e t s y s g n i n o s a e r . m e t s y s t n e g a i t l u m
e g d e l w o n K e h t g n i t i s i v s a w I , 6 9 9 1 r e b m e v o N n I , f f n a B f o n w o t n a i d a n a C e h t n i p o h s k r o W n o i t i s i u q c A t n e g a i t l u m
g n i y f i d o m f l e s a n o r e p a p a t n e s e r p o t
The web-site of CNTS can be visited at http://webger-www.uia.ac.be/webger/ger/cnts/main .html For more information, contact Steven Gillis (
[email protected])
, h c r a e s e r D h P y m f o d n e e h t r a e n n e h t s a w I . m e t s y s e h t f o s t n a p i c i t r a p e r i u q n i y l e t i l o p o t d e t r a t s d n a e l b a l i a v a t u o b a g n i h t y n a w e n k y e h t r e h t e h w p o h s k r o w y n a f o w o n k t o n d i d y n a m e l i h W . s n o i t i s o p c o d t s o p
D A O R B A I A
NVKI-Nieuwsbrief
w e f , s r a c y n a m , s l l a m
g n i p p o h s , a e r a n w o t n w o d
Apart from memory-based learning, we have also investigated the use of symbolic rule learning (decision tree learning, rule induction, ILP) as a tool in linguistic research. In this research, we are mainly interested in using machine learning techniques to discover and evaluate linguistic hypotheses and categories (formulated as generalizations). We apply these techniques mainly to discovering and evaluating theories about phonology and morphology.
196
December 1998
e t i u q s a w n g i s e D & I A d n a s m e t s y s t n e g a i t l u m
o t e m d e t i v n i y l e t a i d e m m i m e h t f o e n o , s n o i t i s o p n e p o
e h t t a c o d t s o p e h t t r a t s o t d e d i c e d s a w t I . e t a i r p o r p p a
n i e c n e i c S r e t u p m o C f o t n e m t r a p e d e h t t a c o d t s o p a o d
. r a e y e n o f o n o i t a r u d e h t r o f , 8 9 9 1 f o g n i n n i g e b
o s l a I , r e v e w o H ! e s r u o c f o d e r e t t a l f s a w I . . . y r a g l a C
t r a t s t o n d l u o c I o s , D h P y m h s i n i f o t d e d e e n
n a : d e s i n a g r o y l t n e i n e v n o c y r e v s i p u o r g e h t f o b a l e h T
, s e i t i l i b i s s o p e m o s d e n i l t u o y l f e i r b y e h T . y l e t a i d e m m i
g n i d n u o r r u s s e c i f f o d n a , s e l c i b u c i m e s h t i w e c a p s n e p o
n i h c r a e s e r n w o r i e h t d n a n o i t a c u d e e c n a t s i d , . g . e
d n a t e e m y l i s a e n a c s t n e d u t s d n a f f a t S . e c a p s n e p o s i h t
. s m e t s y s t n e g a i t l u m
y b e s u r o f y l t s o m m o o r g n i t e e m A . s e u s s i s s u c s i d
: h c r a e s e r y r o t a p i c i t r a p e h t n i s d i a b a l s i h t f o s r e b m e m
g n i r e e n i g n E e r a w t f o S e h t d e t i s i v I 7 9 9 1 r e b m e t p e S n I
s s u c s i d o t d e y o l p m e e b n a c a i d e m i t l u m t r a e h t f o e t a t s
r e t u p m o C f o t n e m t r a p e d e h t f o t r a p , k r o w t e N h c r a e s e R
e h T . s e p y t o t o r p d n a , k r o w g n i o g n o , s r e p a p
s s u c s i d o t , ) C o U ( y r a g l a C f o y t i s r e v i n U e h t t a e c n e i c S
y r e v a s i h t s e k a m h c r a e s e r o t h c a o r p p a e v i t a r o b a l l o c
e h T . y r a g l a C n i g n i v i l r o f s n o i t a r a p e r p e k a m d n a b o j e h t
e k i l h c u m e r e h p s o m t a n a n i k r o w o t e c a l p e l b a t r o f m o c
y l t n e r r u c d n a g n u o y s i , w a h S d e r d l i M y b d e d a e h s i p u o r g
. p u o r g ' d l o ' y m n i o t d e s u s a w I
r e h t o d n a y r t s u d n i h t i w s e i t g n o r t s s a h t I . g n i d n a p x e n e k o p s e v a h I t i s i v s i h t g n i r u D . C o U e h t n i s t n e m t r a p e d
H C R A E S E R E H T
y l t n e r r u c e r a o h w , s r e b m e m f f a t s w e n l a r e v e s h t i w
t n e g a i t l u m , s p a m t p e c n o c n i h c r a e s e r n i d e v l o v n i
n i d n u o r g k c a b y M . c t e , t n e m e g a n a m w o l f k r o w , s m e t s y s
: s a e r a e e r h t n i s i n i d e v l o v n i y l t n e r r u c m a I h c r a e s e r e h T
S E C N E R E F F I D
e c n a t s i d , g n i r e e n i g n e s t n e m e r i u q e r r o f s l e d o m s s e c o r p
a n i ( s m e t s y s t n e g a i t l u m r o f s e r u t c e t i h c r a d n a , n o i t a c u d e
s i h t n e e w t e b s e c n e r e f f i d w e f a e t i u q e r a e r e h T
s i e r e h t s a e r a e e r h t l l a n I . ) g n i t t e s g n i r u t c a f u n a m
. m a d r e t s m A n i t i e t i s r e v i n U e j i r V e h t d n a y t i s r e v i n u
s e v i g h c i h w , s t n e d u t s d n a s r e h c r a e s e r h t i w n o i t a r o b a l l o c
: C o U e h t t a ' g n i t e g d u b ' e r o m h c u m s i e r e h t l l a f o t s r i F
. s n o i n i p o d n a s d n u o r g k c a b e s r e v i d f o x i m t n a s a e l p a
t i , d e e t n a r a u g t o n s i k s e d r u o y n o r e t u p m o c a g n i v a h
e n O . h c r a e s e r f o s a e r a e e r h t l l a n i e d a m s i s s e r g o r P
, s l l a c e n o h p e l e t l l a , y l r a l i m i S . r o f d i a p e b o t s d e e n
e c n a t s i d a d n a l a c o l a h t o b s a t h g u a t g n i e b w o n s i e s r u o c
: t e g d u b a n o d e c a l p e b o t d e e n . c t e s e i p o c o t o h p
r e g g i b a r e t s e m e s t x e n , t n e m i r e p x e n a l l i t s s i s i h T . e s r u o c
y r o t a d n a m e d i w y t i s r e v i n u e h T . l a n o s r e p r o h c r a e s e r
r a l i m i s a n i p u t e s e b l l i w ) s t n e d u t s f o r e b m u n n i ( e s r u o c
y r a l a s r u o y t c a p m i n a c h c i h w s r o t c u r t s n i f o s n o i t a u l a v e
t n e m t r a p e d e h t f o p u o r G e c a f r e t n I c i n o l o H e h T . r e n n a m
7 , y a d r e p s r u o h 4 2 e h T . e c n e r e f f i d r e h t o n a e r a e s a e r c n i
n o h c r a e s e r e h t n i s e t a p i c i t r a p g n i r e e n i g n E l a c i n a h c e M f o
o n : t e s s a t a e r g a s i y t i s r e v i n u e l b i s s e c c a k e e w r e p s y a d
d n a , c i t s i l a e r y n a m e d i v o r p y e h t : s m e t s y s t n e g a i t l u m
y a m h c i h w s r u o h g n i k r o w o t d e n i f n o c u o y e r a r e g n o l
r o f s l e d o m s s e c o r p n o k r o w e h T . s e s a c e s u , e f i l l a e r
t o n m ' I . s e l u d e h c s h c r a e s e r / k r o w r u o y h t i w t c i l f n o c
n o i t a r o b a l l o c t n a s a e l p a s i g n i r e e n i g n e s t n e m e r i u q e r
t s u j , y a d a s r u o h 4 2 k r o w d l u o h s e n o t a h t g n i s o p o r p
n i p u o r g d l o y m d n a p u o r g t n e r r u c y m n e e w t e b
s t i d n a y t i s r e v i n u e h t o t s s e c c a f o d n i k s i h t g n i v a h t a h t
n o g n i k r o w m ' I , C o U e h t t a b o j y m o t t x e N . m a d r e t s m A
t e g , n i p e t s y l f e i r b o t r e i s a e h c u m t i s e k a m t n e m p i u q e
e l b a t ' n s a w I s a , s d n e k e e w d n a s g n i n e v e e h t n i s i s e h t y m
s d n e k e e w n i , . g . e ( n i a g a e v a e l d n a , e n o d s g n i h t e m o s
t o n s i s i h t e s r u o c f o t u b ( e r u t r a p e d y m e r o f e b t i h s i n i f o t
I s a d e s i l a r e n e g e b t o n n a c s g n i d n i f e s e h T . ) s y a d i l o h d n a
. ) s c o d t s o p r o f e r u d e c o r p d e d n e m m o c e r a
e h t n i s e i t i s r e v i n u t n e r e f f i d n i d e y o l p m e n e e b t ' n e v a h
. t h g u o h t r o f d o o f e m o s s e v i g t i t u b , s d n a l r e h t e N
e s r u o c + + C a t h g u a t e v a h I 8 9 9 1 f o r e t s e m e s t s r i f e h t n I
0 5 1 : e c n e i r e p x e n a e t i u q s a w t a h T . s t n e d u t s r a e y t s r i f r o f
m a I ( c p a n o k r o w o t e v a h I t a h t s i e n i m f o e v e e p A
e r i t n e e h T . + + C f o " + + " e h t h c a e t o t d a h I o h w s t n e d u t s
g n i y r t a s i h c i h w , ) S O c a M d n a s i r a l o S n o k r o w o t d e s u
r o t c u r t s n i r e h t o e h t , s t n e d u t s 0 0 3 f o d e t s i s n o c s s a l c
, s t o o b e r y l i a d t s o m l a o t d e n g i s e r w o n m a I . e c n e i r e p x e
s e s s a l c r u o h t o b d n e e h t n I . s t n e d u t s 0 5 1 r e h t o n a t h g u a t
t i t u b , l a c i p y t a
m ' I e b y a M . c t e , s m a r g o r p g n i h s a r c
t n a s a e l p a s a w t a h t o s , e g a r e v a l l a r e v o e m a s e h t d a h
y m n o k r o w h s i n i f o t e m o h o g I t a h t s r u c c o y l t n e u q e r f
s t n e d u t s e h t y b s n o i t a u l a v e ) y r o t a d n a m ( e h T . e m o c t u o
d n a ' y l i s a e ' e n o d s g n i h t t e g o t t s u j , c a M l a n o s r e p
r o f l l a t a d a b t o N . e v i t c u r t s n o c y l t s o m d n a g n i r e t t a l f e r e w
y n a m o s y h w r e d n o w e m s e k a m y l l a e r t I . ' y l k c i u q '
y m ( s e t a u d a r g r e d n u r o f e s r u o c a g n i h c a e t e m i t t s r i f e h t
. . . s ' c p n o k r o w o t e s o o h c e l p o e p
, . g . e , s p u o r g l l a m s g n i h c a e t n i s i e c n e i r e p x e s u o i v e r p
. ) g n i t t e s r a n i m e s a n i s t n e d u t s r a e y h t r u o f
December 1998
197
NVKI-Nieuwsbrief
: e r u t l u c d e s a b e l i b o m o t u a n a s i y r a g l a C . m a d r e t s m A
e r a d a o r b a y a t s s i h t g n i r u d t e g I s e c n e i r e p x e e h T
r o , h g u o n e t n e u q e r f t o n t u b , e l b a l i a v a s i t i s n a r t c i l b u p
d n a h c r a e s e r n o w e i v r e d a o r b a e m e v i g d n a e s r e v i d
e s r u o c f O . l u f e s u y l l a e r t i e k a m o t , h g u o n e e t a l l i t n u
t n e r e f f i d a o t g n i t p a d a , s i h t o t n o i t i d d a n I . s e i t e i c o s
e l c y c i b y l n o , y t i c e h t r e v o l l a s y a w h g i h , s t e e r t s y t p m e
l a c i m o c o t d a e l s a h d n a , n u f s i ) . c t e , s m o t s u c l a c o l
e h t , s t c u d o r p r o f s e m a n d n a r b w e n g n i n r a e l ( e r u t l u c
o Ct d a e l s r a c e s e h t l l a t u b , s r a c g n i s u r o f t p o e l p o e p s e k a m 0 4 s a w o l s a s e r u t a r e p m e t h t i w s r e t n i w g n i v a h
n o d n e p e d y l d r a h I , d e v o r p m i
h s i l g n E h c t u D a
. s g n i d n a t s r e d n u s i m
l e e f n e v e I d n a , y t i c s u o m y n o n a n a s e t a e r c s i h T . s a e r a
e h t , l l a f o t n a t r o p m i t s o m d n A . l l a t a y r a n o i t c i d
a o t t e g o t b r u b u s a h g u o r h t g n i k l a w n e h w ' t c e p s u s '
, k c a b d i a l , d e x a l e r y r e v e r a r a f o s t e m e v ' I s n a i d a n a C
n i e v i l I e t i u s t n e m e s a b e h t , y l e t a n u t r o F . e c a l p r a l u c i t r a p
. d o o g s i e r e h e f i l o s ; e l p o e p s u o r u o m u h
, y t i s r e v i n u f o e c n a t s i d g n i k l a w t a d e t a c o l y l t n e i n e v n o c s i m ' I . b u h t i s n a r t c i l b u p a d n a s l l a m g n i p p o h s f o e l p u o c a
a s i h c i h w , t r o p r i a e h t o t o g o t d e e n I l i t n u . K . O y l l a u s u . . . t i s n a r t c i l b u p a i v e g n e l l a h c
n a i d a n a C e h t f o y t i m i x o r p e h t s i y r a g l a C f o s u l p g i b A
y l e t u l o s b a h t i w e g n a r n i a t n u o m n w o n e r e h t , s e i k c o R
0 9 n i h t i W . s e k a l d n a s n i a t n u o m , s y e l l a v t n e c i f i n g a m
l a n o i t a n e h t f o e n o f o e l d d i m e h t n i e b n a c u o y s e t u n i m
: w o n s s ' e r e h t n e h w d n a ( g n i k i h y n a m e r e h w , s k r a p
: o o z a n i g n i k l a w e k i l s l e e f t I . e l b a l i a v a e r a s l i a r t ) g n i i k s
e l g n i s a n o d e r e t n u o c n e e b n a c e f i l d l i w f o s e i c e p s y n a m
. t e y r a e b a t e m t ' n e v a h I y l e t a n u t r o f e k i h
k r o w t e N h c r a e s e R g n i r e e n i g n E e r a w t f o S
f o ( o h w d n e i r f l r i g y m m o r f y a w a r a f m ' I y l e t a n u t r o f n U
e c n e i c S r e t u p m o C f o t n e m t r a p e D
e k a m s e o d s i h T . n e e v l e t s m A n i b o j r e h d e u n i t n o c ) e s r u o c
y r a g l a C f o y t i s r e v i n U
e n o h p e l e t ( l l e w s a e v i s n e p x e d n a , s e m i t t a t l u c i f f i d e f i l
a d a n a C , y r a g l a C
o s l a s ' t a h T . ) s h t n o m o w t y r e v e h t r o f & k c a b g n i y l f , s l l a c
l n . u v . s c @ k e i n / a c . y r a g l a c u . c s p c @ k e i n
k o o l d n a y r a u n a J n i m a d r e t s m A o t n r u t e r l l ' I n o s a e r e h t
, l a r e n e g n I . s d n a l r e h t e N e h T n i e g n e l l a h c c i f i t n e i c s a r o f
t u o k r o w , D h P u o y h s i n i f ( l l e w e r a p e r p u o y f i t a h t y a s d ' I
g n i k r o w ) . c t e , r e h t o t n a c i f i n g i s r u o y t e e m o t e l u d e h c s a
. d e d n e m m o c e r e b n a c d a o r b a n o i t i s o p c o d t s o p a n i
s d r a a g n j i W k e i N
Section-Editor Radboud Winkels
s a h h s i l g n E y m e s r u o c f O
n a i r t s e d e p o n t s o m l a d n a , e s u l a n o i t a e r c e r r o f s y a w h t a p
PROSA - EEN COMPUTERPROGRAMMA ALS INSTRUCTIEOMGEVING VOOR HET ONDERSTEUNEN VAN HET LEREN OPLOSSEN VAN EEN JURIDISCHE CASUS
SECTION KNOWLEDGE SYSTEMS IN LAW AND COMPUTER SCIENCE
December 1998
198
NVKI-Nieuwsbrief
S E C N E I R E P X E
n i g n i v i l m o r f t n e r e f f i d e t i u q s i y r a g l a C n i g n i v i L ░
Lezing door Antoinette Muntjewerff, Technische Universiteit Twente
(1965), die beschrijft hoe een omgeving dient ingericht te worden om leerprocessen voor een gegeven doel te stimuleren. Al snel bleek deze theorie te globaal om er een afdoende instructietheorie uit te kunnen afleiden en werd gekeken naar Merrill (1983) die ze verder heeft gespecificeerd. Deze laatste verfijnde de instructies om een bepaald leerdoel te bereiken tot op het niveau van een les. Noch de modellen van Gagné als Merrill besteden echter aandacht aan het motivatieaspect binnen het leerproces: daarom werd tenslotte gekeken naar het ARCSmodel van Keller & Suzuki (1988) waarbinnen motivatie wel is opgenomen.
Verlag Raf van Kuyck & Stijn Debaene Katholieke Universiteit Leuven In de rechtspraktijk is het oplossen van juridische casussen een belangrijke bezigheid. Om die reden moet het aanleren van de methodiek hiertoe een plaats vinden in het juridisch onderwijs. Binnen het project PROSA (Probleemsituaties op het terrein van het Administratief procesrecht), dat kadert in het promotieonderzoek van Antoinette Muntjewerff, wordt een computerprogramma ontwikkeld dat moeilijkheden die studenten ondervinden bij het oplossen van juridische casussen moet verhelpen. De toepassing wordt ontwikkeld in het pakket Authorware, een tool voor de ontwikkeling van interactieve educatieve toepassingen met onder andere een goed beheer van gebruikersgegevens en hun vorderingen.
Vervolgens moet de taak van het juridisch casus oplossen worden geanalyseerd om aldus te kunnen specificieren wat wel en wat niet in het instructiemodel op te nemen. Tenslotte werd gekeken naar de moeilijkheden van casus oplossen zelf. Zo werd bijvoorbeeld een vergelijkende studie uitgevoerd waarbij zowel studenten als experten een casus moesten oplossen. Een frappant resultaat was dat experten in een bepaald juridisch domein toch moeilijkheden ondervinden in een ander domein. De hypothese dat de problemen bij het oplossen van casussen vooral voortkomen uit een gebrek aan methode (Crombag) dient dus aangevuld te worden. Zo blijkt naast de methode ook de specifieke domeinkennis of de inhoudelijke ondersteuning van primordiaal belang.
De presentatie gebeurde in twee luiken: eerst werden de ontwerpbeslissingen van het project verduidelijkt en vervolgens werd een korte demonstratie van PROSA gegeven. Een geautomatiseerd leerproject staat of valt met gefundeerde theoretische aannames betreffende leer- en instructietheorie. Daarnaast zijn een analyse van de taak van het juridisch casus oplossen en een inventarisatie van de moeilijkheden hierbij noodzakelijk om de leer- en instructietheorie te specificeren. Vooreerst is uitgegaan van de theorie, opgesteld door Gagné Beide aspecten (methode en inhoudelijke ondersteuning) dienen bijgevolg opgenomen te worden in een systeem dat casus oplossen wil aanleren aan studenten. Daarom werd PROSA gebouwd in twee presentatielagen: de presentatie van de inhoud en de werkvorm enerzijds en de presentatie van de ondersteuning anderzijds. Deze twee lagen zijn aanwezig voor elk van de drie taakonderdelen: de casus, de juridische oplossing en de rechtsregel. Samen resulteert dit in een kleurrijk computerscherm bestaande uit zes onderscheiden vensters.
feiten of het eigenlijke probleem behoort aldus niet tot het systeem. (Bijvoorbeeld: er zitten geen adders onder het gras; een virtueel interview met een denkbeeldige cliënt is niet nodig.) De ondersteuning bij dit taakonderdeel bestaat dan onder andere uit een begrippenlijst, informatie over wat een casus is, het structureren ervan enzovoort. Het tweede taakonderdeel bevat de presentatie van rechtsregels. Binnen dit venster dient de student de naar zijn mening van toepassing zijnde rechtsregel te selecteren. De ondersteuning bij dit taakonderdeel bestaat onder andere uit uitleg betreffende de ordening van rechtsregels, het goed en efficiënt lezen van rechtsregels enzovoort. Door middel van de selectie van delen van de casus en de rechtsregels dient de student in het derde taakonderdeel - 'construeer juridische oplossing' - een oplossing voor de casus te construeren. Dit gebeurt door koppeling ('is' of
Het eerste taakonderdeel bevat de presentatie van de juridische casus. De juridische casussen zijn geordend naar onderwerp binnen het bestuursrecht en binnen elk onderwerp naar moeilijkheidsgraad. De casussen bevatten niet meer en niet minder dan alle gegevens die nodig zijn om tot een oplossing te komen. Ook de vraag horende bij de casus wordt gesteld. Het vinden van
NVKI-Nieuwsbrief
199
December 1998
hard cases in recht en rechtsinformatica.
'is niet') van deze delen. Bij deze taak zijn zowel het proces als het product van belang: ten eerste dient de student één van de drie voorziene routes (waarvan één aanbevolen) bij het oplossen van een casus te volgen (proces) en ten tweede dient hij uiteraard te komen tot een correcte oplossing van de casus, bestaande uit een aantal onderdelen (koppelingen) in een bepaalde volgorde en een antwoord op de gestelde vraag.
Hage heeft ooit beweerd dat als een toepassing van rechtsinformatica gebruikt wordt bij het nemen van een juridische beslissing, deze toepassing aan twee eisen dient te voldoen: het systeem moet betrouwbaar zijn en het moet non-triviale oplossingen bieden. Betrouwbaarheid betekent dat de conclusies die het systeem trekt correct en juridische houdbaar zijn. Non-trivialiteit betekent dat het systeem geen conclusies produceert die de gebruiker zelf ook had kunnen trekken. Men zou ook meer of andere eisen aan een toepassing van rechtsinformatica kunnen stellen. De genoemde eisen vormen echter vaak het uitgangspunt van onderzoek op het gebied van de rechtsinformatica. Een klassiek probleem in de rechtsinformatica is daarom hoe men ervoor kan zorgen dat systemen aan deze eisen kunnen voldoen.
Op elk moment in een PROSA-sessie kan de gebruiker geconfronteerd worden, al dan niet op aanvraag, met feedback. Zo kan de student feedback betreffende het door hem gevolgde proces en het bereikte product opvragen. Het product wordt vergeleken met de normoplossing: zijn alle onderdelen aanwezig en in de juiste volgorde en is het antwoord juist? Globaal houdt PROSA tevens bij welke casussen de student reeds oploste. Ook hieromtrent wordt feedback gegeven, bijvoorbeeld in verband met de graduele stijging van de moeilijkheidsgraad.
Kernprobleem van de rechtsinformatica is de openheid van het recht. Dat wil zeggen: de juridische kennis op basis waarvan conclusies getrokken worden, is niet statisch, maar kan veranderen in de loop van een proces of geschil. In de rechtstheorie spreekt men dan van 'hard cases' of moeilijke gevallen. Voor betrouwbare juridische informatiesystemen heeft men echter juridische kennis nodig die juist wél statisch is. De betrouwbaarheid komt bij hard cases dus in het geding. Als men echter toepassingen zoekt in rechtsgebieden waarop juridische kennis wél statisch is, in de rechtstheorie spreekt men dan van 'clear cases' of eenvoudige gevallen, bestaat het risico dat het systeem triviale conclusies trekt. Het klassieke probleem van de hard en clear cases houdt de rechtsinformatica dan ook al jaren bezig.
Op het einde van de lezing werd nog een korte demonstratie gegeven van PROSA. Een casus van gemiddelde moeilijkheidsgraad werd gedeeltelijk opgelost.
RECHTSINFORMATICA EN HARD CASES Ronald van den Hoogen Recht en Informatisering, UU Verslag van de lezing van Ronald Leenes tijdens de Jurix-vergadering van 23 oktober 1998 op de Universiteit Twente, getiteld: Hercules of Karneades: In de rechtsinformatica heeft men op verschillende manieren geprobeerd dit probleem op te lossen. Meestal wordt daarbij aangesloten bij rechtstheoretische opvattingen over hetgeen 'recht' is. Leenes concludeert dat de klassieke rechtstheoretische opvattingen van Hart of Dworkin, waarin recht als een eindige verzameling regels wordt gezien of juist als een open systeem waarin rechtsbeginselen een rol kunnen spelen, geen afdoende oplossing bieden voor het geschetste probleem. Hij sluit daarom aan bij de dialoogbenadering van Alexy, Aarnio, Peczenik en Lodder. Dat wil zeggen dat recht beschouwd wordt als een dynamisch proces van juridische betekenisgeving dat de vorm heeft van een dialoog. In die dialoog kunnen nieuwe regels en interpretaties van regels worden ingevoerd en vindt discussie plaats over betekenis van feiten. Het onderscheid tussen hard en clear cases is in deze opvatting niet zozeer een
NVKI-Nieuwsbrief
onderscheid in de kenmerken van een probleem, maar in de argumenten die partijen in een dialoog naar voren brengen: hard cases worden gemaakt. De dialogen worden beheerst door juridische 'spelregels'. Het handhaven van die spelregels beschouwt Leenes als een mogelijke toepassing voor een rechtsinformaticasysteem. Deze toepassing wordt een moderator genoemd. Een moderator kan een juridische dialoog begeleiden en waarborgen dat een rationele uitkomst wordt bereikt. Een marginale toetsing van argumenten wordt uitgevoerd in plaats van de lastige inhoudelijke toetsing die traditionele rechtsinformatica-toepassingen uitvoeren. Het ontwikkelen van elektronische moderatoren staat nog in de kinderschoenen en het onderzoek van Leenes kan aan deze ontwikkeling een bijdrage
200
December 1998
leveren. Het onderzoek toont de problemen en de mogelijkheden van moderatoren in het recht. Leenes heeft daartoe in zijn onderzoek een analyse gemaakt van een concrete juridische procedure, de dagvaardingsprocedure voor de rechtbank. Deze analyse werpt een licht op de regels ten aanzien van de zetten van de spelers in het taalspel. Het civiele bewijsrecht speelt daarbij een belangrijke rol. Daarnaast heeft een analyse van een praktijkgeval plaatsgevonden.
die stelt, bewijst' lang niet altijd op gaat. De meest belangwekkende conclusie van Leenes is misschien wel dat de partijen de discussie niet zonder de inbreng van de rechter tot een goed einde kunnen brengen. Een moderator voor de dagvaardingsprocedure zou een marginale toetsing kunnen verrichten van het materiële recht en kunnen controleren 'of er bewijs is en door wie het geleverd is' en 'of het bewijs houdbaar is'. Leenes geeft toe dat hiermee nog altijd niet voldaan is aan de eis van betrouwbaarheid voor rechtsinformaticatoepassingen. Zijn onderzoek levert vooral bouwstenen voor een realistisch model voor moderatoren. Daarnaast beschrijft het de dagvaardingsprocedure op hoofdlijnen, geeft het de aanzet tot de ontwikkelingen van een catalogus voor de bewijslastverdeling en maakt het duidelijk wat de rol van de rechter is en geeft het een beter inzicht in juridische begrippen als claim, erkentenis, bewijsaanbod, betwisten, betwijfelen en herroeping. Verder onderzoek kan wellicht duidelijk maken welke mogelijkheden er zijn voor moderatoren in het recht en wat de belangrijkste problemen zijn die nog opgelost moeten worden.
Uit dit onderzoek is onder andere gebleken dat het er in de juridische praktijk waarin argumenten worden uitgewisseld vaak anders aan toe gaat dan men op basis van de regels in de modellen als die van Lodder en Gordon zou verwachten. Soms blijkt bijvoorbeeld dat de ene partij een groot aantal zetten tegelijkertijd doet en dat de andere partij dan reageert met een groot aantal tegenzetten: niet bepaald de manier waarop een normaal 'spel' verloopt. Uit het praktijkgeval waarin de vraag 'wie moet de eigendom bewijzen van een tenthuisje?' centraal stond, bleek dat de discussie op verschillende niveaus plaatsvindt. Er bleek zowel discussie over de feiten als over het recht en de procedure te bestaan. Daarnaast bleek dat spelers verschillende rollen en bevoegdheden kunnen hebben en dat de hoofdregel van het civiele bewijsrecht 'hij Jammer is het overigens, als ik met een persoonlijke opmerking mag besluiten, dat Leenes geen aandacht besteedt aan het internationale recht. Met name artikel 6 EVRM kan in mijn ogen, ook voor de dagvaardingsprocedure, een belangrijke rol gaan spelen bij de ontwikkeling van toepassingen van de rechtsinformatica omdat dit artikel een interessante dubbelrol speelt in de toepassing en beoordeling van recht in een concreet geval. Het artikel, waarin het recht op een eerlijk proces beschreven wordt, fungeert namelijk als rechtsbeginsel ter aanvulling op het Nederlandse recht, maar ook als rechtsregel van een hogere orde, waaraan het Nederlandse (proces)recht ondergeschikt is. Meestal wordt het internationale recht in het rechtsinformatica-onderzoek achterwege gelaten en worden rechtsbeginselen, als deze al van belang worden geacht, in de eerste betekenis gebruikt, terwijl de juridische ontwikkeling juist meer in de richting van de tweede betekenins wijst.
Artificial Intelligence & Audit Automation Verslag door Arno R. Lodder Computer/Law Institute Vrije Universiteit Amsterdam Tom van Engers is werkzaam bij de Projectorganisatie AI & AA (Artificial Intelligence & Audit Automation) van de Belastingdienst en is daar verantwoordelijk voor het researchbeleid. De Projectorganisatie werd medio jaren tachtig opgericht op het moment dat de AI zich begon af te tekenen als veelbelovende technologie. Het was indertijd de bedoeling dat de Projectorganisatie twee jaar zou blijven bestaan, maar inmiddels is ze toe aan de viering van haar tweede lustrum. De vooralsnog door AI & AA ontwikkelde (kennis)systemen worden door de Belastingdienst-medewerkers geaccepteerd, daadwerkelijk gebruikt en hebben een duidelijke meerwaarde. Het onderwerp van de bijeenkomst was het POWER-project (Programma Ondersteuning Wet En Regelgeving) waar een researchgroep momenteel in samenwerking met onder meer O&I management partners aan werkt. Verder is er op deelterreinen samenwerking met de KUB (Leda en ontologieën) en de UvA (ontologieën).
Leenes, R.E. (1998). Hercules of Karneades; Hard Cases in Recht en Rechtsinformatica. Enschede: Twente University Press.
POWER - PROGRAMMA ONDERSTEUNING WET EN REGELGEVING Tom M. van Engers
NVKI-Nieuwsbrief
201
December 1998
toetsen op de consistentie is een ander doel. Bij de ontwikkeling van het kennissysteem VVV-IBR (Verrekening, Vrijstelling en Verliescompensatie in het Internationaal BelastingRecht, waarbij compensatie voor ondernemingen in verband met in het buitenland betaalde belasting kan worden bepaald, waarmee dubbele belastingheffing wordt voorkomen) heeft het modelleren van de voorgenomen wetgeving geleid tot meer consistentie. De syllabus die als fiscaal-technische achtergrond bij het kennissysteem diende, is als aanschrijving bij de wet opgenomen. Ook het simuleren van de gevolgen van beleidsbeslissingen en een verbeterde voorlichting aan de belastingbetaler streeft men na.
Er zijn verschillende redenen waarom het POWER-project van start is gegaan. In de eerste plaats is ondersteuning bij het maken en uitvoeren van de complexe wet- en regelgeving op het terrein van het belastingrecht gewenst. Een tweede reden is gelegen in het feit dat specialistische kennis vaak beperkt is tot enkele experts. Wanneer deze vertrekken is daarmee ook de betreffende kennis uit de organisatie verdwenen. Als derde kan worden genoemd het streven om misbruik en oneigenlijk gebruik van belastingwetgeving tegen te gaan. Dit is eigenlijk alleen goed mogelijk wanneer de regelgeving voor de belastingbetaler begrijpelijk en voor de Belastingdienst uitvoerbaar is. Tenslotte is het nu vaak ontbrekende inzicht in effecten van gewijzigde regelgeving onontbeerlijk. Voor al de genoemde onderwerpen biedt de technologie, waaronder AI-technieken, mogelijkheden tot verbetering en optimalisering.
Het POWER-project is momenteel nog in het beginstadium. Hierbij is vooral aandacht voor de ondersteuning van de uitvoering van belastingwetgeving. Niet alleen wordt het redeneren met wet- en regelgeving (m.n. uitvoerende dienst) gemodelleerd, ook aan redeneren over wet- en regelgeving (m.n. wetgever) besteedt men aandacht. De vertaling in programmatuur ligt daarmee dicht bij de wetgeving. Bij het redeneren over wet- en regelgeving moeten ondermeer inconsistenties, vaagheden en dergelijke worden opgespoord. Een aardig voorbeeld hiervan, een zogenaamde live lock (cirkel-redenering) treffen we bijvoorbeeld aan in het Loonbelasting-domein: 1. Als tariefgroep is 2 dan is de belastingvrije som de basisaftrek + de bovenbasisaftrek (art. 20 lid 2) voorkomen. Dit is van belang omdat voor een eenduidige uitvoering van wetgeving door de verschillende uitvoerings-organisaties overeenstemming over de betekenis van begrippen nodig is. In eerste instantie was het de bedoeling om mei volgend jaar alle stappen uitgewerkt te hebben waarna de uitvoering van de in het programma opgenomen stappen ter hand zou worden genomen. In het traject tot dit programma was voorzien in een beperkte proefneming waarbij een aantal stappen in een beperkt toepassingsdomein zouden worden genomen (stap 1 genereren standaardspecificaties; stap 2 toetsen consistenties; stap 3 uitvoeren simulaties, etc.). Echter, het POWER-project is als proefneming binnen het kader van de wetgeving 21ste eeuw gepositioneerd en daarom zal proefneming nu plaatsvinden op een onderwerp binnen deze nieuwe wetgeving. Het betekent ook dat delen al geoperationaliseerd moeten worden nog voordat alle stappen volledig zijn uitgewerkt.
De doelstellingen van het POWER-project zijn divers. Zo wil men trachten een transparante vertaling van wet- en regelgeving naar beslissingsondersteunende (kennis)applicaties te bewerkstelligen. Op dit moment kunnen specificaties voor dit type geautomatiseerde hulpmiddelen niet altijd op eenvoudige wijze worden terug herleid tot de bron in wet- en regelgeving. Daarom kan het gebeuren dat de grond van een beslissing niet direct duidelijk is. Ook is bij aanpassing van wet- en regelgeving dan moeilijk in te schatten op welke plaatsen wijzigingen nodig zijn. De wet- en regelgeving 2. In tariefgroep 2 wordt ingedeeld degene die de basisaftrek en de bovenbasisaftrek geniet (art. 22 lid 1). Als de methode zoals deze in POWER wordt ontwikkeld werkt, dan kan de vigerende vertaalsystematiek aanzienlijk worden vereenvoudigd. De aandacht voor het optimaliseren van de kennisinfrastructuur is gegeven deze doelstelling (kennismanagement) niet verwonderlijk. Als punten voor discussie werd naar voren gebracht: traject van informeel naar semi-formeel; traject van semi-formeel naar formeel; gegevensstandaardisatie en (her)gebruik - bv. de term pensioen heeft verschillende betekenissen binnen de sociale zekerheids- en de fiscale wetgeving; jurisprudentie. In de discussie werden deze en andere onderwerpen aan de orde gesteld. Ik zal er hieronder enkele uitlichten. Onder standaardspecificatie wordt begrepen het definiëren en beschrijven van begrippen die in de wet
NVKI-Nieuwsbrief
202
December 1998
R E B M U N
Y T I S R E V I N U D N A M A R G O R P
4 1
a c i t a m r o f n I
Een heftige discussie had als onderwerp de jurisprudentie. Als inzage in beleid en eerdere beslissingen van de Belastingdienst wordt gevraagd wordt dit nochtans door de Belastingdienst geweigerd met als enige argument dat ze zelf de beschikking over die informatie ook niet hebben. Wanneer door POWER deze informatie voor de medewerkers toegankelijk wordt gemaakt gaat het door de Belastingdienst gebruikte argument niet meer op. De vraag was of er problemen te verwachten zijn wanneer de jurisprudentie geautomatiseerd wordt opgeslagen. Hoewel Van Engers niet kan overzien of er in dat opzicht problemen zijn te verwachten, hoopt hij dat dit stadium in ieder geval bereikt wordt.
m a d r e t s m A n a V t i e t i s r e v i n U 0 2
e i t n e g i l l e t n I e g i t a m t s n u K m a d r e t s m A n a V t i e t i s r e v i n U
3 2
a c i t a m r o f n I t i e t i s r e v i n U e j i r V
6
e i t n e g i l l e t n I e g i t a m t s n u K t i e t i s r e v i n U e j i r V
0 5
a c i t a m r o f n I t h c e r t U t i e t i s r e v i n U
7 1
e i t n e g i l l e t n I e g i t a m t s n u K e v e i t i n g o C t h c e r t U t i e t i s r e v i n U
Het POWER-project is ambitieus. Het richt zich niet alleen op ondersteuning van uitvoerders van wetgeving, maar ook op makers ervan. Verder wordt de kennisinfrastructuur geoptimaliseerd en gehanteerde begrippen gestandaardiseerd. Wat het project voor rechtsinformatici zo interessant maakt is dat grootschalig diverse onderzoeksonderwerpen (kennissystemen, kennisrepresentatie, ontologieën) praktisch worden verwerkt en ingezet ter facilitering van juristen. Momenteel is een eerste stap gezet op een op zich veelbelovende weg die uiteindelijk moet leiden tot een optimale inzet van AI en IT ten behoeve van juristen, oftewel rechtsinformatica in optima forma. Om met de woorden van de spreker te spreken: 'Laten we hopen dat we daar uitkomen!'.
1 1
e i g o l o n h c e t s i n n e K t h c i r t s a a M t i e t i s r e v i n U
4 1
a c i t a m r o f n I n e g e m j i N t i e t i s r e v i n U e k e i l o h t a K
9 2
a c i t a m r o f n I n e g n i n o r G t i e t i s r e v i n u s k j i R
1 1
a c i t a m r o f n I n e d i e L t i e t i s r e v i n U
END OF SECTION CALL FOR PARTICIPATION ARTIFICIAL INTELLIGENCE AND BEYOND
KNOWLEDGE SYSTEMS IN LAW AND COMPUTER SCIENCE
a joint program by Flanders Language Valley Education and K.U.Leuven Campus Kortrijk Faculty of Science Monthly : November - May 1998 - 1999 OBJECTIVES
E H T N I S T N E D U T S I A W E N
Artificial Intelligence - the science that builds intelligent agents and artifacts - has developed a number of exciting new techniques and applications over the last decade. The FLV-KULAK seminars will give a survey of these achievements. Participants meet in half-day seminars, each of which focuses on one sub-area of artificial intelligence. The areas include genetic algorithms, neural networks, intelligent software agents, mainstream artificial intelligence and evolving
S D N A L R E H T E N a m t s o P c i r E d n a s n o m e l l e H e k o J . 8 9 9 1 , 1 1 r e b m e c e D r e p e l b a t d e t a d p U NVKI-Nieuwsbrief
203
December 1998
hardware. Each seminar will address both theory and applications and will be taught by leading AI researchers and practitioners from Belgium and abroad. Therefore the series should be of interest to both industry and academia.
Starlab, Riverland, Zaventem Module 4 - 03.02.1999 : Neural Networks By Prof. Dr. Joos Vandewalle, Kuleuven Module 5 - 03.03.1999 : Genetic Algorithms By Manderick, VUB
TARGET AUDIENCE This series is aimed at engineers, computer scientists, linguists, R&D managers and other scientists interested in state-of-the-art artificial intelligence and its applications. It is also recommended to post-graduate and doctoral students in AI-related fields.
Prof.
Dr.
Bernard
Module 6 - 14.04.1999 : Introduction To Artificial Life And Swarm Intelligence By Prof. Dr. Marco Dorigo, ULB Module 7 - 05.05.1999 : Morning Seminar : Virtual And Augmented Reality As Integrator of Knowledge And Information Technology By Prof. Dr. Fernand Van Damme, Bikit and UGent
PROGRAMME See : http://www.kulak.ac.be/facult/wet/flv-kulak or : http://www.flv.be for the abstracts
Module 8 (A + B) - 05.05.1999 : Afternoon Seminar : Evolving Hardware And Cam-brain Machine By Dr. Michael G. Korkin, Genobyte Inc., Boulder Co, USA The Age Of Spiritual Machines By Dr. Ray Kurzweil, Kurzweil Educational Systems, Waltham, USA
Module 1 - 25.11.1998 : Opening Seminar : Welcome by Mr. Jo Lernout, Flv and Rector Marcel Joniau, Kulak Artificial Intelligence : A New Step Forward ? By Prof. Dr. Luc Steels, VUB Introduction to the Seminars - Lecturers' Panel Module 2 - 09.12.1998 : Progress In Traditional Artificial Intelligence : Machine Learning By Prof. Dr. Luc De Raedt, Kuleuven
PROJECT COMMITTEE * prof. dr. Marcel JONIAU, Rector K.U.Leuven Campus Kortrijk * dr. ir. Dirk FRIMOUT, Astronaut STS45, Chairman FLV Education, Ieper Location : Conferentiezaal - Stadhuis Ieper Grote Markt - B-8900 Ieper Phone : +32(0)57-22.85.62
Module 3 - 06.01.1999 : Software Agents : The New Future of AI ? By Dr. Walter van de Velde, * prof. dr. Lea VERMEIRE, K.U.Leuven Campus Kortrijk * prof. dr. Luc DE RAEDT, K.U.Leuven and Fund for Scientific Research Flanders * mr. Patrick MOESICK, Manager Delegate, FLV Education, Ieper * mrs. Virginie COUCKE, Staff Office of the Rector, K.U.Leuven Campus Kortrijk * mr. Jos VERNIEST, PR manager, FLV, Ieper
Seminar Syllabus : An outline will be handed out at the beginning of each seminar. A final seminar syllabus will be available on request.
Timing and Dates : 8 half days : 6 afternoons and a full day, each on Wednesday - afternoon : from 2.30 to 5 pm, with coffee break - closing day : from 10 am to 6 pm, with lunch November 25 and december 9, 1998, January 6, February 3, March 3, April 14 and May 5, 1999
Fee : * Professional : full series (8 modules) = 30.000 BEF; 3 modules (min. registration) = 15.000 BEF; per additional module = 4.500 BEF * Educational staff : full series (8 modules) = 10.000 BEF; 3 modules (min. registration) = 5.000 BEF; per additional module = 1.500 BEF
* Students : full series (8 modules) = 6.600 BEF; 3
modules (min.
PRACTICAL INFORMATION
NVKI-Nieuwsbrief
204
December 1998
n n a s e / e b . c a . l c u . e c i d . w w w / / : p t t h : n o i t a m r o f n I
registration) = 3.300 BEF; per additional module = 1.100 BEF
26-28 april 1999 COORDINATION ’99, Third International Conference on Coordination Models and Language, Amsterdam, The Netherlands. Information: http://www.cs.unibo.it/~coord99
Fee includes participation, outlines and coffee. Lunch at the closing day is optional at 700 BEF. Bank account : 552-2959901-91, FLV Education, B-8900 Ieper. Please add : A.I. Seminars.
25-27 mei 1999 International Conference on Computational Intelligence, Dortmund, Germany. Information: http://ls1-www.cs.uni-dortmund.de/fd6
REGISTRATION
31 mei-3 juni 1999 IEA/AIE-99, The Twelfth International Conference on Industrial & Engineering Applications of Artificial Intelligence & Expert Systems, Cairo, Egypt. Information: Dr. Moonis Ali, E-mail:
[email protected] and Dr. Ibrahim Imam, E-mail:
[email protected]
You can register at these websites : http://www.kulak.ac.be/facult/wet/flv-kulak http://www.flv.be
1-4 juni 1999 IIA’99, Intelligent Industrial Automation and SOCO’99, Soft Computing, Palazzo Ducale, Genova, Italy. Information: http://www.ixsc.ab.ca/iia99.htm
CONFERENTIES,
Hieronder volgt een lijst van data van conferenties SYMPOSIA, en een contactadres. Graag wijzen we onze lezers WORKSHOPS tevens op de aanvullende Calendar 1998, zoals die gepubliceerd wordt in de AI Communications. Voorts hebben we referenties aan SIGART Newsletter ontleend.
2-4 juni 1999 VISUAL99, Third International Conference on Information Systems, Amsterdam, The Netherlands. Information: http://www.wins.uva.nl/events/VISual99
Visual
15-17 juni 1999 ASCI 1999 Conference, Heijen, The Netherlands. Information: http://www.asci.tudelft.nl
25-29 januari 1999 SNN, Stichting Neurale Netwerken Advanced Issues in Neurocomputing Course Information: http://www.wins.uva.nl/ ~krose/asci_nn.html
14-18 juni 1999 ICAIL-99, Seventh International Conference on Artificial Intelligence and Law, University of Oslo, Norway. Information: Program Chair: Mr. Thomas Gordon, E-mail:
[email protected].
17-19 februari 1999 CIMCA’99, Computational Intelligence for Modelling, Control and Auromation, Vienna, Austria. Information: http://www-gscit.fcit.monash.edu.au/conferences/cimca99
22-25 juni 1999 CIMA’99, International ICSC Congress on Computational Intelligence: Methods and Applications, Rochester Institute of Technology, NY, USA. Information: http://www.icsc.ab.ca/cima99.htm
25-27 maart 1999 DGNMR’99, Fourth Dutch-German Workshop on Nonmonotonic Reasoning Techniques and Their Applications Institute of Logic, Language and Information, University of Amsterdam Information: http://pgs.twi.delft.nl/~witt/dgnmr99.htm
18-22 juli 1999 AAAI-99, Sixteenth National Conference on Artificial Intelligence, Orlando, Florida. Information: http://www.aaai.org/Conferences/National/1999
12-14 april 1999 HPCN Europe’99. The 7th International conference on High Performance Computing and Networking Europe. Information: http://www.wins.uva.nl/events/HPCN99
30 juli - 1 augustus 1999 UAI99, the Fifteenth Annual Conference on Uncertainty in Artificial Intelligence, Sweden. Information: http://uai99.iet.com
19-23 april 1999 PA, EXPO 99, The Commonwealth Conference and Events Centre, London, UK Information: http://www.commonwealth.org.uk/ 9 9 9 1 l i r p a 3 2 1 2
16-20 augustus 1999 ESSLLI-99, Eleventh European School in Logic, Language and Information, Utrecht, The Netherlands Information: http://www.wins.uva.nl/research/folli/
s k r o w t e N l a r u e N l a i c i f i t r A n o m u i s o p m y S n a e p o r u E h t 7 , 9 9 ' N N A S E . m u i g l e B , s e g u r B
13-17 september 1999
ECAL99, 5th European Conference on Artificial Life, Swiss Federal Institute of Technology in Lausanne (EPFL), Switzerland Information: http://www.epfl.ch/ecal99
Second World Conference on New Trends in Criminal Investigation, Amsterdam. Information: http://www.eurocongres.com/criminallaw
10-15 december 1999 Workshop on Artificial Intelligence and Judicial Proof at the
NVKI-Nieuwsbrief
205
December 1998
(EMAIL)ADRESSEN BESTUURSLEDEN NVKI
REDACTIE NVKI-NIEUWSBRIEF Dr. E.O. Postma (hoofdredacteur) (Zie adressen bestuursleden)
Prof.dr. J. N. Kok Wiskunde en Natuurwetenschappen, Dept. of Computer Science Rijksuniversiteit Leiden, Niels Bohrweg 1, 2333 CA Leiden Tel: (071) 5277057, E-mail:
[email protected]
Prof. dr. H.J. van den Herik Universiteit Maastricht, Vakgroep Informatica, Postbus 616, 6200 MD Maastricht Tel.: (043) 3883485, E-mail:
[email protected]
Dr. Y.H. Tan EURIDIS, Erasmus Universiteit Rotterdam Postbus 1738, 3000 DR Rotterdam Tel.: (010) 4082255. E-mail:
[email protected]
Dr. C. Witteveen Technische Universiteit Delft, Vakgroep Informatica, Julianalaan 132, 2628 BL Delft Tel.: (015) 2782521, E-mail:
[email protected]
Dr. E.O. Postma Universiteit Maastricht, Department of Computer Science Postbus 616, 6200 MD Maastricht Tel.: (043) 3883493. E-mail:
[email protected] Dr. R. Verbrugge Rijksuniversiteit Groningen, Cognitive Science Engineering Grote Kruisstraat 2/1, 9712 TS Groningen Tel.: (050) 3636334. E-mail:
[email protected]
Dr. R.G.F. Winkels Universiteit van Amsterdam, Vakgroep Rechtsinformatica Postbus 1030, 1000 BA Amsterdam Tel.: (020) 5253485, E-mail:
[email protected]
and
Dr. S.-H. Nienhuys-Cheng Erasmus Universiteit Rotterdam, Vakgroep Informatica Postbus 1738, 3000 DR Rotterdam Tel.: (010) 4081345, E-mail:
[email protected]
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]
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]
Dr. L. de Raedt Department of Computer Science, Katholieke Universiteit Leuven, Celestijnenlaan 200A, B-3001 Heverlee, België Tel.: +32 16 327643. E-mail:
[email protected]
Dr. A. van den Bosch Katholieke Universiteit Brabant, Vakgroep Taal- en Literatuurwetenschap, Postbus 90153, 5000 LE Tilburg Tel.: (013) 4360911, E-mail:
[email protected]
Dr. G.J. Beijer BOLESIAN BV, Steenovenweg 1, 5708 HN Helmond Tel.: (0492) 502525. E-mail:
[email protected] Dr. W. Daelemans Katholieke Universiteit Brabant, Vakgroep TaalLiteratuurwetenscha, Postbus 90153, 5000 LE Tilburg Tel.: (013) 4663070 E-mail:
[email protected]
Technische
Drs. B. de Boer Vrije Universiteit Brussel, AI-Lab Pleinlaan 2, B-1050 Brussel, België Tel.: +32 (0)2 6293703, E-mail:
[email protected]
en
HOE WORD IK LID?
ADVERTENTIES
Het lidmaatschap van de NVKI kost in 1998 fl. 75,- voor gewone leden; fl. 50,- voor AIO's en fl. 40,- voor studenten. Als onderdeel van het NVKI-lidmaatschap zult u in 1998 twee maal het Europese Tijdschrift AI Communications ontvangen, dat beschouwd kan worden als pendant van AI Magazine. Tevens ontvangt u zesmaal per jaar een NVKI-Nieuwsbrief met informatie over conferenties, onderzoeksprojecten, onderzoeksplaatsen, subsidie mogelijkheden, etc., tenminste als er voldoende informatie wordt aangeleverd. Daarom worden alle leden opgeroepen om nieuws en nieuwtjes die zij de moeite waard vinden, op te sturen aan de redactie van de NVKI-Nieuwsbrief. U kunt (uitsluitend) lid worden door overmaking van fl. 75,- (fl. 50,-, fl. 40,-) op RABO-Bankrekeningnummer 11.66.34.200 of op Postgironummer 3102697, ten name van NVKI, FdAW, Vakgroep Informatica, Postbus 616, 6200 MD Maastricht.. Pas als uw betaling ontvangen is, wordt met verzending van de NVKI-Nieuwsbrief begonnen. Indien u uw lidmaatschap wenst te beëindigen dient u dit schriftelijk door te geven aan het redactie-secretariaat vóór 1 december 1998.
Het is mogelijk uw advertentie in de NVKI-Nieuwsbrief te laten opnemen. Voor informatie over prijzen e.d. kunt u contact opnemen met het redactie-secretariaat.
ADRESWIJZIGINGEN De verzending van de NVKI-Nieuwsbrief vindt plaats vanuit Maastricht. Daarom heeft het NVKI-bestuur besloten dat bij het NVKI-Nieuwsbrief secretariaat ook de NVKIledenadministratie bijgehouden wordt. U dient derhalve uw adreswijziging door te geven aan: Redactiesecretariaat NVKI-Nieuwsbrief Universiteit Maastricht, Vakgroep Informatica Postbus 616, 6200 MD Maastricht Tel.: 043-388 3477, E-mail:
[email protected]
KOPIJ De redactie zet haar kolommen ook open voor productaankondigingen, boekbesprekingen, productbesprekingen, overviews van AI-onderzoek in het bedrijfsleven, reviews van nieuwe AI-ontwikkelingen en interviews. Hierbij kunnen natuurlijk controversiële meningen aan het daglicht treden. Om een goed beeld van de ontwikkeling te krijgen is dit geen bezwaar, sterker nog wij moedigen discussies zelfs aan. Liefst ontvangen wij uw kopij per e-mail, of op diskette 3,50" (WP 7.0 of ascii-tekst plus een hard copy).
NVKI-Nieuwsbrief
206
December 1998
advertentiepagina bOLESIAN
NVKI-Nieuwsbrief
207
December 1998
